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/device/devres.h> 45 #include <linux/lockdep.h> 46 #include <linux/idr.h> 47 #include <linux/jhash.h> 48 #include <linux/hashtable.h> 49 #include <linux/rculist.h> 50 #include <linux/nodemask.h> 51 #include <linux/moduleparam.h> 52 #include <linux/uaccess.h> 53 #include <linux/sched/isolation.h> 54 #include <linux/sched/debug.h> 55 #include <linux/nmi.h> 56 #include <linux/kvm_para.h> 57 #include <linux/delay.h> 58 #include <linux/irq_work.h> 59 60 #include "workqueue_internal.h" 61 62 enum worker_pool_flags { 63 /* 64 * worker_pool flags 65 * 66 * A bound pool is either associated or disassociated with its CPU. 67 * While associated (!DISASSOCIATED), all workers are bound to the 68 * CPU and none has %WORKER_UNBOUND set and concurrency management 69 * is in effect. 70 * 71 * While DISASSOCIATED, the cpu may be offline and all workers have 72 * %WORKER_UNBOUND set and concurrency management disabled, and may 73 * be executing on any CPU. The pool behaves as an unbound one. 74 * 75 * Note that DISASSOCIATED should be flipped only while holding 76 * wq_pool_attach_mutex to avoid changing binding state while 77 * worker_attach_to_pool() is in progress. 78 * 79 * As there can only be one concurrent BH execution context per CPU, a 80 * BH pool is per-CPU and always DISASSOCIATED. 81 */ 82 POOL_BH = 1 << 0, /* is a BH pool */ 83 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */ 84 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 85 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */ 86 }; 87 88 enum worker_flags { 89 /* worker flags */ 90 WORKER_DIE = 1 << 1, /* die die die */ 91 WORKER_IDLE = 1 << 2, /* is idle */ 92 WORKER_PREP = 1 << 3, /* preparing to run works */ 93 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 94 WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 95 WORKER_REBOUND = 1 << 8, /* worker was rebound */ 96 97 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 98 WORKER_UNBOUND | WORKER_REBOUND, 99 }; 100 101 enum work_cancel_flags { 102 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */ 103 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */ 104 }; 105 106 enum wq_internal_consts { 107 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 108 109 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 110 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 111 112 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 113 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 114 115 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 116 /* call for help after 10ms 117 (min two ticks) */ 118 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 119 CREATE_COOLDOWN = HZ, /* time to breath after fail */ 120 121 RESCUER_BATCH = 16, /* process items per turn */ 122 123 /* 124 * Rescue workers are used only on emergencies and shared by 125 * all cpus. Give MIN_NICE. 126 */ 127 RESCUER_NICE_LEVEL = MIN_NICE, 128 HIGHPRI_NICE_LEVEL = MIN_NICE, 129 130 WQ_NAME_LEN = 32, 131 WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */ 132 }; 133 134 /* Layout of shards within one LLC pod */ 135 struct llc_shard_layout { 136 int nr_large_shards; /* number of large shards (cores_per_shard + 1) */ 137 int cores_per_shard; /* base number of cores per default shard */ 138 int nr_shards; /* total number of shards */ 139 /* nr_default shards = (nr_shards - nr_large_shards) */ 140 }; 141 142 /* 143 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and 144 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because 145 * msecs_to_jiffies() can't be an initializer. 146 */ 147 #define BH_WORKER_JIFFIES msecs_to_jiffies(2) 148 #define BH_WORKER_RESTARTS 10 149 150 /* 151 * Structure fields follow one of the following exclusion rules. 152 * 153 * I: Modifiable by initialization/destruction paths and read-only for 154 * everyone else. 155 * 156 * P: Preemption protected. Disabling preemption is enough and should 157 * only be modified and accessed from the local cpu. 158 * 159 * L: pool->lock protected. Access with pool->lock held. 160 * 161 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for 162 * reads. 163 * 164 * K: Only modified by worker while holding pool->lock. Can be safely read by 165 * self, while holding pool->lock or from IRQ context if %current is the 166 * kworker. 167 * 168 * S: Only modified by worker self. 169 * 170 * A: wq_pool_attach_mutex protected. 171 * 172 * PL: wq_pool_mutex protected. 173 * 174 * PR: wq_pool_mutex protected for writes. RCU protected for reads. 175 * 176 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 177 * 178 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 179 * RCU for reads. 180 * 181 * WQ: wq->mutex protected. 182 * 183 * WR: wq->mutex protected for writes. RCU protected for reads. 184 * 185 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read 186 * with READ_ONCE() without locking. 187 * 188 * MD: wq_mayday_lock protected. 189 * 190 * WD: Used internally by the watchdog. 191 */ 192 193 /* struct worker is defined in workqueue_internal.h */ 194 195 struct worker_pool { 196 raw_spinlock_t lock; /* the pool lock */ 197 int cpu; /* I: the associated cpu */ 198 int node; /* I: the associated node ID */ 199 int id; /* I: pool ID */ 200 unsigned int flags; /* L: flags */ 201 202 unsigned long last_progress_ts; /* L: last forward progress timestamp */ 203 bool cpu_stall; /* WD: stalled cpu bound pool */ 204 205 /* 206 * The counter is incremented in a process context on the associated CPU 207 * w/ preemption disabled, and decremented or reset in the same context 208 * but w/ pool->lock held. The readers grab pool->lock and are 209 * guaranteed to see if the counter reached zero. 210 */ 211 int nr_running; 212 213 struct list_head worklist; /* L: list of pending works */ 214 215 int nr_workers; /* L: total number of workers */ 216 int nr_idle; /* L: currently idle workers */ 217 218 struct list_head idle_list; /* L: list of idle workers */ 219 struct timer_list idle_timer; /* L: worker idle timeout */ 220 struct work_struct idle_cull_work; /* L: worker idle cleanup */ 221 222 struct timer_list mayday_timer; /* L: SOS timer for workers */ 223 224 /* a workers is either on busy_hash or idle_list, or the manager */ 225 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 226 /* L: hash of busy workers */ 227 228 struct worker *manager; /* L: purely informational */ 229 struct list_head workers; /* A: attached workers */ 230 231 struct ida worker_ida; /* worker IDs for task name */ 232 233 struct workqueue_attrs *attrs; /* I: worker attributes */ 234 struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 235 int refcnt; /* PL: refcnt for unbound pools */ 236 #ifdef CONFIG_PREEMPT_RT 237 spinlock_t cb_lock; /* BH worker cancel lock */ 238 #endif 239 /* 240 * Destruction of pool is RCU protected to allow dereferences 241 * from get_work_pool(). 