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 #ifdef CONFIG_PREEMPT_RT 226 spinlock_t cb_lock; /* BH worker cancel lock */ 227 #endif 228 /* 229 * Destruction of pool is RCU protected to allow dereferences 230 * from get_work_pool(). 231 */ 232 struct rcu_head rcu; 233 }; 234 235 /* 236 * Per-pool_workqueue statistics. These can be monitored using 237 * tools/workqueue/wq_monitor.py. 238 */ 239 enum pool_workqueue_stats { 240 PWQ_STAT_STARTED, /* work items started execution */ 241 PWQ_STAT_COMPLETED, /* work items completed execution */ 242 PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 243 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 244 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 245 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ 246 PWQ_STAT_MAYDAY, /* maydays to rescuer */ 247 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 248 249 PWQ_NR_STATS, 250 }; 251 252 /* 253 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT 254 * of work_struct->data are used for flags and the remaining high bits 255 * point to the pwq; thus, pwqs need to be aligned at two's power of the 256 * number of flag bits. 257 */ 258 struct pool_workqueue { 259 struct worker_pool *pool; /* I: the associated pool */ 260 struct workqueue_struct *wq; /* I: the owning workqueue */ 261 int work_color; /* L: current color */ 262 int flush_color; /* L: flushing color */ 263 int refcnt; /* L: reference count */ 264 int nr_in_flight[WORK_NR_COLORS]; 265 /* L: nr of in_flight works */ 266 bool plugged; /* L: execution suspended */ 267 268 /* 269 * nr_active management and WORK_STRUCT_INACTIVE: 270 * 271 * When pwq->nr_active >= max_active, new work item is queued to 272 * pwq->inactive_works instead of pool->worklist and marked with 273 * WORK_STRUCT_INACTIVE. 274 * 275 * All work items marked with WORK_STRUCT_INACTIVE do not participate in 276 * nr_active and all work items in pwq->inactive_works are marked with 277 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are 278 * in pwq->inactive_works. Some of them are ready to run in 279 * pool->worklist or worker->scheduled. Those work itmes are only struct 280 * wq_barrier which is used for flush_work() and should not participate 281 * in nr_active. For non-barrier work item, it is marked with 282 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 283 */ 284 int nr_active; /* L: nr of active works */ 285 struct list_head inactive_works; /* L: inactive works */ 286 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */ 287 struct list_head pwqs_node; /* WR: node on wq->pwqs */ 288 struct list_head mayday_node; /* MD: node on wq->maydays */ 289 290 u64 stats[PWQ_NR_STATS]; 291 292 /* 293 * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 294 * and pwq_release_workfn() for details. pool_workqueue itself is also 295 * RCU protected so that the first pwq can be determined without 296 * grabbing wq->mutex. 297 */ 298 struct kthread_work release_work; 299 struct rcu_head rcu; 300 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); 301 302 /* 303 * Structure used to wait for workqueue flush. 304 */ 305 struct wq_flusher { 306 struct list_head list; /* WQ: list of flushers */ 307 int flush_color; /* WQ: flush color waiting for */ 308 struct completion done; /* flush completion */ 309 }; 310 311 struct wq_device; 312 313 /* 314 * Unlike in a per-cpu workqueue where max_active limits its concurrency level 315 * on each CPU, in an unbound workqueue, max_active applies to the whole system. 316 * As sharing a single nr_active across multiple sockets can be very expensive, 317 * the counting and enforcement is per NUMA node. 318 * 319 * The following struct is used to enforce per-node max_active. When a pwq wants 320 * to start executing a work item, it should increment ->nr using 321 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over 322 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish 323 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in 324 * round-robin order. 325 */ 326 struct wq_node_nr_active { 327 int max; /* per-node max_active */ 328 atomic_t nr; /* per-node nr_active */ 329 raw_spinlock_t lock; /* nests inside pool locks */ 330 struct list_head pending_pwqs; /* LN: pwqs with inactive works */ 331 }; 332 333 /* 334 * The externally visible workqueue. It relays the issued work items to 335 * the appropriate worker_pool through its pool_workqueues. 336 */ 337 struct workqueue_struct { 338 struct list_head pwqs; /* WR: all pwqs of this wq */ 339 struct list_head list; /* PR: list of all workqueues */ 340 341 struct mutex mutex; /* protects this wq */ 342 int work_color; /* WQ: current work color */ 343 int flush_color; /* WQ: current flush color */ 344 atomic_t nr_pwqs_to_flush; /* flush in progress */ 345 struct wq_flusher *first_flusher; /* WQ: first flusher */ 346 struct list_head flusher_queue; /* WQ: flush waiters */ 347 struct list_head flusher_overflow; /* WQ: flush overflow list */ 348 349 struct list_head maydays; /* MD: pwqs requesting rescue */ 350 struct worker *rescuer; /* MD: rescue worker */ 351 352 int nr_drainers; /* WQ: drain in progress */ 353 354 /* See alloc_workqueue() function comment for info on min/max_active */ 355 int max_active; /* WO: max active works */ 356 int min_active; /* WO: min active works */ 357 int saved_max_active; /* WQ: saved max_active */ 358 int saved_min_active; /* WQ: saved min_active */ 359 360 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 361 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */ 362 363 #ifdef CONFIG_SYSFS 364 struct wq_device *wq_dev; /* I: for sysfs interface */ 365 #endif 366 #ifdef CONFIG_LOCKDEP 367 char *lock_name; 368 struct lock_class_key key; 369 struct lockdep_map __lockdep_map; 370 struct lockdep_map *lockdep_map; 371 #endif 372 char name[WQ_NAME_LEN]; /* I: workqueue name */ 373 374 /* 375 * Destruction of workqueue_struct is RCU protected to allow walking 376 * the workqueues list without grabbing wq_pool_mutex. 377 * This is used to dump all workqueues from sysrq. 378 */ 379 struct rcu_head rcu; 380 381 /* hot fields used during command issue, aligned to cacheline */ 382 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 383 struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */ 384 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */ 385 }; 386 387 /* 388 * Each pod type describes how CPUs should be grouped for unbound workqueues. 389 * See the comment above workqueue_attrs->affn_scope. 390 */ 391 struct wq_pod_type { 392 int nr_pods; /* number of pods */ 393 cpumask_var_t *pod_cpus; /* pod -> cpus */ 394 int *pod_node; /* pod -> node */ 395 int *cpu_pod; /* cpu -> pod */ 396 }; 397 398 struct work_offq_data { 399 u32 pool_id; 400 u32 disable; 401 u32 flags; 402 }; 403 404 static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { 405 [WQ_AFFN_DFL] = "default", 406 [WQ_AFFN_CPU] = "cpu", 407 [WQ_AFFN_SMT] = "smt", 408 [WQ_AFFN_CACHE] = "cache", 409 [WQ_AFFN_NUMA] = "numa", 410 [WQ_AFFN_SYSTEM] = "system", 411 }; 412 413 /* 414 * Per-cpu work items which run for longer than the following threshold are 415 * automatically considered CPU intensive and excluded from concurrency 416 * management to prevent them from noticeably delaying other per-cpu work items. 417 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 418 * The actual value is initialized in wq_cpu_intensive_thresh_init(). 419 */ 420 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 421 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 422 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 423 static unsigned int wq_cpu_intensive_warning_thresh = 4; 424 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644); 425 #endif 426 427 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 428 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 429 module_param_named(power_efficient, wq_power_efficient, bool, 0444); 430 431 static bool wq_online; /* can kworkers be created yet? */ 432 static bool wq_topo_initialized __read_mostly = false; 433 434 static struct kmem_cache *pwq_cache; 435 436 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; 437 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; 438 439 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 440 static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf; 441 442 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 443 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 444 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 445 /* wait for manager to go away */ 446 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 447 448 static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 449 static bool workqueue_freezing; /* PL: have wqs started freezing? */ 450 451 /* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */ 452 static cpumask_var_t wq_online_cpumask; 453 454 /* PL&A: allowable cpus for unbound wqs and work items */ 455 static cpumask_var_t wq_unbound_cpumask; 456 457 /* PL: user requested unbound cpumask via sysfs */ 458 static cpumask_var_t wq_requested_unbound_cpumask; 459 460 /* PL: isolated cpumask to be excluded from unbound cpumask */ 461 static cpumask_var_t wq_isolated_cpumask; 462 463 /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 464 static struct cpumask wq_cmdline_cpumask __initdata; 465 466 /* CPU where unbound work was last round robin scheduled from this CPU */ 467 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 468 469 /* 470 * Local execution of unbound work items is no longer guaranteed. The 471 * following always forces round-robin CPU selection on unbound work items 472 * to uncover usages which depend on it. 473 */ 474 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 475 static bool wq_debug_force_rr_cpu = true; 476 #else 477 static bool wq_debug_force_rr_cpu = false; 478 #endif 479 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 480 481 /* to raise softirq for the BH worker pools on other CPUs */ 482 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works); 483 484 /* the BH worker pools */ 485 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools); 486 487 /* the per-cpu worker pools */ 488 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 489 490 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 491 492 /* PL: hash of all unbound pools keyed by pool->attrs */ 493 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 494 495 /* I: attributes used when instantiating standard unbound pools on demand */ 496 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 497 498 /* I: attributes used when instantiating ordered pools on demand */ 499 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 500 501 /* 502 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 503 * process context while holding a pool lock. Bounce to a dedicated kthread 504 * worker to avoid A-A deadlocks. 505 */ 506 static struct kthread_worker *pwq_release_worker __ro_after_init; 507 508 struct workqueue_struct *system_wq __ro_after_init; 509 EXPORT_SYMBOL(system_wq); 510 struct workqueue_struct *system_percpu_wq __ro_after_init; 511 EXPORT_SYMBOL(system_percpu_wq); 512 struct workqueue_struct *system_highpri_wq __ro_after_init; 513 EXPORT_SYMBOL_GPL(system_highpri_wq); 514 struct workqueue_struct *system_long_wq __ro_after_init; 515 EXPORT_SYMBOL_GPL(system_long_wq); 516 struct workqueue_struct *system_unbound_wq __ro_after_init; 517 EXPORT_SYMBOL_GPL(system_unbound_wq); 518 struct workqueue_struct *system_dfl_wq __ro_after_init; 519 EXPORT_SYMBOL_GPL(system_dfl_wq); 520 struct workqueue_struct *system_freezable_wq __ro_after_init; 521 EXPORT_SYMBOL_GPL(system_freezable_wq); 522 struct workqueue_struct *system_power_efficient_wq __ro_after_init; 523 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 524 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; 525 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 526 struct workqueue_struct *system_bh_wq; 527 EXPORT_SYMBOL_GPL(system_bh_wq); 528 struct workqueue_struct *system_bh_highpri_wq; 529 EXPORT_SYMBOL_GPL(system_bh_highpri_wq); 530 531 static int worker_thread(void *__worker); 532 static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 533 static void show_pwq(struct pool_workqueue *pwq); 534 static void show_one_worker_pool(struct worker_pool *pool); 535 536 #define CREATE_TRACE_POINTS 537 #include <trace/events/workqueue.h> 538 539 #define assert_rcu_or_pool_mutex() \ 540 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ 541 !lockdep_is_held(&wq_pool_mutex), \ 542 "RCU 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 timer_destroy_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 int old = atomic_read(&nna->nr); 1690 1691 do { 1692 if (old >= max) 1693 return false; 1694 } while (!atomic_try_cmpxchg_relaxed(&nna->nr, &old, old + 1)); 1695 1696 return true; 1697 } 1698 1699 /** 1700 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq 1701 * @pwq: pool_workqueue of interest 1702 * @fill: max_active may have increased, try to increase concurrency level 1703 * 1704 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is 1705 * successfully obtained. %false otherwise. 1706 */ 1707 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) 1708 { 1709 struct workqueue_struct *wq = pwq->wq; 1710 struct worker_pool *pool = pwq->pool; 1711 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); 1712 bool obtained = false; 1713 1714 lockdep_assert_held(&pool->lock); 1715 1716 if (!nna) { 1717 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */ 1718 obtained = pwq->nr_active < READ_ONCE(wq->max_active); 1719 goto out; 1720 } 1721 1722 if (unlikely(pwq->plugged)) 1723 return false; 1724 1725 /* 1726 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is 1727 * already waiting on $nna, pwq_dec_nr_active() will maintain the 1728 * concurrency level. Don't jump the line. 1729 * 1730 * We need to ignore the pending test after max_active has increased as 1731 * pwq_dec_nr_active() can only maintain the concurrency level but not 1732 * increase it. This is indicated by @fill. 1733 */ 1734 if (!list_empty(&pwq->pending_node) && likely(!fill)) 1735 goto out; 1736 1737 obtained = tryinc_node_nr_active(nna); 1738 if (obtained) 1739 goto out; 1740 1741 /* 1742 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs 1743 * and try again. The smp_mb() is paired with the implied memory barrier 1744 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either 1745 * we see the decremented $nna->nr or they see non-empty 1746 * $nna->pending_pwqs. 1747 */ 1748 raw_spin_lock(&nna->lock); 1749 1750 if (list_empty(&pwq->pending_node)) 1751 list_add_tail(&pwq->pending_node, &nna->pending_pwqs); 1752 else if (likely(!fill)) 1753 goto out_unlock; 1754 1755 smp_mb(); 1756 1757 obtained = tryinc_node_nr_active(nna); 1758 1759 /* 1760 * If @fill, @pwq might have already been pending. Being spuriously 1761 * pending in cold paths doesn't affect anything. Let's leave it be. 1762 */ 1763 if (obtained && likely(!fill)) 1764 list_del_init(&pwq->pending_node); 1765 1766 out_unlock: 1767 raw_spin_unlock(&nna->lock); 1768 out: 1769 if (obtained) 1770 pwq->nr_active++; 1771 return obtained; 1772 } 1773 1774 /** 1775 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq 1776 * @pwq: pool_workqueue of interest 1777 * @fill: max_active may have increased, try to increase concurrency level 1778 * 1779 * Activate the first inactive work item of @pwq if available and allowed by 1780 * max_active limit. 1781 * 1782 * Returns %true if an inactive work item has been activated. %false if no 1783 * inactive work item is found or max_active limit is reached. 1784 */ 1785 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) 1786 { 1787 struct work_struct *work = 1788 list_first_entry_or_null(&pwq->inactive_works, 1789 struct work_struct, entry); 1790 1791 if (work && pwq_tryinc_nr_active(pwq, fill)) { 1792 __pwq_activate_work(pwq, work); 1793 return true; 1794 } else { 1795 return false; 1796 } 1797 } 1798 1799 /** 1800 * unplug_oldest_pwq - unplug the oldest pool_workqueue 1801 * @wq: workqueue_struct where its oldest pwq is to be unplugged 1802 * 1803 * This function should only be called for ordered workqueues where only the 1804 * oldest pwq is unplugged, the others are plugged to suspend execution to 1805 * ensure proper work item ordering:: 1806 * 1807 * dfl_pwq --------------+ [P] - plugged 1808 * | 1809 * v 1810 * pwqs -> A -> B [P] -> C [P] (newest) 1811 * | | | 1812 * 1 3 5 1813 * | | | 1814 * 2 4 6 1815 * 1816 * When the oldest pwq is drained and removed, this function should be called 1817 * to unplug the next oldest one to start its work item execution. Note that 1818 * pwq's are linked into wq->pwqs with the oldest first, so the first one in 1819 * the list is the oldest. 1820 */ 1821 static void unplug_oldest_pwq(struct workqueue_struct *wq) 1822 { 1823 struct pool_workqueue *pwq; 1824 1825 lockdep_assert_held(&wq->mutex); 1826 1827 /* Caller should make sure that pwqs isn't empty before calling */ 1828 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue, 1829 pwqs_node); 1830 raw_spin_lock_irq(&pwq->pool->lock); 1831 if (pwq->plugged) { 1832 pwq->plugged = false; 1833 if (pwq_activate_first_inactive(pwq, true)) 1834 kick_pool(pwq->pool); 1835 } 1836 raw_spin_unlock_irq(&pwq->pool->lock); 1837 } 1838 1839 /** 1840 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active 1841 * @nna: wq_node_nr_active to activate a pending pwq for 1842 * @caller_pool: worker_pool the caller is locking 1843 * 1844 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. 1845 * @caller_pool may be unlocked and relocked to lock other worker_pools. 1846 */ 1847 static void node_activate_pending_pwq(struct wq_node_nr_active *nna, 1848 struct worker_pool *caller_pool) 1849 { 1850 struct worker_pool *locked_pool = caller_pool; 1851 struct pool_workqueue *pwq; 1852 struct work_struct *work; 1853 1854 lockdep_assert_held(&caller_pool->lock); 1855 1856 raw_spin_lock(&nna->lock); 1857 retry: 1858 pwq = list_first_entry_or_null(&nna->pending_pwqs, 1859 struct pool_workqueue, pending_node); 1860 if (!pwq) 1861 goto out_unlock; 1862 1863 /* 1864 * If @pwq is for a different pool than @locked_pool, we need to lock 1865 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock 1866 * / lock dance. For that, we also need to release @nna->lock as it's 1867 * nested inside pool locks. 1868 */ 1869 if (pwq->pool != locked_pool) { 1870 raw_spin_unlock(&locked_pool->lock); 1871 locked_pool = pwq->pool; 1872 if (!raw_spin_trylock(&locked_pool->lock)) { 1873 raw_spin_unlock(&nna->lock); 1874 raw_spin_lock(&locked_pool->lock); 1875 raw_spin_lock(&nna->lock); 1876 goto retry; 1877 } 1878 } 1879 1880 /* 1881 * $pwq may not have any inactive work items due to e.g. cancellations. 1882 * Drop it from pending_pwqs and see if there's another one. 1883 */ 1884 work = list_first_entry_or_null(&pwq->inactive_works, 1885 struct work_struct, entry); 1886 if (!work) { 1887 list_del_init(&pwq->pending_node); 1888 goto retry; 1889 } 1890 1891 /* 1892 * Acquire an nr_active count and activate the inactive work item. If 1893 * $pwq still has inactive work items, rotate it to the end of the 1894 * pending_pwqs so that we round-robin through them. This means that 1895 * inactive work items are not activated in queueing order which is fine 1896 * given that there has never been any ordering across different pwqs. 1897 */ 1898 if (likely(tryinc_node_nr_active(nna))) { 1899 pwq->nr_active++; 1900 __pwq_activate_work(pwq, work); 1901 1902 if (list_empty(&pwq->inactive_works)) 1903 list_del_init(&pwq->pending_node); 1904 else 1905 list_move_tail(&pwq->pending_node, &nna->pending_pwqs); 1906 1907 /* if activating a foreign pool, make sure it's running */ 1908 if (pwq->pool != caller_pool) 1909 kick_pool(pwq->pool); 1910 } 1911 1912 out_unlock: 1913 raw_spin_unlock(&nna->lock); 1914 if (locked_pool != caller_pool) { 1915 raw_spin_unlock(&locked_pool->lock); 1916 raw_spin_lock(&caller_pool->lock); 1917 } 1918 } 1919 1920 /** 1921 * pwq_dec_nr_active - Retire an active count 1922 * @pwq: pool_workqueue of interest 1923 * 1924 * Decrement @pwq's nr_active and try to activate the first inactive work item. 1925 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. 1926 */ 1927 static void pwq_dec_nr_active(struct pool_workqueue *pwq) 1928 { 1929 struct worker_pool *pool = pwq->pool; 1930 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); 1931 1932 lockdep_assert_held(&pool->lock); 1933 1934 /* 1935 * @pwq->nr_active should be decremented for both percpu and unbound 1936 * workqueues. 1937 */ 1938 pwq->nr_active--; 1939 1940 /* 1941 * For a percpu workqueue, it's simple. Just need to kick the first 1942 * inactive work item on @pwq itself. 1943 */ 1944 if (!nna) { 1945 pwq_activate_first_inactive(pwq, false); 1946 return; 1947 } 1948 1949 /* 1950 * If @pwq is for an unbound workqueue, it's more complicated because 1951 * multiple pwqs and pools may be sharing the nr_active count. When a 1952 * pwq needs to wait for an nr_active count, it puts itself on 1953 * $nna->pending_pwqs. The following atomic_dec_return()'s implied 1954 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to 1955 * guarantee that either we see non-empty pending_pwqs or they see 1956 * decremented $nna->nr. 1957 * 1958 * $nna->max may change as CPUs come online/offline and @pwq->wq's 1959 * max_active gets updated. However, it is guaranteed to be equal to or 1960 * larger than @pwq->wq->min_active which is above zero unless freezing. 1961 * This maintains the forward progress guarantee. 1962 */ 1963 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) 1964 return; 1965 1966 if (!list_empty(&nna->pending_pwqs)) 1967 node_activate_pending_pwq(nna, pool); 1968 } 1969 1970 /** 1971 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1972 * @pwq: pwq of interest 1973 * @work_data: work_data of work which left the queue 1974 * 1975 * A work either has completed or is removed from pending queue, 1976 * decrement nr_in_flight of its pwq and handle workqueue flushing. 1977 * 1978 * NOTE: 1979 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock 1980 * and thus should be called after all other state updates for the in-flight 1981 * work item is complete. 1982 * 1983 * CONTEXT: 1984 * raw_spin_lock_irq(pool->lock). 1985 */ 1986 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1987 { 1988 int color = get_work_color(work_data); 1989 1990 if (!(work_data & WORK_STRUCT_INACTIVE)) 1991 pwq_dec_nr_active(pwq); 1992 1993 pwq->nr_in_flight[color]--; 1994 1995 /* is flush in progress and are we at the flushing tip? */ 1996 if (likely(pwq->flush_color != color)) 1997 goto out_put; 1998 1999 /* are there still in-flight works? */ 2000 if (pwq->nr_in_flight[color]) 2001 goto out_put; 2002 2003 /* this pwq is done, clear flush_color */ 2004 pwq->flush_color = -1; 2005 2006 /* 2007 * If this was the last pwq, wake up the first flusher. It 2008 * will handle the rest. 2009 */ 2010 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 2011 complete(&pwq->wq->first_flusher->done); 2012 out_put: 2013 put_pwq(pwq); 2014 } 2015 2016 /** 2017 * try_to_grab_pending - steal work item from worklist and disable irq 2018 * @work: work item to steal 2019 * @cflags: %WORK_CANCEL_ flags 2020 * @irq_flags: place to store irq state 2021 * 2022 * Try to grab PENDING bit of @work. This function can handle @work in any 2023 * stable state - idle, on timer or on worklist. 2024 * 2025 * Return: 2026 * 2027 * ======== ================================================================ 2028 * 1 if @work was pending and we successfully stole PENDING 2029 * 0 if @work was idle and we claimed PENDING 2030 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 2031 * ======== ================================================================ 2032 * 2033 * Note: 2034 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 2035 * interrupted while holding PENDING and @work off queue, irq must be 2036 * disabled on entry. This, combined with delayed_work->timer being 2037 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 2038 * 2039 * On successful return, >= 0, irq is disabled and the caller is 2040 * responsible for releasing it using local_irq_restore(*@irq_flags). 2041 * 2042 * This function is safe to call from any context including IRQ handler. 2043 */ 2044 static int try_to_grab_pending(struct work_struct *work, u32 cflags, 2045 unsigned long *irq_flags) 2046 { 2047 struct worker_pool *pool; 2048 struct pool_workqueue *pwq; 2049 2050 local_irq_save(*irq_flags); 2051 2052 /* try to steal the timer if it exists */ 2053 if (cflags & WORK_CANCEL_DELAYED) { 2054 struct delayed_work *dwork = to_delayed_work(work); 2055 2056 /* 2057 * dwork->timer is irqsafe. If timer_delete() fails, it's 2058 * guaranteed that the timer is not queued anywhere and not 2059 * running on the local CPU. 2060 */ 2061 if (likely(timer_delete(&dwork->timer))) 2062 return 1; 2063 } 2064 2065 /* try to claim PENDING the normal way */ 2066 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 2067 return 0; 2068 2069 rcu_read_lock(); 2070 /* 2071 * The queueing is in progress, or it is already queued. Try to 2072 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 2073 */ 2074 pool = get_work_pool(work); 2075 if (!pool) 2076 goto fail; 2077 2078 raw_spin_lock(&pool->lock); 2079 /* 2080 * work->data is guaranteed to point to pwq only while the work 2081 * item is queued on pwq->wq, and both updating work->data to point 2082 * to pwq on queueing and to pool on dequeueing are done under 2083 * pwq->pool->lock. This in turn guarantees that, if work->data 2084 * points to pwq which is associated with a locked pool, the work 2085 * item is currently queued on that pool. 2086 */ 2087 pwq = get_work_pwq(work); 2088 if (pwq && pwq->pool == pool) { 2089 unsigned long work_data = *work_data_bits(work); 2090 2091 debug_work_deactivate(work); 2092 2093 /* 2094 * A cancelable inactive work item must be in the 2095 * pwq->inactive_works since a queued barrier can't be 2096 * canceled (see the comments in insert_wq_barrier()). 