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/interrupt.h> 33 #include <linux/signal.h> 34 #include <linux/completion.h> 35 #include <linux/workqueue.h> 36 #include <linux/slab.h> 37 #include <linux/cpu.h> 38 #include <linux/notifier.h> 39 #include <linux/kthread.h> 40 #include <linux/hardirq.h> 41 #include <linux/mempolicy.h> 42 #include <linux/freezer.h> 43 #include <linux/debug_locks.h> 44 #include <linux/lockdep.h> 45 #include <linux/idr.h> 46 #include <linux/jhash.h> 47 #include <linux/hashtable.h> 48 #include <linux/rculist.h> 49 #include <linux/nodemask.h> 50 #include <linux/moduleparam.h> 51 #include <linux/uaccess.h> 52 #include <linux/sched/isolation.h> 53 #include <linux/sched/debug.h> 54 #include <linux/nmi.h> 55 #include <linux/kvm_para.h> 56 #include <linux/delay.h> 57 #include <linux/irq_work.h> 58 59 #include "workqueue_internal.h" 60 61 enum worker_pool_flags { 62 /* 63 * worker_pool flags 64 * 65 * A bound pool is either associated or disassociated with its CPU. 66 * While associated (!DISASSOCIATED), all workers are bound to the 67 * CPU and none has %WORKER_UNBOUND set and concurrency management 68 * is in effect. 69 * 70 * While DISASSOCIATED, the cpu may be offline and all workers have 71 * %WORKER_UNBOUND set and concurrency management disabled, and may 72 * be executing on any CPU. The pool behaves as an unbound one. 73 * 74 * Note that DISASSOCIATED should be flipped only while holding 75 * wq_pool_attach_mutex to avoid changing binding state while 76 * worker_attach_to_pool() is in progress. 77 * 78 * As there can only be one concurrent BH execution context per CPU, a 79 * BH pool is per-CPU and always DISASSOCIATED. 80 */ 81 POOL_BH = 1 << 0, /* is a BH pool */ 82 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */ 83 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 84 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */ 85 }; 86 87 enum worker_flags { 88 /* worker flags */ 89 WORKER_DIE = 1 << 1, /* die die die */ 90 WORKER_IDLE = 1 << 2, /* is idle */ 91 WORKER_PREP = 1 << 3, /* preparing to run works */ 92 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 93 WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 94 WORKER_REBOUND = 1 << 8, /* worker was rebound */ 95 96 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 97 WORKER_UNBOUND | WORKER_REBOUND, 98 }; 99 100 enum work_cancel_flags { 101 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */ 102 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */ 103 }; 104 105 enum wq_internal_consts { 106 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 107 108 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 109 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 110 111 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 112 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 113 114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 115 /* call for help after 10ms 116 (min two ticks) */ 117 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 118 CREATE_COOLDOWN = HZ, /* time to breath after fail */ 119 120 /* 121 * Rescue workers are used only on emergencies and shared by 122 * all cpus. Give MIN_NICE. 123 */ 124 RESCUER_NICE_LEVEL = MIN_NICE, 125 HIGHPRI_NICE_LEVEL = MIN_NICE, 126 127 WQ_NAME_LEN = 32, 128 WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */ 129 }; 130 131 /* 132 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and 133 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because 134 * msecs_to_jiffies() can't be an initializer. 135 */ 136 #define BH_WORKER_JIFFIES msecs_to_jiffies(2) 137 #define BH_WORKER_RESTARTS 10 138 139 /* 140 * Structure fields follow one of the following exclusion rules. 141 * 142 * I: Modifiable by initialization/destruction paths and read-only for 143 * everyone else. 144 * 145 * P: Preemption protected. Disabling preemption is enough and should 146 * only be modified and accessed from the local cpu. 147 * 148 * L: pool->lock protected. Access with pool->lock held. 149 * 150 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for 151 * reads. 152 * 153 * K: Only modified by worker while holding pool->lock. Can be safely read by 154 * self, while holding pool->lock or from IRQ context if %current is the 155 * kworker. 156 * 157 * S: Only modified by worker self. 158 * 159 * A: wq_pool_attach_mutex protected. 160 * 161 * PL: wq_pool_mutex protected. 162 * 163 * PR: wq_pool_mutex protected for writes. RCU protected for reads. 164 * 165 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 166 * 167 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 168 * RCU for reads. 169 * 170 * WQ: wq->mutex protected. 171 * 172 * WR: wq->mutex protected for writes. RCU protected for reads. 173 * 174 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read 175 * with READ_ONCE() without locking. 176 * 177 * MD: wq_mayday_lock protected. 178 * 179 * WD: Used internally by the watchdog. 180 */ 181 182 /* struct worker is defined in workqueue_internal.h */ 183 184 struct worker_pool { 185 raw_spinlock_t lock; /* the pool lock */ 186 int cpu; /* I: the associated cpu */ 187 int node; /* I: the associated node ID */ 188 int id; /* I: pool ID */ 189 unsigned int flags; /* L: flags */ 190 191 unsigned long watchdog_ts; /* L: watchdog timestamp */ 192 bool cpu_stall; /* WD: stalled cpu bound pool */ 193 194 /* 195 * The counter is incremented in a process context on the associated CPU 196 * w/ preemption disabled, and decremented or reset in the same context 197 * but w/ pool->lock held. The readers grab pool->lock and are 198 * guaranteed to see if the counter reached zero. 199 */ 200 int nr_running; 201 202 struct list_head worklist; /* L: list of pending works */ 203 204 int nr_workers; /* L: total number of workers */ 205 int nr_idle; /* L: currently idle workers */ 206 207 struct list_head idle_list; /* L: list of idle workers */ 208 struct timer_list idle_timer; /* L: worker idle timeout */ 209 struct work_struct idle_cull_work; /* L: worker idle cleanup */ 210 211 struct timer_list mayday_timer; /* L: SOS timer for workers */ 212 213 /* a workers is either on busy_hash or idle_list, or the manager */ 214 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 215 /* L: hash of busy workers */ 216 217 struct worker *manager; /* L: purely informational */ 218 struct list_head workers; /* A: attached workers */ 219 220 struct ida worker_ida; /* worker IDs for task name */ 221 222 struct workqueue_attrs *attrs; /* I: worker attributes */ 223 struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 224 int refcnt; /* PL: refcnt for unbound pools */ 225 226 /* 227 * Destruction of pool is RCU protected to allow dereferences 228 * from get_work_pool(). 229 */ 230 struct rcu_head rcu; 231 }; 232 233 /* 234 * Per-pool_workqueue statistics. These can be monitored using 235 * tools/workqueue/wq_monitor.py. 236 */ 237 enum pool_workqueue_stats { 238 PWQ_STAT_STARTED, /* work items started execution */ 239 PWQ_STAT_COMPLETED, /* work items completed execution */ 240 PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 242 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 243 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ 244 PWQ_STAT_MAYDAY, /* maydays to rescuer */ 245 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 246 247 PWQ_NR_STATS, 248 }; 249 250 /* 251 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT 252 * of work_struct->data are used for flags and the remaining high bits 253 * point to the pwq; thus, pwqs need to be aligned at two's power of the 254 * number of flag bits. 255 */ 256 struct pool_workqueue { 257 struct worker_pool *pool; /* I: the associated pool */ 258 struct workqueue_struct *wq; /* I: the owning workqueue */ 259 int work_color; /* L: current color */ 260 int flush_color; /* L: flushing color */ 261 int refcnt; /* L: reference count */ 262 int nr_in_flight[WORK_NR_COLORS]; 263 /* L: nr of in_flight works */ 264 bool plugged; /* L: execution suspended */ 265 266 /* 267 * nr_active management and WORK_STRUCT_INACTIVE: 268 * 269 * When pwq->nr_active >= max_active, new work item is queued to 270 * pwq->inactive_works instead of pool->worklist and marked with 271 * WORK_STRUCT_INACTIVE. 272 * 273 * All work items marked with WORK_STRUCT_INACTIVE do not participate in 274 * nr_active and all work items in pwq->inactive_works are marked with 275 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are 276 * in pwq->inactive_works. Some of them are ready to run in 277 * pool->worklist or worker->scheduled. Those work itmes are only struct 278 * wq_barrier which is used for flush_work() and should not participate 279 * in nr_active. For non-barrier work item, it is marked with 280 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 281 */ 282 int nr_active; /* L: nr of active works */ 283 struct list_head inactive_works; /* L: inactive works */ 284 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */ 285 struct list_head pwqs_node; /* WR: node on wq->pwqs */ 286 struct list_head mayday_node; /* MD: node on wq->maydays */ 287 288 u64 stats[PWQ_NR_STATS]; 289 290 /* 291 * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 292 * and pwq_release_workfn() for details. pool_workqueue itself is also 293 * RCU protected so that the first pwq can be determined without 294 * grabbing wq->mutex. 295 */ 296 struct kthread_work release_work; 297 struct rcu_head rcu; 298 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); 299 300 /* 301 * Structure used to wait for workqueue flush. 302 */ 303 struct wq_flusher { 304 struct list_head list; /* WQ: list of flushers */ 305 int flush_color; /* WQ: flush color waiting for */ 306 struct completion done; /* flush completion */ 307 }; 308 309 struct wq_device; 310 311 /* 312 * Unlike in a per-cpu workqueue where max_active limits its concurrency level 313 * on each CPU, in an unbound workqueue, max_active applies to the whole system. 314 * As sharing a single nr_active across multiple sockets can be very expensive, 315 * the counting and enforcement is per NUMA node. 316 * 317 * The following struct is used to enforce per-node max_active. When a pwq wants 318 * to start executing a work item, it should increment ->nr using 319 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over 320 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish 321 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in 322 * round-robin order. 323 */ 324 struct wq_node_nr_active { 325 int max; /* per-node max_active */ 326 atomic_t nr; /* per-node nr_active */ 327 raw_spinlock_t lock; /* nests inside pool locks */ 328 struct list_head pending_pwqs; /* LN: pwqs with inactive works */ 329 }; 330 331 /* 332 * The externally visible workqueue. It relays the issued work items to 333 * the appropriate worker_pool through its pool_workqueues. 334 */ 335 struct workqueue_struct { 336 struct list_head pwqs; /* WR: all pwqs of this wq */ 337 struct list_head list; /* PR: list of all workqueues */ 338 339 struct mutex mutex; /* protects this wq */ 340 int work_color; /* WQ: current work color */ 341 int flush_color; /* WQ: current flush color */ 342 atomic_t nr_pwqs_to_flush; /* flush in progress */ 343 struct wq_flusher *first_flusher; /* WQ: first flusher */ 344 struct list_head flusher_queue; /* WQ: flush waiters */ 345 struct list_head flusher_overflow; /* WQ: flush overflow list */ 346 347 struct list_head maydays; /* MD: pwqs requesting rescue */ 348 struct worker *rescuer; /* MD: rescue worker */ 349 350 int nr_drainers; /* WQ: drain in progress */ 351 352 /* See alloc_workqueue() function comment for info on min/max_active */ 353 int max_active; /* WO: max active works */ 354 int min_active; /* WO: min active works */ 355 int saved_max_active; /* WQ: saved max_active */ 356 int saved_min_active; /* WQ: saved min_active */ 357 358 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 359 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */ 360 361 #ifdef CONFIG_SYSFS 362 struct wq_device *wq_dev; /* I: for sysfs interface */ 363 #endif 364 #ifdef CONFIG_LOCKDEP 365 char *lock_name; 366 struct lock_class_key key; 367 struct lockdep_map __lockdep_map; 368 struct lockdep_map *lockdep_map; 369 #endif 370 char name[WQ_NAME_LEN]; /* I: workqueue name */ 371 372 /* 373 * Destruction of workqueue_struct is RCU protected to allow walking 374 * the workqueues list without grabbing wq_pool_mutex. 375 * This is used to dump all workqueues from sysrq. 376 */ 377 struct rcu_head rcu; 378 379 /* hot fields used during command issue, aligned to cacheline */ 380 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 381 struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */ 382 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */ 383 }; 384 385 /* 386 * Each pod type describes how CPUs should be grouped for unbound workqueues. 387 * See the comment above workqueue_attrs->affn_scope. 388 */ 389 struct wq_pod_type { 390 int nr_pods; /* number of pods */ 391 cpumask_var_t *pod_cpus; /* pod -> cpus */ 392 int *pod_node; /* pod -> node */ 393 int *cpu_pod; /* cpu -> pod */ 394 }; 395 396 struct work_offq_data { 397 u32 pool_id; 398 u32 disable; 399 u32 flags; 400 }; 401 402 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { 403 [WQ_AFFN_DFL] = "default", 404 [WQ_AFFN_CPU] = "cpu", 405 [WQ_AFFN_SMT] = "smt", 406 [WQ_AFFN_CACHE] = "cache", 407 [WQ_AFFN_NUMA] = "numa", 408 [WQ_AFFN_SYSTEM] = "system", 409 }; 410 411 /* 412 * Per-cpu work items which run for longer than the following threshold are 413 * automatically considered CPU intensive and excluded from concurrency 414 * management to prevent them from noticeably delaying other per-cpu work items. 415 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 416 * The actual value is initialized in wq_cpu_intensive_thresh_init(). 417 */ 418 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 419 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 420 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 421 static unsigned int wq_cpu_intensive_warning_thresh = 4; 422 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644); 423 #endif 424 425 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 426 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 427 module_param_named(power_efficient, wq_power_efficient, bool, 0444); 428 429 static bool wq_online; /* can kworkers be created yet? */ 430 static bool wq_topo_initialized __read_mostly = false; 431 432 static struct kmem_cache *pwq_cache; 433 434 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; 435 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; 436 437 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 438 static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf; 439 440 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 441 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 442 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 443 /* wait for manager to go away */ 444 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 445 446 static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 447 static bool workqueue_freezing; /* PL: have wqs started freezing? */ 448 449 /* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */ 450 static cpumask_var_t wq_online_cpumask; 451 452 /* PL&A: allowable cpus for unbound wqs and work items */ 453 static cpumask_var_t wq_unbound_cpumask; 454 455 /* PL: user requested unbound cpumask via sysfs */ 456 static cpumask_var_t wq_requested_unbound_cpumask; 457 458 /* PL: isolated cpumask to be excluded from unbound cpumask */ 459 static cpumask_var_t wq_isolated_cpumask; 460 461 /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 462 static struct cpumask wq_cmdline_cpumask __initdata; 463 464 /* CPU where unbound work was last round robin scheduled from this CPU */ 465 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 466 467 /* 468 * Local execution of unbound work items is no longer guaranteed. The 469 * following always forces round-robin CPU selection on unbound work items 470 * to uncover usages which depend on it. 471 */ 472 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 473 static bool wq_debug_force_rr_cpu = true; 474 #else 475 static bool wq_debug_force_rr_cpu = false; 476 #endif 477 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 478 479 /* to raise softirq for the BH worker pools on other CPUs */ 480 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works); 481 482 /* the BH worker pools */ 483 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools); 484 485 /* the per-cpu worker pools */ 486 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 487 488 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 489 490 /* PL: hash of all unbound pools keyed by pool->attrs */ 491 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 492 493 /* I: attributes used when instantiating standard unbound pools on demand */ 494 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 495 496 /* I: attributes used when instantiating ordered pools on demand */ 497 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 498 499 /* 500 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 501 * process context while holding a pool lock. Bounce to a dedicated kthread 502 * worker to avoid A-A deadlocks. 503 */ 504 static struct kthread_worker *pwq_release_worker __ro_after_init; 505 506 struct workqueue_struct *system_wq __ro_after_init; 507 EXPORT_SYMBOL(system_wq); 508 struct workqueue_struct *system_percpu_wq __ro_after_init; 509 EXPORT_SYMBOL(system_percpu_wq); 510 struct workqueue_struct *system_highpri_wq __ro_after_init; 511 EXPORT_SYMBOL_GPL(system_highpri_wq); 512 struct workqueue_struct *system_long_wq __ro_after_init; 513 EXPORT_SYMBOL_GPL(system_long_wq); 514 struct workqueue_struct *system_unbound_wq __ro_after_init; 515 EXPORT_SYMBOL_GPL(system_unbound_wq); 516 struct workqueue_struct *system_dfl_wq __ro_after_init; 517 EXPORT_SYMBOL_GPL(system_dfl_wq); 518 struct workqueue_struct *system_freezable_wq __ro_after_init; 519 EXPORT_SYMBOL_GPL(system_freezable_wq); 520 struct workqueue_struct *system_power_efficient_wq __ro_after_init; 521 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 522 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; 523 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 524 struct workqueue_struct *system_bh_wq; 525 EXPORT_SYMBOL_GPL(system_bh_wq); 526 struct workqueue_struct *system_bh_highpri_wq; 527 EXPORT_SYMBOL_GPL(system_bh_highpri_wq); 528 529 static int worker_thread(void *__worker); 530 static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 531 static void show_pwq(struct pool_workqueue *pwq); 532 static void show_one_worker_pool(struct worker_pool *pool); 533 534 #define CREATE_TRACE_POINTS 535 #include <trace/events/workqueue.h> 536 537 #define assert_rcu_or_pool_mutex() \ 538 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ 539 !lockdep_is_held(&wq_pool_mutex), \ 540 "RCU or wq_pool_mutex should be held") 541 542 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 543 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ 544 !lockdep_is_held(&wq->mutex) && \ 545 !lockdep_is_held(&wq_pool_mutex), \ 546 "RCU, wq->mutex or wq_pool_mutex should be held") 547 548 #define for_each_bh_worker_pool(pool, cpu) \ 549 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \ 550 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 551 (pool)++) 552 553 #define for_each_cpu_worker_pool(pool, cpu) \ 554 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 555 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 556 (pool)++) 557 558 /** 559 * for_each_pool - iterate through all worker_pools in the system 560 * @pool: iteration cursor 561 * @pi: integer used for iteration 562 * 563 * This must be called either with wq_pool_mutex held or RCU read 564 * locked. If the pool needs to be used beyond the locking in effect, the 565 * caller is responsible for guaranteeing that the pool stays online. 566 * 567 * The if/else clause exists only for the lockdep assertion and can be 568 * ignored. 569 */ 570 #define for_each_pool(pool, pi) \ 571 idr_for_each_entry(&worker_pool_idr, pool, pi) \ 572 if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 573 else 574 575 /** 576 * for_each_pool_worker - iterate through all workers of a worker_pool 577 * @worker: iteration cursor 578 * @pool: worker_pool to iterate workers of 579 * 580 * This must be called with wq_pool_attach_mutex. 581 * 582 * The if/else clause exists only for the lockdep assertion and can be 583 * ignored. 584 */ 585 #define for_each_pool_worker(worker, pool) \ 586 list_for_each_entry((worker), &(pool)->workers, node) \ 587 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 588 else 589 590 /** 591 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 592 * @pwq: iteration cursor 593 * @wq: the target workqueue 594 * 595 * This must be called either with wq->mutex held or RCU read locked. 596 * If the pwq needs to be used beyond the locking in effect, the caller is 597 * responsible for guaranteeing that the pwq stays online. 598 * 599 * The if/else clause exists only for the lockdep assertion and can be 600 * ignored. 601 */ 602 #define for_each_pwq(pwq, wq) \ 603 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 604 lockdep_is_held(&(wq->mutex))) 605 606 #ifdef CONFIG_DEBUG_OBJECTS_WORK 607 608 static const struct debug_obj_descr work_debug_descr; 609 610 static void *work_debug_hint(void *addr) 611 { 612 return ((struct work_struct *) addr)->func; 613 } 614 615 static bool work_is_static_object(void *addr) 616 { 617 struct work_struct *work = addr; 618 619 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 620 } 621 622 /* 623 * fixup_init is called when: 624 * - an active object is initialized 625 */ 626 static bool work_fixup_init(void *addr, enum debug_obj_state state) 627 { 628 struct work_struct *work = addr; 629 630 switch (state) { 631 case ODEBUG_STATE_ACTIVE: 632 cancel_work_sync(work); 633 debug_object_init(work, &work_debug_descr); 634 return true; 635 default: 636 return false; 637 } 638 } 639 640 /* 641 * fixup_free is called when: 642 * - an active object is freed 643 */ 644 static bool work_fixup_free(void *addr, enum debug_obj_state state) 645 { 646 struct work_struct *work = addr; 647 648 switch (state) { 649 case ODEBUG_STATE_ACTIVE: 650 cancel_work_sync(work); 651 debug_object_free(work, &work_debug_descr); 652 return true; 653 default: 654 return false; 655 } 656 } 657 658 static const struct debug_obj_descr work_debug_descr = { 659 .name = "work_struct", 660 .debug_hint = work_debug_hint, 661 .is_static_object = work_is_static_object, 662 .fixup_init = work_fixup_init, 663 .fixup_free = work_fixup_free, 664 }; 665 666 static inline void debug_work_activate(struct work_struct *work) 667 { 668 debug_object_activate(work, &work_debug_descr); 669 } 670 671 static inline void debug_work_deactivate(struct work_struct *work) 672 { 673 debug_object_deactivate(work, &work_debug_descr); 674 } 675 676 void __init_work(struct work_struct *work, int onstack) 677 { 678 if (onstack) 679 debug_object_init_on_stack(work, &work_debug_descr); 680 else 681 debug_object_init(work, &work_debug_descr); 682 } 683 EXPORT_SYMBOL_GPL(__init_work); 684 685 void destroy_work_on_stack(struct work_struct *work) 686 { 687 debug_object_free(work, &work_debug_descr); 688 } 689 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 690 691 void destroy_delayed_work_on_stack(struct delayed_work *work) 692 { 693 timer_destroy_on_stack(&work->timer); 694 debug_object_free(&work->work, &work_debug_descr); 695 } 696 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 697 698 #else 699 static inline void debug_work_activate(struct work_struct *work) { } 700 static inline void debug_work_deactivate(struct work_struct *work) { } 701 #endif 702 703 /** 704 * worker_pool_assign_id - allocate ID and assign it to @pool 705 * @pool: the pool pointer of interest 706 * 707 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 708 * successfully, -errno on failure. 709 */ 710 static int worker_pool_assign_id(struct worker_pool *pool) 711 { 712 int ret; 713 714 lockdep_assert_held(&wq_pool_mutex); 715 716 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 717 GFP_KERNEL); 718 if (ret >= 0) { 719 pool->id = ret; 720 return 0; 721 } 722 return ret; 723 } 724 725 static struct pool_workqueue __rcu ** 726 unbound_pwq_slot(struct workqueue_struct *wq, int cpu) 727 { 728 if (cpu >= 0) 729 return per_cpu_ptr(wq->cpu_pwq, cpu); 730 else 731 return &wq->dfl_pwq; 732 } 733 734 /* @cpu < 0 for dfl_pwq */ 735 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu) 736 { 737 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu), 738 lockdep_is_held(&wq_pool_mutex) || 739 lockdep_is_held(&wq->mutex)); 740 } 741 742 /** 743 * unbound_effective_cpumask - effective cpumask of an unbound workqueue 744 * @wq: workqueue of interest 745 * 746 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which 747 * is masked with wq_unbound_cpumask to determine the effective cpumask. The 748 * default pwq is always mapped to the pool with the current effective cpumask. 749 */ 750 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq) 751 { 752 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask; 753 } 754 755 static unsigned int work_color_to_flags(int color) 756 { 757 return color << WORK_STRUCT_COLOR_SHIFT; 758 } 759 760 static int get_work_color(unsigned long work_data) 761 { 762 return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 763 ((1 << WORK_STRUCT_COLOR_BITS) - 1); 764 } 765 766 static int work_next_color(int color) 767 { 768 return (color + 1) % WORK_NR_COLORS; 769 } 770 771 static unsigned long pool_offq_flags(struct worker_pool *pool) 772 { 773 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0; 774 } 775 776 /* 777 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 778 * contain the pointer to the queued pwq. Once execution starts, the flag 779 * is cleared and the high bits contain OFFQ flags and pool ID. 780 * 781 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling() 782 * can be used to set the pwq, pool or clear work->data. These functions should 783 * only be called while the work is owned - ie. while the PENDING bit is set. 784 * 785 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 786 * corresponding to a work. Pool is available once the work has been 787 * queued anywhere after initialization until it is sync canceled. pwq is 788 * available only while the work item is queued. 789 */ 790 static inline void set_work_data(struct work_struct *work, unsigned long data) 791 { 792 WARN_ON_ONCE(!work_pending(work)); 793 atomic_long_set(&work->data, data | work_static(work)); 794 } 795 796 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 797 unsigned long flags) 798 { 799 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING | 800 WORK_STRUCT_PWQ | flags); 801 } 802 803 static void set_work_pool_and_keep_pending(struct work_struct *work, 804 int pool_id, unsigned long flags) 805 { 806 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | 807 WORK_STRUCT_PENDING | flags); 808 } 809 810 static void set_work_pool_and_clear_pending(struct work_struct *work, 811 int pool_id, unsigned long flags) 812 { 813 /* 814 * The following wmb is paired with the implied mb in 815 * test_and_set_bit(PENDING) and ensures all updates to @work made 816 * here are visible to and precede any updates by the next PENDING 817 * owner. 818 */ 819 smp_wmb(); 820 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | 821 flags); 822 /* 823 * The following mb guarantees that previous clear of a PENDING bit 824 * will not be reordered with any speculative LOADS or STORES from 825 * work->current_func, which is executed afterwards. This possible 826 * reordering can lead to a missed execution on attempt to queue 827 * the same @work. E.g. consider this case: 828 * 829 * CPU#0 CPU#1 830 * ---------------------------- -------------------------------- 831 * 832 * 1 STORE event_indicated 833 * 2 queue_work_on() { 834 * 3 test_and_set_bit(PENDING) 835 * 4 } set_..._and_clear_pending() { 836 * 5 set_work_data() # clear bit 837 * 6 smp_mb() 838 * 7 work->current_func() { 839 * 8 LOAD event_indicated 840 * } 841 * 842 * Without an explicit full barrier speculative LOAD on line 8 can 843 * be executed before CPU#0 does STORE on line 1. If that happens, 844 * CPU#0 observes the PENDING bit is still set and new execution of 845 * a @work is not queued in a hope, that CPU#1 will eventually 846 * finish the queued @work. Meanwhile CPU#1 does not see 847 * event_indicated is set, because speculative LOAD was executed 848 * before actual STORE. 849 */ 850 smp_mb(); 851 } 852 853 static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 854 { 855 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK); 856 } 857 858 static struct pool_workqueue *get_work_pwq(struct work_struct *work) 859 { 860 unsigned long data = atomic_long_read(&work->data); 861 862 if (data & WORK_STRUCT_PWQ) 863 return work_struct_pwq(data); 864 else 865 return NULL; 866 } 867 868 /** 869 * get_work_pool - return the worker_pool a given work was associated with 870 * @work: the work item of interest 871 * 872 * Pools are created and destroyed under wq_pool_mutex, and allows read 873 * access under RCU read lock. As such, this function should be 874 * called under wq_pool_mutex or inside of a rcu_read_lock() region. 875 * 876 * All fields of the returned pool are accessible as long as the above 877 * mentioned locking is in effect. If the returned pool needs to be used 878 * beyond the critical section, the caller is responsible for ensuring the 879 * returned pool is and stays online. 880 * 881 * Return: The worker_pool @work was last associated with. %NULL if none. 882 */ 883 static struct worker_pool *get_work_pool(struct work_struct *work) 884 { 885 unsigned long data = atomic_long_read(&work->data); 886 int pool_id; 887 888 assert_rcu_or_pool_mutex(); 889 890 if (data & WORK_STRUCT_PWQ) 891 return work_struct_pwq(data)->pool; 892 893 pool_id = data >> WORK_OFFQ_POOL_SHIFT; 894 if (pool_id == WORK_OFFQ_POOL_NONE) 895 return NULL; 896 897 return idr_find(&worker_pool_idr, pool_id); 898 } 899 900 static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits) 901 { 902 return (v >> shift) & ((1U << bits) - 1); 903 } 904 905 static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data) 906 { 907 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); 908 909 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT, 910 WORK_OFFQ_POOL_BITS); 911 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT, 912 WORK_OFFQ_DISABLE_BITS); 913 offqd->flags = data & WORK_OFFQ_FLAG_MASK; 914 } 915 916 static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd) 917 { 918 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) | 919 ((unsigned long)offqd->flags); 920 } 921 922 /* 923 * Policy functions. These define the policies on how the global worker 924 * pools are managed. Unless noted otherwise, these functions assume that 925 * they're being called with pool->lock held. 926 */ 927 928 /* 929 * Need to wake up a worker? Called from anything but currently 930 * running workers. 931 * 932 * Note that, because unbound workers never contribute to nr_running, this 933 * function will always return %true for unbound pools as long as the 934 * worklist isn't empty. 