242 */ 243 struct rcu_head rcu; 244 }; 245 246 /* 247 * Per-pool_workqueue statistics. These can be monitored using 248 * tools/workqueue/wq_monitor.py. 249 */ 250 enum pool_workqueue_stats { 251 PWQ_STAT_STARTED, /* work items started execution */ 252 PWQ_STAT_COMPLETED, /* work items completed execution */ 253 PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 254 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 255 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 256 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ 257 PWQ_STAT_MAYDAY, /* maydays to rescuer */ 258 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 259 260 PWQ_NR_STATS, 261 }; 262 263 /* 264 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT 265 * of work_struct->data are used for flags and the remaining high bits 266 * point to the pwq; thus, pwqs need to be aligned at two's power of the 267 * number of flag bits. 268 */ 269 struct pool_workqueue { 270 struct worker_pool *pool; /* I: the associated pool */ 271 struct workqueue_struct *wq; /* I: the owning workqueue */ 272 int work_color; /* L: current color */ 273 int flush_color; /* L: flushing color */ 274 int refcnt; /* L: reference count */ 275 int nr_in_flight[WORK_NR_COLORS]; 276 /* L: nr of in_flight works */ 277 bool plugged; /* L: execution suspended */ 278 279 /* 280 * nr_active management and WORK_STRUCT_INACTIVE: 281 * 282 * When pwq->nr_active >= max_active, new work item is queued to 283 * pwq->inactive_works instead of pool->worklist and marked with 284 * WORK_STRUCT_INACTIVE. 285 * 286 * All work items marked with WORK_STRUCT_INACTIVE do not participate in 287 * nr_active and all work items in pwq->inactive_works are marked with 288 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are 289 * in pwq->inactive_works. Some of them are ready to run in 290 * pool->worklist or worker->scheduled. Those work itmes are only struct 291 * wq_barrier which is used for flush_work() and should not participate 292 * in nr_active. For non-barrier work item, it is marked with 293 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 294 */ 295 int nr_active; /* L: nr of active works */ 296 struct list_head inactive_works; /* L: inactive works */ 297 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */ 298 struct list_head pwqs_node; /* WR: node on wq->pwqs */ 299 struct list_head mayday_node; /* MD: node on wq->maydays */ 300 struct work_struct mayday_cursor; /* L: cursor on pool->worklist */ 301 302 u64 stats[PWQ_NR_STATS]; 303 304 /* 305 * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 306 * and pwq_release_workfn() for details. pool_workqueue itself is also 307 * RCU protected so that the first pwq can be determined without 308 * grabbing wq->mutex. 309 */ 310 struct kthread_work release_work; 311 struct rcu_head rcu; 312 } __aligned(1 << WORK_STRUCT_PWQ_SHIFT); 313 314 /* 315 * Structure used to wait for workqueue flush. 316 */ 317 struct wq_flusher { 318 struct list_head list; /* WQ: list of flushers */ 319 int flush_color; /* WQ: flush color waiting for */ 320 struct completion done; /* flush completion */ 321 }; 322 323 struct wq_device; 324 325 /* 326 * Unlike in a per-cpu workqueue where max_active limits its concurrency level 327 * on each CPU, in an unbound workqueue, max_active applies to the whole system. 328 * As sharing a single nr_active across multiple sockets can be very expensive, 329 * the counting and enforcement is per NUMA node. 330 * 331 * The following struct is used to enforce per-node max_active. When a pwq wants 332 * to start executing a work item, it should increment ->nr using 333 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over 334 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish 335 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in 336 * round-robin order. 337 */ 338 struct wq_node_nr_active { 339 int max; /* per-node max_active */ 340 atomic_t nr; /* per-node nr_active */ 341 raw_spinlock_t lock; /* nests inside pool locks */ 342 struct list_head pending_pwqs; /* LN: pwqs with inactive works */ 343 }; 344 345 /* 346 * The externally visible workqueue. It relays the issued work items to 347 * the appropriate worker_pool through its pool_workqueues. 348 */ 349 struct workqueue_struct { 350 struct list_head pwqs; /* WR: all pwqs of this wq */ 351 struct list_head list; /* PR: list of all workqueues */ 352 353 struct mutex mutex; /* protects this wq */ 354 int work_color; /* WQ: current work color */ 355 int flush_color; /* WQ: current flush color */ 356 atomic_t nr_pwqs_to_flush; /* flush in progress */ 357 struct wq_flusher *first_flusher; /* WQ: first flusher */ 358 struct list_head flusher_queue; /* WQ: flush waiters */ 359 struct list_head flusher_overflow; /* WQ: flush overflow list */ 360 361 struct list_head maydays; /* MD: pwqs requesting rescue */ 362 struct worker *rescuer; /* MD: rescue worker */ 363 364 int nr_drainers; /* WQ: drain in progress */ 365 366 /* See alloc_workqueue() function comment for info on min/max_active */ 367 int max_active; /* WO: max active works */ 368 int min_active; /* WO: min active works */ 369 int saved_max_active; /* WQ: saved max_active */ 370 int saved_min_active; /* WQ: saved min_active */ 371 372 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 373 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */ 374 375 #ifdef CONFIG_SYSFS 376 struct wq_device *wq_dev; /* I: for sysfs interface */ 377 #endif 378 #ifdef CONFIG_LOCKDEP 379 char *lock_name; 380 struct lock_class_key key; 381 struct lockdep_map __lockdep_map; 382 struct lockdep_map *lockdep_map; 383 #endif 384 char name[WQ_NAME_LEN]; /* I: workqueue name */ 385 386 /* 387 * Destruction of workqueue_struct is RCU protected to allow walking 388 * the workqueues list without grabbing wq_pool_mutex. 