2097 * 2098 * An inactive work item cannot be deleted directly because 2099 * it might have linked barrier work items which, if left 2100 * on the inactive_works list, will confuse pwq->nr_active 2101 * management later on and cause stall. Move the linked 2102 * barrier work items to the worklist when deleting the grabbed 2103 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that 2104 * it doesn't participate in nr_active management in later 2105 * pwq_dec_nr_in_flight(). 2106 */ 2107 if (work_data & WORK_STRUCT_INACTIVE) 2108 move_linked_works(work, &pwq->pool->worklist, NULL); 2109 2110 list_del_init(&work->entry); 2111 2112 /* 2113 * work->data points to pwq iff queued. Let's point to pool. As 2114 * this destroys work->data needed by the next step, stash it. 2115 */ 2116 set_work_pool_and_keep_pending(work, pool->id, 2117 pool_offq_flags(pool)); 2118 2119 /* must be the last step, see the function comment */ 2120 pwq_dec_nr_in_flight(pwq, work_data); 2121 2122 raw_spin_unlock(&pool->lock); 2123 rcu_read_unlock(); 2124 return 1; 2125 } 2126 raw_spin_unlock(&pool->lock); 2127 fail: 2128 rcu_read_unlock(); 2129 local_irq_restore(*irq_flags); 2130 return -EAGAIN; 2131 } 2132 2133 /** 2134 * work_grab_pending - steal work item from worklist and disable irq 2135 * @work: work item to steal 2136 * @cflags: %WORK_CANCEL_ flags 2137 * @irq_flags: place to store IRQ state 2138 * 2139 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer 2140 * or on worklist. 2141 * 2142 * Can be called from any context. IRQ is disabled on return with IRQ state 2143 * stored in *@irq_flags. The caller is responsible for re-enabling it using 2144 * local_irq_restore(). 2145 * 2146 * Returns %true if @work was pending. %false if idle. 2147 */ 2148 static bool work_grab_pending(struct work_struct *work, u32 cflags, 2149 unsigned long *irq_flags) 2150 { 2151 int ret; 2152 2153 while (true) { 2154 ret = try_to_grab_pending(work, cflags, irq_flags); 2155 if (ret >= 0) 2156 return ret; 2157 cpu_relax(); 2158 } 2159 } 2160 2161 /** 2162 * insert_work - insert a work into a pool 2163 * @pwq: pwq @work belongs to 2164 * @work: work to insert 2165 * @head: insertion point 2166 * @extra_flags: extra WORK_STRUCT_* flags to set 2167 * 2168 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 2169 * work_struct flags. 2170 * 2171 * CONTEXT: 2172 * raw_spin_lock_irq(pool->lock). 2173 */ 2174 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 2175 struct list_head *head, unsigned int extra_flags) 2176 { 2177 debug_work_activate(work); 2178 2179 /* record the work call stack in order to print it in KASAN reports */ 2180 kasan_record_aux_stack(work); 2181 2182 /* we own @work, set data and link */ 2183 set_work_pwq(work, pwq, extra_flags); 2184 list_add_tail(&work->entry, head); 2185 get_pwq(pwq); 2186 } 2187 2188 /* 2189 * Test whether @work is being queued from another work executing on the 2190 * same workqueue. 2191 */ 2192 static bool is_chained_work(struct workqueue_struct *wq) 2193 { 2194 struct worker *worker; 2195 2196 worker = current_wq_worker(); 2197 /* 2198 * Return %true iff I'm a worker executing a work item on @wq. If 2199 * I'm @worker, it's safe to dereference it without locking. 2200 */ 2201 return worker && worker->current_pwq->wq == wq; 2202 } 2203 2204 /* 2205 * When queueing an unbound work item to a wq, prefer local CPU if allowed 2206 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 2207 * avoid perturbing sensitive tasks. 2208 */ 2209 static int wq_select_unbound_cpu(int cpu) 2210 { 2211 int new_cpu; 2212 2213 if (likely(!wq_debug_force_rr_cpu)) { 2214 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 2215 return cpu; 2216 } else { 2217 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 2218 } 2219 2220 new_cpu = __this_cpu_read(wq_rr_cpu_last); 2221 new_cpu = cpumask_next_and_wrap(new_cpu, wq_unbound_cpumask, cpu_online_mask); 2222 if (unlikely(new_cpu >= nr_cpu_ids)) 2223 return cpu; 2224 __this_cpu_write(wq_rr_cpu_last, new_cpu); 2225 2226 return new_cpu; 2227 } 2228 2229 static void __queue_work(int cpu, struct workqueue_struct *wq, 2230 struct work_struct *work) 2231 { 2232 struct pool_workqueue *pwq; 2233 struct worker_pool *last_pool, *pool; 2234 unsigned int work_flags; 2235 unsigned int req_cpu = cpu; 2236 2237 /* 2238 * While a work item is PENDING && off queue, a task trying to 2239 * steal the PENDING will busy-loop waiting for it to either get 2240 * queued or lose PENDING. Grabbing PENDING and queueing should 2241 * happen with IRQ disabled. 2242 */ 2243 lockdep_assert_irqs_disabled(); 2244 2245 /* 2246 * For a draining wq, only works from the same workqueue are 2247 * allowed. The __WQ_DESTROYING helps to spot the issue that 2248 * queues a new work item to a wq after destroy_workqueue(wq). 2249 */ 2250 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 2251 WARN_ONCE(!is_chained_work(wq), "workqueue: cannot queue %ps on wq %s\n", 2252 work->func, wq->name))) { 2253 return; 2254 } 2255 rcu_read_lock(); 2256 retry: 2257 /* pwq which will be used unless @work is executing elsewhere */ 2258 if (req_cpu == WORK_CPU_UNBOUND) { 2259 if (wq->flags & WQ_UNBOUND) 2260 cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 2261 else 2262 cpu = raw_smp_processor_id(); 2263 } 2264 2265 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 2266 pool = pwq->pool; 2267 2268 /* 2269 * If @work was previously on a different pool, it might still be 2270 * running there, in which case the work needs to be queued on that 2271 * pool to guarantee non-reentrancy. 2272 * 2273 * For ordered workqueue, work items must be queued on the newest pwq 2274 * for accurate order management. Guaranteed order also guarantees 2275 * non-reentrancy. See the comments above unplug_oldest_pwq(). 2276 */ 2277 last_pool = get_work_pool(work); 2278 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) { 2279 struct worker *worker; 2280 2281 raw_spin_lock(&last_pool->lock); 2282 2283 worker = find_worker_executing_work(last_pool, work); 2284 2285 if (worker && worker->current_pwq->wq == wq) { 2286 pwq = worker->current_pwq; 2287 pool = pwq->pool; 2288 WARN_ON_ONCE(pool != last_pool); 2289 } else { 2290 /* meh... not running there, queue here */ 2291 raw_spin_unlock(&last_pool->lock); 2292 raw_spin_lock(&pool->lock); 2293 } 2294 } else { 2295 raw_spin_lock(&pool->lock); 2296 } 2297 2298 /* 2299 * pwq is determined and locked. For unbound pools, we could have raced 2300 * with pwq release and it could already be dead. If its refcnt is zero, 2301 * repeat pwq selection. Note that unbound pwqs never die without 2302 * another pwq replacing it in cpu_pwq or while work items are executing 2303 * on it, so the retrying is guaranteed to make forward-progress. 2304 */ 2305 if (unlikely(!pwq->refcnt)) { 2306 if (wq->flags & WQ_UNBOUND) { 2307 raw_spin_unlock(&pool->lock); 2308 cpu_relax(); 2309 goto retry; 2310 } 2311 /* oops */ 2312 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 2313 wq->name, cpu); 2314 } 2315 2316 /* pwq determined, queue */ 2317 trace_workqueue_queue_work(req_cpu, pwq, work); 2318 2319 if (WARN_ON(!list_empty(&work->entry))) 2320 goto out; 2321 2322 pwq->nr_in_flight[pwq->work_color]++; 2323 work_flags = work_color_to_flags(pwq->work_color); 2324 2325 /* 2326 * Limit the number of concurrently active work items to max_active. 2327 * @work must also queue behind existing inactive work items to maintain 2328 * ordering when max_active changes. See wq_adjust_max_active(). 2329 */ 2330 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { 2331 if (list_empty(&pool->worklist)) 2332 pool->watchdog_ts = jiffies; 2333 2334 trace_workqueue_activate_work(work); 2335 insert_work(pwq, work, &pool->worklist, work_flags); 2336 kick_pool(pool); 2337 } else { 2338 work_flags |= WORK_STRUCT_INACTIVE; 2339 insert_work(pwq, work, &pwq->inactive_works, work_flags); 2340 } 2341 2342 out: 2343 raw_spin_unlock(&pool->lock); 2344 rcu_read_unlock(); 2345 } 2346 2347 static bool clear_pending_if_disabled(struct work_struct *work) 2348 { 2349 unsigned long data = *work_data_bits(work); 2350 struct work_offq_data offqd; 2351 2352 if (likely((data & WORK_STRUCT_PWQ) || 2353 !(data & WORK_OFFQ_DISABLE_MASK))) 2354 return false; 2355 2356 work_offqd_unpack(&offqd, data); 2357 set_work_pool_and_clear_pending(work, offqd.pool_id, 2358 work_offqd_pack_flags(&offqd)); 2359 return true; 2360 } 2361 2362 /** 2363 * queue_work_on - queue work on specific cpu 2364 * @cpu: CPU number to execute work on 2365 * @wq: workqueue to use 2366 * @work: work to queue 2367 * 2368 * We queue the work to a specific CPU, the caller must ensure it 2369 * can't go away. Callers that fail to ensure that the specified 2370 * CPU cannot go away will execute on a randomly chosen CPU. 2371 * But note well that callers specifying a CPU that never has been 2372 * online will get a splat. 2373 * 2374 * Return: %false if @work was already on a queue, %true otherwise. 2375 */ 2376 bool queue_work_on(int cpu, struct workqueue_struct *wq, 2377 struct work_struct *work) 2378 { 2379 bool ret = false; 2380 unsigned long irq_flags; 2381 2382 local_irq_save(irq_flags); 2383 2384 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2385 !clear_pending_if_disabled(work)) { 2386 __queue_work(cpu, wq, work); 2387 ret = true; 2388 } 2389 2390 local_irq_restore(irq_flags); 2391 return ret; 2392 } 2393 EXPORT_SYMBOL(queue_work_on); 2394 2395 /** 2396 * select_numa_node_cpu - Select a CPU based on NUMA node 2397 * @node: NUMA node ID that we want to select a CPU from 2398 * 2399 * This function will attempt to find a "random" cpu available on a given 2400 * node. If there are no CPUs available on the given node it will return 2401 * WORK_CPU_UNBOUND indicating that we should just schedule to any 2402 * available CPU if we need to schedule this work. 2403 */ 2404 static int select_numa_node_cpu(int node) 2405 { 2406 int cpu; 2407 2408 /* Delay binding to CPU if node is not valid or online */ 2409 if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 2410 return WORK_CPU_UNBOUND; 2411 2412 /* Use local node/cpu if we are already there */ 2413 cpu = raw_smp_processor_id(); 2414 if (node == cpu_to_node(cpu)) 2415 return cpu; 2416 2417 /* Use "random" otherwise know as "first" online CPU of node */ 2418 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 2419 2420 /* If CPU is valid return that, otherwise just defer */ 2421 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 2422 } 2423 2424 /** 2425 * queue_work_node - queue work on a "random" cpu for a given NUMA node 2426 * @node: NUMA node that we are targeting the work for 2427 * @wq: workqueue to use 2428 * @work: work to queue 2429 * 2430 * We queue the work to a "random" CPU within a given NUMA node. The basic 2431 * idea here is to provide a way to somehow associate work with a given 2432 * NUMA node. 2433 * 2434 * This function will only make a best effort attempt at getting this onto 2435 * the right NUMA node. If no node is requested or the requested node is 2436 * offline then we just fall back to standard queue_work behavior. 2437 * 2438 * Currently the "random" CPU ends up being the first available CPU in the 2439 * intersection of cpu_online_mask and the cpumask of the node, unless we 2440 * are running on the node. In that case we just use the current CPU. 2441 * 2442 * Return: %false if @work was already on a queue, %true otherwise. 2443 */ 2444 bool queue_work_node(int node, struct workqueue_struct *wq, 2445 struct work_struct *work) 2446 { 2447 unsigned long irq_flags; 2448 bool ret = false; 2449 2450 /* 2451 * This current implementation is specific to unbound workqueues. 2452 * Specifically we only return the first available CPU for a given 2453 * node instead of cycling through individual CPUs within the node. 2454 * 2455 * If this is used with a per-cpu workqueue then the logic in 2456 * workqueue_select_cpu_near would need to be updated to allow for 2457 * some round robin type logic. 2458 */ 2459 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 2460 2461 local_irq_save(irq_flags); 2462 2463 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2464 !clear_pending_if_disabled(work)) { 2465 int cpu = select_numa_node_cpu(node); 2466 2467 __queue_work(cpu, wq, work); 2468 ret = true; 2469 } 2470 2471 local_irq_restore(irq_flags); 2472 return ret; 2473 } 2474 EXPORT_SYMBOL_GPL(queue_work_node); 2475 2476 void delayed_work_timer_fn(struct timer_list *t) 2477 { 2478 struct delayed_work *dwork = timer_container_of(dwork, t, timer); 2479 2480 /* should have been called from irqsafe timer with irq already off */ 2481 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 2482 } 2483 EXPORT_SYMBOL(delayed_work_timer_fn); 2484 2485 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 2486 struct delayed_work *dwork, unsigned long delay) 2487 { 2488 struct timer_list *timer = &dwork->timer; 2489 struct work_struct *work = &dwork->work; 2490 2491 WARN_ON_ONCE(!wq); 2492 WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 2493 WARN_ON_ONCE(timer_pending(timer)); 2494 WARN_ON_ONCE(!list_empty(&work->entry)); 2495 2496 /* 2497 * If @delay is 0, queue @dwork->work immediately. This is for 2498 * both optimization and correctness. The earliest @timer can 2499 * expire is on the closest next tick and delayed_work users depend 2500 * on that there's no such delay when @delay is 0. 2501 */ 2502 if (!delay) { 2503 __queue_work(cpu, wq, &dwork->work); 2504 return; 2505 } 2506 2507 WARN_ON_ONCE(cpu != WORK_CPU_UNBOUND && !cpu_online(cpu)); 2508 dwork->wq = wq; 2509 dwork->cpu = cpu; 2510 timer->expires = jiffies + delay; 2511 2512 if (housekeeping_enabled(HK_TYPE_TIMER)) { 2513 /* If the current cpu is a housekeeping cpu, use it. */ 2514 cpu = smp_processor_id(); 2515 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) 2516 cpu = housekeeping_any_cpu(HK_TYPE_TIMER); 2517 add_timer_on(timer, cpu); 2518 } else { 2519 if (likely(cpu == WORK_CPU_UNBOUND)) 2520 add_timer_global(timer); 2521 else 2522 add_timer_on(timer, cpu); 2523 } 2524 } 2525 2526 /** 2527 * queue_delayed_work_on - queue work on specific CPU after delay 2528 * @cpu: CPU number to execute work on 2529 * @wq: workqueue to use 2530 * @dwork: work to queue 2531 * @delay: number of jiffies to wait before queueing 2532 * 2533 * We queue the delayed_work to a specific CPU, for non-zero delays the 2534 * caller must ensure it is online and can't go away. Callers that fail 2535 * to ensure this, may get @dwork->timer queued to an offlined CPU and 2536 * this will prevent queueing of @dwork->work unless the offlined CPU 2537 * becomes online again. 2538 * 2539 * Return: %false if @work was already on a queue, %true otherwise. If 2540 * @delay is zero and @dwork is idle, it will be scheduled for immediate 2541 * execution. 2542 */ 2543 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 2544 struct delayed_work *dwork, unsigned long delay) 2545 { 2546 struct work_struct *work = &dwork->work; 2547 bool ret = false; 2548 unsigned long irq_flags; 2549 2550 /* read the comment in __queue_work() */ 2551 local_irq_save(irq_flags); 2552 2553 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2554 !clear_pending_if_disabled(work)) { 2555 __queue_delayed_work(cpu, wq, dwork, delay); 2556 ret = true; 2557 } 2558 2559 local_irq_restore(irq_flags); 2560 return ret; 2561 } 2562 EXPORT_SYMBOL(queue_delayed_work_on); 2563 2564 /** 2565 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 2566 * @cpu: CPU number to execute work on 2567 * @wq: workqueue to use 2568 * @dwork: work to queue 2569 * @delay: number of jiffies to wait before queueing 2570 * 2571 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 2572 * modify @dwork's timer so that it expires after @delay. If @delay is 2573 * zero, @work is guaranteed to be scheduled immediately regardless of its 2574 * current state. 2575 * 2576 * Return: %false if @dwork was idle and queued, %true if @dwork was 2577 * pending and its timer was modified. 2578 * 2579 * This function is safe to call from any context including IRQ handler. 2580 * See try_to_grab_pending() for details. 2581 */ 2582 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 2583 struct delayed_work *dwork, unsigned long delay) 2584 { 2585 unsigned long irq_flags; 2586 bool ret; 2587 2588 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags); 2589 2590 if (!clear_pending_if_disabled(&dwork->work)) 2591 __queue_delayed_work(cpu, wq, dwork, delay); 2592 2593 local_irq_restore(irq_flags); 2594 return ret; 2595 } 2596 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2597 2598 static void rcu_work_rcufn(struct rcu_head *rcu) 2599 { 2600 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 2601 2602 /* read the comment in __queue_work() */ 2603 local_irq_disable(); 2604 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 2605 local_irq_enable(); 2606 } 2607 2608 /** 2609 * queue_rcu_work - queue work after a RCU grace period 2610 * @wq: workqueue to use 2611 * @rwork: work to queue 2612 * 2613 * Return: %false if @rwork was already pending, %true otherwise. Note 2614 * that a full RCU grace period is guaranteed only after a %true return. 2615 * While @rwork is guaranteed to be executed after a %false return, the 2616 * execution may happen before a full RCU grace period has passed. 2617 */ 2618 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 2619 { 2620 struct work_struct *work = &rwork->work; 2621 2622 /* 2623 * rcu_work can't be canceled or disabled. Warn if the user reached 2624 * inside @rwork and disabled the inner work. 2625 */ 2626 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2627 !WARN_ON_ONCE(clear_pending_if_disabled(work))) { 2628 rwork->wq = wq; 2629 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 2630 return true; 2631 } 2632 2633 return false; 2634 } 2635 EXPORT_SYMBOL(queue_rcu_work); 2636 2637 static struct worker *alloc_worker(int node) 2638 { 2639 struct worker *worker; 2640 2641 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2642 if (worker) { 2643 INIT_LIST_HEAD(&worker->entry); 2644 INIT_LIST_HEAD(&worker->scheduled); 2645 INIT_LIST_HEAD(&worker->node); 2646 /* on creation a worker is in !idle && prep state */ 2647 worker->flags = WORKER_PREP; 2648 } 2649 return worker; 2650 } 2651 2652 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2653 { 2654 if (pool->cpu < 0 && pool->attrs->affn_strict) 2655 return pool->attrs->__pod_cpumask; 2656 else 2657 return pool->attrs->cpumask; 2658 } 2659 2660 /** 2661 * worker_attach_to_pool() - attach a worker to a pool 2662 * @worker: worker to be attached 2663 * @pool: the target pool 2664 * 2665 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 2666 * cpu-binding of @worker are kept coordinated with the pool across 2667 * cpu-[un]hotplugs. 2668 */ 2669 static void worker_attach_to_pool(struct worker *worker, 2670 struct worker_pool *pool) 2671 { 2672 mutex_lock(&wq_pool_attach_mutex); 2673 2674 /* 2675 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable 2676 * across this function. See the comments above the flag definition for 2677 * details. BH workers are, while per-CPU, always DISASSOCIATED. 2678 */ 2679 if (pool->flags & POOL_DISASSOCIATED) { 2680 worker->flags |= WORKER_UNBOUND; 2681 } else { 2682 WARN_ON_ONCE(pool->flags & POOL_BH); 2683 kthread_set_per_cpu(worker->task, pool->cpu); 2684 } 2685 2686 if (worker->rescue_wq) 2687 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2688 2689 list_add_tail(&worker->node, &pool->workers); 2690 worker->pool = pool; 2691 2692 mutex_unlock(&wq_pool_attach_mutex); 2693 } 2694 2695 static void unbind_worker(struct worker *worker) 2696 { 2697 lockdep_assert_held(&wq_pool_attach_mutex); 2698 2699 kthread_set_per_cpu(worker->task, -1); 2700 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2701 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2702 else 2703 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2704 } 2705 2706 2707 static void detach_worker(struct worker *worker) 2708 { 2709 lockdep_assert_held(&wq_pool_attach_mutex); 2710 2711 unbind_worker(worker); 2712 list_del(&worker->node); 2713 } 2714 2715 /** 2716 * worker_detach_from_pool() - detach a worker from its pool 2717 * @worker: worker which is attached to its pool 2718 * 2719 * Undo the attaching which had been done in worker_attach_to_pool(). The 2720 * caller worker shouldn't access to the pool after detached except it has 2721 * other reference to the pool. 2722 */ 2723 static void worker_detach_from_pool(struct worker *worker) 2724 { 2725 struct worker_pool *pool = worker->pool; 2726 2727 /* there is one permanent BH worker per CPU which should never detach */ 2728 WARN_ON_ONCE(pool->flags & POOL_BH); 2729 2730 mutex_lock(&wq_pool_attach_mutex); 2731 detach_worker(worker); 2732 worker->pool = NULL; 2733 mutex_unlock(&wq_pool_attach_mutex); 2734 2735 /* clear leftover flags without pool->lock after it is detached */ 2736 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2737 } 2738 2739 static int format_worker_id(char *buf, size_t size, struct worker *worker, 2740 struct worker_pool *pool) 2741 { 2742 if (worker->rescue_wq) 2743 return scnprintf(buf, size, "kworker/R-%s", 2744 worker->rescue_wq->name); 2745 2746 if (pool) { 2747 if (pool->cpu >= 0) 2748 return scnprintf(buf, size, "kworker/%d:%d%s", 2749 pool->cpu, worker->id, 2750 pool->attrs->nice < 0 ? "H" : ""); 2751 else 2752 return scnprintf(buf, size, "kworker/u%d:%d", 2753 pool->id, worker->id); 2754 } else { 2755 return scnprintf(buf, size, "kworker/dying"); 2756 } 2757 } 2758 2759 /** 2760 * create_worker - create a new workqueue worker 2761 * @pool: pool the new worker will belong to 2762 * 2763 * Create and start a new worker which is attached to @pool. 2764 * 2765 * CONTEXT: 2766 * Might sleep. Does GFP_KERNEL allocations. 2767 * 2768 * Return: 2769 * Pointer to the newly created worker. 2770 */ 2771 static struct worker *create_worker(struct worker_pool *pool) 2772 { 2773 struct worker *worker; 2774 int id; 2775 2776 /* ID is needed to determine kthread name */ 2777 id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 2778 if (id < 0) { 2779 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 2780 ERR_PTR(id)); 2781 return NULL; 2782 } 2783 2784 worker = alloc_worker(pool->node); 2785 if (!worker) { 2786 pr_err_once("workqueue: Failed to allocate a worker\n"); 2787 goto fail; 2788 } 2789 2790 worker->id = id; 2791 2792 if (!(pool->flags & POOL_BH)) { 2793 char id_buf[WORKER_ID_LEN]; 2794 2795 format_worker_id(id_buf, sizeof(id_buf), worker, pool); 2796 worker->task = kthread_create_on_node(worker_thread, worker, 2797 pool->node, "%s", id_buf); 2798 if (IS_ERR(worker->task)) { 2799 if (PTR_ERR(worker->task) == -EINTR) { 2800 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n", 2801 id_buf); 2802 } else { 2803 pr_err_once("workqueue: Failed to create a worker thread: %pe", 2804 worker->task); 2805 } 2806 goto fail; 2807 } 2808 2809 set_user_nice(worker->task, pool->attrs->nice); 2810 kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 2811 } 2812 2813 /* successful, attach the worker to the pool */ 2814 worker_attach_to_pool(worker, pool); 2815 2816 /* start the newly created worker */ 2817 raw_spin_lock_irq(&pool->lock); 2818 2819 worker->pool->nr_workers++; 2820 worker_enter_idle(worker); 2821 2822 /* 2823 * @worker is waiting on a completion in kthread() and will trigger hung 2824 * check if not woken up soon. As kick_pool() is noop if @pool is empty, 2825 * wake it up explicitly. 2826 */ 2827 if (worker->task) 2828 wake_up_process(worker->task); 2829 2830 raw_spin_unlock_irq(&pool->lock); 2831 2832 return worker; 2833 2834 fail: 2835 ida_free(&pool->worker_ida, id); 2836 kfree(worker); 2837 return NULL; 2838 } 2839 2840 static void detach_dying_workers(struct list_head *cull_list) 2841 { 2842 struct worker *worker; 2843 2844 list_for_each_entry(worker, cull_list, entry) 2845 detach_worker(worker); 2846 } 2847 2848 static void reap_dying_workers(struct list_head *cull_list) 2849 { 2850 struct worker *worker, *tmp; 2851 2852 list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2853 list_del_init(&worker->entry); 2854 kthread_stop_put(worker->task); 2855 kfree(worker); 2856 } 2857 } 2858 2859 /** 2860 * set_worker_dying - Tag a worker for destruction 2861 * @worker: worker to be destroyed 2862 * @list: transfer worker away from its pool->idle_list and into list 2863 * 2864 * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2865 * should be idle. 2866 * 2867 * CONTEXT: 2868 * raw_spin_lock_irq(pool->lock). 2869 */ 2870 static void set_worker_dying(struct worker *worker, struct list_head *list) 2871 { 2872 struct worker_pool *pool = worker->pool; 2873 2874 lockdep_assert_held(&pool->lock); 2875 lockdep_assert_held(&wq_pool_attach_mutex); 2876 2877 /* sanity check frenzy */ 2878 if (WARN_ON(worker->current_work) || 2879 WARN_ON(!list_empty(&worker->scheduled)) || 2880 WARN_ON(!(worker->flags & WORKER_IDLE))) 2881 return; 2882 2883 pool->nr_workers--; 2884 pool->nr_idle--; 2885 2886 worker->flags |= WORKER_DIE; 2887 2888 list_move(&worker->entry, list); 2889 2890 /* get an extra task struct reference for later kthread_stop_put() */ 2891 get_task_struct(worker->task); 2892 } 2893 2894 /** 2895 * idle_worker_timeout - check if some idle workers can now be deleted. 2896 * @t: The pool's idle_timer that just expired 2897 * 2898 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 2899 * worker_leave_idle(), as a worker flicking between idle and active while its 2900 * pool is at the too_many_workers() tipping point would cause too much timer 2901 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 2902 * it expire and re-evaluate things from there. 