935 */ 936 static bool need_more_worker(struct worker_pool *pool) 937 { 938 return !list_empty(&pool->worklist) && !pool->nr_running; 939 } 940 941 /* Can I start working? Called from busy but !running workers. */ 942 static bool may_start_working(struct worker_pool *pool) 943 { 944 return pool->nr_idle; 945 } 946 947 /* Do I need to keep working? Called from currently running workers. */ 948 static bool keep_working(struct worker_pool *pool) 949 { 950 return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 951 } 952 953 /* Do we need a new worker? Called from manager. */ 954 static bool need_to_create_worker(struct worker_pool *pool) 955 { 956 return need_more_worker(pool) && !may_start_working(pool); 957 } 958 959 /* Do we have too many workers and should some go away? */ 960 static bool too_many_workers(struct worker_pool *pool) 961 { 962 bool managing = pool->flags & POOL_MANAGER_ACTIVE; 963 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 964 int nr_busy = pool->nr_workers - nr_idle; 965 966 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 967 } 968 969 /** 970 * worker_set_flags - set worker flags and adjust nr_running accordingly 971 * @worker: self 972 * @flags: flags to set 973 * 974 * Set @flags in @worker->flags and adjust nr_running accordingly. 975 */ 976 static inline void worker_set_flags(struct worker *worker, unsigned int flags) 977 { 978 struct worker_pool *pool = worker->pool; 979 980 lockdep_assert_held(&pool->lock); 981 982 /* If transitioning into NOT_RUNNING, adjust nr_running. */ 983 if ((flags & WORKER_NOT_RUNNING) && 984 !(worker->flags & WORKER_NOT_RUNNING)) { 985 pool->nr_running--; 986 } 987 988 worker->flags |= flags; 989 } 990 991 /** 992 * worker_clr_flags - clear worker flags and adjust nr_running accordingly 993 * @worker: self 994 * @flags: flags to clear 995 * 996 * Clear @flags in @worker->flags and adjust nr_running accordingly. 997 */ 998 static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 999 { 1000 struct worker_pool *pool = worker->pool; 1001 unsigned int oflags = worker->flags; 1002 1003 lockdep_assert_held(&pool->lock); 1004 1005 worker->flags &= ~flags; 1006 1007 /* 1008 * If transitioning out of NOT_RUNNING, increment nr_running. Note 1009 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 1010 * of multiple flags, not a single flag. 1011 */ 1012 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1013 if (!(worker->flags & WORKER_NOT_RUNNING)) 1014 pool->nr_running++; 1015 } 1016 1017 /* Return the first idle worker. Called with pool->lock held. */ 1018 static struct worker *first_idle_worker(struct worker_pool *pool) 1019 { 1020 if (unlikely(list_empty(&pool->idle_list))) 1021 return NULL; 1022 1023 return list_first_entry(&pool->idle_list, struct worker, entry); 1024 } 1025 1026 /** 1027 * worker_enter_idle - enter idle state 1028 * @worker: worker which is entering idle state 1029 * 1030 * @worker is entering idle state. Update stats and idle timer if 1031 * necessary. 1032 * 1033 * LOCKING: 1034 * raw_spin_lock_irq(pool->lock). 1035 */ 1036 static void worker_enter_idle(struct worker *worker) 1037 { 1038 struct worker_pool *pool = worker->pool; 1039 1040 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 1041 WARN_ON_ONCE(!list_empty(&worker->entry) && 1042 (worker->hentry.next || worker->hentry.pprev))) 1043 return; 1044 1045 /* can't use worker_set_flags(), also called from create_worker() */ 1046 worker->flags |= WORKER_IDLE; 1047 pool->nr_idle++; 1048 worker->last_active = jiffies; 1049 1050 /* idle_list is LIFO */ 1051 list_add(&worker->entry, &pool->idle_list); 1052 1053 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1054 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1055 1056 /* Sanity check nr_running. */ 1057 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 1058 } 1059 1060 /** 1061 * worker_leave_idle - leave idle state 1062 * @worker: worker which is leaving idle state 1063 * 1064 * @worker is leaving idle state. Update stats. 1065 * 1066 * LOCKING: 1067 * raw_spin_lock_irq(pool->lock). 1068 */ 1069 static void worker_leave_idle(struct worker *worker) 1070 { 1071 struct worker_pool *pool = worker->pool; 1072 1073 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 1074 return; 1075 worker_clr_flags(worker, WORKER_IDLE); 1076 pool->nr_idle--; 1077 list_del_init(&worker->entry); 1078 } 1079 1080 /** 1081 * find_worker_executing_work - find worker which is executing a work 1082 * @pool: pool of interest 1083 * @work: work to find worker for 1084 * 1085 * Find a worker which is executing @work on @pool by searching 1086 * @pool->busy_hash which is keyed by the address of @work. For a worker 1087 * to match, its current execution should match the address of @work and 1088 * its work function. This is to avoid unwanted dependency between 1089 * unrelated work executions through a work item being recycled while still 1090 * being executed. 1091 * 1092 * This is a bit tricky. A work item may be freed once its execution 1093 * starts and nothing prevents the freed area from being recycled for 1094 * another work item. If the same work item address ends up being reused 1095 * before the original execution finishes, workqueue will identify the 1096 * recycled work item as currently executing and make it wait until the 1097 * current execution finishes, introducing an unwanted dependency. 1098 * 1099 * This function checks the work item address and work function to avoid 1100 * false positives. Note that this isn't complete as one may construct a 1101 * work function which can introduce dependency onto itself through a 1102 * recycled work item. Well, if somebody wants to shoot oneself in the 1103 * foot that badly, there's only so much we can do, and if such deadlock 1104 * actually occurs, it should be easy to locate the culprit work function. 1105 * 1106 * CONTEXT: 1107 * raw_spin_lock_irq(pool->lock). 1108 * 1109 * Return: 1110 * Pointer to worker which is executing @work if found, %NULL 1111 * otherwise. 1112 */ 1113 static struct worker *find_worker_executing_work(struct worker_pool *pool, 1114 struct work_struct *work) 1115 { 1116 struct worker *worker; 1117 1118 hash_for_each_possible(pool->busy_hash, worker, hentry, 1119 (unsigned long)work) 1120 if (worker->current_work == work && 1121 worker->current_func == work->func) 1122 return worker; 1123 1124 return NULL; 1125 } 1126 1127 /** 1128 * move_linked_works - move linked works to a list 1129 * @work: start of series of works to be scheduled 1130 * @head: target list to append @work to 1131 * @nextp: out parameter for nested worklist walking 1132 * 1133 * Schedule linked works starting from @work to @head. Work series to be 1134 * scheduled starts at @work and includes any consecutive work with 1135 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on 1136 * @nextp. 1137 * 1138 * CONTEXT: 1139 * raw_spin_lock_irq(pool->lock). 1140 */ 1141 static void move_linked_works(struct work_struct *work, struct list_head *head, 1142 struct work_struct **nextp) 1143 { 1144 struct work_struct *n; 1145 1146 /* 1147 * Linked worklist will always end before the end of the list, 1148 * use NULL for list head. 1149 */ 1150 list_for_each_entry_safe_from(work, n, NULL, entry) { 1151 list_move_tail(&work->entry, head); 1152 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1153 break; 1154 } 1155 1156 /* 1157 * If we're already inside safe list traversal and have moved 1158 * multiple works to the scheduled queue, the next position 1159 * needs to be updated. 1160 */ 1161 if (nextp) 1162 *nextp = n; 1163 } 1164 1165 /** 1166 * assign_work - assign a work item and its linked work items to a worker 1167 * @work: work to assign 1168 * @worker: worker to assign to 1169 * @nextp: out parameter for nested worklist walking 1170 * 1171 * Assign @work and its linked work items to @worker. If @work is already being 1172 * executed by another worker in the same pool, it'll be punted there. 1173 * 1174 * If @nextp is not NULL, it's updated to point to the next work of the last 1175 * scheduled work. This allows assign_work() to be nested inside 1176 * list_for_each_entry_safe(). 1177 * 1178 * Returns %true if @work was successfully assigned to @worker. %false if @work 1179 * was punted to another worker already executing it. 1180 */ 1181 static bool assign_work(struct work_struct *work, struct worker *worker, 1182 struct work_struct **nextp) 1183 { 1184 struct worker_pool *pool = worker->pool; 1185 struct worker *collision; 1186 1187 lockdep_assert_held(&pool->lock); 1188 1189 /* 1190 * A single work shouldn't be executed concurrently by multiple workers. 1191 * __queue_work() ensures that @work doesn't jump to a different pool 1192 * while still running in the previous pool. Here, we should ensure that 1193 * @work is not executed concurrently by multiple workers from the same 1194 * pool. Check whether anyone is already processing the work. If so, 1195 * defer the work to the currently executing one. 1196 */ 1197 collision = find_worker_executing_work(pool, work); 1198 if (unlikely(collision)) { 1199 move_linked_works(work, &collision->scheduled, nextp); 1200 return false; 1201 } 1202 1203 move_linked_works(work, &worker->scheduled, nextp); 1204 return true; 1205 } 1206 1207 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool) 1208 { 1209 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0; 1210 1211 return &per_cpu(bh_pool_irq_works, pool->cpu)[high]; 1212 } 1213 1214 static void kick_bh_pool(struct worker_pool *pool) 1215 { 1216 #ifdef CONFIG_SMP 1217 /* see drain_dead_softirq_workfn() for BH_DRAINING */ 1218 if (unlikely(pool->cpu != smp_processor_id() && 1219 !(pool->flags & POOL_BH_DRAINING))) { 1220 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu); 1221 return; 1222 } 1223 #endif 1224 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 1225 raise_softirq_irqoff(HI_SOFTIRQ); 1226 else 1227 raise_softirq_irqoff(TASKLET_SOFTIRQ); 1228 } 1229 1230 /** 1231 * kick_pool - wake up an idle worker if necessary 1232 * @pool: pool to kick 1233 * 1234 * @pool may have pending work items. Wake up worker if necessary. Returns 1235 * whether a worker was woken up. 1236 */ 1237 static bool kick_pool(struct worker_pool *pool) 1238 { 1239 struct worker *worker = first_idle_worker(pool); 1240 struct task_struct *p; 1241 1242 lockdep_assert_held(&pool->lock); 1243 1244 if (!need_more_worker(pool) || !worker) 1245 return false; 1246 1247 if (pool->flags & POOL_BH) { 1248 kick_bh_pool(pool); 1249 return true; 1250 } 1251 1252 p = worker->task; 1253 1254 #ifdef CONFIG_SMP 1255 /* 1256 * Idle @worker is about to execute @work and waking up provides an 1257 * opportunity to migrate @worker at a lower cost by setting the task's 1258 * wake_cpu field. Let's see if we want to move @worker to improve 1259 * execution locality. 1260 * 1261 * We're waking the worker that went idle the latest and there's some 1262 * chance that @worker is marked idle but hasn't gone off CPU yet. If 1263 * so, setting the wake_cpu won't do anything. As this is a best-effort 1264 * optimization and the race window is narrow, let's leave as-is for 1265 * now. If this becomes pronounced, we can skip over workers which are 1266 * still on cpu when picking an idle worker. 1267 * 1268 * If @pool has non-strict affinity, @worker might have ended up outside 1269 * its affinity scope. Repatriate. 1270 */ 1271 if (!pool->attrs->affn_strict && 1272 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { 1273 struct work_struct *work = list_first_entry(&pool->worklist, 1274 struct work_struct, entry); 1275 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask, 1276 cpu_online_mask); 1277 if (wake_cpu < nr_cpu_ids) { 1278 p->wake_cpu = wake_cpu; 1279 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; 1280 } 1281 } 1282 #endif 1283 wake_up_process(p); 1284 return true; 1285 } 1286 1287 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 1288 1289 /* 1290 * Concurrency-managed per-cpu work items that hog CPU for longer than 1291 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 1292 * which prevents them from stalling other concurrency-managed work items. If a 1293 * work function keeps triggering this mechanism, it's likely that the work item 1294 * should be using an unbound workqueue instead. 1295 * 1296 * wq_cpu_intensive_report() tracks work functions which trigger such conditions 1297 * and report them so that they can be examined and converted to use unbound 1298 * workqueues as appropriate. To avoid flooding the console, each violating work 1299 * function is tracked and reported with exponential backoff. 1300 */ 1301 #define WCI_MAX_ENTS 128 1302 1303 struct wci_ent { 1304 work_func_t func; 1305 atomic64_t cnt; 1306 struct hlist_node hash_node; 1307 }; 1308 1309 static struct wci_ent wci_ents[WCI_MAX_ENTS]; 1310 static int wci_nr_ents; 1311 static DEFINE_RAW_SPINLOCK(wci_lock); 1312 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 1313 1314 static struct wci_ent *wci_find_ent(work_func_t func) 1315 { 1316 struct wci_ent *ent; 1317 1318 hash_for_each_possible_rcu(wci_hash, ent, hash_node, 1319 (unsigned long)func) { 1320 if (ent->func == func) 1321 return ent; 1322 } 1323 return NULL; 1324 } 1325 1326 static void wq_cpu_intensive_report(work_func_t func) 1327 { 1328 struct wci_ent *ent; 1329 1330 restart: 1331 ent = wci_find_ent(func); 1332 if (ent) { 1333 u64 cnt; 1334 1335 /* 1336 * Start reporting from the warning_thresh and back off 1337 * exponentially. 1338 */ 1339 cnt = atomic64_inc_return_relaxed(&ent->cnt); 1340 if (wq_cpu_intensive_warning_thresh && 1341 cnt >= wq_cpu_intensive_warning_thresh && 1342 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh)) 1343 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 1344 ent->func, wq_cpu_intensive_thresh_us, 1345 atomic64_read(&ent->cnt)); 1346 return; 1347 } 1348 1349 /* 1350 * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 1351 * is exhausted, something went really wrong and we probably made enough 1352 * noise already. 1353 */ 1354 if (wci_nr_ents >= WCI_MAX_ENTS) 1355 return; 1356 1357 raw_spin_lock(&wci_lock); 1358 1359 if (wci_nr_ents >= WCI_MAX_ENTS) { 1360 raw_spin_unlock(&wci_lock); 1361 return; 1362 } 1363 1364 if (wci_find_ent(func)) { 1365 raw_spin_unlock(&wci_lock); 1366 goto restart; 1367 } 1368 1369 ent = &wci_ents[wci_nr_ents++]; 1370 ent->func = func; 1371 atomic64_set(&ent->cnt, 0); 1372 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 1373 1374 raw_spin_unlock(&wci_lock); 1375 1376 goto restart; 1377 } 1378 1379 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1380 static void wq_cpu_intensive_report(work_func_t func) {} 1381 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1382 1383 /** 1384 * wq_worker_running - a worker is running again 1385 * @task: task waking up 1386 * 1387 * This function is called when a worker returns from schedule() 1388 */ 1389 void wq_worker_running(struct task_struct *task) 1390 { 1391 struct worker *worker = kthread_data(task); 1392 1393 if (!READ_ONCE(worker->sleeping)) 1394 return; 1395 1396 /* 1397 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 1398 * and the nr_running increment below, we may ruin the nr_running reset 1399 * and leave with an unexpected pool->nr_running == 1 on the newly unbound 1400 * pool. Protect against such race. 1401 */ 1402 preempt_disable(); 1403 if (!(worker->flags & WORKER_NOT_RUNNING)) 1404 worker->pool->nr_running++; 1405 preempt_enable(); 1406 1407 /* 1408 * CPU intensive auto-detection cares about how long a work item hogged 1409 * CPU without sleeping. Reset the starting timestamp on wakeup. 1410 */ 1411 worker->current_at = worker->task->se.sum_exec_runtime; 1412 1413 WRITE_ONCE(worker->sleeping, 0); 1414 } 1415 1416 /** 1417 * wq_worker_sleeping - a worker is going to sleep 1418 * @task: task going to sleep 1419 * 1420 * This function is called from schedule() when a busy worker is 1421 * going to sleep. 1422 */ 1423 void wq_worker_sleeping(struct task_struct *task) 1424 { 1425 struct worker *worker = kthread_data(task); 1426 struct worker_pool *pool; 1427 1428 /* 1429 * Rescuers, which may not have all the fields set up like normal 1430 * workers, also reach here, let's not access anything before 1431 * checking NOT_RUNNING. 1432 */ 1433 if (worker->flags & WORKER_NOT_RUNNING) 1434 return; 1435 1436 pool = worker->pool; 1437 1438 /* Return if preempted before wq_worker_running() was reached */ 1439 if (READ_ONCE(worker->sleeping)) 1440 return; 1441 1442 WRITE_ONCE(worker->sleeping, 1); 1443 raw_spin_lock_irq(&pool->lock); 1444 1445 /* 1446 * Recheck in case unbind_workers() preempted us. We don't 1447 * want to decrement nr_running after the worker is unbound 1448 * and nr_running has been reset. 1449 */ 1450 if (worker->flags & WORKER_NOT_RUNNING) { 1451 raw_spin_unlock_irq(&pool->lock); 1452 return; 1453 } 1454 1455 pool->nr_running--; 1456 if (kick_pool(pool)) 1457 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1458 1459 raw_spin_unlock_irq(&pool->lock); 1460 } 1461 1462 /** 1463 * wq_worker_tick - a scheduler tick occurred while a kworker is running 1464 * @task: task currently running 1465 * 1466 * Called from sched_tick(). We're in the IRQ context and the current 1467 * worker's fields which follow the 'K' locking rule can be accessed safely. 1468 */ 1469 void wq_worker_tick(struct task_struct *task) 1470 { 1471 struct worker *worker = kthread_data(task); 1472 struct pool_workqueue *pwq = worker->current_pwq; 1473 struct worker_pool *pool = worker->pool; 1474 1475 if (!pwq) 1476 return; 1477 1478 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 1479 1480 if (!wq_cpu_intensive_thresh_us) 1481 return; 1482 1483 /* 1484 * If the current worker is concurrency managed and hogged the CPU for 1485 * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1486 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1487 * 1488 * Set @worker->sleeping means that @worker is in the process of 1489 * switching out voluntarily and won't be contributing to 1490 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1491 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1492 * double decrements. The task is releasing the CPU anyway. Let's skip. 1493 * We probably want to make this prettier in the future. 1494 */ 1495 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1496 worker->task->se.sum_exec_runtime - worker->current_at < 1497 wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1498 return; 1499 1500 raw_spin_lock(&pool->lock); 1501 1502 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 1503 wq_cpu_intensive_report(worker->current_func); 1504 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1505 1506 if (kick_pool(pool)) 1507 pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1508 1509 raw_spin_unlock(&pool->lock); 1510 } 1511 1512 /** 1513 * wq_worker_last_func - retrieve worker's last work function 1514 * @task: Task to retrieve last work function of. 1515 * 1516 * Determine the last function a worker executed. This is called from 1517 * the scheduler to get a worker's last known identity. 1518 * 1519 * CONTEXT: 1520 * raw_spin_lock_irq(rq->lock) 1521 * 1522 * This function is called during schedule() when a kworker is going 1523 * to sleep. It's used by psi to identify aggregation workers during 1524 * dequeuing, to allow periodic aggregation to shut-off when that 1525 * worker is the last task in the system or cgroup to go to sleep. 1526 * 1527 * As this function doesn't involve any workqueue-related locking, it 1528 * only returns stable values when called from inside the scheduler's 1529 * queuing and dequeuing paths, when @task, which must be a kworker, 1530 * is guaranteed to not be processing any works. 1531 * 1532 * Return: 1533 * The last work function %current executed as a worker, NULL if it 1534 * hasn't executed any work yet. 1535 */ 1536 work_func_t wq_worker_last_func(struct task_struct *task) 1537 { 1538 struct worker *worker = kthread_data(task); 1539 1540 return worker->last_func; 1541 } 1542 1543 /** 1544 * wq_node_nr_active - Determine wq_node_nr_active to use 1545 * @wq: workqueue of interest 1546 * @node: NUMA node, can be %NUMA_NO_NODE 1547 * 1548 * Determine wq_node_nr_active to use for @wq on @node. Returns: 1549 * 1550 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. 1551 * 1552 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. 1553 * 1554 * - Otherwise, node_nr_active[@node]. 1555 */ 1556 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, 1557 int node) 1558 { 1559 if (!(wq->flags & WQ_UNBOUND)) 1560 return NULL; 1561 1562 if (node == NUMA_NO_NODE) 1563 node = nr_node_ids; 1564 1565 return wq->node_nr_active[node]; 1566 } 1567 1568 /** 1569 * wq_update_node_max_active - Update per-node max_actives to use 1570 * @wq: workqueue to update 1571 * @off_cpu: CPU that's going down, -1 if a CPU is not going down 1572 * 1573 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is 1574 * distributed among nodes according to the proportions of numbers of online 1575 * cpus. The result is always between @wq->min_active and max_active. 1576 */ 1577 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) 1578 { 1579 struct cpumask *effective = unbound_effective_cpumask(wq); 1580 int min_active = READ_ONCE(wq->min_active); 1581 int max_active = READ_ONCE(wq->max_active); 1582 int total_cpus, node; 1583 1584 lockdep_assert_held(&wq->mutex); 1585 1586 if (!wq_topo_initialized) 1587 return; 1588 1589 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective)) 1590 off_cpu = -1; 1591 1592 total_cpus = cpumask_weight_and(effective, cpu_online_mask); 1593 if (off_cpu >= 0) 1594 total_cpus--; 1595 1596 /* If all CPUs of the wq get offline, use the default values */ 1597 if (unlikely(!total_cpus)) { 1598 for_each_node(node) 1599 wq_node_nr_active(wq, node)->max = min_active; 1600 1601 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; 1602 return; 1603 } 1604 1605 for_each_node(node) { 1606 int node_cpus; 1607 1608 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); 1609 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) 1610 node_cpus--; 1611 1612 wq_node_nr_active(wq, node)->max = 1613 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), 1614 min_active, max_active); 1615 } 1616 1617 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; 1618 } 1619 1620 /** 1621 * get_pwq - get an extra reference on the specified pool_workqueue 1622 * @pwq: pool_workqueue to get 1623 * 1624 * Obtain an extra reference on @pwq. The caller should guarantee that 1625 * @pwq has positive refcnt and be holding the matching pool->lock. 1626 */ 1627 static void get_pwq(struct pool_workqueue *pwq) 1628 { 1629 lockdep_assert_held(&pwq->pool->lock); 1630 WARN_ON_ONCE(pwq->refcnt <= 0); 1631 pwq->refcnt++; 1632 } 1633 1634 /** 1635 * put_pwq - put a pool_workqueue reference 1636 * @pwq: pool_workqueue to put 1637 * 1638 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 1639 * destruction. The caller should be holding the matching pool->lock. 1640 */ 1641 static void put_pwq(struct pool_workqueue *pwq) 1642 { 1643 lockdep_assert_held(&pwq->pool->lock); 1644 if (likely(--pwq->refcnt)) 1645 return; 1646 /* 1647 * @pwq can't be released under pool->lock, bounce to a dedicated 1648 * kthread_worker to avoid A-A deadlocks. 1649 */ 1650 kthread_queue_work(pwq_release_worker, &pwq->release_work); 1651 } 1652 1653 /** 1654 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1655 * @pwq: pool_workqueue to put (can be %NULL) 1656 * 1657 * put_pwq() with locking. This function also allows %NULL @pwq. 1658 */ 1659 static void put_pwq_unlocked(struct pool_workqueue *pwq) 1660 { 1661 if (pwq) { 1662 /* 1663 * As both pwqs and pools are RCU protected, the 1664 * following lock operations are safe. 1665 */ 1666 raw_spin_lock_irq(&pwq->pool->lock); 1667 put_pwq(pwq); 1668 raw_spin_unlock_irq(&pwq->pool->lock); 1669 } 1670 } 1671 1672 static bool pwq_is_empty(struct pool_workqueue *pwq) 1673 { 1674 return !pwq->nr_active && list_empty(&pwq->inactive_works); 1675 } 1676 1677 static void __pwq_activate_work(struct pool_workqueue *pwq, 1678 struct work_struct *work) 1679 { 1680 unsigned long *wdb = work_data_bits(work); 1681 1682 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); 1683 trace_workqueue_activate_work(work); 1684 if (list_empty(&pwq->pool->worklist)) 1685 pwq->pool->watchdog_ts = jiffies; 1686 move_linked_works(work, &pwq->pool->worklist, NULL); 1687 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); 1688 } 1689 1690 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) 1691 { 1692 int max = READ_ONCE(nna->max); 1693 int old = atomic_read(&nna->nr); 1694 1695 do { 1696 if (old >= max) 1697 return false; 1698 } while (!atomic_try_cmpxchg_relaxed(&nna->nr, &old, old + 1)); 1699 1700 return true; 1701 } 1702 1703 /** 1704 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq 1705 * @pwq: pool_workqueue of interest 1706 * @fill: max_active may have increased, try to increase concurrency level 1707 * 1708 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is 1709 * successfully obtained. %false otherwise. 1710 */ 1711 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) 1712 { 1713 struct workqueue_struct *wq = pwq->wq; 1714 struct worker_pool *pool = pwq->pool; 1715 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); 1716 bool obtained = false; 1717 1718 lockdep_assert_held(&pool->lock); 1719 1720 if (!nna) { 1721 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */ 1722 obtained = pwq->nr_active < READ_ONCE(wq->max_active); 1723 goto out; 1724 } 1725 1726 if (unlikely(pwq->plugged)) 1727 return false; 1728 1729 /* 1730 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is 1731 * already waiting on $nna, pwq_dec_nr_active() will maintain the 1732 * concurrency level. Don't jump the line. 1733 * 1734 * We need to ignore the pending test after max_active has increased as 1735 * pwq_dec_nr_active() can only maintain the concurrency level but not 1736 * increase it. This is indicated by @fill. 1737 */ 1738 if (!list_empty(&pwq->pending_node) && likely(!fill)) 1739 goto out; 1740 1741 obtained = tryinc_node_nr_active(nna); 1742 if (obtained) 1743 goto out; 1744 1745 /* 1746 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs 1747 * and try again. The smp_mb() is paired with the implied memory barrier 1748 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either 1749 * we see the decremented $nna->nr or they see non-empty 1750 * $nna->pending_pwqs. 1751 */ 1752 raw_spin_lock(&nna->lock); 1753 1754 if (list_empty(&pwq->pending_node)) 1755 list_add_tail(&pwq->pending_node, &nna->pending_pwqs); 1756 else if (likely(!fill)) 1757 goto out_unlock; 1758 1759 smp_mb(); 1760 1761 obtained = tryinc_node_nr_active(nna); 1762 1763 /* 1764 * If @fill, @pwq might have already been pending. Being spuriously 1765 * pending in cold paths doesn't affect anything. Let's leave it be. 1766 */ 1767 if (obtained && likely(!fill)) 1768 list_del_init(&pwq->pending_node); 1769 1770 out_unlock: 1771 raw_spin_unlock(&nna->lock); 1772 out: 1773 if (obtained) 1774 pwq->nr_active++; 1775 return obtained; 1776 } 1777 1778 /** 1779 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq 1780 * @pwq: pool_workqueue of interest 1781 * @fill: max_active may have increased, try to increase concurrency level 1782 * 1783 * Activate the first inactive work item of @pwq if available and allowed by 1784 * max_active limit. 1785 * 1786 * Returns %true if an inactive work item has been activated. %false if no 1787 * inactive work item is found or max_active limit is reached. 1788 */ 1789 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) 1790 { 1791 struct work_struct *work = 1792 list_first_entry_or_null(&pwq->inactive_works, 1793 struct work_struct, entry); 1794 1795 if (work && pwq_tryinc_nr_active(pwq, fill)) { 1796 __pwq_activate_work(pwq, work); 1797 return true; 1798 } else { 1799 return false; 1800 } 1801 } 1802 1803 /** 1804 * unplug_oldest_pwq - unplug the oldest pool_workqueue 1805 * @wq: workqueue_struct where its oldest pwq is to be unplugged 1806 * 1807 * This function should only be called for ordered workqueues where only the 1808 * oldest pwq is unplugged, the others are plugged to suspend execution to 1809 * ensure proper work item ordering:: 1810 * 1811 * dfl_pwq --------------+ [P] - plugged 1812 * | 1813 * v 1814 * pwqs -> A -> B [P] -> C [P] (newest) 1815 * | | | 1816 * 1 3 5 1817 * | | | 1818 * 2 4 6 1819 * 1820 * When the oldest pwq is drained and removed, this function should be called 1821 * to unplug the next oldest one to start its work item execution. Note that 1822 * pwq's are linked into wq->pwqs with the oldest first, so the first one in 1823 * the list is the oldest. 1824 */ 1825 static void unplug_oldest_pwq(struct workqueue_struct *wq) 1826 { 1827 struct pool_workqueue *pwq; 1828 1829 lockdep_assert_held(&wq->mutex); 1830 1831 /* Caller should make sure that pwqs isn't empty before calling */ 1832 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue, 1833 pwqs_node); 1834 raw_spin_lock_irq(&pwq->pool->lock); 1835 if (pwq->plugged) { 1836 pwq->plugged = false; 1837 if (pwq_activate_first_inactive(pwq, true)) 1838 kick_pool(pwq->pool); 1839 } 1840 raw_spin_unlock_irq(&pwq->pool->lock); 1841 } 1842 1843 /** 1844 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active 1845 * @nna: wq_node_nr_active to activate a pending pwq for 1846 * @caller_pool: worker_pool the caller is locking 1847 * 1848 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. 1849 * @caller_pool may be unlocked and relocked to lock other worker_pools. 1850 */ 1851 static void node_activate_pending_pwq(struct wq_node_nr_active *nna, 1852 struct worker_pool *caller_pool) 1853 { 1854 struct worker_pool *locked_pool = caller_pool; 1855 struct pool_workqueue *pwq; 1856 struct work_struct *work; 1857 1858 lockdep_assert_held(&caller_pool->lock); 1859 1860 raw_spin_lock(&nna->lock); 1861 retry: 1862 pwq = list_first_entry_or_null(&nna->pending_pwqs, 1863 struct pool_workqueue, pending_node); 1864 if (!pwq) 1865 goto out_unlock; 1866 1867 /* 1868 * If @pwq is for a different pool than @locked_pool, we need to lock 1869 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock 1870 * / lock dance. For that, we also need to release @nna->lock as it's 1871 * nested inside pool locks. 1872 */ 1873 if (pwq->pool != locked_pool) { 1874 raw_spin_unlock(&locked_pool->lock); 1875 locked_pool = pwq->pool; 1876 if (!raw_spin_trylock(&locked_pool->lock)) { 1877 raw_spin_unlock(&nna->lock); 1878 raw_spin_lock(&locked_pool->lock); 1879 raw_spin_lock(&nna->lock); 1880 goto retry; 1881 } 1882 } 1883 1884 /* 1885 * $pwq may not have any inactive work items due to e.g. cancellations. 