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_highpri_wq __ro_after_init; 509 EXPORT_SYMBOL_GPL(system_highpri_wq); 510 struct workqueue_struct *system_long_wq __ro_after_init; 511 EXPORT_SYMBOL_GPL(system_long_wq); 512 struct workqueue_struct *system_unbound_wq __ro_after_init; 513 EXPORT_SYMBOL_GPL(system_unbound_wq); 514 struct workqueue_struct *system_freezable_wq __ro_after_init; 515 EXPORT_SYMBOL_GPL(system_freezable_wq); 516 struct workqueue_struct *system_power_efficient_wq __ro_after_init; 517 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 518 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; 519 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 520 struct workqueue_struct *system_bh_wq; 521 EXPORT_SYMBOL_GPL(system_bh_wq); 522 struct workqueue_struct *system_bh_highpri_wq; 523 EXPORT_SYMBOL_GPL(system_bh_highpri_wq); 524 525 static int worker_thread(void *__worker); 526 static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 527 static void show_pwq(struct pool_workqueue *pwq); 528 static void show_one_worker_pool(struct worker_pool *pool); 529 530 #define CREATE_TRACE_POINTS 531 #include <trace/events/workqueue.h> 532 533 #define assert_rcu_or_pool_mutex() \ 534 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ 535 !lockdep_is_held(&wq_pool_mutex), \ 536 "RCU or wq_pool_mutex should be held") 537 538 #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 539 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ 540 !lockdep_is_held(&wq->mutex) && \ 541 !lockdep_is_held(&wq_pool_mutex), \ 542 "RCU, wq->mutex or wq_pool_mutex should be held") 543 544 #define for_each_bh_worker_pool(pool, cpu) \ 545 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \ 546 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 547 (pool)++) 548 549 #define for_each_cpu_worker_pool(pool, cpu) \ 550 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 551 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 552 (pool)++) 553 554 /** 555 * for_each_pool - iterate through all worker_pools in the system 556 * @pool: iteration cursor 557 * @pi: integer used for iteration 558 * 559 * This must be called either with wq_pool_mutex held or RCU read 560 * locked. If the pool needs to be used beyond the locking in effect, the 561 * caller is responsible for guaranteeing that the pool stays online. 562 * 563 * The if/else clause exists only for the lockdep assertion and can be 564 * ignored. 565 */ 566 #define for_each_pool(pool, pi) \ 567 idr_for_each_entry(&worker_pool_idr, pool, pi) \ 568 if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 569 else 570 571 /** 572 * for_each_pool_worker - iterate through all workers of a worker_pool 573 * @worker: iteration cursor 574 * @pool: worker_pool to iterate workers of 575 * 576 * This must be called with wq_pool_attach_mutex. 577 * 578 * The if/else clause exists only for the lockdep assertion and can be 579 * ignored. 580 */ 581 #define for_each_pool_worker(worker, pool) \ 582 list_for_each_entry((worker), &(pool)->workers, node) \ 583 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 584 else 585 586 /** 587 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 588 * @pwq: iteration cursor 589 * @wq: the target workqueue 590 * 591 * This must be called either with wq->mutex held or RCU read locked. 592 * If the pwq needs to be used beyond the locking in effect, the caller is 593 * responsible for guaranteeing that the pwq stays online. 594 * 595 * The if/else clause exists only for the lockdep assertion and can be 596 * ignored. 597 */ 598 #define for_each_pwq(pwq, wq) \ 599 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 600 lockdep_is_held(&(wq->mutex))) 601 602 #ifdef CONFIG_DEBUG_OBJECTS_WORK 603 604 static const struct debug_obj_descr work_debug_descr; 605 606 static void *work_debug_hint(void *addr) 607 { 608 return ((struct work_struct *) addr)->func; 609 } 610 611 static bool work_is_static_object(void *addr) 612 { 613 struct work_struct *work = addr; 614 615 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 616 } 617 618 /* 619 * fixup_init is called when: 620 * - an active object is initialized 621 */ 622 static bool work_fixup_init(void *addr, enum debug_obj_state state) 623 { 624 struct work_struct *work = addr; 625 626 switch (state) { 627 case ODEBUG_STATE_ACTIVE: 628 cancel_work_sync(work); 629 debug_object_init(work, &work_debug_descr); 630 return true; 631 default: 632 return false; 633 } 634 } 635 636 /* 637 * fixup_free is called when: 638 * - an active object is freed 639 */ 640 static bool work_fixup_free(void *addr, enum debug_obj_state state) 641 { 642 struct work_struct *work = addr; 643 644 switch (state) { 645 case ODEBUG_STATE_ACTIVE: 646 cancel_work_sync(work); 647 debug_object_free(work, &work_debug_descr); 648 return true; 649 default: 650 return false; 651 } 652 } 653 654 static const struct debug_obj_descr work_debug_descr = { 655 .name = "work_struct", 656 .debug_hint = work_debug_hint, 657 .is_static_object = work_is_static_object, 658 .fixup_init = work_fixup_init, 659 .fixup_free = work_fixup_free, 660 }; 661 662 static inline void debug_work_activate(struct work_struct *work) 663 { 664 debug_object_activate(work, &work_debug_descr); 665 } 666 667 static inline void debug_work_deactivate(struct work_struct *work) 668 { 669 debug_object_deactivate(work, &work_debug_descr); 670 } 671 672 void __init_work(struct work_struct *work, int onstack) 673 { 674 if (onstack) 675 debug_object_init_on_stack(work, &work_debug_descr); 676 else 677 debug_object_init(work, &work_debug_descr); 678 } 679 EXPORT_SYMBOL_GPL(__init_work); 680 681 void destroy_work_on_stack(struct work_struct *work) 682 { 683 debug_object_free(work, &work_debug_descr); 684 } 685 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 686 687 void destroy_delayed_work_on_stack(struct delayed_work *work) 688 { 689 destroy_timer_on_stack(&work->timer); 690 debug_object_free(&work->work, &work_debug_descr); 691 } 692 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 693 694 #else 695 static inline void debug_work_activate(struct work_struct *work) { } 696 static inline void debug_work_deactivate(struct work_struct *work) { } 697 #endif 698 699 /** 700 * worker_pool_assign_id - allocate ID and assign it to @pool 701 * @pool: the pool pointer of interest 702 * 703 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 704 * successfully, -errno on failure. 705 */ 706 static int worker_pool_assign_id(struct worker_pool *pool) 707 { 708 int ret; 709 710 lockdep_assert_held(&wq_pool_mutex); 711 712 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 713 GFP_KERNEL); 714 if (ret >= 0) { 715 pool->id = ret; 716 return 0; 717 } 718 return ret; 719 } 720 721 static struct pool_workqueue __rcu ** 722 unbound_pwq_slot(struct workqueue_struct *wq, int cpu) 723 { 724 if (cpu >= 0) 725 return per_cpu_ptr(wq->cpu_pwq, cpu); 726 else 727 return &wq->dfl_pwq; 728 } 729 730 /* @cpu < 0 for dfl_pwq */ 731 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu) 732 { 733 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu), 734 lockdep_is_held(&wq_pool_mutex) || 735 lockdep_is_held(&wq->mutex)); 736 } 737 738 /** 739 * unbound_effective_cpumask - effective cpumask of an unbound workqueue 740 * @wq: workqueue of interest 741 * 742 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which 743 * is masked with wq_unbound_cpumask to determine the effective cpumask. The 744 * default pwq is always mapped to the pool with the current effective cpumask. 745 */ 746 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq) 747 { 748 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask; 749 } 750 751 static unsigned int work_color_to_flags(int color) 752 { 753 return color << WORK_STRUCT_COLOR_SHIFT; 754 } 755 756 static int get_work_color(unsigned long work_data) 757 { 758 return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 759 ((1 << WORK_STRUCT_COLOR_BITS) - 1); 760 } 761 762 static int work_next_color(int color) 763 { 764 return (color + 1) % WORK_NR_COLORS; 765 } 766 767 static unsigned long pool_offq_flags(struct worker_pool *pool) 768 { 769 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0; 770 } 771 772 /* 773 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 774 * contain the pointer to the queued pwq. Once execution starts, the flag 775 * is cleared and the high bits contain OFFQ flags and pool ID. 776 * 777 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling() 778 * can be used to set the pwq, pool or clear work->data. These functions should 779 * only be called while the work is owned - ie. while the PENDING bit is set. 780 * 781 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 782 * corresponding to a work. Pool is available once the work has been 783 * queued anywhere after initialization until it is sync canceled. pwq is 784 * available only while the work item is queued. 785 */ 786 static inline void set_work_data(struct work_struct *work, unsigned long data) 787 { 788 WARN_ON_ONCE(!work_pending(work)); 789 atomic_long_set(&work->data, data | work_static(work)); 790 } 791 792 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 793 unsigned long flags) 794 { 795 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING | 796 WORK_STRUCT_PWQ | flags); 797 } 798 799 static void set_work_pool_and_keep_pending(struct work_struct *work, 800 int pool_id, unsigned long flags) 801 { 802 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | 803 WORK_STRUCT_PENDING | flags); 804 } 805 806 static void set_work_pool_and_clear_pending(struct work_struct *work, 807 int pool_id, unsigned long flags) 808 { 809 /* 810 * The following wmb is paired with the implied mb in 811 * test_and_set_bit(PENDING) and ensures all updates to @work made 812 * here are visible to and precede any updates by the next PENDING 813 * owner. 814 */ 815 smp_wmb(); 816 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | 817 flags); 818 /* 819 * The following mb guarantees that previous clear of a PENDING bit 820 * will not be reordered with any speculative LOADS or STORES from 821 * work->current_func, which is executed afterwards. This possible 822 * reordering can lead to a missed execution on attempt to queue 823 * the same @work. E.g. consider this case: 824 * 825 * CPU#0 CPU#1 826 * ---------------------------- -------------------------------- 827 * 828 * 1 STORE event_indicated 829 * 2 queue_work_on() { 830 * 3 test_and_set_bit(PENDING) 831 * 4 } set_..._and_clear_pending() { 832 * 5 set_work_data() # clear bit 833 * 6 smp_mb() 834 * 7 work->current_func() { 835 * 8 LOAD event_indicated 836 * } 837 * 838 * Without an explicit full barrier speculative LOAD on line 8 can 839 * be executed before CPU#0 does STORE on line 1. If that happens, 840 * CPU#0 observes the PENDING bit is still set and new execution of 841 * a @work is not queued in a hope, that CPU#1 will eventually 842 * finish the queued @work. Meanwhile CPU#1 does not see 843 * event_indicated is set, because speculative LOAD was executed 844 * before actual STORE. 845 */ 846 smp_mb(); 847 } 848 849 static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 850 { 851 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK); 852 } 853 854 static struct pool_workqueue *get_work_pwq(struct work_struct *work) 855 { 856 unsigned long data = atomic_long_read(&work->data); 857 858 if (data & WORK_STRUCT_PWQ) 859 return work_struct_pwq(data); 860 else 861 return NULL; 862 } 863 864 /** 865 * get_work_pool - return the worker_pool a given work was associated with 866 * @work: the work item of interest 867 * 868 * Pools are created and destroyed under wq_pool_mutex, and allows read 869 * access under RCU read lock. As such, this function should be 870 * called under wq_pool_mutex or inside of a rcu_read_lock() region. 871 * 872 * All fields of the returned pool are accessible as long as the above 873 * mentioned locking is in effect. If the returned pool needs to be used 874 * beyond the critical section, the caller is responsible for ensuring the 875 * returned pool is and stays online. 876 * 877 * Return: The worker_pool @work was last associated with. %NULL if none. 878 */ 879 static struct worker_pool *get_work_pool(struct work_struct *work) 880 { 881 unsigned long data = atomic_long_read(&work->data); 882 int pool_id; 883 884 assert_rcu_or_pool_mutex(); 885 886 if (data & WORK_STRUCT_PWQ) 887 return work_struct_pwq(data)->pool; 888 889 pool_id = data >> WORK_OFFQ_POOL_SHIFT; 890 if (pool_id == WORK_OFFQ_POOL_NONE) 891 return NULL; 892 893 return idr_find(&worker_pool_idr, pool_id); 894 } 895 896 static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits) 897 { 898 return (v >> shift) & ((1U << bits) - 1); 899 } 900 901 static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data) 902 { 903 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); 904 905 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT, 906 WORK_OFFQ_POOL_BITS); 907 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT, 908 WORK_OFFQ_DISABLE_BITS); 909 offqd->flags = data & WORK_OFFQ_FLAG_MASK; 910 } 911 912 static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd) 913 { 914 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) | 915 ((unsigned long)offqd->flags); 916 } 917 918 /* 919 * Policy functions. These define the policies on how the global worker 920 * pools are managed. Unless noted otherwise, these functions assume that 921 * they're being called with pool->lock held. 922 */ 923 924 /* 925 * Need to wake up a worker? Called from anything but currently 926 * running workers. 927 * 928 * Note that, because unbound workers never contribute to nr_running, this 929 * function will always return %true for unbound pools as long as the 930 * worklist isn't empty. 931 */ 932 static bool need_more_worker(struct worker_pool *pool) 933 { 934 return !list_empty(&pool->worklist) && !pool->nr_running; 935 } 936 937 /* Can I start working? Called from busy but !running workers. */ 938 static bool may_start_working(struct worker_pool *pool) 939 { 940 return pool->nr_idle; 941 } 942 943 /* Do I need to keep working? Called from currently running workers. */ 944 static bool keep_working(struct worker_pool *pool) 945 { 946 return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 947 } 948 949 /* Do we need a new worker? Called from manager. */ 950 static bool need_to_create_worker(struct worker_pool *pool) 951 { 952 return need_more_worker(pool) && !may_start_working(pool); 953 } 954 955 /* Do we have too many workers and should some go away? */ 956 static bool too_many_workers(struct worker_pool *pool) 957 { 958 bool managing = pool->flags & POOL_MANAGER_ACTIVE; 959 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 960 int nr_busy = pool->nr_workers - nr_idle; 961 962 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 963 } 964 965 /** 966 * worker_set_flags - set worker flags and adjust nr_running accordingly 967 * @worker: self 968 * @flags: flags to set 969 * 970 * Set @flags in @worker->flags and adjust nr_running accordingly. 971 */ 972 static inline void worker_set_flags(struct worker *worker, unsigned int flags) 973 { 974 struct worker_pool *pool = worker->pool; 975 976 lockdep_assert_held(&pool->lock); 977 978 /* If transitioning into NOT_RUNNING, adjust nr_running. */ 979 if ((flags & WORKER_NOT_RUNNING) && 980 !(worker->flags & WORKER_NOT_RUNNING)) { 981 pool->nr_running--; 982 } 983 984 worker->flags |= flags; 985 } 986 987 /** 988 * worker_clr_flags - clear worker flags and adjust nr_running accordingly 989 * @worker: self 990 * @flags: flags to clear 991 * 992 * Clear @flags in @worker->flags and adjust nr_running accordingly. 993 */ 994 static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 995 { 996 struct worker_pool *pool = worker->pool; 997 unsigned int oflags = worker->flags; 998 999 lockdep_assert_held(&pool->lock); 1000 1001 worker->flags &= ~flags; 1002 1003 /* 1004 * If transitioning out of NOT_RUNNING, increment nr_running. Note 1005 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 1006 * of multiple flags, not a single flag. 1007 */ 1008 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1009 if (!(worker->flags & WORKER_NOT_RUNNING)) 1010 pool->nr_running++; 1011 } 1012 1013 /* Return the first idle worker. Called with pool->lock held. */ 1014 static struct worker *first_idle_worker(struct worker_pool *pool) 1015 { 1016 if (unlikely(list_empty(&pool->idle_list))) 1017 return NULL; 1018 1019 return list_first_entry(&pool->idle_list, struct worker, entry); 1020 } 1021 1022 /** 1023 * worker_enter_idle - enter idle state 1024 * @worker: worker which is entering idle state 1025 * 1026 * @worker is entering idle state. Update stats and idle timer if 1027 * necessary. 1028 * 1029 * LOCKING: 1030 * raw_spin_lock_irq(pool->lock). 1031 */ 1032 static void worker_enter_idle(struct worker *worker) 1033 { 1034 struct worker_pool *pool = worker->pool; 1035 1036 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 1037 WARN_ON_ONCE(!list_empty(&worker->entry) && 1038 (worker->hentry.next || worker->hentry.pprev))) 1039 return; 1040 1041 /* can't use worker_set_flags(), also called from create_worker() */ 1042 worker->flags |= WORKER_IDLE; 1043 pool->nr_idle++; 1044 worker->last_active = jiffies; 1045 1046 /* idle_list is LIFO */ 1047 list_add(&worker->entry, &pool->idle_list); 1048 1049 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1050 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1051 1052 /* Sanity check nr_running. */ 1053 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 1054 } 1055 1056 /** 1057 * worker_leave_idle - leave idle state 1058 * @worker: worker which is leaving idle state 1059 * 1060 * @worker is leaving idle state. Update stats. 1061 * 1062 * LOCKING: 1063 * raw_spin_lock_irq(pool->lock). 1064 */ 1065 static void worker_leave_idle(struct worker *worker) 1066 { 1067 struct worker_pool *pool = worker->pool; 1068 1069 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 1070 return; 1071 worker_clr_flags(worker, WORKER_IDLE); 1072 pool->nr_idle--; 1073 list_del_init(&worker->entry); 1074 } 1075 1076 /** 1077 * find_worker_executing_work - find worker which is executing a work 1078 * @pool: pool of interest 1079 * @work: work to find worker for 1080 * 1081 * Find a worker which is executing @work on @pool by searching 1082 * @pool->busy_hash which is keyed by the address of @work. For a worker 1083 * to match, its current execution should match the address of @work and 1084 * its work function. This is to avoid unwanted dependency between 1085 * unrelated work executions through a work item being recycled while still 1086 * being executed. 1087 * 1088 * This is a bit tricky. A work item may be freed once its execution 1089 * starts and nothing prevents the freed area from being recycled for 1090 * another work item. If the same work item address ends up being reused 1091 * before the original execution finishes, workqueue will identify the 1092 * recycled work item as currently executing and make it wait until the 1093 * current execution finishes, introducing an unwanted dependency. 1094 * 1095 * This function checks the work item address and work function to avoid 1096 * false positives. Note that this isn't complete as one may construct a 1097 * work function which can introduce dependency onto itself through a 1098 * recycled work item. Well, if somebody wants to shoot oneself in the 1099 * foot that badly, there's only so much we can do, and if such deadlock 1100 * actually occurs, it should be easy to locate the culprit work function. 1101 * 1102 * CONTEXT: 1103 * raw_spin_lock_irq(pool->lock). 1104 * 1105 * Return: 1106 * Pointer to worker which is executing @work if found, %NULL 1107 * otherwise. 1108 */ 1109 static struct worker *find_worker_executing_work(struct worker_pool *pool, 1110 struct work_struct *work) 1111 { 1112 struct worker *worker; 1113 1114 hash_for_each_possible(pool->busy_hash, worker, hentry, 1115 (unsigned long)work) 1116 if (worker->current_work == work && 1117 worker->current_func == work->func) 1118 return worker; 1119 1120 return NULL; 1121 } 1122 1123 /** 1124 * move_linked_works - move linked works to a list 1125 * @work: start of series of works to be scheduled 1126 * @head: target list to append @work to 1127 * @nextp: out parameter for nested worklist walking 1128 * 1129 * Schedule linked works starting from @work to @head. Work series to be 1130 * scheduled starts at @work and includes any consecutive work with 1131 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on 1132 * @nextp. 1133 * 1134 * CONTEXT: 1135 * raw_spin_lock_irq(pool->lock). 1136 */ 1137 static void move_linked_works(struct work_struct *work, struct list_head *head, 1138 struct work_struct **nextp) 1139 { 1140 struct work_struct *n; 1141 1142 /* 1143 * Linked worklist will always end before the end of the list, 1144 * use NULL for list head. 1145 */ 1146 list_for_each_entry_safe_from(work, n, NULL, entry) { 1147 list_move_tail(&work->entry, head); 1148 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1149 break; 1150 } 1151 1152 /* 1153 * If we're already inside safe list traversal and have moved 1154 * multiple works to the scheduled queue, the next position 1155 * needs to be updated. 1156 */ 1157 if (nextp) 1158 *nextp = n; 1159 } 1160 1161 /** 1162 * assign_work - assign a work item and its linked work items to a worker 1163 * @work: work to assign 1164 * @worker: worker to assign to 1165 * @nextp: out parameter for nested worklist walking 1166 * 1167 * Assign @work and its linked work items to @worker. If @work is already being 1168 * executed by another worker in the same pool, it'll be punted there. 1169 * 1170 * If @nextp is not NULL, it's updated to point to the next work of the last 1171 * scheduled work. This allows assign_work() to be nested inside 1172 * list_for_each_entry_safe(). 1173 * 1174 * Returns %true if @work was successfully assigned to @worker. %false if @work 1175 * was punted to another worker already executing it. 1176 */ 1177 static bool assign_work(struct work_struct *work, struct worker *worker, 1178 struct work_struct **nextp) 1179 { 1180 struct worker_pool *pool = worker->pool; 1181 struct worker *collision; 1182 1183 lockdep_assert_held(&pool->lock); 1184 1185 /* 1186 * A single work shouldn't be executed concurrently by multiple workers. 1187 * __queue_work() ensures that @work doesn't jump to a different pool 1188 * while still running in the previous pool. Here, we should ensure that 1189 * @work is not executed concurrently by multiple workers from the same 1190 * pool. Check whether anyone is already processing the work. If so, 1191 * defer the work to the currently executing one. 1192 */ 1193 collision = find_worker_executing_work(pool, work); 1194 if (unlikely(collision)) { 1195 move_linked_works(work, &collision->scheduled, nextp); 1196 return false; 1197 } 1198 1199 move_linked_works(work, &worker->scheduled, nextp); 1200 return true; 1201 } 1202 1203 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool) 1204 { 1205 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0; 1206 1207 return &per_cpu(bh_pool_irq_works, pool->cpu)[high]; 1208 } 1209 1210 static void kick_bh_pool(struct worker_pool *pool) 1211 { 1212 #ifdef CONFIG_SMP 1213 /* see drain_dead_softirq_workfn() for BH_DRAINING */ 1214 if (unlikely(pool->cpu != smp_processor_id() && 1215 !(pool->flags & POOL_BH_DRAINING))) { 1216 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu); 1217 return; 1218 } 1219 #endif 1220 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 1221 raise_softirq_irqoff(HI_SOFTIRQ); 1222 else 1223 raise_softirq_irqoff(TASKLET_SOFTIRQ); 1224 } 1225 1226 /** 1227 * kick_pool - wake up an idle worker if necessary 1228 * @pool: pool to kick 1229 * 1230 * @pool may have pending work items. Wake up worker if necessary. Returns 1231 * whether a worker was woken up. 1232 */ 1233 static bool kick_pool(struct worker_pool *pool) 1234 { 1235 struct worker *worker = first_idle_worker(pool); 1236 struct task_struct *p; 1237 1238 lockdep_assert_held(&pool->lock); 1239 1240 if (!need_more_worker(pool) || !worker) 1241 return false; 1242 1243 if (pool->flags & POOL_BH) { 1244 kick_bh_pool(pool); 1245 return true; 1246 } 1247 1248 p = worker->task; 1249 1250 #ifdef CONFIG_SMP 1251 /* 1252 * Idle @worker is about to execute @work and waking up provides an 1253 * opportunity to migrate @worker at a lower cost by setting the task's 1254 * wake_cpu field. Let's see if we want to move @worker to improve 1255 * execution locality. 1256 * 1257 * We're waking the worker that went idle the latest and there's some 1258 * chance that @worker is marked idle but hasn't gone off CPU yet. If 1259 * so, setting the wake_cpu won't do anything. As this is a best-effort 1260 * optimization and the race window is narrow, let's leave as-is for 1261 * now. If this becomes pronounced, we can skip over workers which are 1262 * still on cpu when picking an idle worker. 1263 * 1264 * If @pool has non-strict affinity, @worker might have ended up outside 1265 * its affinity scope. Repatriate. 1266 */ 1267 if (!pool->attrs->affn_strict && 1268 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { 1269 struct work_struct *work = list_first_entry(&pool->worklist, 1270 struct work_struct, entry); 1271 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask, 1272 cpu_online_mask); 1273 if (wake_cpu < nr_cpu_ids) { 1274 p->wake_cpu = wake_cpu; 1275 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; 1276 } 1277 } 1278 #endif 1279 wake_up_process(p); 1280 return true; 1281 } 1282 1283 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 1284 1285 /* 1286 * Concurrency-managed per-cpu work items that hog CPU for longer than 1287 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 1288 * which prevents them from stalling other concurrency-managed work items. If a 1289 * work function keeps triggering this mechanism, it's likely that the work item 1290 * should be using an unbound workqueue instead. 1291 * 1292 * wq_cpu_intensive_report() tracks work functions which trigger such conditions 1293 * and report them so that they can be examined and converted to use unbound 1294 * workqueues as appropriate. To avoid flooding the console, each violating work 1295 * function is tracked and reported with exponential backoff. 1296 */ 1297 #define WCI_MAX_ENTS 128 1298 1299 struct wci_ent { 1300 work_func_t func; 1301 atomic64_t cnt; 1302 struct hlist_node hash_node; 1303 }; 1304 1305 static struct wci_ent wci_ents[WCI_MAX_ENTS]; 1306 static int wci_nr_ents; 1307 static DEFINE_RAW_SPINLOCK(wci_lock); 1308 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 1309 1310 static struct wci_ent *wci_find_ent(work_func_t func) 1311 { 1312 struct wci_ent *ent; 1313 1314 hash_for_each_possible_rcu(wci_hash, ent, hash_node, 1315 (unsigned long)func) { 1316 if (ent->func == func) 1317 return ent; 1318 } 1319 return NULL; 1320 } 1321 1322 static void wq_cpu_intensive_report(work_func_t func) 1323 { 1324 struct wci_ent *ent; 1325 1326 restart: 1327 ent = wci_find_ent(func); 1328 if (ent) { 1329 u64 cnt; 1330 1331 /* 1332 * Start reporting from the warning_thresh and back off 1333 * exponentially. 1334 */ 1335 cnt = atomic64_inc_return_relaxed(&ent->cnt); 1336 if (wq_cpu_intensive_warning_thresh && 1337 cnt >= wq_cpu_intensive_warning_thresh && 1338 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh)) 1339 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 1340 ent->func, wq_cpu_intensive_thresh_us, 1341 atomic64_read(&ent->cnt)); 1342 return; 1343 } 1344 1345 /* 1346 * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 1347 * is exhausted, something went really wrong and we probably made enough 1348 * noise already. 