2903 */ 2904 static void idle_worker_timeout(struct timer_list *t) 2905 { 2906 struct worker_pool *pool = timer_container_of(pool, t, idle_timer); 2907 bool do_cull = false; 2908 2909 if (work_pending(&pool->idle_cull_work)) 2910 return; 2911 2912 raw_spin_lock_irq(&pool->lock); 2913 2914 if (too_many_workers(pool)) { 2915 struct worker *worker; 2916 unsigned long expires; 2917 2918 /* idle_list is kept in LIFO order, check the last one */ 2919 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2920 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2921 do_cull = !time_before(jiffies, expires); 2922 2923 if (!do_cull) 2924 mod_timer(&pool->idle_timer, expires); 2925 } 2926 raw_spin_unlock_irq(&pool->lock); 2927 2928 if (do_cull) 2929 queue_work(system_dfl_wq, &pool->idle_cull_work); 2930 } 2931 2932 /** 2933 * idle_cull_fn - cull workers that have been idle for too long. 2934 * @work: the pool's work for handling these idle workers 2935 * 2936 * This goes through a pool's idle workers and gets rid of those that have been 2937 * idle for at least IDLE_WORKER_TIMEOUT seconds. 2938 * 2939 * We don't want to disturb isolated CPUs because of a pcpu kworker being 2940 * culled, so this also resets worker affinity. This requires a sleepable 2941 * context, hence the split between timer callback and work item. 2942 */ 2943 static void idle_cull_fn(struct work_struct *work) 2944 { 2945 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2946 LIST_HEAD(cull_list); 2947 2948 /* 2949 * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2950 * cannot proceed beyong set_pf_worker() in its self-destruct path. 2951 * This is required as a previously-preempted worker could run after 2952 * set_worker_dying() has happened but before detach_dying_workers() did. 2953 */ 2954 mutex_lock(&wq_pool_attach_mutex); 2955 raw_spin_lock_irq(&pool->lock); 2956 2957 while (too_many_workers(pool)) { 2958 struct worker *worker; 2959 unsigned long expires; 2960 2961 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2962 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2963 2964 if (time_before(jiffies, expires)) { 2965 mod_timer(&pool->idle_timer, expires); 2966 break; 2967 } 2968 2969 set_worker_dying(worker, &cull_list); 2970 } 2971 2972 raw_spin_unlock_irq(&pool->lock); 2973 detach_dying_workers(&cull_list); 2974 mutex_unlock(&wq_pool_attach_mutex); 2975 2976 reap_dying_workers(&cull_list); 2977 } 2978 2979 static void send_mayday(struct work_struct *work) 2980 { 2981 struct pool_workqueue *pwq = get_work_pwq(work); 2982 struct workqueue_struct *wq = pwq->wq; 2983 2984 lockdep_assert_held(&wq_mayday_lock); 2985 2986 if (!wq->rescuer) 2987 return; 2988 2989 /* mayday mayday mayday */ 2990 if (list_empty(&pwq->mayday_node)) { 2991 /* 2992 * If @pwq is for an unbound wq, its base ref may be put at 2993 * any time due to an attribute change. Pin @pwq until the 2994 * rescuer is done with it. 2995 */ 2996 get_pwq(pwq); 2997 list_add_tail(&pwq->mayday_node, &wq->maydays); 2998 wake_up_process(wq->rescuer->task); 2999 pwq->stats[PWQ_STAT_MAYDAY]++; 3000 } 3001 } 3002 3003 static void pool_mayday_timeout(struct timer_list *t) 3004 { 3005 struct worker_pool *pool = timer_container_of(pool, t, mayday_timer); 3006 struct work_struct *work; 3007 3008 raw_spin_lock_irq(&pool->lock); 3009 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 3010 3011 if (need_to_create_worker(pool)) { 3012 /* 3013 * We've been trying to create a new worker but 3014 * haven't been successful. We might be hitting an 3015 * allocation deadlock. Send distress signals to 3016 * rescuers. 3017 */ 3018 list_for_each_entry(work, &pool->worklist, entry) 3019 send_mayday(work); 3020 } 3021 3022 raw_spin_unlock(&wq_mayday_lock); 3023 raw_spin_unlock_irq(&pool->lock); 3024 3025 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 3026 } 3027 3028 /** 3029 * maybe_create_worker - create a new worker if necessary 3030 * @pool: pool to create a new worker for 3031 * 3032 * Create a new worker for @pool if necessary. @pool is guaranteed to 3033 * have at least one idle worker on return from this function. If 3034 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 3035 * sent to all rescuers with works scheduled on @pool to resolve 3036 * possible allocation deadlock. 3037 * 3038 * On return, need_to_create_worker() is guaranteed to be %false and 3039 * may_start_working() %true. 3040 * 3041 * LOCKING: 3042 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3043 * multiple times. Does GFP_KERNEL allocations. Called only from 3044 * manager. 3045 */ 3046 static void maybe_create_worker(struct worker_pool *pool) 3047 __releases(&pool->lock) 3048 __acquires(&pool->lock) 3049 { 3050 restart: 3051 raw_spin_unlock_irq(&pool->lock); 3052 3053 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 3054 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 3055 3056 while (true) { 3057 if (create_worker(pool) || !need_to_create_worker(pool)) 3058 break; 3059 3060 schedule_timeout_interruptible(CREATE_COOLDOWN); 3061 3062 if (!need_to_create_worker(pool)) 3063 break; 3064 } 3065 3066 timer_delete_sync(&pool->mayday_timer); 3067 raw_spin_lock_irq(&pool->lock); 3068 /* 3069 * This is necessary even after a new worker was just successfully 3070 * created as @pool->lock was dropped and the new worker might have 3071 * already become busy. 3072 */ 3073 if (need_to_create_worker(pool)) 3074 goto restart; 3075 } 3076 3077 #ifdef CONFIG_PREEMPT_RT 3078 static void worker_lock_callback(struct worker_pool *pool) 3079 { 3080 spin_lock(&pool->cb_lock); 3081 } 3082 3083 static void worker_unlock_callback(struct worker_pool *pool) 3084 { 3085 spin_unlock(&pool->cb_lock); 3086 } 3087 3088 static void workqueue_callback_cancel_wait_running(struct worker_pool *pool) 3089 { 3090 spin_lock(&pool->cb_lock); 3091 spin_unlock(&pool->cb_lock); 3092 } 3093 3094 #else 3095 3096 static void worker_lock_callback(struct worker_pool *pool) { } 3097 static void worker_unlock_callback(struct worker_pool *pool) { } 3098 static void workqueue_callback_cancel_wait_running(struct worker_pool *pool) { } 3099 3100 #endif 3101 3102 /** 3103 * manage_workers - manage worker pool 3104 * @worker: self 3105 * 3106 * Assume the manager role and manage the worker pool @worker belongs 3107 * to. At any given time, there can be only zero or one manager per 3108 * pool. The exclusion is handled automatically by this function. 3109 * 3110 * The caller can safely start processing works on false return. On 3111 * true return, it's guaranteed that need_to_create_worker() is false 3112 * and may_start_working() is true. 3113 * 3114 * CONTEXT: 3115 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3116 * multiple times. Does GFP_KERNEL allocations. 3117 * 3118 * Return: 3119 * %false if the pool doesn't need management and the caller can safely 3120 * start processing works, %true if management function was performed and 3121 * the conditions that the caller verified before calling the function may 3122 * no longer be true. 3123 */ 3124 static bool manage_workers(struct worker *worker) 3125 { 3126 struct worker_pool *pool = worker->pool; 3127 3128 if (pool->flags & POOL_MANAGER_ACTIVE) 3129 return false; 3130 3131 pool->flags |= POOL_MANAGER_ACTIVE; 3132 pool->manager = worker; 3133 3134 maybe_create_worker(pool); 3135 3136 pool->manager = NULL; 3137 pool->flags &= ~POOL_MANAGER_ACTIVE; 3138 rcuwait_wake_up(&manager_wait); 3139 return true; 3140 } 3141 3142 /** 3143 * process_one_work - process single work 3144 * @worker: self 3145 * @work: work to process 3146 * 3147 * Process @work. This function contains all the logics necessary to 3148 * process a single work including synchronization against and 3149 * interaction with other workers on the same cpu, queueing and 3150 * flushing. As long as context requirement is met, any worker can 3151 * call this function to process a work. 3152 * 3153 * CONTEXT: 3154 * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 3155 */ 3156 static void process_one_work(struct worker *worker, struct work_struct *work) 3157 __releases(&pool->lock) 3158 __acquires(&pool->lock) 3159 { 3160 struct pool_workqueue *pwq = get_work_pwq(work); 3161 struct worker_pool *pool = worker->pool; 3162 unsigned long work_data; 3163 int lockdep_start_depth, rcu_start_depth; 3164 bool bh_draining = pool->flags & POOL_BH_DRAINING; 3165 #ifdef CONFIG_LOCKDEP 3166 /* 3167 * It is permissible to free the struct work_struct from 3168 * inside the function that is called from it, this we need to 3169 * take into account for lockdep too. To avoid bogus "held 3170 * lock freed" warnings as well as problems when looking into 3171 * work->lockdep_map, make a copy and use that here. 3172 */ 3173 struct lockdep_map lockdep_map; 3174 3175 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 3176 #endif 3177 /* ensure we're on the correct CPU */ 3178 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 3179 raw_smp_processor_id() != pool->cpu); 3180 3181 /* claim and dequeue */ 3182 debug_work_deactivate(work); 3183 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 3184 worker->current_work = work; 3185 worker->current_func = work->func; 3186 worker->current_pwq = pwq; 3187 if (worker->task) 3188 worker->current_at = worker->task->se.sum_exec_runtime; 3189 work_data = *work_data_bits(work); 3190 worker->current_color = get_work_color(work_data); 3191 3192 /* 3193 * Record wq name for cmdline and debug reporting, may get 3194 * overridden through set_worker_desc(). 3195 */ 3196 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 3197 3198 list_del_init(&work->entry); 3199 3200 /* 3201 * CPU intensive works don't participate in concurrency management. 3202 * They're the scheduler's responsibility. This takes @worker out 3203 * of concurrency management and the next code block will chain 3204 * execution of the pending work items. 3205 */ 3206 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 3207 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 3208 3209 /* 3210 * Kick @pool if necessary. It's always noop for per-cpu worker pools 3211 * since nr_running would always be >= 1 at this point. This is used to 3212 * chain execution of the pending work items for WORKER_NOT_RUNNING 3213 * workers such as the UNBOUND and CPU_INTENSIVE ones. 3214 */ 3215 kick_pool(pool); 3216 3217 /* 3218 * Record the last pool and clear PENDING which should be the last 3219 * update to @work. Also, do this inside @pool->lock so that 3220 * PENDING and queued state changes happen together while IRQ is 3221 * disabled. 3222 */ 3223 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool)); 3224 3225 pwq->stats[PWQ_STAT_STARTED]++; 3226 raw_spin_unlock_irq(&pool->lock); 3227 3228 rcu_start_depth = rcu_preempt_depth(); 3229 lockdep_start_depth = lockdep_depth(current); 3230 /* see drain_dead_softirq_workfn() */ 3231 if (!bh_draining) 3232 lock_map_acquire(pwq->wq->lockdep_map); 3233 lock_map_acquire(&lockdep_map); 3234 /* 3235 * Strictly speaking we should mark the invariant state without holding 3236 * any locks, that is, before these two lock_map_acquire()'s. 3237 * 3238 * However, that would result in: 3239 * 3240 * A(W1) 3241 * WFC(C) 3242 * A(W1) 3243 * C(C) 3244 * 3245 * Which would create W1->C->W1 dependencies, even though there is no 3246 * actual deadlock possible. There are two solutions, using a 3247 * read-recursive acquire on the work(queue) 'locks', but this will then 3248 * hit the lockdep limitation on recursive locks, or simply discard 3249 * these locks. 3250 * 3251 * AFAICT there is no possible deadlock scenario between the 3252 * flush_work() and complete() primitives (except for single-threaded 3253 * workqueues), so hiding them isn't a problem. 3254 */ 3255 lockdep_invariant_state(true); 3256 trace_workqueue_execute_start(work); 3257 worker->current_func(work); 3258 /* 3259 * While we must be careful to not use "work" after this, the trace 3260 * point will only record its address. 3261 */ 3262 trace_workqueue_execute_end(work, worker->current_func); 3263 3264 lock_map_release(&lockdep_map); 3265 if (!bh_draining) 3266 lock_map_release(pwq->wq->lockdep_map); 3267 3268 if (unlikely((worker->task && in_atomic()) || 3269 lockdep_depth(current) != lockdep_start_depth || 3270 rcu_preempt_depth() != rcu_start_depth)) { 3271 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n" 3272 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n", 3273 current->comm, task_pid_nr(current), preempt_count(), 3274 lockdep_start_depth, lockdep_depth(current), 3275 rcu_start_depth, rcu_preempt_depth(), 3276 worker->current_func); 3277 debug_show_held_locks(current); 3278 dump_stack(); 3279 } 3280 3281 /* 3282 * The following prevents a kworker from hogging CPU on !PREEMPTION 3283 * kernels, where a requeueing work item waiting for something to 3284 * happen could deadlock with stop_machine as such work item could 3285 * indefinitely requeue itself while all other CPUs are trapped in 3286 * stop_machine. At the same time, report a quiescent RCU state so 3287 * the same condition doesn't freeze RCU. 3288 */ 3289 if (worker->task) 3290 cond_resched(); 3291 3292 raw_spin_lock_irq(&pool->lock); 3293 3294 pwq->stats[PWQ_STAT_COMPLETED]++; 3295 3296 /* 3297 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 3298 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 3299 * wq_cpu_intensive_thresh_us. Clear it. 3300 */ 3301 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 3302 3303 /* tag the worker for identification in schedule() */ 3304 worker->last_func = worker->current_func; 3305 3306 /* we're done with it, release */ 3307 hash_del(&worker->hentry); 3308 worker->current_work = NULL; 3309 worker->current_func = NULL; 3310 worker->current_pwq = NULL; 3311 worker->current_color = INT_MAX; 3312 3313 /* must be the last step, see the function comment */ 3314 pwq_dec_nr_in_flight(pwq, work_data); 3315 } 3316 3317 /** 3318 * process_scheduled_works - process scheduled works 3319 * @worker: self 3320 * 3321 * Process all scheduled works. Please note that the scheduled list 3322 * may change while processing a work, so this function repeatedly 3323 * fetches a work from the top and executes it. 3324 * 3325 * CONTEXT: 3326 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3327 * multiple times. 3328 */ 3329 static void process_scheduled_works(struct worker *worker) 3330 { 3331 struct work_struct *work; 3332 bool first = true; 3333 3334 while ((work = list_first_entry_or_null(&worker->scheduled, 3335 struct work_struct, entry))) { 3336 if (first) { 3337 worker->pool->watchdog_ts = jiffies; 3338 first = false; 3339 } 3340 process_one_work(worker, work); 3341 } 3342 } 3343 3344 static void set_pf_worker(bool val) 3345 { 3346 mutex_lock(&wq_pool_attach_mutex); 3347 if (val) 3348 current->flags |= PF_WQ_WORKER; 3349 else 3350 current->flags &= ~PF_WQ_WORKER; 3351 mutex_unlock(&wq_pool_attach_mutex); 3352 } 3353 3354 /** 3355 * worker_thread - the worker thread function 3356 * @__worker: self 3357 * 3358 * The worker thread function. All workers belong to a worker_pool - 3359 * either a per-cpu one or dynamic unbound one. These workers process all 3360 * work items regardless of their specific target workqueue. The only 3361 * exception is work items which belong to workqueues with a rescuer which 3362 * will be explained in rescuer_thread(). 3363 * 3364 * Return: 0 3365 */ 3366 static int worker_thread(void *__worker) 3367 { 3368 struct worker *worker = __worker; 3369 struct worker_pool *pool = worker->pool; 3370 3371 /* tell the scheduler that this is a workqueue worker */ 3372 set_pf_worker(true); 3373 woke_up: 3374 raw_spin_lock_irq(&pool->lock); 3375 3376 /* am I supposed to die? */ 3377 if (unlikely(worker->flags & WORKER_DIE)) { 3378 raw_spin_unlock_irq(&pool->lock); 3379 set_pf_worker(false); 3380 /* 3381 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool 3382 * shouldn't be accessed, reset it to NULL in case otherwise. 3383 */ 3384 worker->pool = NULL; 3385 ida_free(&pool->worker_ida, worker->id); 3386 return 0; 3387 } 3388 3389 worker_leave_idle(worker); 3390 recheck: 3391 /* no more worker necessary? */ 3392 if (!need_more_worker(pool)) 3393 goto sleep; 3394 3395 /* do we need to manage? */ 3396 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 3397 goto recheck; 3398 3399 /* 3400 * ->scheduled list can only be filled while a worker is 3401 * preparing to process a work or actually processing it. 3402 * Make sure nobody diddled with it while I was sleeping. 3403 */ 3404 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3405 3406 /* 3407 * Finish PREP stage. We're guaranteed to have at least one idle 3408 * worker or that someone else has already assumed the manager 3409 * role. This is where @worker starts participating in concurrency 3410 * management if applicable and concurrency management is restored 3411 * after being rebound. See rebind_workers() for details. 3412 */ 3413 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3414 3415 do { 3416 struct work_struct *work = 3417 list_first_entry(&pool->worklist, 3418 struct work_struct, entry); 3419 3420 if (assign_work(work, worker, NULL)) 3421 process_scheduled_works(worker); 3422 } while (keep_working(pool)); 3423 3424 worker_set_flags(worker, WORKER_PREP); 3425 sleep: 3426 /* 3427 * pool->lock is held and there's no work to process and no need to 3428 * manage, sleep. Workers are woken up only while holding 3429 * pool->lock or from local cpu, so setting the current state 3430 * before releasing pool->lock is enough to prevent losing any 3431 * event. 3432 */ 3433 worker_enter_idle(worker); 3434 __set_current_state(TASK_IDLE); 3435 raw_spin_unlock_irq(&pool->lock); 3436 schedule(); 3437 goto woke_up; 3438 } 3439 3440 static bool assign_rescuer_work(struct pool_workqueue *pwq, struct worker *rescuer) 3441 { 3442 struct worker_pool *pool = pwq->pool; 3443 struct work_struct *work, *n; 3444 3445 /* need rescue? */ 3446 if (!pwq->nr_active || !need_to_create_worker(pool)) 3447 return false; 3448 3449 /* 3450 * Slurp in all works issued via this workqueue and 3451 * process'em. 3452 */ 3453 list_for_each_entry_safe(work, n, &pool->worklist, entry) { 3454 if (get_work_pwq(work) == pwq && assign_work(work, rescuer, &n)) 3455 pwq->stats[PWQ_STAT_RESCUED]++; 3456 } 3457 3458 return !list_empty(&rescuer->scheduled); 3459 } 3460 3461 /** 3462 * rescuer_thread - the rescuer thread function 3463 * @__rescuer: self 3464 * 3465 * Workqueue rescuer thread function. There's one rescuer for each 3466 * workqueue which has WQ_MEM_RECLAIM set. 3467 * 3468 * Regular work processing on a pool may block trying to create a new 3469 * worker which uses GFP_KERNEL allocation which has slight chance of 3470 * developing into deadlock if some works currently on the same queue 3471 * need to be processed to satisfy the GFP_KERNEL allocation. This is 3472 * the problem rescuer solves. 3473 * 3474 * When such condition is possible, the pool summons rescuers of all 3475 * workqueues which have works queued on the pool and let them process 3476 * those works so that forward progress can be guaranteed. 3477 * 3478 * This should happen rarely. 3479 * 3480 * Return: 0 3481 */ 3482 static int rescuer_thread(void *__rescuer) 3483 { 3484 struct worker *rescuer = __rescuer; 3485 struct workqueue_struct *wq = rescuer->rescue_wq; 3486 bool should_stop; 3487 3488 set_user_nice(current, RESCUER_NICE_LEVEL); 3489 3490 /* 3491 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 3492 * doesn't participate in concurrency management. 3493 */ 3494 set_pf_worker(true); 3495 repeat: 3496 set_current_state(TASK_IDLE); 3497 3498 /* 3499 * By the time the rescuer is requested to stop, the workqueue 3500 * shouldn't have any work pending, but @wq->maydays may still have 3501 * pwq(s) queued. This can happen by non-rescuer workers consuming 3502 * all the work items before the rescuer got to them. Go through 3503 * @wq->maydays processing before acting on should_stop so that the 3504 * list is always empty on exit. 3505 */ 3506 should_stop = kthread_should_stop(); 3507 3508 /* see whether any pwq is asking for help */ 3509 raw_spin_lock_irq(&wq_mayday_lock); 3510 3511 while (!list_empty(&wq->maydays)) { 3512 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 3513 struct pool_workqueue, mayday_node); 3514 struct worker_pool *pool = pwq->pool; 3515 3516 __set_current_state(TASK_RUNNING); 3517 list_del_init(&pwq->mayday_node); 3518 3519 raw_spin_unlock_irq(&wq_mayday_lock); 3520 3521 worker_attach_to_pool(rescuer, pool); 3522 3523 raw_spin_lock_irq(&pool->lock); 3524 3525 WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 3526 3527 if (assign_rescuer_work(pwq, rescuer)) { 3528 process_scheduled_works(rescuer); 3529 3530 /* 3531 * The above execution of rescued work items could 3532 * have created more to rescue through 3533 * pwq_activate_first_inactive() or chained 3534 * queueing. Let's put @pwq back on mayday list so 3535 * that such back-to-back work items, which may be 3536 * being used to relieve memory pressure, don't 3537 * incur MAYDAY_INTERVAL delay inbetween. 3538 */ 3539 if (pwq->nr_active && need_to_create_worker(pool)) { 3540 raw_spin_lock(&wq_mayday_lock); 3541 /* 3542 * Queue iff somebody else hasn't queued it already. 3543 */ 3544 if (list_empty(&pwq->mayday_node)) { 3545 get_pwq(pwq); 3546 list_add_tail(&pwq->mayday_node, &wq->maydays); 3547 } 3548 raw_spin_unlock(&wq_mayday_lock); 3549 } 3550 } 3551 3552 /* 3553 * Leave this pool. Notify regular workers; otherwise, we end up 3554 * with 0 concurrency and stalling the execution. 3555 */ 3556 kick_pool(pool); 3557 3558 raw_spin_unlock_irq(&pool->lock); 3559 3560 worker_detach_from_pool(rescuer); 3561 3562 /* 3563 * Put the reference grabbed by send_mayday(). @pool might 3564 * go away any time after it. 3565 */ 3566 put_pwq_unlocked(pwq); 3567 3568 raw_spin_lock_irq(&wq_mayday_lock); 3569 } 3570 3571 raw_spin_unlock_irq(&wq_mayday_lock); 3572 3573 if (should_stop) { 3574 __set_current_state(TASK_RUNNING); 3575 set_pf_worker(false); 3576 return 0; 3577 } 3578 3579 /* rescuers should never participate in concurrency management */ 3580 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 3581 schedule(); 3582 goto repeat; 3583 } 3584 3585 static void bh_worker(struct worker *worker) 3586 { 3587 struct worker_pool *pool = worker->pool; 3588 int nr_restarts = BH_WORKER_RESTARTS; 3589 unsigned long end = jiffies + BH_WORKER_JIFFIES; 3590 3591 worker_lock_callback(pool); 3592 raw_spin_lock_irq(&pool->lock); 3593 worker_leave_idle(worker); 3594 3595 /* 3596 * This function follows the structure of worker_thread(). See there for 3597 * explanations on each step. 3598 */ 3599 if (!need_more_worker(pool)) 3600 goto done; 3601 3602 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3603 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3604 3605 do { 3606 struct work_struct *work = 3607 list_first_entry(&pool->worklist, 3608 struct work_struct, entry); 3609 3610 if (assign_work(work, worker, NULL)) 3611 process_scheduled_works(worker); 3612 } while (keep_working(pool) && 3613 --nr_restarts && time_before(jiffies, end)); 3614 3615 worker_set_flags(worker, WORKER_PREP); 3616 done: 3617 worker_enter_idle(worker); 3618 kick_pool(pool); 3619 raw_spin_unlock_irq(&pool->lock); 3620 worker_unlock_callback(pool); 3621 } 3622 3623 /* 3624 * TODO: Convert all tasklet users to workqueue and use softirq directly. 3625 * 3626 * This is currently called from tasklet[_hi]action() and thus is also called 3627 * whenever there are tasklets to run. Let's do an early exit if there's nothing 3628 * queued. Once conversion from tasklet is complete, the need_more_worker() test 3629 * can be dropped. 3630 * 3631 * After full conversion, we'll add worker->softirq_action, directly use the 3632 * softirq action and obtain the worker pointer from the softirq_action pointer. 3633 */ 3634 void workqueue_softirq_action(bool highpri) 3635 { 3636 struct worker_pool *pool = 3637 &per_cpu(bh_worker_pools, smp_processor_id())[highpri]; 3638 if (need_more_worker(pool)) 3639 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3640 } 3641 3642 struct wq_drain_dead_softirq_work { 3643 struct work_struct work; 3644 struct worker_pool *pool; 3645 struct completion done; 3646 }; 3647 3648 static void drain_dead_softirq_workfn(struct work_struct *work) 3649 { 3650 struct wq_drain_dead_softirq_work *dead_work = 3651 container_of(work, struct wq_drain_dead_softirq_work, work); 3652 struct worker_pool *pool = dead_work->pool; 3653 bool repeat; 3654 3655 /* 3656 * @pool's CPU is dead and we want to execute its still pending work 3657 * items from this BH work item which is running on a different CPU. As 3658 * its CPU is dead, @pool can't be kicked and, as work execution path 3659 * will be nested, a lockdep annotation needs to be suppressed. Mark 3660 * @pool with %POOL_BH_DRAINING for the special treatments. 3661 */ 3662 raw_spin_lock_irq(&pool->lock); 3663 pool->flags |= POOL_BH_DRAINING; 3664 raw_spin_unlock_irq(&pool->lock); 3665 3666 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3667 3668 raw_spin_lock_irq(&pool->lock); 3669 pool->flags &= ~POOL_BH_DRAINING; 3670 repeat = need_more_worker(pool); 3671 raw_spin_unlock_irq(&pool->lock); 3672 3673 /* 3674 * bh_worker() might hit consecutive execution limit and bail. If there 3675 * still are pending work items, reschedule self and return so that we 3676 * don't hog this CPU's BH. 3677 */ 3678 if (repeat) { 3679 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3680 queue_work(system_bh_highpri_wq, work); 3681 else 3682 queue_work(system_bh_wq, work); 3683 } else { 3684 complete(&dead_work->done); 3685 } 3686 } 3687 3688 /* 3689 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's 3690 * possible to allocate dead_work per CPU and avoid flushing. However, then we 3691 * have to worry about draining overlapping with CPU coming back online or 3692 * nesting (one CPU's dead_work queued on another CPU which is also dead and so 3693 * on). Let's keep it simple and drain them synchronously. These are BH work 3694 * items which shouldn't be requeued on the same pool. Shouldn't take long. 3695 */ 3696 void workqueue_softirq_dead(unsigned int cpu) 3697 { 3698 int i; 3699 3700 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 3701 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i]; 3702 struct wq_drain_dead_softirq_work dead_work; 3703 3704 if (!need_more_worker(pool)) 3705 continue; 3706 3707 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn); 3708 dead_work.pool = pool; 3709 init_completion(&dead_work.done); 3710 3711 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3712 queue_work(system_bh_highpri_wq, &dead_work.work); 3713 else 3714 queue_work(system_bh_wq, &dead_work.work); 3715 3716 wait_for_completion(&dead_work.done); 3717 destroy_work_on_stack(&dead_work.work); 3718 } 3719 } 3720 3721 /** 3722 * check_flush_dependency - check for flush dependency sanity 3723 * @target_wq: workqueue being flushed 3724 * @target_work: work item being flushed (NULL for workqueue flushes) 3725 * @from_cancel: are we called from the work cancel path 3726 * 3727 * %current is trying to flush the whole @target_wq or @target_work on it. 3728 * If this is not the cancel path (which implies work being flushed is either 3729 * already running, or will not be at all), check if @target_wq doesn't have 3730 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running 3731 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward- 3732 * progress guarantee leading to a deadlock. 