389 * This is used to dump all workqueues from sysrq. 390 */ 391 struct rcu_head rcu; 392 393 /* hot fields used during command issue, aligned to cacheline */ 394 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 395 struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */ 396 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */ 397 }; 398 399 /* 400 * Each pod type describes how CPUs should be grouped for unbound workqueues. 401 * See the comment above workqueue_attrs->affn_scope. 402 */ 403 struct wq_pod_type { 404 int nr_pods; /* number of pods */ 405 cpumask_var_t *pod_cpus; /* pod -> cpus */ 406 int *pod_node; /* pod -> node */ 407 int *cpu_pod; /* cpu -> pod */ 408 }; 409 410 struct work_offq_data { 411 u32 pool_id; 412 u32 disable; 413 u32 flags; 414 }; 415 416 static const char * const wq_affn_names[WQ_AFFN_NR_TYPES] = { 417 [WQ_AFFN_DFL] = "default", 418 [WQ_AFFN_CPU] = "cpu", 419 [WQ_AFFN_SMT] = "smt", 420 [WQ_AFFN_CACHE] = "cache", 421 [WQ_AFFN_CACHE_SHARD] = "cache_shard", 422 [WQ_AFFN_NUMA] = "numa", 423 [WQ_AFFN_SYSTEM] = "system", 424 }; 425 426 /* 427 * Per-cpu work items which run for longer than the following threshold are 428 * automatically considered CPU intensive and excluded from concurrency 429 * management to prevent them from noticeably delaying other per-cpu work items. 430 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 431 * The actual value is initialized in wq_cpu_intensive_thresh_init(). 432 */ 433 static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 434 module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 435 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 436 static unsigned int wq_cpu_intensive_warning_thresh = 4; 437 module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644); 438 #endif 439 440 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 441 static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 442 module_param_named(power_efficient, wq_power_efficient, bool, 0444); 443 444 static unsigned int wq_cache_shard_size = 8; 445 module_param_named(cache_shard_size, wq_cache_shard_size, uint, 0444); 446 447 static bool wq_online; /* can kworkers be created yet? */ 448 static bool wq_topo_initialized __read_mostly = false; 449 450 static struct kmem_cache *pwq_cache; 451 452 static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; 453 static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE_SHARD; 454 455 /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 456 static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf; 457 458 static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 459 static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 460 static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 461 /* wait for manager to go away */ 462 static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 463 464 static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 465 static bool workqueue_freezing; /* PL: have wqs started freezing? */ 466 467 /* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */ 468 static cpumask_var_t wq_online_cpumask; 469 470 /* PL&A: allowable cpus for unbound wqs and work items */ 471 static cpumask_var_t wq_unbound_cpumask; 472 473 /* PL: user requested unbound cpumask via sysfs */ 474 static cpumask_var_t wq_requested_unbound_cpumask; 475 476 /* PL: isolated cpumask to be excluded from unbound cpumask */ 477 static cpumask_var_t wq_isolated_cpumask; 478 479 /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 480 static struct cpumask wq_cmdline_cpumask __initdata; 481 482 /* CPU where unbound work was last round robin scheduled from this CPU */ 483 static DEFINE_PER_CPU(int, wq_rr_cpu_last); 484 485 /* 486 * Local execution of unbound work items is no longer guaranteed. The 487 * following always forces round-robin CPU selection on unbound work items 488 * to uncover usages which depend on it. 489 */ 490 #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 491 static bool wq_debug_force_rr_cpu = true; 492 #else 493 static bool wq_debug_force_rr_cpu = false; 494 #endif 495 module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 496 497 /* to raise softirq for the BH worker pools on other CPUs */ 498 static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works); 499 500 /* the BH worker pools */ 501 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools); 502 503 /* the per-cpu worker pools */ 504 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 505 506 static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 507 508 /* PL: hash of all unbound pools keyed by pool->attrs */ 509 static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 510 511 /* I: attributes used when instantiating standard unbound pools on demand */ 512 static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 513 514 /* I: attributes used when instantiating ordered pools on demand */ 515 static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 516 517 /* 518 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 519 * process context while holding a pool lock. Bounce to a dedicated kthread 520 * worker to avoid A-A deadlocks. 