1886 * Drop it from pending_pwqs and see if there's another one. 1887 */ 1888 work = list_first_entry_or_null(&pwq->inactive_works, 1889 struct work_struct, entry); 1890 if (!work) { 1891 list_del_init(&pwq->pending_node); 1892 goto retry; 1893 } 1894 1895 /* 1896 * Acquire an nr_active count and activate the inactive work item. If 1897 * $pwq still has inactive work items, rotate it to the end of the 1898 * pending_pwqs so that we round-robin through them. This means that 1899 * inactive work items are not activated in queueing order which is fine 1900 * given that there has never been any ordering across different pwqs. 1901 */ 1902 if (likely(tryinc_node_nr_active(nna))) { 1903 pwq->nr_active++; 1904 __pwq_activate_work(pwq, work); 1905 1906 if (list_empty(&pwq->inactive_works)) 1907 list_del_init(&pwq->pending_node); 1908 else 1909 list_move_tail(&pwq->pending_node, &nna->pending_pwqs); 1910 1911 /* if activating a foreign pool, make sure it's running */ 1912 if (pwq->pool != caller_pool) 1913 kick_pool(pwq->pool); 1914 } 1915 1916 out_unlock: 1917 raw_spin_unlock(&nna->lock); 1918 if (locked_pool != caller_pool) { 1919 raw_spin_unlock(&locked_pool->lock); 1920 raw_spin_lock(&caller_pool->lock); 1921 } 1922 } 1923 1924 /** 1925 * pwq_dec_nr_active - Retire an active count 1926 * @pwq: pool_workqueue of interest 1927 * 1928 * Decrement @pwq's nr_active and try to activate the first inactive work item. 1929 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. 1930 */ 1931 static void pwq_dec_nr_active(struct pool_workqueue *pwq) 1932 { 1933 struct worker_pool *pool = pwq->pool; 1934 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); 1935 1936 lockdep_assert_held(&pool->lock); 1937 1938 /* 1939 * @pwq->nr_active should be decremented for both percpu and unbound 1940 * workqueues. 1941 */ 1942 pwq->nr_active--; 1943 1944 /* 1945 * For a percpu workqueue, it's simple. Just need to kick the first 1946 * inactive work item on @pwq itself. 1947 */ 1948 if (!nna) { 1949 pwq_activate_first_inactive(pwq, false); 1950 return; 1951 } 1952 1953 /* 1954 * If @pwq is for an unbound workqueue, it's more complicated because 1955 * multiple pwqs and pools may be sharing the nr_active count. When a 1956 * pwq needs to wait for an nr_active count, it puts itself on 1957 * $nna->pending_pwqs. The following atomic_dec_return()'s implied 1958 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to 1959 * guarantee that either we see non-empty pending_pwqs or they see 1960 * decremented $nna->nr. 1961 * 1962 * $nna->max may change as CPUs come online/offline and @pwq->wq's 1963 * max_active gets updated. However, it is guaranteed to be equal to or 1964 * larger than @pwq->wq->min_active which is above zero unless freezing. 1965 * This maintains the forward progress guarantee. 1966 */ 1967 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) 1968 return; 1969 1970 if (!list_empty(&nna->pending_pwqs)) 1971 node_activate_pending_pwq(nna, pool); 1972 } 1973 1974 /** 1975 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1976 * @pwq: pwq of interest 1977 * @work_data: work_data of work which left the queue 1978 * 1979 * A work either has completed or is removed from pending queue, 1980 * decrement nr_in_flight of its pwq and handle workqueue flushing. 1981 * 1982 * NOTE: 1983 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock 1984 * and thus should be called after all other state updates for the in-flight 1985 * work item is complete. 1986 * 1987 * CONTEXT: 1988 * raw_spin_lock_irq(pool->lock). 1989 */ 1990 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1991 { 1992 int color = get_work_color(work_data); 1993 1994 if (!(work_data & WORK_STRUCT_INACTIVE)) 1995 pwq_dec_nr_active(pwq); 1996 1997 pwq->nr_in_flight[color]--; 1998 1999 /* is flush in progress and are we at the flushing tip? */ 2000 if (likely(pwq->flush_color != color)) 2001 goto out_put; 2002 2003 /* are there still in-flight works? */ 2004 if (pwq->nr_in_flight[color]) 2005 goto out_put; 2006 2007 /* this pwq is done, clear flush_color */ 2008 pwq->flush_color = -1; 2009 2010 /* 2011 * If this was the last pwq, wake up the first flusher. It 2012 * will handle the rest. 2013 */ 2014 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 2015 complete(&pwq->wq->first_flusher->done); 2016 out_put: 2017 put_pwq(pwq); 2018 } 2019 2020 /** 2021 * try_to_grab_pending - steal work item from worklist and disable irq 2022 * @work: work item to steal 2023 * @cflags: %WORK_CANCEL_ flags 2024 * @irq_flags: place to store irq state 2025 * 2026 * Try to grab PENDING bit of @work. This function can handle @work in any 2027 * stable state - idle, on timer or on worklist. 2028 * 2029 * Return: 2030 * 2031 * ======== ================================================================ 2032 * 1 if @work was pending and we successfully stole PENDING 2033 * 0 if @work was idle and we claimed PENDING 2034 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 2035 * ======== ================================================================ 2036 * 2037 * Note: 2038 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 2039 * interrupted while holding PENDING and @work off queue, irq must be 2040 * disabled on entry. This, combined with delayed_work->timer being 2041 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 2042 * 2043 * On successful return, >= 0, irq is disabled and the caller is 2044 * responsible for releasing it using local_irq_restore(*@irq_flags). 2045 * 2046 * This function is safe to call from any context including IRQ handler. 2047 */ 2048 static int try_to_grab_pending(struct work_struct *work, u32 cflags, 2049 unsigned long *irq_flags) 2050 { 2051 struct worker_pool *pool; 2052 struct pool_workqueue *pwq; 2053 2054 local_irq_save(*irq_flags); 2055 2056 /* try to steal the timer if it exists */ 2057 if (cflags & WORK_CANCEL_DELAYED) { 2058 struct delayed_work *dwork = to_delayed_work(work); 2059 2060 /* 2061 * dwork->timer is irqsafe. If timer_delete() fails, it's 2062 * guaranteed that the timer is not queued anywhere and not 2063 * running on the local CPU. 2064 */ 2065 if (likely(timer_delete(&dwork->timer))) 2066 return 1; 2067 } 2068 2069 /* try to claim PENDING the normal way */ 2070 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 2071 return 0; 2072 2073 rcu_read_lock(); 2074 /* 2075 * The queueing is in progress, or it is already queued. Try to 2076 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 2077 */ 2078 pool = get_work_pool(work); 2079 if (!pool) 2080 goto fail; 2081 2082 raw_spin_lock(&pool->lock); 2083 /* 2084 * work->data is guaranteed to point to pwq only while the work 2085 * item is queued on pwq->wq, and both updating work->data to point 2086 * to pwq on queueing and to pool on dequeueing are done under 2087 * pwq->pool->lock. This in turn guarantees that, if work->data 2088 * points to pwq which is associated with a locked pool, the work 2089 * item is currently queued on that pool. 2090 */ 2091 pwq = get_work_pwq(work); 2092 if (pwq && pwq->pool == pool) { 2093 unsigned long work_data = *work_data_bits(work); 2094 2095 debug_work_deactivate(work); 2096 2097 /* 2098 * A cancelable inactive work item must be in the 2099 * pwq->inactive_works since a queued barrier can't be 2100 * canceled (see the comments in insert_wq_barrier()). 2101 * 2102 * An inactive work item cannot be deleted directly because 2103 * it might have linked barrier work items which, if left 2104 * on the inactive_works list, will confuse pwq->nr_active 2105 * management later on and cause stall. Move the linked 2106 * barrier work items to the worklist when deleting the grabbed 2107 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that 2108 * it doesn't participate in nr_active management in later 2109 * pwq_dec_nr_in_flight(). 2110 */ 2111 if (work_data & WORK_STRUCT_INACTIVE) 2112 move_linked_works(work, &pwq->pool->worklist, NULL); 2113 2114 list_del_init(&work->entry); 2115 2116 /* 2117 * work->data points to pwq iff queued. Let's point to pool. As 2118 * this destroys work->data needed by the next step, stash it. 2119 */ 2120 set_work_pool_and_keep_pending(work, pool->id, 2121 pool_offq_flags(pool)); 2122 2123 /* must be the last step, see the function comment */ 2124 pwq_dec_nr_in_flight(pwq, work_data); 2125 2126 raw_spin_unlock(&pool->lock); 2127 rcu_read_unlock(); 2128 return 1; 2129 } 2130 raw_spin_unlock(&pool->lock); 2131 fail: 2132 rcu_read_unlock(); 2133 local_irq_restore(*irq_flags); 2134 return -EAGAIN; 2135 } 2136 2137 /** 2138 * work_grab_pending - steal work item from worklist and disable irq 2139 * @work: work item to steal 2140 * @cflags: %WORK_CANCEL_ flags 2141 * @irq_flags: place to store IRQ state 2142 * 2143 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer 2144 * or on worklist. 2145 * 2146 * Can be called from any context. IRQ is disabled on return with IRQ state 2147 * stored in *@irq_flags. The caller is responsible for re-enabling it using 2148 * local_irq_restore(). 2149 * 2150 * Returns %true if @work was pending. %false if idle. 2151 */ 2152 static bool work_grab_pending(struct work_struct *work, u32 cflags, 2153 unsigned long *irq_flags) 2154 { 2155 int ret; 2156 2157 while (true) { 2158 ret = try_to_grab_pending(work, cflags, irq_flags); 2159 if (ret >= 0) 2160 return ret; 2161 cpu_relax(); 2162 } 2163 } 2164 2165 /** 2166 * insert_work - insert a work into a pool 2167 * @pwq: pwq @work belongs to 2168 * @work: work to insert 2169 * @head: insertion point 2170 * @extra_flags: extra WORK_STRUCT_* flags to set 2171 * 2172 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 2173 * work_struct flags. 2174 * 2175 * CONTEXT: 2176 * raw_spin_lock_irq(pool->lock). 2177 */ 2178 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 2179 struct list_head *head, unsigned int extra_flags) 2180 { 2181 debug_work_activate(work); 2182 2183 /* record the work call stack in order to print it in KASAN reports */ 2184 kasan_record_aux_stack(work); 2185 2186 /* we own @work, set data and link */ 2187 set_work_pwq(work, pwq, extra_flags); 2188 list_add_tail(&work->entry, head); 2189 get_pwq(pwq); 2190 } 2191 2192 /* 2193 * Test whether @work is being queued from another work executing on the 2194 * same workqueue. 2195 */ 2196 static bool is_chained_work(struct workqueue_struct *wq) 2197 { 2198 struct worker *worker; 2199 2200 worker = current_wq_worker(); 2201 /* 2202 * Return %true iff I'm a worker executing a work item on @wq. If 2203 * I'm @worker, it's safe to dereference it without locking. 2204 */ 2205 return worker && worker->current_pwq->wq == wq; 2206 } 2207 2208 /* 2209 * When queueing an unbound work item to a wq, prefer local CPU if allowed 2210 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 2211 * avoid perturbing sensitive tasks. 2212 */ 2213 static int wq_select_unbound_cpu(int cpu) 2214 { 2215 int new_cpu; 2216 2217 if (likely(!wq_debug_force_rr_cpu)) { 2218 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 2219 return cpu; 2220 } else { 2221 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 2222 } 2223 2224 new_cpu = __this_cpu_read(wq_rr_cpu_last); 2225 new_cpu = cpumask_next_and_wrap(new_cpu, wq_unbound_cpumask, cpu_online_mask); 2226 if (unlikely(new_cpu >= nr_cpu_ids)) 2227 return cpu; 2228 __this_cpu_write(wq_rr_cpu_last, new_cpu); 2229 2230 return new_cpu; 2231 } 2232 2233 static void __queue_work(int cpu, struct workqueue_struct *wq, 2234 struct work_struct *work) 2235 { 2236 struct pool_workqueue *pwq; 2237 struct worker_pool *last_pool, *pool; 2238 unsigned int work_flags; 2239 unsigned int req_cpu = cpu; 2240 2241 /* 2242 * While a work item is PENDING && off queue, a task trying to 2243 * steal the PENDING will busy-loop waiting for it to either get 2244 * queued or lose PENDING. Grabbing PENDING and queueing should 2245 * happen with IRQ disabled. 2246 */ 2247 lockdep_assert_irqs_disabled(); 2248 2249 /* 2250 * For a draining wq, only works from the same workqueue are 2251 * allowed. The __WQ_DESTROYING helps to spot the issue that 2252 * queues a new work item to a wq after destroy_workqueue(wq). 2253 */ 2254 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 2255 WARN_ONCE(!is_chained_work(wq), "workqueue: cannot queue %ps on wq %s\n", 2256 work->func, wq->name))) { 2257 return; 2258 } 2259 rcu_read_lock(); 2260 retry: 2261 /* pwq which will be used unless @work is executing elsewhere */ 2262 if (req_cpu == WORK_CPU_UNBOUND) { 2263 if (wq->flags & WQ_UNBOUND) 2264 cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 2265 else 2266 cpu = raw_smp_processor_id(); 2267 } 2268 2269 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 2270 pool = pwq->pool; 2271 2272 /* 2273 * If @work was previously on a different pool, it might still be 2274 * running there, in which case the work needs to be queued on that 2275 * pool to guarantee non-reentrancy. 2276 * 2277 * For ordered workqueue, work items must be queued on the newest pwq 2278 * for accurate order management. Guaranteed order also guarantees 2279 * non-reentrancy. See the comments above unplug_oldest_pwq(). 2280 */ 2281 last_pool = get_work_pool(work); 2282 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) { 2283 struct worker *worker; 2284 2285 raw_spin_lock(&last_pool->lock); 2286 2287 worker = find_worker_executing_work(last_pool, work); 2288 2289 if (worker && worker->current_pwq->wq == wq) { 2290 pwq = worker->current_pwq; 2291 pool = pwq->pool; 2292 WARN_ON_ONCE(pool != last_pool); 2293 } else { 2294 /* meh... not running there, queue here */ 2295 raw_spin_unlock(&last_pool->lock); 2296 raw_spin_lock(&pool->lock); 2297 } 2298 } else { 2299 raw_spin_lock(&pool->lock); 2300 } 2301 2302 /* 2303 * pwq is determined and locked. For unbound pools, we could have raced 2304 * with pwq release and it could already be dead. If its refcnt is zero, 2305 * repeat pwq selection. Note that unbound pwqs never die without 2306 * another pwq replacing it in cpu_pwq or while work items are executing 2307 * on it, so the retrying is guaranteed to make forward-progress. 2308 */ 2309 if (unlikely(!pwq->refcnt)) { 2310 if (wq->flags & WQ_UNBOUND) { 2311 raw_spin_unlock(&pool->lock); 2312 cpu_relax(); 2313 goto retry; 2314 } 2315 /* oops */ 2316 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 2317 wq->name, cpu); 2318 } 2319 2320 /* pwq determined, queue */ 2321 trace_workqueue_queue_work(req_cpu, pwq, work); 2322 2323 if (WARN_ON(!list_empty(&work->entry))) 2324 goto out; 2325 2326 pwq->nr_in_flight[pwq->work_color]++; 2327 work_flags = work_color_to_flags(pwq->work_color); 2328 2329 /* 2330 * Limit the number of concurrently active work items to max_active. 2331 * @work must also queue behind existing inactive work items to maintain 2332 * ordering when max_active changes. See wq_adjust_max_active(). 2333 */ 2334 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { 2335 if (list_empty(&pool->worklist)) 2336 pool->watchdog_ts = jiffies; 2337 2338 trace_workqueue_activate_work(work); 2339 insert_work(pwq, work, &pool->worklist, work_flags); 2340 kick_pool(pool); 2341 } else { 2342 work_flags |= WORK_STRUCT_INACTIVE; 2343 insert_work(pwq, work, &pwq->inactive_works, work_flags); 2344 } 2345 2346 out: 2347 raw_spin_unlock(&pool->lock); 2348 rcu_read_unlock(); 2349 } 2350 2351 static bool clear_pending_if_disabled(struct work_struct *work) 2352 { 2353 unsigned long data = *work_data_bits(work); 2354 struct work_offq_data offqd; 2355 2356 if (likely((data & WORK_STRUCT_PWQ) || 2357 !(data & WORK_OFFQ_DISABLE_MASK))) 2358 return false; 2359 2360 work_offqd_unpack(&offqd, data); 2361 set_work_pool_and_clear_pending(work, offqd.pool_id, 2362 work_offqd_pack_flags(&offqd)); 2363 return true; 2364 } 2365 2366 /** 2367 * queue_work_on - queue work on specific cpu 2368 * @cpu: CPU number to execute work on 2369 * @wq: workqueue to use 2370 * @work: work to queue 2371 * 2372 * We queue the work to a specific CPU, the caller must ensure it 2373 * can't go away. Callers that fail to ensure that the specified 2374 * CPU cannot go away will execute on a randomly chosen CPU. 2375 * But note well that callers specifying a CPU that never has been 2376 * online will get a splat. 2377 * 2378 * Return: %false if @work was already on a queue, %true otherwise. 2379 */ 2380 bool queue_work_on(int cpu, struct workqueue_struct *wq, 2381 struct work_struct *work) 2382 { 2383 bool ret = false; 2384 unsigned long irq_flags; 2385 2386 local_irq_save(irq_flags); 2387 2388 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2389 !clear_pending_if_disabled(work)) { 2390 __queue_work(cpu, wq, work); 2391 ret = true; 2392 } 2393 2394 local_irq_restore(irq_flags); 2395 return ret; 2396 } 2397 EXPORT_SYMBOL(queue_work_on); 2398 2399 /** 2400 * select_numa_node_cpu - Select a CPU based on NUMA node 2401 * @node: NUMA node ID that we want to select a CPU from 2402 * 2403 * This function will attempt to find a "random" cpu available on a given 2404 * node. If there are no CPUs available on the given node it will return 2405 * WORK_CPU_UNBOUND indicating that we should just schedule to any 2406 * available CPU if we need to schedule this work. 2407 */ 2408 static int select_numa_node_cpu(int node) 2409 { 2410 int cpu; 2411 2412 /* Delay binding to CPU if node is not valid or online */ 2413 if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 2414 return WORK_CPU_UNBOUND; 2415 2416 /* Use local node/cpu if we are already there */ 2417 cpu = raw_smp_processor_id(); 2418 if (node == cpu_to_node(cpu)) 2419 return cpu; 2420 2421 /* Use "random" otherwise know as "first" online CPU of node */ 2422 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 2423 2424 /* If CPU is valid return that, otherwise just defer */ 2425 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 2426 } 2427 2428 /** 2429 * queue_work_node - queue work on a "random" cpu for a given NUMA node 2430 * @node: NUMA node that we are targeting the work for 2431 * @wq: workqueue to use 2432 * @work: work to queue 2433 * 2434 * We queue the work to a "random" CPU within a given NUMA node. The basic 2435 * idea here is to provide a way to somehow associate work with a given 2436 * NUMA node. 2437 * 2438 * This function will only make a best effort attempt at getting this onto 2439 * the right NUMA node. If no node is requested or the requested node is 2440 * offline then we just fall back to standard queue_work behavior. 2441 * 2442 * Currently the "random" CPU ends up being the first available CPU in the 2443 * intersection of cpu_online_mask and the cpumask of the node, unless we 2444 * are running on the node. In that case we just use the current CPU. 2445 * 2446 * Return: %false if @work was already on a queue, %true otherwise. 2447 */ 2448 bool queue_work_node(int node, struct workqueue_struct *wq, 2449 struct work_struct *work) 2450 { 2451 unsigned long irq_flags; 2452 bool ret = false; 2453 2454 /* 2455 * This current implementation is specific to unbound workqueues. 2456 * Specifically we only return the first available CPU for a given 2457 * node instead of cycling through individual CPUs within the node. 2458 * 2459 * If this is used with a per-cpu workqueue then the logic in 2460 * workqueue_select_cpu_near would need to be updated to allow for 2461 * some round robin type logic. 2462 */ 2463 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 2464 2465 local_irq_save(irq_flags); 2466 2467 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2468 !clear_pending_if_disabled(work)) { 2469 int cpu = select_numa_node_cpu(node); 2470 2471 __queue_work(cpu, wq, work); 2472 ret = true; 2473 } 2474 2475 local_irq_restore(irq_flags); 2476 return ret; 2477 } 2478 EXPORT_SYMBOL_GPL(queue_work_node); 2479 2480 void delayed_work_timer_fn(struct timer_list *t) 2481 { 2482 struct delayed_work *dwork = timer_container_of(dwork, t, timer); 2483 2484 /* should have been called from irqsafe timer with irq already off */ 2485 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 2486 } 2487 EXPORT_SYMBOL(delayed_work_timer_fn); 2488 2489 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 2490 struct delayed_work *dwork, unsigned long delay) 2491 { 2492 struct timer_list *timer = &dwork->timer; 2493 struct work_struct *work = &dwork->work; 2494 2495 WARN_ON_ONCE(!wq); 2496 WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 2497 WARN_ON_ONCE(timer_pending(timer)); 2498 WARN_ON_ONCE(!list_empty(&work->entry)); 2499 2500 /* 2501 * If @delay is 0, queue @dwork->work immediately. This is for 2502 * both optimization and correctness. The earliest @timer can 2503 * expire is on the closest next tick and delayed_work users depend 2504 * on that there's no such delay when @delay is 0. 2505 */ 2506 if (!delay) { 2507 __queue_work(cpu, wq, &dwork->work); 2508 return; 2509 } 2510 2511 WARN_ON_ONCE(cpu != WORK_CPU_UNBOUND && !cpu_online(cpu)); 2512 dwork->wq = wq; 2513 dwork->cpu = cpu; 2514 timer->expires = jiffies + delay; 2515 2516 if (housekeeping_enabled(HK_TYPE_TIMER)) { 2517 /* If the current cpu is a housekeeping cpu, use it. */ 2518 cpu = smp_processor_id(); 2519 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) 2520 cpu = housekeeping_any_cpu(HK_TYPE_TIMER); 2521 add_timer_on(timer, cpu); 2522 } else { 2523 if (likely(cpu == WORK_CPU_UNBOUND)) 2524 add_timer_global(timer); 2525 else 2526 add_timer_on(timer, cpu); 2527 } 2528 } 2529 2530 /** 2531 * queue_delayed_work_on - queue work on specific CPU after delay 2532 * @cpu: CPU number to execute work on 2533 * @wq: workqueue to use 2534 * @dwork: work to queue 2535 * @delay: number of jiffies to wait before queueing 2536 * 2537 * We queue the delayed_work to a specific CPU, for non-zero delays the 2538 * caller must ensure it is online and can't go away. Callers that fail 2539 * to ensure this, may get @dwork->timer queued to an offlined CPU and 2540 * this will prevent queueing of @dwork->work unless the offlined CPU 2541 * becomes online again. 2542 * 2543 * Return: %false if @work was already on a queue, %true otherwise. If 2544 * @delay is zero and @dwork is idle, it will be scheduled for immediate 2545 * execution. 2546 */ 2547 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 2548 struct delayed_work *dwork, unsigned long delay) 2549 { 2550 struct work_struct *work = &dwork->work; 2551 bool ret = false; 2552 unsigned long irq_flags; 2553 2554 /* read the comment in __queue_work() */ 2555 local_irq_save(irq_flags); 2556 2557 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2558 !clear_pending_if_disabled(work)) { 2559 __queue_delayed_work(cpu, wq, dwork, delay); 2560 ret = true; 2561 } 2562 2563 local_irq_restore(irq_flags); 2564 return ret; 2565 } 2566 EXPORT_SYMBOL(queue_delayed_work_on); 2567 2568 /** 2569 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 2570 * @cpu: CPU number to execute work on 2571 * @wq: workqueue to use 2572 * @dwork: work to queue 2573 * @delay: number of jiffies to wait before queueing 2574 * 2575 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 2576 * modify @dwork's timer so that it expires after @delay. If @delay is 2577 * zero, @work is guaranteed to be scheduled immediately regardless of its 2578 * current state. 2579 * 2580 * Return: %false if @dwork was idle and queued, %true if @dwork was 2581 * pending and its timer was modified. 2582 * 2583 * This function is safe to call from any context including IRQ handler. 2584 * See try_to_grab_pending() for details. 2585 */ 2586 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 2587 struct delayed_work *dwork, unsigned long delay) 2588 { 2589 unsigned long irq_flags; 2590 bool ret; 2591 2592 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags); 2593 2594 if (!clear_pending_if_disabled(&dwork->work)) 2595 __queue_delayed_work(cpu, wq, dwork, delay); 2596 2597 local_irq_restore(irq_flags); 2598 return ret; 2599 } 2600 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2601 2602 static void rcu_work_rcufn(struct rcu_head *rcu) 2603 { 2604 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 2605 2606 /* read the comment in __queue_work() */ 2607 local_irq_disable(); 2608 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 2609 local_irq_enable(); 2610 } 2611 2612 /** 2613 * queue_rcu_work - queue work after a RCU grace period 2614 * @wq: workqueue to use 2615 * @rwork: work to queue 2616 * 2617 * Return: %false if @rwork was already pending, %true otherwise. Note 2618 * that a full RCU grace period is guaranteed only after a %true return. 2619 * While @rwork is guaranteed to be executed after a %false return, the 2620 * execution may happen before a full RCU grace period has passed. 2621 */ 2622 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 2623 { 2624 struct work_struct *work = &rwork->work; 2625 2626 /* 2627 * rcu_work can't be canceled or disabled. Warn if the user reached 2628 * inside @rwork and disabled the inner work. 2629 */ 2630 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2631 !WARN_ON_ONCE(clear_pending_if_disabled(work))) { 2632 rwork->wq = wq; 2633 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 2634 return true; 2635 } 2636 2637 return false; 2638 } 2639 EXPORT_SYMBOL(queue_rcu_work); 2640 2641 static struct worker *alloc_worker(int node) 2642 { 2643 struct worker *worker; 2644 2645 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2646 if (worker) { 2647 INIT_LIST_HEAD(&worker->entry); 2648 INIT_LIST_HEAD(&worker->scheduled); 2649 INIT_LIST_HEAD(&worker->node); 2650 /* on creation a worker is in !idle && prep state */ 2651 worker->flags = WORKER_PREP; 2652 } 2653 return worker; 2654 } 2655 2656 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2657 { 2658 if (pool->cpu < 0 && pool->attrs->affn_strict) 2659 return pool->attrs->__pod_cpumask; 2660 else 2661 return pool->attrs->cpumask; 2662 } 2663 2664 /** 2665 * worker_attach_to_pool() - attach a worker to a pool 2666 * @worker: worker to be attached 2667 * @pool: the target pool 2668 * 2669 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 2670 * cpu-binding of @worker are kept coordinated with the pool across 2671 * cpu-[un]hotplugs. 2672 */ 2673 static void worker_attach_to_pool(struct worker *worker, 2674 struct worker_pool *pool) 2675 { 2676 mutex_lock(&wq_pool_attach_mutex); 2677 2678 /* 2679 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable 2680 * across this function. See the comments above the flag definition for 2681 * details. BH workers are, while per-CPU, always DISASSOCIATED. 2682 */ 2683 if (pool->flags & POOL_DISASSOCIATED) { 2684 worker->flags |= WORKER_UNBOUND; 2685 } else { 2686 WARN_ON_ONCE(pool->flags & POOL_BH); 2687 kthread_set_per_cpu(worker->task, pool->cpu); 2688 } 2689 2690 if (worker->rescue_wq) 2691 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2692 2693 list_add_tail(&worker->node, &pool->workers); 2694 worker->pool = pool; 2695 2696 mutex_unlock(&wq_pool_attach_mutex); 2697 } 2698 2699 static void unbind_worker(struct worker *worker) 2700 { 2701 lockdep_assert_held(&wq_pool_attach_mutex); 2702 2703 kthread_set_per_cpu(worker->task, -1); 2704 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2705 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2706 else 2707 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2708 } 2709 2710 2711 static void detach_worker(struct worker *worker) 2712 { 2713 lockdep_assert_held(&wq_pool_attach_mutex); 2714 2715 unbind_worker(worker); 2716 list_del(&worker->node); 2717 } 2718 2719 /** 2720 * worker_detach_from_pool() - detach a worker from its pool 2721 * @worker: worker which is attached to its pool 2722 * 2723 * Undo the attaching which had been done in worker_attach_to_pool(). The 2724 * caller worker shouldn't access to the pool after detached except it has 2725 * other reference to the pool. 2726 */ 2727 static void worker_detach_from_pool(struct worker *worker) 2728 { 2729 struct worker_pool *pool = worker->pool; 2730 2731 /* there is one permanent BH worker per CPU which should never detach */ 2732 WARN_ON_ONCE(pool->flags & POOL_BH); 2733 2734 mutex_lock(&wq_pool_attach_mutex); 2735 detach_worker(worker); 2736 worker->pool = NULL; 2737 mutex_unlock(&wq_pool_attach_mutex); 2738 2739 /* clear leftover flags without pool->lock after it is detached */ 2740 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2741 } 2742 2743 static int format_worker_id(char *buf, size_t size, struct worker *worker, 2744 struct worker_pool *pool) 2745 { 2746 if (worker->rescue_wq) 2747 return scnprintf(buf, size, "kworker/R-%s", 2748 worker->rescue_wq->name); 2749 2750 if (pool) { 2751 if (pool->cpu >= 0) 2752 return scnprintf(buf, size, "kworker/%d:%d%s", 2753 pool->cpu, worker->id, 2754 pool->attrs->nice < 0 ? "H" : ""); 2755 else 2756 return scnprintf(buf, size, "kworker/u%d:%d", 2757 pool->id, worker->id); 2758 } else { 2759 return scnprintf(buf, size, "kworker/dying"); 2760 } 2761 } 2762 2763 /** 2764 * create_worker - create a new workqueue worker 2765 * @pool: pool the new worker will belong to 2766 * 2767 * Create and start a new worker which is attached to @pool. 2768 * 2769 * CONTEXT: 2770 * Might sleep. Does GFP_KERNEL allocations. 2771 * 2772 * Return: 2773 * Pointer to the newly created worker. 2774 */ 2775 static struct worker *create_worker(struct worker_pool *pool) 2776 { 2777 struct worker *worker; 2778 int id; 2779 2780 /* ID is needed to determine kthread name */ 2781 id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 2782 if (id < 0) { 2783 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 2784 ERR_PTR(id)); 2785 return NULL; 2786 } 2787 2788 worker = alloc_worker(pool->node); 2789 if (!worker) { 2790 pr_err_once("workqueue: Failed to allocate a worker\n"); 2791 goto fail; 2792 } 2793 2794 worker->id = id; 2795 2796 if (!(pool->flags & POOL_BH)) { 2797 char id_buf[WORKER_ID_LEN]; 2798 2799 format_worker_id(id_buf, sizeof(id_buf), worker, pool); 2800 worker->task = kthread_create_on_node(worker_thread, worker, 2801 pool->node, "%s", id_buf); 2802 if (IS_ERR(worker->task)) { 2803 if (PTR_ERR(worker->task) == -EINTR) { 2804 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n", 2805 id_buf); 2806 } else { 2807 pr_err_once("workqueue: Failed to create a worker thread: %pe", 2808 worker->task); 2809 } 2810 goto fail; 2811 } 2812 2813 set_user_nice(worker->task, pool->attrs->nice); 2814 kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 2815 } 2816 2817 /* successful, attach the worker to the pool */ 2818 worker_attach_to_pool(worker, pool); 2819 2820 /* start the newly created worker */ 2821 raw_spin_lock_irq(&pool->lock); 2822 2823 worker->pool->nr_workers++; 2824 worker_enter_idle(worker); 2825 2826 /* 2827 * @worker is waiting on a completion in kthread() and will trigger hung 2828 * check if not woken up soon. As kick_pool() is noop if @pool is empty, 2829 * wake it up explicitly. 2830 */ 2831 if (worker->task) 2832 wake_up_process(worker->task); 2833 2834 raw_spin_unlock_irq(&pool->lock); 2835 2836 return worker; 2837 2838 fail: 2839 ida_free(&pool->worker_ida, id); 2840 kfree(worker); 2841 return NULL; 2842 } 2843 2844 static void detach_dying_workers(struct list_head *cull_list) 2845 { 2846 struct worker *worker; 2847 2848 list_for_each_entry(worker, cull_list, entry) 2849 detach_worker(worker); 2850 } 2851 2852 static void reap_dying_workers(struct list_head *cull_list) 2853 { 2854 struct worker *worker, *tmp; 2855 2856 list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2857 list_del_init(&worker->entry); 2858 kthread_stop_put(worker->task); 2859 kfree(worker); 2860 } 2861 } 2862 2863 /** 2864 * set_worker_dying - Tag a worker for destruction 2865 * @worker: worker to be destroyed 2866 * @list: transfer worker away from its pool->idle_list and into list 2867 * 2868 * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2869 * should be idle. 2870 * 2871 * CONTEXT: 2872 * raw_spin_lock_irq(pool->lock). 