3733 */ 3734 static void check_flush_dependency(struct workqueue_struct *target_wq, 3735 struct work_struct *target_work, 3736 bool from_cancel) 3737 { 3738 work_func_t target_func; 3739 struct worker *worker; 3740 3741 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM) 3742 return; 3743 3744 worker = current_wq_worker(); 3745 target_func = target_work ? target_work->func : NULL; 3746 3747 WARN_ONCE(current->flags & PF_MEMALLOC, 3748 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 3749 current->pid, current->comm, target_wq->name, target_func); 3750 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 3751 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 3752 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 3753 worker->current_pwq->wq->name, worker->current_func, 3754 target_wq->name, target_func); 3755 } 3756 3757 struct wq_barrier { 3758 struct work_struct work; 3759 struct completion done; 3760 struct task_struct *task; /* purely informational */ 3761 }; 3762 3763 static void wq_barrier_func(struct work_struct *work) 3764 { 3765 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 3766 complete(&barr->done); 3767 } 3768 3769 /** 3770 * insert_wq_barrier - insert a barrier work 3771 * @pwq: pwq to insert barrier into 3772 * @barr: wq_barrier to insert 3773 * @target: target work to attach @barr to 3774 * @worker: worker currently executing @target, NULL if @target is not executing 3775 * 3776 * @barr is linked to @target such that @barr is completed only after 3777 * @target finishes execution. Please note that the ordering 3778 * guarantee is observed only with respect to @target and on the local 3779 * cpu. 3780 * 3781 * Currently, a queued barrier can't be canceled. This is because 3782 * try_to_grab_pending() can't determine whether the work to be 3783 * grabbed is at the head of the queue and thus can't clear LINKED 3784 * flag of the previous work while there must be a valid next work 3785 * after a work with LINKED flag set. 3786 * 3787 * Note that when @worker is non-NULL, @target may be modified 3788 * underneath us, so we can't reliably determine pwq from @target. 3789 * 3790 * CONTEXT: 3791 * raw_spin_lock_irq(pool->lock). 3792 */ 3793 static void insert_wq_barrier(struct pool_workqueue *pwq, 3794 struct wq_barrier *barr, 3795 struct work_struct *target, struct worker *worker) 3796 { 3797 static __maybe_unused struct lock_class_key bh_key, thr_key; 3798 unsigned int work_flags = 0; 3799 unsigned int work_color; 3800 struct list_head *head; 3801 3802 /* 3803 * debugobject calls are safe here even with pool->lock locked 3804 * as we know for sure that this will not trigger any of the 3805 * checks and call back into the fixup functions where we 3806 * might deadlock. 3807 * 3808 * BH and threaded workqueues need separate lockdep keys to avoid 3809 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} 3810 * usage". 3811 */ 3812 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func, 3813 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key); 3814 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 3815 3816 init_completion_map(&barr->done, &target->lockdep_map); 3817 3818 barr->task = current; 3819 3820 /* The barrier work item does not participate in nr_active. */ 3821 work_flags |= WORK_STRUCT_INACTIVE; 3822 3823 /* 3824 * If @target is currently being executed, schedule the 3825 * barrier to the worker; otherwise, put it after @target. 3826 */ 3827 if (worker) { 3828 head = worker->scheduled.next; 3829 work_color = worker->current_color; 3830 } else { 3831 unsigned long *bits = work_data_bits(target); 3832 3833 head = target->entry.next; 3834 /* there can already be other linked works, inherit and set */ 3835 work_flags |= *bits & WORK_STRUCT_LINKED; 3836 work_color = get_work_color(*bits); 3837 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3838 } 3839 3840 pwq->nr_in_flight[work_color]++; 3841 work_flags |= work_color_to_flags(work_color); 3842 3843 insert_work(pwq, &barr->work, head, work_flags); 3844 } 3845 3846 /** 3847 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 3848 * @wq: workqueue being flushed 3849 * @flush_color: new flush color, < 0 for no-op 3850 * @work_color: new work color, < 0 for no-op 3851 * 3852 * Prepare pwqs for workqueue flushing. 3853 * 3854 * If @flush_color is non-negative, flush_color on all pwqs should be 3855 * -1. If no pwq has in-flight commands at the specified color, all 3856 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3857 * has in flight commands, its pwq->flush_color is set to 3858 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 3859 * wakeup logic is armed and %true is returned. 3860 * 3861 * The caller should have initialized @wq->first_flusher prior to 3862 * calling this function with non-negative @flush_color. If 3863 * @flush_color is negative, no flush color update is done and %false 3864 * is returned. 3865 * 3866 * If @work_color is non-negative, all pwqs should have the same 3867 * work_color which is previous to @work_color and all will be 3868 * advanced to @work_color. 3869 * 3870 * CONTEXT: 3871 * mutex_lock(wq->mutex). 3872 * 3873 * Return: 3874 * %true if @flush_color >= 0 and there's something to flush. %false 3875 * otherwise. 3876 */ 3877 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 3878 int flush_color, int work_color) 3879 { 3880 bool wait = false; 3881 struct pool_workqueue *pwq; 3882 struct worker_pool *current_pool = NULL; 3883 3884 if (flush_color >= 0) { 3885 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3886 atomic_set(&wq->nr_pwqs_to_flush, 1); 3887 } 3888 3889 /* 3890 * For unbound workqueue, pwqs will map to only a few pools. 3891 * Most of the time, pwqs within the same pool will be linked 3892 * sequentially to wq->pwqs by cpu index. So in the majority 3893 * of pwq iters, the pool is the same, only doing lock/unlock 3894 * if the pool has changed. This can largely reduce expensive 3895 * lock operations. 3896 */ 3897 for_each_pwq(pwq, wq) { 3898 if (current_pool != pwq->pool) { 3899 if (likely(current_pool)) 3900 raw_spin_unlock_irq(¤t_pool->lock); 3901 current_pool = pwq->pool; 3902 raw_spin_lock_irq(¤t_pool->lock); 3903 } 3904 3905 if (flush_color >= 0) { 3906 WARN_ON_ONCE(pwq->flush_color != -1); 3907 3908 if (pwq->nr_in_flight[flush_color]) { 3909 pwq->flush_color = flush_color; 3910 atomic_inc(&wq->nr_pwqs_to_flush); 3911 wait = true; 3912 } 3913 } 3914 3915 if (work_color >= 0) { 3916 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3917 pwq->work_color = work_color; 3918 } 3919 3920 } 3921 3922 if (current_pool) 3923 raw_spin_unlock_irq(¤t_pool->lock); 3924 3925 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 3926 complete(&wq->first_flusher->done); 3927 3928 return wait; 3929 } 3930 3931 static void touch_wq_lockdep_map(struct workqueue_struct *wq) 3932 { 3933 #ifdef CONFIG_LOCKDEP 3934 if (unlikely(!wq->lockdep_map)) 3935 return; 3936 3937 if (wq->flags & WQ_BH) 3938 local_bh_disable(); 3939 3940 lock_map_acquire(wq->lockdep_map); 3941 lock_map_release(wq->lockdep_map); 3942 3943 if (wq->flags & WQ_BH) 3944 local_bh_enable(); 3945 #endif 3946 } 3947 3948 static void touch_work_lockdep_map(struct work_struct *work, 3949 struct workqueue_struct *wq) 3950 { 3951 #ifdef CONFIG_LOCKDEP 3952 if (wq->flags & WQ_BH) 3953 local_bh_disable(); 3954 3955 lock_map_acquire(&work->lockdep_map); 3956 lock_map_release(&work->lockdep_map); 3957 3958 if (wq->flags & WQ_BH) 3959 local_bh_enable(); 3960 #endif 3961 } 3962 3963 /** 3964 * __flush_workqueue - ensure that any scheduled work has run to completion. 3965 * @wq: workqueue to flush 3966 * 3967 * This function sleeps until all work items which were queued on entry 3968 * have finished execution, but it is not livelocked by new incoming ones. 3969 */ 3970 void __flush_workqueue(struct workqueue_struct *wq) 3971 { 3972 struct wq_flusher this_flusher = { 3973 .list = LIST_HEAD_INIT(this_flusher.list), 3974 .flush_color = -1, 3975 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)), 3976 }; 3977 int next_color; 3978 3979 if (WARN_ON(!wq_online)) 3980 return; 3981 3982 touch_wq_lockdep_map(wq); 3983 3984 mutex_lock(&wq->mutex); 3985 3986 /* 3987 * Start-to-wait phase 3988 */ 3989 next_color = work_next_color(wq->work_color); 3990 3991 if (next_color != wq->flush_color) { 3992 /* 3993 * Color space is not full. The current work_color 3994 * becomes our flush_color and work_color is advanced 3995 * by one. 3996 */ 3997 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 3998 this_flusher.flush_color = wq->work_color; 3999 wq->work_color = next_color; 4000 4001 if (!wq->first_flusher) { 4002 /* no flush in progress, become the first flusher */ 4003 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 4004 4005 wq->first_flusher = &this_flusher; 4006 4007 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 4008 wq->work_color)) { 4009 /* nothing to flush, done */ 4010 wq->flush_color = next_color; 4011 wq->first_flusher = NULL; 4012 goto out_unlock; 4013 } 4014 } else { 4015 /* wait in queue */ 4016 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 4017 list_add_tail(&this_flusher.list, &wq->flusher_queue); 4018 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 4019 } 4020 } else { 4021 /* 4022 * Oops, color space is full, wait on overflow queue. 4023 * The next flush completion will assign us 4024 * flush_color and transfer to flusher_queue. 4025 */ 4026 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 4027 } 4028 4029 check_flush_dependency(wq, NULL, false); 4030 4031 mutex_unlock(&wq->mutex); 4032 4033 wait_for_completion(&this_flusher.done); 4034 4035 /* 4036 * Wake-up-and-cascade phase 4037 * 4038 * First flushers are responsible for cascading flushes and 4039 * handling overflow. Non-first flushers can simply return. 4040 */ 4041 if (READ_ONCE(wq->first_flusher) != &this_flusher) 4042 return; 4043 4044 mutex_lock(&wq->mutex); 4045 4046 /* we might have raced, check again with mutex held */ 4047 if (wq->first_flusher != &this_flusher) 4048 goto out_unlock; 4049 4050 WRITE_ONCE(wq->first_flusher, NULL); 4051 4052 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 4053 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 4054 4055 while (true) { 4056 struct wq_flusher *next, *tmp; 4057 4058 /* complete all the flushers sharing the current flush color */ 4059 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 4060 if (next->flush_color != wq->flush_color) 4061 break; 4062 list_del_init(&next->list); 4063 complete(&next->done); 4064 } 4065 4066 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 4067 wq->flush_color != work_next_color(wq->work_color)); 4068 4069 /* this flush_color is finished, advance by one */ 4070 wq->flush_color = work_next_color(wq->flush_color); 4071 4072 /* one color has been freed, handle overflow queue */ 4073 if (!list_empty(&wq->flusher_overflow)) { 4074 /* 4075 * Assign the same color to all overflowed 4076 * flushers, advance work_color and append to 4077 * flusher_queue. This is the start-to-wait 4078 * phase for these overflowed flushers. 4079 */ 4080 list_for_each_entry(tmp, &wq->flusher_overflow, list) 4081 tmp->flush_color = wq->work_color; 4082 4083 wq->work_color = work_next_color(wq->work_color); 4084 4085 list_splice_tail_init(&wq->flusher_overflow, 4086 &wq->flusher_queue); 4087 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 4088 } 4089 4090 if (list_empty(&wq->flusher_queue)) { 4091 WARN_ON_ONCE(wq->flush_color != wq->work_color); 4092 break; 4093 } 4094 4095 /* 4096 * Need to flush more colors. Make the next flusher 4097 * the new first flusher and arm pwqs. 4098 */ 4099 WARN_ON_ONCE(wq->flush_color == wq->work_color); 4100 WARN_ON_ONCE(wq->flush_color != next->flush_color); 4101 4102 list_del_init(&next->list); 4103 wq->first_flusher = next; 4104 4105 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 4106 break; 4107 4108 /* 4109 * Meh... this color is already done, clear first 4110 * flusher and repeat cascading. 4111 */ 4112 wq->first_flusher = NULL; 4113 } 4114 4115 out_unlock: 4116 mutex_unlock(&wq->mutex); 4117 } 4118 EXPORT_SYMBOL(__flush_workqueue); 4119 4120 /** 4121 * drain_workqueue - drain a workqueue 4122 * @wq: workqueue to drain 4123 * 4124 * Wait until the workqueue becomes empty. While draining is in progress, 4125 * only chain queueing is allowed. IOW, only currently pending or running 4126 * work items on @wq can queue further work items on it. @wq is flushed 4127 * repeatedly until it becomes empty. The number of flushing is determined 4128 * by the depth of chaining and should be relatively short. Whine if it 4129 * takes too long. 4130 */ 4131 void drain_workqueue(struct workqueue_struct *wq) 4132 { 4133 unsigned int flush_cnt = 0; 4134 struct pool_workqueue *pwq; 4135 4136 /* 4137 * __queue_work() needs to test whether there are drainers, is much 4138 * hotter than drain_workqueue() and already looks at @wq->flags. 4139 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 4140 */ 4141 mutex_lock(&wq->mutex); 4142 if (!wq->nr_drainers++) 4143 wq->flags |= __WQ_DRAINING; 4144 mutex_unlock(&wq->mutex); 4145 reflush: 4146 __flush_workqueue(wq); 4147 4148 mutex_lock(&wq->mutex); 4149 4150 for_each_pwq(pwq, wq) { 4151 bool drained; 4152 4153 raw_spin_lock_irq(&pwq->pool->lock); 4154 drained = pwq_is_empty(pwq); 4155 raw_spin_unlock_irq(&pwq->pool->lock); 4156 4157 if (drained) 4158 continue; 4159 4160 if (++flush_cnt == 10 || 4161 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 4162 pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 4163 wq->name, __func__, flush_cnt); 4164 4165 mutex_unlock(&wq->mutex); 4166 goto reflush; 4167 } 4168 4169 if (!--wq->nr_drainers) 4170 wq->flags &= ~__WQ_DRAINING; 4171 mutex_unlock(&wq->mutex); 4172 } 4173 EXPORT_SYMBOL_GPL(drain_workqueue); 4174 4175 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 4176 bool from_cancel) 4177 { 4178 struct worker *worker = NULL; 4179 struct worker_pool *pool; 4180 struct pool_workqueue *pwq; 4181 struct workqueue_struct *wq; 4182 4183 rcu_read_lock(); 4184 pool = get_work_pool(work); 4185 if (!pool) { 4186 rcu_read_unlock(); 4187 return false; 4188 } 4189 4190 raw_spin_lock_irq(&pool->lock); 4191 /* see the comment in try_to_grab_pending() with the same code */ 4192 pwq = get_work_pwq(work); 4193 if (pwq) { 4194 if (unlikely(pwq->pool != pool)) 4195 goto already_gone; 4196 } else { 4197 worker = find_worker_executing_work(pool, work); 4198 if (!worker) 4199 goto already_gone; 4200 pwq = worker->current_pwq; 4201 } 4202 4203 wq = pwq->wq; 4204 check_flush_dependency(wq, work, from_cancel); 4205 4206 insert_wq_barrier(pwq, barr, work, worker); 4207 raw_spin_unlock_irq(&pool->lock); 4208 4209 touch_work_lockdep_map(work, wq); 4210 4211 /* 4212 * Force a lock recursion deadlock when using flush_work() inside a 4213 * single-threaded or rescuer equipped workqueue. 4214 * 4215 * For single threaded workqueues the deadlock happens when the work 4216 * is after the work issuing the flush_work(). For rescuer equipped 4217 * workqueues the deadlock happens when the rescuer stalls, blocking 4218 * forward progress. 4219 */ 4220 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer)) 4221 touch_wq_lockdep_map(wq); 4222 4223 rcu_read_unlock(); 4224 return true; 4225 already_gone: 4226 raw_spin_unlock_irq(&pool->lock); 4227 rcu_read_unlock(); 4228 return false; 4229 } 4230 4231 static bool __flush_work(struct work_struct *work, bool from_cancel) 4232 { 4233 struct wq_barrier barr; 4234 4235 if (WARN_ON(!wq_online)) 4236 return false; 4237 4238 if (WARN_ON(!work->func)) 4239 return false; 4240 4241 if (!start_flush_work(work, &barr, from_cancel)) 4242 return false; 4243 4244 /* 4245 * start_flush_work() returned %true. If @from_cancel is set, we know 4246 * that @work must have been executing during start_flush_work() and 4247 * can't currently be queued. Its data must contain OFFQ bits. If @work 4248 * was queued on a BH workqueue, we also know that it was running in the 4249 * BH context and thus can be busy-waited. 4250 */ 4251 if (from_cancel) { 4252 unsigned long data = *work_data_bits(work); 4253 4254 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && 4255 (data & WORK_OFFQ_BH)) { 4256 /* 4257 * On RT, prevent a live lock when %current preempted 4258 * soft interrupt processing by blocking on lock which 4259 * is owned by the thread invoking the callback. 4260 */ 4261 while (!try_wait_for_completion(&barr.done)) { 4262 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 4263 struct worker_pool *pool; 4264 4265 guard(rcu)(); 4266 pool = get_work_pool(work); 4267 if (pool) 4268 workqueue_callback_cancel_wait_running(pool); 4269 } else { 4270 cpu_relax(); 4271 } 4272 } 4273 goto out_destroy; 4274 } 4275 } 4276 4277 wait_for_completion(&barr.done); 4278 4279 out_destroy: 4280 destroy_work_on_stack(&barr.work); 4281 return true; 4282 } 4283 4284 /** 4285 * flush_work - wait for a work to finish executing the last queueing instance 4286 * @work: the work to flush 4287 * 4288 * Wait until @work has finished execution. @work is guaranteed to be idle 4289 * on return if it hasn't been requeued since flush started. 4290 * 4291 * Return: 4292 * %true if flush_work() waited for the work to finish execution, 4293 * %false if it was already idle. 4294 */ 4295 bool flush_work(struct work_struct *work) 4296 { 4297 might_sleep(); 4298 return __flush_work(work, false); 4299 } 4300 EXPORT_SYMBOL_GPL(flush_work); 4301 4302 /** 4303 * flush_delayed_work - wait for a dwork to finish executing the last queueing 4304 * @dwork: the delayed work to flush 4305 * 4306 * Delayed timer is cancelled and the pending work is queued for 4307 * immediate execution. Like flush_work(), this function only 4308 * considers the last queueing instance of @dwork. 4309 * 4310 * Return: 4311 * %true if flush_work() waited for the work to finish execution, 4312 * %false if it was already idle. 4313 */ 4314 bool flush_delayed_work(struct delayed_work *dwork) 4315 { 4316 local_irq_disable(); 4317 if (timer_delete_sync(&dwork->timer)) 4318 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 4319 local_irq_enable(); 4320 return flush_work(&dwork->work); 4321 } 4322 EXPORT_SYMBOL(flush_delayed_work); 4323 4324 /** 4325 * flush_rcu_work - wait for a rwork to finish executing the last queueing 4326 * @rwork: the rcu work to flush 4327 * 4328 * Return: 4329 * %true if flush_rcu_work() waited for the work to finish execution, 4330 * %false if it was already idle. 4331 */ 4332 bool flush_rcu_work(struct rcu_work *rwork) 4333 { 4334 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 4335 rcu_barrier(); 4336 flush_work(&rwork->work); 4337 return true; 4338 } else { 4339 return flush_work(&rwork->work); 4340 } 4341 } 4342 EXPORT_SYMBOL(flush_rcu_work); 4343 4344 static void work_offqd_disable(struct work_offq_data *offqd) 4345 { 4346 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1; 4347 4348 if (likely(offqd->disable < max)) 4349 offqd->disable++; 4350 else 4351 WARN_ONCE(true, "workqueue: work disable count overflowed\n"); 4352 } 4353 4354 static void work_offqd_enable(struct work_offq_data *offqd) 4355 { 4356 if (likely(offqd->disable > 0)) 4357 offqd->disable--; 4358 else 4359 WARN_ONCE(true, "workqueue: work disable count underflowed\n"); 4360 } 4361 4362 static bool __cancel_work(struct work_struct *work, u32 cflags) 4363 { 4364 struct work_offq_data offqd; 4365 unsigned long irq_flags; 4366 int ret; 4367 4368 ret = work_grab_pending(work, cflags, &irq_flags); 4369 4370 work_offqd_unpack(&offqd, *work_data_bits(work)); 4371 4372 if (cflags & WORK_CANCEL_DISABLE) 4373 work_offqd_disable(&offqd); 4374 4375 set_work_pool_and_clear_pending(work, offqd.pool_id, 4376 work_offqd_pack_flags(&offqd)); 4377 local_irq_restore(irq_flags); 4378 return ret; 4379 } 4380 4381 static bool __cancel_work_sync(struct work_struct *work, u32 cflags) 4382 { 4383 bool ret; 4384 4385 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE); 4386 4387 if (*work_data_bits(work) & WORK_OFFQ_BH) 4388 WARN_ON_ONCE(in_hardirq()); 4389 else 4390 might_sleep(); 4391 4392 /* 4393 * Skip __flush_work() during early boot when we know that @work isn't 4394 * executing. This allows canceling during early boot. 4395 */ 4396 if (wq_online) 4397 __flush_work(work, true); 4398 4399 if (!(cflags & WORK_CANCEL_DISABLE)) 4400 enable_work(work); 4401 4402 return ret; 4403 } 4404 4405 /* 4406 * See cancel_delayed_work() 4407 */ 4408 bool cancel_work(struct work_struct *work) 4409 { 4410 return __cancel_work(work, 0); 4411 } 4412 EXPORT_SYMBOL(cancel_work); 4413 4414 /** 4415 * cancel_work_sync - cancel a work and wait for it to finish 4416 * @work: the work to cancel 4417 * 4418 * Cancel @work and wait for its execution to finish. This function can be used 4419 * even if the work re-queues itself or migrates to another workqueue. On return 4420 * from this function, @work is guaranteed to be not pending or executing on any 4421 * CPU as long as there aren't racing enqueues. 4422 * 4423 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's. 4424 * Use cancel_delayed_work_sync() instead. 4425 * 4426 * Must be called from a sleepable context if @work was last queued on a non-BH 4427 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4428 * if @work was last queued on a BH workqueue. 4429 * 4430 * Returns %true if @work was pending, %false otherwise. 4431 */ 4432 bool cancel_work_sync(struct work_struct *work) 4433 { 4434 return __cancel_work_sync(work, 0); 4435 } 4436 EXPORT_SYMBOL_GPL(cancel_work_sync); 4437 4438 /** 4439 * cancel_delayed_work - cancel a delayed work 4440 * @dwork: delayed_work to cancel 4441 * 4442 * Kill off a pending delayed_work. 4443 * 4444 * Return: %true if @dwork was pending and canceled; %false if it wasn't 4445 * pending. 4446 * 4447 * Note: 4448 * The work callback function may still be running on return, unless 4449 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 4450 * use cancel_delayed_work_sync() to wait on it. 4451 * 4452 * This function is safe to call from any context including IRQ handler. 4453 */ 4454 bool cancel_delayed_work(struct delayed_work *dwork) 4455 { 4456 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED); 4457 } 4458 EXPORT_SYMBOL(cancel_delayed_work); 4459 4460 /** 4461 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 4462 * @dwork: the delayed work cancel 4463 * 4464 * This is cancel_work_sync() for delayed works. 4465 * 4466 * Return: 4467 * %true if @dwork was pending, %false otherwise. 4468 */ 4469 bool cancel_delayed_work_sync(struct delayed_work *dwork) 4470 { 4471 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED); 4472 } 4473 EXPORT_SYMBOL(cancel_delayed_work_sync); 4474 4475 /** 4476 * disable_work - Disable and cancel a work item 4477 * @work: work item to disable 4478 * 4479 * Disable @work by incrementing its disable count and cancel it if currently 4480 * pending. As long as the disable count is non-zero, any attempt to queue @work 4481 * will fail and return %false. The maximum supported disable depth is 2 to the 4482 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536. 4483 * 4484 * Can be called from any context. Returns %true if @work was pending, %false 4485 * otherwise. 4486 */ 4487 bool disable_work(struct work_struct *work) 4488 { 4489 return __cancel_work(work, WORK_CANCEL_DISABLE); 4490 } 4491 EXPORT_SYMBOL_GPL(disable_work); 4492 4493 /** 4494 * disable_work_sync - Disable, cancel and drain a work item 4495 * @work: work item to disable 4496 * 4497 * Similar to disable_work() but also wait for @work to finish if currently 4498 * executing. 4499 * 4500 * Must be called from a sleepable context if @work was last queued on a non-BH 4501 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4502 * if @work was last queued on a BH workqueue. 4503 * 4504 * Returns %true if @work was pending, %false otherwise. 4505 */ 4506 bool disable_work_sync(struct work_struct *work) 4507 { 4508 return __cancel_work_sync(work, WORK_CANCEL_DISABLE); 4509 } 4510 EXPORT_SYMBOL_GPL(disable_work_sync); 4511 4512 /** 4513 * enable_work - Enable a work item 4514 * @work: work item to enable 4515 * 4516 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can 4517 * only be queued if its disable count is 0. 4518 * 4519 * Can be called from any context. Returns %true if the disable count reached 0. 4520 * Otherwise, %false. 4521 */ 4522 bool enable_work(struct work_struct *work) 4523 { 4524 struct work_offq_data offqd; 4525 unsigned long irq_flags; 4526 4527 work_grab_pending(work, 0, &irq_flags); 4528 4529 work_offqd_unpack(&offqd, *work_data_bits(work)); 4530 work_offqd_enable(&offqd); 4531 set_work_pool_and_clear_pending(work, offqd.pool_id, 4532 work_offqd_pack_flags(&offqd)); 4533 local_irq_restore(irq_flags); 4534 4535 return !offqd.disable; 4536 } 4537 EXPORT_SYMBOL_GPL(enable_work); 4538 4539 /** 4540 * disable_delayed_work - Disable and cancel a delayed work item 4541 * @dwork: delayed work item to disable 4542 * 4543 * disable_work() for delayed work items. 