521 */ 522 static struct kthread_worker *pwq_release_worker __ro_after_init; 523 524 struct workqueue_struct *system_wq __ro_after_init; 525 EXPORT_SYMBOL(system_wq); 526 struct workqueue_struct *system_percpu_wq __ro_after_init; 527 EXPORT_SYMBOL(system_percpu_wq); 528 struct workqueue_struct *system_highpri_wq __ro_after_init; 529 EXPORT_SYMBOL_GPL(system_highpri_wq); 530 struct workqueue_struct *system_long_wq __ro_after_init; 531 EXPORT_SYMBOL_GPL(system_long_wq); 532 struct workqueue_struct *system_unbound_wq __ro_after_init; 533 EXPORT_SYMBOL_GPL(system_unbound_wq); 534 struct workqueue_struct *system_dfl_wq __ro_after_init; 535 EXPORT_SYMBOL_GPL(system_dfl_wq); 536 struct workqueue_struct *system_freezable_wq __ro_after_init; 537 EXPORT_SYMBOL_GPL(system_freezable_wq); 538 struct workqueue_struct *system_power_efficient_wq __ro_after_init; 539 EXPORT_SYMBOL_GPL(system_power_efficient_wq); 540 struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init; 541 EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 542 struct workqueue_struct *system_bh_wq; 543 EXPORT_SYMBOL_GPL(system_bh_wq); 544 struct workqueue_struct *system_bh_highpri_wq; 545 EXPORT_SYMBOL_GPL(system_bh_highpri_wq); 546 struct workqueue_struct *system_dfl_long_wq __ro_after_init; 547 EXPORT_SYMBOL_GPL(system_dfl_long_wq); 548 549 static int worker_thread(void *__worker); 550 static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 551 static void show_pwq(struct pool_workqueue *pwq); 552 static void show_one_worker_pool(struct worker_pool *pool); 553 554 #define CREATE_TRACE_POINTS 555 #include <trace/events/workqueue.h> 556 557 #define assert_rcu_or_pool_mutex() \ 558 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \ 559 !lockdep_is_held(&wq_pool_mutex), \ 560 "RCU or wq_pool_mutex should be held") 561 562 #define for_each_bh_worker_pool(pool, cpu) \ 563 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \ 564 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 565 (pool)++) 566 567 #define for_each_cpu_worker_pool(pool, cpu) \ 568 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 569 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 570 (pool)++) 571 572 /** 573 * for_each_pool - iterate through all worker_pools in the system 574 * @pool: iteration cursor 575 * @pi: integer used for iteration 576 * 577 * This must be called either with wq_pool_mutex held or RCU read 578 * locked. If the pool needs to be used beyond the locking in effect, the 579 * caller is responsible for guaranteeing that the pool stays online. 580 * 581 * The if/else clause exists only for the lockdep assertion and can be 582 * ignored. 583 */ 584 #define for_each_pool(pool, pi) \ 585 idr_for_each_entry(&worker_pool_idr, pool, pi) \ 586 if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 587 else 588 589 /** 590 * for_each_pool_worker - iterate through all workers of a worker_pool 591 * @worker: iteration cursor 592 * @pool: worker_pool to iterate workers of 593 * 594 * This must be called with wq_pool_attach_mutex. 595 * 596 * The if/else clause exists only for the lockdep assertion and can be 597 * ignored. 598 */ 599 #define for_each_pool_worker(worker, pool) \ 600 list_for_each_entry((worker), &(pool)->workers, node) \ 601 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 602 else 603 604 /** 605 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 606 * @pwq: iteration cursor 607 * @wq: the target workqueue 608 * 609 * This must be called either with wq->mutex held or RCU read locked. 610 * If the pwq needs to be used beyond the locking in effect, the caller is 611 * responsible for guaranteeing that the pwq stays online. 612 * 613 * The if/else clause exists only for the lockdep assertion and can be 614 * ignored. 615 */ 616 #define for_each_pwq(pwq, wq) \ 617 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 618 lockdep_is_held(&(wq->mutex))) 619 620 #ifdef CONFIG_DEBUG_OBJECTS_WORK 621 622 static const struct debug_obj_descr work_debug_descr; 623 624 static void *work_debug_hint(void *addr) 625 { 626 return ((struct work_struct *) addr)->func; 627 } 628 629 static bool work_is_static_object(void *addr) 630 { 631 struct work_struct *work = addr; 632 633 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 634 } 635 636 /* 637 * fixup_init is called when: 638 * - an active object is initialized 639 */ 640 static bool work_fixup_init(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_init(work, &work_debug_descr); 648 return true; 649 default: 650 return false; 651 } 652 } 653 654 /* 655 * fixup_free is called when: 656 * - an active object is freed 657 */ 658 static bool work_fixup_free(void *addr, enum debug_obj_state state) 659 { 660 struct work_struct *work = addr; 661 662 switch (state) { 663 case ODEBUG_STATE_ACTIVE: 664 cancel_work_sync(work); 665 debug_object_free(work, &work_debug_descr); 666 return true; 667 default: 668 return false; 669 } 670 } 671 672 static const struct debug_obj_descr work_debug_descr = { 673 .name = "work_struct", 674 .debug_hint = work_debug_hint, 675 .is_static_object = work_is_static_object, 676 .fixup_init = work_fixup_init, 677 .fixup_free = work_fixup_free, 678 }; 679 680 static inline void debug_work_activate(struct work_struct *work) 681 { 682 debug_object_activate(work, &work_debug_descr); 683 } 684 685 static inline void debug_work_deactivate(struct work_struct *work) 686 { 687 debug_object_deactivate(work, &work_debug_descr); 688 } 689 690 void __init_work(struct work_struct *work, int onstack) 691 { 692 if (onstack) 693 debug_object_init_on_stack(work, &work_debug_descr); 694 else 695 debug_object_init(work, &work_debug_descr); 696 } 697 EXPORT_SYMBOL_GPL(__init_work); 698 699 void destroy_work_on_stack(struct work_struct *work) 700 { 701 debug_object_free(work, &work_debug_descr); 702 } 703 EXPORT_SYMBOL_GPL(destroy_work_on_stack); 704 705 void destroy_delayed_work_on_stack(struct delayed_work *work) 706 { 707 timer_destroy_on_stack(&work->timer); 708 debug_object_free(&work->work, &work_debug_descr); 709 } 710 EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 711 712 #else 713 static inline void debug_work_activate(struct work_struct *work) { } 714 static inline void debug_work_deactivate(struct work_struct *work) { } 715 #endif 716 717 /** 718 * worker_pool_assign_id - allocate ID and assign it to @pool 719 * @pool: the pool pointer of interest 720 * 721 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 722 * successfully, -errno on failure. 