2873 */ 2874 static void set_worker_dying(struct worker *worker, struct list_head *list) 2875 { 2876 struct worker_pool *pool = worker->pool; 2877 2878 lockdep_assert_held(&pool->lock); 2879 lockdep_assert_held(&wq_pool_attach_mutex); 2880 2881 /* sanity check frenzy */ 2882 if (WARN_ON(worker->current_work) || 2883 WARN_ON(!list_empty(&worker->scheduled)) || 2884 WARN_ON(!(worker->flags & WORKER_IDLE))) 2885 return; 2886 2887 pool->nr_workers--; 2888 pool->nr_idle--; 2889 2890 worker->flags |= WORKER_DIE; 2891 2892 list_move(&worker->entry, list); 2893 2894 /* get an extra task struct reference for later kthread_stop_put() */ 2895 get_task_struct(worker->task); 2896 } 2897 2898 /** 2899 * idle_worker_timeout - check if some idle workers can now be deleted. 2900 * @t: The pool's idle_timer that just expired 2901 * 2902 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 2903 * worker_leave_idle(), as a worker flicking between idle and active while its 2904 * pool is at the too_many_workers() tipping point would cause too much timer 2905 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 2906 * it expire and re-evaluate things from there. 2907 */ 2908 static void idle_worker_timeout(struct timer_list *t) 2909 { 2910 struct worker_pool *pool = timer_container_of(pool, t, idle_timer); 2911 bool do_cull = false; 2912 2913 if (work_pending(&pool->idle_cull_work)) 2914 return; 2915 2916 raw_spin_lock_irq(&pool->lock); 2917 2918 if (too_many_workers(pool)) { 2919 struct worker *worker; 2920 unsigned long expires; 2921 2922 /* idle_list is kept in LIFO order, check the last one */ 2923 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2924 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2925 do_cull = !time_before(jiffies, expires); 2926 2927 if (!do_cull) 2928 mod_timer(&pool->idle_timer, expires); 2929 } 2930 raw_spin_unlock_irq(&pool->lock); 2931 2932 if (do_cull) 2933 queue_work(system_unbound_wq, &pool->idle_cull_work); 2934 } 2935 2936 /** 2937 * idle_cull_fn - cull workers that have been idle for too long. 2938 * @work: the pool's work for handling these idle workers 2939 * 2940 * This goes through a pool's idle workers and gets rid of those that have been 2941 * idle for at least IDLE_WORKER_TIMEOUT seconds. 2942 * 2943 * We don't want to disturb isolated CPUs because of a pcpu kworker being 2944 * culled, so this also resets worker affinity. This requires a sleepable 2945 * context, hence the split between timer callback and work item. 2946 */ 2947 static void idle_cull_fn(struct work_struct *work) 2948 { 2949 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2950 LIST_HEAD(cull_list); 2951 2952 /* 2953 * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2954 * cannot proceed beyong set_pf_worker() in its self-destruct path. 2955 * This is required as a previously-preempted worker could run after 2956 * set_worker_dying() has happened but before detach_dying_workers() did. 2957 */ 2958 mutex_lock(&wq_pool_attach_mutex); 2959 raw_spin_lock_irq(&pool->lock); 2960 2961 while (too_many_workers(pool)) { 2962 struct worker *worker; 2963 unsigned long expires; 2964 2965 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2966 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2967 2968 if (time_before(jiffies, expires)) { 2969 mod_timer(&pool->idle_timer, expires); 2970 break; 2971 } 2972 2973 set_worker_dying(worker, &cull_list); 2974 } 2975 2976 raw_spin_unlock_irq(&pool->lock); 2977 detach_dying_workers(&cull_list); 2978 mutex_unlock(&wq_pool_attach_mutex); 2979 2980 reap_dying_workers(&cull_list); 2981 } 2982 2983 static void send_mayday(struct work_struct *work) 2984 { 2985 struct pool_workqueue *pwq = get_work_pwq(work); 2986 struct workqueue_struct *wq = pwq->wq; 2987 2988 lockdep_assert_held(&wq_mayday_lock); 2989 2990 if (!wq->rescuer) 2991 return; 2992 2993 /* mayday mayday mayday */ 2994 if (list_empty(&pwq->mayday_node)) { 2995 /* 2996 * If @pwq is for an unbound wq, its base ref may be put at 2997 * any time due to an attribute change. Pin @pwq until the 2998 * rescuer is done with it. 2999 */ 3000 get_pwq(pwq); 3001 list_add_tail(&pwq->mayday_node, &wq->maydays); 3002 wake_up_process(wq->rescuer->task); 3003 pwq->stats[PWQ_STAT_MAYDAY]++; 3004 } 3005 } 3006 3007 static void pool_mayday_timeout(struct timer_list *t) 3008 { 3009 struct worker_pool *pool = timer_container_of(pool, t, mayday_timer); 3010 struct work_struct *work; 3011 3012 raw_spin_lock_irq(&pool->lock); 3013 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 3014 3015 if (need_to_create_worker(pool)) { 3016 /* 3017 * We've been trying to create a new worker but 3018 * haven't been successful. We might be hitting an 3019 * allocation deadlock. Send distress signals to 3020 * rescuers. 3021 */ 3022 list_for_each_entry(work, &pool->worklist, entry) 3023 send_mayday(work); 3024 } 3025 3026 raw_spin_unlock(&wq_mayday_lock); 3027 raw_spin_unlock_irq(&pool->lock); 3028 3029 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 3030 } 3031 3032 /** 3033 * maybe_create_worker - create a new worker if necessary 3034 * @pool: pool to create a new worker for 3035 * 3036 * Create a new worker for @pool if necessary. @pool is guaranteed to 3037 * have at least one idle worker on return from this function. If 3038 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 3039 * sent to all rescuers with works scheduled on @pool to resolve 3040 * possible allocation deadlock. 3041 * 3042 * On return, need_to_create_worker() is guaranteed to be %false and 3043 * may_start_working() %true. 3044 * 3045 * LOCKING: 3046 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3047 * multiple times. Does GFP_KERNEL allocations. Called only from 3048 * manager. 3049 */ 3050 static void maybe_create_worker(struct worker_pool *pool) 3051 __releases(&pool->lock) 3052 __acquires(&pool->lock) 3053 { 3054 restart: 3055 raw_spin_unlock_irq(&pool->lock); 3056 3057 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 3058 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 3059 3060 while (true) { 3061 if (create_worker(pool) || !need_to_create_worker(pool)) 3062 break; 3063 3064 schedule_timeout_interruptible(CREATE_COOLDOWN); 3065 3066 if (!need_to_create_worker(pool)) 3067 break; 3068 } 3069 3070 timer_delete_sync(&pool->mayday_timer); 3071 raw_spin_lock_irq(&pool->lock); 3072 /* 3073 * This is necessary even after a new worker was just successfully 3074 * created as @pool->lock was dropped and the new worker might have 3075 * already become busy. 3076 */ 3077 if (need_to_create_worker(pool)) 3078 goto restart; 3079 } 3080 3081 /** 3082 * manage_workers - manage worker pool 3083 * @worker: self 3084 * 3085 * Assume the manager role and manage the worker pool @worker belongs 3086 * to. At any given time, there can be only zero or one manager per 3087 * pool. The exclusion is handled automatically by this function. 3088 * 3089 * The caller can safely start processing works on false return. On 3090 * true return, it's guaranteed that need_to_create_worker() is false 3091 * and may_start_working() is true. 3092 * 3093 * CONTEXT: 3094 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3095 * multiple times. Does GFP_KERNEL allocations. 3096 * 3097 * Return: 3098 * %false if the pool doesn't need management and the caller can safely 3099 * start processing works, %true if management function was performed and 3100 * the conditions that the caller verified before calling the function may 3101 * no longer be true. 3102 */ 3103 static bool manage_workers(struct worker *worker) 3104 { 3105 struct worker_pool *pool = worker->pool; 3106 3107 if (pool->flags & POOL_MANAGER_ACTIVE) 3108 return false; 3109 3110 pool->flags |= POOL_MANAGER_ACTIVE; 3111 pool->manager = worker; 3112 3113 maybe_create_worker(pool); 3114 3115 pool->manager = NULL; 3116 pool->flags &= ~POOL_MANAGER_ACTIVE; 3117 rcuwait_wake_up(&manager_wait); 3118 return true; 3119 } 3120 3121 /** 3122 * process_one_work - process single work 3123 * @worker: self 3124 * @work: work to process 3125 * 3126 * Process @work. This function contains all the logics necessary to 3127 * process a single work including synchronization against and 3128 * interaction with other workers on the same cpu, queueing and 3129 * flushing. As long as context requirement is met, any worker can 3130 * call this function to process a work. 3131 * 3132 * CONTEXT: 3133 * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 3134 */ 3135 static void process_one_work(struct worker *worker, struct work_struct *work) 3136 __releases(&pool->lock) 3137 __acquires(&pool->lock) 3138 { 3139 struct pool_workqueue *pwq = get_work_pwq(work); 3140 struct worker_pool *pool = worker->pool; 3141 unsigned long work_data; 3142 int lockdep_start_depth, rcu_start_depth; 3143 bool bh_draining = pool->flags & POOL_BH_DRAINING; 3144 #ifdef CONFIG_LOCKDEP 3145 /* 3146 * It is permissible to free the struct work_struct from 3147 * inside the function that is called from it, this we need to 3148 * take into account for lockdep too. To avoid bogus "held 3149 * lock freed" warnings as well as problems when looking into 3150 * work->lockdep_map, make a copy and use that here. 3151 */ 3152 struct lockdep_map lockdep_map; 3153 3154 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 3155 #endif 3156 /* ensure we're on the correct CPU */ 3157 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 3158 raw_smp_processor_id() != pool->cpu); 3159 3160 /* claim and dequeue */ 3161 debug_work_deactivate(work); 3162 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 3163 worker->current_work = work; 3164 worker->current_func = work->func; 3165 worker->current_pwq = pwq; 3166 if (worker->task) 3167 worker->current_at = worker->task->se.sum_exec_runtime; 3168 work_data = *work_data_bits(work); 3169 worker->current_color = get_work_color(work_data); 3170 3171 /* 3172 * Record wq name for cmdline and debug reporting, may get 3173 * overridden through set_worker_desc(). 3174 */ 3175 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 3176 3177 list_del_init(&work->entry); 3178 3179 /* 3180 * CPU intensive works don't participate in concurrency management. 3181 * They're the scheduler's responsibility. This takes @worker out 3182 * of concurrency management and the next code block will chain 3183 * execution of the pending work items. 3184 */ 3185 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 3186 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 3187 3188 /* 3189 * Kick @pool if necessary. It's always noop for per-cpu worker pools 3190 * since nr_running would always be >= 1 at this point. This is used to 3191 * chain execution of the pending work items for WORKER_NOT_RUNNING 3192 * workers such as the UNBOUND and CPU_INTENSIVE ones. 3193 */ 3194 kick_pool(pool); 3195 3196 /* 3197 * Record the last pool and clear PENDING which should be the last 3198 * update to @work. Also, do this inside @pool->lock so that 3199 * PENDING and queued state changes happen together while IRQ is 3200 * disabled. 3201 */ 3202 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool)); 3203 3204 pwq->stats[PWQ_STAT_STARTED]++; 3205 raw_spin_unlock_irq(&pool->lock); 3206 3207 rcu_start_depth = rcu_preempt_depth(); 3208 lockdep_start_depth = lockdep_depth(current); 3209 /* see drain_dead_softirq_workfn() */ 3210 if (!bh_draining) 3211 lock_map_acquire(pwq->wq->lockdep_map); 3212 lock_map_acquire(&lockdep_map); 3213 /* 3214 * Strictly speaking we should mark the invariant state without holding 3215 * any locks, that is, before these two lock_map_acquire()'s. 3216 * 3217 * However, that would result in: 3218 * 3219 * A(W1) 3220 * WFC(C) 3221 * A(W1) 3222 * C(C) 3223 * 3224 * Which would create W1->C->W1 dependencies, even though there is no 3225 * actual deadlock possible. There are two solutions, using a 3226 * read-recursive acquire on the work(queue) 'locks', but this will then 3227 * hit the lockdep limitation on recursive locks, or simply discard 3228 * these locks. 3229 * 3230 * AFAICT there is no possible deadlock scenario between the 3231 * flush_work() and complete() primitives (except for single-threaded 3232 * workqueues), so hiding them isn't a problem. 3233 */ 3234 lockdep_invariant_state(true); 3235 trace_workqueue_execute_start(work); 3236 worker->current_func(work); 3237 /* 3238 * While we must be careful to not use "work" after this, the trace 3239 * point will only record its address. 3240 */ 3241 trace_workqueue_execute_end(work, worker->current_func); 3242 3243 lock_map_release(&lockdep_map); 3244 if (!bh_draining) 3245 lock_map_release(pwq->wq->lockdep_map); 3246 3247 if (unlikely((worker->task && in_atomic()) || 3248 lockdep_depth(current) != lockdep_start_depth || 3249 rcu_preempt_depth() != rcu_start_depth)) { 3250 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n" 3251 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n", 3252 current->comm, task_pid_nr(current), preempt_count(), 3253 lockdep_start_depth, lockdep_depth(current), 3254 rcu_start_depth, rcu_preempt_depth(), 3255 worker->current_func); 3256 debug_show_held_locks(current); 3257 dump_stack(); 3258 } 3259 3260 /* 3261 * The following prevents a kworker from hogging CPU on !PREEMPTION 3262 * kernels, where a requeueing work item waiting for something to 3263 * happen could deadlock with stop_machine as such work item could 3264 * indefinitely requeue itself while all other CPUs are trapped in 3265 * stop_machine. At the same time, report a quiescent RCU state so 3266 * the same condition doesn't freeze RCU. 3267 */ 3268 if (worker->task) 3269 cond_resched(); 3270 3271 raw_spin_lock_irq(&pool->lock); 3272 3273 pwq->stats[PWQ_STAT_COMPLETED]++; 3274 3275 /* 3276 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 3277 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 3278 * wq_cpu_intensive_thresh_us. Clear it. 3279 */ 3280 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 3281 3282 /* tag the worker for identification in schedule() */ 3283 worker->last_func = worker->current_func; 3284 3285 /* we're done with it, release */ 3286 hash_del(&worker->hentry); 3287 worker->current_work = NULL; 3288 worker->current_func = NULL; 3289 worker->current_pwq = NULL; 3290 worker->current_color = INT_MAX; 3291 3292 /* must be the last step, see the function comment */ 3293 pwq_dec_nr_in_flight(pwq, work_data); 3294 } 3295 3296 /** 3297 * process_scheduled_works - process scheduled works 3298 * @worker: self 3299 * 3300 * Process all scheduled works. Please note that the scheduled list 3301 * may change while processing a work, so this function repeatedly 3302 * fetches a work from the top and executes it. 3303 * 3304 * CONTEXT: 3305 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3306 * multiple times. 3307 */ 3308 static void process_scheduled_works(struct worker *worker) 3309 { 3310 struct work_struct *work; 3311 bool first = true; 3312 3313 while ((work = list_first_entry_or_null(&worker->scheduled, 3314 struct work_struct, entry))) { 3315 if (first) { 3316 worker->pool->watchdog_ts = jiffies; 3317 first = false; 3318 } 3319 process_one_work(worker, work); 3320 } 3321 } 3322 3323 static void set_pf_worker(bool val) 3324 { 3325 mutex_lock(&wq_pool_attach_mutex); 3326 if (val) 3327 current->flags |= PF_WQ_WORKER; 3328 else 3329 current->flags &= ~PF_WQ_WORKER; 3330 mutex_unlock(&wq_pool_attach_mutex); 3331 } 3332 3333 /** 3334 * worker_thread - the worker thread function 3335 * @__worker: self 3336 * 3337 * The worker thread function. All workers belong to a worker_pool - 3338 * either a per-cpu one or dynamic unbound one. These workers process all 3339 * work items regardless of their specific target workqueue. The only 3340 * exception is work items which belong to workqueues with a rescuer which 3341 * will be explained in rescuer_thread(). 3342 * 3343 * Return: 0 3344 */ 3345 static int worker_thread(void *__worker) 3346 { 3347 struct worker *worker = __worker; 3348 struct worker_pool *pool = worker->pool; 3349 3350 /* tell the scheduler that this is a workqueue worker */ 3351 set_pf_worker(true); 3352 woke_up: 3353 raw_spin_lock_irq(&pool->lock); 3354 3355 /* am I supposed to die? */ 3356 if (unlikely(worker->flags & WORKER_DIE)) { 3357 raw_spin_unlock_irq(&pool->lock); 3358 set_pf_worker(false); 3359 /* 3360 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool 3361 * shouldn't be accessed, reset it to NULL in case otherwise. 3362 */ 3363 worker->pool = NULL; 3364 ida_free(&pool->worker_ida, worker->id); 3365 return 0; 3366 } 3367 3368 worker_leave_idle(worker); 3369 recheck: 3370 /* no more worker necessary? */ 3371 if (!need_more_worker(pool)) 3372 goto sleep; 3373 3374 /* do we need to manage? */ 3375 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 3376 goto recheck; 3377 3378 /* 3379 * ->scheduled list can only be filled while a worker is 3380 * preparing to process a work or actually processing it. 3381 * Make sure nobody diddled with it while I was sleeping. 3382 */ 3383 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3384 3385 /* 3386 * Finish PREP stage. We're guaranteed to have at least one idle 3387 * worker or that someone else has already assumed the manager 3388 * role. This is where @worker starts participating in concurrency 3389 * management if applicable and concurrency management is restored 3390 * after being rebound. See rebind_workers() for details. 3391 */ 3392 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3393 3394 do { 3395 struct work_struct *work = 3396 list_first_entry(&pool->worklist, 3397 struct work_struct, entry); 3398 3399 if (assign_work(work, worker, NULL)) 3400 process_scheduled_works(worker); 3401 } while (keep_working(pool)); 3402 3403 worker_set_flags(worker, WORKER_PREP); 3404 sleep: 3405 /* 3406 * pool->lock is held and there's no work to process and no need to 3407 * manage, sleep. Workers are woken up only while holding 3408 * pool->lock or from local cpu, so setting the current state 3409 * before releasing pool->lock is enough to prevent losing any 3410 * event. 3411 */ 3412 worker_enter_idle(worker); 3413 __set_current_state(TASK_IDLE); 3414 raw_spin_unlock_irq(&pool->lock); 3415 schedule(); 3416 goto woke_up; 3417 } 3418 3419 /** 3420 * rescuer_thread - the rescuer thread function 3421 * @__rescuer: self 3422 * 3423 * Workqueue rescuer thread function. There's one rescuer for each 3424 * workqueue which has WQ_MEM_RECLAIM set. 3425 * 3426 * Regular work processing on a pool may block trying to create a new 3427 * worker which uses GFP_KERNEL allocation which has slight chance of 3428 * developing into deadlock if some works currently on the same queue 3429 * need to be processed to satisfy the GFP_KERNEL allocation. This is 3430 * the problem rescuer solves. 3431 * 3432 * When such condition is possible, the pool summons rescuers of all 3433 * workqueues which have works queued on the pool and let them process 3434 * those works so that forward progress can be guaranteed. 3435 * 3436 * This should happen rarely. 3437 * 3438 * Return: 0 3439 */ 3440 static int rescuer_thread(void *__rescuer) 3441 { 3442 struct worker *rescuer = __rescuer; 3443 struct workqueue_struct *wq = rescuer->rescue_wq; 3444 bool should_stop; 3445 3446 set_user_nice(current, RESCUER_NICE_LEVEL); 3447 3448 /* 3449 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 3450 * doesn't participate in concurrency management. 3451 */ 3452 set_pf_worker(true); 3453 repeat: 3454 set_current_state(TASK_IDLE); 3455 3456 /* 3457 * By the time the rescuer is requested to stop, the workqueue 3458 * shouldn't have any work pending, but @wq->maydays may still have 3459 * pwq(s) queued. This can happen by non-rescuer workers consuming 3460 * all the work items before the rescuer got to them. Go through 3461 * @wq->maydays processing before acting on should_stop so that the 3462 * list is always empty on exit. 3463 */ 3464 should_stop = kthread_should_stop(); 3465 3466 /* see whether any pwq is asking for help */ 3467 raw_spin_lock_irq(&wq_mayday_lock); 3468 3469 while (!list_empty(&wq->maydays)) { 3470 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 3471 struct pool_workqueue, mayday_node); 3472 struct worker_pool *pool = pwq->pool; 3473 struct work_struct *work, *n; 3474 3475 __set_current_state(TASK_RUNNING); 3476 list_del_init(&pwq->mayday_node); 3477 3478 raw_spin_unlock_irq(&wq_mayday_lock); 3479 3480 worker_attach_to_pool(rescuer, pool); 3481 3482 raw_spin_lock_irq(&pool->lock); 3483 3484 /* 3485 * Slurp in all works issued via this workqueue and 3486 * process'em. 3487 */ 3488 WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 3489 list_for_each_entry_safe(work, n, &pool->worklist, entry) { 3490 if (get_work_pwq(work) == pwq && 3491 assign_work(work, rescuer, &n)) 3492 pwq->stats[PWQ_STAT_RESCUED]++; 3493 } 3494 3495 if (!list_empty(&rescuer->scheduled)) { 3496 process_scheduled_works(rescuer); 3497 3498 /* 3499 * The above execution of rescued work items could 3500 * have created more to rescue through 3501 * pwq_activate_first_inactive() or chained 3502 * queueing. Let's put @pwq back on mayday list so 3503 * that such back-to-back work items, which may be 3504 * being used to relieve memory pressure, don't 3505 * incur MAYDAY_INTERVAL delay inbetween. 3506 */ 3507 if (pwq->nr_active && need_to_create_worker(pool)) { 3508 raw_spin_lock(&wq_mayday_lock); 3509 /* 3510 * Queue iff we aren't racing destruction 3511 * and somebody else hasn't queued it already. 3512 */ 3513 if (wq->rescuer && list_empty(&pwq->mayday_node)) { 3514 get_pwq(pwq); 3515 list_add_tail(&pwq->mayday_node, &wq->maydays); 3516 } 3517 raw_spin_unlock(&wq_mayday_lock); 3518 } 3519 } 3520 3521 /* 3522 * Leave this pool. Notify regular workers; otherwise, we end up 3523 * with 0 concurrency and stalling the execution. 3524 */ 3525 kick_pool(pool); 3526 3527 raw_spin_unlock_irq(&pool->lock); 3528 3529 worker_detach_from_pool(rescuer); 3530 3531 /* 3532 * Put the reference grabbed by send_mayday(). @pool might 3533 * go away any time after it. 3534 */ 3535 put_pwq_unlocked(pwq); 3536 3537 raw_spin_lock_irq(&wq_mayday_lock); 3538 } 3539 3540 raw_spin_unlock_irq(&wq_mayday_lock); 3541 3542 if (should_stop) { 3543 __set_current_state(TASK_RUNNING); 3544 set_pf_worker(false); 3545 return 0; 3546 } 3547 3548 /* rescuers should never participate in concurrency management */ 3549 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 3550 schedule(); 3551 goto repeat; 3552 } 3553 3554 static void bh_worker(struct worker *worker) 3555 { 3556 struct worker_pool *pool = worker->pool; 3557 int nr_restarts = BH_WORKER_RESTARTS; 3558 unsigned long end = jiffies + BH_WORKER_JIFFIES; 3559 3560 raw_spin_lock_irq(&pool->lock); 3561 worker_leave_idle(worker); 3562 3563 /* 3564 * This function follows the structure of worker_thread(). See there for 3565 * explanations on each step. 3566 */ 3567 if (!need_more_worker(pool)) 3568 goto done; 3569 3570 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3571 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3572 3573 do { 3574 struct work_struct *work = 3575 list_first_entry(&pool->worklist, 3576 struct work_struct, entry); 3577 3578 if (assign_work(work, worker, NULL)) 3579 process_scheduled_works(worker); 3580 } while (keep_working(pool) && 3581 --nr_restarts && time_before(jiffies, end)); 3582 3583 worker_set_flags(worker, WORKER_PREP); 3584 done: 3585 worker_enter_idle(worker); 3586 kick_pool(pool); 3587 raw_spin_unlock_irq(&pool->lock); 3588 } 3589 3590 /* 3591 * TODO: Convert all tasklet users to workqueue and use softirq directly. 3592 * 3593 * This is currently called from tasklet[_hi]action() and thus is also called 3594 * whenever there are tasklets to run. Let's do an early exit if there's nothing 3595 * queued. Once conversion from tasklet is complete, the need_more_worker() test 3596 * can be dropped. 3597 * 3598 * After full conversion, we'll add worker->softirq_action, directly use the 3599 * softirq action and obtain the worker pointer from the softirq_action pointer. 3600 */ 3601 void workqueue_softirq_action(bool highpri) 3602 { 3603 struct worker_pool *pool = 3604 &per_cpu(bh_worker_pools, smp_processor_id())[highpri]; 3605 if (need_more_worker(pool)) 3606 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3607 } 3608 3609 struct wq_drain_dead_softirq_work { 3610 struct work_struct work; 3611 struct worker_pool *pool; 3612 struct completion done; 3613 }; 3614 3615 static void drain_dead_softirq_workfn(struct work_struct *work) 3616 { 3617 struct wq_drain_dead_softirq_work *dead_work = 3618 container_of(work, struct wq_drain_dead_softirq_work, work); 3619 struct worker_pool *pool = dead_work->pool; 3620 bool repeat; 3621 3622 /* 3623 * @pool's CPU is dead and we want to execute its still pending work 3624 * items from this BH work item which is running on a different CPU. As 3625 * its CPU is dead, @pool can't be kicked and, as work execution path 3626 * will be nested, a lockdep annotation needs to be suppressed. Mark 3627 * @pool with %POOL_BH_DRAINING for the special treatments. 3628 */ 3629 raw_spin_lock_irq(&pool->lock); 3630 pool->flags |= POOL_BH_DRAINING; 3631 raw_spin_unlock_irq(&pool->lock); 3632 3633 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3634 3635 raw_spin_lock_irq(&pool->lock); 3636 pool->flags &= ~POOL_BH_DRAINING; 3637 repeat = need_more_worker(pool); 3638 raw_spin_unlock_irq(&pool->lock); 3639 3640 /* 3641 * bh_worker() might hit consecutive execution limit and bail. If there 3642 * still are pending work items, reschedule self and return so that we 3643 * don't hog this CPU's BH. 3644 */ 3645 if (repeat) { 3646 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3647 queue_work(system_bh_highpri_wq, work); 3648 else 3649 queue_work(system_bh_wq, work); 3650 } else { 3651 complete(&dead_work->done); 3652 } 3653 } 3654 3655 /* 3656 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's 3657 * possible to allocate dead_work per CPU and avoid flushing. However, then we 3658 * have to worry about draining overlapping with CPU coming back online or 3659 * nesting (one CPU's dead_work queued on another CPU which is also dead and so 3660 * on). Let's keep it simple and drain them synchronously. These are BH work 3661 * items which shouldn't be requeued on the same pool. Shouldn't take long. 3662 */ 3663 void workqueue_softirq_dead(unsigned int cpu) 3664 { 3665 int i; 3666 3667 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 3668 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i]; 3669 struct wq_drain_dead_softirq_work dead_work; 3670 3671 if (!need_more_worker(pool)) 3672 continue; 3673 3674 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn); 3675 dead_work.pool = pool; 3676 init_completion(&dead_work.done); 3677 3678 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3679 queue_work(system_bh_highpri_wq, &dead_work.work); 3680 else 3681 queue_work(system_bh_wq, &dead_work.work); 3682 3683 wait_for_completion(&dead_work.done); 3684 destroy_work_on_stack(&dead_work.work); 3685 } 3686 } 3687 3688 /** 3689 * check_flush_dependency - check for flush dependency sanity 3690 * @target_wq: workqueue being flushed 3691 * @target_work: work item being flushed (NULL for workqueue flushes) 3692 * @from_cancel: are we called from the work cancel path 3693 * 3694 * %current is trying to flush the whole @target_wq or @target_work on it. 3695 * If this is not the cancel path (which implies work being flushed is either 3696 * already running, or will not be at all), check if @target_wq doesn't have 3697 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running 3698 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward- 3699 * progress guarantee leading to a deadlock. 3700 */ 3701 static void check_flush_dependency(struct workqueue_struct *target_wq, 3702 struct work_struct *target_work, 3703 bool from_cancel) 3704 { 3705 work_func_t target_func; 3706 struct worker *worker; 3707 3708 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM) 3709 return; 3710 3711 worker = current_wq_worker(); 3712 target_func = target_work ? target_work->func : NULL; 3713 3714 WARN_ONCE(current->flags & PF_MEMALLOC, 3715 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 3716 current->pid, current->comm, target_wq->name, target_func); 3717 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 3718 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 3719 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 3720 worker->current_pwq->wq->name, worker->current_func, 3721 target_wq->name, target_func); 3722 } 3723 3724 struct wq_barrier { 3725 struct work_struct work; 3726 struct completion done; 3727 struct task_struct *task; /* purely informational */ 3728 }; 3729 3730 static void wq_barrier_func(struct work_struct *work) 3731 { 3732 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 3733 complete(&barr->done); 3734 } 3735 3736 /** 3737 * insert_wq_barrier - insert a barrier work 3738 * @pwq: pwq to insert barrier into 3739 * @barr: wq_barrier to insert 3740 * @target: target work to attach @barr to 3741 * @worker: worker currently executing @target, NULL if @target is not executing 3742 * 3743 * @barr is linked to @target such that @barr is completed only after 3744 * @target finishes execution. Please note that the ordering 3745 * guarantee is observed only with respect to @target and on the local 3746 * cpu. 3747 * 3748 * Currently, a queued barrier can't be canceled. This is because 3749 * try_to_grab_pending() can't determine whether the work to be 3750 * grabbed is at the head of the queue and thus can't clear LINKED 3751 * flag of the previous work while there must be a valid next work 3752 * after a work with LINKED flag set. 3753 * 3754 * Note that when @worker is non-NULL, @target may be modified 3755 * underneath us, so we can't reliably determine pwq from @target. 3756 * 3757 * CONTEXT: 3758 * raw_spin_lock_irq(pool->lock). 3759 */ 3760 static void insert_wq_barrier(struct pool_workqueue *pwq, 3761 struct wq_barrier *barr, 3762 struct work_struct *target, struct worker *worker) 3763 { 3764 static __maybe_unused struct lock_class_key bh_key, thr_key; 3765 unsigned int work_flags = 0; 3766 unsigned int work_color; 3767 struct list_head *head; 3768 3769 /* 3770 * debugobject calls are safe here even with pool->lock locked 3771 * as we know for sure that this will not trigger any of the 3772 * checks and call back into the fixup functions where we 3773 * might deadlock. 3774 * 3775 * BH and threaded workqueues need separate lockdep keys to avoid 3776 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} 3777 * usage". 3778 */ 3779 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func, 3780 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key); 3781 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 3782 3783 init_completion_map(&barr->done, &target->lockdep_map); 3784 3785 barr->task = current; 3786 3787 /* The barrier work item does not participate in nr_active. */ 3788 work_flags |= WORK_STRUCT_INACTIVE; 3789 3790 /* 3791 * If @target is currently being executed, schedule the 3792 * barrier to the worker; otherwise, put it after @target. 