4544 */ 4545 bool disable_delayed_work(struct delayed_work *dwork) 4546 { 4547 return __cancel_work(&dwork->work, 4548 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4549 } 4550 EXPORT_SYMBOL_GPL(disable_delayed_work); 4551 4552 /** 4553 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item 4554 * @dwork: delayed work item to disable 4555 * 4556 * disable_work_sync() for delayed work items. 4557 */ 4558 bool disable_delayed_work_sync(struct delayed_work *dwork) 4559 { 4560 return __cancel_work_sync(&dwork->work, 4561 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4562 } 4563 EXPORT_SYMBOL_GPL(disable_delayed_work_sync); 4564 4565 /** 4566 * enable_delayed_work - Enable a delayed work item 4567 * @dwork: delayed work item to enable 4568 * 4569 * enable_work() for delayed work items. 4570 */ 4571 bool enable_delayed_work(struct delayed_work *dwork) 4572 { 4573 return enable_work(&dwork->work); 4574 } 4575 EXPORT_SYMBOL_GPL(enable_delayed_work); 4576 4577 /** 4578 * schedule_on_each_cpu - execute a function synchronously on each online CPU 4579 * @func: the function to call 4580 * 4581 * schedule_on_each_cpu() executes @func on each online CPU using the 4582 * system workqueue and blocks until all CPUs have completed. 4583 * schedule_on_each_cpu() is very slow. 4584 * 4585 * Return: 4586 * 0 on success, -errno on failure. 4587 */ 4588 int schedule_on_each_cpu(work_func_t func) 4589 { 4590 int cpu; 4591 struct work_struct __percpu *works; 4592 4593 works = alloc_percpu(struct work_struct); 4594 if (!works) 4595 return -ENOMEM; 4596 4597 cpus_read_lock(); 4598 4599 for_each_online_cpu(cpu) { 4600 struct work_struct *work = per_cpu_ptr(works, cpu); 4601 4602 INIT_WORK(work, func); 4603 schedule_work_on(cpu, work); 4604 } 4605 4606 for_each_online_cpu(cpu) 4607 flush_work(per_cpu_ptr(works, cpu)); 4608 4609 cpus_read_unlock(); 4610 free_percpu(works); 4611 return 0; 4612 } 4613 4614 /** 4615 * execute_in_process_context - reliably execute the routine with user context 4616 * @fn: the function to execute 4617 * @ew: guaranteed storage for the execute work structure (must 4618 * be available when the work executes) 4619 * 4620 * Executes the function immediately if process context is available, 4621 * otherwise schedules the function for delayed execution. 4622 * 4623 * Return: 0 - function was executed 4624 * 1 - function was scheduled for execution 4625 */ 4626 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 4627 { 4628 if (!in_interrupt()) { 4629 fn(&ew->work); 4630 return 0; 4631 } 4632 4633 INIT_WORK(&ew->work, fn); 4634 schedule_work(&ew->work); 4635 4636 return 1; 4637 } 4638 EXPORT_SYMBOL_GPL(execute_in_process_context); 4639 4640 /** 4641 * free_workqueue_attrs - free a workqueue_attrs 4642 * @attrs: workqueue_attrs to free 4643 * 4644 * Undo alloc_workqueue_attrs(). 4645 */ 4646 void free_workqueue_attrs(struct workqueue_attrs *attrs) 4647 { 4648 if (attrs) { 4649 free_cpumask_var(attrs->cpumask); 4650 free_cpumask_var(attrs->__pod_cpumask); 4651 kfree(attrs); 4652 } 4653 } 4654 4655 /** 4656 * alloc_workqueue_attrs - allocate a workqueue_attrs 4657 * 4658 * Allocate a new workqueue_attrs, initialize with default settings and 4659 * return it. 4660 * 4661 * Return: The allocated new workqueue_attr on success. %NULL on failure. 4662 */ 4663 struct workqueue_attrs *alloc_workqueue_attrs_noprof(void) 4664 { 4665 struct workqueue_attrs *attrs; 4666 4667 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 4668 if (!attrs) 4669 goto fail; 4670 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 4671 goto fail; 4672 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 4673 goto fail; 4674 4675 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4676 attrs->affn_scope = WQ_AFFN_DFL; 4677 return attrs; 4678 fail: 4679 free_workqueue_attrs(attrs); 4680 return NULL; 4681 } 4682 4683 static void copy_workqueue_attrs(struct workqueue_attrs *to, 4684 const struct workqueue_attrs *from) 4685 { 4686 to->nice = from->nice; 4687 cpumask_copy(to->cpumask, from->cpumask); 4688 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 4689 to->affn_strict = from->affn_strict; 4690 4691 /* 4692 * Unlike hash and equality test, copying shouldn't ignore wq-only 4693 * fields as copying is used for both pool and wq attrs. Instead, 4694 * get_unbound_pool() explicitly clears the fields. 4695 */ 4696 to->affn_scope = from->affn_scope; 4697 to->ordered = from->ordered; 4698 } 4699 4700 /* 4701 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 4702 * comments in 'struct workqueue_attrs' definition. 4703 */ 4704 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 4705 { 4706 attrs->affn_scope = WQ_AFFN_NR_TYPES; 4707 attrs->ordered = false; 4708 if (attrs->affn_strict) 4709 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4710 } 4711 4712 /* hash value of the content of @attr */ 4713 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 4714 { 4715 u32 hash = 0; 4716 4717 hash = jhash_1word(attrs->nice, hash); 4718 hash = jhash_1word(attrs->affn_strict, hash); 4719 hash = jhash(cpumask_bits(attrs->__pod_cpumask), 4720 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4721 if (!attrs->affn_strict) 4722 hash = jhash(cpumask_bits(attrs->cpumask), 4723 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4724 return hash; 4725 } 4726 4727 /* content equality test */ 4728 static bool wqattrs_equal(const struct workqueue_attrs *a, 4729 const struct workqueue_attrs *b) 4730 { 4731 if (a->nice != b->nice) 4732 return false; 4733 if (a->affn_strict != b->affn_strict) 4734 return false; 4735 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 4736 return false; 4737 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask)) 4738 return false; 4739 return true; 4740 } 4741 4742 /* Update @attrs with actually available CPUs */ 4743 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 4744 const cpumask_t *unbound_cpumask) 4745 { 4746 /* 4747 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 4748 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 4749 * @unbound_cpumask. 4750 */ 4751 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 4752 if (unlikely(cpumask_empty(attrs->cpumask))) 4753 cpumask_copy(attrs->cpumask, unbound_cpumask); 4754 } 4755 4756 /* find wq_pod_type to use for @attrs */ 4757 static const struct wq_pod_type * 4758 wqattrs_pod_type(const struct workqueue_attrs *attrs) 4759 { 4760 enum wq_affn_scope scope; 4761 struct wq_pod_type *pt; 4762 4763 /* to synchronize access to wq_affn_dfl */ 4764 lockdep_assert_held(&wq_pool_mutex); 4765 4766 if (attrs->affn_scope == WQ_AFFN_DFL) 4767 scope = wq_affn_dfl; 4768 else 4769 scope = attrs->affn_scope; 4770 4771 pt = &wq_pod_types[scope]; 4772 4773 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 4774 likely(pt->nr_pods)) 4775 return pt; 4776 4777 /* 4778 * Before workqueue_init_topology(), only SYSTEM is available which is 4779 * initialized in workqueue_init_early(). 4780 */ 4781 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 4782 BUG_ON(!pt->nr_pods); 4783 return pt; 4784 } 4785 4786 /** 4787 * init_worker_pool - initialize a newly zalloc'd worker_pool 4788 * @pool: worker_pool to initialize 4789 * 4790 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 4791 * 4792 * Return: 0 on success, -errno on failure. Even on failure, all fields 4793 * inside @pool proper are initialized and put_unbound_pool() can be called 4794 * on @pool safely to release it. 4795 */ 4796 static int init_worker_pool(struct worker_pool *pool) 4797 { 4798 raw_spin_lock_init(&pool->lock); 4799 pool->id = -1; 4800 pool->cpu = -1; 4801 pool->node = NUMA_NO_NODE; 4802 pool->flags |= POOL_DISASSOCIATED; 4803 pool->watchdog_ts = jiffies; 4804 INIT_LIST_HEAD(&pool->worklist); 4805 INIT_LIST_HEAD(&pool->idle_list); 4806 hash_init(pool->busy_hash); 4807 4808 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 4809 INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 4810 4811 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 4812 4813 INIT_LIST_HEAD(&pool->workers); 4814 4815 ida_init(&pool->worker_ida); 4816 INIT_HLIST_NODE(&pool->hash_node); 4817 pool->refcnt = 1; 4818 #ifdef CONFIG_PREEMPT_RT 4819 spin_lock_init(&pool->cb_lock); 4820 #endif 4821 4822 /* shouldn't fail above this point */ 4823 pool->attrs = alloc_workqueue_attrs(); 4824 if (!pool->attrs) 4825 return -ENOMEM; 4826 4827 wqattrs_clear_for_pool(pool->attrs); 4828 4829 return 0; 4830 } 4831 4832 #ifdef CONFIG_LOCKDEP 4833 static void wq_init_lockdep(struct workqueue_struct *wq) 4834 { 4835 char *lock_name; 4836 4837 lockdep_register_key(&wq->key); 4838 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 4839 if (!lock_name) 4840 lock_name = wq->name; 4841 4842 wq->lock_name = lock_name; 4843 wq->lockdep_map = &wq->__lockdep_map; 4844 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0); 4845 } 4846 4847 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4848 { 4849 if (wq->lockdep_map != &wq->__lockdep_map) 4850 return; 4851 4852 lockdep_unregister_key(&wq->key); 4853 } 4854 4855 static void wq_free_lockdep(struct workqueue_struct *wq) 4856 { 4857 if (wq->lockdep_map != &wq->__lockdep_map) 4858 return; 4859 4860 if (wq->lock_name != wq->name) 4861 kfree(wq->lock_name); 4862 } 4863 #else 4864 static void wq_init_lockdep(struct workqueue_struct *wq) 4865 { 4866 } 4867 4868 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4869 { 4870 } 4871 4872 static void wq_free_lockdep(struct workqueue_struct *wq) 4873 { 4874 } 4875 #endif 4876 4877 static void free_node_nr_active(struct wq_node_nr_active **nna_ar) 4878 { 4879 int node; 4880 4881 for_each_node(node) { 4882 kfree(nna_ar[node]); 4883 nna_ar[node] = NULL; 4884 } 4885 4886 kfree(nna_ar[nr_node_ids]); 4887 nna_ar[nr_node_ids] = NULL; 4888 } 4889 4890 static void init_node_nr_active(struct wq_node_nr_active *nna) 4891 { 4892 nna->max = WQ_DFL_MIN_ACTIVE; 4893 atomic_set(&nna->nr, 0); 4894 raw_spin_lock_init(&nna->lock); 4895 INIT_LIST_HEAD(&nna->pending_pwqs); 4896 } 4897 4898 /* 4899 * Each node's nr_active counter will be accessed mostly from its own node and 4900 * should be allocated in the node. 4901 */ 4902 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) 4903 { 4904 struct wq_node_nr_active *nna; 4905 int node; 4906 4907 for_each_node(node) { 4908 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); 4909 if (!nna) 4910 goto err_free; 4911 init_node_nr_active(nna); 4912 nna_ar[node] = nna; 4913 } 4914 4915 /* [nr_node_ids] is used as the fallback */ 4916 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); 4917 if (!nna) 4918 goto err_free; 4919 init_node_nr_active(nna); 4920 nna_ar[nr_node_ids] = nna; 4921 4922 return 0; 4923 4924 err_free: 4925 free_node_nr_active(nna_ar); 4926 return -ENOMEM; 4927 } 4928 4929 static void rcu_free_wq(struct rcu_head *rcu) 4930 { 4931 struct workqueue_struct *wq = 4932 container_of(rcu, struct workqueue_struct, rcu); 4933 4934 if (wq->flags & WQ_UNBOUND) 4935 free_node_nr_active(wq->node_nr_active); 4936 4937 wq_free_lockdep(wq); 4938 free_percpu(wq->cpu_pwq); 4939 free_workqueue_attrs(wq->unbound_attrs); 4940 kfree(wq); 4941 } 4942 4943 static void rcu_free_pool(struct rcu_head *rcu) 4944 { 4945 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 4946 4947 ida_destroy(&pool->worker_ida); 4948 free_workqueue_attrs(pool->attrs); 4949 kfree(pool); 4950 } 4951 4952 /** 4953 * put_unbound_pool - put a worker_pool 4954 * @pool: worker_pool to put 4955 * 4956 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 4957 * safe manner. get_unbound_pool() calls this function on its failure path 4958 * and this function should be able to release pools which went through, 4959 * successfully or not, init_worker_pool(). 4960 * 4961 * Should be called with wq_pool_mutex held. 4962 */ 4963 static void put_unbound_pool(struct worker_pool *pool) 4964 { 4965 struct worker *worker; 4966 LIST_HEAD(cull_list); 4967 4968 lockdep_assert_held(&wq_pool_mutex); 4969 4970 if (--pool->refcnt) 4971 return; 4972 4973 /* sanity checks */ 4974 if (WARN_ON(!(pool->cpu < 0)) || 4975 WARN_ON(!list_empty(&pool->worklist))) 4976 return; 4977 4978 /* release id and unhash */ 4979 if (pool->id >= 0) 4980 idr_remove(&worker_pool_idr, pool->id); 4981 hash_del(&pool->hash_node); 4982 4983 /* 4984 * Become the manager and destroy all workers. This prevents 4985 * @pool's workers from blocking on attach_mutex. We're the last 4986 * manager and @pool gets freed with the flag set. 4987 * 4988 * Having a concurrent manager is quite unlikely to happen as we can 4989 * only get here with 4990 * pwq->refcnt == pool->refcnt == 0 4991 * which implies no work queued to the pool, which implies no worker can 4992 * become the manager. However a worker could have taken the role of 4993 * manager before the refcnts dropped to 0, since maybe_create_worker() 4994 * drops pool->lock 4995 */ 4996 while (true) { 4997 rcuwait_wait_event(&manager_wait, 4998 !(pool->flags & POOL_MANAGER_ACTIVE), 4999 TASK_UNINTERRUPTIBLE); 5000 5001 mutex_lock(&wq_pool_attach_mutex); 5002 raw_spin_lock_irq(&pool->lock); 5003 if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 5004 pool->flags |= POOL_MANAGER_ACTIVE; 5005 break; 5006 } 5007 raw_spin_unlock_irq(&pool->lock); 5008 mutex_unlock(&wq_pool_attach_mutex); 5009 } 5010 5011 while ((worker = first_idle_worker(pool))) 5012 set_worker_dying(worker, &cull_list); 5013 WARN_ON(pool->nr_workers || pool->nr_idle); 5014 raw_spin_unlock_irq(&pool->lock); 5015 5016 detach_dying_workers(&cull_list); 5017 5018 mutex_unlock(&wq_pool_attach_mutex); 5019 5020 reap_dying_workers(&cull_list); 5021 5022 /* shut down the timers */ 5023 timer_delete_sync(&pool->idle_timer); 5024 cancel_work_sync(&pool->idle_cull_work); 5025 timer_delete_sync(&pool->mayday_timer); 5026 5027 /* RCU protected to allow dereferences from get_work_pool() */ 5028 call_rcu(&pool->rcu, rcu_free_pool); 5029 } 5030 5031 /** 5032 * get_unbound_pool - get a worker_pool with the specified attributes 5033 * @attrs: the attributes of the worker_pool to get 5034 * 5035 * Obtain a worker_pool which has the same attributes as @attrs, bump the 5036 * reference count and return it. If there already is a matching 5037 * worker_pool, it will be used; otherwise, this function attempts to 5038 * create a new one. 5039 * 5040 * Should be called with wq_pool_mutex held. 5041 * 5042 * Return: On success, a worker_pool with the same attributes as @attrs. 5043 * On failure, %NULL. 5044 */ 5045 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 5046 { 5047 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 5048 u32 hash = wqattrs_hash(attrs); 5049 struct worker_pool *pool; 5050 int pod, node = NUMA_NO_NODE; 5051 5052 lockdep_assert_held(&wq_pool_mutex); 5053 5054 /* do we already have a matching pool? */ 5055 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 5056 if (wqattrs_equal(pool->attrs, attrs)) { 5057 pool->refcnt++; 5058 return pool; 5059 } 5060 } 5061 5062 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 5063 for (pod = 0; pod < pt->nr_pods; pod++) { 5064 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 5065 node = pt->pod_node[pod]; 5066 break; 5067 } 5068 } 5069 5070 /* nope, create a new one */ 5071 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 5072 if (!pool || init_worker_pool(pool) < 0) 5073 goto fail; 5074 5075 pool->node = node; 5076 copy_workqueue_attrs(pool->attrs, attrs); 5077 wqattrs_clear_for_pool(pool->attrs); 5078 5079 if (worker_pool_assign_id(pool) < 0) 5080 goto fail; 5081 5082 /* create and start the initial worker */ 5083 if (wq_online && !create_worker(pool)) 5084 goto fail; 5085 5086 /* install */ 5087 hash_add(unbound_pool_hash, &pool->hash_node, hash); 5088 5089 return pool; 5090 fail: 5091 if (pool) 5092 put_unbound_pool(pool); 5093 return NULL; 5094 } 5095 5096 /* 5097 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 5098 * refcnt and needs to be destroyed. 5099 */ 5100 static void pwq_release_workfn(struct kthread_work *work) 5101 { 5102 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 5103 release_work); 5104 struct workqueue_struct *wq = pwq->wq; 5105 struct worker_pool *pool = pwq->pool; 5106 bool is_last = false; 5107 5108 /* 5109 * When @pwq is not linked, it doesn't hold any reference to the 5110 * @wq, and @wq is invalid to access. 5111 */ 5112 if (!list_empty(&pwq->pwqs_node)) { 5113 mutex_lock(&wq->mutex); 5114 list_del_rcu(&pwq->pwqs_node); 5115 is_last = list_empty(&wq->pwqs); 5116 5117 /* 5118 * For ordered workqueue with a plugged dfl_pwq, restart it now. 5119 */ 5120 if (!is_last && (wq->flags & __WQ_ORDERED)) 5121 unplug_oldest_pwq(wq); 5122 5123 mutex_unlock(&wq->mutex); 5124 } 5125 5126 if (wq->flags & WQ_UNBOUND) { 5127 mutex_lock(&wq_pool_mutex); 5128 put_unbound_pool(pool); 5129 mutex_unlock(&wq_pool_mutex); 5130 } 5131 5132 if (!list_empty(&pwq->pending_node)) { 5133 struct wq_node_nr_active *nna = 5134 wq_node_nr_active(pwq->wq, pwq->pool->node); 5135 5136 raw_spin_lock_irq(&nna->lock); 5137 list_del_init(&pwq->pending_node); 5138 raw_spin_unlock_irq(&nna->lock); 5139 } 5140 5141 kfree_rcu(pwq, rcu); 5142 5143 /* 5144 * If we're the last pwq going away, @wq is already dead and no one 5145 * is gonna access it anymore. Schedule RCU free. 5146 */ 5147 if (is_last) { 5148 wq_unregister_lockdep(wq); 5149 call_rcu(&wq->rcu, rcu_free_wq); 5150 } 5151 } 5152 5153 /* initialize newly allocated @pwq which is associated with @wq and @pool */ 5154 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 5155 struct worker_pool *pool) 5156 { 5157 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK); 5158 5159 memset(pwq, 0, sizeof(*pwq)); 5160 5161 pwq->pool = pool; 5162 pwq->wq = wq; 5163 pwq->flush_color = -1; 5164 pwq->refcnt = 1; 5165 INIT_LIST_HEAD(&pwq->inactive_works); 5166 INIT_LIST_HEAD(&pwq->pending_node); 5167 INIT_LIST_HEAD(&pwq->pwqs_node); 5168 INIT_LIST_HEAD(&pwq->mayday_node); 5169 kthread_init_work(&pwq->release_work, pwq_release_workfn); 5170 } 5171 5172 /* sync @pwq with the current state of its associated wq and link it */ 5173 static void link_pwq(struct pool_workqueue *pwq) 5174 { 5175 struct workqueue_struct *wq = pwq->wq; 5176 5177 lockdep_assert_held(&wq->mutex); 5178 5179 /* may be called multiple times, ignore if already linked */ 5180 if (!list_empty(&pwq->pwqs_node)) 5181 return; 5182 5183 /* set the matching work_color */ 5184 pwq->work_color = wq->work_color; 5185 5186 /* link in @pwq */ 5187 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); 5188 } 5189 5190 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 5191 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 5192 const struct workqueue_attrs *attrs) 5193 { 5194 struct worker_pool *pool; 5195 struct pool_workqueue *pwq; 5196 5197 lockdep_assert_held(&wq_pool_mutex); 5198 5199 pool = get_unbound_pool(attrs); 5200 if (!pool) 5201 return NULL; 5202 5203 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 5204 if (!pwq) { 5205 put_unbound_pool(pool); 5206 return NULL; 5207 } 5208 5209 init_pwq(pwq, wq, pool); 5210 return pwq; 5211 } 5212 5213 static void apply_wqattrs_lock(void) 5214 { 5215 mutex_lock(&wq_pool_mutex); 5216 } 5217 5218 static void apply_wqattrs_unlock(void) 5219 { 5220 mutex_unlock(&wq_pool_mutex); 5221 } 5222 5223 /** 5224 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 5225 * @attrs: the wq_attrs of the default pwq of the target workqueue 5226 * @cpu: the target CPU 5227 * 5228 * Calculate the cpumask a workqueue with @attrs should use on @pod. 5229 * The result is stored in @attrs->__pod_cpumask. 5230 * 5231 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 5232 * and @pod has online CPUs requested by @attrs, the returned cpumask is the 5233 * intersection of the possible CPUs of @pod and @attrs->cpumask. 5234 * 5235 * The caller is responsible for ensuring that the cpumask of @pod stays stable. 5236 */ 5237 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu) 5238 { 5239 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5240 int pod = pt->cpu_pod[cpu]; 5241 5242 /* calculate possible CPUs in @pod that @attrs wants */ 5243 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 5244 /* does @pod have any online CPUs @attrs wants? */ 5245 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) { 5246 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 5247 return; 5248 } 5249 } 5250 5251 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ 5252 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 5253 int cpu, struct pool_workqueue *pwq) 5254 { 5255 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); 5256 struct pool_workqueue *old_pwq; 5257 5258 lockdep_assert_held(&wq_pool_mutex); 5259 lockdep_assert_held(&wq->mutex); 5260 5261 /* link_pwq() can handle duplicate calls */ 5262 link_pwq(pwq); 5263 5264 old_pwq = rcu_access_pointer(*slot); 5265 rcu_assign_pointer(*slot, pwq); 5266 return old_pwq; 5267 } 5268 5269 /* context to store the prepared attrs & pwqs before applying */ 5270 struct apply_wqattrs_ctx { 5271 struct workqueue_struct *wq; /* target workqueue */ 5272 struct workqueue_attrs *attrs; /* attrs to apply */ 5273 struct list_head list; /* queued for batching commit */ 5274 struct pool_workqueue *dfl_pwq; 5275 struct pool_workqueue *pwq_tbl[]; 5276 }; 5277 5278 /* free the resources after success or abort */ 5279 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 5280 { 5281 if (ctx) { 5282 int cpu; 5283 5284 for_each_possible_cpu(cpu) 5285 put_pwq_unlocked(ctx->pwq_tbl[cpu]); 5286 put_pwq_unlocked(ctx->dfl_pwq); 5287 5288 free_workqueue_attrs(ctx->attrs); 5289 5290 kfree(ctx); 5291 } 5292 } 5293 5294 /* allocate the attrs and pwqs for later installation */ 5295 static struct apply_wqattrs_ctx * 5296 apply_wqattrs_prepare(struct workqueue_struct *wq, 5297 const struct workqueue_attrs *attrs, 5298 const cpumask_var_t unbound_cpumask) 5299 { 5300 struct apply_wqattrs_ctx *ctx; 5301 struct workqueue_attrs *new_attrs; 5302 int cpu; 5303 5304 lockdep_assert_held(&wq_pool_mutex); 5305 5306 if (WARN_ON(attrs->affn_scope < 0 || 5307 attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 5308 return ERR_PTR(-EINVAL); 5309 5310 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 5311 5312 new_attrs = alloc_workqueue_attrs(); 5313 if (!ctx || !new_attrs) 5314 goto out_free; 5315 5316 /* 5317 * If something goes wrong during CPU up/down, we'll fall back to 5318 * the default pwq covering whole @attrs->cpumask. Always create 5319 * it even if we don't use it immediately. 5320 */ 5321 copy_workqueue_attrs(new_attrs, attrs); 5322 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 5323 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5324 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 5325 if (!ctx->dfl_pwq) 5326 goto out_free; 5327 5328 for_each_possible_cpu(cpu) { 5329 if (new_attrs->ordered) { 5330 ctx->dfl_pwq->refcnt++; 5331 ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 5332 } else { 5333 wq_calc_pod_cpumask(new_attrs, cpu); 5334 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 5335 if (!ctx->pwq_tbl[cpu]) 5336 goto out_free; 5337 } 5338 } 5339 5340 /* save the user configured attrs and sanitize it. */ 5341 copy_workqueue_attrs(new_attrs, attrs); 5342 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 5343 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5344 ctx->attrs = new_attrs; 5345 5346 /* 5347 * For initialized ordered workqueues, there should only be one pwq 5348 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution 5349 * of newly queued work items until execution of older work items in 5350 * the old pwq's have completed. 5351 */ 5352 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)) 5353 ctx->dfl_pwq->plugged = true; 5354 5355 ctx->wq = wq; 5356 return ctx; 5357 5358 out_free: 5359 free_workqueue_attrs(new_attrs); 5360 apply_wqattrs_cleanup(ctx); 5361 return ERR_PTR(-ENOMEM); 5362 } 5363 5364 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 5365 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 5366 { 5367 int cpu; 5368 5369 /* all pwqs have been created successfully, let's install'em */ 5370 mutex_lock(&ctx->wq->mutex); 5371 5372 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 5373 5374 /* save the previous pwqs and install the new ones */ 5375 for_each_possible_cpu(cpu) 5376 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 5377 ctx->pwq_tbl[cpu]); 5378 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); 5379 5380 /* update node_nr_active->max */ 5381 wq_update_node_max_active(ctx->wq, -1); 5382 5383 mutex_unlock(&ctx->wq->mutex); 5384 } 5385 5386 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 5387 const struct workqueue_attrs *attrs) 5388 { 5389 struct apply_wqattrs_ctx *ctx; 5390 5391 /* only unbound workqueues can change attributes */ 5392 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 5393 return -EINVAL; 5394 5395 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 5396 if (IS_ERR(ctx)) 5397 return PTR_ERR(ctx); 5398 5399 /* the ctx has been prepared successfully, let's commit it */ 5400 apply_wqattrs_commit(ctx); 5401 apply_wqattrs_cleanup(ctx); 5402 5403 return 0; 5404 } 5405 5406 /** 5407 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 5408 * @wq: the target workqueue 5409 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 5410 * 5411 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 5412 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 5413 * work items are affine to the pod it was issued on. Older pwqs are released as 5414 * in-flight work items finish. Note that a work item which repeatedly requeues 5415 * itself back-to-back will stay on its current pwq. 5416 * 5417 * Performs GFP_KERNEL allocations. 5418 * 5419 * Return: 0 on success and -errno on failure. 