1349 */ 1350 if (wci_nr_ents >= WCI_MAX_ENTS) 1351 return; 1352 1353 raw_spin_lock(&wci_lock); 1354 1355 if (wci_nr_ents >= WCI_MAX_ENTS) { 1356 raw_spin_unlock(&wci_lock); 1357 return; 1358 } 1359 1360 if (wci_find_ent(func)) { 1361 raw_spin_unlock(&wci_lock); 1362 goto restart; 1363 } 1364 1365 ent = &wci_ents[wci_nr_ents++]; 1366 ent->func = func; 1367 atomic64_set(&ent->cnt, 0); 1368 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 1369 1370 raw_spin_unlock(&wci_lock); 1371 1372 goto restart; 1373 } 1374 1375 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1376 static void wq_cpu_intensive_report(work_func_t func) {} 1377 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1378 1379 /** 1380 * wq_worker_running - a worker is running again 1381 * @task: task waking up 1382 * 1383 * This function is called when a worker returns from schedule() 1384 */ 1385 void wq_worker_running(struct task_struct *task) 1386 { 1387 struct worker *worker = kthread_data(task); 1388 1389 if (!READ_ONCE(worker->sleeping)) 1390 return; 1391 1392 /* 1393 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 1394 * and the nr_running increment below, we may ruin the nr_running reset 1395 * and leave with an unexpected pool->nr_running == 1 on the newly unbound 1396 * pool. Protect against such race. 1397 */ 1398 preempt_disable(); 1399 if (!(worker->flags & WORKER_NOT_RUNNING)) 1400 worker->pool->nr_running++; 1401 preempt_enable(); 1402 1403 /* 1404 * CPU intensive auto-detection cares about how long a work item hogged 1405 * CPU without sleeping. Reset the starting timestamp on wakeup. 1406 */ 1407 worker->current_at = worker->task->se.sum_exec_runtime; 1408 1409 WRITE_ONCE(worker->sleeping, 0); 1410 } 1411 1412 /** 1413 * wq_worker_sleeping - a worker is going to sleep 1414 * @task: task going to sleep 1415 * 1416 * This function is called from schedule() when a busy worker is 1417 * going to sleep. 1418 */ 1419 void wq_worker_sleeping(struct task_struct *task) 1420 { 1421 struct worker *worker = kthread_data(task); 1422 struct worker_pool *pool; 1423 1424 /* 1425 * Rescuers, which may not have all the fields set up like normal 1426 * workers, also reach here, let's not access anything before 1427 * checking NOT_RUNNING. 1428 */ 1429 if (worker->flags & WORKER_NOT_RUNNING) 1430 return; 1431 1432 pool = worker->pool; 1433 1434 /* Return if preempted before wq_worker_running() was reached */ 1435 if (READ_ONCE(worker->sleeping)) 1436 return; 1437 1438 WRITE_ONCE(worker->sleeping, 1); 1439 raw_spin_lock_irq(&pool->lock); 1440 1441 /* 1442 * Recheck in case unbind_workers() preempted us. We don't 1443 * want to decrement nr_running after the worker is unbound 1444 * and nr_running has been reset. 1445 */ 1446 if (worker->flags & WORKER_NOT_RUNNING) { 1447 raw_spin_unlock_irq(&pool->lock); 1448 return; 1449 } 1450 1451 pool->nr_running--; 1452 if (kick_pool(pool)) 1453 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1454 1455 raw_spin_unlock_irq(&pool->lock); 1456 } 1457 1458 /** 1459 * wq_worker_tick - a scheduler tick occurred while a kworker is running 1460 * @task: task currently running 1461 * 1462 * Called from sched_tick(). We're in the IRQ context and the current 1463 * worker's fields which follow the 'K' locking rule can be accessed safely. 1464 */ 1465 void wq_worker_tick(struct task_struct *task) 1466 { 1467 struct worker *worker = kthread_data(task); 1468 struct pool_workqueue *pwq = worker->current_pwq; 1469 struct worker_pool *pool = worker->pool; 1470 1471 if (!pwq) 1472 return; 1473 1474 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 1475 1476 if (!wq_cpu_intensive_thresh_us) 1477 return; 1478 1479 /* 1480 * If the current worker is concurrency managed and hogged the CPU for 1481 * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1482 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1483 * 1484 * Set @worker->sleeping means that @worker is in the process of 1485 * switching out voluntarily and won't be contributing to 1486 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1487 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1488 * double decrements. The task is releasing the CPU anyway. Let's skip. 1489 * We probably want to make this prettier in the future. 1490 */ 1491 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1492 worker->task->se.sum_exec_runtime - worker->current_at < 1493 wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1494 return; 1495 1496 raw_spin_lock(&pool->lock); 1497 1498 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 1499 wq_cpu_intensive_report(worker->current_func); 1500 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1501 1502 if (kick_pool(pool)) 1503 pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1504 1505 raw_spin_unlock(&pool->lock); 1506 } 1507 1508 /** 1509 * wq_worker_last_func - retrieve worker's last work function 1510 * @task: Task to retrieve last work function of. 1511 * 1512 * Determine the last function a worker executed. This is called from 1513 * the scheduler to get a worker's last known identity. 1514 * 1515 * CONTEXT: 1516 * raw_spin_lock_irq(rq->lock) 1517 * 1518 * This function is called during schedule() when a kworker is going 1519 * to sleep. It's used by psi to identify aggregation workers during 1520 * dequeuing, to allow periodic aggregation to shut-off when that 1521 * worker is the last task in the system or cgroup to go to sleep. 1522 * 1523 * As this function doesn't involve any workqueue-related locking, it 1524 * only returns stable values when called from inside the scheduler's 1525 * queuing and dequeuing paths, when @task, which must be a kworker, 1526 * is guaranteed to not be processing any works. 1527 * 1528 * Return: 1529 * The last work function %current executed as a worker, NULL if it 1530 * hasn't executed any work yet. 1531 */ 1532 work_func_t wq_worker_last_func(struct task_struct *task) 1533 { 1534 struct worker *worker = kthread_data(task); 1535 1536 return worker->last_func; 1537 } 1538 1539 /** 1540 * wq_node_nr_active - Determine wq_node_nr_active to use 1541 * @wq: workqueue of interest 1542 * @node: NUMA node, can be %NUMA_NO_NODE 1543 * 1544 * Determine wq_node_nr_active to use for @wq on @node. Returns: 1545 * 1546 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. 1547 * 1548 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. 1549 * 1550 * - Otherwise, node_nr_active[@node]. 1551 */ 1552 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, 1553 int node) 1554 { 1555 if (!(wq->flags & WQ_UNBOUND)) 1556 return NULL; 1557 1558 if (node == NUMA_NO_NODE) 1559 node = nr_node_ids; 1560 1561 return wq->node_nr_active[node]; 1562 } 1563 1564 /** 1565 * wq_update_node_max_active - Update per-node max_actives to use 1566 * @wq: workqueue to update 1567 * @off_cpu: CPU that's going down, -1 if a CPU is not going down 1568 * 1569 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is 1570 * distributed among nodes according to the proportions of numbers of online 1571 * cpus. The result is always between @wq->min_active and max_active. 1572 */ 1573 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) 1574 { 1575 struct cpumask *effective = unbound_effective_cpumask(wq); 1576 int min_active = READ_ONCE(wq->min_active); 1577 int max_active = READ_ONCE(wq->max_active); 1578 int total_cpus, node; 1579 1580 lockdep_assert_held(&wq->mutex); 1581 1582 if (!wq_topo_initialized) 1583 return; 1584 1585 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective)) 1586 off_cpu = -1; 1587 1588 total_cpus = cpumask_weight_and(effective, cpu_online_mask); 1589 if (off_cpu >= 0) 1590 total_cpus--; 1591 1592 /* If all CPUs of the wq get offline, use the default values */ 1593 if (unlikely(!total_cpus)) { 1594 for_each_node(node) 1595 wq_node_nr_active(wq, node)->max = min_active; 1596 1597 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; 1598 return; 1599 } 1600 1601 for_each_node(node) { 1602 int node_cpus; 1603 1604 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); 1605 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) 1606 node_cpus--; 1607 1608 wq_node_nr_active(wq, node)->max = 1609 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), 1610 min_active, max_active); 1611 } 1612 1613 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; 1614 } 1615 1616 /** 1617 * get_pwq - get an extra reference on the specified pool_workqueue 1618 * @pwq: pool_workqueue to get 1619 * 1620 * Obtain an extra reference on @pwq. The caller should guarantee that 1621 * @pwq has positive refcnt and be holding the matching pool->lock. 1622 */ 1623 static void get_pwq(struct pool_workqueue *pwq) 1624 { 1625 lockdep_assert_held(&pwq->pool->lock); 1626 WARN_ON_ONCE(pwq->refcnt <= 0); 1627 pwq->refcnt++; 1628 } 1629 1630 /** 1631 * put_pwq - put a pool_workqueue reference 1632 * @pwq: pool_workqueue to put 1633 * 1634 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 1635 * destruction. The caller should be holding the matching pool->lock. 1636 */ 1637 static void put_pwq(struct pool_workqueue *pwq) 1638 { 1639 lockdep_assert_held(&pwq->pool->lock); 1640 if (likely(--pwq->refcnt)) 1641 return; 1642 /* 1643 * @pwq can't be released under pool->lock, bounce to a dedicated 1644 * kthread_worker to avoid A-A deadlocks. 1645 */ 1646 kthread_queue_work(pwq_release_worker, &pwq->release_work); 1647 } 1648 1649 /** 1650 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1651 * @pwq: pool_workqueue to put (can be %NULL) 1652 * 1653 * put_pwq() with locking. This function also allows %NULL @pwq. 1654 */ 1655 static void put_pwq_unlocked(struct pool_workqueue *pwq) 1656 { 1657 if (pwq) { 1658 /* 1659 * As both pwqs and pools are RCU protected, the 1660 * following lock operations are safe. 1661 */ 1662 raw_spin_lock_irq(&pwq->pool->lock); 1663 put_pwq(pwq); 1664 raw_spin_unlock_irq(&pwq->pool->lock); 1665 } 1666 } 1667 1668 static bool pwq_is_empty(struct pool_workqueue *pwq) 1669 { 1670 return !pwq->nr_active && list_empty(&pwq->inactive_works); 1671 } 1672 1673 static void __pwq_activate_work(struct pool_workqueue *pwq, 1674 struct work_struct *work) 1675 { 1676 unsigned long *wdb = work_data_bits(work); 1677 1678 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); 1679 trace_workqueue_activate_work(work); 1680 if (list_empty(&pwq->pool->worklist)) 1681 pwq->pool->watchdog_ts = jiffies; 1682 move_linked_works(work, &pwq->pool->worklist, NULL); 1683 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); 1684 } 1685 1686 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) 1687 { 1688 int max = READ_ONCE(nna->max); 1689 1690 while (true) { 1691 int old, tmp; 1692 1693 old = atomic_read(&nna->nr); 1694 if (old >= max) 1695 return false; 1696 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1); 1697 if (tmp == old) 1698 return true; 1699 } 1700 } 1701 1702 /** 1703 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq 1704 * @pwq: pool_workqueue of interest 1705 * @fill: max_active may have increased, try to increase concurrency level 1706 * 1707 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is 1708 * successfully obtained. %false otherwise. 1709 */ 1710 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) 1711 { 1712 struct workqueue_struct *wq = pwq->wq; 1713 struct worker_pool *pool = pwq->pool; 1714 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); 1715 bool obtained = false; 1716 1717 lockdep_assert_held(&pool->lock); 1718 1719 if (!nna) { 1720 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */ 1721 obtained = pwq->nr_active < READ_ONCE(wq->max_active); 1722 goto out; 1723 } 1724 1725 if (unlikely(pwq->plugged)) 1726 return false; 1727 1728 /* 1729 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is 1730 * already waiting on $nna, pwq_dec_nr_active() will maintain the 1731 * concurrency level. Don't jump the line. 1732 * 1733 * We need to ignore the pending test after max_active has increased as 1734 * pwq_dec_nr_active() can only maintain the concurrency level but not 1735 * increase it. This is indicated by @fill. 1736 */ 1737 if (!list_empty(&pwq->pending_node) && likely(!fill)) 1738 goto out; 1739 1740 obtained = tryinc_node_nr_active(nna); 1741 if (obtained) 1742 goto out; 1743 1744 /* 1745 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs 1746 * and try again. The smp_mb() is paired with the implied memory barrier 1747 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either 1748 * we see the decremented $nna->nr or they see non-empty 1749 * $nna->pending_pwqs. 1750 */ 1751 raw_spin_lock(&nna->lock); 1752 1753 if (list_empty(&pwq->pending_node)) 1754 list_add_tail(&pwq->pending_node, &nna->pending_pwqs); 1755 else if (likely(!fill)) 1756 goto out_unlock; 1757 1758 smp_mb(); 1759 1760 obtained = tryinc_node_nr_active(nna); 1761 1762 /* 1763 * If @fill, @pwq might have already been pending. Being spuriously 1764 * pending in cold paths doesn't affect anything. Let's leave it be. 1765 */ 1766 if (obtained && likely(!fill)) 1767 list_del_init(&pwq->pending_node); 1768 1769 out_unlock: 1770 raw_spin_unlock(&nna->lock); 1771 out: 1772 if (obtained) 1773 pwq->nr_active++; 1774 return obtained; 1775 } 1776 1777 /** 1778 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq 1779 * @pwq: pool_workqueue of interest 1780 * @fill: max_active may have increased, try to increase concurrency level 1781 * 1782 * Activate the first inactive work item of @pwq if available and allowed by 1783 * max_active limit. 1784 * 1785 * Returns %true if an inactive work item has been activated. %false if no 1786 * inactive work item is found or max_active limit is reached. 1787 */ 1788 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) 1789 { 1790 struct work_struct *work = 1791 list_first_entry_or_null(&pwq->inactive_works, 1792 struct work_struct, entry); 1793 1794 if (work && pwq_tryinc_nr_active(pwq, fill)) { 1795 __pwq_activate_work(pwq, work); 1796 return true; 1797 } else { 1798 return false; 1799 } 1800 } 1801 1802 /** 1803 * unplug_oldest_pwq - unplug the oldest pool_workqueue 1804 * @wq: workqueue_struct where its oldest pwq is to be unplugged 1805 * 1806 * This function should only be called for ordered workqueues where only the 1807 * oldest pwq is unplugged, the others are plugged to suspend execution to 1808 * ensure proper work item ordering:: 1809 * 1810 * dfl_pwq --------------+ [P] - plugged 1811 * | 1812 * v 1813 * pwqs -> A -> B [P] -> C [P] (newest) 1814 * | | | 1815 * 1 3 5 1816 * | | | 1817 * 2 4 6 1818 * 1819 * When the oldest pwq is drained and removed, this function should be called 1820 * to unplug the next oldest one to start its work item execution. Note that 1821 * pwq's are linked into wq->pwqs with the oldest first, so the first one in 1822 * the list is the oldest. 1823 */ 1824 static void unplug_oldest_pwq(struct workqueue_struct *wq) 1825 { 1826 struct pool_workqueue *pwq; 1827 1828 lockdep_assert_held(&wq->mutex); 1829 1830 /* Caller should make sure that pwqs isn't empty before calling */ 1831 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue, 1832 pwqs_node); 1833 raw_spin_lock_irq(&pwq->pool->lock); 1834 if (pwq->plugged) { 1835 pwq->plugged = false; 1836 if (pwq_activate_first_inactive(pwq, true)) 1837 kick_pool(pwq->pool); 1838 } 1839 raw_spin_unlock_irq(&pwq->pool->lock); 1840 } 1841 1842 /** 1843 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active 1844 * @nna: wq_node_nr_active to activate a pending pwq for 1845 * @caller_pool: worker_pool the caller is locking 1846 * 1847 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. 1848 * @caller_pool may be unlocked and relocked to lock other worker_pools. 1849 */ 1850 static void node_activate_pending_pwq(struct wq_node_nr_active *nna, 1851 struct worker_pool *caller_pool) 1852 { 1853 struct worker_pool *locked_pool = caller_pool; 1854 struct pool_workqueue *pwq; 1855 struct work_struct *work; 1856 1857 lockdep_assert_held(&caller_pool->lock); 1858 1859 raw_spin_lock(&nna->lock); 1860 retry: 1861 pwq = list_first_entry_or_null(&nna->pending_pwqs, 1862 struct pool_workqueue, pending_node); 1863 if (!pwq) 1864 goto out_unlock; 1865 1866 /* 1867 * If @pwq is for a different pool than @locked_pool, we need to lock 1868 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock 1869 * / lock dance. For that, we also need to release @nna->lock as it's 1870 * nested inside pool locks. 1871 */ 1872 if (pwq->pool != locked_pool) { 1873 raw_spin_unlock(&locked_pool->lock); 1874 locked_pool = pwq->pool; 1875 if (!raw_spin_trylock(&locked_pool->lock)) { 1876 raw_spin_unlock(&nna->lock); 1877 raw_spin_lock(&locked_pool->lock); 1878 raw_spin_lock(&nna->lock); 1879 goto retry; 1880 } 1881 } 1882 1883 /* 1884 * $pwq may not have any inactive work items due to e.g. cancellations. 1885 * Drop it from pending_pwqs and see if there's another one. 1886 */ 1887 work = list_first_entry_or_null(&pwq->inactive_works, 1888 struct work_struct, entry); 1889 if (!work) { 1890 list_del_init(&pwq->pending_node); 1891 goto retry; 1892 } 1893 1894 /* 1895 * Acquire an nr_active count and activate the inactive work item. If 1896 * $pwq still has inactive work items, rotate it to the end of the 1897 * pending_pwqs so that we round-robin through them. This means that 1898 * inactive work items are not activated in queueing order which is fine 1899 * given that there has never been any ordering across different pwqs. 1900 */ 1901 if (likely(tryinc_node_nr_active(nna))) { 1902 pwq->nr_active++; 1903 __pwq_activate_work(pwq, work); 1904 1905 if (list_empty(&pwq->inactive_works)) 1906 list_del_init(&pwq->pending_node); 1907 else 1908 list_move_tail(&pwq->pending_node, &nna->pending_pwqs); 1909 1910 /* if activating a foreign pool, make sure it's running */ 1911 if (pwq->pool != caller_pool) 1912 kick_pool(pwq->pool); 1913 } 1914 1915 out_unlock: 1916 raw_spin_unlock(&nna->lock); 1917 if (locked_pool != caller_pool) { 1918 raw_spin_unlock(&locked_pool->lock); 1919 raw_spin_lock(&caller_pool->lock); 1920 } 1921 } 1922 1923 /** 1924 * pwq_dec_nr_active - Retire an active count 1925 * @pwq: pool_workqueue of interest 1926 * 1927 * Decrement @pwq's nr_active and try to activate the first inactive work item. 1928 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. 1929 */ 1930 static void pwq_dec_nr_active(struct pool_workqueue *pwq) 1931 { 1932 struct worker_pool *pool = pwq->pool; 1933 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); 1934 1935 lockdep_assert_held(&pool->lock); 1936 1937 /* 1938 * @pwq->nr_active should be decremented for both percpu and unbound 1939 * workqueues. 1940 */ 1941 pwq->nr_active--; 1942 1943 /* 1944 * For a percpu workqueue, it's simple. Just need to kick the first 1945 * inactive work item on @pwq itself. 1946 */ 1947 if (!nna) { 1948 pwq_activate_first_inactive(pwq, false); 1949 return; 1950 } 1951 1952 /* 1953 * If @pwq is for an unbound workqueue, it's more complicated because 1954 * multiple pwqs and pools may be sharing the nr_active count. When a 1955 * pwq needs to wait for an nr_active count, it puts itself on 1956 * $nna->pending_pwqs. The following atomic_dec_return()'s implied 1957 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to 1958 * guarantee that either we see non-empty pending_pwqs or they see 1959 * decremented $nna->nr. 1960 * 1961 * $nna->max may change as CPUs come online/offline and @pwq->wq's 1962 * max_active gets updated. However, it is guaranteed to be equal to or 1963 * larger than @pwq->wq->min_active which is above zero unless freezing. 1964 * This maintains the forward progress guarantee. 1965 */ 1966 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) 1967 return; 1968 1969 if (!list_empty(&nna->pending_pwqs)) 1970 node_activate_pending_pwq(nna, pool); 1971 } 1972 1973 /** 1974 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1975 * @pwq: pwq of interest 1976 * @work_data: work_data of work which left the queue 1977 * 1978 * A work either has completed or is removed from pending queue, 1979 * decrement nr_in_flight of its pwq and handle workqueue flushing. 1980 * 1981 * NOTE: 1982 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock 1983 * and thus should be called after all other state updates for the in-flight 1984 * work item is complete. 1985 * 1986 * CONTEXT: 1987 * raw_spin_lock_irq(pool->lock). 1988 */ 1989 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1990 { 1991 int color = get_work_color(work_data); 1992 1993 if (!(work_data & WORK_STRUCT_INACTIVE)) 1994 pwq_dec_nr_active(pwq); 1995 1996 pwq->nr_in_flight[color]--; 1997 1998 /* is flush in progress and are we at the flushing tip? */ 1999 if (likely(pwq->flush_color != color)) 2000 goto out_put; 2001 2002 /* are there still in-flight works? */ 2003 if (pwq->nr_in_flight[color]) 2004 goto out_put; 2005 2006 /* this pwq is done, clear flush_color */ 2007 pwq->flush_color = -1; 2008 2009 /* 2010 * If this was the last pwq, wake up the first flusher. It 2011 * will handle the rest. 2012 */ 2013 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 2014 complete(&pwq->wq->first_flusher->done); 2015 out_put: 2016 put_pwq(pwq); 2017 } 2018 2019 /** 2020 * try_to_grab_pending - steal work item from worklist and disable irq 2021 * @work: work item to steal 2022 * @cflags: %WORK_CANCEL_ flags 2023 * @irq_flags: place to store irq state 2024 * 2025 * Try to grab PENDING bit of @work. This function can handle @work in any 2026 * stable state - idle, on timer or on worklist. 2027 * 2028 * Return: 2029 * 2030 * ======== ================================================================ 2031 * 1 if @work was pending and we successfully stole PENDING 2032 * 0 if @work was idle and we claimed PENDING 2033 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 2034 * ======== ================================================================ 2035 * 2036 * Note: 2037 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 2038 * interrupted while holding PENDING and @work off queue, irq must be 2039 * disabled on entry. This, combined with delayed_work->timer being 2040 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 2041 * 2042 * On successful return, >= 0, irq is disabled and the caller is 2043 * responsible for releasing it using local_irq_restore(*@irq_flags). 2044 * 2045 * This function is safe to call from any context including IRQ handler. 2046 */ 2047 static int try_to_grab_pending(struct work_struct *work, u32 cflags, 2048 unsigned long *irq_flags) 2049 { 2050 struct worker_pool *pool; 2051 struct pool_workqueue *pwq; 2052 2053 local_irq_save(*irq_flags); 2054 2055 /* try to steal the timer if it exists */ 2056 if (cflags & WORK_CANCEL_DELAYED) { 2057 struct delayed_work *dwork = to_delayed_work(work); 2058 2059 /* 2060 * dwork->timer is irqsafe. If del_timer() fails, it's 2061 * guaranteed that the timer is not queued anywhere and not 2062 * running on the local CPU. 2063 */ 2064 if (likely(del_timer(&dwork->timer))) 2065 return 1; 2066 } 2067 2068 /* try to claim PENDING the normal way */ 2069 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 2070 return 0; 2071 2072 rcu_read_lock(); 2073 /* 2074 * The queueing is in progress, or it is already queued. Try to 2075 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 2076 */ 2077 pool = get_work_pool(work); 2078 if (!pool) 2079 goto fail; 2080 2081 raw_spin_lock(&pool->lock); 2082 /* 2083 * work->data is guaranteed to point to pwq only while the work 2084 * item is queued on pwq->wq, and both updating work->data to point 2085 * to pwq on queueing and to pool on dequeueing are done under 2086 * pwq->pool->lock. This in turn guarantees that, if work->data 2087 * points to pwq which is associated with a locked pool, the work 2088 * item is currently queued on that pool. 2089 */ 2090 pwq = get_work_pwq(work); 2091 if (pwq && pwq->pool == pool) { 2092 unsigned long work_data = *work_data_bits(work); 2093 2094 debug_work_deactivate(work); 2095 2096 /* 2097 * A cancelable inactive work item must be in the 2098 * pwq->inactive_works since a queued barrier can't be 2099 * canceled (see the comments in insert_wq_barrier()). 2100 * 2101 * An inactive work item cannot be deleted directly because 2102 * it might have linked barrier work items which, if left 2103 * on the inactive_works list, will confuse pwq->nr_active 2104 * management later on and cause stall. Move the linked 2105 * barrier work items to the worklist when deleting the grabbed 2106 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that 2107 * it doesn't participate in nr_active management in later 2108 * pwq_dec_nr_in_flight(). 2109 */ 2110 if (work_data & WORK_STRUCT_INACTIVE) 2111 move_linked_works(work, &pwq->pool->worklist, NULL); 2112 2113 list_del_init(&work->entry); 2114 2115 /* 2116 * work->data points to pwq iff queued. Let's point to pool. As 2117 * this destroys work->data needed by the next step, stash it. 2118 */ 2119 set_work_pool_and_keep_pending(work, pool->id, 2120 pool_offq_flags(pool)); 2121 2122 /* must be the last step, see the function comment */ 2123 pwq_dec_nr_in_flight(pwq, work_data); 2124 2125 raw_spin_unlock(&pool->lock); 2126 rcu_read_unlock(); 2127 return 1; 2128 } 2129 raw_spin_unlock(&pool->lock); 2130 fail: 2131 rcu_read_unlock(); 2132 local_irq_restore(*irq_flags); 2133 return -EAGAIN; 2134 } 2135 2136 /** 2137 * work_grab_pending - steal work item from worklist and disable irq 2138 * @work: work item to steal 2139 * @cflags: %WORK_CANCEL_ flags 2140 * @irq_flags: place to store IRQ state 2141 * 2142 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer 2143 * or on worklist. 2144 * 2145 * Can be called from any context. IRQ is disabled on return with IRQ state 2146 * stored in *@irq_flags. The caller is responsible for re-enabling it using 2147 * local_irq_restore(). 2148 * 2149 * Returns %true if @work was pending. %false if idle. 2150 */ 2151 static bool work_grab_pending(struct work_struct *work, u32 cflags, 2152 unsigned long *irq_flags) 2153 { 2154 int ret; 2155 2156 while (true) { 2157 ret = try_to_grab_pending(work, cflags, irq_flags); 2158 if (ret >= 0) 2159 return ret; 2160 cpu_relax(); 2161 } 2162 } 2163 2164 /** 2165 * insert_work - insert a work into a pool 2166 * @pwq: pwq @work belongs to 2167 * @work: work to insert 2168 * @head: insertion point 2169 * @extra_flags: extra WORK_STRUCT_* flags to set 2170 * 2171 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 2172 * work_struct flags. 2173 * 2174 * CONTEXT: 2175 * raw_spin_lock_irq(pool->lock). 2176 */ 2177 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 2178 struct list_head *head, unsigned int extra_flags) 2179 { 2180 debug_work_activate(work); 2181 2182 /* record the work call stack in order to print it in KASAN reports */ 2183 kasan_record_aux_stack_noalloc(work); 2184 2185 /* we own @work, set data and link */ 2186 set_work_pwq(work, pwq, extra_flags); 2187 list_add_tail(&work->entry, head); 2188 get_pwq(pwq); 2189 } 2190 2191 /* 2192 * Test whether @work is being queued from another work executing on the 2193 * same workqueue. 2194 */ 2195 static bool is_chained_work(struct workqueue_struct *wq) 2196 { 2197 struct worker *worker; 2198 2199 worker = current_wq_worker(); 2200 /* 2201 * Return %true iff I'm a worker executing a work item on @wq. If 2202 * I'm @worker, it's safe to dereference it without locking. 2203 */ 2204 return worker && worker->current_pwq->wq == wq; 2205 } 2206 2207 /* 2208 * When queueing an unbound work item to a wq, prefer local CPU if allowed 2209 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 2210 * avoid perturbing sensitive tasks. 2211 */ 2212 static int wq_select_unbound_cpu(int cpu) 2213 { 2214 int new_cpu; 2215 2216 if (likely(!wq_debug_force_rr_cpu)) { 2217 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 2218 return cpu; 2219 } else { 2220 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 2221 } 2222 2223 new_cpu = __this_cpu_read(wq_rr_cpu_last); 2224 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 2225 if (unlikely(new_cpu >= nr_cpu_ids)) { 2226 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 2227 if (unlikely(new_cpu >= nr_cpu_ids)) 2228 return cpu; 2229 } 2230 __this_cpu_write(wq_rr_cpu_last, new_cpu); 2231 2232 return new_cpu; 2233 } 2234 2235 static void __queue_work(int cpu, struct workqueue_struct *wq, 2236 struct work_struct *work) 2237 { 2238 struct pool_workqueue *pwq; 2239 struct worker_pool *last_pool, *pool; 2240 unsigned int work_flags; 2241 unsigned int req_cpu = cpu; 2242 2243 /* 2244 * While a work item is PENDING && off queue, a task trying to 2245 * steal the PENDING will busy-loop waiting for it to either get 2246 * queued or lose PENDING. Grabbing PENDING and queueing should 2247 * happen with IRQ disabled. 