723 */ 724 static int worker_pool_assign_id(struct worker_pool *pool) 725 { 726 int ret; 727 728 lockdep_assert_held(&wq_pool_mutex); 729 730 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 731 GFP_KERNEL); 732 if (ret >= 0) { 733 pool->id = ret; 734 return 0; 735 } 736 return ret; 737 } 738 739 static struct pool_workqueue __rcu ** 740 unbound_pwq_slot(struct workqueue_struct *wq, int cpu) 741 { 742 if (cpu >= 0) 743 return per_cpu_ptr(wq->cpu_pwq, cpu); 744 else 745 return &wq->dfl_pwq; 746 } 747 748 /* @cpu < 0 for dfl_pwq */ 749 static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu) 750 { 751 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu), 752 lockdep_is_held(&wq_pool_mutex) || 753 lockdep_is_held(&wq->mutex)); 754 } 755 756 /** 757 * unbound_effective_cpumask - effective cpumask of an unbound workqueue 758 * @wq: workqueue of interest 759 * 760 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which 761 * is masked with wq_unbound_cpumask to determine the effective cpumask. The 762 * default pwq is always mapped to the pool with the current effective cpumask. 763 */ 764 static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq) 765 { 766 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask; 767 } 768 769 static unsigned int work_color_to_flags(int color) 770 { 771 return color << WORK_STRUCT_COLOR_SHIFT; 772 } 773 774 static int get_work_color(unsigned long work_data) 775 { 776 return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 777 ((1 << WORK_STRUCT_COLOR_BITS) - 1); 778 } 779 780 static int work_next_color(int color) 781 { 782 return (color + 1) % WORK_NR_COLORS; 783 } 784 785 static unsigned long pool_offq_flags(struct worker_pool *pool) 786 { 787 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0; 788 } 789 790 /* 791 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 792 * contain the pointer to the queued pwq. Once execution starts, the flag 793 * is cleared and the high bits contain OFFQ flags and pool ID. 794 * 795 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling() 796 * can be used to set the pwq, pool or clear work->data. These functions should 797 * only be called while the work is owned - ie. while the PENDING bit is set. 798 * 799 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 800 * corresponding to a work. Pool is available once the work has been 801 * queued anywhere after initialization until it is sync canceled. pwq is 802 * available only while the work item is queued. 803 */ 804 static inline void set_work_data(struct work_struct *work, unsigned long data) 805 { 806 WARN_ON_ONCE(!work_pending(work)); 807 atomic_long_set(&work->data, data | work_static(work)); 808 } 809 810 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 811 unsigned long flags) 812 { 813 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING | 814 WORK_STRUCT_PWQ | flags); 815 } 816 817 static void set_work_pool_and_keep_pending(struct work_struct *work, 818 int pool_id, unsigned long flags) 819 { 820 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | 821 WORK_STRUCT_PENDING | flags); 822 } 823 824 static void set_work_pool_and_clear_pending(struct work_struct *work, 825 int pool_id, unsigned long flags) 826 { 827 /* 828 * The following wmb is paired with the implied mb in 829 * test_and_set_bit(PENDING) and ensures all updates to @work made 830 * here are visible to and precede any updates by the next PENDING 831 * owner. 832 */ 833 smp_wmb(); 834 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) | 835 flags); 836 /* 837 * The following mb guarantees that previous clear of a PENDING bit 838 * will not be reordered with any speculative LOADS or STORES from 839 * work->current_func, which is executed afterwards. This possible 840 * reordering can lead to a missed execution on attempt to queue 841 * the same @work. E.g. consider this case: 842 * 843 * CPU#0 CPU#1 844 * ---------------------------- -------------------------------- 845 * 846 * 1 STORE event_indicated 847 * 2 queue_work_on() { 848 * 3 test_and_set_bit(PENDING) 849 * 4 } set_..._and_clear_pending() { 850 * 5 set_work_data() # clear bit 851 * 6 smp_mb() 852 * 7 work->current_func() { 853 * 8 LOAD event_indicated 854 * } 855 * 856 * Without an explicit full barrier speculative LOAD on line 8 can 857 * be executed before CPU#0 does STORE on line 1. If that happens, 858 * CPU#0 observes the PENDING bit is still set and new execution of 859 * a @work is not queued in a hope, that CPU#1 will eventually 860 * finish the queued @work. Meanwhile CPU#1 does not see 861 * event_indicated is set, because speculative LOAD was executed 862 * before actual STORE. 863 */ 864 smp_mb(); 865 } 866 867 static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 868 { 869 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK); 870 } 871 872 static struct pool_workqueue *get_work_pwq(struct work_struct *work) 873 { 874 unsigned long data = atomic_long_read(&work->data); 875 876 if (data & WORK_STRUCT_PWQ) 877 return work_struct_pwq(data); 878 else 879 return NULL; 880 } 881 882 /** 883 * get_work_pool - return the worker_pool a given work was associated with 884 * @work: the work item of interest 885 * 886 * Pools are created and destroyed under wq_pool_mutex, and allows read 887 * access under RCU read lock. As such, this function should be 888 * called under wq_pool_mutex or inside of a rcu_read_lock() region. 889 * 890 * All fields of the returned pool are accessible as long as the above 891 * mentioned locking is in effect. If the returned pool needs to be used 892 * beyond the critical section, the caller is responsible for ensuring the 893 * returned pool is and stays online. 