5420 */ 5421 int apply_workqueue_attrs(struct workqueue_struct *wq, 5422 const struct workqueue_attrs *attrs) 5423 { 5424 int ret; 5425 5426 mutex_lock(&wq_pool_mutex); 5427 ret = apply_workqueue_attrs_locked(wq, attrs); 5428 mutex_unlock(&wq_pool_mutex); 5429 5430 return ret; 5431 } 5432 5433 /** 5434 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug 5435 * @wq: the target workqueue 5436 * @cpu: the CPU to update the pwq slot for 5437 * 5438 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 5439 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged. 5440 * 5441 * 5442 * If pod affinity can't be adjusted due to memory allocation failure, it falls 5443 * back to @wq->dfl_pwq which may not be optimal but is always correct. 5444 * 5445 * Note that when the last allowed CPU of a pod goes offline for a workqueue 5446 * with a cpumask spanning multiple pods, the workers which were already 5447 * executing the work items for the workqueue will lose their CPU affinity and 5448 * may execute on any CPU. This is similar to how per-cpu workqueues behave on 5449 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 5450 * responsibility to flush the work item from CPU_DOWN_PREPARE. 5451 */ 5452 static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu) 5453 { 5454 struct pool_workqueue *old_pwq = NULL, *pwq; 5455 struct workqueue_attrs *target_attrs; 5456 5457 lockdep_assert_held(&wq_pool_mutex); 5458 5459 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 5460 return; 5461 5462 /* 5463 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 5464 * Let's use a preallocated one. The following buf is protected by 5465 * CPU hotplug exclusion. 5466 */ 5467 target_attrs = unbound_wq_update_pwq_attrs_buf; 5468 5469 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 5470 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 5471 5472 /* nothing to do if the target cpumask matches the current pwq */ 5473 wq_calc_pod_cpumask(target_attrs, cpu); 5474 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) 5475 return; 5476 5477 /* create a new pwq */ 5478 pwq = alloc_unbound_pwq(wq, target_attrs); 5479 if (!pwq) { 5480 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 5481 wq->name); 5482 goto use_dfl_pwq; 5483 } 5484 5485 /* Install the new pwq. */ 5486 mutex_lock(&wq->mutex); 5487 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5488 goto out_unlock; 5489 5490 use_dfl_pwq: 5491 mutex_lock(&wq->mutex); 5492 pwq = unbound_pwq(wq, -1); 5493 raw_spin_lock_irq(&pwq->pool->lock); 5494 get_pwq(pwq); 5495 raw_spin_unlock_irq(&pwq->pool->lock); 5496 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5497 out_unlock: 5498 mutex_unlock(&wq->mutex); 5499 put_pwq_unlocked(old_pwq); 5500 } 5501 5502 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 5503 { 5504 bool highpri = wq->flags & WQ_HIGHPRI; 5505 int cpu, ret; 5506 5507 lockdep_assert_held(&wq_pool_mutex); 5508 5509 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 5510 if (!wq->cpu_pwq) 5511 goto enomem; 5512 5513 if (!(wq->flags & WQ_UNBOUND)) { 5514 struct worker_pool __percpu *pools; 5515 5516 if (wq->flags & WQ_BH) 5517 pools = bh_worker_pools; 5518 else 5519 pools = cpu_worker_pools; 5520 5521 for_each_possible_cpu(cpu) { 5522 struct pool_workqueue **pwq_p; 5523 struct worker_pool *pool; 5524 5525 pool = &(per_cpu_ptr(pools, cpu)[highpri]); 5526 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu); 5527 5528 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 5529 pool->node); 5530 if (!*pwq_p) 5531 goto enomem; 5532 5533 init_pwq(*pwq_p, wq, pool); 5534 5535 mutex_lock(&wq->mutex); 5536 link_pwq(*pwq_p); 5537 mutex_unlock(&wq->mutex); 5538 } 5539 return 0; 5540 } 5541 5542 if (wq->flags & __WQ_ORDERED) { 5543 struct pool_workqueue *dfl_pwq; 5544 5545 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]); 5546 /* there should only be single pwq for ordering guarantee */ 5547 dfl_pwq = rcu_access_pointer(wq->dfl_pwq); 5548 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || 5549 wq->pwqs.prev != &dfl_pwq->pwqs_node), 5550 "ordering guarantee broken for workqueue %s\n", wq->name); 5551 } else { 5552 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]); 5553 } 5554 5555 return ret; 5556 5557 enomem: 5558 if (wq->cpu_pwq) { 5559 for_each_possible_cpu(cpu) { 5560 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 5561 5562 if (pwq) 5563 kmem_cache_free(pwq_cache, pwq); 5564 } 5565 free_percpu(wq->cpu_pwq); 5566 wq->cpu_pwq = NULL; 5567 } 5568 return -ENOMEM; 5569 } 5570 5571 static int wq_clamp_max_active(int max_active, unsigned int flags, 5572 const char *name) 5573 { 5574 if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 5575 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 5576 max_active, name, 1, WQ_MAX_ACTIVE); 5577 5578 return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 5579 } 5580 5581 /* 5582 * Workqueues which may be used during memory reclaim should have a rescuer 5583 * to guarantee forward progress. 5584 */ 5585 static int init_rescuer(struct workqueue_struct *wq) 5586 { 5587 struct worker *rescuer; 5588 char id_buf[WORKER_ID_LEN]; 5589 int ret; 5590 5591 lockdep_assert_held(&wq_pool_mutex); 5592 5593 if (!(wq->flags & WQ_MEM_RECLAIM)) 5594 return 0; 5595 5596 rescuer = alloc_worker(NUMA_NO_NODE); 5597 if (!rescuer) { 5598 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 5599 wq->name); 5600 return -ENOMEM; 5601 } 5602 5603 rescuer->rescue_wq = wq; 5604 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL); 5605 5606 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf); 5607 if (IS_ERR(rescuer->task)) { 5608 ret = PTR_ERR(rescuer->task); 5609 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 5610 wq->name, ERR_PTR(ret)); 5611 kfree(rescuer); 5612 return ret; 5613 } 5614 5615 wq->rescuer = rescuer; 5616 5617 /* initial cpumask is consistent with the detached rescuer and unbind_worker() */ 5618 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 5619 kthread_bind_mask(rescuer->task, wq_unbound_cpumask); 5620 else 5621 kthread_bind_mask(rescuer->task, cpu_possible_mask); 5622 5623 wake_up_process(rescuer->task); 5624 5625 return 0; 5626 } 5627 5628 /** 5629 * wq_adjust_max_active - update a wq's max_active to the current setting 5630 * @wq: target workqueue 5631 * 5632 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and 5633 * activate inactive work items accordingly. If @wq is freezing, clear 5634 * @wq->max_active to zero. 5635 */ 5636 static void wq_adjust_max_active(struct workqueue_struct *wq) 5637 { 5638 bool activated; 5639 int new_max, new_min; 5640 5641 lockdep_assert_held(&wq->mutex); 5642 5643 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { 5644 new_max = 0; 5645 new_min = 0; 5646 } else { 5647 new_max = wq->saved_max_active; 5648 new_min = wq->saved_min_active; 5649 } 5650 5651 if (wq->max_active == new_max && wq->min_active == new_min) 5652 return; 5653 5654 /* 5655 * Update @wq->max/min_active and then kick inactive work items if more 5656 * active work items are allowed. This doesn't break work item ordering 5657 * because new work items are always queued behind existing inactive 5658 * work items if there are any. 5659 */ 5660 WRITE_ONCE(wq->max_active, new_max); 5661 WRITE_ONCE(wq->min_active, new_min); 5662 5663 if (wq->flags & WQ_UNBOUND) 5664 wq_update_node_max_active(wq, -1); 5665 5666 if (new_max == 0) 5667 return; 5668 5669 /* 5670 * Round-robin through pwq's activating the first inactive work item 5671 * until max_active is filled. 5672 */ 5673 do { 5674 struct pool_workqueue *pwq; 5675 5676 activated = false; 5677 for_each_pwq(pwq, wq) { 5678 unsigned long irq_flags; 5679 5680 /* can be called during early boot w/ irq disabled */ 5681 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 5682 if (pwq_activate_first_inactive(pwq, true)) { 5683 activated = true; 5684 kick_pool(pwq->pool); 5685 } 5686 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 5687 } 5688 } while (activated); 5689 } 5690 5691 __printf(1, 0) 5692 static struct workqueue_struct *__alloc_workqueue(const char *fmt, 5693 unsigned int flags, 5694 int max_active, va_list args) 5695 { 5696 struct workqueue_struct *wq; 5697 size_t wq_size; 5698 int name_len; 5699 5700 if (flags & WQ_BH) { 5701 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS)) 5702 return NULL; 5703 if (WARN_ON_ONCE(max_active)) 5704 return NULL; 5705 } 5706 5707 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 5708 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 5709 flags |= WQ_UNBOUND; 5710 5711 /* allocate wq and format name */ 5712 if (flags & WQ_UNBOUND) 5713 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); 5714 else 5715 wq_size = sizeof(*wq); 5716 5717 wq = kzalloc_noprof(wq_size, GFP_KERNEL); 5718 if (!wq) 5719 return NULL; 5720 5721 if (flags & WQ_UNBOUND) { 5722 wq->unbound_attrs = alloc_workqueue_attrs_noprof(); 5723 if (!wq->unbound_attrs) 5724 goto err_free_wq; 5725 } 5726 5727 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); 5728 5729 if (name_len >= WQ_NAME_LEN) 5730 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", 5731 wq->name); 5732 5733 if (flags & WQ_BH) { 5734 /* 5735 * BH workqueues always share a single execution context per CPU 5736 * and don't impose any max_active limit. 5737 */ 5738 max_active = INT_MAX; 5739 } else { 5740 max_active = max_active ?: WQ_DFL_ACTIVE; 5741 max_active = wq_clamp_max_active(max_active, flags, wq->name); 5742 } 5743 5744 /* init wq */ 5745 wq->flags = flags; 5746 wq->max_active = max_active; 5747 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); 5748 wq->saved_max_active = wq->max_active; 5749 wq->saved_min_active = wq->min_active; 5750 mutex_init(&wq->mutex); 5751 atomic_set(&wq->nr_pwqs_to_flush, 0); 5752 INIT_LIST_HEAD(&wq->pwqs); 5753 INIT_LIST_HEAD(&wq->flusher_queue); 5754 INIT_LIST_HEAD(&wq->flusher_overflow); 5755 INIT_LIST_HEAD(&wq->maydays); 5756 5757 INIT_LIST_HEAD(&wq->list); 5758 5759 if (flags & WQ_UNBOUND) { 5760 if (alloc_node_nr_active(wq->node_nr_active) < 0) 5761 goto err_free_wq; 5762 } 5763 5764 /* 5765 * wq_pool_mutex protects the workqueues list, allocations of PWQs, 5766 * and the global freeze state. 5767 */ 5768 apply_wqattrs_lock(); 5769 5770 if (alloc_and_link_pwqs(wq) < 0) 5771 goto err_unlock_free_node_nr_active; 5772 5773 mutex_lock(&wq->mutex); 5774 wq_adjust_max_active(wq); 5775 mutex_unlock(&wq->mutex); 5776 5777 list_add_tail_rcu(&wq->list, &workqueues); 5778 5779 if (wq_online && init_rescuer(wq) < 0) 5780 goto err_unlock_destroy; 5781 5782 apply_wqattrs_unlock(); 5783 5784 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 5785 goto err_destroy; 5786 5787 return wq; 5788 5789 err_unlock_free_node_nr_active: 5790 apply_wqattrs_unlock(); 5791 /* 5792 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work, 5793 * flushing the pwq_release_worker ensures that the pwq_release_workfn() 5794 * completes before calling kfree(wq). 5795 */ 5796 if (wq->flags & WQ_UNBOUND) { 5797 kthread_flush_worker(pwq_release_worker); 5798 free_node_nr_active(wq->node_nr_active); 5799 } 5800 err_free_wq: 5801 free_workqueue_attrs(wq->unbound_attrs); 5802 kfree(wq); 5803 return NULL; 5804 err_unlock_destroy: 5805 apply_wqattrs_unlock(); 5806 err_destroy: 5807 destroy_workqueue(wq); 5808 return NULL; 5809 } 5810 5811 __printf(1, 4) 5812 struct workqueue_struct *alloc_workqueue_noprof(const char *fmt, 5813 unsigned int flags, 5814 int max_active, ...) 5815 { 5816 struct workqueue_struct *wq; 5817 va_list args; 5818 5819 va_start(args, max_active); 5820 wq = __alloc_workqueue(fmt, flags, max_active, args); 5821 va_end(args); 5822 if (!wq) 5823 return NULL; 5824 5825 wq_init_lockdep(wq); 5826 5827 return wq; 5828 } 5829 EXPORT_SYMBOL_GPL(alloc_workqueue_noprof); 5830 5831 #ifdef CONFIG_LOCKDEP 5832 __printf(1, 5) 5833 struct workqueue_struct * 5834 alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags, 5835 int max_active, struct lockdep_map *lockdep_map, ...) 5836 { 5837 struct workqueue_struct *wq; 5838 va_list args; 5839 5840 va_start(args, lockdep_map); 5841 wq = __alloc_workqueue(fmt, flags, max_active, args); 5842 va_end(args); 5843 if (!wq) 5844 return NULL; 5845 5846 wq->lockdep_map = lockdep_map; 5847 5848 return wq; 5849 } 5850 EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map); 5851 #endif 5852 5853 static bool pwq_busy(struct pool_workqueue *pwq) 5854 { 5855 int i; 5856 5857 for (i = 0; i < WORK_NR_COLORS; i++) 5858 if (pwq->nr_in_flight[i]) 5859 return true; 5860 5861 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) 5862 return true; 5863 if (!pwq_is_empty(pwq)) 5864 return true; 5865 5866 return false; 5867 } 5868 5869 /** 5870 * destroy_workqueue - safely terminate a workqueue 5871 * @wq: target workqueue 5872 * 5873 * Safely destroy a workqueue. All work currently pending will be done first. 5874 * 5875 * This function does NOT guarantee that non-pending work that has been 5876 * submitted with queue_delayed_work() and similar functions will be done 5877 * before destroying the workqueue. The fundamental problem is that, currently, 5878 * the workqueue has no way of accessing non-pending delayed_work. delayed_work 5879 * is only linked on the timer-side. All delayed_work must, therefore, be 5880 * canceled before calling this function. 5881 * 5882 * TODO: It would be better if the problem described above wouldn't exist and 5883 * destroy_workqueue() would cleanly cancel all pending and non-pending 5884 * delayed_work. 5885 */ 5886 void destroy_workqueue(struct workqueue_struct *wq) 5887 { 5888 struct pool_workqueue *pwq; 5889 int cpu; 5890 5891 /* 5892 * Remove it from sysfs first so that sanity check failure doesn't 5893 * lead to sysfs name conflicts. 5894 */ 5895 workqueue_sysfs_unregister(wq); 5896 5897 /* mark the workqueue destruction is in progress */ 5898 mutex_lock(&wq->mutex); 5899 wq->flags |= __WQ_DESTROYING; 5900 mutex_unlock(&wq->mutex); 5901 5902 /* drain it before proceeding with destruction */ 5903 drain_workqueue(wq); 5904 5905 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 5906 if (wq->rescuer) { 5907 /* rescuer will empty maydays list before exiting */ 5908 kthread_stop(wq->rescuer->task); 5909 kfree(wq->rescuer); 5910 wq->rescuer = NULL; 5911 } 5912 5913 /* 5914 * Sanity checks - grab all the locks so that we wait for all 5915 * in-flight operations which may do put_pwq(). 5916 */ 5917 mutex_lock(&wq_pool_mutex); 5918 mutex_lock(&wq->mutex); 5919 for_each_pwq(pwq, wq) { 5920 raw_spin_lock_irq(&pwq->pool->lock); 5921 if (WARN_ON(pwq_busy(pwq))) { 5922 pr_warn("%s: %s has the following busy pwq\n", 5923 __func__, wq->name); 5924 show_pwq(pwq); 5925 raw_spin_unlock_irq(&pwq->pool->lock); 5926 mutex_unlock(&wq->mutex); 5927 mutex_unlock(&wq_pool_mutex); 5928 show_one_workqueue(wq); 5929 return; 5930 } 5931 raw_spin_unlock_irq(&pwq->pool->lock); 5932 } 5933 mutex_unlock(&wq->mutex); 5934 5935 /* 5936 * wq list is used to freeze wq, remove from list after 5937 * flushing is complete in case freeze races us. 5938 */ 5939 list_del_rcu(&wq->list); 5940 mutex_unlock(&wq_pool_mutex); 5941 5942 /* 5943 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 5944 * to put the base refs. @wq will be auto-destroyed from the last 5945 * pwq_put. RCU read lock prevents @wq from going away from under us. 5946 */ 5947 rcu_read_lock(); 5948 5949 for_each_possible_cpu(cpu) { 5950 put_pwq_unlocked(unbound_pwq(wq, cpu)); 5951 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); 5952 } 5953 5954 put_pwq_unlocked(unbound_pwq(wq, -1)); 5955 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); 5956 5957 rcu_read_unlock(); 5958 } 5959 EXPORT_SYMBOL_GPL(destroy_workqueue); 5960 5961 /** 5962 * workqueue_set_max_active - adjust max_active of a workqueue 5963 * @wq: target workqueue 5964 * @max_active: new max_active value. 5965 * 5966 * Set max_active of @wq to @max_active. See the alloc_workqueue() function 5967 * comment. 5968 * 5969 * CONTEXT: 5970 * Don't call from IRQ context. 5971 */ 5972 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 5973 { 5974 /* max_active doesn't mean anything for BH workqueues */ 5975 if (WARN_ON(wq->flags & WQ_BH)) 5976 return; 5977 /* disallow meddling with max_active for ordered workqueues */ 5978 if (WARN_ON(wq->flags & __WQ_ORDERED)) 5979 return; 5980 5981 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 5982 5983 mutex_lock(&wq->mutex); 5984 5985 wq->saved_max_active = max_active; 5986 if (wq->flags & WQ_UNBOUND) 5987 wq->saved_min_active = min(wq->saved_min_active, max_active); 5988 5989 wq_adjust_max_active(wq); 5990 5991 mutex_unlock(&wq->mutex); 5992 } 5993 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 5994 5995 /** 5996 * workqueue_set_min_active - adjust min_active of an unbound workqueue 5997 * @wq: target unbound workqueue 5998 * @min_active: new min_active value 5999 * 6000 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an 6001 * unbound workqueue is not guaranteed to be able to process max_active 6002 * interdependent work items. Instead, an unbound workqueue is guaranteed to be 6003 * able to process min_active number of interdependent work items which is 6004 * %WQ_DFL_MIN_ACTIVE by default. 6005 * 6006 * Use this function to adjust the min_active value between 0 and the current 6007 * max_active. 6008 */ 6009 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active) 6010 { 6011 /* min_active is only meaningful for non-ordered unbound workqueues */ 6012 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) != 6013 WQ_UNBOUND)) 6014 return; 6015 6016 mutex_lock(&wq->mutex); 6017 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active); 6018 wq_adjust_max_active(wq); 6019 mutex_unlock(&wq->mutex); 6020 } 6021 6022 /** 6023 * current_work - retrieve %current task's work struct 6024 * 6025 * Determine if %current task is a workqueue worker and what it's working on. 6026 * Useful to find out the context that the %current task is running in. 6027 * 6028 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 6029 */ 6030 struct work_struct *current_work(void) 6031 { 6032 struct worker *worker = current_wq_worker(); 6033 6034 return worker ? worker->current_work : NULL; 6035 } 6036 EXPORT_SYMBOL(current_work); 6037 6038 /** 6039 * current_is_workqueue_rescuer - is %current workqueue rescuer? 6040 * 6041 * Determine whether %current is a workqueue rescuer. Can be used from 6042 * work functions to determine whether it's being run off the rescuer task. 6043 * 6044 * Return: %true if %current is a workqueue rescuer. %false otherwise. 6045 */ 6046 bool current_is_workqueue_rescuer(void) 6047 { 6048 struct worker *worker = current_wq_worker(); 6049 6050 return worker && worker->rescue_wq; 6051 } 6052 6053 /** 6054 * workqueue_congested - test whether a workqueue is congested 6055 * @cpu: CPU in question 6056 * @wq: target workqueue 6057 * 6058 * Test whether @wq's cpu workqueue for @cpu is congested. There is 6059 * no synchronization around this function and the test result is 6060 * unreliable and only useful as advisory hints or for debugging. 6061 * 6062 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 6063 * 6064 * With the exception of ordered workqueues, all workqueues have per-cpu 6065 * pool_workqueues, each with its own congested state. A workqueue being 6066 * congested on one CPU doesn't mean that the workqueue is contested on any 6067 * other CPUs. 6068 * 6069 * Return: 6070 * %true if congested, %false otherwise. 6071 */ 6072 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 6073 { 6074 struct pool_workqueue *pwq; 6075 bool ret; 6076 6077 preempt_disable(); 6078 6079 if (cpu == WORK_CPU_UNBOUND) 6080 cpu = smp_processor_id(); 6081 6082 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 6083 ret = !list_empty(&pwq->inactive_works); 6084 6085 preempt_enable(); 6086 6087 return ret; 6088 } 6089 EXPORT_SYMBOL_GPL(workqueue_congested); 6090 6091 /** 6092 * work_busy - test whether a work is currently pending or running 6093 * @work: the work to be tested 6094 * 6095 * Test whether @work is currently pending or running. There is no 6096 * synchronization around this function and the test result is 6097 * unreliable and only useful as advisory hints or for debugging. 6098 * 6099 * Return: 6100 * OR'd bitmask of WORK_BUSY_* bits. 6101 */ 6102 unsigned int work_busy(struct work_struct *work) 6103 { 6104 struct worker_pool *pool; 6105 unsigned long irq_flags; 6106 unsigned int ret = 0; 6107 6108 if (work_pending(work)) 6109 ret |= WORK_BUSY_PENDING; 6110 6111 rcu_read_lock(); 6112 pool = get_work_pool(work); 6113 if (pool) { 6114 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6115 if (find_worker_executing_work(pool, work)) 6116 ret |= WORK_BUSY_RUNNING; 6117 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6118 } 6119 rcu_read_unlock(); 6120 6121 return ret; 6122 } 6123 EXPORT_SYMBOL_GPL(work_busy); 6124 6125 /** 6126 * set_worker_desc - set description for the current work item 6127 * @fmt: printf-style format string 6128 * @...: arguments for the format string 6129 * 6130 * This function can be called by a running work function to describe what 6131 * the work item is about. If the worker task gets dumped, this 6132 * information will be printed out together to help debugging. The 6133 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 6134 */ 6135 void set_worker_desc(const char *fmt, ...) 6136 { 6137 struct worker *worker = current_wq_worker(); 6138 va_list args; 6139 6140 if (worker) { 6141 va_start(args, fmt); 6142 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 6143 va_end(args); 6144 } 6145 } 6146 EXPORT_SYMBOL_GPL(set_worker_desc); 6147 6148 /** 6149 * print_worker_info - print out worker information and description 6150 * @log_lvl: the log level to use when printing 6151 * @task: target task 6152 * 6153 * If @task is a worker and currently executing a work item, print out the 6154 * name of the workqueue being serviced and worker description set with 6155 * set_worker_desc() by the currently executing work item. 6156 * 6157 * This function can be safely called on any task as long as the 6158 * task_struct itself is accessible. While safe, this function isn't 6159 * synchronized and may print out mixups or garbages of limited length. 6160 */ 6161 void print_worker_info(const char *log_lvl, struct task_struct *task) 6162 { 6163 work_func_t *fn = NULL; 6164 char name[WQ_NAME_LEN] = { }; 6165 char desc[WORKER_DESC_LEN] = { }; 6166 struct pool_workqueue *pwq = NULL; 6167 struct workqueue_struct *wq = NULL; 6168 struct worker *worker; 6169 6170 if (!(task->flags & PF_WQ_WORKER)) 6171 return; 6172 6173 /* 6174 * This function is called without any synchronization and @task 6175 * could be in any state. Be careful with dereferences. 6176 */ 6177 worker = kthread_probe_data(task); 6178 6179 /* 6180 * Carefully copy the associated workqueue's workfn, name and desc. 6181 * Keep the original last '\0' in case the original is garbage. 6182 */ 6183 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 6184 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 6185 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 6186 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 6187 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 6188 6189 if (fn || name[0] || desc[0]) { 6190 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 6191 if (strcmp(name, desc)) 6192 pr_cont(" (%s)", desc); 6193 pr_cont("\n"); 6194 } 6195 } 6196 6197 static void pr_cont_pool_info(struct worker_pool *pool) 6198 { 6199 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 6200 if (pool->node != NUMA_NO_NODE) 6201 pr_cont(" node=%d", pool->node); 6202 pr_cont(" flags=0x%x", pool->flags); 6203 if (pool->flags & POOL_BH) 6204 pr_cont(" bh%s", 6205 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6206 else 6207 pr_cont(" nice=%d", pool->attrs->nice); 6208 } 6209 6210 static void pr_cont_worker_id(struct worker *worker) 6211 { 6212 struct worker_pool *pool = worker->pool; 6213 6214 if (pool->flags & WQ_BH) 6215 pr_cont("bh%s", 6216 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6217 else 6218 pr_cont("%d%s", task_pid_nr(worker->task), 6219 worker->rescue_wq ? "(RESCUER)" : ""); 6220 } 6221 6222 struct pr_cont_work_struct { 6223 bool comma; 6224 work_func_t func; 6225 long ctr; 6226 }; 6227 6228 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 6229 { 6230 if (!pcwsp->ctr) 6231 goto out_record; 6232 if (func == pcwsp->func) { 6233 pcwsp->ctr++; 6234 return; 6235 } 6236 if (pcwsp->ctr == 1) 6237 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 6238 else 6239 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 6240 pcwsp->ctr = 0; 6241 out_record: 6242 if ((long)func == -1L) 6243 return; 6244 pcwsp->comma = comma; 6245 pcwsp->func = func; 6246 pcwsp->ctr = 1; 6247 } 6248 6249 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 6250 { 6251 if (work->func == wq_barrier_func) { 6252 struct wq_barrier *barr; 6253 6254 barr = container_of(work, struct wq_barrier, work); 6255 6256 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6257 pr_cont("%s BAR(%d)", comma ? "," : "", 6258 task_pid_nr(barr->task)); 6259 } else { 6260 if (!comma) 6261 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6262 pr_cont_work_flush(comma, work->func, pcwsp); 6263 } 6264 } 6265 6266 static void show_pwq(struct pool_workqueue *pwq) 6267 { 6268 struct pr_cont_work_struct pcws = { .ctr = 0, }; 6269 struct worker_pool *pool = pwq->pool; 6270 struct work_struct *work; 6271 struct worker *worker; 6272 bool has_in_flight = false, has_pending = false; 6273 int bkt; 6274 6275 pr_info(" pwq %d:", pool->id); 6276 pr_cont_pool_info(pool); 6277 6278 pr_cont(" active=%d refcnt=%d%s\n", 6279 pwq->nr_active, pwq->refcnt, 6280 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 6281 6282 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6283 if (worker->current_pwq == pwq) { 6284 has_in_flight = true; 6285 break; 6286 } 6287 } 6288 if (has_in_flight) { 6289 bool comma = false; 6290 6291 pr_info(" in-flight:"); 6292 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6293 if (worker->current_pwq != pwq) 6294 continue; 6295 6296 pr_cont(" %s", comma ? "," : ""); 6297 pr_cont_worker_id(worker); 6298 pr_cont(":%ps", worker->current_func); 6299 list_for_each_entry(work, &worker->scheduled, entry) 6300 pr_cont_work(false, work, &pcws); 6301 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6302 comma = true; 6303 } 6304 pr_cont("\n"); 6305 } 6306 6307 list_for_each_entry(work, &pool->worklist, entry) { 6308 if (get_work_pwq(work) == pwq) { 6309 has_pending = true; 6310 break; 6311 } 6312 } 6313 if (has_pending) { 6314 bool comma = false; 6315 6316 pr_info(" pending:"); 6317 list_for_each_entry(work, &pool->worklist, entry) { 6318 if (get_work_pwq(work) != pwq) 6319 continue; 6320 6321 pr_cont_work(comma, work, &pcws); 6322 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6323 } 6324 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6325 pr_cont("\n"); 6326 } 6327 6328 if (!list_empty(&pwq->inactive_works)) { 6329 bool comma = false; 6330 6331 pr_info(" inactive:"); 6332 list_for_each_entry(work, &pwq->inactive_works, entry) { 6333 pr_cont_work(comma, work, &pcws); 6334 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6335 } 6336 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6337 pr_cont("\n"); 6338 } 6339 } 6340 6341 /** 6342 * show_one_workqueue - dump state of specified workqueue 6343 * @wq: workqueue whose state will be printed 6344 */ 6345 void show_one_workqueue(struct workqueue_struct *wq) 6346 { 6347 struct pool_workqueue *pwq; 6348 bool idle = true; 6349 unsigned long irq_flags; 6350 6351 for_each_pwq(pwq, wq) { 6352 if (!pwq_is_empty(pwq)) { 6353 idle = false; 6354 break; 6355 } 6356 } 6357 if (idle) /* Nothing to print for idle workqueue */ 6358 return; 6359 6360 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 6361 6362 for_each_pwq(pwq, wq) { 6363 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 6364 if (!pwq_is_empty(pwq)) { 6365 /* 6366 * Defer printing to avoid deadlocks in console 6367 * drivers that queue work while holding locks 6368 * also taken in their write paths. 