3793 */ 3794 if (worker) { 3795 head = worker->scheduled.next; 3796 work_color = worker->current_color; 3797 } else { 3798 unsigned long *bits = work_data_bits(target); 3799 3800 head = target->entry.next; 3801 /* there can already be other linked works, inherit and set */ 3802 work_flags |= *bits & WORK_STRUCT_LINKED; 3803 work_color = get_work_color(*bits); 3804 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3805 } 3806 3807 pwq->nr_in_flight[work_color]++; 3808 work_flags |= work_color_to_flags(work_color); 3809 3810 insert_work(pwq, &barr->work, head, work_flags); 3811 } 3812 3813 /** 3814 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 3815 * @wq: workqueue being flushed 3816 * @flush_color: new flush color, < 0 for no-op 3817 * @work_color: new work color, < 0 for no-op 3818 * 3819 * Prepare pwqs for workqueue flushing. 3820 * 3821 * If @flush_color is non-negative, flush_color on all pwqs should be 3822 * -1. If no pwq has in-flight commands at the specified color, all 3823 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3824 * has in flight commands, its pwq->flush_color is set to 3825 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 3826 * wakeup logic is armed and %true is returned. 3827 * 3828 * The caller should have initialized @wq->first_flusher prior to 3829 * calling this function with non-negative @flush_color. If 3830 * @flush_color is negative, no flush color update is done and %false 3831 * is returned. 3832 * 3833 * If @work_color is non-negative, all pwqs should have the same 3834 * work_color which is previous to @work_color and all will be 3835 * advanced to @work_color. 3836 * 3837 * CONTEXT: 3838 * mutex_lock(wq->mutex). 3839 * 3840 * Return: 3841 * %true if @flush_color >= 0 and there's something to flush. %false 3842 * otherwise. 3843 */ 3844 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 3845 int flush_color, int work_color) 3846 { 3847 bool wait = false; 3848 struct pool_workqueue *pwq; 3849 struct worker_pool *current_pool = NULL; 3850 3851 if (flush_color >= 0) { 3852 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3853 atomic_set(&wq->nr_pwqs_to_flush, 1); 3854 } 3855 3856 /* 3857 * For unbound workqueue, pwqs will map to only a few pools. 3858 * Most of the time, pwqs within the same pool will be linked 3859 * sequentially to wq->pwqs by cpu index. So in the majority 3860 * of pwq iters, the pool is the same, only doing lock/unlock 3861 * if the pool has changed. This can largely reduce expensive 3862 * lock operations. 3863 */ 3864 for_each_pwq(pwq, wq) { 3865 if (current_pool != pwq->pool) { 3866 if (likely(current_pool)) 3867 raw_spin_unlock_irq(¤t_pool->lock); 3868 current_pool = pwq->pool; 3869 raw_spin_lock_irq(¤t_pool->lock); 3870 } 3871 3872 if (flush_color >= 0) { 3873 WARN_ON_ONCE(pwq->flush_color != -1); 3874 3875 if (pwq->nr_in_flight[flush_color]) { 3876 pwq->flush_color = flush_color; 3877 atomic_inc(&wq->nr_pwqs_to_flush); 3878 wait = true; 3879 } 3880 } 3881 3882 if (work_color >= 0) { 3883 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3884 pwq->work_color = work_color; 3885 } 3886 3887 } 3888 3889 if (current_pool) 3890 raw_spin_unlock_irq(¤t_pool->lock); 3891 3892 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 3893 complete(&wq->first_flusher->done); 3894 3895 return wait; 3896 } 3897 3898 static void touch_wq_lockdep_map(struct workqueue_struct *wq) 3899 { 3900 #ifdef CONFIG_LOCKDEP 3901 if (unlikely(!wq->lockdep_map)) 3902 return; 3903 3904 if (wq->flags & WQ_BH) 3905 local_bh_disable(); 3906 3907 lock_map_acquire(wq->lockdep_map); 3908 lock_map_release(wq->lockdep_map); 3909 3910 if (wq->flags & WQ_BH) 3911 local_bh_enable(); 3912 #endif 3913 } 3914 3915 static void touch_work_lockdep_map(struct work_struct *work, 3916 struct workqueue_struct *wq) 3917 { 3918 #ifdef CONFIG_LOCKDEP 3919 if (wq->flags & WQ_BH) 3920 local_bh_disable(); 3921 3922 lock_map_acquire(&work->lockdep_map); 3923 lock_map_release(&work->lockdep_map); 3924 3925 if (wq->flags & WQ_BH) 3926 local_bh_enable(); 3927 #endif 3928 } 3929 3930 /** 3931 * __flush_workqueue - ensure that any scheduled work has run to completion. 3932 * @wq: workqueue to flush 3933 * 3934 * This function sleeps until all work items which were queued on entry 3935 * have finished execution, but it is not livelocked by new incoming ones. 3936 */ 3937 void __flush_workqueue(struct workqueue_struct *wq) 3938 { 3939 struct wq_flusher this_flusher = { 3940 .list = LIST_HEAD_INIT(this_flusher.list), 3941 .flush_color = -1, 3942 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)), 3943 }; 3944 int next_color; 3945 3946 if (WARN_ON(!wq_online)) 3947 return; 3948 3949 touch_wq_lockdep_map(wq); 3950 3951 mutex_lock(&wq->mutex); 3952 3953 /* 3954 * Start-to-wait phase 3955 */ 3956 next_color = work_next_color(wq->work_color); 3957 3958 if (next_color != wq->flush_color) { 3959 /* 3960 * Color space is not full. The current work_color 3961 * becomes our flush_color and work_color is advanced 3962 * by one. 3963 */ 3964 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 3965 this_flusher.flush_color = wq->work_color; 3966 wq->work_color = next_color; 3967 3968 if (!wq->first_flusher) { 3969 /* no flush in progress, become the first flusher */ 3970 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 3971 3972 wq->first_flusher = &this_flusher; 3973 3974 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 3975 wq->work_color)) { 3976 /* nothing to flush, done */ 3977 wq->flush_color = next_color; 3978 wq->first_flusher = NULL; 3979 goto out_unlock; 3980 } 3981 } else { 3982 /* wait in queue */ 3983 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 3984 list_add_tail(&this_flusher.list, &wq->flusher_queue); 3985 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 3986 } 3987 } else { 3988 /* 3989 * Oops, color space is full, wait on overflow queue. 3990 * The next flush completion will assign us 3991 * flush_color and transfer to flusher_queue. 3992 */ 3993 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 3994 } 3995 3996 check_flush_dependency(wq, NULL, false); 3997 3998 mutex_unlock(&wq->mutex); 3999 4000 wait_for_completion(&this_flusher.done); 4001 4002 /* 4003 * Wake-up-and-cascade phase 4004 * 4005 * First flushers are responsible for cascading flushes and 4006 * handling overflow. Non-first flushers can simply return. 4007 */ 4008 if (READ_ONCE(wq->first_flusher) != &this_flusher) 4009 return; 4010 4011 mutex_lock(&wq->mutex); 4012 4013 /* we might have raced, check again with mutex held */ 4014 if (wq->first_flusher != &this_flusher) 4015 goto out_unlock; 4016 4017 WRITE_ONCE(wq->first_flusher, NULL); 4018 4019 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 4020 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 4021 4022 while (true) { 4023 struct wq_flusher *next, *tmp; 4024 4025 /* complete all the flushers sharing the current flush color */ 4026 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 4027 if (next->flush_color != wq->flush_color) 4028 break; 4029 list_del_init(&next->list); 4030 complete(&next->done); 4031 } 4032 4033 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 4034 wq->flush_color != work_next_color(wq->work_color)); 4035 4036 /* this flush_color is finished, advance by one */ 4037 wq->flush_color = work_next_color(wq->flush_color); 4038 4039 /* one color has been freed, handle overflow queue */ 4040 if (!list_empty(&wq->flusher_overflow)) { 4041 /* 4042 * Assign the same color to all overflowed 4043 * flushers, advance work_color and append to 4044 * flusher_queue. This is the start-to-wait 4045 * phase for these overflowed flushers. 4046 */ 4047 list_for_each_entry(tmp, &wq->flusher_overflow, list) 4048 tmp->flush_color = wq->work_color; 4049 4050 wq->work_color = work_next_color(wq->work_color); 4051 4052 list_splice_tail_init(&wq->flusher_overflow, 4053 &wq->flusher_queue); 4054 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 4055 } 4056 4057 if (list_empty(&wq->flusher_queue)) { 4058 WARN_ON_ONCE(wq->flush_color != wq->work_color); 4059 break; 4060 } 4061 4062 /* 4063 * Need to flush more colors. Make the next flusher 4064 * the new first flusher and arm pwqs. 4065 */ 4066 WARN_ON_ONCE(wq->flush_color == wq->work_color); 4067 WARN_ON_ONCE(wq->flush_color != next->flush_color); 4068 4069 list_del_init(&next->list); 4070 wq->first_flusher = next; 4071 4072 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 4073 break; 4074 4075 /* 4076 * Meh... this color is already done, clear first 4077 * flusher and repeat cascading. 4078 */ 4079 wq->first_flusher = NULL; 4080 } 4081 4082 out_unlock: 4083 mutex_unlock(&wq->mutex); 4084 } 4085 EXPORT_SYMBOL(__flush_workqueue); 4086 4087 /** 4088 * drain_workqueue - drain a workqueue 4089 * @wq: workqueue to drain 4090 * 4091 * Wait until the workqueue becomes empty. While draining is in progress, 4092 * only chain queueing is allowed. IOW, only currently pending or running 4093 * work items on @wq can queue further work items on it. @wq is flushed 4094 * repeatedly until it becomes empty. The number of flushing is determined 4095 * by the depth of chaining and should be relatively short. Whine if it 4096 * takes too long. 4097 */ 4098 void drain_workqueue(struct workqueue_struct *wq) 4099 { 4100 unsigned int flush_cnt = 0; 4101 struct pool_workqueue *pwq; 4102 4103 /* 4104 * __queue_work() needs to test whether there are drainers, is much 4105 * hotter than drain_workqueue() and already looks at @wq->flags. 4106 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 4107 */ 4108 mutex_lock(&wq->mutex); 4109 if (!wq->nr_drainers++) 4110 wq->flags |= __WQ_DRAINING; 4111 mutex_unlock(&wq->mutex); 4112 reflush: 4113 __flush_workqueue(wq); 4114 4115 mutex_lock(&wq->mutex); 4116 4117 for_each_pwq(pwq, wq) { 4118 bool drained; 4119 4120 raw_spin_lock_irq(&pwq->pool->lock); 4121 drained = pwq_is_empty(pwq); 4122 raw_spin_unlock_irq(&pwq->pool->lock); 4123 4124 if (drained) 4125 continue; 4126 4127 if (++flush_cnt == 10 || 4128 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 4129 pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 4130 wq->name, __func__, flush_cnt); 4131 4132 mutex_unlock(&wq->mutex); 4133 goto reflush; 4134 } 4135 4136 if (!--wq->nr_drainers) 4137 wq->flags &= ~__WQ_DRAINING; 4138 mutex_unlock(&wq->mutex); 4139 } 4140 EXPORT_SYMBOL_GPL(drain_workqueue); 4141 4142 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 4143 bool from_cancel) 4144 { 4145 struct worker *worker = NULL; 4146 struct worker_pool *pool; 4147 struct pool_workqueue *pwq; 4148 struct workqueue_struct *wq; 4149 4150 rcu_read_lock(); 4151 pool = get_work_pool(work); 4152 if (!pool) { 4153 rcu_read_unlock(); 4154 return false; 4155 } 4156 4157 raw_spin_lock_irq(&pool->lock); 4158 /* see the comment in try_to_grab_pending() with the same code */ 4159 pwq = get_work_pwq(work); 4160 if (pwq) { 4161 if (unlikely(pwq->pool != pool)) 4162 goto already_gone; 4163 } else { 4164 worker = find_worker_executing_work(pool, work); 4165 if (!worker) 4166 goto already_gone; 4167 pwq = worker->current_pwq; 4168 } 4169 4170 wq = pwq->wq; 4171 check_flush_dependency(wq, work, from_cancel); 4172 4173 insert_wq_barrier(pwq, barr, work, worker); 4174 raw_spin_unlock_irq(&pool->lock); 4175 4176 touch_work_lockdep_map(work, wq); 4177 4178 /* 4179 * Force a lock recursion deadlock when using flush_work() inside a 4180 * single-threaded or rescuer equipped workqueue. 4181 * 4182 * For single threaded workqueues the deadlock happens when the work 4183 * is after the work issuing the flush_work(). For rescuer equipped 4184 * workqueues the deadlock happens when the rescuer stalls, blocking 4185 * forward progress. 4186 */ 4187 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer)) 4188 touch_wq_lockdep_map(wq); 4189 4190 rcu_read_unlock(); 4191 return true; 4192 already_gone: 4193 raw_spin_unlock_irq(&pool->lock); 4194 rcu_read_unlock(); 4195 return false; 4196 } 4197 4198 static bool __flush_work(struct work_struct *work, bool from_cancel) 4199 { 4200 struct wq_barrier barr; 4201 4202 if (WARN_ON(!wq_online)) 4203 return false; 4204 4205 if (WARN_ON(!work->func)) 4206 return false; 4207 4208 if (!start_flush_work(work, &barr, from_cancel)) 4209 return false; 4210 4211 /* 4212 * start_flush_work() returned %true. If @from_cancel is set, we know 4213 * that @work must have been executing during start_flush_work() and 4214 * can't currently be queued. Its data must contain OFFQ bits. If @work 4215 * was queued on a BH workqueue, we also know that it was running in the 4216 * BH context and thus can be busy-waited. 4217 */ 4218 if (from_cancel) { 4219 unsigned long data = *work_data_bits(work); 4220 4221 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && 4222 (data & WORK_OFFQ_BH)) { 4223 /* 4224 * On RT, prevent a live lock when %current preempted 4225 * soft interrupt processing or prevents ksoftirqd from 4226 * running by keeping flipping BH. If the BH work item 4227 * runs on a different CPU then this has no effect other 4228 * than doing the BH disable/enable dance for nothing. 4229 * This is copied from 4230 * kernel/softirq.c::tasklet_unlock_spin_wait(). 4231 */ 4232 while (!try_wait_for_completion(&barr.done)) { 4233 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 4234 local_bh_disable(); 4235 local_bh_enable(); 4236 } else { 4237 cpu_relax(); 4238 } 4239 } 4240 goto out_destroy; 4241 } 4242 } 4243 4244 wait_for_completion(&barr.done); 4245 4246 out_destroy: 4247 destroy_work_on_stack(&barr.work); 4248 return true; 4249 } 4250 4251 /** 4252 * flush_work - wait for a work to finish executing the last queueing instance 4253 * @work: the work to flush 4254 * 4255 * Wait until @work has finished execution. @work is guaranteed to be idle 4256 * on return if it hasn't been requeued since flush started. 4257 * 4258 * Return: 4259 * %true if flush_work() waited for the work to finish execution, 4260 * %false if it was already idle. 4261 */ 4262 bool flush_work(struct work_struct *work) 4263 { 4264 might_sleep(); 4265 return __flush_work(work, false); 4266 } 4267 EXPORT_SYMBOL_GPL(flush_work); 4268 4269 /** 4270 * flush_delayed_work - wait for a dwork to finish executing the last queueing 4271 * @dwork: the delayed work to flush 4272 * 4273 * Delayed timer is cancelled and the pending work is queued for 4274 * immediate execution. Like flush_work(), this function only 4275 * considers the last queueing instance of @dwork. 4276 * 4277 * Return: 4278 * %true if flush_work() waited for the work to finish execution, 4279 * %false if it was already idle. 4280 */ 4281 bool flush_delayed_work(struct delayed_work *dwork) 4282 { 4283 local_irq_disable(); 4284 if (timer_delete_sync(&dwork->timer)) 4285 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 4286 local_irq_enable(); 4287 return flush_work(&dwork->work); 4288 } 4289 EXPORT_SYMBOL(flush_delayed_work); 4290 4291 /** 4292 * flush_rcu_work - wait for a rwork to finish executing the last queueing 4293 * @rwork: the rcu work to flush 4294 * 4295 * Return: 4296 * %true if flush_rcu_work() waited for the work to finish execution, 4297 * %false if it was already idle. 4298 */ 4299 bool flush_rcu_work(struct rcu_work *rwork) 4300 { 4301 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 4302 rcu_barrier(); 4303 flush_work(&rwork->work); 4304 return true; 4305 } else { 4306 return flush_work(&rwork->work); 4307 } 4308 } 4309 EXPORT_SYMBOL(flush_rcu_work); 4310 4311 static void work_offqd_disable(struct work_offq_data *offqd) 4312 { 4313 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1; 4314 4315 if (likely(offqd->disable < max)) 4316 offqd->disable++; 4317 else 4318 WARN_ONCE(true, "workqueue: work disable count overflowed\n"); 4319 } 4320 4321 static void work_offqd_enable(struct work_offq_data *offqd) 4322 { 4323 if (likely(offqd->disable > 0)) 4324 offqd->disable--; 4325 else 4326 WARN_ONCE(true, "workqueue: work disable count underflowed\n"); 4327 } 4328 4329 static bool __cancel_work(struct work_struct *work, u32 cflags) 4330 { 4331 struct work_offq_data offqd; 4332 unsigned long irq_flags; 4333 int ret; 4334 4335 ret = work_grab_pending(work, cflags, &irq_flags); 4336 4337 work_offqd_unpack(&offqd, *work_data_bits(work)); 4338 4339 if (cflags & WORK_CANCEL_DISABLE) 4340 work_offqd_disable(&offqd); 4341 4342 set_work_pool_and_clear_pending(work, offqd.pool_id, 4343 work_offqd_pack_flags(&offqd)); 4344 local_irq_restore(irq_flags); 4345 return ret; 4346 } 4347 4348 static bool __cancel_work_sync(struct work_struct *work, u32 cflags) 4349 { 4350 bool ret; 4351 4352 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE); 4353 4354 if (*work_data_bits(work) & WORK_OFFQ_BH) 4355 WARN_ON_ONCE(in_hardirq()); 4356 else 4357 might_sleep(); 4358 4359 /* 4360 * Skip __flush_work() during early boot when we know that @work isn't 4361 * executing. This allows canceling during early boot. 4362 */ 4363 if (wq_online) 4364 __flush_work(work, true); 4365 4366 if (!(cflags & WORK_CANCEL_DISABLE)) 4367 enable_work(work); 4368 4369 return ret; 4370 } 4371 4372 /* 4373 * See cancel_delayed_work() 4374 */ 4375 bool cancel_work(struct work_struct *work) 4376 { 4377 return __cancel_work(work, 0); 4378 } 4379 EXPORT_SYMBOL(cancel_work); 4380 4381 /** 4382 * cancel_work_sync - cancel a work and wait for it to finish 4383 * @work: the work to cancel 4384 * 4385 * Cancel @work and wait for its execution to finish. This function can be used 4386 * even if the work re-queues itself or migrates to another workqueue. On return 4387 * from this function, @work is guaranteed to be not pending or executing on any 4388 * CPU as long as there aren't racing enqueues. 4389 * 4390 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's. 4391 * Use cancel_delayed_work_sync() instead. 4392 * 4393 * Must be called from a sleepable context if @work was last queued on a non-BH 4394 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4395 * if @work was last queued on a BH workqueue. 4396 * 4397 * Returns %true if @work was pending, %false otherwise. 4398 */ 4399 bool cancel_work_sync(struct work_struct *work) 4400 { 4401 return __cancel_work_sync(work, 0); 4402 } 4403 EXPORT_SYMBOL_GPL(cancel_work_sync); 4404 4405 /** 4406 * cancel_delayed_work - cancel a delayed work 4407 * @dwork: delayed_work to cancel 4408 * 4409 * Kill off a pending delayed_work. 4410 * 4411 * Return: %true if @dwork was pending and canceled; %false if it wasn't 4412 * pending. 4413 * 4414 * Note: 4415 * The work callback function may still be running on return, unless 4416 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 4417 * use cancel_delayed_work_sync() to wait on it. 4418 * 4419 * This function is safe to call from any context including IRQ handler. 4420 */ 4421 bool cancel_delayed_work(struct delayed_work *dwork) 4422 { 4423 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED); 4424 } 4425 EXPORT_SYMBOL(cancel_delayed_work); 4426 4427 /** 4428 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 4429 * @dwork: the delayed work cancel 4430 * 4431 * This is cancel_work_sync() for delayed works. 4432 * 4433 * Return: 4434 * %true if @dwork was pending, %false otherwise. 4435 */ 4436 bool cancel_delayed_work_sync(struct delayed_work *dwork) 4437 { 4438 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED); 4439 } 4440 EXPORT_SYMBOL(cancel_delayed_work_sync); 4441 4442 /** 4443 * disable_work - Disable and cancel a work item 4444 * @work: work item to disable 4445 * 4446 * Disable @work by incrementing its disable count and cancel it if currently 4447 * pending. As long as the disable count is non-zero, any attempt to queue @work 4448 * will fail and return %false. The maximum supported disable depth is 2 to the 4449 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536. 4450 * 4451 * Can be called from any context. Returns %true if @work was pending, %false 4452 * otherwise. 4453 */ 4454 bool disable_work(struct work_struct *work) 4455 { 4456 return __cancel_work(work, WORK_CANCEL_DISABLE); 4457 } 4458 EXPORT_SYMBOL_GPL(disable_work); 4459 4460 /** 4461 * disable_work_sync - Disable, cancel and drain a work item 4462 * @work: work item to disable 4463 * 4464 * Similar to disable_work() but also wait for @work to finish if currently 4465 * executing. 4466 * 4467 * Must be called from a sleepable context if @work was last queued on a non-BH 4468 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4469 * if @work was last queued on a BH workqueue. 4470 * 4471 * Returns %true if @work was pending, %false otherwise. 4472 */ 4473 bool disable_work_sync(struct work_struct *work) 4474 { 4475 return __cancel_work_sync(work, WORK_CANCEL_DISABLE); 4476 } 4477 EXPORT_SYMBOL_GPL(disable_work_sync); 4478 4479 /** 4480 * enable_work - Enable a work item 4481 * @work: work item to enable 4482 * 4483 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can 4484 * only be queued if its disable count is 0. 4485 * 4486 * Can be called from any context. Returns %true if the disable count reached 0. 4487 * Otherwise, %false. 4488 */ 4489 bool enable_work(struct work_struct *work) 4490 { 4491 struct work_offq_data offqd; 4492 unsigned long irq_flags; 4493 4494 work_grab_pending(work, 0, &irq_flags); 4495 4496 work_offqd_unpack(&offqd, *work_data_bits(work)); 4497 work_offqd_enable(&offqd); 4498 set_work_pool_and_clear_pending(work, offqd.pool_id, 4499 work_offqd_pack_flags(&offqd)); 4500 local_irq_restore(irq_flags); 4501 4502 return !offqd.disable; 4503 } 4504 EXPORT_SYMBOL_GPL(enable_work); 4505 4506 /** 4507 * disable_delayed_work - Disable and cancel a delayed work item 4508 * @dwork: delayed work item to disable 4509 * 4510 * disable_work() for delayed work items. 4511 */ 4512 bool disable_delayed_work(struct delayed_work *dwork) 4513 { 4514 return __cancel_work(&dwork->work, 4515 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4516 } 4517 EXPORT_SYMBOL_GPL(disable_delayed_work); 4518 4519 /** 4520 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item 4521 * @dwork: delayed work item to disable 4522 * 4523 * disable_work_sync() for delayed work items. 4524 */ 4525 bool disable_delayed_work_sync(struct delayed_work *dwork) 4526 { 4527 return __cancel_work_sync(&dwork->work, 4528 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4529 } 4530 EXPORT_SYMBOL_GPL(disable_delayed_work_sync); 4531 4532 /** 4533 * enable_delayed_work - Enable a delayed work item 4534 * @dwork: delayed work item to enable 4535 * 4536 * enable_work() for delayed work items. 4537 */ 4538 bool enable_delayed_work(struct delayed_work *dwork) 4539 { 4540 return enable_work(&dwork->work); 4541 } 4542 EXPORT_SYMBOL_GPL(enable_delayed_work); 4543 4544 /** 4545 * schedule_on_each_cpu - execute a function synchronously on each online CPU 4546 * @func: the function to call 4547 * 4548 * schedule_on_each_cpu() executes @func on each online CPU using the 4549 * system workqueue and blocks until all CPUs have completed. 4550 * schedule_on_each_cpu() is very slow. 4551 * 4552 * Return: 4553 * 0 on success, -errno on failure. 4554 */ 4555 int schedule_on_each_cpu(work_func_t func) 4556 { 4557 int cpu; 4558 struct work_struct __percpu *works; 4559 4560 works = alloc_percpu(struct work_struct); 4561 if (!works) 4562 return -ENOMEM; 4563 4564 cpus_read_lock(); 4565 4566 for_each_online_cpu(cpu) { 4567 struct work_struct *work = per_cpu_ptr(works, cpu); 4568 4569 INIT_WORK(work, func); 4570 schedule_work_on(cpu, work); 4571 } 4572 4573 for_each_online_cpu(cpu) 4574 flush_work(per_cpu_ptr(works, cpu)); 4575 4576 cpus_read_unlock(); 4577 free_percpu(works); 4578 return 0; 4579 } 4580 4581 /** 4582 * execute_in_process_context - reliably execute the routine with user context 4583 * @fn: the function to execute 4584 * @ew: guaranteed storage for the execute work structure (must 4585 * be available when the work executes) 4586 * 4587 * Executes the function immediately if process context is available, 4588 * otherwise schedules the function for delayed execution. 4589 * 4590 * Return: 0 - function was executed 4591 * 1 - function was scheduled for execution 4592 */ 4593 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 4594 { 4595 if (!in_interrupt()) { 4596 fn(&ew->work); 4597 return 0; 4598 } 4599 4600 INIT_WORK(&ew->work, fn); 4601 schedule_work(&ew->work); 4602 4603 return 1; 4604 } 4605 EXPORT_SYMBOL_GPL(execute_in_process_context); 4606 4607 /** 4608 * free_workqueue_attrs - free a workqueue_attrs 4609 * @attrs: workqueue_attrs to free 4610 * 4611 * Undo alloc_workqueue_attrs(). 4612 */ 4613 void free_workqueue_attrs(struct workqueue_attrs *attrs) 4614 { 4615 if (attrs) { 4616 free_cpumask_var(attrs->cpumask); 4617 free_cpumask_var(attrs->__pod_cpumask); 4618 kfree(attrs); 4619 } 4620 } 4621 4622 /** 4623 * alloc_workqueue_attrs - allocate a workqueue_attrs 4624 * 4625 * Allocate a new workqueue_attrs, initialize with default settings and 4626 * return it. 4627 * 4628 * Return: The allocated new workqueue_attr on success. %NULL on failure. 4629 */ 4630 struct workqueue_attrs *alloc_workqueue_attrs_noprof(void) 4631 { 4632 struct workqueue_attrs *attrs; 4633 4634 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 4635 if (!attrs) 4636 goto fail; 4637 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 4638 goto fail; 4639 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 4640 goto fail; 4641 4642 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4643 attrs->affn_scope = WQ_AFFN_DFL; 4644 return attrs; 4645 fail: 4646 free_workqueue_attrs(attrs); 4647 return NULL; 4648 } 4649 4650 static void copy_workqueue_attrs(struct workqueue_attrs *to, 4651 const struct workqueue_attrs *from) 4652 { 4653 to->nice = from->nice; 4654 cpumask_copy(to->cpumask, from->cpumask); 4655 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 4656 to->affn_strict = from->affn_strict; 4657 4658 /* 4659 * Unlike hash and equality test, copying shouldn't ignore wq-only 4660 * fields as copying is used for both pool and wq attrs. Instead, 4661 * get_unbound_pool() explicitly clears the fields. 4662 */ 4663 to->affn_scope = from->affn_scope; 4664 to->ordered = from->ordered; 4665 } 4666 4667 /* 4668 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 4669 * comments in 'struct workqueue_attrs' definition. 4670 */ 4671 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 4672 { 4673 attrs->affn_scope = WQ_AFFN_NR_TYPES; 4674 attrs->ordered = false; 4675 if (attrs->affn_strict) 4676 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4677 } 4678 4679 /* hash value of the content of @attr */ 4680 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 4681 { 4682 u32 hash = 0; 4683 4684 hash = jhash_1word(attrs->nice, hash); 4685 hash = jhash_1word(attrs->affn_strict, hash); 4686 hash = jhash(cpumask_bits(attrs->__pod_cpumask), 4687 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4688 if (!attrs->affn_strict) 4689 hash = jhash(cpumask_bits(attrs->cpumask), 4690 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4691 return hash; 4692 } 4693 4694 /* content equality test */ 4695 static bool wqattrs_equal(const struct workqueue_attrs *a, 4696 const struct workqueue_attrs *b) 4697 { 4698 if (a->nice != b->nice) 4699 return false; 4700 if (a->affn_strict != b->affn_strict) 4701 return false; 4702 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 4703 return false; 4704 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask)) 4705 return false; 4706 return true; 4707 } 4708 4709 /* Update @attrs with actually available CPUs */ 4710 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 4711 const cpumask_t *unbound_cpumask) 4712 { 4713 /* 4714 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 4715 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 4716 * @unbound_cpumask. 4717 */ 4718 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 4719 if (unlikely(cpumask_empty(attrs->cpumask))) 4720 cpumask_copy(attrs->cpumask, unbound_cpumask); 4721 } 4722 4723 /* find wq_pod_type to use for @attrs */ 4724 static const struct wq_pod_type * 4725 wqattrs_pod_type(const struct workqueue_attrs *attrs) 4726 { 4727 enum wq_affn_scope scope; 4728 struct wq_pod_type *pt; 4729 4730 /* to synchronize access to wq_affn_dfl */ 4731 lockdep_assert_held(&wq_pool_mutex); 4732 4733 if (attrs->affn_scope == WQ_AFFN_DFL) 4734 scope = wq_affn_dfl; 4735 else 4736 scope = attrs->affn_scope; 4737 4738 pt = &wq_pod_types[scope]; 4739 4740 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 4741 likely(pt->nr_pods)) 4742 return pt; 4743 4744 /* 4745 * Before workqueue_init_topology(), only SYSTEM is available which is 4746 * initialized in workqueue_init_early(). 4747 */ 4748 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 4749 BUG_ON(!pt->nr_pods); 4750 return pt; 4751 } 4752 4753 /** 4754 * init_worker_pool - initialize a newly zalloc'd worker_pool 4755 * @pool: worker_pool to initialize 4756 * 4757 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 4758 * 4759 * Return: 0 on success, -errno on failure. Even on failure, all fields 4760 * inside @pool proper are initialized and put_unbound_pool() can be called 4761 * on @pool safely to release it. 4762 */ 4763 static int init_worker_pool(struct worker_pool *pool) 4764 { 4765 raw_spin_lock_init(&pool->lock); 4766 pool->id = -1; 4767 pool->cpu = -1; 4768 pool->node = NUMA_NO_NODE; 4769 pool->flags |= POOL_DISASSOCIATED; 4770 pool->watchdog_ts = jiffies; 4771 INIT_LIST_HEAD(&pool->worklist); 4772 INIT_LIST_HEAD(&pool->idle_list); 4773 hash_init(pool->busy_hash); 4774 4775 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 4776 INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 4777 4778 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 4779 4780 INIT_LIST_HEAD(&pool->workers); 4781 4782 ida_init(&pool->worker_ida); 4783 INIT_HLIST_NODE(&pool->hash_node); 4784 pool->refcnt = 1; 4785 4786 /* shouldn't fail above this point */ 4787 pool->attrs = alloc_workqueue_attrs(); 4788 if (!pool->attrs) 4789 return -ENOMEM; 4790 4791 wqattrs_clear_for_pool(pool->attrs); 4792 4793 return 0; 4794 } 4795 4796 #ifdef CONFIG_LOCKDEP 4797 static void wq_init_lockdep(struct workqueue_struct *wq) 4798 { 4799 char *lock_name; 4800 4801 lockdep_register_key(&wq->key); 4802 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 4803 if (!lock_name) 4804 lock_name = wq->name; 4805 4806 wq->lock_name = lock_name; 4807 wq->lockdep_map = &wq->__lockdep_map; 4808 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0); 4809 } 4810 4811 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4812 { 4813 if (wq->lockdep_map != &wq->__lockdep_map) 4814 return; 4815 4816 lockdep_unregister_key(&wq->key); 4817 } 4818 4819 static void wq_free_lockdep(struct workqueue_struct *wq) 4820 { 4821 if (wq->lockdep_map != &wq->__lockdep_map) 4822 return; 4823 4824 if (wq->lock_name != wq->name) 4825 kfree(wq->lock_name); 4826 } 4827 #else 4828 static void wq_init_lockdep(struct workqueue_struct *wq) 4829 { 4830 } 4831 4832 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4833 { 4834 } 4835 4836 static void wq_free_lockdep(struct workqueue_struct *wq) 4837 { 4838 } 4839 #endif 4840 4841 static void free_node_nr_active(struct wq_node_nr_active **nna_ar) 4842 { 4843 int node; 4844 4845 for_each_node(node) { 4846 kfree(nna_ar[node]); 4847 nna_ar[node] = NULL; 4848 } 4849 4850 kfree(nna_ar[nr_node_ids]); 4851 nna_ar[nr_node_ids] = NULL; 4852 } 4853 4854 static void init_node_nr_active(struct wq_node_nr_active *nna) 4855 { 4856 nna->max = WQ_DFL_MIN_ACTIVE; 4857 atomic_set(&nna->nr, 0); 4858 raw_spin_lock_init(&nna->lock); 4859 INIT_LIST_HEAD(&nna->pending_pwqs); 4860 } 4861 4862 /* 4863 * Each node's nr_active counter will be accessed mostly from its own node and 4864 * should be allocated in the node. 4865 */ 4866 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) 4867 { 4868 struct wq_node_nr_active *nna; 4869 int node; 4870 4871 for_each_node(node) { 4872 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); 4873 if (!nna) 4874 goto err_free; 4875 init_node_nr_active(nna); 4876 nna_ar[node] = nna; 4877 } 4878 4879 /* [nr_node_ids] is used as the fallback */ 4880 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); 4881 if (!nna) 4882 goto err_free; 4883 init_node_nr_active(nna); 4884 nna_ar[nr_node_ids] = nna; 4885 4886 return 0; 4887 4888 err_free: 4889 free_node_nr_active(nna_ar); 4890 return -ENOMEM; 4891 } 4892 4893 static void rcu_free_wq(struct rcu_head *rcu) 4894 { 4895 struct workqueue_struct *wq = 4896 container_of(rcu, struct workqueue_struct, rcu); 4897 4898 if (wq->flags & WQ_UNBOUND) 4899 free_node_nr_active(wq->node_nr_active); 4900 4901 wq_free_lockdep(wq); 4902 free_percpu(wq->cpu_pwq); 4903 free_workqueue_attrs(wq->unbound_attrs); 4904 kfree(wq); 4905 } 4906 4907 static void rcu_free_pool(struct rcu_head *rcu) 4908 { 4909 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 4910 4911 ida_destroy(&pool->worker_ida); 4912 free_workqueue_attrs(pool->attrs); 4913 kfree(pool); 4914 } 4915 4916 /** 4917 * put_unbound_pool - put a worker_pool 4918 * @pool: worker_pool to put 4919 * 4920 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 4921 * safe manner. get_unbound_pool() calls this function on its failure path 4922 * and this function should be able to release pools which went through, 4923 * successfully or not, init_worker_pool(). 