894 * 895 * Return: The worker_pool @work was last associated with. %NULL if none. 896 */ 897 static struct worker_pool *get_work_pool(struct work_struct *work) 898 { 899 unsigned long data = atomic_long_read(&work->data); 900 int pool_id; 901 902 assert_rcu_or_pool_mutex(); 903 904 if (data & WORK_STRUCT_PWQ) 905 return work_struct_pwq(data)->pool; 906 907 pool_id = data >> WORK_OFFQ_POOL_SHIFT; 908 if (pool_id == WORK_OFFQ_POOL_NONE) 909 return NULL; 910 911 return idr_find(&worker_pool_idr, pool_id); 912 } 913 914 static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits) 915 { 916 return (v >> shift) & ((1U << bits) - 1); 917 } 918 919 static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data) 920 { 921 WARN_ON_ONCE(data & WORK_STRUCT_PWQ); 922 923 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT, 924 WORK_OFFQ_POOL_BITS); 925 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT, 926 WORK_OFFQ_DISABLE_BITS); 927 offqd->flags = data & WORK_OFFQ_FLAG_MASK; 928 } 929 930 static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd) 931 { 932 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) | 933 ((unsigned long)offqd->flags); 934 } 935 936 /* 937 * Policy functions. These define the policies on how the global worker 938 * pools are managed. Unless noted otherwise, these functions assume that 939 * they're being called with pool->lock held. 940 */ 941 942 /* 943 * Need to wake up a worker? Called from anything but currently 944 * running workers. 945 * 946 * Note that, because unbound workers never contribute to nr_running, this 947 * function will always return %true for unbound pools as long as the 948 * worklist isn't empty. 949 */ 950 static bool need_more_worker(struct worker_pool *pool) 951 { 952 return !list_empty(&pool->worklist) && !pool->nr_running; 953 } 954 955 /* Can I start working? Called from busy but !running workers. */ 956 static bool may_start_working(struct worker_pool *pool) 957 { 958 return pool->nr_idle; 959 } 960 961 /* Do I need to keep working? Called from currently running workers. */ 962 static bool keep_working(struct worker_pool *pool) 963 { 964 return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 965 } 966 967 /* Do we need a new worker? Called from manager. */ 968 static bool need_to_create_worker(struct worker_pool *pool) 969 { 970 return need_more_worker(pool) && !may_start_working(pool); 971 } 972 973 /* Do we have too many workers and should some go away? */ 974 static bool too_many_workers(struct worker_pool *pool) 975 { 976 bool managing = pool->flags & POOL_MANAGER_ACTIVE; 977 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 978 int nr_busy = pool->nr_workers - nr_idle; 979 980 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 981 } 982 983 /** 984 * worker_set_flags - set worker flags and adjust nr_running accordingly 985 * @worker: self 986 * @flags: flags to set 987 * 988 * Set @flags in @worker->flags and adjust nr_running accordingly. 989 */ 990 static inline void worker_set_flags(struct worker *worker, unsigned int flags) 991 { 992 struct worker_pool *pool = worker->pool; 993 994 lockdep_assert_held(&pool->lock); 995 996 /* If transitioning into NOT_RUNNING, adjust nr_running. */ 997 if ((flags & WORKER_NOT_RUNNING) && 998 !(worker->flags & WORKER_NOT_RUNNING)) { 999 pool->nr_running--; 1000 } 1001 1002 worker->flags |= flags; 1003 } 1004 1005 /** 1006 * worker_clr_flags - clear worker flags and adjust nr_running accordingly 1007 * @worker: self 1008 * @flags: flags to clear 1009 * 1010 * Clear @flags in @worker->flags and adjust nr_running accordingly. 1011 */ 1012 static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 1013 { 1014 struct worker_pool *pool = worker->pool; 1015 unsigned int oflags = worker->flags; 1016 1017 lockdep_assert_held(&pool->lock); 1018 1019 worker->flags &= ~flags; 1020 1021 /* 1022 * If transitioning out of NOT_RUNNING, increment nr_running. Note 1023 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 1024 * of multiple flags, not a single flag. 1025 */ 1026 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1027 if (!(worker->flags & WORKER_NOT_RUNNING)) 1028 pool->nr_running++; 1029 } 1030 1031 /* Return the first idle worker. Called with pool->lock held. */ 1032 static struct worker *first_idle_worker(struct worker_pool *pool) 1033 { 1034 if (unlikely(list_empty(&pool->idle_list))) 1035 return NULL; 1036 1037 return list_first_entry(&pool->idle_list, struct worker, entry); 1038 } 1039 1040 /** 1041 * worker_enter_idle - enter idle state 1042 * @worker: worker which is entering idle state 1043 * 1044 * @worker is entering idle state. Update stats and idle timer if 1045 * necessary. 1046 * 1047 * LOCKING: 1048 * raw_spin_lock_irq(pool->lock). 1049 */ 1050 static void worker_enter_idle(struct worker *worker) 1051 { 1052 struct worker_pool *pool = worker->pool; 1053 1054 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 1055 WARN_ON_ONCE(!list_empty(&worker->entry) && 1056 (worker->hentry.next || worker->hentry.pprev))) 1057 return; 1058 1059 /* can't use worker_set_flags(), also called from create_worker() */ 1060 worker->flags |= WORKER_IDLE; 1061 pool->nr_idle++; 1062 worker->last_active = jiffies; 1063 1064 /* idle_list is LIFO */ 1065 list_add(&worker->entry, &pool->idle_list); 1066 1067 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1068 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1069 1070 /* Sanity check nr_running. */ 1071 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 1072 } 1073 1074 /** 1075 * worker_leave_idle - leave idle state 1076 * @worker: worker which is leaving idle state 1077 * 1078 * @worker is leaving idle state. Update stats. 1079 * 1080 * LOCKING: 1081 * raw_spin_lock_irq(pool->lock). 