2248 */ 2249 lockdep_assert_irqs_disabled(); 2250 2251 /* 2252 * For a draining wq, only works from the same workqueue are 2253 * allowed. The __WQ_DESTROYING helps to spot the issue that 2254 * queues a new work item to a wq after destroy_workqueue(wq). 2255 */ 2256 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 2257 WARN_ON_ONCE(!is_chained_work(wq)))) 2258 return; 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 = from_timer(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 dwork->wq = wq; 2512 dwork->cpu = cpu; 2513 timer->expires = jiffies + delay; 2514 2515 if (housekeeping_enabled(HK_TYPE_TIMER)) { 2516 /* If the current cpu is a housekeeping cpu, use it. */ 2517 cpu = smp_processor_id(); 2518 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) 2519 cpu = housekeeping_any_cpu(HK_TYPE_TIMER); 2520 add_timer_on(timer, cpu); 2521 } else { 2522 if (likely(cpu == WORK_CPU_UNBOUND)) 2523 add_timer_global(timer); 2524 else 2525 add_timer_on(timer, cpu); 2526 } 2527 } 2528 2529 /** 2530 * queue_delayed_work_on - queue work on specific CPU after delay 2531 * @cpu: CPU number to execute work on 2532 * @wq: workqueue to use 2533 * @dwork: work to queue 2534 * @delay: number of jiffies to wait before queueing 2535 * 2536 * Return: %false if @work was already on a queue, %true otherwise. If 2537 * @delay is zero and @dwork is idle, it will be scheduled for immediate 2538 * execution. 2539 */ 2540 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 2541 struct delayed_work *dwork, unsigned long delay) 2542 { 2543 struct work_struct *work = &dwork->work; 2544 bool ret = false; 2545 unsigned long irq_flags; 2546 2547 /* read the comment in __queue_work() */ 2548 local_irq_save(irq_flags); 2549 2550 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2551 !clear_pending_if_disabled(work)) { 2552 __queue_delayed_work(cpu, wq, dwork, delay); 2553 ret = true; 2554 } 2555 2556 local_irq_restore(irq_flags); 2557 return ret; 2558 } 2559 EXPORT_SYMBOL(queue_delayed_work_on); 2560 2561 /** 2562 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 2563 * @cpu: CPU number to execute work on 2564 * @wq: workqueue to use 2565 * @dwork: work to queue 2566 * @delay: number of jiffies to wait before queueing 2567 * 2568 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 2569 * modify @dwork's timer so that it expires after @delay. If @delay is 2570 * zero, @work is guaranteed to be scheduled immediately regardless of its 2571 * current state. 2572 * 2573 * Return: %false if @dwork was idle and queued, %true if @dwork was 2574 * pending and its timer was modified. 2575 * 2576 * This function is safe to call from any context including IRQ handler. 2577 * See try_to_grab_pending() for details. 2578 */ 2579 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 2580 struct delayed_work *dwork, unsigned long delay) 2581 { 2582 unsigned long irq_flags; 2583 bool ret; 2584 2585 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags); 2586 2587 if (!clear_pending_if_disabled(&dwork->work)) 2588 __queue_delayed_work(cpu, wq, dwork, delay); 2589 2590 local_irq_restore(irq_flags); 2591 return ret; 2592 } 2593 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2594 2595 static void rcu_work_rcufn(struct rcu_head *rcu) 2596 { 2597 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 2598 2599 /* read the comment in __queue_work() */ 2600 local_irq_disable(); 2601 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 2602 local_irq_enable(); 2603 } 2604 2605 /** 2606 * queue_rcu_work - queue work after a RCU grace period 2607 * @wq: workqueue to use 2608 * @rwork: work to queue 2609 * 2610 * Return: %false if @rwork was already pending, %true otherwise. Note 2611 * that a full RCU grace period is guaranteed only after a %true return. 2612 * While @rwork is guaranteed to be executed after a %false return, the 2613 * execution may happen before a full RCU grace period has passed. 2614 */ 2615 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 2616 { 2617 struct work_struct *work = &rwork->work; 2618 2619 /* 2620 * rcu_work can't be canceled or disabled. Warn if the user reached 2621 * inside @rwork and disabled the inner work. 2622 */ 2623 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2624 !WARN_ON_ONCE(clear_pending_if_disabled(work))) { 2625 rwork->wq = wq; 2626 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 2627 return true; 2628 } 2629 2630 return false; 2631 } 2632 EXPORT_SYMBOL(queue_rcu_work); 2633 2634 static struct worker *alloc_worker(int node) 2635 { 2636 struct worker *worker; 2637 2638 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2639 if (worker) { 2640 INIT_LIST_HEAD(&worker->entry); 2641 INIT_LIST_HEAD(&worker->scheduled); 2642 INIT_LIST_HEAD(&worker->node); 2643 /* on creation a worker is in !idle && prep state */ 2644 worker->flags = WORKER_PREP; 2645 } 2646 return worker; 2647 } 2648 2649 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2650 { 2651 if (pool->cpu < 0 && pool->attrs->affn_strict) 2652 return pool->attrs->__pod_cpumask; 2653 else 2654 return pool->attrs->cpumask; 2655 } 2656 2657 /** 2658 * worker_attach_to_pool() - attach a worker to a pool 2659 * @worker: worker to be attached 2660 * @pool: the target pool 2661 * 2662 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 2663 * cpu-binding of @worker are kept coordinated with the pool across 2664 * cpu-[un]hotplugs. 2665 */ 2666 static void worker_attach_to_pool(struct worker *worker, 2667 struct worker_pool *pool) 2668 { 2669 mutex_lock(&wq_pool_attach_mutex); 2670 2671 /* 2672 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable 2673 * across this function. See the comments above the flag definition for 2674 * details. BH workers are, while per-CPU, always DISASSOCIATED. 2675 */ 2676 if (pool->flags & POOL_DISASSOCIATED) { 2677 worker->flags |= WORKER_UNBOUND; 2678 } else { 2679 WARN_ON_ONCE(pool->flags & POOL_BH); 2680 kthread_set_per_cpu(worker->task, pool->cpu); 2681 } 2682 2683 if (worker->rescue_wq) 2684 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2685 2686 list_add_tail(&worker->node, &pool->workers); 2687 worker->pool = pool; 2688 2689 mutex_unlock(&wq_pool_attach_mutex); 2690 } 2691 2692 static void unbind_worker(struct worker *worker) 2693 { 2694 lockdep_assert_held(&wq_pool_attach_mutex); 2695 2696 kthread_set_per_cpu(worker->task, -1); 2697 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2698 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2699 else 2700 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2701 } 2702 2703 2704 static void detach_worker(struct worker *worker) 2705 { 2706 lockdep_assert_held(&wq_pool_attach_mutex); 2707 2708 unbind_worker(worker); 2709 list_del(&worker->node); 2710 } 2711 2712 /** 2713 * worker_detach_from_pool() - detach a worker from its pool 2714 * @worker: worker which is attached to its pool 2715 * 2716 * Undo the attaching which had been done in worker_attach_to_pool(). The 2717 * caller worker shouldn't access to the pool after detached except it has 2718 * other reference to the pool. 2719 */ 2720 static void worker_detach_from_pool(struct worker *worker) 2721 { 2722 struct worker_pool *pool = worker->pool; 2723 2724 /* there is one permanent BH worker per CPU which should never detach */ 2725 WARN_ON_ONCE(pool->flags & POOL_BH); 2726 2727 mutex_lock(&wq_pool_attach_mutex); 2728 detach_worker(worker); 2729 worker->pool = NULL; 2730 mutex_unlock(&wq_pool_attach_mutex); 2731 2732 /* clear leftover flags without pool->lock after it is detached */ 2733 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2734 } 2735 2736 static int format_worker_id(char *buf, size_t size, struct worker *worker, 2737 struct worker_pool *pool) 2738 { 2739 if (worker->rescue_wq) 2740 return scnprintf(buf, size, "kworker/R-%s", 2741 worker->rescue_wq->name); 2742 2743 if (pool) { 2744 if (pool->cpu >= 0) 2745 return scnprintf(buf, size, "kworker/%d:%d%s", 2746 pool->cpu, worker->id, 2747 pool->attrs->nice < 0 ? "H" : ""); 2748 else 2749 return scnprintf(buf, size, "kworker/u%d:%d", 2750 pool->id, worker->id); 2751 } else { 2752 return scnprintf(buf, size, "kworker/dying"); 2753 } 2754 } 2755 2756 /** 2757 * create_worker - create a new workqueue worker 2758 * @pool: pool the new worker will belong to 2759 * 2760 * Create and start a new worker which is attached to @pool. 2761 * 2762 * CONTEXT: 2763 * Might sleep. Does GFP_KERNEL allocations. 2764 * 2765 * Return: 2766 * Pointer to the newly created worker. 2767 */ 2768 static struct worker *create_worker(struct worker_pool *pool) 2769 { 2770 struct worker *worker; 2771 int id; 2772 2773 /* ID is needed to determine kthread name */ 2774 id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 2775 if (id < 0) { 2776 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 2777 ERR_PTR(id)); 2778 return NULL; 2779 } 2780 2781 worker = alloc_worker(pool->node); 2782 if (!worker) { 2783 pr_err_once("workqueue: Failed to allocate a worker\n"); 2784 goto fail; 2785 } 2786 2787 worker->id = id; 2788 2789 if (!(pool->flags & POOL_BH)) { 2790 char id_buf[WORKER_ID_LEN]; 2791 2792 format_worker_id(id_buf, sizeof(id_buf), worker, pool); 2793 worker->task = kthread_create_on_node(worker_thread, worker, 2794 pool->node, "%s", id_buf); 2795 if (IS_ERR(worker->task)) { 2796 if (PTR_ERR(worker->task) == -EINTR) { 2797 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n", 2798 id_buf); 2799 } else { 2800 pr_err_once("workqueue: Failed to create a worker thread: %pe", 2801 worker->task); 2802 } 2803 goto fail; 2804 } 2805 2806 set_user_nice(worker->task, pool->attrs->nice); 2807 kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 2808 } 2809 2810 /* successful, attach the worker to the pool */ 2811 worker_attach_to_pool(worker, pool); 2812 2813 /* start the newly created worker */ 2814 raw_spin_lock_irq(&pool->lock); 2815 2816 worker->pool->nr_workers++; 2817 worker_enter_idle(worker); 2818 2819 /* 2820 * @worker is waiting on a completion in kthread() and will trigger hung 2821 * check if not woken up soon. As kick_pool() is noop if @pool is empty, 2822 * wake it up explicitly. 2823 */ 2824 if (worker->task) 2825 wake_up_process(worker->task); 2826 2827 raw_spin_unlock_irq(&pool->lock); 2828 2829 return worker; 2830 2831 fail: 2832 ida_free(&pool->worker_ida, id); 2833 kfree(worker); 2834 return NULL; 2835 } 2836 2837 static void detach_dying_workers(struct list_head *cull_list) 2838 { 2839 struct worker *worker; 2840 2841 list_for_each_entry(worker, cull_list, entry) 2842 detach_worker(worker); 2843 } 2844 2845 static void reap_dying_workers(struct list_head *cull_list) 2846 { 2847 struct worker *worker, *tmp; 2848 2849 list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2850 list_del_init(&worker->entry); 2851 kthread_stop_put(worker->task); 2852 kfree(worker); 2853 } 2854 } 2855 2856 /** 2857 * set_worker_dying - Tag a worker for destruction 2858 * @worker: worker to be destroyed 2859 * @list: transfer worker away from its pool->idle_list and into list 2860 * 2861 * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2862 * should be idle. 2863 * 2864 * CONTEXT: 2865 * raw_spin_lock_irq(pool->lock). 2866 */ 2867 static void set_worker_dying(struct worker *worker, struct list_head *list) 2868 { 2869 struct worker_pool *pool = worker->pool; 2870 2871 lockdep_assert_held(&pool->lock); 2872 lockdep_assert_held(&wq_pool_attach_mutex); 2873 2874 /* sanity check frenzy */ 2875 if (WARN_ON(worker->current_work) || 2876 WARN_ON(!list_empty(&worker->scheduled)) || 2877 WARN_ON(!(worker->flags & WORKER_IDLE))) 2878 return; 2879 2880 pool->nr_workers--; 2881 pool->nr_idle--; 2882 2883 worker->flags |= WORKER_DIE; 2884 2885 list_move(&worker->entry, list); 2886 2887 /* get an extra task struct reference for later kthread_stop_put() */ 2888 get_task_struct(worker->task); 2889 } 2890 2891 /** 2892 * idle_worker_timeout - check if some idle workers can now be deleted. 2893 * @t: The pool's idle_timer that just expired 2894 * 2895 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 2896 * worker_leave_idle(), as a worker flicking between idle and active while its 2897 * pool is at the too_many_workers() tipping point would cause too much timer 2898 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 2899 * it expire and re-evaluate things from there. 2900 */ 2901 static void idle_worker_timeout(struct timer_list *t) 2902 { 2903 struct worker_pool *pool = from_timer(pool, t, idle_timer); 2904 bool do_cull = false; 2905 2906 if (work_pending(&pool->idle_cull_work)) 2907 return; 2908 2909 raw_spin_lock_irq(&pool->lock); 2910 2911 if (too_many_workers(pool)) { 2912 struct worker *worker; 2913 unsigned long expires; 2914 2915 /* idle_list is kept in LIFO order, check the last one */ 2916 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2917 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2918 do_cull = !time_before(jiffies, expires); 2919 2920 if (!do_cull) 2921 mod_timer(&pool->idle_timer, expires); 2922 } 2923 raw_spin_unlock_irq(&pool->lock); 2924 2925 if (do_cull) 2926 queue_work(system_unbound_wq, &pool->idle_cull_work); 2927 } 2928 2929 /** 2930 * idle_cull_fn - cull workers that have been idle for too long. 2931 * @work: the pool's work for handling these idle workers 2932 * 2933 * This goes through a pool's idle workers and gets rid of those that have been 2934 * idle for at least IDLE_WORKER_TIMEOUT seconds. 2935 * 2936 * We don't want to disturb isolated CPUs because of a pcpu kworker being 2937 * culled, so this also resets worker affinity. This requires a sleepable 2938 * context, hence the split between timer callback and work item. 2939 */ 2940 static void idle_cull_fn(struct work_struct *work) 2941 { 2942 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2943 LIST_HEAD(cull_list); 2944 2945 /* 2946 * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2947 * cannot proceed beyong set_pf_worker() in its self-destruct path. 2948 * This is required as a previously-preempted worker could run after 2949 * set_worker_dying() has happened but before detach_dying_workers() did. 2950 */ 2951 mutex_lock(&wq_pool_attach_mutex); 2952 raw_spin_lock_irq(&pool->lock); 2953 2954 while (too_many_workers(pool)) { 2955 struct worker *worker; 2956 unsigned long expires; 2957 2958 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2959 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2960 2961 if (time_before(jiffies, expires)) { 2962 mod_timer(&pool->idle_timer, expires); 2963 break; 2964 } 2965 2966 set_worker_dying(worker, &cull_list); 2967 } 2968 2969 raw_spin_unlock_irq(&pool->lock); 2970 detach_dying_workers(&cull_list); 2971 mutex_unlock(&wq_pool_attach_mutex); 2972 2973 reap_dying_workers(&cull_list); 2974 } 2975 2976 static void send_mayday(struct work_struct *work) 2977 { 2978 struct pool_workqueue *pwq = get_work_pwq(work); 2979 struct workqueue_struct *wq = pwq->wq; 2980 2981 lockdep_assert_held(&wq_mayday_lock); 2982 2983 if (!wq->rescuer) 2984 return; 2985 2986 /* mayday mayday mayday */ 2987 if (list_empty(&pwq->mayday_node)) { 2988 /* 2989 * If @pwq is for an unbound wq, its base ref may be put at 2990 * any time due to an attribute change. Pin @pwq until the 2991 * rescuer is done with it. 2992 */ 2993 get_pwq(pwq); 2994 list_add_tail(&pwq->mayday_node, &wq->maydays); 2995 wake_up_process(wq->rescuer->task); 2996 pwq->stats[PWQ_STAT_MAYDAY]++; 2997 } 2998 } 2999 3000 static void pool_mayday_timeout(struct timer_list *t) 3001 { 3002 struct worker_pool *pool = from_timer(pool, t, mayday_timer); 3003 struct work_struct *work; 3004 3005 raw_spin_lock_irq(&pool->lock); 3006 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 3007 3008 if (need_to_create_worker(pool)) { 3009 /* 3010 * We've been trying to create a new worker but 3011 * haven't been successful. We might be hitting an 3012 * allocation deadlock. Send distress signals to 3013 * rescuers. 3014 */ 3015 list_for_each_entry(work, &pool->worklist, entry) 3016 send_mayday(work); 3017 } 3018 3019 raw_spin_unlock(&wq_mayday_lock); 3020 raw_spin_unlock_irq(&pool->lock); 3021 3022 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 3023 } 3024 3025 /** 3026 * maybe_create_worker - create a new worker if necessary 3027 * @pool: pool to create a new worker for 3028 * 3029 * Create a new worker for @pool if necessary. @pool is guaranteed to 3030 * have at least one idle worker on return from this function. If 3031 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 3032 * sent to all rescuers with works scheduled on @pool to resolve 3033 * possible allocation deadlock. 3034 * 3035 * On return, need_to_create_worker() is guaranteed to be %false and 3036 * may_start_working() %true. 3037 * 3038 * LOCKING: 3039 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3040 * multiple times. Does GFP_KERNEL allocations. Called only from 3041 * manager. 3042 */ 3043 static void maybe_create_worker(struct worker_pool *pool) 3044 __releases(&pool->lock) 3045 __acquires(&pool->lock) 3046 { 3047 restart: 3048 raw_spin_unlock_irq(&pool->lock); 3049 3050 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 3051 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 3052 3053 while (true) { 3054 if (create_worker(pool) || !need_to_create_worker(pool)) 3055 break; 3056 3057 schedule_timeout_interruptible(CREATE_COOLDOWN); 3058 3059 if (!need_to_create_worker(pool)) 3060 break; 3061 } 3062 3063 del_timer_sync(&pool->mayday_timer); 3064 raw_spin_lock_irq(&pool->lock); 3065 /* 3066 * This is necessary even after a new worker was just successfully 3067 * created as @pool->lock was dropped and the new worker might have 3068 * already become busy. 3069 */ 3070 if (need_to_create_worker(pool)) 3071 goto restart; 3072 } 3073 3074 /** 3075 * manage_workers - manage worker pool 3076 * @worker: self 3077 * 3078 * Assume the manager role and manage the worker pool @worker belongs 3079 * to. At any given time, there can be only zero or one manager per 3080 * pool. The exclusion is handled automatically by this function. 3081 * 3082 * The caller can safely start processing works on false return. On 3083 * true return, it's guaranteed that need_to_create_worker() is false 3084 * and may_start_working() is true. 3085 * 3086 * CONTEXT: 3087 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3088 * multiple times. Does GFP_KERNEL allocations. 3089 * 3090 * Return: 3091 * %false if the pool doesn't need management and the caller can safely 3092 * start processing works, %true if management function was performed and 3093 * the conditions that the caller verified before calling the function may 3094 * no longer be true. 3095 */ 3096 static bool manage_workers(struct worker *worker) 3097 { 3098 struct worker_pool *pool = worker->pool; 3099 3100 if (pool->flags & POOL_MANAGER_ACTIVE) 3101 return false; 3102 3103 pool->flags |= POOL_MANAGER_ACTIVE; 3104 pool->manager = worker; 3105 3106 maybe_create_worker(pool); 3107 3108 pool->manager = NULL; 3109 pool->flags &= ~POOL_MANAGER_ACTIVE; 3110 rcuwait_wake_up(&manager_wait); 3111 return true; 3112 } 3113 3114 /** 3115 * process_one_work - process single work 3116 * @worker: self 3117 * @work: work to process 3118 * 3119 * Process @work. This function contains all the logics necessary to 3120 * process a single work including synchronization against and 3121 * interaction with other workers on the same cpu, queueing and 3122 * flushing. As long as context requirement is met, any worker can 3123 * call this function to process a work. 3124 * 3125 * CONTEXT: 3126 * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 3127 */ 3128 static void process_one_work(struct worker *worker, struct work_struct *work) 3129 __releases(&pool->lock) 3130 __acquires(&pool->lock) 3131 { 3132 struct pool_workqueue *pwq = get_work_pwq(work); 3133 struct worker_pool *pool = worker->pool; 3134 unsigned long work_data; 3135 int lockdep_start_depth, rcu_start_depth; 3136 bool bh_draining = pool->flags & POOL_BH_DRAINING; 3137 #ifdef CONFIG_LOCKDEP 3138 /* 3139 * It is permissible to free the struct work_struct from 3140 * inside the function that is called from it, this we need to 3141 * take into account for lockdep too. To avoid bogus "held 3142 * lock freed" warnings as well as problems when looking into 3143 * work->lockdep_map, make a copy and use that here. 3144 */ 3145 struct lockdep_map lockdep_map; 3146 3147 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 3148 #endif 3149 /* ensure we're on the correct CPU */ 3150 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 3151 raw_smp_processor_id() != pool->cpu); 3152 3153 /* claim and dequeue */ 3154 debug_work_deactivate(work); 3155 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 3156 worker->current_work = work; 3157 worker->current_func = work->func; 3158 worker->current_pwq = pwq; 3159 if (worker->task) 3160 worker->current_at = worker->task->se.sum_exec_runtime; 3161 work_data = *work_data_bits(work); 3162 worker->current_color = get_work_color(work_data); 3163 3164 /* 3165 * Record wq name for cmdline and debug reporting, may get 3166 * overridden through set_worker_desc(). 3167 */ 3168 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 3169 3170 list_del_init(&work->entry); 3171 3172 /* 3173 * CPU intensive works don't participate in concurrency management. 3174 * They're the scheduler's responsibility. This takes @worker out 3175 * of concurrency management and the next code block will chain 3176 * execution of the pending work items. 3177 */ 3178 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 3179 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 3180 3181 /* 3182 * Kick @pool if necessary. It's always noop for per-cpu worker pools 3183 * since nr_running would always be >= 1 at this point. This is used to 3184 * chain execution of the pending work items for WORKER_NOT_RUNNING 3185 * workers such as the UNBOUND and CPU_INTENSIVE ones. 3186 */ 3187 kick_pool(pool); 3188 3189 /* 3190 * Record the last pool and clear PENDING which should be the last 3191 * update to @work. Also, do this inside @pool->lock so that 3192 * PENDING and queued state changes happen together while IRQ is 3193 * disabled. 3194 */ 3195 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool)); 3196 3197 pwq->stats[PWQ_STAT_STARTED]++; 3198 raw_spin_unlock_irq(&pool->lock); 3199 3200 rcu_start_depth = rcu_preempt_depth(); 3201 lockdep_start_depth = lockdep_depth(current); 3202 /* see drain_dead_softirq_workfn() */ 3203 if (!bh_draining) 3204 lock_map_acquire(pwq->wq->lockdep_map); 3205 lock_map_acquire(&lockdep_map); 3206 /* 3207 * Strictly speaking we should mark the invariant state without holding 3208 * any locks, that is, before these two lock_map_acquire()'s. 3209 * 3210 * However, that would result in: 3211 * 3212 * A(W1) 3213 * WFC(C) 3214 * A(W1) 3215 * C(C) 3216 * 3217 * Which would create W1->C->W1 dependencies, even though there is no 3218 * actual deadlock possible. There are two solutions, using a 3219 * read-recursive acquire on the work(queue) 'locks', but this will then 3220 * hit the lockdep limitation on recursive locks, or simply discard 3221 * these locks. 3222 * 3223 * AFAICT there is no possible deadlock scenario between the 3224 * flush_work() and complete() primitives (except for single-threaded 3225 * workqueues), so hiding them isn't a problem. 3226 */ 3227 lockdep_invariant_state(true); 3228 trace_workqueue_execute_start(work); 3229 worker->current_func(work); 3230 /* 3231 * While we must be careful to not use "work" after this, the trace 3232 * point will only record its address. 3233 */ 3234 trace_workqueue_execute_end(work, worker->current_func); 3235 pwq->stats[PWQ_STAT_COMPLETED]++; 3236 lock_map_release(&lockdep_map); 3237 if (!bh_draining) 3238 lock_map_release(pwq->wq->lockdep_map); 3239 3240 if (unlikely((worker->task && in_atomic()) || 3241 lockdep_depth(current) != lockdep_start_depth || 3242 rcu_preempt_depth() != rcu_start_depth)) { 3243 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n" 3244 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n", 3245 current->comm, task_pid_nr(current), preempt_count(), 3246 lockdep_start_depth, lockdep_depth(current), 3247 rcu_start_depth, rcu_preempt_depth(), 3248 worker->current_func); 3249 debug_show_held_locks(current); 3250 dump_stack(); 3251 } 3252 3253 /* 3254 * The following prevents a kworker from hogging CPU on !PREEMPTION 3255 * kernels, where a requeueing work item waiting for something to 3256 * happen could deadlock with stop_machine as such work item could 3257 * indefinitely requeue itself while all other CPUs are trapped in 3258 * stop_machine. At the same time, report a quiescent RCU state so 3259 * the same condition doesn't freeze RCU. 3260 */ 3261 if (worker->task) 3262 cond_resched(); 3263 3264 raw_spin_lock_irq(&pool->lock); 3265 3266 /* 3267 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 3268 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 3269 * wq_cpu_intensive_thresh_us. Clear it. 3270 */ 3271 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 3272 3273 /* tag the worker for identification in schedule() */ 3274 worker->last_func = worker->current_func; 3275 3276 /* we're done with it, release */ 3277 hash_del(&worker->hentry); 3278 worker->current_work = NULL; 3279 worker->current_func = NULL; 3280 worker->current_pwq = NULL; 3281 worker->current_color = INT_MAX; 3282 3283 /* must be the last step, see the function comment */ 3284 pwq_dec_nr_in_flight(pwq, work_data); 3285 } 3286 3287 /** 3288 * process_scheduled_works - process scheduled works 3289 * @worker: self 3290 * 3291 * Process all scheduled works. Please note that the scheduled list 3292 * may change while processing a work, so this function repeatedly 3293 * fetches a work from the top and executes it. 3294 * 3295 * CONTEXT: 3296 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3297 * multiple times. 3298 */ 3299 static void process_scheduled_works(struct worker *worker) 3300 { 3301 struct work_struct *work; 3302 bool first = true; 3303 3304 while ((work = list_first_entry_or_null(&worker->scheduled, 3305 struct work_struct, entry))) { 3306 if (first) { 3307 worker->pool->watchdog_ts = jiffies; 3308 first = false; 3309 } 3310 process_one_work(worker, work); 3311 } 3312 } 3313 3314 static void set_pf_worker(bool val) 3315 { 3316 mutex_lock(&wq_pool_attach_mutex); 3317 if (val) 3318 current->flags |= PF_WQ_WORKER; 3319 else 3320 current->flags &= ~PF_WQ_WORKER; 3321 mutex_unlock(&wq_pool_attach_mutex); 3322 } 3323 3324 /** 3325 * worker_thread - the worker thread function 3326 * @__worker: self 3327 * 3328 * The worker thread function. All workers belong to a worker_pool - 3329 * either a per-cpu one or dynamic unbound one. These workers process all 3330 * work items regardless of their specific target workqueue. The only 3331 * exception is work items which belong to workqueues with a rescuer which 3332 * will be explained in rescuer_thread(). 3333 * 3334 * Return: 0 3335 */ 3336 static int worker_thread(void *__worker) 3337 { 3338 struct worker *worker = __worker; 3339 struct worker_pool *pool = worker->pool; 3340 3341 /* tell the scheduler that this is a workqueue worker */ 3342 set_pf_worker(true); 3343 woke_up: 3344 raw_spin_lock_irq(&pool->lock); 3345 3346 /* am I supposed to die? */ 3347 if (unlikely(worker->flags & WORKER_DIE)) { 3348 raw_spin_unlock_irq(&pool->lock); 3349 set_pf_worker(false); 3350 /* 3351 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool 3352 * shouldn't be accessed, reset it to NULL in case otherwise. 