6369 */ 6370 printk_deferred_enter(); 6371 show_pwq(pwq); 6372 printk_deferred_exit(); 6373 } 6374 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 6375 /* 6376 * We could be printing a lot from atomic context, e.g. 6377 * sysrq-t -> show_all_workqueues(). Avoid triggering 6378 * hard lockup. 6379 */ 6380 touch_nmi_watchdog(); 6381 } 6382 6383 } 6384 6385 /** 6386 * show_one_worker_pool - dump state of specified worker pool 6387 * @pool: worker pool whose state will be printed 6388 */ 6389 static void show_one_worker_pool(struct worker_pool *pool) 6390 { 6391 struct worker *worker; 6392 bool first = true; 6393 unsigned long irq_flags; 6394 unsigned long hung = 0; 6395 6396 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6397 if (pool->nr_workers == pool->nr_idle) 6398 goto next_pool; 6399 6400 /* How long the first pending work is waiting for a worker. */ 6401 if (!list_empty(&pool->worklist)) 6402 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 6403 6404 /* 6405 * Defer printing to avoid deadlocks in console drivers that 6406 * queue work while holding locks also taken in their write 6407 * paths. 6408 */ 6409 printk_deferred_enter(); 6410 pr_info("pool %d:", pool->id); 6411 pr_cont_pool_info(pool); 6412 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 6413 if (pool->manager) 6414 pr_cont(" manager: %d", 6415 task_pid_nr(pool->manager->task)); 6416 list_for_each_entry(worker, &pool->idle_list, entry) { 6417 pr_cont(" %s", first ? "idle: " : ""); 6418 pr_cont_worker_id(worker); 6419 first = false; 6420 } 6421 pr_cont("\n"); 6422 printk_deferred_exit(); 6423 next_pool: 6424 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6425 /* 6426 * We could be printing a lot from atomic context, e.g. 6427 * sysrq-t -> show_all_workqueues(). Avoid triggering 6428 * hard lockup. 6429 */ 6430 touch_nmi_watchdog(); 6431 6432 } 6433 6434 /** 6435 * show_all_workqueues - dump workqueue state 6436 * 6437 * Called from a sysrq handler and prints out all busy workqueues and pools. 6438 */ 6439 void show_all_workqueues(void) 6440 { 6441 struct workqueue_struct *wq; 6442 struct worker_pool *pool; 6443 int pi; 6444 6445 rcu_read_lock(); 6446 6447 pr_info("Showing busy workqueues and worker pools:\n"); 6448 6449 list_for_each_entry_rcu(wq, &workqueues, list) 6450 show_one_workqueue(wq); 6451 6452 for_each_pool(pool, pi) 6453 show_one_worker_pool(pool); 6454 6455 rcu_read_unlock(); 6456 } 6457 6458 /** 6459 * show_freezable_workqueues - dump freezable workqueue state 6460 * 6461 * Called from try_to_freeze_tasks() and prints out all freezable workqueues 6462 * still busy. 6463 */ 6464 void show_freezable_workqueues(void) 6465 { 6466 struct workqueue_struct *wq; 6467 6468 rcu_read_lock(); 6469 6470 pr_info("Showing freezable workqueues that are still busy:\n"); 6471 6472 list_for_each_entry_rcu(wq, &workqueues, list) { 6473 if (!(wq->flags & WQ_FREEZABLE)) 6474 continue; 6475 show_one_workqueue(wq); 6476 } 6477 6478 rcu_read_unlock(); 6479 } 6480 6481 /* used to show worker information through /proc/PID/{comm,stat,status} */ 6482 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 6483 { 6484 /* stabilize PF_WQ_WORKER and worker pool association */ 6485 mutex_lock(&wq_pool_attach_mutex); 6486 6487 if (task->flags & PF_WQ_WORKER) { 6488 struct worker *worker = kthread_data(task); 6489 struct worker_pool *pool = worker->pool; 6490 int off; 6491 6492 off = format_worker_id(buf, size, worker, pool); 6493 6494 if (pool) { 6495 raw_spin_lock_irq(&pool->lock); 6496 /* 6497 * ->desc tracks information (wq name or 6498 * set_worker_desc()) for the latest execution. If 6499 * current, prepend '+', otherwise '-'. 6500 */ 6501 if (worker->desc[0] != '\0') { 6502 if (worker->current_work) 6503 scnprintf(buf + off, size - off, "+%s", 6504 worker->desc); 6505 else 6506 scnprintf(buf + off, size - off, "-%s", 6507 worker->desc); 6508 } 6509 raw_spin_unlock_irq(&pool->lock); 6510 } 6511 } else { 6512 strscpy(buf, task->comm, size); 6513 } 6514 6515 mutex_unlock(&wq_pool_attach_mutex); 6516 } 6517 6518 #ifdef CONFIG_SMP 6519 6520 /* 6521 * CPU hotplug. 6522 * 6523 * There are two challenges in supporting CPU hotplug. Firstly, there 6524 * are a lot of assumptions on strong associations among work, pwq and 6525 * pool which make migrating pending and scheduled works very 6526 * difficult to implement without impacting hot paths. Secondly, 6527 * worker pools serve mix of short, long and very long running works making 6528 * blocked draining impractical. 6529 * 6530 * This is solved by allowing the pools to be disassociated from the CPU 6531 * running as an unbound one and allowing it to be reattached later if the 6532 * cpu comes back online. 6533 */ 6534 6535 static void unbind_workers(int cpu) 6536 { 6537 struct worker_pool *pool; 6538 struct worker *worker; 6539 6540 for_each_cpu_worker_pool(pool, cpu) { 6541 mutex_lock(&wq_pool_attach_mutex); 6542 raw_spin_lock_irq(&pool->lock); 6543 6544 /* 6545 * We've blocked all attach/detach operations. Make all workers 6546 * unbound and set DISASSOCIATED. Before this, all workers 6547 * must be on the cpu. After this, they may become diasporas. 6548 * And the preemption disabled section in their sched callbacks 6549 * are guaranteed to see WORKER_UNBOUND since the code here 6550 * is on the same cpu. 6551 */ 6552 for_each_pool_worker(worker, pool) 6553 worker->flags |= WORKER_UNBOUND; 6554 6555 pool->flags |= POOL_DISASSOCIATED; 6556 6557 /* 6558 * The handling of nr_running in sched callbacks are disabled 6559 * now. Zap nr_running. After this, nr_running stays zero and 6560 * need_more_worker() and keep_working() are always true as 6561 * long as the worklist is not empty. This pool now behaves as 6562 * an unbound (in terms of concurrency management) pool which 6563 * are served by workers tied to the pool. 6564 */ 6565 pool->nr_running = 0; 6566 6567 /* 6568 * With concurrency management just turned off, a busy 6569 * worker blocking could lead to lengthy stalls. Kick off 6570 * unbound chain execution of currently pending work items. 6571 */ 6572 kick_pool(pool); 6573 6574 raw_spin_unlock_irq(&pool->lock); 6575 6576 for_each_pool_worker(worker, pool) 6577 unbind_worker(worker); 6578 6579 mutex_unlock(&wq_pool_attach_mutex); 6580 } 6581 } 6582 6583 /** 6584 * rebind_workers - rebind all workers of a pool to the associated CPU 6585 * @pool: pool of interest 6586 * 6587 * @pool->cpu is coming online. Rebind all workers to the CPU. 6588 */ 6589 static void rebind_workers(struct worker_pool *pool) 6590 { 6591 struct worker *worker; 6592 6593 lockdep_assert_held(&wq_pool_attach_mutex); 6594 6595 /* 6596 * Restore CPU affinity of all workers. As all idle workers should 6597 * be on the run-queue of the associated CPU before any local 6598 * wake-ups for concurrency management happen, restore CPU affinity 6599 * of all workers first and then clear UNBOUND. As we're called 6600 * from CPU_ONLINE, the following shouldn't fail. 6601 */ 6602 for_each_pool_worker(worker, pool) { 6603 kthread_set_per_cpu(worker->task, pool->cpu); 6604 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 6605 pool_allowed_cpus(pool)) < 0); 6606 } 6607 6608 raw_spin_lock_irq(&pool->lock); 6609 6610 pool->flags &= ~POOL_DISASSOCIATED; 6611 6612 for_each_pool_worker(worker, pool) { 6613 unsigned int worker_flags = worker->flags; 6614 6615 /* 6616 * We want to clear UNBOUND but can't directly call 6617 * worker_clr_flags() or adjust nr_running. Atomically 6618 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 6619 * @worker will clear REBOUND using worker_clr_flags() when 6620 * it initiates the next execution cycle thus restoring 6621 * concurrency management. Note that when or whether 6622 * @worker clears REBOUND doesn't affect correctness. 6623 * 6624 * WRITE_ONCE() is necessary because @worker->flags may be 6625 * tested without holding any lock in 6626 * wq_worker_running(). Without it, NOT_RUNNING test may 6627 * fail incorrectly leading to premature concurrency 6628 * management operations. 6629 */ 6630 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 6631 worker_flags |= WORKER_REBOUND; 6632 worker_flags &= ~WORKER_UNBOUND; 6633 WRITE_ONCE(worker->flags, worker_flags); 6634 } 6635 6636 raw_spin_unlock_irq(&pool->lock); 6637 } 6638 6639 /** 6640 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 6641 * @pool: unbound pool of interest 6642 * @cpu: the CPU which is coming up 6643 * 6644 * An unbound pool may end up with a cpumask which doesn't have any online 6645 * CPUs. When a worker of such pool get scheduled, the scheduler resets 6646 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 6647 * online CPU before, cpus_allowed of all its workers should be restored. 6648 */ 6649 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 6650 { 6651 static cpumask_t cpumask; 6652 struct worker *worker; 6653 6654 lockdep_assert_held(&wq_pool_attach_mutex); 6655 6656 /* is @cpu allowed for @pool? */ 6657 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 6658 return; 6659 6660 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 6661 6662 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 6663 for_each_pool_worker(worker, pool) 6664 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 6665 } 6666 6667 int workqueue_prepare_cpu(unsigned int cpu) 6668 { 6669 struct worker_pool *pool; 6670 6671 for_each_cpu_worker_pool(pool, cpu) { 6672 if (pool->nr_workers) 6673 continue; 6674 if (!create_worker(pool)) 6675 return -ENOMEM; 6676 } 6677 return 0; 6678 } 6679 6680 int workqueue_online_cpu(unsigned int cpu) 6681 { 6682 struct worker_pool *pool; 6683 struct workqueue_struct *wq; 6684 int pi; 6685 6686 mutex_lock(&wq_pool_mutex); 6687 6688 cpumask_set_cpu(cpu, wq_online_cpumask); 6689 6690 for_each_pool(pool, pi) { 6691 /* BH pools aren't affected by hotplug */ 6692 if (pool->flags & POOL_BH) 6693 continue; 6694 6695 mutex_lock(&wq_pool_attach_mutex); 6696 if (pool->cpu == cpu) 6697 rebind_workers(pool); 6698 else if (pool->cpu < 0) 6699 restore_unbound_workers_cpumask(pool, cpu); 6700 mutex_unlock(&wq_pool_attach_mutex); 6701 } 6702 6703 /* update pod affinity of unbound workqueues */ 6704 list_for_each_entry(wq, &workqueues, list) { 6705 struct workqueue_attrs *attrs = wq->unbound_attrs; 6706 6707 if (attrs) { 6708 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6709 int tcpu; 6710 6711 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6712 unbound_wq_update_pwq(wq, tcpu); 6713 6714 mutex_lock(&wq->mutex); 6715 wq_update_node_max_active(wq, -1); 6716 mutex_unlock(&wq->mutex); 6717 } 6718 } 6719 6720 mutex_unlock(&wq_pool_mutex); 6721 return 0; 6722 } 6723 6724 int workqueue_offline_cpu(unsigned int cpu) 6725 { 6726 struct workqueue_struct *wq; 6727 6728 /* unbinding per-cpu workers should happen on the local CPU */ 6729 if (WARN_ON(cpu != smp_processor_id())) 6730 return -1; 6731 6732 unbind_workers(cpu); 6733 6734 /* update pod affinity of unbound workqueues */ 6735 mutex_lock(&wq_pool_mutex); 6736 6737 cpumask_clear_cpu(cpu, wq_online_cpumask); 6738 6739 list_for_each_entry(wq, &workqueues, list) { 6740 struct workqueue_attrs *attrs = wq->unbound_attrs; 6741 6742 if (attrs) { 6743 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6744 int tcpu; 6745 6746 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6747 unbound_wq_update_pwq(wq, tcpu); 6748 6749 mutex_lock(&wq->mutex); 6750 wq_update_node_max_active(wq, cpu); 6751 mutex_unlock(&wq->mutex); 6752 } 6753 } 6754 mutex_unlock(&wq_pool_mutex); 6755 6756 return 0; 6757 } 6758 6759 struct work_for_cpu { 6760 struct work_struct work; 6761 long (*fn)(void *); 6762 void *arg; 6763 long ret; 6764 }; 6765 6766 static void work_for_cpu_fn(struct work_struct *work) 6767 { 6768 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 6769 6770 wfc->ret = wfc->fn(wfc->arg); 6771 } 6772 6773 /** 6774 * work_on_cpu_key - run a function in thread context on a particular cpu 6775 * @cpu: the cpu to run on 6776 * @fn: the function to run 6777 * @arg: the function arg 6778 * @key: The lock class key for lock debugging purposes 6779 * 6780 * It is up to the caller to ensure that the cpu doesn't go offline. 6781 * The caller must not hold any locks which would prevent @fn from completing. 6782 * 6783 * Return: The value @fn returns. 6784 */ 6785 long work_on_cpu_key(int cpu, long (*fn)(void *), 6786 void *arg, struct lock_class_key *key) 6787 { 6788 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 6789 6790 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); 6791 schedule_work_on(cpu, &wfc.work); 6792 flush_work(&wfc.work); 6793 destroy_work_on_stack(&wfc.work); 6794 return wfc.ret; 6795 } 6796 EXPORT_SYMBOL_GPL(work_on_cpu_key); 6797 #endif /* CONFIG_SMP */ 6798 6799 #ifdef CONFIG_FREEZER 6800 6801 /** 6802 * freeze_workqueues_begin - begin freezing workqueues 6803 * 6804 * Start freezing workqueues. After this function returns, all freezable 6805 * workqueues will queue new works to their inactive_works list instead of 6806 * pool->worklist. 6807 * 6808 * CONTEXT: 6809 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6810 */ 6811 void freeze_workqueues_begin(void) 6812 { 6813 struct workqueue_struct *wq; 6814 6815 mutex_lock(&wq_pool_mutex); 6816 6817 WARN_ON_ONCE(workqueue_freezing); 6818 workqueue_freezing = true; 6819 6820 list_for_each_entry(wq, &workqueues, list) { 6821 mutex_lock(&wq->mutex); 6822 wq_adjust_max_active(wq); 6823 mutex_unlock(&wq->mutex); 6824 } 6825 6826 mutex_unlock(&wq_pool_mutex); 6827 } 6828 6829 /** 6830 * freeze_workqueues_busy - are freezable workqueues still busy? 6831 * 6832 * Check whether freezing is complete. This function must be called 6833 * between freeze_workqueues_begin() and thaw_workqueues(). 6834 * 6835 * CONTEXT: 6836 * Grabs and releases wq_pool_mutex. 6837 * 6838 * Return: 6839 * %true if some freezable workqueues are still busy. %false if freezing 6840 * is complete. 6841 */ 6842 bool freeze_workqueues_busy(void) 6843 { 6844 bool busy = false; 6845 struct workqueue_struct *wq; 6846 struct pool_workqueue *pwq; 6847 6848 mutex_lock(&wq_pool_mutex); 6849 6850 WARN_ON_ONCE(!workqueue_freezing); 6851 6852 list_for_each_entry(wq, &workqueues, list) { 6853 if (!(wq->flags & WQ_FREEZABLE)) 6854 continue; 6855 /* 6856 * nr_active is monotonically decreasing. It's safe 6857 * to peek without lock. 6858 */ 6859 rcu_read_lock(); 6860 for_each_pwq(pwq, wq) { 6861 WARN_ON_ONCE(pwq->nr_active < 0); 6862 if (pwq->nr_active) { 6863 busy = true; 6864 rcu_read_unlock(); 6865 goto out_unlock; 6866 } 6867 } 6868 rcu_read_unlock(); 6869 } 6870 out_unlock: 6871 mutex_unlock(&wq_pool_mutex); 6872 return busy; 6873 } 6874 6875 /** 6876 * thaw_workqueues - thaw workqueues 6877 * 6878 * Thaw workqueues. Normal queueing is restored and all collected 6879 * frozen works are transferred to their respective pool worklists. 6880 * 6881 * CONTEXT: 6882 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6883 */ 6884 void thaw_workqueues(void) 6885 { 6886 struct workqueue_struct *wq; 6887 6888 mutex_lock(&wq_pool_mutex); 6889 6890 if (!workqueue_freezing) 6891 goto out_unlock; 6892 6893 workqueue_freezing = false; 6894 6895 /* restore max_active and repopulate worklist */ 6896 list_for_each_entry(wq, &workqueues, list) { 6897 mutex_lock(&wq->mutex); 6898 wq_adjust_max_active(wq); 6899 mutex_unlock(&wq->mutex); 6900 } 6901 6902 out_unlock: 6903 mutex_unlock(&wq_pool_mutex); 6904 } 6905 #endif /* CONFIG_FREEZER */ 6906 6907 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 6908 { 6909 LIST_HEAD(ctxs); 6910 int ret = 0; 6911 struct workqueue_struct *wq; 6912 struct apply_wqattrs_ctx *ctx, *n; 6913 6914 lockdep_assert_held(&wq_pool_mutex); 6915 6916 list_for_each_entry(wq, &workqueues, list) { 6917 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING)) 6918 continue; 6919 6920 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 6921 if (IS_ERR(ctx)) { 6922 ret = PTR_ERR(ctx); 6923 break; 6924 } 6925 6926 list_add_tail(&ctx->list, &ctxs); 6927 } 6928 6929 list_for_each_entry_safe(ctx, n, &ctxs, list) { 6930 if (!ret) 6931 apply_wqattrs_commit(ctx); 6932 apply_wqattrs_cleanup(ctx); 6933 } 6934 6935 if (!ret) { 6936 int cpu; 6937 struct worker_pool *pool; 6938 struct worker *worker; 6939 6940 mutex_lock(&wq_pool_attach_mutex); 6941 cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 6942 /* rescuer needs to respect cpumask changes when it is not attached */ 6943 list_for_each_entry(wq, &workqueues, list) { 6944 if (wq->rescuer && !wq->rescuer->pool) 6945 unbind_worker(wq->rescuer); 6946 } 6947 /* DISASSOCIATED worker needs to respect wq_unbound_cpumask */ 6948 for_each_possible_cpu(cpu) { 6949 for_each_cpu_worker_pool(pool, cpu) { 6950 if (!(pool->flags & POOL_DISASSOCIATED)) 6951 continue; 6952 for_each_pool_worker(worker, pool) 6953 unbind_worker(worker); 6954 } 6955 } 6956 mutex_unlock(&wq_pool_attach_mutex); 6957 } 6958 return ret; 6959 } 6960 6961 /** 6962 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask 6963 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask 6964 * 6965 * This function can be called from cpuset code to provide a set of isolated 6966 * CPUs that should be excluded from wq_unbound_cpumask. 6967 */ 6968 int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask) 6969 { 6970 cpumask_var_t cpumask; 6971 int ret = 0; 6972 6973 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6974 return -ENOMEM; 6975 6976 mutex_lock(&wq_pool_mutex); 6977 6978 /* 6979 * If the operation fails, it will fall back to 6980 * wq_requested_unbound_cpumask which is initially set to 6981 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten 6982 * by any subsequent write to workqueue/cpumask sysfs file. 6983 */ 6984 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask)) 6985 cpumask_copy(cpumask, wq_requested_unbound_cpumask); 6986 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 6987 ret = workqueue_apply_unbound_cpumask(cpumask); 6988 6989 /* Save the current isolated cpumask & export it via sysfs */ 6990 if (!ret) 6991 cpumask_copy(wq_isolated_cpumask, exclude_cpumask); 6992 6993 mutex_unlock(&wq_pool_mutex); 6994 free_cpumask_var(cpumask); 6995 return ret; 6996 } 6997 6998 static int parse_affn_scope(const char *val) 6999 { 7000 int i; 7001 7002 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 7003 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 7004 return i; 7005 } 7006 return -EINVAL; 7007 } 7008 7009 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 7010 { 7011 struct workqueue_struct *wq; 7012 int affn, cpu; 7013 7014 affn = parse_affn_scope(val); 7015 if (affn < 0) 7016 return affn; 7017 if (affn == WQ_AFFN_DFL) 7018 return -EINVAL; 7019 7020 cpus_read_lock(); 7021 mutex_lock(&wq_pool_mutex); 7022 7023 wq_affn_dfl = affn; 7024 7025 list_for_each_entry(wq, &workqueues, list) { 7026 for_each_online_cpu(cpu) 7027 unbound_wq_update_pwq(wq, cpu); 7028 } 7029 7030 mutex_unlock(&wq_pool_mutex); 7031 cpus_read_unlock(); 7032 7033 return 0; 7034 } 7035 7036 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 7037 { 7038 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 7039 } 7040 7041 static const struct kernel_param_ops wq_affn_dfl_ops = { 7042 .set = wq_affn_dfl_set, 7043 .get = wq_affn_dfl_get, 7044 }; 7045 7046 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 7047 7048 #ifdef CONFIG_SYSFS 7049 /* 7050 * Workqueues with WQ_SYSFS flag set is visible to userland via 7051 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 7052 * following attributes. 7053 * 7054 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 7055 * max_active RW int : maximum number of in-flight work items 7056 * 7057 * Unbound workqueues have the following extra attributes. 7058 * 7059 * nice RW int : nice value of the workers 7060 * cpumask RW mask : bitmask of allowed CPUs for the workers 7061 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 7062 * affinity_strict RW bool : worker CPU affinity is strict 7063 */ 7064 struct wq_device { 7065 struct workqueue_struct *wq; 7066 struct device dev; 7067 }; 7068 7069 static struct workqueue_struct *dev_to_wq(struct device *dev) 7070 { 7071 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7072 7073 return wq_dev->wq; 7074 } 7075 7076 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 7077 char *buf) 7078 { 7079 struct workqueue_struct *wq = dev_to_wq(dev); 7080 7081 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 7082 } 7083 static DEVICE_ATTR_RO(per_cpu); 7084 7085 static ssize_t max_active_show(struct device *dev, 7086 struct device_attribute *attr, char *buf) 7087 { 7088 struct workqueue_struct *wq = dev_to_wq(dev); 7089 7090 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 7091 } 7092 7093 static ssize_t max_active_store(struct device *dev, 7094 struct device_attribute *attr, const char *buf, 7095 size_t count) 7096 { 7097 struct workqueue_struct *wq = dev_to_wq(dev); 7098 int val; 7099 7100 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 7101 return -EINVAL; 7102 7103 workqueue_set_max_active(wq, val); 7104 return count; 7105 } 7106 static DEVICE_ATTR_RW(max_active); 7107 7108 static struct attribute *wq_sysfs_attrs[] = { 7109 &dev_attr_per_cpu.attr, 7110 &dev_attr_max_active.attr, 7111 NULL, 7112 }; 7113 ATTRIBUTE_GROUPS(wq_sysfs); 7114 7115 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 7116 char *buf) 7117 { 7118 struct workqueue_struct *wq = dev_to_wq(dev); 7119 int written; 7120 7121 mutex_lock(&wq->mutex); 7122 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 7123 mutex_unlock(&wq->mutex); 7124 7125 return written; 7126 } 7127 7128 /* prepare workqueue_attrs for sysfs store operations */ 7129 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 7130 { 7131 struct workqueue_attrs *attrs; 7132 7133 lockdep_assert_held(&wq_pool_mutex); 7134 7135 attrs = alloc_workqueue_attrs(); 7136 if (!attrs) 7137 return NULL; 7138 7139 copy_workqueue_attrs(attrs, wq->unbound_attrs); 7140 return attrs; 7141 } 7142 7143 static ssize_t wq_nice_store(struct device *dev, 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 if (sscanf(buf, "%d", &attrs->nice) == 1 && 7157 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 7158 ret = apply_workqueue_attrs_locked(wq, attrs); 7159 else 7160 ret = -EINVAL; 7161 7162 out_unlock: 7163 apply_wqattrs_unlock(); 7164 free_workqueue_attrs(attrs); 7165 return ret ?: count; 7166 } 7167 7168 static ssize_t wq_cpumask_show(struct device *dev, 7169 struct device_attribute *attr, char *buf) 7170 { 7171 struct workqueue_struct *wq = dev_to_wq(dev); 7172 int written; 7173 7174 mutex_lock(&wq->mutex); 7175 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 7176 cpumask_pr_args(wq->unbound_attrs->cpumask)); 7177 mutex_unlock(&wq->mutex); 7178 return written; 7179 } 7180 7181 static ssize_t wq_cpumask_store(struct device *dev, 7182 struct device_attribute *attr, 7183 const char *buf, size_t count) 7184 { 7185 struct workqueue_struct *wq = dev_to_wq(dev); 7186 struct workqueue_attrs *attrs; 7187 int ret = -ENOMEM; 7188 7189 apply_wqattrs_lock(); 7190 7191 attrs = wq_sysfs_prep_attrs(wq); 7192 if (!attrs) 7193 goto out_unlock; 7194 7195 ret = cpumask_parse(buf, attrs->cpumask); 7196 if (!ret) 7197 ret = apply_workqueue_attrs_locked(wq, attrs); 7198 7199 out_unlock: 7200 apply_wqattrs_unlock(); 7201 free_workqueue_attrs(attrs); 7202 return ret ?: count; 7203 } 7204 7205 static ssize_t wq_affn_scope_show(struct device *dev, 7206 struct device_attribute *attr, char *buf) 7207 { 7208 struct workqueue_struct *wq = dev_to_wq(dev); 7209 int written; 7210 7211 mutex_lock(&wq->mutex); 7212 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 7213 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 7214 wq_affn_names[WQ_AFFN_DFL], 7215 wq_affn_names[wq_affn_dfl]); 7216 else 7217 written = scnprintf(buf, PAGE_SIZE, "%s\n", 7218 wq_affn_names[wq->unbound_attrs->affn_scope]); 7219 mutex_unlock(&wq->mutex); 7220 7221 return written; 7222 } 7223 7224 static ssize_t wq_affn_scope_store(struct device *dev, 7225 struct device_attribute *attr, 7226 const char *buf, size_t count) 7227 { 7228 struct workqueue_struct *wq = dev_to_wq(dev); 7229 struct workqueue_attrs *attrs; 7230 int affn, ret = -ENOMEM; 7231 7232 affn = parse_affn_scope(buf); 7233 if (affn < 0) 7234 return affn; 7235 7236 apply_wqattrs_lock(); 7237 attrs = wq_sysfs_prep_attrs(wq); 7238 if (attrs) { 7239 attrs->affn_scope = affn; 7240 ret = apply_workqueue_attrs_locked(wq, attrs); 7241 } 7242 apply_wqattrs_unlock(); 7243 free_workqueue_attrs(attrs); 7244 return ret ?: count; 7245 } 7246 7247 static ssize_t wq_affinity_strict_show(struct device *dev, 7248 struct device_attribute *attr, char *buf) 7249 { 7250 struct workqueue_struct *wq = dev_to_wq(dev); 7251 7252 return scnprintf(buf, PAGE_SIZE, "%d\n", 7253 wq->unbound_attrs->affn_strict); 7254 } 7255 7256 static ssize_t wq_affinity_strict_store(struct device *dev, 7257 struct device_attribute *attr, 7258 const char *buf, size_t count) 7259 { 7260 struct workqueue_struct *wq = dev_to_wq(dev); 7261 struct workqueue_attrs *attrs; 7262 int v, ret = -ENOMEM; 7263 7264 if (sscanf(buf, "%d", &v) != 1) 7265 return -EINVAL; 7266 7267 apply_wqattrs_lock(); 7268 attrs = wq_sysfs_prep_attrs(wq); 7269 if (attrs) { 7270 attrs->affn_strict = (bool)v; 7271 ret = apply_workqueue_attrs_locked(wq, attrs); 7272 } 7273 apply_wqattrs_unlock(); 7274 free_workqueue_attrs(attrs); 7275 return ret ?: count; 7276 } 7277 7278 static struct device_attribute wq_sysfs_unbound_attrs[] = { 7279 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 7280 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 7281 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 7282 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 7283 __ATTR_NULL, 7284 }; 7285 7286 static const struct bus_type wq_subsys = { 7287 .name = "workqueue", 7288 .dev_groups = wq_sysfs_groups, 7289 }; 7290 7291 /** 7292 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 7293 * @cpumask: the cpumask to set 7294 * 7295 * The low-level workqueues cpumask is a global cpumask that limits 7296 * the affinity of all unbound workqueues. This function check the @cpumask 7297 * and apply it to all unbound workqueues and updates all pwqs of them. 7298 * 7299 * Return: 0 - Success 7300 * -EINVAL - Invalid @cpumask 7301 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 7302 */ 7303 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 7304 { 7305 int ret = -EINVAL; 7306 7307 /* 7308 * Not excluding isolated cpus on purpose. 