1082 */ 1083 static void worker_leave_idle(struct worker *worker) 1084 { 1085 struct worker_pool *pool = worker->pool; 1086 1087 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 1088 return; 1089 worker_clr_flags(worker, WORKER_IDLE); 1090 pool->nr_idle--; 1091 list_del_init(&worker->entry); 1092 } 1093 1094 /** 1095 * find_worker_executing_work - find worker which is executing a work 1096 * @pool: pool of interest 1097 * @work: work to find worker for 1098 * 1099 * Find a worker which is executing @work on @pool by searching 1100 * @pool->busy_hash which is keyed by the address of @work. For a worker 1101 * to match, its current execution should match the address of @work and 1102 * its work function. This is to avoid unwanted dependency between 1103 * unrelated work executions through a work item being recycled while still 1104 * being executed. 1105 * 1106 * This is a bit tricky. A work item may be freed once its execution 1107 * starts and nothing prevents the freed area from being recycled for 1108 * another work item. If the same work item address ends up being reused 1109 * before the original execution finishes, workqueue will identify the 1110 * recycled work item as currently executing and make it wait until the 1111 * current execution finishes, introducing an unwanted dependency. 1112 * 1113 * This function checks the work item address and work function to avoid 1114 * false positives. Note that this isn't complete as one may construct a 1115 * work function which can introduce dependency onto itself through a 1116 * recycled work item. Well, if somebody wants to shoot oneself in the 1117 * foot that badly, there's only so much we can do, and if such deadlock 1118 * actually occurs, it should be easy to locate the culprit work function. 1119 * 1120 * CONTEXT: 1121 * raw_spin_lock_irq(pool->lock). 1122 * 1123 * Return: 1124 * Pointer to worker which is executing @work if found, %NULL 1125 * otherwise. 1126 */ 1127 static struct worker *find_worker_executing_work(struct worker_pool *pool, 1128 struct work_struct *work) 1129 { 1130 struct worker *worker; 1131 1132 hash_for_each_possible(pool->busy_hash, worker, hentry, 1133 (unsigned long)work) 1134 if (worker->current_work == work && 1135 worker->current_func == work->func) 1136 return worker; 1137 1138 return NULL; 1139 } 1140 1141 static void mayday_cursor_func(struct work_struct *work) 1142 { 1143 /* should not be processed, only for marking position */ 1144 BUG(); 1145 } 1146 1147 /** 1148 * move_linked_works - move linked works to a list 1149 * @work: start of series of works to be scheduled 1150 * @head: target list to append @work to 1151 * @nextp: out parameter for nested worklist walking 1152 * 1153 * Schedule linked works starting from @work to @head. Work series to be 1154 * scheduled starts at @work and includes any consecutive work with 1155 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on 1156 * @nextp. 1157 * 1158 * CONTEXT: 1159 * raw_spin_lock_irq(pool->lock). 1160 */ 1161 static void move_linked_works(struct work_struct *work, struct list_head *head, 1162 struct work_struct **nextp) 1163 { 1164 struct work_struct *n; 1165 1166 /* 1167 * Linked worklist will always end before the end of the list, 1168 * use NULL for list head. 1169 */ 1170 list_for_each_entry_safe_from(work, n, NULL, entry) { 1171 list_move_tail(&work->entry, head); 1172 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1173 break; 1174 } 1175 1176 /* 1177 * If we're already inside safe list traversal and have moved 1178 * multiple works to the scheduled queue, the next position 1179 * needs to be updated. 1180 */ 1181 if (nextp) 1182 *nextp = n; 1183 } 1184 1185 /** 1186 * assign_work - assign a work item and its linked work items to a worker 1187 * @work: work to assign 1188 * @worker: worker to assign to 1189 * @nextp: out parameter for nested worklist walking 1190 * 1191 * Assign @work and its linked work items to @worker. If @work is already being 1192 * executed by another worker in the same pool, it'll be punted there. 1193 * 1194 * If @nextp is not NULL, it's updated to point to the next work of the last 1195 * scheduled work. This allows assign_work() to be nested inside 1196 * list_for_each_entry_safe(). 1197 * 1198 * Returns %true if @work was successfully assigned to @worker. %false if @work 1199 * was punted to another worker already executing it. 1200 */ 1201 static bool assign_work(struct work_struct *work, struct worker *worker, 1202 struct work_struct **nextp) 1203 { 1204 struct worker_pool *pool = worker->pool; 1205 struct worker *collision; 1206 1207 lockdep_assert_held(&pool->lock); 1208 1209 /* The cursor work should not be processed */ 1210 if (unlikely(work->func == mayday_cursor_func)) { 1211 /* only worker_thread() can possibly take this branch */ 1212 WARN_ON_ONCE(worker->rescue_wq); 1213 if (nextp) 1214 *nextp = list_next_entry(work, entry); 1215 list_del_init(&work->entry); 1216 return false; 1217 } 1218 1219 /* 1220 * A single work shouldn't be executed concurrently by multiple workers. 1221 * __queue_work() ensures that @work doesn't jump to a different pool 1222 * while still running in the previous pool. Here, we should ensure that 1223 * @work is not executed concurrently by multiple workers from the same 1224 * pool. Check whether anyone is already processing the work. If so, 1225 * defer the work to the currently executing one. 1226 */ 1227 collision = find_worker_executing_work(pool, work); 1228 if (unlikely(collision)) { 1229 move_linked_works(work, &collision->scheduled, nextp); 1230 return false; 1231 } 1232 1233 move_linked_works(work, &worker->scheduled, nextp); 1234 return true; 1235 } 1236 1237 static struct irq_work *bh_pool_irq_work(struct worker_pool *pool) 1238 { 1239 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0; 1240 1241 return &per_cpu(bh_pool_irq_works, pool->cpu)[high]; 1242 } 1243 1244 static void kick_bh_pool(struct worker_pool *pool) 1245 { 1246 #ifdef CONFIG_SMP 1247 /* see drain_dead_softirq_workfn() for BH_DRAINING */ 1248 if (unlikely(pool->cpu != smp_processor_id() && 1249 !