3353 */ 3354 worker->pool = NULL; 3355 ida_free(&pool->worker_ida, worker->id); 3356 return 0; 3357 } 3358 3359 worker_leave_idle(worker); 3360 recheck: 3361 /* no more worker necessary? */ 3362 if (!need_more_worker(pool)) 3363 goto sleep; 3364 3365 /* do we need to manage? */ 3366 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 3367 goto recheck; 3368 3369 /* 3370 * ->scheduled list can only be filled while a worker is 3371 * preparing to process a work or actually processing it. 3372 * Make sure nobody diddled with it while I was sleeping. 3373 */ 3374 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3375 3376 /* 3377 * Finish PREP stage. We're guaranteed to have at least one idle 3378 * worker or that someone else has already assumed the manager 3379 * role. This is where @worker starts participating in concurrency 3380 * management if applicable and concurrency management is restored 3381 * after being rebound. See rebind_workers() for details. 3382 */ 3383 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3384 3385 do { 3386 struct work_struct *work = 3387 list_first_entry(&pool->worklist, 3388 struct work_struct, entry); 3389 3390 if (assign_work(work, worker, NULL)) 3391 process_scheduled_works(worker); 3392 } while (keep_working(pool)); 3393 3394 worker_set_flags(worker, WORKER_PREP); 3395 sleep: 3396 /* 3397 * pool->lock is held and there's no work to process and no need to 3398 * manage, sleep. Workers are woken up only while holding 3399 * pool->lock or from local cpu, so setting the current state 3400 * before releasing pool->lock is enough to prevent losing any 3401 * event. 3402 */ 3403 worker_enter_idle(worker); 3404 __set_current_state(TASK_IDLE); 3405 raw_spin_unlock_irq(&pool->lock); 3406 schedule(); 3407 goto woke_up; 3408 } 3409 3410 /** 3411 * rescuer_thread - the rescuer thread function 3412 * @__rescuer: self 3413 * 3414 * Workqueue rescuer thread function. There's one rescuer for each 3415 * workqueue which has WQ_MEM_RECLAIM set. 3416 * 3417 * Regular work processing on a pool may block trying to create a new 3418 * worker which uses GFP_KERNEL allocation which has slight chance of 3419 * developing into deadlock if some works currently on the same queue 3420 * need to be processed to satisfy the GFP_KERNEL allocation. This is 3421 * the problem rescuer solves. 3422 * 3423 * When such condition is possible, the pool summons rescuers of all 3424 * workqueues which have works queued on the pool and let them process 3425 * those works so that forward progress can be guaranteed. 3426 * 3427 * This should happen rarely. 3428 * 3429 * Return: 0 3430 */ 3431 static int rescuer_thread(void *__rescuer) 3432 { 3433 struct worker *rescuer = __rescuer; 3434 struct workqueue_struct *wq = rescuer->rescue_wq; 3435 bool should_stop; 3436 3437 set_user_nice(current, RESCUER_NICE_LEVEL); 3438 3439 /* 3440 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 3441 * doesn't participate in concurrency management. 3442 */ 3443 set_pf_worker(true); 3444 repeat: 3445 set_current_state(TASK_IDLE); 3446 3447 /* 3448 * By the time the rescuer is requested to stop, the workqueue 3449 * shouldn't have any work pending, but @wq->maydays may still have 3450 * pwq(s) queued. This can happen by non-rescuer workers consuming 3451 * all the work items before the rescuer got to them. Go through 3452 * @wq->maydays processing before acting on should_stop so that the 3453 * list is always empty on exit. 3454 */ 3455 should_stop = kthread_should_stop(); 3456 3457 /* see whether any pwq is asking for help */ 3458 raw_spin_lock_irq(&wq_mayday_lock); 3459 3460 while (!list_empty(&wq->maydays)) { 3461 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 3462 struct pool_workqueue, mayday_node); 3463 struct worker_pool *pool = pwq->pool; 3464 struct work_struct *work, *n; 3465 3466 __set_current_state(TASK_RUNNING); 3467 list_del_init(&pwq->mayday_node); 3468 3469 raw_spin_unlock_irq(&wq_mayday_lock); 3470 3471 worker_attach_to_pool(rescuer, pool); 3472 3473 raw_spin_lock_irq(&pool->lock); 3474 3475 /* 3476 * Slurp in all works issued via this workqueue and 3477 * process'em. 3478 */ 3479 WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 3480 list_for_each_entry_safe(work, n, &pool->worklist, entry) { 3481 if (get_work_pwq(work) == pwq && 3482 assign_work(work, rescuer, &n)) 3483 pwq->stats[PWQ_STAT_RESCUED]++; 3484 } 3485 3486 if (!list_empty(&rescuer->scheduled)) { 3487 process_scheduled_works(rescuer); 3488 3489 /* 3490 * The above execution of rescued work items could 3491 * have created more to rescue through 3492 * pwq_activate_first_inactive() or chained 3493 * queueing. Let's put @pwq back on mayday list so 3494 * that such back-to-back work items, which may be 3495 * being used to relieve memory pressure, don't 3496 * incur MAYDAY_INTERVAL delay inbetween. 3497 */ 3498 if (pwq->nr_active && need_to_create_worker(pool)) { 3499 raw_spin_lock(&wq_mayday_lock); 3500 /* 3501 * Queue iff we aren't racing destruction 3502 * and somebody else hasn't queued it already. 3503 */ 3504 if (wq->rescuer && list_empty(&pwq->mayday_node)) { 3505 get_pwq(pwq); 3506 list_add_tail(&pwq->mayday_node, &wq->maydays); 3507 } 3508 raw_spin_unlock(&wq_mayday_lock); 3509 } 3510 } 3511 3512 /* 3513 * Put the reference grabbed by send_mayday(). @pool won't 3514 * go away while we're still attached to it. 3515 */ 3516 put_pwq(pwq); 3517 3518 /* 3519 * Leave this pool. Notify regular workers; otherwise, we end up 3520 * with 0 concurrency and stalling the execution. 3521 */ 3522 kick_pool(pool); 3523 3524 raw_spin_unlock_irq(&pool->lock); 3525 3526 worker_detach_from_pool(rescuer); 3527 3528 raw_spin_lock_irq(&wq_mayday_lock); 3529 } 3530 3531 raw_spin_unlock_irq(&wq_mayday_lock); 3532 3533 if (should_stop) { 3534 __set_current_state(TASK_RUNNING); 3535 set_pf_worker(false); 3536 return 0; 3537 } 3538 3539 /* rescuers should never participate in concurrency management */ 3540 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 3541 schedule(); 3542 goto repeat; 3543 } 3544 3545 static void bh_worker(struct worker *worker) 3546 { 3547 struct worker_pool *pool = worker->pool; 3548 int nr_restarts = BH_WORKER_RESTARTS; 3549 unsigned long end = jiffies + BH_WORKER_JIFFIES; 3550 3551 raw_spin_lock_irq(&pool->lock); 3552 worker_leave_idle(worker); 3553 3554 /* 3555 * This function follows the structure of worker_thread(). See there for 3556 * explanations on each step. 3557 */ 3558 if (!need_more_worker(pool)) 3559 goto done; 3560 3561 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3562 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3563 3564 do { 3565 struct work_struct *work = 3566 list_first_entry(&pool->worklist, 3567 struct work_struct, entry); 3568 3569 if (assign_work(work, worker, NULL)) 3570 process_scheduled_works(worker); 3571 } while (keep_working(pool) && 3572 --nr_restarts && time_before(jiffies, end)); 3573 3574 worker_set_flags(worker, WORKER_PREP); 3575 done: 3576 worker_enter_idle(worker); 3577 kick_pool(pool); 3578 raw_spin_unlock_irq(&pool->lock); 3579 } 3580 3581 /* 3582 * TODO: Convert all tasklet users to workqueue and use softirq directly. 3583 * 3584 * This is currently called from tasklet[_hi]action() and thus is also called 3585 * whenever there are tasklets to run. Let's do an early exit if there's nothing 3586 * queued. Once conversion from tasklet is complete, the need_more_worker() test 3587 * can be dropped. 3588 * 3589 * After full conversion, we'll add worker->softirq_action, directly use the 3590 * softirq action and obtain the worker pointer from the softirq_action pointer. 3591 */ 3592 void workqueue_softirq_action(bool highpri) 3593 { 3594 struct worker_pool *pool = 3595 &per_cpu(bh_worker_pools, smp_processor_id())[highpri]; 3596 if (need_more_worker(pool)) 3597 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3598 } 3599 3600 struct wq_drain_dead_softirq_work { 3601 struct work_struct work; 3602 struct worker_pool *pool; 3603 struct completion done; 3604 }; 3605 3606 static void drain_dead_softirq_workfn(struct work_struct *work) 3607 { 3608 struct wq_drain_dead_softirq_work *dead_work = 3609 container_of(work, struct wq_drain_dead_softirq_work, work); 3610 struct worker_pool *pool = dead_work->pool; 3611 bool repeat; 3612 3613 /* 3614 * @pool's CPU is dead and we want to execute its still pending work 3615 * items from this BH work item which is running on a different CPU. As 3616 * its CPU is dead, @pool can't be kicked and, as work execution path 3617 * will be nested, a lockdep annotation needs to be suppressed. Mark 3618 * @pool with %POOL_BH_DRAINING for the special treatments. 3619 */ 3620 raw_spin_lock_irq(&pool->lock); 3621 pool->flags |= POOL_BH_DRAINING; 3622 raw_spin_unlock_irq(&pool->lock); 3623 3624 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3625 3626 raw_spin_lock_irq(&pool->lock); 3627 pool->flags &= ~POOL_BH_DRAINING; 3628 repeat = need_more_worker(pool); 3629 raw_spin_unlock_irq(&pool->lock); 3630 3631 /* 3632 * bh_worker() might hit consecutive execution limit and bail. If there 3633 * still are pending work items, reschedule self and return so that we 3634 * don't hog this CPU's BH. 3635 */ 3636 if (repeat) { 3637 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3638 queue_work(system_bh_highpri_wq, work); 3639 else 3640 queue_work(system_bh_wq, work); 3641 } else { 3642 complete(&dead_work->done); 3643 } 3644 } 3645 3646 /* 3647 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's 3648 * possible to allocate dead_work per CPU and avoid flushing. However, then we 3649 * have to worry about draining overlapping with CPU coming back online or 3650 * nesting (one CPU's dead_work queued on another CPU which is also dead and so 3651 * on). Let's keep it simple and drain them synchronously. These are BH work 3652 * items which shouldn't be requeued on the same pool. Shouldn't take long. 3653 */ 3654 void workqueue_softirq_dead(unsigned int cpu) 3655 { 3656 int i; 3657 3658 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 3659 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i]; 3660 struct wq_drain_dead_softirq_work dead_work; 3661 3662 if (!need_more_worker(pool)) 3663 continue; 3664 3665 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn); 3666 dead_work.pool = pool; 3667 init_completion(&dead_work.done); 3668 3669 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3670 queue_work(system_bh_highpri_wq, &dead_work.work); 3671 else 3672 queue_work(system_bh_wq, &dead_work.work); 3673 3674 wait_for_completion(&dead_work.done); 3675 destroy_work_on_stack(&dead_work.work); 3676 } 3677 } 3678 3679 /** 3680 * check_flush_dependency - check for flush dependency sanity 3681 * @target_wq: workqueue being flushed 3682 * @target_work: work item being flushed (NULL for workqueue flushes) 3683 * @from_cancel: are we called from the work cancel path 3684 * 3685 * %current is trying to flush the whole @target_wq or @target_work on it. 3686 * If this is not the cancel path (which implies work being flushed is either 3687 * already running, or will not be at all), check if @target_wq doesn't have 3688 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running 3689 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward- 3690 * progress guarantee leading to a deadlock. 3691 */ 3692 static void check_flush_dependency(struct workqueue_struct *target_wq, 3693 struct work_struct *target_work, 3694 bool from_cancel) 3695 { 3696 work_func_t target_func; 3697 struct worker *worker; 3698 3699 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM) 3700 return; 3701 3702 worker = current_wq_worker(); 3703 target_func = target_work ? target_work->func : NULL; 3704 3705 WARN_ONCE(current->flags & PF_MEMALLOC, 3706 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 3707 current->pid, current->comm, target_wq->name, target_func); 3708 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 3709 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 3710 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 3711 worker->current_pwq->wq->name, worker->current_func, 3712 target_wq->name, target_func); 3713 } 3714 3715 struct wq_barrier { 3716 struct work_struct work; 3717 struct completion done; 3718 struct task_struct *task; /* purely informational */ 3719 }; 3720 3721 static void wq_barrier_func(struct work_struct *work) 3722 { 3723 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 3724 complete(&barr->done); 3725 } 3726 3727 /** 3728 * insert_wq_barrier - insert a barrier work 3729 * @pwq: pwq to insert barrier into 3730 * @barr: wq_barrier to insert 3731 * @target: target work to attach @barr to 3732 * @worker: worker currently executing @target, NULL if @target is not executing 3733 * 3734 * @barr is linked to @target such that @barr is completed only after 3735 * @target finishes execution. Please note that the ordering 3736 * guarantee is observed only with respect to @target and on the local 3737 * cpu. 3738 * 3739 * Currently, a queued barrier can't be canceled. This is because 3740 * try_to_grab_pending() can't determine whether the work to be 3741 * grabbed is at the head of the queue and thus can't clear LINKED 3742 * flag of the previous work while there must be a valid next work 3743 * after a work with LINKED flag set. 3744 * 3745 * Note that when @worker is non-NULL, @target may be modified 3746 * underneath us, so we can't reliably determine pwq from @target. 3747 * 3748 * CONTEXT: 3749 * raw_spin_lock_irq(pool->lock). 3750 */ 3751 static void insert_wq_barrier(struct pool_workqueue *pwq, 3752 struct wq_barrier *barr, 3753 struct work_struct *target, struct worker *worker) 3754 { 3755 static __maybe_unused struct lock_class_key bh_key, thr_key; 3756 unsigned int work_flags = 0; 3757 unsigned int work_color; 3758 struct list_head *head; 3759 3760 /* 3761 * debugobject calls are safe here even with pool->lock locked 3762 * as we know for sure that this will not trigger any of the 3763 * checks and call back into the fixup functions where we 3764 * might deadlock. 3765 * 3766 * BH and threaded workqueues need separate lockdep keys to avoid 3767 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} 3768 * usage". 3769 */ 3770 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func, 3771 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key); 3772 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 3773 3774 init_completion_map(&barr->done, &target->lockdep_map); 3775 3776 barr->task = current; 3777 3778 /* The barrier work item does not participate in nr_active. */ 3779 work_flags |= WORK_STRUCT_INACTIVE; 3780 3781 /* 3782 * If @target is currently being executed, schedule the 3783 * barrier to the worker; otherwise, put it after @target. 3784 */ 3785 if (worker) { 3786 head = worker->scheduled.next; 3787 work_color = worker->current_color; 3788 } else { 3789 unsigned long *bits = work_data_bits(target); 3790 3791 head = target->entry.next; 3792 /* there can already be other linked works, inherit and set */ 3793 work_flags |= *bits & WORK_STRUCT_LINKED; 3794 work_color = get_work_color(*bits); 3795 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3796 } 3797 3798 pwq->nr_in_flight[work_color]++; 3799 work_flags |= work_color_to_flags(work_color); 3800 3801 insert_work(pwq, &barr->work, head, work_flags); 3802 } 3803 3804 /** 3805 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 3806 * @wq: workqueue being flushed 3807 * @flush_color: new flush color, < 0 for no-op 3808 * @work_color: new work color, < 0 for no-op 3809 * 3810 * Prepare pwqs for workqueue flushing. 3811 * 3812 * If @flush_color is non-negative, flush_color on all pwqs should be 3813 * -1. If no pwq has in-flight commands at the specified color, all 3814 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3815 * has in flight commands, its pwq->flush_color is set to 3816 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 3817 * wakeup logic is armed and %true is returned. 3818 * 3819 * The caller should have initialized @wq->first_flusher prior to 3820 * calling this function with non-negative @flush_color. If 3821 * @flush_color is negative, no flush color update is done and %false 3822 * is returned. 3823 * 3824 * If @work_color is non-negative, all pwqs should have the same 3825 * work_color which is previous to @work_color and all will be 3826 * advanced to @work_color. 3827 * 3828 * CONTEXT: 3829 * mutex_lock(wq->mutex). 3830 * 3831 * Return: 3832 * %true if @flush_color >= 0 and there's something to flush. %false 3833 * otherwise. 3834 */ 3835 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 3836 int flush_color, int work_color) 3837 { 3838 bool wait = false; 3839 struct pool_workqueue *pwq; 3840 struct worker_pool *current_pool = NULL; 3841 3842 if (flush_color >= 0) { 3843 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3844 atomic_set(&wq->nr_pwqs_to_flush, 1); 3845 } 3846 3847 /* 3848 * For unbound workqueue, pwqs will map to only a few pools. 3849 * Most of the time, pwqs within the same pool will be linked 3850 * sequentially to wq->pwqs by cpu index. So in the majority 3851 * of pwq iters, the pool is the same, only doing lock/unlock 3852 * if the pool has changed. This can largely reduce expensive 3853 * lock operations. 3854 */ 3855 for_each_pwq(pwq, wq) { 3856 if (current_pool != pwq->pool) { 3857 if (likely(current_pool)) 3858 raw_spin_unlock_irq(¤t_pool->lock); 3859 current_pool = pwq->pool; 3860 raw_spin_lock_irq(¤t_pool->lock); 3861 } 3862 3863 if (flush_color >= 0) { 3864 WARN_ON_ONCE(pwq->flush_color != -1); 3865 3866 if (pwq->nr_in_flight[flush_color]) { 3867 pwq->flush_color = flush_color; 3868 atomic_inc(&wq->nr_pwqs_to_flush); 3869 wait = true; 3870 } 3871 } 3872 3873 if (work_color >= 0) { 3874 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3875 pwq->work_color = work_color; 3876 } 3877 3878 } 3879 3880 if (current_pool) 3881 raw_spin_unlock_irq(¤t_pool->lock); 3882 3883 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 3884 complete(&wq->first_flusher->done); 3885 3886 return wait; 3887 } 3888 3889 static void touch_wq_lockdep_map(struct workqueue_struct *wq) 3890 { 3891 #ifdef CONFIG_LOCKDEP 3892 if (unlikely(!wq->lockdep_map)) 3893 return; 3894 3895 if (wq->flags & WQ_BH) 3896 local_bh_disable(); 3897 3898 lock_map_acquire(wq->lockdep_map); 3899 lock_map_release(wq->lockdep_map); 3900 3901 if (wq->flags & WQ_BH) 3902 local_bh_enable(); 3903 #endif 3904 } 3905 3906 static void touch_work_lockdep_map(struct work_struct *work, 3907 struct workqueue_struct *wq) 3908 { 3909 #ifdef CONFIG_LOCKDEP 3910 if (wq->flags & WQ_BH) 3911 local_bh_disable(); 3912 3913 lock_map_acquire(&work->lockdep_map); 3914 lock_map_release(&work->lockdep_map); 3915 3916 if (wq->flags & WQ_BH) 3917 local_bh_enable(); 3918 #endif 3919 } 3920 3921 /** 3922 * __flush_workqueue - ensure that any scheduled work has run to completion. 3923 * @wq: workqueue to flush 3924 * 3925 * This function sleeps until all work items which were queued on entry 3926 * have finished execution, but it is not livelocked by new incoming ones. 3927 */ 3928 void __flush_workqueue(struct workqueue_struct *wq) 3929 { 3930 struct wq_flusher this_flusher = { 3931 .list = LIST_HEAD_INIT(this_flusher.list), 3932 .flush_color = -1, 3933 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)), 3934 }; 3935 int next_color; 3936 3937 if (WARN_ON(!wq_online)) 3938 return; 3939 3940 touch_wq_lockdep_map(wq); 3941 3942 mutex_lock(&wq->mutex); 3943 3944 /* 3945 * Start-to-wait phase 3946 */ 3947 next_color = work_next_color(wq->work_color); 3948 3949 if (next_color != wq->flush_color) { 3950 /* 3951 * Color space is not full. The current work_color 3952 * becomes our flush_color and work_color is advanced 3953 * by one. 3954 */ 3955 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 3956 this_flusher.flush_color = wq->work_color; 3957 wq->work_color = next_color; 3958 3959 if (!wq->first_flusher) { 3960 /* no flush in progress, become the first flusher */ 3961 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 3962 3963 wq->first_flusher = &this_flusher; 3964 3965 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 3966 wq->work_color)) { 3967 /* nothing to flush, done */ 3968 wq->flush_color = next_color; 3969 wq->first_flusher = NULL; 3970 goto out_unlock; 3971 } 3972 } else { 3973 /* wait in queue */ 3974 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 3975 list_add_tail(&this_flusher.list, &wq->flusher_queue); 3976 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 3977 } 3978 } else { 3979 /* 3980 * Oops, color space is full, wait on overflow queue. 3981 * The next flush completion will assign us 3982 * flush_color and transfer to flusher_queue. 3983 */ 3984 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 3985 } 3986 3987 check_flush_dependency(wq, NULL, false); 3988 3989 mutex_unlock(&wq->mutex); 3990 3991 wait_for_completion(&this_flusher.done); 3992 3993 /* 3994 * Wake-up-and-cascade phase 3995 * 3996 * First flushers are responsible for cascading flushes and 3997 * handling overflow. Non-first flushers can simply return. 3998 */ 3999 if (READ_ONCE(wq->first_flusher) != &this_flusher) 4000 return; 4001 4002 mutex_lock(&wq->mutex); 4003 4004 /* we might have raced, check again with mutex held */ 4005 if (wq->first_flusher != &this_flusher) 4006 goto out_unlock; 4007 4008 WRITE_ONCE(wq->first_flusher, NULL); 4009 4010 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 4011 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 4012 4013 while (true) { 4014 struct wq_flusher *next, *tmp; 4015 4016 /* complete all the flushers sharing the current flush color */ 4017 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 4018 if (next->flush_color != wq->flush_color) 4019 break; 4020 list_del_init(&next->list); 4021 complete(&next->done); 4022 } 4023 4024 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 4025 wq->flush_color != work_next_color(wq->work_color)); 4026 4027 /* this flush_color is finished, advance by one */ 4028 wq->flush_color = work_next_color(wq->flush_color); 4029 4030 /* one color has been freed, handle overflow queue */ 4031 if (!list_empty(&wq->flusher_overflow)) { 4032 /* 4033 * Assign the same color to all overflowed 4034 * flushers, advance work_color and append to 4035 * flusher_queue. This is the start-to-wait 4036 * phase for these overflowed flushers. 4037 */ 4038 list_for_each_entry(tmp, &wq->flusher_overflow, list) 4039 tmp->flush_color = wq->work_color; 4040 4041 wq->work_color = work_next_color(wq->work_color); 4042 4043 list_splice_tail_init(&wq->flusher_overflow, 4044 &wq->flusher_queue); 4045 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 4046 } 4047 4048 if (list_empty(&wq->flusher_queue)) { 4049 WARN_ON_ONCE(wq->flush_color != wq->work_color); 4050 break; 4051 } 4052 4053 /* 4054 * Need to flush more colors. Make the next flusher 4055 * the new first flusher and arm pwqs. 4056 */ 4057 WARN_ON_ONCE(wq->flush_color == wq->work_color); 4058 WARN_ON_ONCE(wq->flush_color != next->flush_color); 4059 4060 list_del_init(&next->list); 4061 wq->first_flusher = next; 4062 4063 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 4064 break; 4065 4066 /* 4067 * Meh... this color is already done, clear first 4068 * flusher and repeat cascading. 4069 */ 4070 wq->first_flusher = NULL; 4071 } 4072 4073 out_unlock: 4074 mutex_unlock(&wq->mutex); 4075 } 4076 EXPORT_SYMBOL(__flush_workqueue); 4077 4078 /** 4079 * drain_workqueue - drain a workqueue 4080 * @wq: workqueue to drain 4081 * 4082 * Wait until the workqueue becomes empty. While draining is in progress, 4083 * only chain queueing is allowed. IOW, only currently pending or running 4084 * work items on @wq can queue further work items on it. @wq is flushed 4085 * repeatedly until it becomes empty. The number of flushing is determined 4086 * by the depth of chaining and should be relatively short. Whine if it 4087 * takes too long. 4088 */ 4089 void drain_workqueue(struct workqueue_struct *wq) 4090 { 4091 unsigned int flush_cnt = 0; 4092 struct pool_workqueue *pwq; 4093 4094 /* 4095 * __queue_work() needs to test whether there are drainers, is much 4096 * hotter than drain_workqueue() and already looks at @wq->flags. 4097 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 4098 */ 4099 mutex_lock(&wq->mutex); 4100 if (!wq->nr_drainers++) 4101 wq->flags |= __WQ_DRAINING; 4102 mutex_unlock(&wq->mutex); 4103 reflush: 4104 __flush_workqueue(wq); 4105 4106 mutex_lock(&wq->mutex); 4107 4108 for_each_pwq(pwq, wq) { 4109 bool drained; 4110 4111 raw_spin_lock_irq(&pwq->pool->lock); 4112 drained = pwq_is_empty(pwq); 4113 raw_spin_unlock_irq(&pwq->pool->lock); 4114 4115 if (drained) 4116 continue; 4117 4118 if (++flush_cnt == 10 || 4119 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 4120 pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 4121 wq->name, __func__, flush_cnt); 4122 4123 mutex_unlock(&wq->mutex); 4124 goto reflush; 4125 } 4126 4127 if (!--wq->nr_drainers) 4128 wq->flags &= ~__WQ_DRAINING; 4129 mutex_unlock(&wq->mutex); 4130 } 4131 EXPORT_SYMBOL_GPL(drain_workqueue); 4132 4133 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 4134 bool from_cancel) 4135 { 4136 struct worker *worker = NULL; 4137 struct worker_pool *pool; 4138 struct pool_workqueue *pwq; 4139 struct workqueue_struct *wq; 4140 4141 rcu_read_lock(); 4142 pool = get_work_pool(work); 4143 if (!pool) { 4144 rcu_read_unlock(); 4145 return false; 4146 } 4147 4148 raw_spin_lock_irq(&pool->lock); 4149 /* see the comment in try_to_grab_pending() with the same code */ 4150 pwq = get_work_pwq(work); 4151 if (pwq) { 4152 if (unlikely(pwq->pool != pool)) 4153 goto already_gone; 4154 } else { 4155 worker = find_worker_executing_work(pool, work); 4156 if (!worker) 4157 goto already_gone; 4158 pwq = worker->current_pwq; 4159 } 4160 4161 wq = pwq->wq; 4162 check_flush_dependency(wq, work, from_cancel); 4163 4164 insert_wq_barrier(pwq, barr, work, worker); 4165 raw_spin_unlock_irq(&pool->lock); 4166 4167 touch_work_lockdep_map(work, wq); 4168 4169 /* 4170 * Force a lock recursion deadlock when using flush_work() inside a 4171 * single-threaded or rescuer equipped workqueue. 4172 * 4173 * For single threaded workqueues the deadlock happens when the work 4174 * is after the work issuing the flush_work(). For rescuer equipped 4175 * workqueues the deadlock happens when the rescuer stalls, blocking 4176 * forward progress. 4177 */ 4178 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer)) 4179 touch_wq_lockdep_map(wq); 4180 4181 rcu_read_unlock(); 4182 return true; 4183 already_gone: 4184 raw_spin_unlock_irq(&pool->lock); 4185 rcu_read_unlock(); 4186 return false; 4187 } 4188 4189 static bool __flush_work(struct work_struct *work, bool from_cancel) 4190 { 4191 struct wq_barrier barr; 4192 4193 if (WARN_ON(!wq_online)) 4194 return false; 4195 4196 if (WARN_ON(!work->func)) 4197 return false; 4198 4199 if (!start_flush_work(work, &barr, from_cancel)) 4200 return false; 4201 4202 /* 4203 * start_flush_work() returned %true. If @from_cancel is set, we know 4204 * that @work must have been executing during start_flush_work() and 4205 * can't currently be queued. Its data must contain OFFQ bits. If @work 4206 * was queued on a BH workqueue, we also know that it was running in the 4207 * BH context and thus can be busy-waited. 4208 */ 4209 if (from_cancel) { 4210 unsigned long data = *work_data_bits(work); 4211 4212 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && 4213 (data & WORK_OFFQ_BH)) { 4214 /* 4215 * On RT, prevent a live lock when %current preempted 4216 * soft interrupt processing or prevents ksoftirqd from 4217 * running by keeping flipping BH. If the BH work item 4218 * runs on a different CPU then this has no effect other 4219 * than doing the BH disable/enable dance for nothing. 4220 * This is copied from 4221 * kernel/softirq.c::tasklet_unlock_spin_wait(). 