7309 * If the user wishes to include them, we allow that. 7310 */ 7311 cpumask_and(cpumask, cpumask, cpu_possible_mask); 7312 if (!cpumask_empty(cpumask)) { 7313 ret = 0; 7314 apply_wqattrs_lock(); 7315 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 7316 ret = workqueue_apply_unbound_cpumask(cpumask); 7317 if (!ret) 7318 cpumask_copy(wq_requested_unbound_cpumask, cpumask); 7319 apply_wqattrs_unlock(); 7320 } 7321 7322 return ret; 7323 } 7324 7325 static ssize_t __wq_cpumask_show(struct device *dev, 7326 struct device_attribute *attr, char *buf, cpumask_var_t mask) 7327 { 7328 int written; 7329 7330 mutex_lock(&wq_pool_mutex); 7331 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); 7332 mutex_unlock(&wq_pool_mutex); 7333 7334 return written; 7335 } 7336 7337 static ssize_t cpumask_requested_show(struct device *dev, 7338 struct device_attribute *attr, char *buf) 7339 { 7340 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); 7341 } 7342 static DEVICE_ATTR_RO(cpumask_requested); 7343 7344 static ssize_t cpumask_isolated_show(struct device *dev, 7345 struct device_attribute *attr, char *buf) 7346 { 7347 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); 7348 } 7349 static DEVICE_ATTR_RO(cpumask_isolated); 7350 7351 static ssize_t cpumask_show(struct device *dev, 7352 struct device_attribute *attr, char *buf) 7353 { 7354 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); 7355 } 7356 7357 static ssize_t cpumask_store(struct device *dev, 7358 struct device_attribute *attr, const char *buf, size_t count) 7359 { 7360 cpumask_var_t cpumask; 7361 int ret; 7362 7363 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 7364 return -ENOMEM; 7365 7366 ret = cpumask_parse(buf, cpumask); 7367 if (!ret) 7368 ret = workqueue_set_unbound_cpumask(cpumask); 7369 7370 free_cpumask_var(cpumask); 7371 return ret ? ret : count; 7372 } 7373 static DEVICE_ATTR_RW(cpumask); 7374 7375 static struct attribute *wq_sysfs_cpumask_attrs[] = { 7376 &dev_attr_cpumask.attr, 7377 &dev_attr_cpumask_requested.attr, 7378 &dev_attr_cpumask_isolated.attr, 7379 NULL, 7380 }; 7381 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); 7382 7383 static int __init wq_sysfs_init(void) 7384 { 7385 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups); 7386 } 7387 core_initcall(wq_sysfs_init); 7388 7389 static void wq_device_release(struct device *dev) 7390 { 7391 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7392 7393 kfree(wq_dev); 7394 } 7395 7396 /** 7397 * workqueue_sysfs_register - make a workqueue visible in sysfs 7398 * @wq: the workqueue to register 7399 * 7400 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 7401 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 7402 * which is the preferred method. 7403 * 7404 * Workqueue user should use this function directly iff it wants to apply 7405 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 7406 * apply_workqueue_attrs() may race against userland updating the 7407 * attributes. 7408 * 7409 * Return: 0 on success, -errno on failure. 7410 */ 7411 int workqueue_sysfs_register(struct workqueue_struct *wq) 7412 { 7413 struct wq_device *wq_dev; 7414 int ret; 7415 7416 /* 7417 * Adjusting max_active breaks ordering guarantee. Disallow exposing 7418 * ordered workqueues. 7419 */ 7420 if (WARN_ON(wq->flags & __WQ_ORDERED)) 7421 return -EINVAL; 7422 7423 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 7424 if (!wq_dev) 7425 return -ENOMEM; 7426 7427 wq_dev->wq = wq; 7428 wq_dev->dev.bus = &wq_subsys; 7429 wq_dev->dev.release = wq_device_release; 7430 dev_set_name(&wq_dev->dev, "%s", wq->name); 7431 7432 /* 7433 * unbound_attrs are created separately. Suppress uevent until 7434 * everything is ready. 7435 */ 7436 dev_set_uevent_suppress(&wq_dev->dev, true); 7437 7438 ret = device_register(&wq_dev->dev); 7439 if (ret) { 7440 put_device(&wq_dev->dev); 7441 wq->wq_dev = NULL; 7442 return ret; 7443 } 7444 7445 if (wq->flags & WQ_UNBOUND) { 7446 struct device_attribute *attr; 7447 7448 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 7449 ret = device_create_file(&wq_dev->dev, attr); 7450 if (ret) { 7451 device_unregister(&wq_dev->dev); 7452 wq->wq_dev = NULL; 7453 return ret; 7454 } 7455 } 7456 } 7457 7458 dev_set_uevent_suppress(&wq_dev->dev, false); 7459 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 7460 return 0; 7461 } 7462 7463 /** 7464 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 7465 * @wq: the workqueue to unregister 7466 * 7467 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 7468 */ 7469 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 7470 { 7471 struct wq_device *wq_dev = wq->wq_dev; 7472 7473 if (!wq->wq_dev) 7474 return; 7475 7476 wq->wq_dev = NULL; 7477 device_unregister(&wq_dev->dev); 7478 } 7479 #else /* CONFIG_SYSFS */ 7480 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 7481 #endif /* CONFIG_SYSFS */ 7482 7483 /* 7484 * Workqueue watchdog. 7485 * 7486 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 7487 * flush dependency, a concurrency managed work item which stays RUNNING 7488 * indefinitely. Workqueue stalls can be very difficult to debug as the 7489 * usual warning mechanisms don't trigger and internal workqueue state is 7490 * largely opaque. 7491 * 7492 * Workqueue watchdog monitors all worker pools periodically and dumps 7493 * state if some pools failed to make forward progress for a while where 7494 * forward progress is defined as the first item on ->worklist changing. 7495 * 7496 * This mechanism is controlled through the kernel parameter 7497 * "workqueue.watchdog_thresh" which can be updated at runtime through the 7498 * corresponding sysfs parameter file. 7499 */ 7500 #ifdef CONFIG_WQ_WATCHDOG 7501 7502 static unsigned long wq_watchdog_thresh = 30; 7503 static struct timer_list wq_watchdog_timer; 7504 7505 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 7506 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 7507 7508 static unsigned int wq_panic_on_stall; 7509 module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644); 7510 7511 /* 7512 * Show workers that might prevent the processing of pending work items. 7513 * The only candidates are CPU-bound workers in the running state. 7514 * Pending work items should be handled by another idle worker 7515 * in all other situations. 7516 */ 7517 static void show_cpu_pool_hog(struct worker_pool *pool) 7518 { 7519 struct worker *worker; 7520 unsigned long irq_flags; 7521 int bkt; 7522 7523 raw_spin_lock_irqsave(&pool->lock, irq_flags); 7524 7525 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 7526 if (task_is_running(worker->task)) { 7527 /* 7528 * Defer printing to avoid deadlocks in console 7529 * drivers that queue work while holding locks 7530 * also taken in their write paths. 7531 */ 7532 printk_deferred_enter(); 7533 7534 pr_info("pool %d:\n", pool->id); 7535 sched_show_task(worker->task); 7536 7537 printk_deferred_exit(); 7538 } 7539 } 7540 7541 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 7542 } 7543 7544 static void show_cpu_pools_hogs(void) 7545 { 7546 struct worker_pool *pool; 7547 int pi; 7548 7549 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 7550 7551 rcu_read_lock(); 7552 7553 for_each_pool(pool, pi) { 7554 if (pool->cpu_stall) 7555 show_cpu_pool_hog(pool); 7556 7557 } 7558 7559 rcu_read_unlock(); 7560 } 7561 7562 static void panic_on_wq_watchdog(void) 7563 { 7564 static unsigned int wq_stall; 7565 7566 if (wq_panic_on_stall) { 7567 wq_stall++; 7568 BUG_ON(wq_stall >= wq_panic_on_stall); 7569 } 7570 } 7571 7572 static void wq_watchdog_reset_touched(void) 7573 { 7574 int cpu; 7575 7576 wq_watchdog_touched = jiffies; 7577 for_each_possible_cpu(cpu) 7578 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 7579 } 7580 7581 static void wq_watchdog_timer_fn(struct timer_list *unused) 7582 { 7583 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7584 bool lockup_detected = false; 7585 bool cpu_pool_stall = false; 7586 unsigned long now = jiffies; 7587 struct worker_pool *pool; 7588 int pi; 7589 7590 if (!thresh) 7591 return; 7592 7593 for_each_pool(pool, pi) { 7594 unsigned long pool_ts, touched, ts; 7595 7596 pool->cpu_stall = false; 7597 if (list_empty(&pool->worklist)) 7598 continue; 7599 7600 /* 7601 * If a virtual machine is stopped by the host it can look to 7602 * the watchdog like a stall. 7603 */ 7604 kvm_check_and_clear_guest_paused(); 7605 7606 /* get the latest of pool and touched timestamps */ 7607 if (pool->cpu >= 0) 7608 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 7609 else 7610 touched = READ_ONCE(wq_watchdog_touched); 7611 pool_ts = READ_ONCE(pool->watchdog_ts); 7612 7613 if (time_after(pool_ts, touched)) 7614 ts = pool_ts; 7615 else 7616 ts = touched; 7617 7618 /* did we stall? */ 7619 if (time_after(now, ts + thresh)) { 7620 lockup_detected = true; 7621 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) { 7622 pool->cpu_stall = true; 7623 cpu_pool_stall = true; 7624 } 7625 pr_emerg("BUG: workqueue lockup - pool"); 7626 pr_cont_pool_info(pool); 7627 pr_cont(" stuck for %us!\n", 7628 jiffies_to_msecs(now - pool_ts) / 1000); 7629 } 7630 7631 7632 } 7633 7634 if (lockup_detected) 7635 show_all_workqueues(); 7636 7637 if (cpu_pool_stall) 7638 show_cpu_pools_hogs(); 7639 7640 if (lockup_detected) 7641 panic_on_wq_watchdog(); 7642 7643 wq_watchdog_reset_touched(); 7644 mod_timer(&wq_watchdog_timer, jiffies + thresh); 7645 } 7646 7647 notrace void wq_watchdog_touch(int cpu) 7648 { 7649 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7650 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched); 7651 unsigned long now = jiffies; 7652 7653 if (cpu >= 0) 7654 per_cpu(wq_watchdog_touched_cpu, cpu) = now; 7655 else 7656 WARN_ONCE(1, "%s should be called with valid CPU", __func__); 7657 7658 /* Don't unnecessarily store to global cacheline */ 7659 if (time_after(now, touch_ts + thresh / 4)) 7660 WRITE_ONCE(wq_watchdog_touched, jiffies); 7661 } 7662 7663 static void wq_watchdog_set_thresh(unsigned long thresh) 7664 { 7665 wq_watchdog_thresh = 0; 7666 timer_delete_sync(&wq_watchdog_timer); 7667 7668 if (thresh) { 7669 wq_watchdog_thresh = thresh; 7670 wq_watchdog_reset_touched(); 7671 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 7672 } 7673 } 7674 7675 static int wq_watchdog_param_set_thresh(const char *val, 7676 const struct kernel_param *kp) 7677 { 7678 unsigned long thresh; 7679 int ret; 7680 7681 ret = kstrtoul(val, 0, &thresh); 7682 if (ret) 7683 return ret; 7684 7685 if (system_percpu_wq) 7686 wq_watchdog_set_thresh(thresh); 7687 else 7688 wq_watchdog_thresh = thresh; 7689 7690 return 0; 7691 } 7692 7693 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 7694 .set = wq_watchdog_param_set_thresh, 7695 .get = param_get_ulong, 7696 }; 7697 7698 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 7699 0644); 7700 7701 static void wq_watchdog_init(void) 7702 { 7703 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 7704 wq_watchdog_set_thresh(wq_watchdog_thresh); 7705 } 7706 7707 #else /* CONFIG_WQ_WATCHDOG */ 7708 7709 static inline void wq_watchdog_init(void) { } 7710 7711 #endif /* CONFIG_WQ_WATCHDOG */ 7712 7713 static void bh_pool_kick_normal(struct irq_work *irq_work) 7714 { 7715 raise_softirq_irqoff(TASKLET_SOFTIRQ); 7716 } 7717 7718 static void bh_pool_kick_highpri(struct irq_work *irq_work) 7719 { 7720 raise_softirq_irqoff(HI_SOFTIRQ); 7721 } 7722 7723 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) 7724 { 7725 if (!cpumask_intersects(wq_unbound_cpumask, mask)) { 7726 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", 7727 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); 7728 return; 7729 } 7730 7731 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); 7732 } 7733 7734 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice) 7735 { 7736 BUG_ON(init_worker_pool(pool)); 7737 pool->cpu = cpu; 7738 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 7739 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 7740 pool->attrs->nice = nice; 7741 pool->attrs->affn_strict = true; 7742 pool->node = cpu_to_node(cpu); 7743 7744 /* alloc pool ID */ 7745 mutex_lock(&wq_pool_mutex); 7746 BUG_ON(worker_pool_assign_id(pool)); 7747 mutex_unlock(&wq_pool_mutex); 7748 } 7749 7750 /** 7751 * workqueue_init_early - early init for workqueue subsystem 7752 * 7753 * This is the first step of three-staged workqueue subsystem initialization and 7754 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 7755 * up. It sets up all the data structures and system workqueues and allows early 7756 * boot code to create workqueues and queue/cancel work items. Actual work item 7757 * execution starts only after kthreads can be created and scheduled right 7758 * before early initcalls. 7759 */ 7760 void __init workqueue_init_early(void) 7761 { 7762 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 7763 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 7764 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal, 7765 bh_pool_kick_highpri }; 7766 int i, cpu; 7767 7768 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 7769 7770 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL)); 7771 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 7772 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); 7773 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); 7774 7775 cpumask_copy(wq_online_cpumask, cpu_online_mask); 7776 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 7777 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ)); 7778 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); 7779 if (!cpumask_empty(&wq_cmdline_cpumask)) 7780 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); 7781 7782 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); 7783 cpumask_andnot(wq_isolated_cpumask, cpu_possible_mask, 7784 housekeeping_cpumask(HK_TYPE_DOMAIN)); 7785 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 7786 7787 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs(); 7788 BUG_ON(!unbound_wq_update_pwq_attrs_buf); 7789 7790 /* 7791 * If nohz_full is enabled, set power efficient workqueue as unbound. 7792 * This allows workqueue items to be moved to HK CPUs. 7793 */ 7794 if (housekeeping_enabled(HK_TYPE_TICK)) 7795 wq_power_efficient = true; 7796 7797 /* initialize WQ_AFFN_SYSTEM pods */ 7798 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 7799 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 7800 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 7801 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 7802 7803 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 7804 7805 pt->nr_pods = 1; 7806 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 7807 pt->pod_node[0] = NUMA_NO_NODE; 7808 pt->cpu_pod[0] = 0; 7809 7810 /* initialize BH and CPU pools */ 7811 for_each_possible_cpu(cpu) { 7812 struct worker_pool *pool; 7813 7814 i = 0; 7815 for_each_bh_worker_pool(pool, cpu) { 7816 init_cpu_worker_pool(pool, cpu, std_nice[i]); 7817 pool->flags |= POOL_BH; 7818 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]); 7819 i++; 7820 } 7821 7822 i = 0; 7823 for_each_cpu_worker_pool(pool, cpu) 7824 init_cpu_worker_pool(pool, cpu, std_nice[i++]); 7825 } 7826 7827 /* create default unbound and ordered wq attrs */ 7828 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 7829 struct workqueue_attrs *attrs; 7830 7831 BUG_ON(!(attrs = alloc_workqueue_attrs())); 7832 attrs->nice = std_nice[i]; 7833 unbound_std_wq_attrs[i] = attrs; 7834 7835 /* 7836 * An ordered wq should have only one pwq as ordering is 7837 * guaranteed by max_active which is enforced by pwqs. 7838 */ 7839 BUG_ON(!(attrs = alloc_workqueue_attrs())); 7840 attrs->nice = std_nice[i]; 7841 attrs->ordered = true; 7842 ordered_wq_attrs[i] = attrs; 7843 } 7844 7845 system_wq = alloc_workqueue("events", WQ_PERCPU, 0); 7846 system_percpu_wq = alloc_workqueue("events", WQ_PERCPU, 0); 7847 system_highpri_wq = alloc_workqueue("events_highpri", 7848 WQ_HIGHPRI | WQ_PERCPU, 0); 7849 system_long_wq = alloc_workqueue("events_long", WQ_PERCPU, 0); 7850 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE); 7851 system_dfl_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE); 7852 system_freezable_wq = alloc_workqueue("events_freezable", 7853 WQ_FREEZABLE | WQ_PERCPU, 0); 7854 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 7855 WQ_POWER_EFFICIENT | WQ_PERCPU, 0); 7856 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", 7857 WQ_FREEZABLE | WQ_POWER_EFFICIENT | WQ_PERCPU, 0); 7858 system_bh_wq = alloc_workqueue("events_bh", WQ_BH | WQ_PERCPU, 0); 7859 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri", 7860 WQ_BH | WQ_HIGHPRI | WQ_PERCPU, 0); 7861 BUG_ON(!system_wq || !system_percpu_wq|| !system_highpri_wq || !system_long_wq || 7862 !system_unbound_wq || !system_freezable_wq || !system_dfl_wq || 7863 !system_power_efficient_wq || 7864 !system_freezable_power_efficient_wq || 7865 !system_bh_wq || !system_bh_highpri_wq); 7866 } 7867 7868 static void __init wq_cpu_intensive_thresh_init(void) 7869 { 7870 unsigned long thresh; 7871 unsigned long bogo; 7872 7873 pwq_release_worker = kthread_run_worker(0, "pool_workqueue_release"); 7874 BUG_ON(IS_ERR(pwq_release_worker)); 7875 7876 /* if the user set it to a specific value, keep it */ 7877 if (wq_cpu_intensive_thresh_us != ULONG_MAX) 7878 return; 7879 7880 /* 7881 * The default of 10ms is derived from the fact that most modern (as of 7882 * 2023) processors can do a lot in 10ms and that it's just below what 7883 * most consider human-perceivable. However, the kernel also runs on a 7884 * lot slower CPUs including microcontrollers where the threshold is way 7885 * too low. 7886 * 7887 * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 7888 * This is by no means accurate but it doesn't have to be. The mechanism 7889 * is still useful even when the threshold is fully scaled up. Also, as 7890 * the reports would usually be applicable to everyone, some machines 7891 * operating on longer thresholds won't significantly diminish their 7892 * usefulness. 7893 */ 7894 thresh = 10 * USEC_PER_MSEC; 7895 7896 /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 7897 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 7898 if (bogo < 4000) 7899 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 7900 7901 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 7902 loops_per_jiffy, bogo, thresh); 7903 7904 wq_cpu_intensive_thresh_us = thresh; 7905 } 7906 7907 /** 7908 * workqueue_init - bring workqueue subsystem fully online 7909 * 7910 * This is the second step of three-staged workqueue subsystem initialization 7911 * and invoked as soon as kthreads can be created and scheduled. Workqueues have 7912 * been created and work items queued on them, but there are no kworkers 7913 * executing the work items yet. Populate the worker pools with the initial 7914 * workers and enable future kworker creations. 7915 */ 7916 void __init workqueue_init(void) 7917 { 7918 struct workqueue_struct *wq; 7919 struct worker_pool *pool; 7920 int cpu, bkt; 7921 7922 wq_cpu_intensive_thresh_init(); 7923 7924 mutex_lock(&wq_pool_mutex); 7925 7926 /* 7927 * Per-cpu pools created earlier could be missing node hint. Fix them 7928 * up. Also, create a rescuer for workqueues that requested it. 7929 */ 7930 for_each_possible_cpu(cpu) { 7931 for_each_bh_worker_pool(pool, cpu) 7932 pool->node = cpu_to_node(cpu); 7933 for_each_cpu_worker_pool(pool, cpu) 7934 pool->node = cpu_to_node(cpu); 7935 } 7936 7937 list_for_each_entry(wq, &workqueues, list) { 7938 WARN(init_rescuer(wq), 7939 "workqueue: failed to create early rescuer for %s", 7940 wq->name); 7941 } 7942 7943 mutex_unlock(&wq_pool_mutex); 7944 7945 /* 7946 * Create the initial workers. A BH pool has one pseudo worker that 7947 * represents the shared BH execution context and thus doesn't get 7948 * affected by hotplug events. Create the BH pseudo workers for all 7949 * possible CPUs here. 7950 */ 7951 for_each_possible_cpu(cpu) 7952 for_each_bh_worker_pool(pool, cpu) 7953 BUG_ON(!create_worker(pool)); 7954 7955 for_each_online_cpu(cpu) { 7956 for_each_cpu_worker_pool(pool, cpu) { 7957 pool->flags &= ~POOL_DISASSOCIATED; 7958 BUG_ON(!create_worker(pool)); 7959 } 7960 } 7961 7962 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 7963 BUG_ON(!create_worker(pool)); 7964 7965 wq_online = true; 7966 wq_watchdog_init(); 7967 } 7968 7969 /* 7970 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 7971 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 7972 * and consecutive pod ID. The rest of @pt is initialized accordingly. 7973 */ 7974 static void __init init_pod_type(struct wq_pod_type *pt, 7975 bool (*cpus_share_pod)(int, int)) 7976 { 7977 int cur, pre, cpu, pod; 7978 7979 pt->nr_pods = 0; 7980 7981 /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 7982 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 7983 BUG_ON(!pt->cpu_pod); 7984 7985 for_each_possible_cpu(cur) { 7986 for_each_possible_cpu(pre) { 7987 if (pre >= cur) { 7988 pt->cpu_pod[cur] = pt->nr_pods++; 7989 break; 7990 } 7991 if (cpus_share_pod(cur, pre)) { 7992 pt->cpu_pod[cur] = pt->cpu_pod[pre]; 7993 break; 7994 } 7995 } 7996 } 7997 7998 /* init the rest to match @pt->cpu_pod[] */ 7999 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 8000 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 8001 BUG_ON(!pt->pod_cpus || !pt->pod_node); 8002 8003 for (pod = 0; pod < pt->nr_pods; pod++) 8004 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 8005 8006 for_each_possible_cpu(cpu) { 8007 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 8008 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 8009 } 8010 } 8011 8012 static bool __init cpus_dont_share(int cpu0, int cpu1) 8013 { 8014 return false; 8015 } 8016 8017 static bool __init cpus_share_smt(int cpu0, int cpu1) 8018 { 8019 #ifdef CONFIG_SCHED_SMT 8020 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 8021 #else 8022 return false; 8023 #endif 8024 } 8025 8026 static bool __init cpus_share_numa(int cpu0, int cpu1) 8027 { 8028 return cpu_to_node(cpu0) == cpu_to_node(cpu1); 8029 } 8030 8031 /** 8032 * workqueue_init_topology - initialize CPU pods for unbound workqueues 8033 * 8034 * This is the third step of three-staged workqueue subsystem initialization and 8035 * invoked after SMP and topology information are fully initialized. It 8036 * initializes the unbound CPU pods accordingly. 8037 */ 8038 void __init workqueue_init_topology(void) 8039 { 8040 struct workqueue_struct *wq; 8041 int cpu; 8042 8043 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 8044 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 8045 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 8046 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 8047 8048 wq_topo_initialized = true; 8049 8050 mutex_lock(&wq_pool_mutex); 8051 8052 /* 8053 * Workqueues allocated earlier would have all CPUs sharing the default 8054 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue 8055 * and CPU combinations to apply per-pod sharing. 8056 */ 8057 list_for_each_entry(wq, &workqueues, list) { 8058 for_each_online_cpu(cpu) 8059 unbound_wq_update_pwq(wq, cpu); 8060 if (wq->flags & WQ_UNBOUND) { 8061 mutex_lock(&wq->mutex); 8062 wq_update_node_max_active(wq, -1); 8063 mutex_unlock(&wq->mutex); 8064 } 8065 } 8066 8067 mutex_unlock(&wq_pool_mutex); 8068 } 8069 8070 void __warn_flushing_systemwide_wq(void) 8071 { 8072 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 8073 dump_stack(); 8074 } 8075 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 8076 8077 static int __init workqueue_unbound_cpus_setup(char *str) 8078 { 8079 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 8080 cpumask_clear(&wq_cmdline_cpumask); 8081 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 8082 } 8083 8084 return 1; 8085 } 8086 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 8087