4924 * 4925 * Should be called with wq_pool_mutex held. 4926 */ 4927 static void put_unbound_pool(struct worker_pool *pool) 4928 { 4929 struct worker *worker; 4930 LIST_HEAD(cull_list); 4931 4932 lockdep_assert_held(&wq_pool_mutex); 4933 4934 if (--pool->refcnt) 4935 return; 4936 4937 /* sanity checks */ 4938 if (WARN_ON(!(pool->cpu < 0)) || 4939 WARN_ON(!list_empty(&pool->worklist))) 4940 return; 4941 4942 /* release id and unhash */ 4943 if (pool->id >= 0) 4944 idr_remove(&worker_pool_idr, pool->id); 4945 hash_del(&pool->hash_node); 4946 4947 /* 4948 * Become the manager and destroy all workers. This prevents 4949 * @pool's workers from blocking on attach_mutex. We're the last 4950 * manager and @pool gets freed with the flag set. 4951 * 4952 * Having a concurrent manager is quite unlikely to happen as we can 4953 * only get here with 4954 * pwq->refcnt == pool->refcnt == 0 4955 * which implies no work queued to the pool, which implies no worker can 4956 * become the manager. However a worker could have taken the role of 4957 * manager before the refcnts dropped to 0, since maybe_create_worker() 4958 * drops pool->lock 4959 */ 4960 while (true) { 4961 rcuwait_wait_event(&manager_wait, 4962 !(pool->flags & POOL_MANAGER_ACTIVE), 4963 TASK_UNINTERRUPTIBLE); 4964 4965 mutex_lock(&wq_pool_attach_mutex); 4966 raw_spin_lock_irq(&pool->lock); 4967 if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 4968 pool->flags |= POOL_MANAGER_ACTIVE; 4969 break; 4970 } 4971 raw_spin_unlock_irq(&pool->lock); 4972 mutex_unlock(&wq_pool_attach_mutex); 4973 } 4974 4975 while ((worker = first_idle_worker(pool))) 4976 set_worker_dying(worker, &cull_list); 4977 WARN_ON(pool->nr_workers || pool->nr_idle); 4978 raw_spin_unlock_irq(&pool->lock); 4979 4980 detach_dying_workers(&cull_list); 4981 4982 mutex_unlock(&wq_pool_attach_mutex); 4983 4984 reap_dying_workers(&cull_list); 4985 4986 /* shut down the timers */ 4987 timer_delete_sync(&pool->idle_timer); 4988 cancel_work_sync(&pool->idle_cull_work); 4989 timer_delete_sync(&pool->mayday_timer); 4990 4991 /* RCU protected to allow dereferences from get_work_pool() */ 4992 call_rcu(&pool->rcu, rcu_free_pool); 4993 } 4994 4995 /** 4996 * get_unbound_pool - get a worker_pool with the specified attributes 4997 * @attrs: the attributes of the worker_pool to get 4998 * 4999 * Obtain a worker_pool which has the same attributes as @attrs, bump the 5000 * reference count and return it. If there already is a matching 5001 * worker_pool, it will be used; otherwise, this function attempts to 5002 * create a new one. 5003 * 5004 * Should be called with wq_pool_mutex held. 5005 * 5006 * Return: On success, a worker_pool with the same attributes as @attrs. 5007 * On failure, %NULL. 5008 */ 5009 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 5010 { 5011 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 5012 u32 hash = wqattrs_hash(attrs); 5013 struct worker_pool *pool; 5014 int pod, node = NUMA_NO_NODE; 5015 5016 lockdep_assert_held(&wq_pool_mutex); 5017 5018 /* do we already have a matching pool? */ 5019 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 5020 if (wqattrs_equal(pool->attrs, attrs)) { 5021 pool->refcnt++; 5022 return pool; 5023 } 5024 } 5025 5026 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 5027 for (pod = 0; pod < pt->nr_pods; pod++) { 5028 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 5029 node = pt->pod_node[pod]; 5030 break; 5031 } 5032 } 5033 5034 /* nope, create a new one */ 5035 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 5036 if (!pool || init_worker_pool(pool) < 0) 5037 goto fail; 5038 5039 pool->node = node; 5040 copy_workqueue_attrs(pool->attrs, attrs); 5041 wqattrs_clear_for_pool(pool->attrs); 5042 5043 if (worker_pool_assign_id(pool) < 0) 5044 goto fail; 5045 5046 /* create and start the initial worker */ 5047 if (wq_online && !create_worker(pool)) 5048 goto fail; 5049 5050 /* install */ 5051 hash_add(unbound_pool_hash, &pool->hash_node, hash); 5052 5053 return pool; 5054 fail: 5055 if (pool) 5056 put_unbound_pool(pool); 5057 return NULL; 5058 } 5059 5060 /* 5061 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 5062 * refcnt and needs to be destroyed. 5063 */ 5064 static void pwq_release_workfn(struct kthread_work *work) 5065 { 5066 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 5067 release_work); 5068 struct workqueue_struct *wq = pwq->wq; 5069 struct worker_pool *pool = pwq->pool; 5070 bool is_last = false; 5071 5072 /* 5073 * When @pwq is not linked, it doesn't hold any reference to the 5074 * @wq, and @wq is invalid to access. 5075 */ 5076 if (!list_empty(&pwq->pwqs_node)) { 5077 mutex_lock(&wq->mutex); 5078 list_del_rcu(&pwq->pwqs_node); 5079 is_last = list_empty(&wq->pwqs); 5080 5081 /* 5082 * For ordered workqueue with a plugged dfl_pwq, restart it now. 5083 */ 5084 if (!is_last && (wq->flags & __WQ_ORDERED)) 5085 unplug_oldest_pwq(wq); 5086 5087 mutex_unlock(&wq->mutex); 5088 } 5089 5090 if (wq->flags & WQ_UNBOUND) { 5091 mutex_lock(&wq_pool_mutex); 5092 put_unbound_pool(pool); 5093 mutex_unlock(&wq_pool_mutex); 5094 } 5095 5096 if (!list_empty(&pwq->pending_node)) { 5097 struct wq_node_nr_active *nna = 5098 wq_node_nr_active(pwq->wq, pwq->pool->node); 5099 5100 raw_spin_lock_irq(&nna->lock); 5101 list_del_init(&pwq->pending_node); 5102 raw_spin_unlock_irq(&nna->lock); 5103 } 5104 5105 kfree_rcu(pwq, rcu); 5106 5107 /* 5108 * If we're the last pwq going away, @wq is already dead and no one 5109 * is gonna access it anymore. Schedule RCU free. 5110 */ 5111 if (is_last) { 5112 wq_unregister_lockdep(wq); 5113 call_rcu(&wq->rcu, rcu_free_wq); 5114 } 5115 } 5116 5117 /* initialize newly allocated @pwq which is associated with @wq and @pool */ 5118 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 5119 struct worker_pool *pool) 5120 { 5121 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK); 5122 5123 memset(pwq, 0, sizeof(*pwq)); 5124 5125 pwq->pool = pool; 5126 pwq->wq = wq; 5127 pwq->flush_color = -1; 5128 pwq->refcnt = 1; 5129 INIT_LIST_HEAD(&pwq->inactive_works); 5130 INIT_LIST_HEAD(&pwq->pending_node); 5131 INIT_LIST_HEAD(&pwq->pwqs_node); 5132 INIT_LIST_HEAD(&pwq->mayday_node); 5133 kthread_init_work(&pwq->release_work, pwq_release_workfn); 5134 } 5135 5136 /* sync @pwq with the current state of its associated wq and link it */ 5137 static void link_pwq(struct pool_workqueue *pwq) 5138 { 5139 struct workqueue_struct *wq = pwq->wq; 5140 5141 lockdep_assert_held(&wq->mutex); 5142 5143 /* may be called multiple times, ignore if already linked */ 5144 if (!list_empty(&pwq->pwqs_node)) 5145 return; 5146 5147 /* set the matching work_color */ 5148 pwq->work_color = wq->work_color; 5149 5150 /* link in @pwq */ 5151 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); 5152 } 5153 5154 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 5155 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 5156 const struct workqueue_attrs *attrs) 5157 { 5158 struct worker_pool *pool; 5159 struct pool_workqueue *pwq; 5160 5161 lockdep_assert_held(&wq_pool_mutex); 5162 5163 pool = get_unbound_pool(attrs); 5164 if (!pool) 5165 return NULL; 5166 5167 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 5168 if (!pwq) { 5169 put_unbound_pool(pool); 5170 return NULL; 5171 } 5172 5173 init_pwq(pwq, wq, pool); 5174 return pwq; 5175 } 5176 5177 static void apply_wqattrs_lock(void) 5178 { 5179 mutex_lock(&wq_pool_mutex); 5180 } 5181 5182 static void apply_wqattrs_unlock(void) 5183 { 5184 mutex_unlock(&wq_pool_mutex); 5185 } 5186 5187 /** 5188 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 5189 * @attrs: the wq_attrs of the default pwq of the target workqueue 5190 * @cpu: the target CPU 5191 * 5192 * Calculate the cpumask a workqueue with @attrs should use on @pod. 5193 * The result is stored in @attrs->__pod_cpumask. 5194 * 5195 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 5196 * and @pod has online CPUs requested by @attrs, the returned cpumask is the 5197 * intersection of the possible CPUs of @pod and @attrs->cpumask. 5198 * 5199 * The caller is responsible for ensuring that the cpumask of @pod stays stable. 5200 */ 5201 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu) 5202 { 5203 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5204 int pod = pt->cpu_pod[cpu]; 5205 5206 /* calculate possible CPUs in @pod that @attrs wants */ 5207 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 5208 /* does @pod have any online CPUs @attrs wants? */ 5209 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) { 5210 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 5211 return; 5212 } 5213 } 5214 5215 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ 5216 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 5217 int cpu, struct pool_workqueue *pwq) 5218 { 5219 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); 5220 struct pool_workqueue *old_pwq; 5221 5222 lockdep_assert_held(&wq_pool_mutex); 5223 lockdep_assert_held(&wq->mutex); 5224 5225 /* link_pwq() can handle duplicate calls */ 5226 link_pwq(pwq); 5227 5228 old_pwq = rcu_access_pointer(*slot); 5229 rcu_assign_pointer(*slot, pwq); 5230 return old_pwq; 5231 } 5232 5233 /* context to store the prepared attrs & pwqs before applying */ 5234 struct apply_wqattrs_ctx { 5235 struct workqueue_struct *wq; /* target workqueue */ 5236 struct workqueue_attrs *attrs; /* attrs to apply */ 5237 struct list_head list; /* queued for batching commit */ 5238 struct pool_workqueue *dfl_pwq; 5239 struct pool_workqueue *pwq_tbl[]; 5240 }; 5241 5242 /* free the resources after success or abort */ 5243 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 5244 { 5245 if (ctx) { 5246 int cpu; 5247 5248 for_each_possible_cpu(cpu) 5249 put_pwq_unlocked(ctx->pwq_tbl[cpu]); 5250 put_pwq_unlocked(ctx->dfl_pwq); 5251 5252 free_workqueue_attrs(ctx->attrs); 5253 5254 kfree(ctx); 5255 } 5256 } 5257 5258 /* allocate the attrs and pwqs for later installation */ 5259 static struct apply_wqattrs_ctx * 5260 apply_wqattrs_prepare(struct workqueue_struct *wq, 5261 const struct workqueue_attrs *attrs, 5262 const cpumask_var_t unbound_cpumask) 5263 { 5264 struct apply_wqattrs_ctx *ctx; 5265 struct workqueue_attrs *new_attrs; 5266 int cpu; 5267 5268 lockdep_assert_held(&wq_pool_mutex); 5269 5270 if (WARN_ON(attrs->affn_scope < 0 || 5271 attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 5272 return ERR_PTR(-EINVAL); 5273 5274 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 5275 5276 new_attrs = alloc_workqueue_attrs(); 5277 if (!ctx || !new_attrs) 5278 goto out_free; 5279 5280 /* 5281 * If something goes wrong during CPU up/down, we'll fall back to 5282 * the default pwq covering whole @attrs->cpumask. Always create 5283 * it even if we don't use it immediately. 5284 */ 5285 copy_workqueue_attrs(new_attrs, attrs); 5286 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 5287 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5288 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 5289 if (!ctx->dfl_pwq) 5290 goto out_free; 5291 5292 for_each_possible_cpu(cpu) { 5293 if (new_attrs->ordered) { 5294 ctx->dfl_pwq->refcnt++; 5295 ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 5296 } else { 5297 wq_calc_pod_cpumask(new_attrs, cpu); 5298 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 5299 if (!ctx->pwq_tbl[cpu]) 5300 goto out_free; 5301 } 5302 } 5303 5304 /* save the user configured attrs and sanitize it. */ 5305 copy_workqueue_attrs(new_attrs, attrs); 5306 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 5307 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5308 ctx->attrs = new_attrs; 5309 5310 /* 5311 * For initialized ordered workqueues, there should only be one pwq 5312 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution 5313 * of newly queued work items until execution of older work items in 5314 * the old pwq's have completed. 5315 */ 5316 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)) 5317 ctx->dfl_pwq->plugged = true; 5318 5319 ctx->wq = wq; 5320 return ctx; 5321 5322 out_free: 5323 free_workqueue_attrs(new_attrs); 5324 apply_wqattrs_cleanup(ctx); 5325 return ERR_PTR(-ENOMEM); 5326 } 5327 5328 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 5329 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 5330 { 5331 int cpu; 5332 5333 /* all pwqs have been created successfully, let's install'em */ 5334 mutex_lock(&ctx->wq->mutex); 5335 5336 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 5337 5338 /* save the previous pwqs and install the new ones */ 5339 for_each_possible_cpu(cpu) 5340 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 5341 ctx->pwq_tbl[cpu]); 5342 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); 5343 5344 /* update node_nr_active->max */ 5345 wq_update_node_max_active(ctx->wq, -1); 5346 5347 /* rescuer needs to respect wq cpumask changes */ 5348 if (ctx->wq->rescuer) 5349 set_cpus_allowed_ptr(ctx->wq->rescuer->task, 5350 unbound_effective_cpumask(ctx->wq)); 5351 5352 mutex_unlock(&ctx->wq->mutex); 5353 } 5354 5355 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 5356 const struct workqueue_attrs *attrs) 5357 { 5358 struct apply_wqattrs_ctx *ctx; 5359 5360 /* only unbound workqueues can change attributes */ 5361 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 5362 return -EINVAL; 5363 5364 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 5365 if (IS_ERR(ctx)) 5366 return PTR_ERR(ctx); 5367 5368 /* the ctx has been prepared successfully, let's commit it */ 5369 apply_wqattrs_commit(ctx); 5370 apply_wqattrs_cleanup(ctx); 5371 5372 return 0; 5373 } 5374 5375 /** 5376 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 5377 * @wq: the target workqueue 5378 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 5379 * 5380 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 5381 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 5382 * work items are affine to the pod it was issued on. Older pwqs are released as 5383 * in-flight work items finish. Note that a work item which repeatedly requeues 5384 * itself back-to-back will stay on its current pwq. 5385 * 5386 * Performs GFP_KERNEL allocations. 5387 * 5388 * Return: 0 on success and -errno on failure. 5389 */ 5390 int apply_workqueue_attrs(struct workqueue_struct *wq, 5391 const struct workqueue_attrs *attrs) 5392 { 5393 int ret; 5394 5395 mutex_lock(&wq_pool_mutex); 5396 ret = apply_workqueue_attrs_locked(wq, attrs); 5397 mutex_unlock(&wq_pool_mutex); 5398 5399 return ret; 5400 } 5401 5402 /** 5403 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug 5404 * @wq: the target workqueue 5405 * @cpu: the CPU to update the pwq slot for 5406 * 5407 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 5408 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged. 5409 * 5410 * 5411 * If pod affinity can't be adjusted due to memory allocation failure, it falls 5412 * back to @wq->dfl_pwq which may not be optimal but is always correct. 5413 * 5414 * Note that when the last allowed CPU of a pod goes offline for a workqueue 5415 * with a cpumask spanning multiple pods, the workers which were already 5416 * executing the work items for the workqueue will lose their CPU affinity and 5417 * may execute on any CPU. This is similar to how per-cpu workqueues behave on 5418 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 5419 * responsibility to flush the work item from CPU_DOWN_PREPARE. 5420 */ 5421 static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu) 5422 { 5423 struct pool_workqueue *old_pwq = NULL, *pwq; 5424 struct workqueue_attrs *target_attrs; 5425 5426 lockdep_assert_held(&wq_pool_mutex); 5427 5428 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 5429 return; 5430 5431 /* 5432 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 5433 * Let's use a preallocated one. The following buf is protected by 5434 * CPU hotplug exclusion. 5435 */ 5436 target_attrs = unbound_wq_update_pwq_attrs_buf; 5437 5438 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 5439 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 5440 5441 /* nothing to do if the target cpumask matches the current pwq */ 5442 wq_calc_pod_cpumask(target_attrs, cpu); 5443 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) 5444 return; 5445 5446 /* create a new pwq */ 5447 pwq = alloc_unbound_pwq(wq, target_attrs); 5448 if (!pwq) { 5449 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 5450 wq->name); 5451 goto use_dfl_pwq; 5452 } 5453 5454 /* Install the new pwq. */ 5455 mutex_lock(&wq->mutex); 5456 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5457 goto out_unlock; 5458 5459 use_dfl_pwq: 5460 mutex_lock(&wq->mutex); 5461 pwq = unbound_pwq(wq, -1); 5462 raw_spin_lock_irq(&pwq->pool->lock); 5463 get_pwq(pwq); 5464 raw_spin_unlock_irq(&pwq->pool->lock); 5465 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5466 out_unlock: 5467 mutex_unlock(&wq->mutex); 5468 put_pwq_unlocked(old_pwq); 5469 } 5470 5471 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 5472 { 5473 bool highpri = wq->flags & WQ_HIGHPRI; 5474 int cpu, ret; 5475 5476 lockdep_assert_held(&wq_pool_mutex); 5477 5478 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 5479 if (!wq->cpu_pwq) 5480 goto enomem; 5481 5482 if (!(wq->flags & WQ_UNBOUND)) { 5483 struct worker_pool __percpu *pools; 5484 5485 if (wq->flags & WQ_BH) 5486 pools = bh_worker_pools; 5487 else 5488 pools = cpu_worker_pools; 5489 5490 for_each_possible_cpu(cpu) { 5491 struct pool_workqueue **pwq_p; 5492 struct worker_pool *pool; 5493 5494 pool = &(per_cpu_ptr(pools, cpu)[highpri]); 5495 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu); 5496 5497 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 5498 pool->node); 5499 if (!*pwq_p) 5500 goto enomem; 5501 5502 init_pwq(*pwq_p, wq, pool); 5503 5504 mutex_lock(&wq->mutex); 5505 link_pwq(*pwq_p); 5506 mutex_unlock(&wq->mutex); 5507 } 5508 return 0; 5509 } 5510 5511 if (wq->flags & __WQ_ORDERED) { 5512 struct pool_workqueue *dfl_pwq; 5513 5514 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]); 5515 /* there should only be single pwq for ordering guarantee */ 5516 dfl_pwq = rcu_access_pointer(wq->dfl_pwq); 5517 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || 5518 wq->pwqs.prev != &dfl_pwq->pwqs_node), 5519 "ordering guarantee broken for workqueue %s\n", wq->name); 5520 } else { 5521 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]); 5522 } 5523 5524 return ret; 5525 5526 enomem: 5527 if (wq->cpu_pwq) { 5528 for_each_possible_cpu(cpu) { 5529 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 5530 5531 if (pwq) 5532 kmem_cache_free(pwq_cache, pwq); 5533 } 5534 free_percpu(wq->cpu_pwq); 5535 wq->cpu_pwq = NULL; 5536 } 5537 return -ENOMEM; 5538 } 5539 5540 static int wq_clamp_max_active(int max_active, unsigned int flags, 5541 const char *name) 5542 { 5543 if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 5544 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 5545 max_active, name, 1, WQ_MAX_ACTIVE); 5546 5547 return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 5548 } 5549 5550 /* 5551 * Workqueues which may be used during memory reclaim should have a rescuer 5552 * to guarantee forward progress. 5553 */ 5554 static int init_rescuer(struct workqueue_struct *wq) 5555 { 5556 struct worker *rescuer; 5557 char id_buf[WORKER_ID_LEN]; 5558 int ret; 5559 5560 lockdep_assert_held(&wq_pool_mutex); 5561 5562 if (!(wq->flags & WQ_MEM_RECLAIM)) 5563 return 0; 5564 5565 rescuer = alloc_worker(NUMA_NO_NODE); 5566 if (!rescuer) { 5567 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 5568 wq->name); 5569 return -ENOMEM; 5570 } 5571 5572 rescuer->rescue_wq = wq; 5573 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL); 5574 5575 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf); 5576 if (IS_ERR(rescuer->task)) { 5577 ret = PTR_ERR(rescuer->task); 5578 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 5579 wq->name, ERR_PTR(ret)); 5580 kfree(rescuer); 5581 return ret; 5582 } 5583 5584 wq->rescuer = rescuer; 5585 if (wq->flags & WQ_UNBOUND) 5586 kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq)); 5587 else 5588 kthread_bind_mask(rescuer->task, cpu_possible_mask); 5589 wake_up_process(rescuer->task); 5590 5591 return 0; 5592 } 5593 5594 /** 5595 * wq_adjust_max_active - update a wq's max_active to the current setting 5596 * @wq: target workqueue 5597 * 5598 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and 5599 * activate inactive work items accordingly. If @wq is freezing, clear 5600 * @wq->max_active to zero. 5601 */ 5602 static void wq_adjust_max_active(struct workqueue_struct *wq) 5603 { 5604 bool activated; 5605 int new_max, new_min; 5606 5607 lockdep_assert_held(&wq->mutex); 5608 5609 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { 5610 new_max = 0; 5611 new_min = 0; 5612 } else { 5613 new_max = wq->saved_max_active; 5614 new_min = wq->saved_min_active; 5615 } 5616 5617 if (wq->max_active == new_max && wq->min_active == new_min) 5618 return; 5619 5620 /* 5621 * Update @wq->max/min_active and then kick inactive work items if more 5622 * active work items are allowed. This doesn't break work item ordering 5623 * because new work items are always queued behind existing inactive 5624 * work items if there are any. 5625 */ 5626 WRITE_ONCE(wq->max_active, new_max); 5627 WRITE_ONCE(wq->min_active, new_min); 5628 5629 if (wq->flags & WQ_UNBOUND) 5630 wq_update_node_max_active(wq, -1); 5631 5632 if (new_max == 0) 5633 return; 5634 5635 /* 5636 * Round-robin through pwq's activating the first inactive work item 5637 * until max_active is filled. 5638 */ 5639 do { 5640 struct pool_workqueue *pwq; 5641 5642 activated = false; 5643 for_each_pwq(pwq, wq) { 5644 unsigned long irq_flags; 5645 5646 /* can be called during early boot w/ irq disabled */ 5647 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 5648 if (pwq_activate_first_inactive(pwq, true)) { 5649 activated = true; 5650 kick_pool(pwq->pool); 5651 } 5652 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 5653 } 5654 } while (activated); 5655 } 5656 5657 __printf(1, 0) 5658 static struct workqueue_struct *__alloc_workqueue(const char *fmt, 5659 unsigned int flags, 5660 int max_active, va_list args) 5661 { 5662 struct workqueue_struct *wq; 5663 size_t wq_size; 5664 int name_len; 5665 5666 if (flags & WQ_BH) { 5667 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS)) 5668 return NULL; 5669 if (WARN_ON_ONCE(max_active)) 5670 return NULL; 5671 } 5672 5673 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 5674 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 5675 flags |= WQ_UNBOUND; 5676 5677 /* allocate wq and format name */ 5678 if (flags & WQ_UNBOUND) 5679 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); 5680 else 5681 wq_size = sizeof(*wq); 5682 5683 wq = kzalloc_noprof(wq_size, GFP_KERNEL); 5684 if (!wq) 5685 return NULL; 5686 5687 if (flags & WQ_UNBOUND) { 5688 wq->unbound_attrs = alloc_workqueue_attrs_noprof(); 5689 if (!wq->unbound_attrs) 5690 goto err_free_wq; 5691 } 5692 5693 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); 5694 5695 if (name_len >= WQ_NAME_LEN) 5696 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", 5697 wq->name); 5698 5699 if (flags & WQ_BH) { 5700 /* 5701 * BH workqueues always share a single execution context per CPU 5702 * and don't impose any max_active limit. 5703 */ 5704 max_active = INT_MAX; 5705 } else { 5706 max_active = max_active ?: WQ_DFL_ACTIVE; 5707 max_active = wq_clamp_max_active(max_active, flags, wq->name); 5708 } 5709 5710 /* init wq */ 5711 wq->flags = flags; 5712 wq->max_active = max_active; 5713 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); 5714 wq->saved_max_active = wq->max_active; 5715 wq->saved_min_active = wq->min_active; 5716 mutex_init(&wq->mutex); 5717 atomic_set(&wq->nr_pwqs_to_flush, 0); 5718 INIT_LIST_HEAD(&wq->pwqs); 5719 INIT_LIST_HEAD(&wq->flusher_queue); 5720 INIT_LIST_HEAD(&wq->flusher_overflow); 5721 INIT_LIST_HEAD(&wq->maydays); 5722 5723 INIT_LIST_HEAD(&wq->list); 5724 5725 if (flags & WQ_UNBOUND) { 5726 if (alloc_node_nr_active(wq->node_nr_active) < 0) 5727 goto err_free_wq; 5728 } 5729 5730 /* 5731 * wq_pool_mutex protects the workqueues list, allocations of PWQs, 5732 * and the global freeze state. 5733 */ 5734 apply_wqattrs_lock(); 5735 5736 if (alloc_and_link_pwqs(wq) < 0) 5737 goto err_unlock_free_node_nr_active; 5738 5739 mutex_lock(&wq->mutex); 5740 wq_adjust_max_active(wq); 5741 mutex_unlock(&wq->mutex); 5742 5743 list_add_tail_rcu(&wq->list, &workqueues); 5744 5745 if (wq_online && init_rescuer(wq) < 0) 5746 goto err_unlock_destroy; 5747 5748 apply_wqattrs_unlock(); 5749 5750 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 5751 goto err_destroy; 5752 5753 return wq; 5754 5755 err_unlock_free_node_nr_active: 5756 apply_wqattrs_unlock(); 5757 /* 5758 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work, 5759 * flushing the pwq_release_worker ensures that the pwq_release_workfn() 5760 * completes before calling kfree(wq). 5761 */ 5762 if (wq->flags & WQ_UNBOUND) { 5763 kthread_flush_worker(pwq_release_worker); 5764 free_node_nr_active(wq->node_nr_active); 5765 } 5766 err_free_wq: 5767 free_workqueue_attrs(wq->unbound_attrs); 5768 kfree(wq); 5769 return NULL; 5770 err_unlock_destroy: 5771 apply_wqattrs_unlock(); 5772 err_destroy: 5773 destroy_workqueue(wq); 5774 return NULL; 5775 } 5776 5777 __printf(1, 4) 5778 struct workqueue_struct *alloc_workqueue_noprof(const char *fmt, 5779 unsigned int flags, 5780 int max_active, ...) 5781 { 5782 struct workqueue_struct *wq; 5783 va_list args; 5784 5785 va_start(args, max_active); 5786 wq = __alloc_workqueue(fmt, flags, max_active, args); 5787 va_end(args); 5788 if (!wq) 5789 return NULL; 5790 5791 wq_init_lockdep(wq); 5792 5793 return wq; 5794 } 5795 EXPORT_SYMBOL_GPL(alloc_workqueue_noprof); 5796 5797 #ifdef CONFIG_LOCKDEP 5798 __printf(1, 5) 5799 struct workqueue_struct * 5800 alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags, 5801 int max_active, struct lockdep_map *lockdep_map, ...) 5802 { 5803 struct workqueue_struct *wq; 5804 va_list args; 5805 5806 va_start(args, lockdep_map); 5807 wq = __alloc_workqueue(fmt, flags, max_active, args); 5808 va_end(args); 5809 if (!wq) 5810 return NULL; 5811 5812 wq->lockdep_map = lockdep_map; 5813 5814 return wq; 5815 } 5816 EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map); 5817 #endif 5818 5819 static bool pwq_busy(struct pool_workqueue *pwq) 5820 { 5821 int i; 5822 5823 for (i = 0; i < WORK_NR_COLORS; i++) 5824 if (pwq->nr_in_flight[i]) 5825 return true; 5826 5827 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) 5828 return true; 5829 if (!pwq_is_empty(pwq)) 5830 return true; 5831 5832 return false; 5833 } 5834 5835 /** 5836 * destroy_workqueue - safely terminate a workqueue 5837 * @wq: target workqueue 5838 * 5839 * Safely destroy a workqueue. All work currently pending will be done first. 5840 * 5841 * This function does NOT guarantee that non-pending work that has been 5842 * submitted with queue_delayed_work() and similar functions will be done 5843 * before destroying the workqueue. The fundamental problem is that, currently, 5844 * the workqueue has no way of accessing non-pending delayed_work. delayed_work 5845 * is only linked on the timer-side. All delayed_work must, therefore, be 5846 * canceled before calling this function. 5847 * 5848 * TODO: It would be better if the problem described above wouldn't exist and 5849 * destroy_workqueue() would cleanly cancel all pending and non-pending 5850 * delayed_work. 5851 */ 5852 void destroy_workqueue(struct workqueue_struct *wq) 5853 { 5854 struct pool_workqueue *pwq; 5855 int cpu; 5856 5857 /* 5858 * Remove it from sysfs first so that sanity check failure doesn't 5859 * lead to sysfs name conflicts. 5860 */ 5861 workqueue_sysfs_unregister(wq); 5862 5863 /* mark the workqueue destruction is in progress */ 5864 mutex_lock(&wq->mutex); 5865 wq->flags |= __WQ_DESTROYING; 5866 mutex_unlock(&wq->mutex); 5867 5868 /* drain it before proceeding with destruction */ 5869 drain_workqueue(wq); 5870 5871 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 5872 if (wq->rescuer) { 5873 struct worker *rescuer = wq->rescuer; 5874 5875 /* this prevents new queueing */ 5876 raw_spin_lock_irq(&wq_mayday_lock); 5877 wq->rescuer = NULL; 5878 raw_spin_unlock_irq(&wq_mayday_lock); 5879 5880 /* rescuer will empty maydays list before exiting */ 5881 kthread_stop(rescuer->task); 5882 kfree(rescuer); 5883 } 5884 5885 /* 5886 * Sanity checks - grab all the locks so that we wait for all 5887 * in-flight operations which may do put_pwq(). 5888 */ 5889 mutex_lock(&wq_pool_mutex); 5890 mutex_lock(&wq->mutex); 5891 for_each_pwq(pwq, wq) { 5892 raw_spin_lock_irq(&pwq->pool->lock); 5893 if (WARN_ON(pwq_busy(pwq))) { 5894 pr_warn("%s: %s has the following busy pwq\n", 5895 __func__, wq->name); 5896 show_pwq(pwq); 5897 raw_spin_unlock_irq(&pwq->pool->lock); 5898 mutex_unlock(&wq->mutex); 5899 mutex_unlock(&wq_pool_mutex); 5900 show_one_workqueue(wq); 5901 return; 5902 } 5903 raw_spin_unlock_irq(&pwq->pool->lock); 5904 } 5905 mutex_unlock(&wq->mutex); 5906 5907 /* 5908 * wq list is used to freeze wq, remove from list after 5909 * flushing is complete in case freeze races us. 5910 */ 5911 list_del_rcu(&wq->list); 5912 mutex_unlock(&wq_pool_mutex); 5913 5914 /* 5915 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 5916 * to put the base refs. @wq will be auto-destroyed from the last 5917 * pwq_put. RCU read lock prevents @wq from going away from under us. 5918 */ 5919 rcu_read_lock(); 5920 5921 for_each_possible_cpu(cpu) { 5922 put_pwq_unlocked(unbound_pwq(wq, cpu)); 5923 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); 5924 } 5925 5926 put_pwq_unlocked(unbound_pwq(wq, -1)); 5927 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); 5928 5929 rcu_read_unlock(); 5930 } 5931 EXPORT_SYMBOL_GPL(destroy_workqueue); 5932 5933 /** 5934 * workqueue_set_max_active - adjust max_active of a workqueue 5935 * @wq: target workqueue 5936 * @max_active: new max_active value. 