4222 */ 4223 while (!try_wait_for_completion(&barr.done)) { 4224 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 4225 local_bh_disable(); 4226 local_bh_enable(); 4227 } else { 4228 cpu_relax(); 4229 } 4230 } 4231 goto out_destroy; 4232 } 4233 } 4234 4235 wait_for_completion(&barr.done); 4236 4237 out_destroy: 4238 destroy_work_on_stack(&barr.work); 4239 return true; 4240 } 4241 4242 /** 4243 * flush_work - wait for a work to finish executing the last queueing instance 4244 * @work: the work to flush 4245 * 4246 * Wait until @work has finished execution. @work is guaranteed to be idle 4247 * on return if it hasn't been requeued since flush started. 4248 * 4249 * Return: 4250 * %true if flush_work() waited for the work to finish execution, 4251 * %false if it was already idle. 4252 */ 4253 bool flush_work(struct work_struct *work) 4254 { 4255 might_sleep(); 4256 return __flush_work(work, false); 4257 } 4258 EXPORT_SYMBOL_GPL(flush_work); 4259 4260 /** 4261 * flush_delayed_work - wait for a dwork to finish executing the last queueing 4262 * @dwork: the delayed work to flush 4263 * 4264 * Delayed timer is cancelled and the pending work is queued for 4265 * immediate execution. Like flush_work(), this function only 4266 * considers the last queueing instance of @dwork. 4267 * 4268 * Return: 4269 * %true if flush_work() waited for the work to finish execution, 4270 * %false if it was already idle. 4271 */ 4272 bool flush_delayed_work(struct delayed_work *dwork) 4273 { 4274 local_irq_disable(); 4275 if (del_timer_sync(&dwork->timer)) 4276 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 4277 local_irq_enable(); 4278 return flush_work(&dwork->work); 4279 } 4280 EXPORT_SYMBOL(flush_delayed_work); 4281 4282 /** 4283 * flush_rcu_work - wait for a rwork to finish executing the last queueing 4284 * @rwork: the rcu work to flush 4285 * 4286 * Return: 4287 * %true if flush_rcu_work() waited for the work to finish execution, 4288 * %false if it was already idle. 4289 */ 4290 bool flush_rcu_work(struct rcu_work *rwork) 4291 { 4292 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 4293 rcu_barrier(); 4294 flush_work(&rwork->work); 4295 return true; 4296 } else { 4297 return flush_work(&rwork->work); 4298 } 4299 } 4300 EXPORT_SYMBOL(flush_rcu_work); 4301 4302 static void work_offqd_disable(struct work_offq_data *offqd) 4303 { 4304 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1; 4305 4306 if (likely(offqd->disable < max)) 4307 offqd->disable++; 4308 else 4309 WARN_ONCE(true, "workqueue: work disable count overflowed\n"); 4310 } 4311 4312 static void work_offqd_enable(struct work_offq_data *offqd) 4313 { 4314 if (likely(offqd->disable > 0)) 4315 offqd->disable--; 4316 else 4317 WARN_ONCE(true, "workqueue: work disable count underflowed\n"); 4318 } 4319 4320 static bool __cancel_work(struct work_struct *work, u32 cflags) 4321 { 4322 struct work_offq_data offqd; 4323 unsigned long irq_flags; 4324 int ret; 4325 4326 ret = work_grab_pending(work, cflags, &irq_flags); 4327 4328 work_offqd_unpack(&offqd, *work_data_bits(work)); 4329 4330 if (cflags & WORK_CANCEL_DISABLE) 4331 work_offqd_disable(&offqd); 4332 4333 set_work_pool_and_clear_pending(work, offqd.pool_id, 4334 work_offqd_pack_flags(&offqd)); 4335 local_irq_restore(irq_flags); 4336 return ret; 4337 } 4338 4339 static bool __cancel_work_sync(struct work_struct *work, u32 cflags) 4340 { 4341 bool ret; 4342 4343 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE); 4344 4345 if (*work_data_bits(work) & WORK_OFFQ_BH) 4346 WARN_ON_ONCE(in_hardirq()); 4347 else 4348 might_sleep(); 4349 4350 /* 4351 * Skip __flush_work() during early boot when we know that @work isn't 4352 * executing. This allows canceling during early boot. 4353 */ 4354 if (wq_online) 4355 __flush_work(work, true); 4356 4357 if (!(cflags & WORK_CANCEL_DISABLE)) 4358 enable_work(work); 4359 4360 return ret; 4361 } 4362 4363 /* 4364 * See cancel_delayed_work() 4365 */ 4366 bool cancel_work(struct work_struct *work) 4367 { 4368 return __cancel_work(work, 0); 4369 } 4370 EXPORT_SYMBOL(cancel_work); 4371 4372 /** 4373 * cancel_work_sync - cancel a work and wait for it to finish 4374 * @work: the work to cancel 4375 * 4376 * Cancel @work and wait for its execution to finish. This function can be used 4377 * even if the work re-queues itself or migrates to another workqueue. On return 4378 * from this function, @work is guaranteed to be not pending or executing on any 4379 * CPU as long as there aren't racing enqueues. 4380 * 4381 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's. 4382 * Use cancel_delayed_work_sync() instead. 4383 * 4384 * Must be called from a sleepable context if @work was last queued on a non-BH 4385 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4386 * if @work was last queued on a BH workqueue. 4387 * 4388 * Returns %true if @work was pending, %false otherwise. 4389 */ 4390 bool cancel_work_sync(struct work_struct *work) 4391 { 4392 return __cancel_work_sync(work, 0); 4393 } 4394 EXPORT_SYMBOL_GPL(cancel_work_sync); 4395 4396 /** 4397 * cancel_delayed_work - cancel a delayed work 4398 * @dwork: delayed_work to cancel 4399 * 4400 * Kill off a pending delayed_work. 4401 * 4402 * Return: %true if @dwork was pending and canceled; %false if it wasn't 4403 * pending. 4404 * 4405 * Note: 4406 * The work callback function may still be running on return, unless 4407 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 4408 * use cancel_delayed_work_sync() to wait on it. 4409 * 4410 * This function is safe to call from any context including IRQ handler. 4411 */ 4412 bool cancel_delayed_work(struct delayed_work *dwork) 4413 { 4414 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED); 4415 } 4416 EXPORT_SYMBOL(cancel_delayed_work); 4417 4418 /** 4419 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 4420 * @dwork: the delayed work cancel 4421 * 4422 * This is cancel_work_sync() for delayed works. 4423 * 4424 * Return: 4425 * %true if @dwork was pending, %false otherwise. 4426 */ 4427 bool cancel_delayed_work_sync(struct delayed_work *dwork) 4428 { 4429 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED); 4430 } 4431 EXPORT_SYMBOL(cancel_delayed_work_sync); 4432 4433 /** 4434 * disable_work - Disable and cancel a work item 4435 * @work: work item to disable 4436 * 4437 * Disable @work by incrementing its disable count and cancel it if currently 4438 * pending. As long as the disable count is non-zero, any attempt to queue @work 4439 * will fail and return %false. The maximum supported disable depth is 2 to the 4440 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536. 4441 * 4442 * Can be called from any context. Returns %true if @work was pending, %false 4443 * otherwise. 4444 */ 4445 bool disable_work(struct work_struct *work) 4446 { 4447 return __cancel_work(work, WORK_CANCEL_DISABLE); 4448 } 4449 EXPORT_SYMBOL_GPL(disable_work); 4450 4451 /** 4452 * disable_work_sync - Disable, cancel and drain a work item 4453 * @work: work item to disable 4454 * 4455 * Similar to disable_work() but also wait for @work to finish if currently 4456 * executing. 4457 * 4458 * Must be called from a sleepable context if @work was last queued on a non-BH 4459 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4460 * if @work was last queued on a BH workqueue. 4461 * 4462 * Returns %true if @work was pending, %false otherwise. 4463 */ 4464 bool disable_work_sync(struct work_struct *work) 4465 { 4466 return __cancel_work_sync(work, WORK_CANCEL_DISABLE); 4467 } 4468 EXPORT_SYMBOL_GPL(disable_work_sync); 4469 4470 /** 4471 * enable_work - Enable a work item 4472 * @work: work item to enable 4473 * 4474 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can 4475 * only be queued if its disable count is 0. 4476 * 4477 * Can be called from any context. Returns %true if the disable count reached 0. 4478 * Otherwise, %false. 4479 */ 4480 bool enable_work(struct work_struct *work) 4481 { 4482 struct work_offq_data offqd; 4483 unsigned long irq_flags; 4484 4485 work_grab_pending(work, 0, &irq_flags); 4486 4487 work_offqd_unpack(&offqd, *work_data_bits(work)); 4488 work_offqd_enable(&offqd); 4489 set_work_pool_and_clear_pending(work, offqd.pool_id, 4490 work_offqd_pack_flags(&offqd)); 4491 local_irq_restore(irq_flags); 4492 4493 return !offqd.disable; 4494 } 4495 EXPORT_SYMBOL_GPL(enable_work); 4496 4497 /** 4498 * disable_delayed_work - Disable and cancel a delayed work item 4499 * @dwork: delayed work item to disable 4500 * 4501 * disable_work() for delayed work items. 4502 */ 4503 bool disable_delayed_work(struct delayed_work *dwork) 4504 { 4505 return __cancel_work(&dwork->work, 4506 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4507 } 4508 EXPORT_SYMBOL_GPL(disable_delayed_work); 4509 4510 /** 4511 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item 4512 * @dwork: delayed work item to disable 4513 * 4514 * disable_work_sync() for delayed work items. 4515 */ 4516 bool disable_delayed_work_sync(struct delayed_work *dwork) 4517 { 4518 return __cancel_work_sync(&dwork->work, 4519 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4520 } 4521 EXPORT_SYMBOL_GPL(disable_delayed_work_sync); 4522 4523 /** 4524 * enable_delayed_work - Enable a delayed work item 4525 * @dwork: delayed work item to enable 4526 * 4527 * enable_work() for delayed work items. 4528 */ 4529 bool enable_delayed_work(struct delayed_work *dwork) 4530 { 4531 return enable_work(&dwork->work); 4532 } 4533 EXPORT_SYMBOL_GPL(enable_delayed_work); 4534 4535 /** 4536 * schedule_on_each_cpu - execute a function synchronously on each online CPU 4537 * @func: the function to call 4538 * 4539 * schedule_on_each_cpu() executes @func on each online CPU using the 4540 * system workqueue and blocks until all CPUs have completed. 4541 * schedule_on_each_cpu() is very slow. 4542 * 4543 * Return: 4544 * 0 on success, -errno on failure. 4545 */ 4546 int schedule_on_each_cpu(work_func_t func) 4547 { 4548 int cpu; 4549 struct work_struct __percpu *works; 4550 4551 works = alloc_percpu(struct work_struct); 4552 if (!works) 4553 return -ENOMEM; 4554 4555 cpus_read_lock(); 4556 4557 for_each_online_cpu(cpu) { 4558 struct work_struct *work = per_cpu_ptr(works, cpu); 4559 4560 INIT_WORK(work, func); 4561 schedule_work_on(cpu, work); 4562 } 4563 4564 for_each_online_cpu(cpu) 4565 flush_work(per_cpu_ptr(works, cpu)); 4566 4567 cpus_read_unlock(); 4568 free_percpu(works); 4569 return 0; 4570 } 4571 4572 /** 4573 * execute_in_process_context - reliably execute the routine with user context 4574 * @fn: the function to execute 4575 * @ew: guaranteed storage for the execute work structure (must 4576 * be available when the work executes) 4577 * 4578 * Executes the function immediately if process context is available, 4579 * otherwise schedules the function for delayed execution. 4580 * 4581 * Return: 0 - function was executed 4582 * 1 - function was scheduled for execution 4583 */ 4584 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 4585 { 4586 if (!in_interrupt()) { 4587 fn(&ew->work); 4588 return 0; 4589 } 4590 4591 INIT_WORK(&ew->work, fn); 4592 schedule_work(&ew->work); 4593 4594 return 1; 4595 } 4596 EXPORT_SYMBOL_GPL(execute_in_process_context); 4597 4598 /** 4599 * free_workqueue_attrs - free a workqueue_attrs 4600 * @attrs: workqueue_attrs to free 4601 * 4602 * Undo alloc_workqueue_attrs(). 4603 */ 4604 void free_workqueue_attrs(struct workqueue_attrs *attrs) 4605 { 4606 if (attrs) { 4607 free_cpumask_var(attrs->cpumask); 4608 free_cpumask_var(attrs->__pod_cpumask); 4609 kfree(attrs); 4610 } 4611 } 4612 4613 /** 4614 * alloc_workqueue_attrs - allocate a workqueue_attrs 4615 * 4616 * Allocate a new workqueue_attrs, initialize with default settings and 4617 * return it. 4618 * 4619 * Return: The allocated new workqueue_attr on success. %NULL on failure. 4620 */ 4621 struct workqueue_attrs *alloc_workqueue_attrs(void) 4622 { 4623 struct workqueue_attrs *attrs; 4624 4625 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 4626 if (!attrs) 4627 goto fail; 4628 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 4629 goto fail; 4630 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 4631 goto fail; 4632 4633 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4634 attrs->affn_scope = WQ_AFFN_DFL; 4635 return attrs; 4636 fail: 4637 free_workqueue_attrs(attrs); 4638 return NULL; 4639 } 4640 4641 static void copy_workqueue_attrs(struct workqueue_attrs *to, 4642 const struct workqueue_attrs *from) 4643 { 4644 to->nice = from->nice; 4645 cpumask_copy(to->cpumask, from->cpumask); 4646 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 4647 to->affn_strict = from->affn_strict; 4648 4649 /* 4650 * Unlike hash and equality test, copying shouldn't ignore wq-only 4651 * fields as copying is used for both pool and wq attrs. Instead, 4652 * get_unbound_pool() explicitly clears the fields. 4653 */ 4654 to->affn_scope = from->affn_scope; 4655 to->ordered = from->ordered; 4656 } 4657 4658 /* 4659 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 4660 * comments in 'struct workqueue_attrs' definition. 4661 */ 4662 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 4663 { 4664 attrs->affn_scope = WQ_AFFN_NR_TYPES; 4665 attrs->ordered = false; 4666 if (attrs->affn_strict) 4667 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4668 } 4669 4670 /* hash value of the content of @attr */ 4671 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 4672 { 4673 u32 hash = 0; 4674 4675 hash = jhash_1word(attrs->nice, hash); 4676 hash = jhash_1word(attrs->affn_strict, hash); 4677 hash = jhash(cpumask_bits(attrs->__pod_cpumask), 4678 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4679 if (!attrs->affn_strict) 4680 hash = jhash(cpumask_bits(attrs->cpumask), 4681 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4682 return hash; 4683 } 4684 4685 /* content equality test */ 4686 static bool wqattrs_equal(const struct workqueue_attrs *a, 4687 const struct workqueue_attrs *b) 4688 { 4689 if (a->nice != b->nice) 4690 return false; 4691 if (a->affn_strict != b->affn_strict) 4692 return false; 4693 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 4694 return false; 4695 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask)) 4696 return false; 4697 return true; 4698 } 4699 4700 /* Update @attrs with actually available CPUs */ 4701 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 4702 const cpumask_t *unbound_cpumask) 4703 { 4704 /* 4705 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 4706 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 4707 * @unbound_cpumask. 4708 */ 4709 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 4710 if (unlikely(cpumask_empty(attrs->cpumask))) 4711 cpumask_copy(attrs->cpumask, unbound_cpumask); 4712 } 4713 4714 /* find wq_pod_type to use for @attrs */ 4715 static const struct wq_pod_type * 4716 wqattrs_pod_type(const struct workqueue_attrs *attrs) 4717 { 4718 enum wq_affn_scope scope; 4719 struct wq_pod_type *pt; 4720 4721 /* to synchronize access to wq_affn_dfl */ 4722 lockdep_assert_held(&wq_pool_mutex); 4723 4724 if (attrs->affn_scope == WQ_AFFN_DFL) 4725 scope = wq_affn_dfl; 4726 else 4727 scope = attrs->affn_scope; 4728 4729 pt = &wq_pod_types[scope]; 4730 4731 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 4732 likely(pt->nr_pods)) 4733 return pt; 4734 4735 /* 4736 * Before workqueue_init_topology(), only SYSTEM is available which is 4737 * initialized in workqueue_init_early(). 4738 */ 4739 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 4740 BUG_ON(!pt->nr_pods); 4741 return pt; 4742 } 4743 4744 /** 4745 * init_worker_pool - initialize a newly zalloc'd worker_pool 4746 * @pool: worker_pool to initialize 4747 * 4748 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 4749 * 4750 * Return: 0 on success, -errno on failure. Even on failure, all fields 4751 * inside @pool proper are initialized and put_unbound_pool() can be called 4752 * on @pool safely to release it. 4753 */ 4754 static int init_worker_pool(struct worker_pool *pool) 4755 { 4756 raw_spin_lock_init(&pool->lock); 4757 pool->id = -1; 4758 pool->cpu = -1; 4759 pool->node = NUMA_NO_NODE; 4760 pool->flags |= POOL_DISASSOCIATED; 4761 pool->watchdog_ts = jiffies; 4762 INIT_LIST_HEAD(&pool->worklist); 4763 INIT_LIST_HEAD(&pool->idle_list); 4764 hash_init(pool->busy_hash); 4765 4766 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 4767 INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 4768 4769 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 4770 4771 INIT_LIST_HEAD(&pool->workers); 4772 4773 ida_init(&pool->worker_ida); 4774 INIT_HLIST_NODE(&pool->hash_node); 4775 pool->refcnt = 1; 4776 4777 /* shouldn't fail above this point */ 4778 pool->attrs = alloc_workqueue_attrs(); 4779 if (!pool->attrs) 4780 return -ENOMEM; 4781 4782 wqattrs_clear_for_pool(pool->attrs); 4783 4784 return 0; 4785 } 4786 4787 #ifdef CONFIG_LOCKDEP 4788 static void wq_init_lockdep(struct workqueue_struct *wq) 4789 { 4790 char *lock_name; 4791 4792 lockdep_register_key(&wq->key); 4793 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 4794 if (!lock_name) 4795 lock_name = wq->name; 4796 4797 wq->lock_name = lock_name; 4798 wq->lockdep_map = &wq->__lockdep_map; 4799 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0); 4800 } 4801 4802 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4803 { 4804 if (wq->lockdep_map != &wq->__lockdep_map) 4805 return; 4806 4807 lockdep_unregister_key(&wq->key); 4808 } 4809 4810 static void wq_free_lockdep(struct workqueue_struct *wq) 4811 { 4812 if (wq->lockdep_map != &wq->__lockdep_map) 4813 return; 4814 4815 if (wq->lock_name != wq->name) 4816 kfree(wq->lock_name); 4817 } 4818 #else 4819 static void wq_init_lockdep(struct workqueue_struct *wq) 4820 { 4821 } 4822 4823 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4824 { 4825 } 4826 4827 static void wq_free_lockdep(struct workqueue_struct *wq) 4828 { 4829 } 4830 #endif 4831 4832 static void free_node_nr_active(struct wq_node_nr_active **nna_ar) 4833 { 4834 int node; 4835 4836 for_each_node(node) { 4837 kfree(nna_ar[node]); 4838 nna_ar[node] = NULL; 4839 } 4840 4841 kfree(nna_ar[nr_node_ids]); 4842 nna_ar[nr_node_ids] = NULL; 4843 } 4844 4845 static void init_node_nr_active(struct wq_node_nr_active *nna) 4846 { 4847 nna->max = WQ_DFL_MIN_ACTIVE; 4848 atomic_set(&nna->nr, 0); 4849 raw_spin_lock_init(&nna->lock); 4850 INIT_LIST_HEAD(&nna->pending_pwqs); 4851 } 4852 4853 /* 4854 * Each node's nr_active counter will be accessed mostly from its own node and 4855 * should be allocated in the node. 4856 */ 4857 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) 4858 { 4859 struct wq_node_nr_active *nna; 4860 int node; 4861 4862 for_each_node(node) { 4863 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); 4864 if (!nna) 4865 goto err_free; 4866 init_node_nr_active(nna); 4867 nna_ar[node] = nna; 4868 } 4869 4870 /* [nr_node_ids] is used as the fallback */ 4871 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); 4872 if (!nna) 4873 goto err_free; 4874 init_node_nr_active(nna); 4875 nna_ar[nr_node_ids] = nna; 4876 4877 return 0; 4878 4879 err_free: 4880 free_node_nr_active(nna_ar); 4881 return -ENOMEM; 4882 } 4883 4884 static void rcu_free_wq(struct rcu_head *rcu) 4885 { 4886 struct workqueue_struct *wq = 4887 container_of(rcu, struct workqueue_struct, rcu); 4888 4889 if (wq->flags & WQ_UNBOUND) 4890 free_node_nr_active(wq->node_nr_active); 4891 4892 wq_free_lockdep(wq); 4893 free_percpu(wq->cpu_pwq); 4894 free_workqueue_attrs(wq->unbound_attrs); 4895 kfree(wq); 4896 } 4897 4898 static void rcu_free_pool(struct rcu_head *rcu) 4899 { 4900 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 4901 4902 ida_destroy(&pool->worker_ida); 4903 free_workqueue_attrs(pool->attrs); 4904 kfree(pool); 4905 } 4906 4907 /** 4908 * put_unbound_pool - put a worker_pool 4909 * @pool: worker_pool to put 4910 * 4911 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 4912 * safe manner. get_unbound_pool() calls this function on its failure path 4913 * and this function should be able to release pools which went through, 4914 * successfully or not, init_worker_pool(). 4915 * 4916 * Should be called with wq_pool_mutex held. 4917 */ 4918 static void put_unbound_pool(struct worker_pool *pool) 4919 { 4920 struct worker *worker; 4921 LIST_HEAD(cull_list); 4922 4923 lockdep_assert_held(&wq_pool_mutex); 4924 4925 if (--pool->refcnt) 4926 return; 4927 4928 /* sanity checks */ 4929 if (WARN_ON(!(pool->cpu < 0)) || 4930 WARN_ON(!list_empty(&pool->worklist))) 4931 return; 4932 4933 /* release id and unhash */ 4934 if (pool->id >= 0) 4935 idr_remove(&worker_pool_idr, pool->id); 4936 hash_del(&pool->hash_node); 4937 4938 /* 4939 * Become the manager and destroy all workers. This prevents 4940 * @pool's workers from blocking on attach_mutex. We're the last 4941 * manager and @pool gets freed with the flag set. 4942 * 4943 * Having a concurrent manager is quite unlikely to happen as we can 4944 * only get here with 4945 * pwq->refcnt == pool->refcnt == 0 4946 * which implies no work queued to the pool, which implies no worker can 4947 * become the manager. However a worker could have taken the role of 4948 * manager before the refcnts dropped to 0, since maybe_create_worker() 4949 * drops pool->lock 4950 */ 4951 while (true) { 4952 rcuwait_wait_event(&manager_wait, 4953 !(pool->flags & POOL_MANAGER_ACTIVE), 4954 TASK_UNINTERRUPTIBLE); 4955 4956 mutex_lock(&wq_pool_attach_mutex); 4957 raw_spin_lock_irq(&pool->lock); 4958 if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 4959 pool->flags |= POOL_MANAGER_ACTIVE; 4960 break; 4961 } 4962 raw_spin_unlock_irq(&pool->lock); 4963 mutex_unlock(&wq_pool_attach_mutex); 4964 } 4965 4966 while ((worker = first_idle_worker(pool))) 4967 set_worker_dying(worker, &cull_list); 4968 WARN_ON(pool->nr_workers || pool->nr_idle); 4969 raw_spin_unlock_irq(&pool->lock); 4970 4971 detach_dying_workers(&cull_list); 4972 4973 mutex_unlock(&wq_pool_attach_mutex); 4974 4975 reap_dying_workers(&cull_list); 4976 4977 /* shut down the timers */ 4978 del_timer_sync(&pool->idle_timer); 4979 cancel_work_sync(&pool->idle_cull_work); 4980 del_timer_sync(&pool->mayday_timer); 4981 4982 /* RCU protected to allow dereferences from get_work_pool() */ 4983 call_rcu(&pool->rcu, rcu_free_pool); 4984 } 4985 4986 /** 4987 * get_unbound_pool - get a worker_pool with the specified attributes 4988 * @attrs: the attributes of the worker_pool to get 4989 * 4990 * Obtain a worker_pool which has the same attributes as @attrs, bump the 4991 * reference count and return it. If there already is a matching 4992 * worker_pool, it will be used; otherwise, this function attempts to 4993 * create a new one. 4994 * 4995 * Should be called with wq_pool_mutex held. 4996 * 4997 * Return: On success, a worker_pool with the same attributes as @attrs. 4998 * On failure, %NULL. 4999 */ 5000 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 5001 { 5002 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 5003 u32 hash = wqattrs_hash(attrs); 5004 struct worker_pool *pool; 5005 int pod, node = NUMA_NO_NODE; 5006 5007 lockdep_assert_held(&wq_pool_mutex); 5008 5009 /* do we already have a matching pool? */ 5010 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 5011 if (wqattrs_equal(pool->attrs, attrs)) { 5012 pool->refcnt++; 5013 return pool; 5014 } 5015 } 5016 5017 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 5018 for (pod = 0; pod < pt->nr_pods; pod++) { 5019 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 5020 node = pt->pod_node[pod]; 5021 break; 5022 } 5023 } 5024 5025 /* nope, create a new one */ 5026 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 5027 if (!pool || init_worker_pool(pool) < 0) 5028 goto fail; 5029 5030 pool->node = node; 5031 copy_workqueue_attrs(pool->attrs, attrs); 5032 wqattrs_clear_for_pool(pool->attrs); 5033 5034 if (worker_pool_assign_id(pool) < 0) 5035 goto fail; 5036 5037 /* create and start the initial worker */ 5038 if (wq_online && !create_worker(pool)) 5039 goto fail; 5040 5041 /* install */ 5042 hash_add(unbound_pool_hash, &pool->hash_node, hash); 5043 5044 return pool; 5045 fail: 5046 if (pool) 5047 put_unbound_pool(pool); 5048 return NULL; 5049 } 5050 5051 /* 5052 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 5053 * refcnt and needs to be destroyed. 5054 */ 5055 static void pwq_release_workfn(struct kthread_work *work) 5056 { 5057 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 5058 release_work); 5059 struct workqueue_struct *wq = pwq->wq; 5060 struct worker_pool *pool = pwq->pool; 5061 bool is_last = false; 5062 5063 /* 5064 * When @pwq is not linked, it doesn't hold any reference to the 5065 * @wq, and @wq is invalid to access. 5066 */ 5067 if (!list_empty(&pwq->pwqs_node)) { 5068 mutex_lock(&wq->mutex); 5069 list_del_rcu(&pwq->pwqs_node); 5070 is_last = list_empty(&wq->pwqs); 5071 5072 /* 5073 * For ordered workqueue with a plugged dfl_pwq, restart it now. 5074 */ 5075 if (!is_last && (wq->flags & __WQ_ORDERED)) 5076 unplug_oldest_pwq(wq); 5077 5078 mutex_unlock(&wq->mutex); 5079 } 5080 5081 if (wq->flags & WQ_UNBOUND) { 5082 mutex_lock(&wq_pool_mutex); 5083 put_unbound_pool(pool); 5084 mutex_unlock(&wq_pool_mutex); 5085 } 5086 5087 if (!list_empty(&pwq->pending_node)) { 5088 struct wq_node_nr_active *nna = 5089 wq_node_nr_active(pwq->wq, pwq->pool->node); 5090 5091 raw_spin_lock_irq(&nna->lock); 5092 list_del_init(&pwq->pending_node); 5093 raw_spin_unlock_irq(&nna->lock); 5094 } 5095 5096 kfree_rcu(pwq, rcu); 5097 5098 /* 5099 * If we're the last pwq going away, @wq is already dead and no one 5100 * is gonna access it anymore. Schedule RCU free. 5101 */ 5102 if (is_last) { 5103 wq_unregister_lockdep(wq); 5104 call_rcu(&wq->rcu, rcu_free_wq); 5105 } 5106 } 5107 5108 /* initialize newly allocated @pwq which is associated with @wq and @pool */ 5109 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 5110 struct worker_pool *pool) 5111 { 5112 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK); 5113 5114 memset(pwq, 0, sizeof(*pwq)); 5115 5116 pwq->pool = pool; 5117 pwq->wq = wq; 5118 pwq->flush_color = -1; 5119 pwq->refcnt = 1; 5120 INIT_LIST_HEAD(&pwq->inactive_works); 5121 INIT_LIST_HEAD(&pwq->pending_node); 5122 INIT_LIST_HEAD(&pwq->pwqs_node); 5123 INIT_LIST_HEAD(&pwq->mayday_node); 5124 kthread_init_work(&pwq->release_work, pwq_release_workfn); 5125 } 5126 5127 /* sync @pwq with the current state of its associated wq and link it */ 5128 static void link_pwq(struct pool_workqueue *pwq) 5129 { 5130 struct workqueue_struct *wq = pwq->wq; 5131 5132 lockdep_assert_held(&wq->mutex); 5133 5134 /* may be called multiple times, ignore if already linked */ 5135 if (!list_empty(&pwq->pwqs_node)) 5136 return; 5137 5138 /* set the matching work_color */ 5139 pwq->work_color = wq->work_color; 5140 5141 /* link in @pwq */ 5142 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); 5143 } 5144 5145 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 5146 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 5147 const struct workqueue_attrs *attrs) 5148 { 5149 struct worker_pool *pool; 5150 struct pool_workqueue *pwq; 5151 5152 lockdep_assert_held(&wq_pool_mutex); 5153 5154 pool = get_unbound_pool(attrs); 5155 if (!pool) 5156 return NULL; 5157 5158 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 5159 if (!pwq) { 5160 put_unbound_pool(pool); 5161 return NULL; 5162 } 5163 5164 init_pwq(pwq, wq, pool); 5165 return pwq; 5166 } 5167 5168 static void apply_wqattrs_lock(void) 5169 { 5170 mutex_lock(&wq_pool_mutex); 5171 } 5172 5173 static void apply_wqattrs_unlock(void) 5174 { 5175 mutex_unlock(&wq_pool_mutex); 5176 } 5177 5178 /** 5179 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 5180 * @attrs: the wq_attrs of the default pwq of the target workqueue 5181 * @cpu: the target CPU 5182 * 5183 * Calculate the cpumask a workqueue with @attrs should use on @pod. 5184 * The result is stored in @attrs->__pod_cpumask. 5185 * 5186 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 5187 * and @pod has online CPUs requested by @attrs, the returned cpumask is the 5188 * intersection of the possible CPUs of @pod and @attrs->cpumask. 5189 * 5190 * The caller is responsible for ensuring that the cpumask of @pod stays stable. 5191 */ 5192 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu) 5193 { 5194 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5195 int pod = pt->cpu_pod[cpu]; 5196 5197 /* calculate possible CPUs in @pod that @attrs wants */ 5198 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 5199 /* does @pod have any online CPUs @attrs wants? */ 5200 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) { 5201 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 5202 return; 5203 } 5204 } 5205 5206 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ 5207 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 5208 int cpu, struct pool_workqueue *pwq) 5209 { 5210 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); 5211 struct pool_workqueue *old_pwq; 5212 5213 lockdep_assert_held(&wq_pool_mutex); 5214 lockdep_assert_held(&wq->mutex); 5215 5216 /* link_pwq() can handle duplicate calls */ 5217 link_pwq(pwq); 5218 5219 old_pwq = rcu_access_pointer(*slot); 5220 rcu_assign_pointer(*slot, pwq); 5221 return old_pwq; 5222 } 5223 5224 /* context to store the prepared attrs & pwqs before applying */ 5225 struct apply_wqattrs_ctx { 5226 struct workqueue_struct *wq; /* target workqueue */ 5227 struct workqueue_attrs *attrs; /* attrs to apply */ 5228 struct list_head list; /* queued for batching commit */ 5229 struct pool_workqueue *dfl_pwq; 5230 struct pool_workqueue *pwq_tbl[]; 5231 }; 5232 5233 /* free the resources after success or abort */ 5234 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 5235 { 5236 if (ctx) { 5237 int cpu; 5238 5239 for_each_possible_cpu(cpu) 5240 put_pwq_unlocked(ctx->pwq_tbl[cpu]); 5241 put_pwq_unlocked(ctx->dfl_pwq); 5242 5243 free_workqueue_attrs(ctx->attrs); 5244 5245 kfree(ctx); 5246 } 5247 } 5248 5249 /* allocate the attrs and pwqs for later installation */ 5250 static struct apply_wqattrs_ctx * 5251 apply_wqattrs_prepare(struct workqueue_struct *wq, 5252 const struct workqueue_attrs *attrs, 5253 const cpumask_var_t unbound_cpumask) 5254 { 5255 struct apply_wqattrs_ctx *ctx; 5256 struct workqueue_attrs *new_attrs; 5257 int cpu; 5258 5259 lockdep_assert_held(&wq_pool_mutex); 5260 5261 if (WARN_ON(attrs->affn_scope < 0 || 5262 attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 5263 return ERR_PTR(-EINVAL); 5264 5265 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 5266 5267 new_attrs = alloc_workqueue_attrs(); 5268 if (!ctx || !new_attrs) 5269 goto out_free; 5270 5271 /* 5272 * If something goes wrong during CPU up/down, we'll fall back to 5273 * the default pwq covering whole @attrs->cpumask. Always create 5274 * it even if we don't use it immediately. 5275 */ 5276 copy_workqueue_attrs(new_attrs, attrs); 5277 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 5278 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5279 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 5280 if (!ctx->dfl_pwq) 5281 goto out_free; 5282 5283 for_each_possible_cpu(cpu) { 5284 if (new_attrs->ordered) { 5285 ctx->dfl_pwq->refcnt++; 5286 ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 5287 } else { 5288 wq_calc_pod_cpumask(new_attrs, cpu); 5289 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 5290 if (!ctx->pwq_tbl[cpu]) 5291 goto out_free; 5292 } 5293 } 5294 5295 /* save the user configured attrs and sanitize it. */ 5296 copy_workqueue_attrs(new_attrs, attrs); 5297 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 5298 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5299 ctx->attrs = new_attrs; 5300 5301 /* 5302 * For initialized ordered workqueues, there should only be one pwq 5303 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution 5304 * of newly queued work items until execution of older work items in 5305 * the old pwq's have completed. 5306 */ 5307 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)) 5308 ctx->dfl_pwq->plugged = true; 5309 5310 ctx->wq = wq; 5311 return ctx; 5312 5313 out_free: 5314 free_workqueue_attrs(new_attrs); 5315 apply_wqattrs_cleanup(ctx); 5316 return ERR_PTR(-ENOMEM); 5317 } 5318 5319 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 5320 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 5321 { 5322 int cpu; 5323 5324 /* all pwqs have been created successfully, let's install'em */ 5325 mutex_lock(&ctx->wq->mutex); 5326 5327 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 5328 5329 /* save the previous pwqs and install the new ones */ 5330 for_each_possible_cpu(cpu) 5331 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 5332 ctx->pwq_tbl[cpu]); 5333 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); 5334 5335 /* update node_nr_active->max */ 5336 wq_update_node_max_active(ctx->wq, -1); 5337 5338 /* rescuer needs to respect wq cpumask changes */ 5339 if (ctx->wq->rescuer) 5340 set_cpus_allowed_ptr(ctx->wq->rescuer->task, 5341 unbound_effective_cpumask(ctx->wq)); 5342 5343 mutex_unlock(&ctx->wq->mutex); 5344 } 5345 5346 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 5347 const struct workqueue_attrs *attrs) 5348 { 5349 struct apply_wqattrs_ctx *ctx; 5350 5351 /* only unbound workqueues can change attributes */ 5352 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 5353 return -EINVAL; 5354 5355 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 5356 if (IS_ERR(ctx)) 5357 return PTR_ERR(ctx); 5358 5359 /* the ctx has been prepared successfully, let's commit it */ 5360 apply_wqattrs_commit(ctx); 5361 apply_wqattrs_cleanup(ctx); 5362 5363 return 0; 5364 } 5365 5366 /** 5367 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 5368 * @wq: the target workqueue 5369 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 5370 * 5371 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 5372 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 5373 * work items are affine to the pod it was issued on. Older pwqs are released as 5374 * in-flight work items finish. Note that a work item which repeatedly requeues 5375 * itself back-to-back will stay on its current pwq. 5376 * 5377 * Performs GFP_KERNEL allocations. 5378 * 5379 * Return: 0 on success and -errno on failure. 5380 */ 5381 int apply_workqueue_attrs(struct workqueue_struct *wq, 5382 const struct workqueue_attrs *attrs) 5383 { 5384 int ret; 5385 5386 mutex_lock(&wq_pool_mutex); 5387 ret = apply_workqueue_attrs_locked(wq, attrs); 5388 mutex_unlock(&wq_pool_mutex); 5389 5390 return ret; 5391 } 5392 5393 /** 5394 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug 5395 * @wq: the target workqueue 5396 * @cpu: the CPU to update the pwq slot for 5397 * 5398 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 5399 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged. 5400 * 5401 * 5402 * If pod affinity can't be adjusted due to memory allocation failure, it falls 5403 * back to @wq->dfl_pwq which may not be optimal but is always correct. 5404 * 5405 * Note that when the last allowed CPU of a pod goes offline for a workqueue 5406 * with a cpumask spanning multiple pods, the workers which were already 5407 * executing the work items for the workqueue will lose their CPU affinity and 5408 * may execute on any CPU. This is similar to how per-cpu workqueues behave on 5409 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 5410 * responsibility to flush the work item from CPU_DOWN_PREPARE. 5411 */ 5412 static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu) 5413 { 5414 struct pool_workqueue *old_pwq = NULL, *pwq; 5415 struct workqueue_attrs *target_attrs; 5416 5417 lockdep_assert_held(&wq_pool_mutex); 5418 5419 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 5420 return; 5421 5422 /* 5423 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 5424 * Let's use a preallocated one. The following buf is protected by 5425 * CPU hotplug exclusion. 5426 */ 5427 target_attrs = unbound_wq_update_pwq_attrs_buf; 5428 5429 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 5430 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 5431 5432 /* nothing to do if the target cpumask matches the current pwq */ 5433 wq_calc_pod_cpumask(target_attrs, cpu); 5434 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) 5435 return; 5436 5437 /* create a new pwq */ 5438 pwq = alloc_unbound_pwq(wq, target_attrs); 5439 if (!pwq) { 5440 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 5441 wq->name); 5442 goto use_dfl_pwq; 5443 } 5444 5445 /* Install the new pwq. */ 5446 mutex_lock(&wq->mutex); 5447 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5448 goto out_unlock; 5449 5450 use_dfl_pwq: 5451 mutex_lock(&wq->mutex); 5452 pwq = unbound_pwq(wq, -1); 5453 raw_spin_lock_irq(&pwq->pool->lock); 5454 get_pwq(pwq); 5455 raw_spin_unlock_irq(&pwq->pool->lock); 5456 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5457 out_unlock: 5458 mutex_unlock(&wq->mutex); 5459 put_pwq_unlocked(old_pwq); 5460 } 5461 5462 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 5463 { 5464 bool highpri = wq->flags & WQ_HIGHPRI; 5465 int cpu, ret; 5466 5467 lockdep_assert_held(&wq_pool_mutex); 5468 5469 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 5470 if (!wq->cpu_pwq) 5471 goto enomem; 5472 5473 if (!(wq->flags & WQ_UNBOUND)) { 5474 struct worker_pool __percpu *pools; 5475 5476 if (wq->flags & WQ_BH) 5477 pools = bh_worker_pools; 5478 else 5479 pools = cpu_worker_pools; 5480 5481 for_each_possible_cpu(cpu) { 5482 struct pool_workqueue **pwq_p; 5483 struct worker_pool *pool; 5484 5485 pool = &(per_cpu_ptr(pools, cpu)[highpri]); 5486 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu); 5487 5488 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 5489 pool->node); 5490 if (!*pwq_p) 5491 goto enomem; 5492 5493 init_pwq(*pwq_p, wq, pool); 5494 5495 mutex_lock(&wq->mutex); 5496 link_pwq(*pwq_p); 5497 mutex_unlock(&wq->mutex); 5498 } 5499 return 0; 5500 } 5501 5502 if (wq->flags & __WQ_ORDERED) { 5503 struct pool_workqueue *dfl_pwq; 5504 5505 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]); 5506 /* there should only be single pwq for ordering guarantee */ 5507 dfl_pwq = rcu_access_pointer(wq->dfl_pwq); 5508 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || 5509 wq->pwqs.prev != &dfl_pwq->pwqs_node), 5510 "ordering guarantee broken for workqueue %s\n", wq->name); 5511 } else { 5512 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]); 5513 } 5514 5515 return ret; 5516 5517 enomem: 5518 if (wq->cpu_pwq) { 5519 for_each_possible_cpu(cpu) { 5520 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 5521 5522 if (pwq) 5523 kmem_cache_free(pwq_cache, pwq); 5524 } 5525 free_percpu(wq->cpu_pwq); 5526 wq->cpu_pwq = NULL; 5527 } 5528 return -ENOMEM; 5529 } 5530 5531 static int wq_clamp_max_active(int max_active, unsigned int flags, 5532 const char *name) 5533 { 5534 if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 5535 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 5536 max_active, name, 1, WQ_MAX_ACTIVE); 5537 5538 return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 5539 } 5540 5541 /* 5542 * Workqueues which may be used during memory reclaim should have a rescuer 5543 * to guarantee forward progress. 5544 */ 5545 static int init_rescuer(struct workqueue_struct *wq) 5546 { 5547 struct worker *rescuer; 5548 char id_buf[WORKER_ID_LEN]; 5549 int ret; 5550 5551 lockdep_assert_held(&wq_pool_mutex); 5552 5553 if (!(wq->flags & WQ_MEM_RECLAIM)) 5554 return 0; 5555 5556 rescuer = alloc_worker(NUMA_NO_NODE); 5557 if (!rescuer) { 5558 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 5559 wq->name); 5560 return -ENOMEM; 5561 } 5562 5563 rescuer->rescue_wq = wq; 5564 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL); 5565 5566 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf); 5567 if (IS_ERR(rescuer->task)) { 5568 ret = PTR_ERR(rescuer->task); 5569 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 5570 wq->name, ERR_PTR(ret)); 5571 kfree(rescuer); 5572 return ret; 5573 } 5574 5575 wq->rescuer = rescuer; 5576 if (wq->flags & WQ_UNBOUND) 5577 kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq)); 5578 else 5579 kthread_bind_mask(rescuer->task, cpu_possible_mask); 5580 wake_up_process(rescuer->task); 5581 5582 return 0; 5583 } 5584 5585 /** 5586 * wq_adjust_max_active - update a wq's max_active to the current setting 5587 * @wq: target workqueue 5588 * 5589 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and 5590 * activate inactive work items accordingly. If @wq is freezing, clear 5591 * @wq->max_active to zero. 5592 */ 5593 static void wq_adjust_max_active(struct workqueue_struct *wq) 5594 { 5595 bool activated; 5596 int new_max, new_min; 5597 5598 lockdep_assert_held(&wq->mutex); 5599 5600 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { 5601 new_max = 0; 5602 new_min = 0; 5603 } else { 5604 new_max = wq->saved_max_active; 5605 new_min = wq->saved_min_active; 5606 } 5607 5608 if (wq->max_active == new_max && wq->min_active == new_min) 5609 return; 5610 5611 /* 5612 * Update @wq->max/min_active and then kick inactive work items if more 5613 * active work items are allowed. This doesn't break work item ordering 5614 * because new work items are always queued behind existing inactive 5615 * work items if there are any. 5616 */ 5617 WRITE_ONCE(wq->max_active, new_max); 5618 WRITE_ONCE(wq->min_active, new_min); 5619 5620 if (wq->flags & WQ_UNBOUND) 5621 wq_update_node_max_active(wq, -1); 5622 5623 if (new_max == 0) 5624 return; 5625 5626 /* 5627 * Round-robin through pwq's activating the first inactive work item 5628 * until max_active is filled. 5629 */ 5630 do { 5631 struct pool_workqueue *pwq; 5632 5633 activated = false; 5634 for_each_pwq(pwq, wq) { 5635 unsigned long irq_flags; 5636 5637 /* can be called during early boot w/ irq disabled */ 5638 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 5639 if (pwq_activate_first_inactive(pwq, true)) { 5640 activated = true; 5641 kick_pool(pwq->pool); 5642 } 5643 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 5644 } 5645 } while (activated); 5646 } 5647 5648 __printf(1, 0) 5649 static struct workqueue_struct *__alloc_workqueue(const char *fmt, 5650 unsigned int flags, 5651 int max_active, va_list args) 5652 { 5653 struct workqueue_struct *wq; 5654 size_t wq_size; 5655 int name_len; 5656 5657 if (flags & WQ_BH) { 5658 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS)) 5659 return NULL; 5660 if (WARN_ON_ONCE(max_active)) 5661 return NULL; 5662 } 5663 5664 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 5665 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 5666 flags |= WQ_UNBOUND; 5667 5668 /* allocate wq and format name */ 5669 if (flags & WQ_UNBOUND) 5670 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); 5671 else 5672 wq_size = sizeof(*wq); 5673 5674 wq = kzalloc(wq_size, GFP_KERNEL); 5675 if (!wq) 5676 return NULL; 5677 5678 if (flags & WQ_UNBOUND) { 5679 wq->unbound_attrs = alloc_workqueue_attrs(); 5680 if (!wq->unbound_attrs) 5681 goto err_free_wq; 5682 } 5683 5684 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); 5685 5686 if (name_len >= WQ_NAME_LEN) 5687 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", 5688 wq->name); 5689 5690 if (flags & WQ_BH) { 5691 /* 5692 * BH workqueues always share a single execution context per CPU 5693 * and don't impose any max_active limit. 5694 */ 5695 max_active = INT_MAX; 5696 } else { 5697 max_active = max_active ?: WQ_DFL_ACTIVE; 5698 max_active = wq_clamp_max_active(max_active, flags, wq->name); 5699 } 5700 5701 /* init wq */ 5702 wq->flags = flags; 5703 wq->max_active = max_active; 5704 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); 5705 wq->saved_max_active = wq->max_active; 5706 wq->saved_min_active = wq->min_active; 5707 mutex_init(&wq->mutex); 5708 atomic_set(&wq->nr_pwqs_to_flush, 0); 5709 INIT_LIST_HEAD(&wq->pwqs); 5710 INIT_LIST_HEAD(&wq->flusher_queue); 5711 INIT_LIST_HEAD(&wq->flusher_overflow); 5712 INIT_LIST_HEAD(&wq->maydays); 5713 5714 INIT_LIST_HEAD(&wq->list); 5715 5716 if (flags & WQ_UNBOUND) { 5717 if (alloc_node_nr_active(wq->node_nr_active) < 0) 5718 goto err_free_wq; 5719 } 5720 5721 /* 5722 * wq_pool_mutex protects the workqueues list, allocations of PWQs, 5723 * and the global freeze state. 5724 */ 5725 apply_wqattrs_lock(); 5726 5727 if (alloc_and_link_pwqs(wq) < 0) 5728 goto err_unlock_free_node_nr_active; 5729 5730 mutex_lock(&wq->mutex); 5731 wq_adjust_max_active(wq); 5732 mutex_unlock(&wq->mutex); 5733 5734 list_add_tail_rcu(&wq->list, &workqueues); 5735 5736 if (wq_online && init_rescuer(wq) < 0) 5737 goto err_unlock_destroy; 5738 5739 apply_wqattrs_unlock(); 5740 5741 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 5742 goto err_destroy; 5743 5744 return wq; 5745 5746 err_unlock_free_node_nr_active: 5747 apply_wqattrs_unlock(); 5748 /* 5749 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work, 5750 * flushing the pwq_release_worker ensures that the pwq_release_workfn() 5751 * completes before calling kfree(wq). 5752 */ 5753 if (wq->flags & WQ_UNBOUND) { 5754 kthread_flush_worker(pwq_release_worker); 5755 free_node_nr_active(wq->node_nr_active); 5756 } 5757 err_free_wq: 5758 free_workqueue_attrs(wq->unbound_attrs); 5759 kfree(wq); 5760 return NULL; 5761 err_unlock_destroy: 5762 apply_wqattrs_unlock(); 5763 err_destroy: 5764 destroy_workqueue(wq); 5765 return NULL; 5766 } 5767 5768 __printf(1, 4) 5769 struct workqueue_struct *alloc_workqueue(const char *fmt, 5770 unsigned int flags, 5771 int max_active, ...) 5772 { 5773 struct workqueue_struct *wq; 5774 va_list args; 5775 5776 va_start(args, max_active); 5777 wq = __alloc_workqueue(fmt, flags, max_active, args); 5778 va_end(args); 5779 if (!wq) 5780 return NULL; 5781 5782 wq_init_lockdep(wq); 5783 5784 return wq; 5785 } 5786 EXPORT_SYMBOL_GPL(alloc_workqueue); 5787 5788 #ifdef CONFIG_LOCKDEP 5789 __printf(1, 5) 5790 struct workqueue_struct * 5791 alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags, 5792 int max_active, struct lockdep_map *lockdep_map, ...) 5793 { 5794 struct workqueue_struct *wq; 5795 va_list args; 5796 5797 va_start(args, lockdep_map); 5798 wq = __alloc_workqueue(fmt, flags, max_active, args); 5799 va_end(args); 5800 if (!wq) 5801 return NULL; 5802 5803 wq->lockdep_map = lockdep_map; 5804 5805 return wq; 5806 } 5807 EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map); 5808 #endif 5809 5810 static bool pwq_busy(struct pool_workqueue *pwq) 5811 { 5812 int i; 5813 5814 for (i = 0; i < WORK_NR_COLORS; i++) 5815 if (pwq->nr_in_flight[i]) 5816 return true; 5817 5818 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) 5819 return true; 5820 if (!pwq_is_empty(pwq)) 5821 return true; 5822 5823 return false; 5824 } 5825 5826 /** 5827 * destroy_workqueue - safely terminate a workqueue 5828 * @wq: target workqueue 5829 * 5830 * Safely destroy a workqueue. All work currently pending will be done first. 5831 */ 5832 void destroy_workqueue(struct workqueue_struct *wq) 5833 { 5834 struct pool_workqueue *pwq; 5835 int cpu; 5836 5837 /* 5838 * Remove it from sysfs first so that sanity check failure doesn't 5839 * lead to sysfs name conflicts. 5840 */ 5841 workqueue_sysfs_unregister(wq); 5842 5843 /* mark the workqueue destruction is in progress */ 5844 mutex_lock(&wq->mutex); 5845 wq->flags |= __WQ_DESTROYING; 5846 mutex_unlock(&wq->mutex); 5847 5848 /* drain it before proceeding with destruction */ 5849 drain_workqueue(wq); 5850 5851 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 5852 if (wq->rescuer) { 5853 struct worker *rescuer = wq->rescuer; 5854 5855 /* this prevents new queueing */ 5856 raw_spin_lock_irq(&wq_mayday_lock); 5857 wq->rescuer = NULL; 5858 raw_spin_unlock_irq(&wq_mayday_lock); 5859 5860 /* rescuer will empty maydays list before exiting */ 5861 kthread_stop(rescuer->task); 5862 kfree(rescuer); 5863 } 5864 5865 /* 5866 * Sanity checks - grab all the locks so that we wait for all 5867 * in-flight operations which may do put_pwq(). 5868 */ 5869 mutex_lock(&wq_pool_mutex); 5870 mutex_lock(&wq->mutex); 5871 for_each_pwq(pwq, wq) { 5872 raw_spin_lock_irq(&pwq->pool->lock); 5873 if (WARN_ON(pwq_busy(pwq))) { 5874 pr_warn("%s: %s has the following busy pwq\n", 5875 __func__, wq->name); 5876 show_pwq(pwq); 5877 raw_spin_unlock_irq(&pwq->pool->lock); 5878 mutex_unlock(&wq->mutex); 5879 mutex_unlock(&wq_pool_mutex); 5880 show_one_workqueue(wq); 5881 return; 5882 } 5883 raw_spin_unlock_irq(&pwq->pool->lock); 5884 } 5885 mutex_unlock(&wq->mutex); 5886 5887 /* 5888 * wq list is used to freeze wq, remove from list after 5889 * flushing is complete in case freeze races us. 5890 */ 5891 list_del_rcu(&wq->list); 5892 mutex_unlock(&wq_pool_mutex); 5893 5894 /* 5895 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 5896 * to put the base refs. @wq will be auto-destroyed from the last 5897 * pwq_put. RCU read lock prevents @wq from going away from under us. 5898 */ 5899 rcu_read_lock(); 5900 5901 for_each_possible_cpu(cpu) { 5902 put_pwq_unlocked(unbound_pwq(wq, cpu)); 5903 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); 5904 } 5905 5906 put_pwq_unlocked(unbound_pwq(wq, -1)); 5907 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); 5908 5909 rcu_read_unlock(); 5910 } 5911 EXPORT_SYMBOL_GPL(destroy_workqueue); 5912 5913 /** 5914 * workqueue_set_max_active - adjust max_active of a workqueue 5915 * @wq: target workqueue 5916 * @max_active: new max_active value. 5917 * 5918 * Set max_active of @wq to @max_active. See the alloc_workqueue() function 5919 * comment. 5920 * 5921 * CONTEXT: 5922 * Don't call from IRQ context. 5923 */ 5924 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 5925 { 5926 /* max_active doesn't mean anything for BH workqueues */ 5927 if (WARN_ON(wq->flags & WQ_BH)) 5928 return; 5929 /* disallow meddling with max_active for ordered workqueues */ 5930 if (WARN_ON(wq->flags & __WQ_ORDERED)) 5931 return; 5932 5933 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 5934 5935 mutex_lock(&wq->mutex); 5936 5937 wq->saved_max_active = max_active; 5938 if (wq->flags & WQ_UNBOUND) 5939 wq->saved_min_active = min(wq->saved_min_active, max_active); 5940 5941 wq_adjust_max_active(wq); 5942 5943 mutex_unlock(&wq->mutex); 5944 } 5945 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 5946 5947 /** 5948 * workqueue_set_min_active - adjust min_active of an unbound workqueue 5949 * @wq: target unbound workqueue 5950 * @min_active: new min_active value 5951 * 5952 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an 5953 * unbound workqueue is not guaranteed to be able to process max_active 5954 * interdependent work items. Instead, an unbound workqueue is guaranteed to be 5955 * able to process min_active number of interdependent work items which is 5956 * %WQ_DFL_MIN_ACTIVE by default. 5957 * 5958 * Use this function to adjust the min_active value between 0 and the current 5959 * max_active. 5960 */ 5961 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active) 5962 { 5963 /* min_active is only meaningful for non-ordered unbound workqueues */ 5964 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) != 5965 WQ_UNBOUND)) 5966 return; 5967 5968 mutex_lock(&wq->mutex); 5969 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active); 5970 wq_adjust_max_active(wq); 5971 mutex_unlock(&wq->mutex); 5972 } 5973 5974 /** 5975 * current_work - retrieve %current task's work struct 5976 * 5977 * Determine if %current task is a workqueue worker and what it's working on. 5978 * Useful to find out the context that the %current task is running in. 5979 * 5980 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 5981 */ 5982 struct work_struct *current_work(void) 5983 { 5984 struct worker *worker = current_wq_worker(); 5985 5986 return worker ? worker->current_work : NULL; 5987 } 5988 EXPORT_SYMBOL(current_work); 5989 5990 /** 5991 * current_is_workqueue_rescuer - is %current workqueue rescuer? 5992 * 5993 * Determine whether %current is a workqueue rescuer. Can be used from 5994 * work functions to determine whether it's being run off the rescuer task. 5995 * 5996 * Return: %true if %current is a workqueue rescuer. %false otherwise. 5997 */ 5998 bool current_is_workqueue_rescuer(void) 5999 { 6000 struct worker *worker = current_wq_worker(); 6001 6002 return worker && worker->rescue_wq; 6003 } 6004 6005 /** 6006 * workqueue_congested - test whether a workqueue is congested 6007 * @cpu: CPU in question 6008 * @wq: target workqueue 6009 * 6010 * Test whether @wq's cpu workqueue for @cpu is congested. There is 6011 * no synchronization around this function and the test result is 6012 * unreliable and only useful as advisory hints or for debugging. 6013 * 6014 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 6015 * 6016 * With the exception of ordered workqueues, all workqueues have per-cpu 6017 * pool_workqueues, each with its own congested state. A workqueue being 6018 * congested on one CPU doesn't mean that the workqueue is contested on any 6019 * other CPUs. 6020 * 6021 * Return: 6022 * %true if congested, %false otherwise. 6023 */ 6024 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 6025 { 6026 struct pool_workqueue *pwq; 6027 bool ret; 6028 6029 rcu_read_lock(); 6030 preempt_disable(); 6031 6032 if (cpu == WORK_CPU_UNBOUND) 6033 cpu = smp_processor_id(); 6034 6035 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 6036 ret = !list_empty(&pwq->inactive_works); 6037 6038 preempt_enable(); 6039 rcu_read_unlock(); 6040 6041 return ret; 6042 } 6043 EXPORT_SYMBOL_GPL(workqueue_congested); 6044 6045 /** 6046 * work_busy - test whether a work is currently pending or running 6047 * @work: the work to be tested 6048 * 6049 * Test whether @work is currently pending or running. There is no 6050 * synchronization around this function and the test result is 6051 * unreliable and only useful as advisory hints or for debugging. 6052 * 6053 * Return: 6054 * OR'd bitmask of WORK_BUSY_* bits. 6055 */ 6056 unsigned int work_busy(struct work_struct *work) 6057 { 6058 struct worker_pool *pool; 6059 unsigned long irq_flags; 6060 unsigned int ret = 0; 6061 6062 if (work_pending(work)) 6063 ret |= WORK_BUSY_PENDING; 6064 6065 rcu_read_lock(); 6066 pool = get_work_pool(work); 6067 if (pool) { 6068 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6069 if (find_worker_executing_work(pool, work)) 6070 ret |= WORK_BUSY_RUNNING; 6071 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6072 } 6073 rcu_read_unlock(); 6074 6075 return ret; 6076 } 6077 EXPORT_SYMBOL_GPL(work_busy); 6078 6079 /** 6080 * set_worker_desc - set description for the current work item 6081 * @fmt: printf-style format string 6082 * @...: arguments for the format string 6083 * 6084 * This function can be called by a running work function to describe what 6085 * the work item is about. If the worker task gets dumped, this 6086 * information will be printed out together to help debugging. The 6087 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 6088 */ 6089 void set_worker_desc(const char *fmt, ...) 6090 { 6091 struct worker *worker = current_wq_worker(); 6092 va_list args; 6093 6094 if (worker) { 6095 va_start(args, fmt); 6096 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 6097 va_end(args); 6098 } 6099 } 6100 EXPORT_SYMBOL_GPL(set_worker_desc); 6101 6102 /** 6103 * print_worker_info - print out worker information and description 6104 * @log_lvl: the log level to use when printing 6105 * @task: target task 6106 * 6107 * If @task is a worker and currently executing a work item, print out the 6108 * name of the workqueue being serviced and worker description set with 6109 * set_worker_desc() by the currently executing work item. 6110 * 6111 * This function can be safely called on any task as long as the 6112 * task_struct itself is accessible. While safe, this function isn't 6113 * synchronized and may print out mixups or garbages of limited length. 6114 */ 6115 void print_worker_info(const char *log_lvl, struct task_struct *task) 6116 { 6117 work_func_t *fn = NULL; 6118 char name[WQ_NAME_LEN] = { }; 6119 char desc[WORKER_DESC_LEN] = { }; 6120 struct pool_workqueue *pwq = NULL; 6121 struct workqueue_struct *wq = NULL; 6122 struct worker *worker; 6123 6124 if (!