5937 * 5938 * Set max_active of @wq to @max_active. See the alloc_workqueue() function 5939 * comment. 5940 * 5941 * CONTEXT: 5942 * Don't call from IRQ context. 5943 */ 5944 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 5945 { 5946 /* max_active doesn't mean anything for BH workqueues */ 5947 if (WARN_ON(wq->flags & WQ_BH)) 5948 return; 5949 /* disallow meddling with max_active for ordered workqueues */ 5950 if (WARN_ON(wq->flags & __WQ_ORDERED)) 5951 return; 5952 5953 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 5954 5955 mutex_lock(&wq->mutex); 5956 5957 wq->saved_max_active = max_active; 5958 if (wq->flags & WQ_UNBOUND) 5959 wq->saved_min_active = min(wq->saved_min_active, max_active); 5960 5961 wq_adjust_max_active(wq); 5962 5963 mutex_unlock(&wq->mutex); 5964 } 5965 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 5966 5967 /** 5968 * workqueue_set_min_active - adjust min_active of an unbound workqueue 5969 * @wq: target unbound workqueue 5970 * @min_active: new min_active value 5971 * 5972 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an 5973 * unbound workqueue is not guaranteed to be able to process max_active 5974 * interdependent work items. Instead, an unbound workqueue is guaranteed to be 5975 * able to process min_active number of interdependent work items which is 5976 * %WQ_DFL_MIN_ACTIVE by default. 5977 * 5978 * Use this function to adjust the min_active value between 0 and the current 5979 * max_active. 5980 */ 5981 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active) 5982 { 5983 /* min_active is only meaningful for non-ordered unbound workqueues */ 5984 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) != 5985 WQ_UNBOUND)) 5986 return; 5987 5988 mutex_lock(&wq->mutex); 5989 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active); 5990 wq_adjust_max_active(wq); 5991 mutex_unlock(&wq->mutex); 5992 } 5993 5994 /** 5995 * current_work - retrieve %current task's work struct 5996 * 5997 * Determine if %current task is a workqueue worker and what it's working on. 5998 * Useful to find out the context that the %current task is running in. 5999 * 6000 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 6001 */ 6002 struct work_struct *current_work(void) 6003 { 6004 struct worker *worker = current_wq_worker(); 6005 6006 return worker ? worker->current_work : NULL; 6007 } 6008 EXPORT_SYMBOL(current_work); 6009 6010 /** 6011 * current_is_workqueue_rescuer - is %current workqueue rescuer? 6012 * 6013 * Determine whether %current is a workqueue rescuer. Can be used from 6014 * work functions to determine whether it's being run off the rescuer task. 6015 * 6016 * Return: %true if %current is a workqueue rescuer. %false otherwise. 6017 */ 6018 bool current_is_workqueue_rescuer(void) 6019 { 6020 struct worker *worker = current_wq_worker(); 6021 6022 return worker && worker->rescue_wq; 6023 } 6024 6025 /** 6026 * workqueue_congested - test whether a workqueue is congested 6027 * @cpu: CPU in question 6028 * @wq: target workqueue 6029 * 6030 * Test whether @wq's cpu workqueue for @cpu is congested. There is 6031 * no synchronization around this function and the test result is 6032 * unreliable and only useful as advisory hints or for debugging. 6033 * 6034 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 6035 * 6036 * With the exception of ordered workqueues, all workqueues have per-cpu 6037 * pool_workqueues, each with its own congested state. A workqueue being 6038 * congested on one CPU doesn't mean that the workqueue is contested on any 6039 * other CPUs. 6040 * 6041 * Return: 6042 * %true if congested, %false otherwise. 6043 */ 6044 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 6045 { 6046 struct pool_workqueue *pwq; 6047 bool ret; 6048 6049 rcu_read_lock(); 6050 preempt_disable(); 6051 6052 if (cpu == WORK_CPU_UNBOUND) 6053 cpu = smp_processor_id(); 6054 6055 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 6056 ret = !list_empty(&pwq->inactive_works); 6057 6058 preempt_enable(); 6059 rcu_read_unlock(); 6060 6061 return ret; 6062 } 6063 EXPORT_SYMBOL_GPL(workqueue_congested); 6064 6065 /** 6066 * work_busy - test whether a work is currently pending or running 6067 * @work: the work to be tested 6068 * 6069 * Test whether @work is currently pending or running. There is no 6070 * synchronization around this function and the test result is 6071 * unreliable and only useful as advisory hints or for debugging. 6072 * 6073 * Return: 6074 * OR'd bitmask of WORK_BUSY_* bits. 6075 */ 6076 unsigned int work_busy(struct work_struct *work) 6077 { 6078 struct worker_pool *pool; 6079 unsigned long irq_flags; 6080 unsigned int ret = 0; 6081 6082 if (work_pending(work)) 6083 ret |= WORK_BUSY_PENDING; 6084 6085 rcu_read_lock(); 6086 pool = get_work_pool(work); 6087 if (pool) { 6088 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6089 if (find_worker_executing_work(pool, work)) 6090 ret |= WORK_BUSY_RUNNING; 6091 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6092 } 6093 rcu_read_unlock(); 6094 6095 return ret; 6096 } 6097 EXPORT_SYMBOL_GPL(work_busy); 6098 6099 /** 6100 * set_worker_desc - set description for the current work item 6101 * @fmt: printf-style format string 6102 * @...: arguments for the format string 6103 * 6104 * This function can be called by a running work function to describe what 6105 * the work item is about. If the worker task gets dumped, this 6106 * information will be printed out together to help debugging. The 6107 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 6108 */ 6109 void set_worker_desc(const char *fmt, ...) 6110 { 6111 struct worker *worker = current_wq_worker(); 6112 va_list args; 6113 6114 if (worker) { 6115 va_start(args, fmt); 6116 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 6117 va_end(args); 6118 } 6119 } 6120 EXPORT_SYMBOL_GPL(set_worker_desc); 6121 6122 /** 6123 * print_worker_info - print out worker information and description 6124 * @log_lvl: the log level to use when printing 6125 * @task: target task 6126 * 6127 * If @task is a worker and currently executing a work item, print out the 6128 * name of the workqueue being serviced and worker description set with 6129 * set_worker_desc() by the currently executing work item. 6130 * 6131 * This function can be safely called on any task as long as the 6132 * task_struct itself is accessible. While safe, this function isn't 6133 * synchronized and may print out mixups or garbages of limited length. 6134 */ 6135 void print_worker_info(const char *log_lvl, struct task_struct *task) 6136 { 6137 work_func_t *fn = NULL; 6138 char name[WQ_NAME_LEN] = { }; 6139 char desc[WORKER_DESC_LEN] = { }; 6140 struct pool_workqueue *pwq = NULL; 6141 struct workqueue_struct *wq = NULL; 6142 struct worker *worker; 6143 6144 if (!(task->flags & PF_WQ_WORKER)) 6145 return; 6146 6147 /* 6148 * This function is called without any synchronization and @task 6149 * could be in any state. Be careful with dereferences. 6150 */ 6151 worker = kthread_probe_data(task); 6152 6153 /* 6154 * Carefully copy the associated workqueue's workfn, name and desc. 6155 * Keep the original last '\0' in case the original is garbage. 6156 */ 6157 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 6158 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 6159 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 6160 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 6161 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 6162 6163 if (fn || name[0] || desc[0]) { 6164 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 6165 if (strcmp(name, desc)) 6166 pr_cont(" (%s)", desc); 6167 pr_cont("\n"); 6168 } 6169 } 6170 6171 static void pr_cont_pool_info(struct worker_pool *pool) 6172 { 6173 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 6174 if (pool->node != NUMA_NO_NODE) 6175 pr_cont(" node=%d", pool->node); 6176 pr_cont(" flags=0x%x", pool->flags); 6177 if (pool->flags & POOL_BH) 6178 pr_cont(" bh%s", 6179 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6180 else 6181 pr_cont(" nice=%d", pool->attrs->nice); 6182 } 6183 6184 static void pr_cont_worker_id(struct worker *worker) 6185 { 6186 struct worker_pool *pool = worker->pool; 6187 6188 if (pool->flags & WQ_BH) 6189 pr_cont("bh%s", 6190 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6191 else 6192 pr_cont("%d%s", task_pid_nr(worker->task), 6193 worker->rescue_wq ? "(RESCUER)" : ""); 6194 } 6195 6196 struct pr_cont_work_struct { 6197 bool comma; 6198 work_func_t func; 6199 long ctr; 6200 }; 6201 6202 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 6203 { 6204 if (!pcwsp->ctr) 6205 goto out_record; 6206 if (func == pcwsp->func) { 6207 pcwsp->ctr++; 6208 return; 6209 } 6210 if (pcwsp->ctr == 1) 6211 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 6212 else 6213 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 6214 pcwsp->ctr = 0; 6215 out_record: 6216 if ((long)func == -1L) 6217 return; 6218 pcwsp->comma = comma; 6219 pcwsp->func = func; 6220 pcwsp->ctr = 1; 6221 } 6222 6223 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 6224 { 6225 if (work->func == wq_barrier_func) { 6226 struct wq_barrier *barr; 6227 6228 barr = container_of(work, struct wq_barrier, work); 6229 6230 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6231 pr_cont("%s BAR(%d)", comma ? "," : "", 6232 task_pid_nr(barr->task)); 6233 } else { 6234 if (!comma) 6235 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6236 pr_cont_work_flush(comma, work->func, pcwsp); 6237 } 6238 } 6239 6240 static void show_pwq(struct pool_workqueue *pwq) 6241 { 6242 struct pr_cont_work_struct pcws = { .ctr = 0, }; 6243 struct worker_pool *pool = pwq->pool; 6244 struct work_struct *work; 6245 struct worker *worker; 6246 bool has_in_flight = false, has_pending = false; 6247 int bkt; 6248 6249 pr_info(" pwq %d:", pool->id); 6250 pr_cont_pool_info(pool); 6251 6252 pr_cont(" active=%d refcnt=%d%s\n", 6253 pwq->nr_active, pwq->refcnt, 6254 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 6255 6256 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6257 if (worker->current_pwq == pwq) { 6258 has_in_flight = true; 6259 break; 6260 } 6261 } 6262 if (has_in_flight) { 6263 bool comma = false; 6264 6265 pr_info(" in-flight:"); 6266 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6267 if (worker->current_pwq != pwq) 6268 continue; 6269 6270 pr_cont(" %s", comma ? "," : ""); 6271 pr_cont_worker_id(worker); 6272 pr_cont(":%ps", worker->current_func); 6273 list_for_each_entry(work, &worker->scheduled, entry) 6274 pr_cont_work(false, work, &pcws); 6275 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6276 comma = true; 6277 } 6278 pr_cont("\n"); 6279 } 6280 6281 list_for_each_entry(work, &pool->worklist, entry) { 6282 if (get_work_pwq(work) == pwq) { 6283 has_pending = true; 6284 break; 6285 } 6286 } 6287 if (has_pending) { 6288 bool comma = false; 6289 6290 pr_info(" pending:"); 6291 list_for_each_entry(work, &pool->worklist, entry) { 6292 if (get_work_pwq(work) != pwq) 6293 continue; 6294 6295 pr_cont_work(comma, work, &pcws); 6296 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6297 } 6298 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6299 pr_cont("\n"); 6300 } 6301 6302 if (!list_empty(&pwq->inactive_works)) { 6303 bool comma = false; 6304 6305 pr_info(" inactive:"); 6306 list_for_each_entry(work, &pwq->inactive_works, entry) { 6307 pr_cont_work(comma, work, &pcws); 6308 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6309 } 6310 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6311 pr_cont("\n"); 6312 } 6313 } 6314 6315 /** 6316 * show_one_workqueue - dump state of specified workqueue 6317 * @wq: workqueue whose state will be printed 6318 */ 6319 void show_one_workqueue(struct workqueue_struct *wq) 6320 { 6321 struct pool_workqueue *pwq; 6322 bool idle = true; 6323 unsigned long irq_flags; 6324 6325 for_each_pwq(pwq, wq) { 6326 if (!pwq_is_empty(pwq)) { 6327 idle = false; 6328 break; 6329 } 6330 } 6331 if (idle) /* Nothing to print for idle workqueue */ 6332 return; 6333 6334 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 6335 6336 for_each_pwq(pwq, wq) { 6337 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 6338 if (!pwq_is_empty(pwq)) { 6339 /* 6340 * Defer printing to avoid deadlocks in console 6341 * drivers that queue work while holding locks 6342 * also taken in their write paths. 6343 */ 6344 printk_deferred_enter(); 6345 show_pwq(pwq); 6346 printk_deferred_exit(); 6347 } 6348 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 6349 /* 6350 * We could be printing a lot from atomic context, e.g. 6351 * sysrq-t -> show_all_workqueues(). Avoid triggering 6352 * hard lockup. 6353 */ 6354 touch_nmi_watchdog(); 6355 } 6356 6357 } 6358 6359 /** 6360 * show_one_worker_pool - dump state of specified worker pool 6361 * @pool: worker pool whose state will be printed 6362 */ 6363 static void show_one_worker_pool(struct worker_pool *pool) 6364 { 6365 struct worker *worker; 6366 bool first = true; 6367 unsigned long irq_flags; 6368 unsigned long hung = 0; 6369 6370 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6371 if (pool->nr_workers == pool->nr_idle) 6372 goto next_pool; 6373 6374 /* How long the first pending work is waiting for a worker. */ 6375 if (!list_empty(&pool->worklist)) 6376 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 6377 6378 /* 6379 * Defer printing to avoid deadlocks in console drivers that 6380 * queue work while holding locks also taken in their write 6381 * paths. 6382 */ 6383 printk_deferred_enter(); 6384 pr_info("pool %d:", pool->id); 6385 pr_cont_pool_info(pool); 6386 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 6387 if (pool->manager) 6388 pr_cont(" manager: %d", 6389 task_pid_nr(pool->manager->task)); 6390 list_for_each_entry(worker, &pool->idle_list, entry) { 6391 pr_cont(" %s", first ? "idle: " : ""); 6392 pr_cont_worker_id(worker); 6393 first = false; 6394 } 6395 pr_cont("\n"); 6396 printk_deferred_exit(); 6397 next_pool: 6398 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6399 /* 6400 * We could be printing a lot from atomic context, e.g. 6401 * sysrq-t -> show_all_workqueues(). Avoid triggering 6402 * hard lockup. 6403 */ 6404 touch_nmi_watchdog(); 6405 6406 } 6407 6408 /** 6409 * show_all_workqueues - dump workqueue state 6410 * 6411 * Called from a sysrq handler and prints out all busy workqueues and pools. 6412 */ 6413 void show_all_workqueues(void) 6414 { 6415 struct workqueue_struct *wq; 6416 struct worker_pool *pool; 6417 int pi; 6418 6419 rcu_read_lock(); 6420 6421 pr_info("Showing busy workqueues and worker pools:\n"); 6422 6423 list_for_each_entry_rcu(wq, &workqueues, list) 6424 show_one_workqueue(wq); 6425 6426 for_each_pool(pool, pi) 6427 show_one_worker_pool(pool); 6428 6429 rcu_read_unlock(); 6430 } 6431 6432 /** 6433 * show_freezable_workqueues - dump freezable workqueue state 6434 * 6435 * Called from try_to_freeze_tasks() and prints out all freezable workqueues 6436 * still busy. 6437 */ 6438 void show_freezable_workqueues(void) 6439 { 6440 struct workqueue_struct *wq; 6441 6442 rcu_read_lock(); 6443 6444 pr_info("Showing freezable workqueues that are still busy:\n"); 6445 6446 list_for_each_entry_rcu(wq, &workqueues, list) { 6447 if (!(wq->flags & WQ_FREEZABLE)) 6448 continue; 6449 show_one_workqueue(wq); 6450 } 6451 6452 rcu_read_unlock(); 6453 } 6454 6455 /* used to show worker information through /proc/PID/{comm,stat,status} */ 6456 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 6457 { 6458 /* stabilize PF_WQ_WORKER and worker pool association */ 6459 mutex_lock(&wq_pool_attach_mutex); 6460 6461 if (task->flags & PF_WQ_WORKER) { 6462 struct worker *worker = kthread_data(task); 6463 struct worker_pool *pool = worker->pool; 6464 int off; 6465 6466 off = format_worker_id(buf, size, worker, pool); 6467 6468 if (pool) { 6469 raw_spin_lock_irq(&pool->lock); 6470 /* 6471 * ->desc tracks information (wq name or 6472 * set_worker_desc()) for the latest execution. If 6473 * current, prepend '+', otherwise '-'. 6474 */ 6475 if (worker->desc[0] != '\0') { 6476 if (worker->current_work) 6477 scnprintf(buf + off, size - off, "+%s", 6478 worker->desc); 6479 else 6480 scnprintf(buf + off, size - off, "-%s", 6481 worker->desc); 6482 } 6483 raw_spin_unlock_irq(&pool->lock); 6484 } 6485 } else { 6486 strscpy(buf, task->comm, size); 6487 } 6488 6489 mutex_unlock(&wq_pool_attach_mutex); 6490 } 6491 6492 #ifdef CONFIG_SMP 6493 6494 /* 6495 * CPU hotplug. 6496 * 6497 * There are two challenges in supporting CPU hotplug. Firstly, there 6498 * are a lot of assumptions on strong associations among work, pwq and 6499 * pool which make migrating pending and scheduled works very 6500 * difficult to implement without impacting hot paths. Secondly, 6501 * worker pools serve mix of short, long and very long running works making 6502 * blocked draining impractical. 6503 * 6504 * This is solved by allowing the pools to be disassociated from the CPU 6505 * running as an unbound one and allowing it to be reattached later if the 6506 * cpu comes back online. 6507 */ 6508 6509 static void unbind_workers(int cpu) 6510 { 6511 struct worker_pool *pool; 6512 struct worker *worker; 6513 6514 for_each_cpu_worker_pool(pool, cpu) { 6515 mutex_lock(&wq_pool_attach_mutex); 6516 raw_spin_lock_irq(&pool->lock); 6517 6518 /* 6519 * We've blocked all attach/detach operations. Make all workers 6520 * unbound and set DISASSOCIATED. Before this, all workers 6521 * must be on the cpu. After this, they may become diasporas. 6522 * And the preemption disabled section in their sched callbacks 6523 * are guaranteed to see WORKER_UNBOUND since the code here 6524 * is on the same cpu. 6525 */ 6526 for_each_pool_worker(worker, pool) 6527 worker->flags |= WORKER_UNBOUND; 6528 6529 pool->flags |= POOL_DISASSOCIATED; 6530 6531 /* 6532 * The handling of nr_running in sched callbacks are disabled 6533 * now. Zap nr_running. After this, nr_running stays zero and 6534 * need_more_worker() and keep_working() are always true as 6535 * long as the worklist is not empty. This pool now behaves as 6536 * an unbound (in terms of concurrency management) pool which 6537 * are served by workers tied to the pool. 6538 */ 6539 pool->nr_running = 0; 6540 6541 /* 6542 * With concurrency management just turned off, a busy 6543 * worker blocking could lead to lengthy stalls. Kick off 6544 * unbound chain execution of currently pending work items. 6545 */ 6546 kick_pool(pool); 6547 6548 raw_spin_unlock_irq(&pool->lock); 6549 6550 for_each_pool_worker(worker, pool) 6551 unbind_worker(worker); 6552 6553 mutex_unlock(&wq_pool_attach_mutex); 6554 } 6555 } 6556 6557 /** 6558 * rebind_workers - rebind all workers of a pool to the associated CPU 6559 * @pool: pool of interest 6560 * 6561 * @pool->cpu is coming online. Rebind all workers to the CPU. 6562 */ 6563 static void rebind_workers(struct worker_pool *pool) 6564 { 6565 struct worker *worker; 6566 6567 lockdep_assert_held(&wq_pool_attach_mutex); 6568 6569 /* 6570 * Restore CPU affinity of all workers. As all idle workers should 6571 * be on the run-queue of the associated CPU before any local 6572 * wake-ups for concurrency management happen, restore CPU affinity 6573 * of all workers first and then clear UNBOUND. As we're called 6574 * from CPU_ONLINE, the following shouldn't fail. 6575 */ 6576 for_each_pool_worker(worker, pool) { 6577 kthread_set_per_cpu(worker->task, pool->cpu); 6578 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 6579 pool_allowed_cpus(pool)) < 0); 6580 } 6581 6582 raw_spin_lock_irq(&pool->lock); 6583 6584 pool->flags &= ~POOL_DISASSOCIATED; 6585 6586 for_each_pool_worker(worker, pool) { 6587 unsigned int worker_flags = worker->flags; 6588 6589 /* 6590 * We want to clear UNBOUND but can't directly call 6591 * worker_clr_flags() or adjust nr_running. Atomically 6592 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 6593 * @worker will clear REBOUND using worker_clr_flags() when 6594 * it initiates the next execution cycle thus restoring 6595 * concurrency management. Note that when or whether 6596 * @worker clears REBOUND doesn't affect correctness. 6597 * 6598 * WRITE_ONCE() is necessary because @worker->flags may be 6599 * tested without holding any lock in 6600 * wq_worker_running(). Without it, NOT_RUNNING test may 6601 * fail incorrectly leading to premature concurrency 6602 * management operations. 6603 */ 6604 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 6605 worker_flags |= WORKER_REBOUND; 6606 worker_flags &= ~WORKER_UNBOUND; 6607 WRITE_ONCE(worker->flags, worker_flags); 6608 } 6609 6610 raw_spin_unlock_irq(&pool->lock); 6611 } 6612 6613 /** 6614 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 6615 * @pool: unbound pool of interest 6616 * @cpu: the CPU which is coming up 6617 * 6618 * An unbound pool may end up with a cpumask which doesn't have any online 6619 * CPUs. When a worker of such pool get scheduled, the scheduler resets 6620 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 6621 * online CPU before, cpus_allowed of all its workers should be restored. 6622 */ 6623 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 6624 { 6625 static cpumask_t cpumask; 6626 struct worker *worker; 6627 6628 lockdep_assert_held(&wq_pool_attach_mutex); 6629 6630 /* is @cpu allowed for @pool? */ 6631 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 6632 return; 6633 6634 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 6635 6636 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 6637 for_each_pool_worker(worker, pool) 6638 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 6639 } 6640 6641 int workqueue_prepare_cpu(unsigned int cpu) 6642 { 6643 struct worker_pool *pool; 6644 6645 for_each_cpu_worker_pool(pool, cpu) { 6646 if (pool->nr_workers) 6647 continue; 6648 if (!create_worker(pool)) 6649 return -ENOMEM; 6650 } 6651 return 0; 6652 } 6653 6654 int workqueue_online_cpu(unsigned int cpu) 6655 { 6656 struct worker_pool *pool; 6657 struct workqueue_struct *wq; 6658 int pi; 6659 6660 mutex_lock(&wq_pool_mutex); 6661 6662 cpumask_set_cpu(cpu, wq_online_cpumask); 6663 6664 for_each_pool(pool, pi) { 6665 /* BH pools aren't affected by hotplug */ 6666 if (pool->flags & POOL_BH) 6667 continue; 6668 6669 mutex_lock(&wq_pool_attach_mutex); 6670 if (pool->cpu == cpu) 6671 rebind_workers(pool); 6672 else if (pool->cpu < 0) 6673 restore_unbound_workers_cpumask(pool, cpu); 6674 mutex_unlock(&wq_pool_attach_mutex); 6675 } 6676 6677 /* update pod affinity of unbound workqueues */ 6678 list_for_each_entry(wq, &workqueues, list) { 6679 struct workqueue_attrs *attrs = wq->unbound_attrs; 6680 6681 if (attrs) { 6682 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6683 int tcpu; 6684 6685 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6686 unbound_wq_update_pwq(wq, tcpu); 6687 6688 mutex_lock(&wq->mutex); 6689 wq_update_node_max_active(wq, -1); 6690 mutex_unlock(&wq->mutex); 6691 } 6692 } 6693 6694 mutex_unlock(&wq_pool_mutex); 6695 return 0; 6696 } 6697 6698 int workqueue_offline_cpu(unsigned int cpu) 6699 { 6700 struct workqueue_struct *wq; 6701 6702 /* unbinding per-cpu workers should happen on the local CPU */ 6703 if (WARN_ON(cpu != smp_processor_id())) 6704 return -1; 6705 6706 unbind_workers(cpu); 6707 6708 /* update pod affinity of unbound workqueues */ 6709 mutex_lock(&wq_pool_mutex); 6710 6711 cpumask_clear_cpu(cpu, wq_online_cpumask); 6712 6713 list_for_each_entry(wq, &workqueues, list) { 6714 struct workqueue_attrs *attrs = wq->unbound_attrs; 6715 6716 if (attrs) { 6717 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6718 int tcpu; 6719 6720 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6721 unbound_wq_update_pwq(wq, tcpu); 6722 6723 mutex_lock(&wq->mutex); 6724 wq_update_node_max_active(wq, cpu); 6725 mutex_unlock(&wq->mutex); 6726 } 6727 } 6728 mutex_unlock(&wq_pool_mutex); 6729 6730 return 0; 6731 } 6732 6733 struct work_for_cpu { 6734 struct work_struct work; 6735 long (*fn)(void *); 6736 void *arg; 6737 long ret; 6738 }; 6739 6740 static void work_for_cpu_fn(struct work_struct *work) 6741 { 6742 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 6743 6744 wfc->ret = wfc->fn(wfc->arg); 6745 } 6746 6747 /** 6748 * work_on_cpu_key - run a function in thread context on a particular cpu 6749 * @cpu: the cpu to run on 6750 * @fn: the function to run 6751 * @arg: the function arg 6752 * @key: The lock class key for lock debugging purposes 6753 * 6754 * It is up to the caller to ensure that the cpu doesn't go offline. 6755 * The caller must not hold any locks which would prevent @fn from completing. 6756 * 6757 * Return: The value @fn returns. 6758 */ 6759 long work_on_cpu_key(int cpu, long (*fn)(void *), 6760 void *arg, struct lock_class_key *key) 6761 { 6762 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 6763 6764 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); 6765 schedule_work_on(cpu, &wfc.work); 6766 flush_work(&wfc.work); 6767 destroy_work_on_stack(&wfc.work); 6768 return wfc.ret; 6769 } 6770 EXPORT_SYMBOL_GPL(work_on_cpu_key); 6771 #endif /* CONFIG_SMP */ 6772 6773 #ifdef CONFIG_FREEZER 6774 6775 /** 6776 * freeze_workqueues_begin - begin freezing workqueues 6777 * 6778 * Start freezing workqueues. After this function returns, all freezable 6779 * workqueues will queue new works to their inactive_works list instead of 6780 * pool->worklist. 6781 * 6782 * CONTEXT: 6783 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6784 */ 6785 void freeze_workqueues_begin(void) 6786 { 6787 struct workqueue_struct *wq; 6788 6789 mutex_lock(&wq_pool_mutex); 6790 6791 WARN_ON_ONCE(workqueue_freezing); 6792 workqueue_freezing = true; 6793 6794 list_for_each_entry(wq, &workqueues, list) { 6795 mutex_lock(&wq->mutex); 6796 wq_adjust_max_active(wq); 6797 mutex_unlock(&wq->mutex); 6798 } 6799 6800 mutex_unlock(&wq_pool_mutex); 6801 } 6802 6803 /** 6804 * freeze_workqueues_busy - are freezable workqueues still busy? 6805 * 6806 * Check whether freezing is complete. This function must be called 6807 * between freeze_workqueues_begin() and thaw_workqueues(). 6808 * 6809 * CONTEXT: 6810 * Grabs and releases wq_pool_mutex. 6811 * 6812 * Return: 6813 * %true if some freezable workqueues are still busy. %false if freezing 6814 * is complete. 6815 */ 6816 bool freeze_workqueues_busy(void) 6817 { 6818 bool busy = false; 6819 struct workqueue_struct *wq; 6820 struct pool_workqueue *pwq; 6821 6822 mutex_lock(&wq_pool_mutex); 6823 6824 WARN_ON_ONCE(!workqueue_freezing); 6825 6826 list_for_each_entry(wq, &workqueues, list) { 6827 if (!(wq->flags & WQ_FREEZABLE)) 6828 continue; 6829 /* 6830 * nr_active is monotonically decreasing. It's safe 6831 * to peek without lock. 6832 */ 6833 rcu_read_lock(); 6834 for_each_pwq(pwq, wq) { 6835 WARN_ON_ONCE(pwq->nr_active < 0); 6836 if (pwq->nr_active) { 6837 busy = true; 6838 rcu_read_unlock(); 6839 goto out_unlock; 6840 } 6841 } 6842 rcu_read_unlock(); 6843 } 6844 out_unlock: 6845 mutex_unlock(&wq_pool_mutex); 6846 return busy; 6847 } 6848 6849 /** 6850 * thaw_workqueues - thaw workqueues 6851 * 6852 * Thaw workqueues. Normal queueing is restored and all collected 6853 * frozen works are transferred to their respective pool worklists. 6854 * 6855 * CONTEXT: 6856 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6857 */ 6858 void thaw_workqueues(void) 6859 { 6860 struct workqueue_struct *wq; 6861 6862 mutex_lock(&wq_pool_mutex); 6863 6864 if (!workqueue_freezing) 6865 goto out_unlock; 6866 6867 workqueue_freezing = false; 6868 6869 /* restore max_active and repopulate worklist */ 6870 list_for_each_entry(wq, &workqueues, list) { 6871 mutex_lock(&wq->mutex); 6872 wq_adjust_max_active(wq); 6873 mutex_unlock(&wq->mutex); 6874 } 6875 6876 out_unlock: 6877 mutex_unlock(&wq_pool_mutex); 6878 } 6879 #endif /* CONFIG_FREEZER */ 6880 6881 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 6882 { 6883 LIST_HEAD(ctxs); 6884 int ret = 0; 6885 struct workqueue_struct *wq; 6886 struct apply_wqattrs_ctx *ctx, *n; 6887 6888 lockdep_assert_held(&wq_pool_mutex); 6889 6890 list_for_each_entry(wq, &workqueues, list) { 6891 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING)) 6892 continue; 6893 6894 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 6895 if (IS_ERR(ctx)) { 6896 ret = PTR_ERR(ctx); 6897 break; 6898 } 6899 6900 list_add_tail(&ctx->list, &ctxs); 6901 } 6902 6903 list_for_each_entry_safe(ctx, n, &ctxs, list) { 6904 if (!ret) 6905 apply_wqattrs_commit(ctx); 6906 apply_wqattrs_cleanup(ctx); 6907 } 6908 6909 if (!ret) { 6910 mutex_lock(&wq_pool_attach_mutex); 6911 cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 6912 mutex_unlock(&wq_pool_attach_mutex); 6913 } 6914 return ret; 6915 } 6916 6917 /** 6918 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask 6919 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask 6920 * 6921 * This function can be called from cpuset code to provide a set of isolated 6922 * CPUs that should be excluded from wq_unbound_cpumask. 6923 */ 6924 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) 6925 { 6926 cpumask_var_t cpumask; 6927 int ret = 0; 6928 6929 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6930 return -ENOMEM; 6931 6932 mutex_lock(&wq_pool_mutex); 6933 6934 /* 6935 * If the operation fails, it will fall back to 6936 * wq_requested_unbound_cpumask which is initially set to 6937 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten 6938 * by any subsequent write to workqueue/cpumask sysfs file. 6939 */ 6940 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) 6941 cpumask_copy(cpumask, wq_requested_unbound_cpumask); 6942 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 6943 ret = workqueue_apply_unbound_cpumask(cpumask); 6944 6945 /* Save the current isolated cpumask & export it via sysfs */ 6946 if (!ret) 6947 cpumask_copy(wq_isolated_cpumask, exclude_cpumask); 6948 6949 mutex_unlock(&wq_pool_mutex); 6950 free_cpumask_var(cpumask); 6951 return ret; 6952 } 6953 6954 static int parse_affn_scope(const char *val) 6955 { 6956 int i; 6957 6958 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 6959 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 6960 return i; 6961 } 6962 return -EINVAL; 6963 } 6964 6965 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 6966 { 6967 struct workqueue_struct *wq; 6968 int affn, cpu; 6969 6970 affn = parse_affn_scope(val); 6971 if (affn < 0) 6972 return affn; 6973 if (affn == WQ_AFFN_DFL) 6974 return -EINVAL; 6975 6976 cpus_read_lock(); 6977 mutex_lock(&wq_pool_mutex); 6978 6979 wq_affn_dfl = affn; 6980 6981 list_for_each_entry(wq, &workqueues, list) { 6982 for_each_online_cpu(cpu) 6983 unbound_wq_update_pwq(wq, cpu); 6984 } 6985 6986 mutex_unlock(&wq_pool_mutex); 6987 cpus_read_unlock(); 6988 6989 return 0; 6990 } 6991 6992 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 6993 { 6994 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 6995 } 6996 6997 static const struct kernel_param_ops wq_affn_dfl_ops = { 6998 .set = wq_affn_dfl_set, 6999 .get = wq_affn_dfl_get, 7000 }; 7001 7002 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 7003 7004 #ifdef CONFIG_SYSFS 7005 /* 7006 * Workqueues with WQ_SYSFS flag set is visible to userland via 7007 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 7008 * following attributes. 