(task->flags & PF_WQ_WORKER)) 6125 return; 6126 6127 /* 6128 * This function is called without any synchronization and @task 6129 * could be in any state. Be careful with dereferences. 6130 */ 6131 worker = kthread_probe_data(task); 6132 6133 /* 6134 * Carefully copy the associated workqueue's workfn, name and desc. 6135 * Keep the original last '\0' in case the original is garbage. 6136 */ 6137 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 6138 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 6139 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 6140 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 6141 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 6142 6143 if (fn || name[0] || desc[0]) { 6144 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 6145 if (strcmp(name, desc)) 6146 pr_cont(" (%s)", desc); 6147 pr_cont("\n"); 6148 } 6149 } 6150 6151 static void pr_cont_pool_info(struct worker_pool *pool) 6152 { 6153 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 6154 if (pool->node != NUMA_NO_NODE) 6155 pr_cont(" node=%d", pool->node); 6156 pr_cont(" flags=0x%x", pool->flags); 6157 if (pool->flags & POOL_BH) 6158 pr_cont(" bh%s", 6159 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6160 else 6161 pr_cont(" nice=%d", pool->attrs->nice); 6162 } 6163 6164 static void pr_cont_worker_id(struct worker *worker) 6165 { 6166 struct worker_pool *pool = worker->pool; 6167 6168 if (pool->flags & WQ_BH) 6169 pr_cont("bh%s", 6170 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6171 else 6172 pr_cont("%d%s", task_pid_nr(worker->task), 6173 worker->rescue_wq ? "(RESCUER)" : ""); 6174 } 6175 6176 struct pr_cont_work_struct { 6177 bool comma; 6178 work_func_t func; 6179 long ctr; 6180 }; 6181 6182 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 6183 { 6184 if (!pcwsp->ctr) 6185 goto out_record; 6186 if (func == pcwsp->func) { 6187 pcwsp->ctr++; 6188 return; 6189 } 6190 if (pcwsp->ctr == 1) 6191 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 6192 else 6193 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 6194 pcwsp->ctr = 0; 6195 out_record: 6196 if ((long)func == -1L) 6197 return; 6198 pcwsp->comma = comma; 6199 pcwsp->func = func; 6200 pcwsp->ctr = 1; 6201 } 6202 6203 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 6204 { 6205 if (work->func == wq_barrier_func) { 6206 struct wq_barrier *barr; 6207 6208 barr = container_of(work, struct wq_barrier, work); 6209 6210 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6211 pr_cont("%s BAR(%d)", comma ? "," : "", 6212 task_pid_nr(barr->task)); 6213 } else { 6214 if (!comma) 6215 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6216 pr_cont_work_flush(comma, work->func, pcwsp); 6217 } 6218 } 6219 6220 static void show_pwq(struct pool_workqueue *pwq) 6221 { 6222 struct pr_cont_work_struct pcws = { .ctr = 0, }; 6223 struct worker_pool *pool = pwq->pool; 6224 struct work_struct *work; 6225 struct worker *worker; 6226 bool has_in_flight = false, has_pending = false; 6227 int bkt; 6228 6229 pr_info(" pwq %d:", pool->id); 6230 pr_cont_pool_info(pool); 6231 6232 pr_cont(" active=%d refcnt=%d%s\n", 6233 pwq->nr_active, pwq->refcnt, 6234 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 6235 6236 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6237 if (worker->current_pwq == pwq) { 6238 has_in_flight = true; 6239 break; 6240 } 6241 } 6242 if (has_in_flight) { 6243 bool comma = false; 6244 6245 pr_info(" in-flight:"); 6246 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6247 if (worker->current_pwq != pwq) 6248 continue; 6249 6250 pr_cont(" %s", comma ? "," : ""); 6251 pr_cont_worker_id(worker); 6252 pr_cont(":%ps", worker->current_func); 6253 list_for_each_entry(work, &worker->scheduled, entry) 6254 pr_cont_work(false, work, &pcws); 6255 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6256 comma = true; 6257 } 6258 pr_cont("\n"); 6259 } 6260 6261 list_for_each_entry(work, &pool->worklist, entry) { 6262 if (get_work_pwq(work) == pwq) { 6263 has_pending = true; 6264 break; 6265 } 6266 } 6267 if (has_pending) { 6268 bool comma = false; 6269 6270 pr_info(" pending:"); 6271 list_for_each_entry(work, &pool->worklist, entry) { 6272 if (get_work_pwq(work) != pwq) 6273 continue; 6274 6275 pr_cont_work(comma, work, &pcws); 6276 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6277 } 6278 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6279 pr_cont("\n"); 6280 } 6281 6282 if (!list_empty(&pwq->inactive_works)) { 6283 bool comma = false; 6284 6285 pr_info(" inactive:"); 6286 list_for_each_entry(work, &pwq->inactive_works, entry) { 6287 pr_cont_work(comma, work, &pcws); 6288 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6289 } 6290 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6291 pr_cont("\n"); 6292 } 6293 } 6294 6295 /** 6296 * show_one_workqueue - dump state of specified workqueue 6297 * @wq: workqueue whose state will be printed 6298 */ 6299 void show_one_workqueue(struct workqueue_struct *wq) 6300 { 6301 struct pool_workqueue *pwq; 6302 bool idle = true; 6303 unsigned long irq_flags; 6304 6305 for_each_pwq(pwq, wq) { 6306 if (!pwq_is_empty(pwq)) { 6307 idle = false; 6308 break; 6309 } 6310 } 6311 if (idle) /* Nothing to print for idle workqueue */ 6312 return; 6313 6314 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 6315 6316 for_each_pwq(pwq, wq) { 6317 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 6318 if (!pwq_is_empty(pwq)) { 6319 /* 6320 * Defer printing to avoid deadlocks in console 6321 * drivers that queue work while holding locks 6322 * also taken in their write paths. 6323 */ 6324 printk_deferred_enter(); 6325 show_pwq(pwq); 6326 printk_deferred_exit(); 6327 } 6328 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 6329 /* 6330 * We could be printing a lot from atomic context, e.g. 6331 * sysrq-t -> show_all_workqueues(). Avoid triggering 6332 * hard lockup. 6333 */ 6334 touch_nmi_watchdog(); 6335 } 6336 6337 } 6338 6339 /** 6340 * show_one_worker_pool - dump state of specified worker pool 6341 * @pool: worker pool whose state will be printed 6342 */ 6343 static void show_one_worker_pool(struct worker_pool *pool) 6344 { 6345 struct worker *worker; 6346 bool first = true; 6347 unsigned long irq_flags; 6348 unsigned long hung = 0; 6349 6350 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6351 if (pool->nr_workers == pool->nr_idle) 6352 goto next_pool; 6353 6354 /* How long the first pending work is waiting for a worker. */ 6355 if (!list_empty(&pool->worklist)) 6356 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 6357 6358 /* 6359 * Defer printing to avoid deadlocks in console drivers that 6360 * queue work while holding locks also taken in their write 6361 * paths. 6362 */ 6363 printk_deferred_enter(); 6364 pr_info("pool %d:", pool->id); 6365 pr_cont_pool_info(pool); 6366 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 6367 if (pool->manager) 6368 pr_cont(" manager: %d", 6369 task_pid_nr(pool->manager->task)); 6370 list_for_each_entry(worker, &pool->idle_list, entry) { 6371 pr_cont(" %s", first ? "idle: " : ""); 6372 pr_cont_worker_id(worker); 6373 first = false; 6374 } 6375 pr_cont("\n"); 6376 printk_deferred_exit(); 6377 next_pool: 6378 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6379 /* 6380 * We could be printing a lot from atomic context, e.g. 6381 * sysrq-t -> show_all_workqueues(). Avoid triggering 6382 * hard lockup. 6383 */ 6384 touch_nmi_watchdog(); 6385 6386 } 6387 6388 /** 6389 * show_all_workqueues - dump workqueue state 6390 * 6391 * Called from a sysrq handler and prints out all busy workqueues and pools. 6392 */ 6393 void show_all_workqueues(void) 6394 { 6395 struct workqueue_struct *wq; 6396 struct worker_pool *pool; 6397 int pi; 6398 6399 rcu_read_lock(); 6400 6401 pr_info("Showing busy workqueues and worker pools:\n"); 6402 6403 list_for_each_entry_rcu(wq, &workqueues, list) 6404 show_one_workqueue(wq); 6405 6406 for_each_pool(pool, pi) 6407 show_one_worker_pool(pool); 6408 6409 rcu_read_unlock(); 6410 } 6411 6412 /** 6413 * show_freezable_workqueues - dump freezable workqueue state 6414 * 6415 * Called from try_to_freeze_tasks() and prints out all freezable workqueues 6416 * still busy. 6417 */ 6418 void show_freezable_workqueues(void) 6419 { 6420 struct workqueue_struct *wq; 6421 6422 rcu_read_lock(); 6423 6424 pr_info("Showing freezable workqueues that are still busy:\n"); 6425 6426 list_for_each_entry_rcu(wq, &workqueues, list) { 6427 if (!(wq->flags & WQ_FREEZABLE)) 6428 continue; 6429 show_one_workqueue(wq); 6430 } 6431 6432 rcu_read_unlock(); 6433 } 6434 6435 /* used to show worker information through /proc/PID/{comm,stat,status} */ 6436 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 6437 { 6438 /* stabilize PF_WQ_WORKER and worker pool association */ 6439 mutex_lock(&wq_pool_attach_mutex); 6440 6441 if (task->flags & PF_WQ_WORKER) { 6442 struct worker *worker = kthread_data(task); 6443 struct worker_pool *pool = worker->pool; 6444 int off; 6445 6446 off = format_worker_id(buf, size, worker, pool); 6447 6448 if (pool) { 6449 raw_spin_lock_irq(&pool->lock); 6450 /* 6451 * ->desc tracks information (wq name or 6452 * set_worker_desc()) for the latest execution. If 6453 * current, prepend '+', otherwise '-'. 6454 */ 6455 if (worker->desc[0] != '\0') { 6456 if (worker->current_work) 6457 scnprintf(buf + off, size - off, "+%s", 6458 worker->desc); 6459 else 6460 scnprintf(buf + off, size - off, "-%s", 6461 worker->desc); 6462 } 6463 raw_spin_unlock_irq(&pool->lock); 6464 } 6465 } else { 6466 strscpy(buf, task->comm, size); 6467 } 6468 6469 mutex_unlock(&wq_pool_attach_mutex); 6470 } 6471 6472 #ifdef CONFIG_SMP 6473 6474 /* 6475 * CPU hotplug. 6476 * 6477 * There are two challenges in supporting CPU hotplug. Firstly, there 6478 * are a lot of assumptions on strong associations among work, pwq and 6479 * pool which make migrating pending and scheduled works very 6480 * difficult to implement without impacting hot paths. Secondly, 6481 * worker pools serve mix of short, long and very long running works making 6482 * blocked draining impractical. 6483 * 6484 * This is solved by allowing the pools to be disassociated from the CPU 6485 * running as an unbound one and allowing it to be reattached later if the 6486 * cpu comes back online. 6487 */ 6488 6489 static void unbind_workers(int cpu) 6490 { 6491 struct worker_pool *pool; 6492 struct worker *worker; 6493 6494 for_each_cpu_worker_pool(pool, cpu) { 6495 mutex_lock(&wq_pool_attach_mutex); 6496 raw_spin_lock_irq(&pool->lock); 6497 6498 /* 6499 * We've blocked all attach/detach operations. Make all workers 6500 * unbound and set DISASSOCIATED. Before this, all workers 6501 * must be on the cpu. After this, they may become diasporas. 6502 * And the preemption disabled section in their sched callbacks 6503 * are guaranteed to see WORKER_UNBOUND since the code here 6504 * is on the same cpu. 6505 */ 6506 for_each_pool_worker(worker, pool) 6507 worker->flags |= WORKER_UNBOUND; 6508 6509 pool->flags |= POOL_DISASSOCIATED; 6510 6511 /* 6512 * The handling of nr_running in sched callbacks are disabled 6513 * now. Zap nr_running. After this, nr_running stays zero and 6514 * need_more_worker() and keep_working() are always true as 6515 * long as the worklist is not empty. This pool now behaves as 6516 * an unbound (in terms of concurrency management) pool which 6517 * are served by workers tied to the pool. 6518 */ 6519 pool->nr_running = 0; 6520 6521 /* 6522 * With concurrency management just turned off, a busy 6523 * worker blocking could lead to lengthy stalls. Kick off 6524 * unbound chain execution of currently pending work items. 6525 */ 6526 kick_pool(pool); 6527 6528 raw_spin_unlock_irq(&pool->lock); 6529 6530 for_each_pool_worker(worker, pool) 6531 unbind_worker(worker); 6532 6533 mutex_unlock(&wq_pool_attach_mutex); 6534 } 6535 } 6536 6537 /** 6538 * rebind_workers - rebind all workers of a pool to the associated CPU 6539 * @pool: pool of interest 6540 * 6541 * @pool->cpu is coming online. Rebind all workers to the CPU. 6542 */ 6543 static void rebind_workers(struct worker_pool *pool) 6544 { 6545 struct worker *worker; 6546 6547 lockdep_assert_held(&wq_pool_attach_mutex); 6548 6549 /* 6550 * Restore CPU affinity of all workers. As all idle workers should 6551 * be on the run-queue of the associated CPU before any local 6552 * wake-ups for concurrency management happen, restore CPU affinity 6553 * of all workers first and then clear UNBOUND. As we're called 6554 * from CPU_ONLINE, the following shouldn't fail. 6555 */ 6556 for_each_pool_worker(worker, pool) { 6557 kthread_set_per_cpu(worker->task, pool->cpu); 6558 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 6559 pool_allowed_cpus(pool)) < 0); 6560 } 6561 6562 raw_spin_lock_irq(&pool->lock); 6563 6564 pool->flags &= ~POOL_DISASSOCIATED; 6565 6566 for_each_pool_worker(worker, pool) { 6567 unsigned int worker_flags = worker->flags; 6568 6569 /* 6570 * We want to clear UNBOUND but can't directly call 6571 * worker_clr_flags() or adjust nr_running. Atomically 6572 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 6573 * @worker will clear REBOUND using worker_clr_flags() when 6574 * it initiates the next execution cycle thus restoring 6575 * concurrency management. Note that when or whether 6576 * @worker clears REBOUND doesn't affect correctness. 6577 * 6578 * WRITE_ONCE() is necessary because @worker->flags may be 6579 * tested without holding any lock in 6580 * wq_worker_running(). Without it, NOT_RUNNING test may 6581 * fail incorrectly leading to premature concurrency 6582 * management operations. 6583 */ 6584 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 6585 worker_flags |= WORKER_REBOUND; 6586 worker_flags &= ~WORKER_UNBOUND; 6587 WRITE_ONCE(worker->flags, worker_flags); 6588 } 6589 6590 raw_spin_unlock_irq(&pool->lock); 6591 } 6592 6593 /** 6594 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 6595 * @pool: unbound pool of interest 6596 * @cpu: the CPU which is coming up 6597 * 6598 * An unbound pool may end up with a cpumask which doesn't have any online 6599 * CPUs. When a worker of such pool get scheduled, the scheduler resets 6600 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 6601 * online CPU before, cpus_allowed of all its workers should be restored. 6602 */ 6603 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 6604 { 6605 static cpumask_t cpumask; 6606 struct worker *worker; 6607 6608 lockdep_assert_held(&wq_pool_attach_mutex); 6609 6610 /* is @cpu allowed for @pool? */ 6611 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 6612 return; 6613 6614 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 6615 6616 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 6617 for_each_pool_worker(worker, pool) 6618 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 6619 } 6620 6621 int workqueue_prepare_cpu(unsigned int cpu) 6622 { 6623 struct worker_pool *pool; 6624 6625 for_each_cpu_worker_pool(pool, cpu) { 6626 if (pool->nr_workers) 6627 continue; 6628 if (!create_worker(pool)) 6629 return -ENOMEM; 6630 } 6631 return 0; 6632 } 6633 6634 int workqueue_online_cpu(unsigned int cpu) 6635 { 6636 struct worker_pool *pool; 6637 struct workqueue_struct *wq; 6638 int pi; 6639 6640 mutex_lock(&wq_pool_mutex); 6641 6642 cpumask_set_cpu(cpu, wq_online_cpumask); 6643 6644 for_each_pool(pool, pi) { 6645 /* BH pools aren't affected by hotplug */ 6646 if (pool->flags & POOL_BH) 6647 continue; 6648 6649 mutex_lock(&wq_pool_attach_mutex); 6650 if (pool->cpu == cpu) 6651 rebind_workers(pool); 6652 else if (pool->cpu < 0) 6653 restore_unbound_workers_cpumask(pool, cpu); 6654 mutex_unlock(&wq_pool_attach_mutex); 6655 } 6656 6657 /* update pod affinity of unbound workqueues */ 6658 list_for_each_entry(wq, &workqueues, list) { 6659 struct workqueue_attrs *attrs = wq->unbound_attrs; 6660 6661 if (attrs) { 6662 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6663 int tcpu; 6664 6665 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6666 unbound_wq_update_pwq(wq, tcpu); 6667 6668 mutex_lock(&wq->mutex); 6669 wq_update_node_max_active(wq, -1); 6670 mutex_unlock(&wq->mutex); 6671 } 6672 } 6673 6674 mutex_unlock(&wq_pool_mutex); 6675 return 0; 6676 } 6677 6678 int workqueue_offline_cpu(unsigned int cpu) 6679 { 6680 struct workqueue_struct *wq; 6681 6682 /* unbinding per-cpu workers should happen on the local CPU */ 6683 if (WARN_ON(cpu != smp_processor_id())) 6684 return -1; 6685 6686 unbind_workers(cpu); 6687 6688 /* update pod affinity of unbound workqueues */ 6689 mutex_lock(&wq_pool_mutex); 6690 6691 cpumask_clear_cpu(cpu, wq_online_cpumask); 6692 6693 list_for_each_entry(wq, &workqueues, list) { 6694 struct workqueue_attrs *attrs = wq->unbound_attrs; 6695 6696 if (attrs) { 6697 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6698 int tcpu; 6699 6700 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6701 unbound_wq_update_pwq(wq, tcpu); 6702 6703 mutex_lock(&wq->mutex); 6704 wq_update_node_max_active(wq, cpu); 6705 mutex_unlock(&wq->mutex); 6706 } 6707 } 6708 mutex_unlock(&wq_pool_mutex); 6709 6710 return 0; 6711 } 6712 6713 struct work_for_cpu { 6714 struct work_struct work; 6715 long (*fn)(void *); 6716 void *arg; 6717 long ret; 6718 }; 6719 6720 static void work_for_cpu_fn(struct work_struct *work) 6721 { 6722 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 6723 6724 wfc->ret = wfc->fn(wfc->arg); 6725 } 6726 6727 /** 6728 * work_on_cpu_key - run a function in thread context on a particular cpu 6729 * @cpu: the cpu to run on 6730 * @fn: the function to run 6731 * @arg: the function arg 6732 * @key: The lock class key for lock debugging purposes 6733 * 6734 * It is up to the caller to ensure that the cpu doesn't go offline. 6735 * The caller must not hold any locks which would prevent @fn from completing. 6736 * 6737 * Return: The value @fn returns. 6738 */ 6739 long work_on_cpu_key(int cpu, long (*fn)(void *), 6740 void *arg, struct lock_class_key *key) 6741 { 6742 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 6743 6744 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); 6745 schedule_work_on(cpu, &wfc.work); 6746 flush_work(&wfc.work); 6747 destroy_work_on_stack(&wfc.work); 6748 return wfc.ret; 6749 } 6750 EXPORT_SYMBOL_GPL(work_on_cpu_key); 6751 6752 /** 6753 * work_on_cpu_safe_key - run a function in thread context on a particular cpu 6754 * @cpu: the cpu to run on 6755 * @fn: the function to run 6756 * @arg: the function argument 6757 * @key: The lock class key for lock debugging purposes 6758 * 6759 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 6760 * any locks which would prevent @fn from completing. 6761 * 6762 * Return: The value @fn returns. 6763 */ 6764 long work_on_cpu_safe_key(int cpu, long (*fn)(void *), 6765 void *arg, struct lock_class_key *key) 6766 { 6767 long ret = -ENODEV; 6768 6769 cpus_read_lock(); 6770 if (cpu_online(cpu)) 6771 ret = work_on_cpu_key(cpu, fn, arg, key); 6772 cpus_read_unlock(); 6773 return ret; 6774 } 6775 EXPORT_SYMBOL_GPL(work_on_cpu_safe_key); 6776 #endif /* CONFIG_SMP */ 6777 6778 #ifdef CONFIG_FREEZER 6779 6780 /** 6781 * freeze_workqueues_begin - begin freezing workqueues 6782 * 6783 * Start freezing workqueues. After this function returns, all freezable 6784 * workqueues will queue new works to their inactive_works list instead of 6785 * pool->worklist. 6786 * 6787 * CONTEXT: 6788 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6789 */ 6790 void freeze_workqueues_begin(void) 6791 { 6792 struct workqueue_struct *wq; 6793 6794 mutex_lock(&wq_pool_mutex); 6795 6796 WARN_ON_ONCE(workqueue_freezing); 6797 workqueue_freezing = true; 6798 6799 list_for_each_entry(wq, &workqueues, list) { 6800 mutex_lock(&wq->mutex); 6801 wq_adjust_max_active(wq); 6802 mutex_unlock(&wq->mutex); 6803 } 6804 6805 mutex_unlock(&wq_pool_mutex); 6806 } 6807 6808 /** 6809 * freeze_workqueues_busy - are freezable workqueues still busy? 6810 * 6811 * Check whether freezing is complete. This function must be called 6812 * between freeze_workqueues_begin() and thaw_workqueues(). 6813 * 6814 * CONTEXT: 6815 * Grabs and releases wq_pool_mutex. 6816 * 6817 * Return: 6818 * %true if some freezable workqueues are still busy. %false if freezing 6819 * is complete. 6820 */ 6821 bool freeze_workqueues_busy(void) 6822 { 6823 bool busy = false; 6824 struct workqueue_struct *wq; 6825 struct pool_workqueue *pwq; 6826 6827 mutex_lock(&wq_pool_mutex); 6828 6829 WARN_ON_ONCE(!workqueue_freezing); 6830 6831 list_for_each_entry(wq, &workqueues, list) { 6832 if (!(wq->flags & WQ_FREEZABLE)) 6833 continue; 6834 /* 6835 * nr_active is monotonically decreasing. It's safe 6836 * to peek without lock. 6837 */ 6838 rcu_read_lock(); 6839 for_each_pwq(pwq, wq) { 6840 WARN_ON_ONCE(pwq->nr_active < 0); 6841 if (pwq->nr_active) { 6842 busy = true; 6843 rcu_read_unlock(); 6844 goto out_unlock; 6845 } 6846 } 6847 rcu_read_unlock(); 6848 } 6849 out_unlock: 6850 mutex_unlock(&wq_pool_mutex); 6851 return busy; 6852 } 6853 6854 /** 6855 * thaw_workqueues - thaw workqueues 6856 * 6857 * Thaw workqueues. Normal queueing is restored and all collected 6858 * frozen works are transferred to their respective pool worklists. 6859 * 6860 * CONTEXT: 6861 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6862 */ 6863 void thaw_workqueues(void) 6864 { 6865 struct workqueue_struct *wq; 6866 6867 mutex_lock(&wq_pool_mutex); 6868 6869 if (!workqueue_freezing) 6870 goto out_unlock; 6871 6872 workqueue_freezing = false; 6873 6874 /* restore max_active and repopulate worklist */ 6875 list_for_each_entry(wq, &workqueues, list) { 6876 mutex_lock(&wq->mutex); 6877 wq_adjust_max_active(wq); 6878 mutex_unlock(&wq->mutex); 6879 } 6880 6881 out_unlock: 6882 mutex_unlock(&wq_pool_mutex); 6883 } 6884 #endif /* CONFIG_FREEZER */ 6885 6886 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 6887 { 6888 LIST_HEAD(ctxs); 6889 int ret = 0; 6890 struct workqueue_struct *wq; 6891 struct apply_wqattrs_ctx *ctx, *n; 6892 6893 lockdep_assert_held(&wq_pool_mutex); 6894 6895 list_for_each_entry(wq, &workqueues, list) { 6896 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING)) 6897 continue; 6898 6899 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 6900 if (IS_ERR(ctx)) { 6901 ret = PTR_ERR(ctx); 6902 break; 6903 } 6904 6905 list_add_tail(&ctx->list, &ctxs); 6906 } 6907 6908 list_for_each_entry_safe(ctx, n, &ctxs, list) { 6909 if (!ret) 6910 apply_wqattrs_commit(ctx); 6911 apply_wqattrs_cleanup(ctx); 6912 } 6913 6914 if (!ret) { 6915 mutex_lock(&wq_pool_attach_mutex); 6916 cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 6917 mutex_unlock(&wq_pool_attach_mutex); 6918 } 6919 return ret; 6920 } 6921 6922 /** 6923 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask 6924 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask 6925 * 6926 * This function can be called from cpuset code to provide a set of isolated 6927 * CPUs that should be excluded from wq_unbound_cpumask. 6928 */ 6929 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) 6930 { 6931 cpumask_var_t cpumask; 6932 int ret = 0; 6933 6934 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6935 return -ENOMEM; 6936 6937 mutex_lock(&wq_pool_mutex); 6938 6939 /* 6940 * If the operation fails, it will fall back to 6941 * wq_requested_unbound_cpumask which is initially set to 6942 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten 6943 * by any subsequent write to workqueue/cpumask sysfs file. 6944 */ 6945 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) 6946 cpumask_copy(cpumask, wq_requested_unbound_cpumask); 6947 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 6948 ret = workqueue_apply_unbound_cpumask(cpumask); 6949 6950 /* Save the current isolated cpumask & export it via sysfs */ 6951 if (!ret) 6952 cpumask_copy(wq_isolated_cpumask, exclude_cpumask); 6953 6954 mutex_unlock(&wq_pool_mutex); 6955 free_cpumask_var(cpumask); 6956 return ret; 6957 } 6958 6959 static int parse_affn_scope(const char *val) 6960 { 6961 int i; 6962 6963 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 6964 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 6965 return i; 6966 } 6967 return -EINVAL; 6968 } 6969 6970 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 6971 { 6972 struct workqueue_struct *wq; 6973 int affn, cpu; 6974 6975 affn = parse_affn_scope(val); 6976 if (affn < 0) 6977 return affn; 6978 if (affn == WQ_AFFN_DFL) 6979 return -EINVAL; 6980 6981 cpus_read_lock(); 6982 mutex_lock(&wq_pool_mutex); 6983 6984 wq_affn_dfl = affn; 6985 6986 list_for_each_entry(wq, &workqueues, list) { 6987 for_each_online_cpu(cpu) 6988 unbound_wq_update_pwq(wq, cpu); 6989 } 6990 6991 mutex_unlock(&wq_pool_mutex); 6992 cpus_read_unlock(); 6993 6994 return 0; 6995 } 6996 6997 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 6998 { 6999 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 7000 } 7001 7002 static const struct kernel_param_ops wq_affn_dfl_ops = { 7003 .set = wq_affn_dfl_set, 7004 .get = wq_affn_dfl_get, 7005 }; 7006 7007 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 7008 7009 #ifdef CONFIG_SYSFS 7010 /* 7011 * Workqueues with WQ_SYSFS flag set is visible to userland via 7012 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 7013 * following attributes. 7014 * 7015 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 7016 * max_active RW int : maximum number of in-flight work items 7017 * 7018 * Unbound workqueues have the following extra attributes. 