7009 * 7010 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 7011 * max_active RW int : maximum number of in-flight work items 7012 * 7013 * Unbound workqueues have the following extra attributes. 7014 * 7015 * nice RW int : nice value of the workers 7016 * cpumask RW mask : bitmask of allowed CPUs for the workers 7017 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 7018 * affinity_strict RW bool : worker CPU affinity is strict 7019 */ 7020 struct wq_device { 7021 struct workqueue_struct *wq; 7022 struct device dev; 7023 }; 7024 7025 static struct workqueue_struct *dev_to_wq(struct device *dev) 7026 { 7027 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7028 7029 return wq_dev->wq; 7030 } 7031 7032 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 7033 char *buf) 7034 { 7035 struct workqueue_struct *wq = dev_to_wq(dev); 7036 7037 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 7038 } 7039 static DEVICE_ATTR_RO(per_cpu); 7040 7041 static ssize_t max_active_show(struct device *dev, 7042 struct device_attribute *attr, char *buf) 7043 { 7044 struct workqueue_struct *wq = dev_to_wq(dev); 7045 7046 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 7047 } 7048 7049 static ssize_t max_active_store(struct device *dev, 7050 struct device_attribute *attr, const char *buf, 7051 size_t count) 7052 { 7053 struct workqueue_struct *wq = dev_to_wq(dev); 7054 int val; 7055 7056 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 7057 return -EINVAL; 7058 7059 workqueue_set_max_active(wq, val); 7060 return count; 7061 } 7062 static DEVICE_ATTR_RW(max_active); 7063 7064 static struct attribute *wq_sysfs_attrs[] = { 7065 &dev_attr_per_cpu.attr, 7066 &dev_attr_max_active.attr, 7067 NULL, 7068 }; 7069 ATTRIBUTE_GROUPS(wq_sysfs); 7070 7071 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 7072 char *buf) 7073 { 7074 struct workqueue_struct *wq = dev_to_wq(dev); 7075 int written; 7076 7077 mutex_lock(&wq->mutex); 7078 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 7079 mutex_unlock(&wq->mutex); 7080 7081 return written; 7082 } 7083 7084 /* prepare workqueue_attrs for sysfs store operations */ 7085 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 7086 { 7087 struct workqueue_attrs *attrs; 7088 7089 lockdep_assert_held(&wq_pool_mutex); 7090 7091 attrs = alloc_workqueue_attrs(); 7092 if (!attrs) 7093 return NULL; 7094 7095 copy_workqueue_attrs(attrs, wq->unbound_attrs); 7096 return attrs; 7097 } 7098 7099 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 7100 const char *buf, size_t count) 7101 { 7102 struct workqueue_struct *wq = dev_to_wq(dev); 7103 struct workqueue_attrs *attrs; 7104 int ret = -ENOMEM; 7105 7106 apply_wqattrs_lock(); 7107 7108 attrs = wq_sysfs_prep_attrs(wq); 7109 if (!attrs) 7110 goto out_unlock; 7111 7112 if (sscanf(buf, "%d", &attrs->nice) == 1 && 7113 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 7114 ret = apply_workqueue_attrs_locked(wq, attrs); 7115 else 7116 ret = -EINVAL; 7117 7118 out_unlock: 7119 apply_wqattrs_unlock(); 7120 free_workqueue_attrs(attrs); 7121 return ret ?: count; 7122 } 7123 7124 static ssize_t wq_cpumask_show(struct device *dev, 7125 struct device_attribute *attr, char *buf) 7126 { 7127 struct workqueue_struct *wq = dev_to_wq(dev); 7128 int written; 7129 7130 mutex_lock(&wq->mutex); 7131 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 7132 cpumask_pr_args(wq->unbound_attrs->cpumask)); 7133 mutex_unlock(&wq->mutex); 7134 return written; 7135 } 7136 7137 static ssize_t wq_cpumask_store(struct device *dev, 7138 struct device_attribute *attr, 7139 const char *buf, size_t count) 7140 { 7141 struct workqueue_struct *wq = dev_to_wq(dev); 7142 struct workqueue_attrs *attrs; 7143 int ret = -ENOMEM; 7144 7145 apply_wqattrs_lock(); 7146 7147 attrs = wq_sysfs_prep_attrs(wq); 7148 if (!attrs) 7149 goto out_unlock; 7150 7151 ret = cpumask_parse(buf, attrs->cpumask); 7152 if (!ret) 7153 ret = apply_workqueue_attrs_locked(wq, attrs); 7154 7155 out_unlock: 7156 apply_wqattrs_unlock(); 7157 free_workqueue_attrs(attrs); 7158 return ret ?: count; 7159 } 7160 7161 static ssize_t wq_affn_scope_show(struct device *dev, 7162 struct device_attribute *attr, char *buf) 7163 { 7164 struct workqueue_struct *wq = dev_to_wq(dev); 7165 int written; 7166 7167 mutex_lock(&wq->mutex); 7168 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 7169 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 7170 wq_affn_names[WQ_AFFN_DFL], 7171 wq_affn_names[wq_affn_dfl]); 7172 else 7173 written = scnprintf(buf, PAGE_SIZE, "%s\n", 7174 wq_affn_names[wq->unbound_attrs->affn_scope]); 7175 mutex_unlock(&wq->mutex); 7176 7177 return written; 7178 } 7179 7180 static ssize_t wq_affn_scope_store(struct device *dev, 7181 struct device_attribute *attr, 7182 const char *buf, size_t count) 7183 { 7184 struct workqueue_struct *wq = dev_to_wq(dev); 7185 struct workqueue_attrs *attrs; 7186 int affn, ret = -ENOMEM; 7187 7188 affn = parse_affn_scope(buf); 7189 if (affn < 0) 7190 return affn; 7191 7192 apply_wqattrs_lock(); 7193 attrs = wq_sysfs_prep_attrs(wq); 7194 if (attrs) { 7195 attrs->affn_scope = affn; 7196 ret = apply_workqueue_attrs_locked(wq, attrs); 7197 } 7198 apply_wqattrs_unlock(); 7199 free_workqueue_attrs(attrs); 7200 return ret ?: count; 7201 } 7202 7203 static ssize_t wq_affinity_strict_show(struct device *dev, 7204 struct device_attribute *attr, char *buf) 7205 { 7206 struct workqueue_struct *wq = dev_to_wq(dev); 7207 7208 return scnprintf(buf, PAGE_SIZE, "%d\n", 7209 wq->unbound_attrs->affn_strict); 7210 } 7211 7212 static ssize_t wq_affinity_strict_store(struct device *dev, 7213 struct device_attribute *attr, 7214 const char *buf, size_t count) 7215 { 7216 struct workqueue_struct *wq = dev_to_wq(dev); 7217 struct workqueue_attrs *attrs; 7218 int v, ret = -ENOMEM; 7219 7220 if (sscanf(buf, "%d", &v) != 1) 7221 return -EINVAL; 7222 7223 apply_wqattrs_lock(); 7224 attrs = wq_sysfs_prep_attrs(wq); 7225 if (attrs) { 7226 attrs->affn_strict = (bool)v; 7227 ret = apply_workqueue_attrs_locked(wq, attrs); 7228 } 7229 apply_wqattrs_unlock(); 7230 free_workqueue_attrs(attrs); 7231 return ret ?: count; 7232 } 7233 7234 static struct device_attribute wq_sysfs_unbound_attrs[] = { 7235 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 7236 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 7237 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 7238 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 7239 __ATTR_NULL, 7240 }; 7241 7242 static const struct bus_type wq_subsys = { 7243 .name = "workqueue", 7244 .dev_groups = wq_sysfs_groups, 7245 }; 7246 7247 /** 7248 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 7249 * @cpumask: the cpumask to set 7250 * 7251 * The low-level workqueues cpumask is a global cpumask that limits 7252 * the affinity of all unbound workqueues. This function check the @cpumask 7253 * and apply it to all unbound workqueues and updates all pwqs of them. 7254 * 7255 * Return: 0 - Success 7256 * -EINVAL - Invalid @cpumask 7257 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 7258 */ 7259 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 7260 { 7261 int ret = -EINVAL; 7262 7263 /* 7264 * Not excluding isolated cpus on purpose. 7265 * If the user wishes to include them, we allow that. 7266 */ 7267 cpumask_and(cpumask, cpumask, cpu_possible_mask); 7268 if (!cpumask_empty(cpumask)) { 7269 ret = 0; 7270 apply_wqattrs_lock(); 7271 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 7272 ret = workqueue_apply_unbound_cpumask(cpumask); 7273 if (!ret) 7274 cpumask_copy(wq_requested_unbound_cpumask, cpumask); 7275 apply_wqattrs_unlock(); 7276 } 7277 7278 return ret; 7279 } 7280 7281 static ssize_t __wq_cpumask_show(struct device *dev, 7282 struct device_attribute *attr, char *buf, cpumask_var_t mask) 7283 { 7284 int written; 7285 7286 mutex_lock(&wq_pool_mutex); 7287 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); 7288 mutex_unlock(&wq_pool_mutex); 7289 7290 return written; 7291 } 7292 7293 static ssize_t cpumask_requested_show(struct device *dev, 7294 struct device_attribute *attr, char *buf) 7295 { 7296 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); 7297 } 7298 static DEVICE_ATTR_RO(cpumask_requested); 7299 7300 static ssize_t cpumask_isolated_show(struct device *dev, 7301 struct device_attribute *attr, char *buf) 7302 { 7303 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); 7304 } 7305 static DEVICE_ATTR_RO(cpumask_isolated); 7306 7307 static ssize_t cpumask_show(struct device *dev, 7308 struct device_attribute *attr, char *buf) 7309 { 7310 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); 7311 } 7312 7313 static ssize_t cpumask_store(struct device *dev, 7314 struct device_attribute *attr, const char *buf, size_t count) 7315 { 7316 cpumask_var_t cpumask; 7317 int ret; 7318 7319 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 7320 return -ENOMEM; 7321 7322 ret = cpumask_parse(buf, cpumask); 7323 if (!ret) 7324 ret = workqueue_set_unbound_cpumask(cpumask); 7325 7326 free_cpumask_var(cpumask); 7327 return ret ? ret : count; 7328 } 7329 static DEVICE_ATTR_RW(cpumask); 7330 7331 static struct attribute *wq_sysfs_cpumask_attrs[] = { 7332 &dev_attr_cpumask.attr, 7333 &dev_attr_cpumask_requested.attr, 7334 &dev_attr_cpumask_isolated.attr, 7335 NULL, 7336 }; 7337 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); 7338 7339 static int __init wq_sysfs_init(void) 7340 { 7341 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups); 7342 } 7343 core_initcall(wq_sysfs_init); 7344 7345 static void wq_device_release(struct device *dev) 7346 { 7347 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7348 7349 kfree(wq_dev); 7350 } 7351 7352 /** 7353 * workqueue_sysfs_register - make a workqueue visible in sysfs 7354 * @wq: the workqueue to register 7355 * 7356 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 7357 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 7358 * which is the preferred method. 7359 * 7360 * Workqueue user should use this function directly iff it wants to apply 7361 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 7362 * apply_workqueue_attrs() may race against userland updating the 7363 * attributes. 7364 * 7365 * Return: 0 on success, -errno on failure. 7366 */ 7367 int workqueue_sysfs_register(struct workqueue_struct *wq) 7368 { 7369 struct wq_device *wq_dev; 7370 int ret; 7371 7372 /* 7373 * Adjusting max_active breaks ordering guarantee. Disallow exposing 7374 * ordered workqueues. 7375 */ 7376 if (WARN_ON(wq->flags & __WQ_ORDERED)) 7377 return -EINVAL; 7378 7379 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 7380 if (!wq_dev) 7381 return -ENOMEM; 7382 7383 wq_dev->wq = wq; 7384 wq_dev->dev.bus = &wq_subsys; 7385 wq_dev->dev.release = wq_device_release; 7386 dev_set_name(&wq_dev->dev, "%s", wq->name); 7387 7388 /* 7389 * unbound_attrs are created separately. Suppress uevent until 7390 * everything is ready. 7391 */ 7392 dev_set_uevent_suppress(&wq_dev->dev, true); 7393 7394 ret = device_register(&wq_dev->dev); 7395 if (ret) { 7396 put_device(&wq_dev->dev); 7397 wq->wq_dev = NULL; 7398 return ret; 7399 } 7400 7401 if (wq->flags & WQ_UNBOUND) { 7402 struct device_attribute *attr; 7403 7404 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 7405 ret = device_create_file(&wq_dev->dev, attr); 7406 if (ret) { 7407 device_unregister(&wq_dev->dev); 7408 wq->wq_dev = NULL; 7409 return ret; 7410 } 7411 } 7412 } 7413 7414 dev_set_uevent_suppress(&wq_dev->dev, false); 7415 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 7416 return 0; 7417 } 7418 7419 /** 7420 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 7421 * @wq: the workqueue to unregister 7422 * 7423 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 7424 */ 7425 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 7426 { 7427 struct wq_device *wq_dev = wq->wq_dev; 7428 7429 if (!wq->wq_dev) 7430 return; 7431 7432 wq->wq_dev = NULL; 7433 device_unregister(&wq_dev->dev); 7434 } 7435 #else /* CONFIG_SYSFS */ 7436 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 7437 #endif /* CONFIG_SYSFS */ 7438 7439 /* 7440 * Workqueue watchdog. 7441 * 7442 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 7443 * flush dependency, a concurrency managed work item which stays RUNNING 7444 * indefinitely. Workqueue stalls can be very difficult to debug as the 7445 * usual warning mechanisms don't trigger and internal workqueue state is 7446 * largely opaque. 7447 * 7448 * Workqueue watchdog monitors all worker pools periodically and dumps 7449 * state if some pools failed to make forward progress for a while where 7450 * forward progress is defined as the first item on ->worklist changing. 7451 * 7452 * This mechanism is controlled through the kernel parameter 7453 * "workqueue.watchdog_thresh" which can be updated at runtime through the 7454 * corresponding sysfs parameter file. 7455 */ 7456 #ifdef CONFIG_WQ_WATCHDOG 7457 7458 static unsigned long wq_watchdog_thresh = 30; 7459 static struct timer_list wq_watchdog_timer; 7460 7461 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 7462 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 7463 7464 static unsigned int wq_panic_on_stall; 7465 module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644); 7466 7467 /* 7468 * Show workers that might prevent the processing of pending work items. 7469 * The only candidates are CPU-bound workers in the running state. 7470 * Pending work items should be handled by another idle worker 7471 * in all other situations. 7472 */ 7473 static void show_cpu_pool_hog(struct worker_pool *pool) 7474 { 7475 struct worker *worker; 7476 unsigned long irq_flags; 7477 int bkt; 7478 7479 raw_spin_lock_irqsave(&pool->lock, irq_flags); 7480 7481 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 7482 if (task_is_running(worker->task)) { 7483 /* 7484 * Defer printing to avoid deadlocks in console 7485 * drivers that queue work while holding locks 7486 * also taken in their write paths. 7487 */ 7488 printk_deferred_enter(); 7489 7490 pr_info("pool %d:\n", pool->id); 7491 sched_show_task(worker->task); 7492 7493 printk_deferred_exit(); 7494 } 7495 } 7496 7497 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 7498 } 7499 7500 static void show_cpu_pools_hogs(void) 7501 { 7502 struct worker_pool *pool; 7503 int pi; 7504 7505 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 7506 7507 rcu_read_lock(); 7508 7509 for_each_pool(pool, pi) { 7510 if (pool->cpu_stall) 7511 show_cpu_pool_hog(pool); 7512 7513 } 7514 7515 rcu_read_unlock(); 7516 } 7517 7518 static void panic_on_wq_watchdog(void) 7519 { 7520 static unsigned int wq_stall; 7521 7522 if (wq_panic_on_stall) { 7523 wq_stall++; 7524 BUG_ON(wq_stall >= wq_panic_on_stall); 7525 } 7526 } 7527 7528 static void wq_watchdog_reset_touched(void) 7529 { 7530 int cpu; 7531 7532 wq_watchdog_touched = jiffies; 7533 for_each_possible_cpu(cpu) 7534 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 7535 } 7536 7537 static void wq_watchdog_timer_fn(struct timer_list *unused) 7538 { 7539 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7540 bool lockup_detected = false; 7541 bool cpu_pool_stall = false; 7542 unsigned long now = jiffies; 7543 struct worker_pool *pool; 7544 int pi; 7545 7546 if (!thresh) 7547 return; 7548 7549 rcu_read_lock(); 7550 7551 for_each_pool(pool, pi) { 7552 unsigned long pool_ts, touched, ts; 7553 7554 pool->cpu_stall = false; 7555 if (list_empty(&pool->worklist)) 7556 continue; 7557 7558 /* 7559 * If a virtual machine is stopped by the host it can look to 7560 * the watchdog like a stall. 7561 */ 7562 kvm_check_and_clear_guest_paused(); 7563 7564 /* get the latest of pool and touched timestamps */ 7565 if (pool->cpu >= 0) 7566 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 7567 else 7568 touched = READ_ONCE(wq_watchdog_touched); 7569 pool_ts = READ_ONCE(pool->watchdog_ts); 7570 7571 if (time_after(pool_ts, touched)) 7572 ts = pool_ts; 7573 else 7574 ts = touched; 7575 7576 /* did we stall? */ 7577 if (time_after(now, ts + thresh)) { 7578 lockup_detected = true; 7579 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) { 7580 pool->cpu_stall = true; 7581 cpu_pool_stall = true; 7582 } 7583 pr_emerg("BUG: workqueue lockup - pool"); 7584 pr_cont_pool_info(pool); 7585 pr_cont(" stuck for %us!\n", 7586 jiffies_to_msecs(now - pool_ts) / 1000); 7587 } 7588 7589 7590 } 7591 7592 rcu_read_unlock(); 7593 7594 if (lockup_detected) 7595 show_all_workqueues(); 7596 7597 if (cpu_pool_stall) 7598 show_cpu_pools_hogs(); 7599 7600 if (lockup_detected) 7601 panic_on_wq_watchdog(); 7602 7603 wq_watchdog_reset_touched(); 7604 mod_timer(&wq_watchdog_timer, jiffies + thresh); 7605 } 7606 7607 notrace void wq_watchdog_touch(int cpu) 7608 { 7609 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7610 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched); 7611 unsigned long now = jiffies; 7612 7613 if (cpu >= 0) 7614 per_cpu(wq_watchdog_touched_cpu, cpu) = now; 7615 else 7616 WARN_ONCE(1, "%s should be called with valid CPU", __func__); 7617 7618 /* Don't unnecessarily store to global cacheline */ 7619 if (time_after(now, touch_ts + thresh / 4)) 7620 WRITE_ONCE(wq_watchdog_touched, jiffies); 7621 } 7622 7623 static void wq_watchdog_set_thresh(unsigned long thresh) 7624 { 7625 wq_watchdog_thresh = 0; 7626 timer_delete_sync(&wq_watchdog_timer); 7627 7628 if (thresh) { 7629 wq_watchdog_thresh = thresh; 7630 wq_watchdog_reset_touched(); 7631 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 7632 } 7633 } 7634 7635 static int wq_watchdog_param_set_thresh(const char *val, 7636 const struct kernel_param *kp) 7637 { 7638 unsigned long thresh; 7639 int ret; 7640 7641 ret = kstrtoul(val, 0, &thresh); 7642 if (ret) 7643 return ret; 7644 7645 if (system_wq) 7646 wq_watchdog_set_thresh(thresh); 7647 else 7648 wq_watchdog_thresh = thresh; 7649 7650 return 0; 7651 } 7652 7653 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 7654 .set = wq_watchdog_param_set_thresh, 7655 .get = param_get_ulong, 7656 }; 7657 7658 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 7659 0644); 7660 7661 static void wq_watchdog_init(void) 7662 { 7663 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 7664 wq_watchdog_set_thresh(wq_watchdog_thresh); 7665 } 7666 7667 #else /* CONFIG_WQ_WATCHDOG */ 7668 7669 static inline void wq_watchdog_init(void) { } 7670 7671 #endif /* CONFIG_WQ_WATCHDOG */ 7672 7673 static void bh_pool_kick_normal(struct irq_work *irq_work) 7674 { 7675 raise_softirq_irqoff(TASKLET_SOFTIRQ); 7676 } 7677 7678 static void bh_pool_kick_highpri(struct irq_work *irq_work) 7679 { 7680 raise_softirq_irqoff(HI_SOFTIRQ); 7681 } 7682 7683 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) 7684 { 7685 if (!cpumask_intersects(wq_unbound_cpumask, mask)) { 7686 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", 7687 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); 7688 return; 7689 } 7690 7691 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); 7692 } 7693 7694 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice) 7695 { 7696 BUG_ON(init_worker_pool(pool)); 7697 pool->cpu = cpu; 7698 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 7699 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 7700 pool->attrs->nice = nice; 7701 pool->attrs->affn_strict = true; 7702 pool->node = cpu_to_node(cpu); 7703 7704 /* alloc pool ID */ 7705 mutex_lock(&wq_pool_mutex); 7706 BUG_ON(worker_pool_assign_id(pool)); 7707 mutex_unlock(&wq_pool_mutex); 7708 } 7709 7710 /** 7711 * workqueue_init_early - early init for workqueue subsystem 7712 * 7713 * This is the first step of three-staged workqueue subsystem initialization and 7714 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 7715 * up. It sets up all the data structures and system workqueues and allows early 7716 * boot code to create workqueues and queue/cancel work items. Actual work item 7717 * execution starts only after kthreads can be created and scheduled right 7718 * before early initcalls. 7719 */ 7720 void __init workqueue_init_early(void) 7721 { 7722 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 7723 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 7724 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal, 7725 bh_pool_kick_highpri }; 7726 int i, cpu; 7727 7728 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 7729 7730 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL)); 7731 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 7732 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); 7733 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); 7734 7735 cpumask_copy(wq_online_cpumask, cpu_online_mask); 7736 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 7737 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); 7738 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); 7739 if (!cpumask_empty(&wq_cmdline_cpumask)) 7740 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); 7741 7742 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); 7743 cpumask_andnot(wq_isolated_cpumask, cpu_possible_mask, 7744 housekeeping_cpumask(HK_TYPE_DOMAIN)); 7745 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 7746 7747 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs(); 7748 BUG_ON(!unbound_wq_update_pwq_attrs_buf); 7749 7750 /* 7751 * If nohz_full is enabled, set power efficient workqueue as unbound. 7752 * This allows workqueue items to be moved to HK CPUs. 7753 */ 7754 if (housekeeping_enabled(HK_TYPE_TICK)) 7755 wq_power_efficient = true; 7756 7757 /* initialize WQ_AFFN_SYSTEM pods */ 7758 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 7759 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 7760 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 7761 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 7762 7763 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 7764 7765 pt->nr_pods = 1; 7766 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 7767 pt->pod_node[0] = NUMA_NO_NODE; 7768 pt->cpu_pod[0] = 0; 7769 7770 /* initialize BH and CPU pools */ 7771 for_each_possible_cpu(cpu) { 7772 struct worker_pool *pool; 7773 7774 i = 0; 7775 for_each_bh_worker_pool(pool, cpu) { 7776 init_cpu_worker_pool(pool, cpu, std_nice[i]); 7777 pool->flags |= POOL_BH; 7778 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]); 7779 i++; 7780 } 7781 7782 i = 0; 7783 for_each_cpu_worker_pool(pool, cpu) 7784 init_cpu_worker_pool(pool, cpu, std_nice[i++]); 7785 } 7786 7787 /* create default unbound and ordered wq attrs */ 7788 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 7789 struct workqueue_attrs *attrs; 7790 7791 BUG_ON(!(attrs = alloc_workqueue_attrs())); 7792 attrs->nice = std_nice[i]; 7793 unbound_std_wq_attrs[i] = attrs; 7794 7795 /* 7796 * An ordered wq should have only one pwq as ordering is 7797 * guaranteed by max_active which is enforced by pwqs. 7798 */ 7799 BUG_ON(!(attrs = alloc_workqueue_attrs())); 7800 attrs->nice = std_nice[i]; 7801 attrs->ordered = true; 7802 ordered_wq_attrs[i] = attrs; 7803 } 7804 7805 system_wq = alloc_workqueue("events", 0, 0); 7806 system_percpu_wq = alloc_workqueue("events", 0, 0); 7807 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 7808 system_long_wq = alloc_workqueue("events_long", 0, 0); 7809 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE); 7810 system_dfl_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE); 7811 system_freezable_wq = alloc_workqueue("events_freezable", 7812 WQ_FREEZABLE, 0); 7813 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 7814 WQ_POWER_EFFICIENT, 0); 7815 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", 7816 WQ_FREEZABLE | WQ_POWER_EFFICIENT, 7817 0); 7818 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0); 7819 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri", 7820 WQ_BH | WQ_HIGHPRI, 0); 7821 BUG_ON(!system_wq || !system_percpu_wq|| !system_highpri_wq || !system_long_wq || 7822 !system_unbound_wq || !system_freezable_wq || !system_dfl_wq || 7823 !system_power_efficient_wq || 7824 !system_freezable_power_efficient_wq || 7825 !system_bh_wq || !system_bh_highpri_wq); 7826 } 7827 7828 static void __init wq_cpu_intensive_thresh_init(void) 7829 { 7830 unsigned long thresh; 7831 unsigned long bogo; 7832 7833 pwq_release_worker = kthread_run_worker(0, "pool_workqueue_release"); 7834 BUG_ON(IS_ERR(pwq_release_worker)); 7835 7836 /* if the user set it to a specific value, keep it */ 7837 if (wq_cpu_intensive_thresh_us != ULONG_MAX) 7838 return; 7839 7840 /* 7841 * The default of 10ms is derived from the fact that most modern (as of 7842 * 2023) processors can do a lot in 10ms and that it's just below what 7843 * most consider human-perceivable. However, the kernel also runs on a 7844 * lot slower CPUs including microcontrollers where the threshold is way 7845 * too low. 7846 * 7847 * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 7848 * This is by no means accurate but it doesn't have to be. The mechanism 7849 * is still useful even when the threshold is fully scaled up. Also, as 7850 * the reports would usually be applicable to everyone, some machines 7851 * operating on longer thresholds won't significantly diminish their 7852 * usefulness. 7853 */ 7854 thresh = 10 * USEC_PER_MSEC; 7855 7856 /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 7857 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 7858 if (bogo < 4000) 7859 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 7860 7861 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 7862 loops_per_jiffy, bogo, thresh); 7863 7864 wq_cpu_intensive_thresh_us = thresh; 7865 } 7866 7867 /** 7868 * workqueue_init - bring workqueue subsystem fully online 7869 * 7870 * This is the second step of three-staged workqueue subsystem initialization 7871 * and invoked as soon as kthreads can be created and scheduled. Workqueues have 7872 * been created and work items queued on them, but there are no kworkers 7873 * executing the work items yet. Populate the worker pools with the initial 7874 * workers and enable future kworker creations. 7875 */ 7876 void __init workqueue_init(void) 7877 { 7878 struct workqueue_struct *wq; 7879 struct worker_pool *pool; 7880 int cpu, bkt; 7881 7882 wq_cpu_intensive_thresh_init(); 7883 7884 mutex_lock(&wq_pool_mutex); 7885 7886 /* 7887 * Per-cpu pools created earlier could be missing node hint. Fix them 7888 * up. Also, create a rescuer for workqueues that requested it. 7889 */ 7890 for_each_possible_cpu(cpu) { 7891 for_each_bh_worker_pool(pool, cpu) 7892 pool->node = cpu_to_node(cpu); 7893 for_each_cpu_worker_pool(pool, cpu) 7894 pool->node = cpu_to_node(cpu); 7895 } 7896 7897 list_for_each_entry(wq, &workqueues, list) { 7898 WARN(init_rescuer(wq), 7899 "workqueue: failed to create early rescuer for %s", 7900 wq->name); 7901 } 7902 7903 mutex_unlock(&wq_pool_mutex); 7904 7905 /* 7906 * Create the initial workers. A BH pool has one pseudo worker that 7907 * represents the shared BH execution context and thus doesn't get 7908 * affected by hotplug events. Create the BH pseudo workers for all 7909 * possible CPUs here. 7910 */ 7911 for_each_possible_cpu(cpu) 7912 for_each_bh_worker_pool(pool, cpu) 7913 BUG_ON(!create_worker(pool)); 7914 7915 for_each_online_cpu(cpu) { 7916 for_each_cpu_worker_pool(pool, cpu) { 7917 pool->flags &= ~POOL_DISASSOCIATED; 7918 BUG_ON(!create_worker(pool)); 7919 } 7920 } 7921 7922 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 7923 BUG_ON(!create_worker(pool)); 7924 7925 wq_online = true; 7926 wq_watchdog_init(); 7927 } 7928 7929 /* 7930 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 7931 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 7932 * and consecutive pod ID. The rest of @pt is initialized accordingly. 7933 */ 7934 static void __init init_pod_type(struct wq_pod_type *pt, 7935 bool (*cpus_share_pod)(int, int)) 7936 { 7937 int cur, pre, cpu, pod; 7938 7939 pt->nr_pods = 0; 7940 7941 /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 7942 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 7943 BUG_ON(!pt->cpu_pod); 7944 7945 for_each_possible_cpu(cur) { 7946 for_each_possible_cpu(pre) { 7947 if (pre >= cur) { 7948 pt->cpu_pod[cur] = pt->nr_pods++; 7949 break; 7950 } 7951 if (cpus_share_pod(cur, pre)) { 7952 pt->cpu_pod[cur] = pt->cpu_pod[pre]; 7953 break; 7954 } 7955 } 7956 } 7957 7958 /* init the rest to match @pt->cpu_pod[] */ 7959 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 7960 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 7961 BUG_ON(!pt->pod_cpus || !pt->pod_node); 7962 7963 for (pod = 0; pod < pt->nr_pods; pod++) 7964 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 7965 7966 for_each_possible_cpu(cpu) { 7967 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 7968 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 7969 } 7970 } 7971 7972 static bool __init cpus_dont_share(int cpu0, int cpu1) 7973 { 7974 return false; 7975 } 7976 7977 static bool __init cpus_share_smt(int cpu0, int cpu1) 7978 { 7979 #ifdef CONFIG_SCHED_SMT 7980 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 7981 #else 7982 return false; 7983 #endif 7984 } 7985 7986 static bool __init cpus_share_numa(int cpu0, int cpu1) 7987 { 7988 return cpu_to_node(cpu0) == cpu_to_node(cpu1); 7989 } 7990 7991 /** 7992 * workqueue_init_topology - initialize CPU pods for unbound workqueues 7993 * 7994 * This is the third step of three-staged workqueue subsystem initialization and 7995 * invoked after SMP and topology information are fully initialized. It 7996 * initializes the unbound CPU pods accordingly. 7997 */ 7998 void __init workqueue_init_topology(void) 7999 { 8000 struct workqueue_struct *wq; 8001 int cpu; 8002 8003 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 8004 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 8005 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 8006 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 8007 8008 wq_topo_initialized = true; 8009 8010 mutex_lock(&wq_pool_mutex); 8011 8012 /* 8013 * Workqueues allocated earlier would have all CPUs sharing the default 8014 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue 8015 * and CPU combinations to apply per-pod sharing. 8016 */ 8017 list_for_each_entry(wq, &workqueues, list) { 8018 for_each_online_cpu(cpu) 8019 unbound_wq_update_pwq(wq, cpu); 8020 if (wq->flags & WQ_UNBOUND) { 8021 mutex_lock(&wq->mutex); 8022 wq_update_node_max_active(wq, -1); 8023 mutex_unlock(&wq->mutex); 8024 } 8025 } 8026 8027 mutex_unlock(&wq_pool_mutex); 8028 } 8029 8030 void __warn_flushing_systemwide_wq(void) 8031 { 8032 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 8033 dump_stack(); 8034 } 8035 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 8036 8037 static int __init workqueue_unbound_cpus_setup(char *str) 8038 { 8039 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 8040 cpumask_clear(&wq_cmdline_cpumask); 8041 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 8042 } 8043 8044 return 1; 8045 } 8046 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 8047