7019 * 7020 * nice RW int : nice value of the workers 7021 * cpumask RW mask : bitmask of allowed CPUs for the workers 7022 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 7023 * affinity_strict RW bool : worker CPU affinity is strict 7024 */ 7025 struct wq_device { 7026 struct workqueue_struct *wq; 7027 struct device dev; 7028 }; 7029 7030 static struct workqueue_struct *dev_to_wq(struct device *dev) 7031 { 7032 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7033 7034 return wq_dev->wq; 7035 } 7036 7037 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 7038 char *buf) 7039 { 7040 struct workqueue_struct *wq = dev_to_wq(dev); 7041 7042 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 7043 } 7044 static DEVICE_ATTR_RO(per_cpu); 7045 7046 static ssize_t max_active_show(struct device *dev, 7047 struct device_attribute *attr, char *buf) 7048 { 7049 struct workqueue_struct *wq = dev_to_wq(dev); 7050 7051 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 7052 } 7053 7054 static ssize_t max_active_store(struct device *dev, 7055 struct device_attribute *attr, const char *buf, 7056 size_t count) 7057 { 7058 struct workqueue_struct *wq = dev_to_wq(dev); 7059 int val; 7060 7061 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 7062 return -EINVAL; 7063 7064 workqueue_set_max_active(wq, val); 7065 return count; 7066 } 7067 static DEVICE_ATTR_RW(max_active); 7068 7069 static struct attribute *wq_sysfs_attrs[] = { 7070 &dev_attr_per_cpu.attr, 7071 &dev_attr_max_active.attr, 7072 NULL, 7073 }; 7074 ATTRIBUTE_GROUPS(wq_sysfs); 7075 7076 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 7077 char *buf) 7078 { 7079 struct workqueue_struct *wq = dev_to_wq(dev); 7080 int written; 7081 7082 mutex_lock(&wq->mutex); 7083 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 7084 mutex_unlock(&wq->mutex); 7085 7086 return written; 7087 } 7088 7089 /* prepare workqueue_attrs for sysfs store operations */ 7090 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 7091 { 7092 struct workqueue_attrs *attrs; 7093 7094 lockdep_assert_held(&wq_pool_mutex); 7095 7096 attrs = alloc_workqueue_attrs(); 7097 if (!attrs) 7098 return NULL; 7099 7100 copy_workqueue_attrs(attrs, wq->unbound_attrs); 7101 return attrs; 7102 } 7103 7104 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 7105 const char *buf, size_t count) 7106 { 7107 struct workqueue_struct *wq = dev_to_wq(dev); 7108 struct workqueue_attrs *attrs; 7109 int ret = -ENOMEM; 7110 7111 apply_wqattrs_lock(); 7112 7113 attrs = wq_sysfs_prep_attrs(wq); 7114 if (!attrs) 7115 goto out_unlock; 7116 7117 if (sscanf(buf, "%d", &attrs->nice) == 1 && 7118 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 7119 ret = apply_workqueue_attrs_locked(wq, attrs); 7120 else 7121 ret = -EINVAL; 7122 7123 out_unlock: 7124 apply_wqattrs_unlock(); 7125 free_workqueue_attrs(attrs); 7126 return ret ?: count; 7127 } 7128 7129 static ssize_t wq_cpumask_show(struct device *dev, 7130 struct device_attribute *attr, char *buf) 7131 { 7132 struct workqueue_struct *wq = dev_to_wq(dev); 7133 int written; 7134 7135 mutex_lock(&wq->mutex); 7136 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 7137 cpumask_pr_args(wq->unbound_attrs->cpumask)); 7138 mutex_unlock(&wq->mutex); 7139 return written; 7140 } 7141 7142 static ssize_t wq_cpumask_store(struct device *dev, 7143 struct device_attribute *attr, 7144 const char *buf, size_t count) 7145 { 7146 struct workqueue_struct *wq = dev_to_wq(dev); 7147 struct workqueue_attrs *attrs; 7148 int ret = -ENOMEM; 7149 7150 apply_wqattrs_lock(); 7151 7152 attrs = wq_sysfs_prep_attrs(wq); 7153 if (!attrs) 7154 goto out_unlock; 7155 7156 ret = cpumask_parse(buf, attrs->cpumask); 7157 if (!ret) 7158 ret = apply_workqueue_attrs_locked(wq, attrs); 7159 7160 out_unlock: 7161 apply_wqattrs_unlock(); 7162 free_workqueue_attrs(attrs); 7163 return ret ?: count; 7164 } 7165 7166 static ssize_t wq_affn_scope_show(struct device *dev, 7167 struct device_attribute *attr, char *buf) 7168 { 7169 struct workqueue_struct *wq = dev_to_wq(dev); 7170 int written; 7171 7172 mutex_lock(&wq->mutex); 7173 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 7174 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 7175 wq_affn_names[WQ_AFFN_DFL], 7176 wq_affn_names[wq_affn_dfl]); 7177 else 7178 written = scnprintf(buf, PAGE_SIZE, "%s\n", 7179 wq_affn_names[wq->unbound_attrs->affn_scope]); 7180 mutex_unlock(&wq->mutex); 7181 7182 return written; 7183 } 7184 7185 static ssize_t wq_affn_scope_store(struct device *dev, 7186 struct device_attribute *attr, 7187 const char *buf, size_t count) 7188 { 7189 struct workqueue_struct *wq = dev_to_wq(dev); 7190 struct workqueue_attrs *attrs; 7191 int affn, ret = -ENOMEM; 7192 7193 affn = parse_affn_scope(buf); 7194 if (affn < 0) 7195 return affn; 7196 7197 apply_wqattrs_lock(); 7198 attrs = wq_sysfs_prep_attrs(wq); 7199 if (attrs) { 7200 attrs->affn_scope = affn; 7201 ret = apply_workqueue_attrs_locked(wq, attrs); 7202 } 7203 apply_wqattrs_unlock(); 7204 free_workqueue_attrs(attrs); 7205 return ret ?: count; 7206 } 7207 7208 static ssize_t wq_affinity_strict_show(struct device *dev, 7209 struct device_attribute *attr, char *buf) 7210 { 7211 struct workqueue_struct *wq = dev_to_wq(dev); 7212 7213 return scnprintf(buf, PAGE_SIZE, "%d\n", 7214 wq->unbound_attrs->affn_strict); 7215 } 7216 7217 static ssize_t wq_affinity_strict_store(struct device *dev, 7218 struct device_attribute *attr, 7219 const char *buf, size_t count) 7220 { 7221 struct workqueue_struct *wq = dev_to_wq(dev); 7222 struct workqueue_attrs *attrs; 7223 int v, ret = -ENOMEM; 7224 7225 if (sscanf(buf, "%d", &v) != 1) 7226 return -EINVAL; 7227 7228 apply_wqattrs_lock(); 7229 attrs = wq_sysfs_prep_attrs(wq); 7230 if (attrs) { 7231 attrs->affn_strict = (bool)v; 7232 ret = apply_workqueue_attrs_locked(wq, attrs); 7233 } 7234 apply_wqattrs_unlock(); 7235 free_workqueue_attrs(attrs); 7236 return ret ?: count; 7237 } 7238 7239 static struct device_attribute wq_sysfs_unbound_attrs[] = { 7240 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 7241 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 7242 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 7243 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 7244 __ATTR_NULL, 7245 }; 7246 7247 static const struct bus_type wq_subsys = { 7248 .name = "workqueue", 7249 .dev_groups = wq_sysfs_groups, 7250 }; 7251 7252 /** 7253 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 7254 * @cpumask: the cpumask to set 7255 * 7256 * The low-level workqueues cpumask is a global cpumask that limits 7257 * the affinity of all unbound workqueues. This function check the @cpumask 7258 * and apply it to all unbound workqueues and updates all pwqs of them. 7259 * 7260 * Return: 0 - Success 7261 * -EINVAL - Invalid @cpumask 7262 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 7263 */ 7264 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 7265 { 7266 int ret = -EINVAL; 7267 7268 /* 7269 * Not excluding isolated cpus on purpose. 7270 * If the user wishes to include them, we allow that. 7271 */ 7272 cpumask_and(cpumask, cpumask, cpu_possible_mask); 7273 if (!cpumask_empty(cpumask)) { 7274 ret = 0; 7275 apply_wqattrs_lock(); 7276 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 7277 ret = workqueue_apply_unbound_cpumask(cpumask); 7278 if (!ret) 7279 cpumask_copy(wq_requested_unbound_cpumask, cpumask); 7280 apply_wqattrs_unlock(); 7281 } 7282 7283 return ret; 7284 } 7285 7286 static ssize_t __wq_cpumask_show(struct device *dev, 7287 struct device_attribute *attr, char *buf, cpumask_var_t mask) 7288 { 7289 int written; 7290 7291 mutex_lock(&wq_pool_mutex); 7292 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); 7293 mutex_unlock(&wq_pool_mutex); 7294 7295 return written; 7296 } 7297 7298 static ssize_t cpumask_requested_show(struct device *dev, 7299 struct device_attribute *attr, char *buf) 7300 { 7301 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); 7302 } 7303 static DEVICE_ATTR_RO(cpumask_requested); 7304 7305 static ssize_t cpumask_isolated_show(struct device *dev, 7306 struct device_attribute *attr, char *buf) 7307 { 7308 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); 7309 } 7310 static DEVICE_ATTR_RO(cpumask_isolated); 7311 7312 static ssize_t cpumask_show(struct device *dev, 7313 struct device_attribute *attr, char *buf) 7314 { 7315 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); 7316 } 7317 7318 static ssize_t cpumask_store(struct device *dev, 7319 struct device_attribute *attr, const char *buf, size_t count) 7320 { 7321 cpumask_var_t cpumask; 7322 int ret; 7323 7324 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 7325 return -ENOMEM; 7326 7327 ret = cpumask_parse(buf, cpumask); 7328 if (!ret) 7329 ret = workqueue_set_unbound_cpumask(cpumask); 7330 7331 free_cpumask_var(cpumask); 7332 return ret ? ret : count; 7333 } 7334 static DEVICE_ATTR_RW(cpumask); 7335 7336 static struct attribute *wq_sysfs_cpumask_attrs[] = { 7337 &dev_attr_cpumask.attr, 7338 &dev_attr_cpumask_requested.attr, 7339 &dev_attr_cpumask_isolated.attr, 7340 NULL, 7341 }; 7342 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); 7343 7344 static int __init wq_sysfs_init(void) 7345 { 7346 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups); 7347 } 7348 core_initcall(wq_sysfs_init); 7349 7350 static void wq_device_release(struct device *dev) 7351 { 7352 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7353 7354 kfree(wq_dev); 7355 } 7356 7357 /** 7358 * workqueue_sysfs_register - make a workqueue visible in sysfs 7359 * @wq: the workqueue to register 7360 * 7361 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 7362 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 7363 * which is the preferred method. 7364 * 7365 * Workqueue user should use this function directly iff it wants to apply 7366 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 7367 * apply_workqueue_attrs() may race against userland updating the 7368 * attributes. 7369 * 7370 * Return: 0 on success, -errno on failure. 7371 */ 7372 int workqueue_sysfs_register(struct workqueue_struct *wq) 7373 { 7374 struct wq_device *wq_dev; 7375 int ret; 7376 7377 /* 7378 * Adjusting max_active breaks ordering guarantee. Disallow exposing 7379 * ordered workqueues. 7380 */ 7381 if (WARN_ON(wq->flags & __WQ_ORDERED)) 7382 return -EINVAL; 7383 7384 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 7385 if (!wq_dev) 7386 return -ENOMEM; 7387 7388 wq_dev->wq = wq; 7389 wq_dev->dev.bus = &wq_subsys; 7390 wq_dev->dev.release = wq_device_release; 7391 dev_set_name(&wq_dev->dev, "%s", wq->name); 7392 7393 /* 7394 * unbound_attrs are created separately. Suppress uevent until 7395 * everything is ready. 7396 */ 7397 dev_set_uevent_suppress(&wq_dev->dev, true); 7398 7399 ret = device_register(&wq_dev->dev); 7400 if (ret) { 7401 put_device(&wq_dev->dev); 7402 wq->wq_dev = NULL; 7403 return ret; 7404 } 7405 7406 if (wq->flags & WQ_UNBOUND) { 7407 struct device_attribute *attr; 7408 7409 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 7410 ret = device_create_file(&wq_dev->dev, attr); 7411 if (ret) { 7412 device_unregister(&wq_dev->dev); 7413 wq->wq_dev = NULL; 7414 return ret; 7415 } 7416 } 7417 } 7418 7419 dev_set_uevent_suppress(&wq_dev->dev, false); 7420 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 7421 return 0; 7422 } 7423 7424 /** 7425 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 7426 * @wq: the workqueue to unregister 7427 * 7428 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 7429 */ 7430 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 7431 { 7432 struct wq_device *wq_dev = wq->wq_dev; 7433 7434 if (!wq->wq_dev) 7435 return; 7436 7437 wq->wq_dev = NULL; 7438 device_unregister(&wq_dev->dev); 7439 } 7440 #else /* CONFIG_SYSFS */ 7441 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 7442 #endif /* CONFIG_SYSFS */ 7443 7444 /* 7445 * Workqueue watchdog. 7446 * 7447 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 7448 * flush dependency, a concurrency managed work item which stays RUNNING 7449 * indefinitely. Workqueue stalls can be very difficult to debug as the 7450 * usual warning mechanisms don't trigger and internal workqueue state is 7451 * largely opaque. 7452 * 7453 * Workqueue watchdog monitors all worker pools periodically and dumps 7454 * state if some pools failed to make forward progress for a while where 7455 * forward progress is defined as the first item on ->worklist changing. 7456 * 7457 * This mechanism is controlled through the kernel parameter 7458 * "workqueue.watchdog_thresh" which can be updated at runtime through the 7459 * corresponding sysfs parameter file. 7460 */ 7461 #ifdef CONFIG_WQ_WATCHDOG 7462 7463 static unsigned long wq_watchdog_thresh = 30; 7464 static struct timer_list wq_watchdog_timer; 7465 7466 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 7467 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 7468 7469 static unsigned int wq_panic_on_stall; 7470 module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644); 7471 7472 /* 7473 * Show workers that might prevent the processing of pending work items. 7474 * The only candidates are CPU-bound workers in the running state. 7475 * Pending work items should be handled by another idle worker 7476 * in all other situations. 7477 */ 7478 static void show_cpu_pool_hog(struct worker_pool *pool) 7479 { 7480 struct worker *worker; 7481 unsigned long irq_flags; 7482 int bkt; 7483 7484 raw_spin_lock_irqsave(&pool->lock, irq_flags); 7485 7486 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 7487 if (task_is_running(worker->task)) { 7488 /* 7489 * Defer printing to avoid deadlocks in console 7490 * drivers that queue work while holding locks 7491 * also taken in their write paths. 7492 */ 7493 printk_deferred_enter(); 7494 7495 pr_info("pool %d:\n", pool->id); 7496 sched_show_task(worker->task); 7497 7498 printk_deferred_exit(); 7499 } 7500 } 7501 7502 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 7503 } 7504 7505 static void show_cpu_pools_hogs(void) 7506 { 7507 struct worker_pool *pool; 7508 int pi; 7509 7510 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 7511 7512 rcu_read_lock(); 7513 7514 for_each_pool(pool, pi) { 7515 if (pool->cpu_stall) 7516 show_cpu_pool_hog(pool); 7517 7518 } 7519 7520 rcu_read_unlock(); 7521 } 7522 7523 static void panic_on_wq_watchdog(void) 7524 { 7525 static unsigned int wq_stall; 7526 7527 if (wq_panic_on_stall) { 7528 wq_stall++; 7529 BUG_ON(wq_stall >= wq_panic_on_stall); 7530 } 7531 } 7532 7533 static void wq_watchdog_reset_touched(void) 7534 { 7535 int cpu; 7536 7537 wq_watchdog_touched = jiffies; 7538 for_each_possible_cpu(cpu) 7539 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 7540 } 7541 7542 static void wq_watchdog_timer_fn(struct timer_list *unused) 7543 { 7544 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7545 bool lockup_detected = false; 7546 bool cpu_pool_stall = false; 7547 unsigned long now = jiffies; 7548 struct worker_pool *pool; 7549 int pi; 7550 7551 if (!thresh) 7552 return; 7553 7554 rcu_read_lock(); 7555 7556 for_each_pool(pool, pi) { 7557 unsigned long pool_ts, touched, ts; 7558 7559 pool->cpu_stall = false; 7560 if (list_empty(&pool->worklist)) 7561 continue; 7562 7563 /* 7564 * If a virtual machine is stopped by the host it can look to 7565 * the watchdog like a stall. 7566 */ 7567 kvm_check_and_clear_guest_paused(); 7568 7569 /* get the latest of pool and touched timestamps */ 7570 if (pool->cpu >= 0) 7571 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 7572 else 7573 touched = READ_ONCE(wq_watchdog_touched); 7574 pool_ts = READ_ONCE(pool->watchdog_ts); 7575 7576 if (time_after(pool_ts, touched)) 7577 ts = pool_ts; 7578 else 7579 ts = touched; 7580 7581 /* did we stall? */ 7582 if (time_after(now, ts + thresh)) { 7583 lockup_detected = true; 7584 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) { 7585 pool->cpu_stall = true; 7586 cpu_pool_stall = true; 7587 } 7588 pr_emerg("BUG: workqueue lockup - pool"); 7589 pr_cont_pool_info(pool); 7590 pr_cont(" stuck for %us!\n", 7591 jiffies_to_msecs(now - pool_ts) / 1000); 7592 } 7593 7594 7595 } 7596 7597 rcu_read_unlock(); 7598 7599 if (lockup_detected) 7600 show_all_workqueues(); 7601 7602 if (cpu_pool_stall) 7603 show_cpu_pools_hogs(); 7604 7605 if (lockup_detected) 7606 panic_on_wq_watchdog(); 7607 7608 wq_watchdog_reset_touched(); 7609 mod_timer(&wq_watchdog_timer, jiffies + thresh); 7610 } 7611 7612 notrace void wq_watchdog_touch(int cpu) 7613 { 7614 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7615 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched); 7616 unsigned long now = jiffies; 7617 7618 if (cpu >= 0) 7619 per_cpu(wq_watchdog_touched_cpu, cpu) = now; 7620 else 7621 WARN_ONCE(1, "%s should be called with valid CPU", __func__); 7622 7623 /* Don't unnecessarily store to global cacheline */ 7624 if (time_after(now, touch_ts + thresh / 4)) 7625 WRITE_ONCE(wq_watchdog_touched, jiffies); 7626 } 7627 7628 static void wq_watchdog_set_thresh(unsigned long thresh) 7629 { 7630 wq_watchdog_thresh = 0; 7631 del_timer_sync(&wq_watchdog_timer); 7632 7633 if (thresh) { 7634 wq_watchdog_thresh = thresh; 7635 wq_watchdog_reset_touched(); 7636 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 7637 } 7638 } 7639 7640 static int wq_watchdog_param_set_thresh(const char *val, 7641 const struct kernel_param *kp) 7642 { 7643 unsigned long thresh; 7644 int ret; 7645 7646 ret = kstrtoul(val, 0, &thresh); 7647 if (ret) 7648 return ret; 7649 7650 if (system_wq) 7651 wq_watchdog_set_thresh(thresh); 7652 else 7653 wq_watchdog_thresh = thresh; 7654 7655 return 0; 7656 } 7657 7658 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 7659 .set = wq_watchdog_param_set_thresh, 7660 .get = param_get_ulong, 7661 }; 7662 7663 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 7664 0644); 7665 7666 static void wq_watchdog_init(void) 7667 { 7668 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 7669 wq_watchdog_set_thresh(wq_watchdog_thresh); 7670 } 7671 7672 #else /* CONFIG_WQ_WATCHDOG */ 7673 7674 static inline void wq_watchdog_init(void) { } 7675 7676 #endif /* CONFIG_WQ_WATCHDOG */ 7677 7678 static void bh_pool_kick_normal(struct irq_work *irq_work) 7679 { 7680 raise_softirq_irqoff(TASKLET_SOFTIRQ); 7681 } 7682 7683 static void bh_pool_kick_highpri(struct irq_work *irq_work) 7684 { 7685 raise_softirq_irqoff(HI_SOFTIRQ); 7686 } 7687 7688 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) 7689 { 7690 if (!cpumask_intersects(wq_unbound_cpumask, mask)) { 7691 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", 7692 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); 7693 return; 7694 } 7695 7696 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); 7697 } 7698 7699 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice) 7700 { 7701 BUG_ON(init_worker_pool(pool)); 7702 pool->cpu = cpu; 7703 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 7704 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 7705 pool->attrs->nice = nice; 7706 pool->attrs->affn_strict = true; 7707 pool->node = cpu_to_node(cpu); 7708 7709 /* alloc pool ID */ 7710 mutex_lock(&wq_pool_mutex); 7711 BUG_ON(worker_pool_assign_id(pool)); 7712 mutex_unlock(&wq_pool_mutex); 7713 } 7714 7715 /** 7716 * workqueue_init_early - early init for workqueue subsystem 7717 * 7718 * This is the first step of three-staged workqueue subsystem initialization and 7719 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 7720 * up. It sets up all the data structures and system workqueues and allows early 7721 * boot code to create workqueues and queue/cancel work items. Actual work item 7722 * execution starts only after kthreads can be created and scheduled right 7723 * before early initcalls. 7724 */ 7725 void __init workqueue_init_early(void) 7726 { 7727 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 7728 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 7729 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal, 7730 bh_pool_kick_highpri }; 7731 int i, cpu; 7732 7733 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 7734 7735 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL)); 7736 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 7737 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); 7738 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); 7739 7740 cpumask_copy(wq_online_cpumask, cpu_online_mask); 7741 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 7742 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); 7743 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); 7744 if (!cpumask_empty(&wq_cmdline_cpumask)) 7745 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); 7746 7747 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); 7748 7749 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 7750 7751 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs(); 7752 BUG_ON(!unbound_wq_update_pwq_attrs_buf); 7753 7754 /* 7755 * If nohz_full is enabled, set power efficient workqueue as unbound. 7756 * This allows workqueue items to be moved to HK CPUs. 7757 */ 7758 if (housekeeping_enabled(HK_TYPE_TICK)) 7759 wq_power_efficient = true; 7760 7761 /* initialize WQ_AFFN_SYSTEM pods */ 7762 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 7763 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 7764 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 7765 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 7766 7767 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 7768 7769 pt->nr_pods = 1; 7770 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 7771 pt->pod_node[0] = NUMA_NO_NODE; 7772 pt->cpu_pod[0] = 0; 7773 7774 /* initialize BH and CPU pools */ 7775 for_each_possible_cpu(cpu) { 7776 struct worker_pool *pool; 7777 7778 i = 0; 7779 for_each_bh_worker_pool(pool, cpu) { 7780 init_cpu_worker_pool(pool, cpu, std_nice[i]); 7781 pool->flags |= POOL_BH; 7782 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]); 7783 i++; 7784 } 7785 7786 i = 0; 7787 for_each_cpu_worker_pool(pool, cpu) 7788 init_cpu_worker_pool(pool, cpu, std_nice[i++]); 7789 } 7790 7791 /* create default unbound and ordered wq attrs */ 7792 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 7793 struct workqueue_attrs *attrs; 7794 7795 BUG_ON(!(attrs = alloc_workqueue_attrs())); 7796 attrs->nice = std_nice[i]; 7797 unbound_std_wq_attrs[i] = attrs; 7798 7799 /* 7800 * An ordered wq should have only one pwq as ordering is 7801 * guaranteed by max_active which is enforced by pwqs. 7802 */ 7803 BUG_ON(!(attrs = alloc_workqueue_attrs())); 7804 attrs->nice = std_nice[i]; 7805 attrs->ordered = true; 7806 ordered_wq_attrs[i] = attrs; 7807 } 7808 7809 system_wq = alloc_workqueue("events", 0, 0); 7810 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 7811 system_long_wq = alloc_workqueue("events_long", 0, 0); 7812 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 7813 WQ_MAX_ACTIVE); 7814 system_freezable_wq = alloc_workqueue("events_freezable", 7815 WQ_FREEZABLE, 0); 7816 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 7817 WQ_POWER_EFFICIENT, 0); 7818 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", 7819 WQ_FREEZABLE | WQ_POWER_EFFICIENT, 7820 0); 7821 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0); 7822 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri", 7823 WQ_BH | WQ_HIGHPRI, 0); 7824 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 7825 !system_unbound_wq || !system_freezable_wq || 7826 !system_power_efficient_wq || 7827 !system_freezable_power_efficient_wq || 7828 !system_bh_wq || !system_bh_highpri_wq); 7829 } 7830 7831 static void __init wq_cpu_intensive_thresh_init(void) 7832 { 7833 unsigned long thresh; 7834 unsigned long bogo; 7835 7836 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 7837 BUG_ON(IS_ERR(pwq_release_worker)); 7838 7839 /* if the user set it to a specific value, keep it */ 7840 if (wq_cpu_intensive_thresh_us != ULONG_MAX) 7841 return; 7842 7843 /* 7844 * The default of 10ms is derived from the fact that most modern (as of 7845 * 2023) processors can do a lot in 10ms and that it's just below what 7846 * most consider human-perceivable. However, the kernel also runs on a 7847 * lot slower CPUs including microcontrollers where the threshold is way 7848 * too low. 7849 * 7850 * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 7851 * This is by no means accurate but it doesn't have to be. The mechanism 7852 * is still useful even when the threshold is fully scaled up. Also, as 7853 * the reports would usually be applicable to everyone, some machines 7854 * operating on longer thresholds won't significantly diminish their 7855 * usefulness. 7856 */ 7857 thresh = 10 * USEC_PER_MSEC; 7858 7859 /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 7860 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 7861 if (bogo < 4000) 7862 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 7863 7864 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 7865 loops_per_jiffy, bogo, thresh); 7866 7867 wq_cpu_intensive_thresh_us = thresh; 7868 } 7869 7870 /** 7871 * workqueue_init - bring workqueue subsystem fully online 7872 * 7873 * This is the second step of three-staged workqueue subsystem initialization 7874 * and invoked as soon as kthreads can be created and scheduled. Workqueues have 7875 * been created and work items queued on them, but there are no kworkers 7876 * executing the work items yet. Populate the worker pools with the initial 7877 * workers and enable future kworker creations. 7878 */ 7879 void __init workqueue_init(void) 7880 { 7881 struct workqueue_struct *wq; 7882 struct worker_pool *pool; 7883 int cpu, bkt; 7884 7885 wq_cpu_intensive_thresh_init(); 7886 7887 mutex_lock(&wq_pool_mutex); 7888 7889 /* 7890 * Per-cpu pools created earlier could be missing node hint. Fix them 7891 * up. Also, create a rescuer for workqueues that requested it. 7892 */ 7893 for_each_possible_cpu(cpu) { 7894 for_each_bh_worker_pool(pool, cpu) 7895 pool->node = cpu_to_node(cpu); 7896 for_each_cpu_worker_pool(pool, cpu) 7897 pool->node = cpu_to_node(cpu); 7898 } 7899 7900 list_for_each_entry(wq, &workqueues, list) { 7901 WARN(init_rescuer(wq), 7902 "workqueue: failed to create early rescuer for %s", 7903 wq->name); 7904 } 7905 7906 mutex_unlock(&wq_pool_mutex); 7907 7908 /* 7909 * Create the initial workers. A BH pool has one pseudo worker that 7910 * represents the shared BH execution context and thus doesn't get 7911 * affected by hotplug events. Create the BH pseudo workers for all 7912 * possible CPUs here. 7913 */ 7914 for_each_possible_cpu(cpu) 7915 for_each_bh_worker_pool(pool, cpu) 7916 BUG_ON(!create_worker(pool)); 7917 7918 for_each_online_cpu(cpu) { 7919 for_each_cpu_worker_pool(pool, cpu) { 7920 pool->flags &= ~POOL_DISASSOCIATED; 7921 BUG_ON(!create_worker(pool)); 7922 } 7923 } 7924 7925 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 7926 BUG_ON(!create_worker(pool)); 7927 7928 wq_online = true; 7929 wq_watchdog_init(); 7930 } 7931 7932 /* 7933 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 7934 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 7935 * and consecutive pod ID. The rest of @pt is initialized accordingly. 7936 */ 7937 static void __init init_pod_type(struct wq_pod_type *pt, 7938 bool (*cpus_share_pod)(int, int)) 7939 { 7940 int cur, pre, cpu, pod; 7941 7942 pt->nr_pods = 0; 7943 7944 /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 7945 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 7946 BUG_ON(!pt->cpu_pod); 7947 7948 for_each_possible_cpu(cur) { 7949 for_each_possible_cpu(pre) { 7950 if (pre >= cur) { 7951 pt->cpu_pod[cur] = pt->nr_pods++; 7952 break; 7953 } 7954 if (cpus_share_pod(cur, pre)) { 7955 pt->cpu_pod[cur] = pt->cpu_pod[pre]; 7956 break; 7957 } 7958 } 7959 } 7960 7961 /* init the rest to match @pt->cpu_pod[] */ 7962 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 7963 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 7964 BUG_ON(!pt->pod_cpus || !pt->pod_node); 7965 7966 for (pod = 0; pod < pt->nr_pods; pod++) 7967 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 7968 7969 for_each_possible_cpu(cpu) { 7970 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 7971 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 7972 } 7973 } 7974 7975 static bool __init cpus_dont_share(int cpu0, int cpu1) 7976 { 7977 return false; 7978 } 7979 7980 static bool __init cpus_share_smt(int cpu0, int cpu1) 7981 { 7982 #ifdef CONFIG_SCHED_SMT 7983 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 7984 #else 7985 return false; 7986 #endif 7987 } 7988 7989 static bool __init cpus_share_numa(int cpu0, int cpu1) 7990 { 7991 return cpu_to_node(cpu0) == cpu_to_node(cpu1); 7992 } 7993 7994 /** 7995 * workqueue_init_topology - initialize CPU pods for unbound workqueues 7996 * 7997 * This is the third step of three-staged workqueue subsystem initialization and 7998 * invoked after SMP and topology information are fully initialized. It 7999 * initializes the unbound CPU pods accordingly. 8000 */ 8001 void __init workqueue_init_topology(void) 8002 { 8003 struct workqueue_struct *wq; 8004 int cpu; 8005 8006 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 8007 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 8008 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 8009 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 8010 8011 wq_topo_initialized = true; 8012 8013 mutex_lock(&wq_pool_mutex); 8014 8015 /* 8016 * Workqueues allocated earlier would have all CPUs sharing the default 8017 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue 8018 * and CPU combinations to apply per-pod sharing. 8019 */ 8020 list_for_each_entry(wq, &workqueues, list) { 8021 for_each_online_cpu(cpu) 8022 unbound_wq_update_pwq(wq, cpu); 8023 if (wq->flags & WQ_UNBOUND) { 8024 mutex_lock(&wq->mutex); 8025 wq_update_node_max_active(wq, -1); 8026 mutex_unlock(&wq->mutex); 8027 } 8028 } 8029 8030 mutex_unlock(&wq_pool_mutex); 8031 } 8032 8033 void __warn_flushing_systemwide_wq(void) 8034 { 8035 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 8036 dump_stack(); 8037 } 8038 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 8039 8040 static int __init workqueue_unbound_cpus_setup(char *str) 8041 { 8042 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 8043 cpumask_clear(&wq_cmdline_cpumask); 8044 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 8045 } 8046 8047 return 1; 8048 } 8049 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 8050