(pool->flags & POOL_BH_DRAINING))) { 1250 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu); 1251 return; 1252 } 1253 #endif 1254 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 1255 raise_softirq_irqoff(HI_SOFTIRQ); 1256 else 1257 raise_softirq_irqoff(TASKLET_SOFTIRQ); 1258 } 1259 1260 /** 1261 * kick_pool - wake up an idle worker if necessary 1262 * @pool: pool to kick 1263 * 1264 * @pool may have pending work items. Wake up worker if necessary. Returns 1265 * whether a worker was woken up. 1266 */ 1267 static bool kick_pool(struct worker_pool *pool) 1268 { 1269 struct worker *worker = first_idle_worker(pool); 1270 struct task_struct *p; 1271 1272 lockdep_assert_held(&pool->lock); 1273 1274 if (!need_more_worker(pool) || !worker) 1275 return false; 1276 1277 if (pool->flags & POOL_BH) { 1278 kick_bh_pool(pool); 1279 return true; 1280 } 1281 1282 p = worker->task; 1283 1284 #ifdef CONFIG_SMP 1285 /* 1286 * Idle @worker is about to execute @work and waking up provides an 1287 * opportunity to migrate @worker at a lower cost by setting the task's 1288 * wake_cpu field. Let's see if we want to move @worker to improve 1289 * execution locality. 1290 * 1291 * We're waking the worker that went idle the latest and there's some 1292 * chance that @worker is marked idle but hasn't gone off CPU yet. If 1293 * so, setting the wake_cpu won't do anything. As this is a best-effort 1294 * optimization and the race window is narrow, let's leave as-is for 1295 * now. If this becomes pronounced, we can skip over workers which are 1296 * still on cpu when picking an idle worker. 1297 * 1298 * If @pool has non-strict affinity, @worker might have ended up outside 1299 * its affinity scope. Repatriate. 1300 */ 1301 if (!pool->attrs->affn_strict && 1302 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { 1303 struct work_struct *work = list_first_entry(&pool->worklist, 1304 struct work_struct, entry); 1305 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask, 1306 cpu_online_mask); 1307 if (wake_cpu < nr_cpu_ids) { 1308 p->wake_cpu = wake_cpu; 1309 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; 1310 } 1311 } 1312 #endif 1313 wake_up_process(p); 1314 return true; 1315 } 1316 1317 #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 1318 1319 /* 1320 * Concurrency-managed per-cpu work items that hog CPU for longer than 1321 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 1322 * which prevents them from stalling other concurrency-managed work items. If a 1323 * work function keeps triggering this mechanism, it's likely that the work item 1324 * should be using an unbound workqueue instead. 1325 * 1326 * wq_cpu_intensive_report() tracks work functions which trigger such conditions 1327 * and report them so that they can be examined and converted to use unbound 1328 * workqueues as appropriate. To avoid flooding the console, each violating work 1329 * function is tracked and reported with exponential backoff. 1330 */ 1331 #define WCI_MAX_ENTS 128 1332 1333 struct wci_ent { 1334 work_func_t func; 1335 atomic64_t cnt; 1336 struct hlist_node hash_node; 1337 }; 1338 1339 static struct wci_ent wci_ents[WCI_MAX_ENTS]; 1340 static int wci_nr_ents; 1341 static DEFINE_RAW_SPINLOCK(wci_lock); 1342 static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 1343 1344 static struct wci_ent *wci_find_ent(work_func_t func) 1345 { 1346 struct wci_ent *ent; 1347 1348 hash_for_each_possible_rcu(wci_hash, ent, hash_node, 1349 (unsigned long)func) { 1350 if (ent->func == func) 1351 return ent; 1352 } 1353 return NULL; 1354 } 1355 1356 static void wq_cpu_intensive_report(work_func_t func) 1357 { 1358 struct wci_ent *ent; 1359 1360 restart: 1361 ent = wci_find_ent(func); 1362 if (ent) { 1363 u64 cnt; 1364 1365 /* 1366 * Start reporting from the warning_thresh and back off 1367 * exponentially. 1368 */ 1369 cnt = atomic64_inc_return_relaxed(&ent->cnt); 1370 if (wq_cpu_intensive_warning_thresh && 1371 cnt >= wq_cpu_intensive_warning_thresh && 1372 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh)) 1373 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 1374 ent->func, wq_cpu_intensive_thresh_us, 1375 atomic64_read(&ent->cnt)); 1376 return; 1377 } 1378 1379 /* 1380 * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 1381 * is exhausted, something went really wrong and we probably made enough 1382 * noise already. 1383 */ 1384 if (wci_nr_ents >= WCI_MAX_ENTS) 1385 return; 1386 1387 raw_spin_lock(&wci_lock); 1388 1389 if (wci_nr_ents >= WCI_MAX_ENTS) { 1390 raw_spin_unlock(&wci_lock); 1391 return; 1392 } 1393 1394 if (wci_find_ent(func)) { 1395 raw_spin_unlock(&wci_lock); 1396 goto restart; 1397 } 1398 1399 ent = &wci_ents[wci_nr_ents++]; 1400 ent->func = func; 1401 atomic64_set(&ent->cnt, 0); 1402 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 1403 1404 raw_spin_unlock(&wci_lock); 1405 1406 goto restart; 1407 } 1408 1409 #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1410 static void wq_cpu_intensive_report(work_func_t func) {} 1411 #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 1412 1413 /** 1414 * wq_worker_running - a worker is running again 1415 * @task: task waking up 1416 * 1417 * This function is called when a worker returns from schedule() 1418 */ 1419 void wq_worker_running(struct task_struct *task) 1420 { 1421 struct worker *worker = kthread_data(task); 1422 1423 if (!READ_ONCE(worker->sleeping)) 1424 return; 1425 1426 /* 1427 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 1428 * and the nr_running increment below, we may ruin the nr_running reset 1429 * and leave with an unexpected pool->nr_running == 1 on the newly unbound 1430 * pool. Protect against such race. 1431 */ 1432 preempt_disable(); 1433 if (!(worker->flags & WORKER_NOT_RUNNING)) 1434 worker->pool->nr_running++; 1435 preempt_enable(); 1436 1437 /* 1438 * CPU intensive auto-detection cares about how long a work item hogged 1439 * CPU without sleeping. Reset the starting timestamp on wakeup. 1440 */ 1441 worker->current_at = worker->task->se.sum_exec_runtime; 1442 1443 WRITE_ONCE(worker->sleeping, 0); 1444 } 1445 1446 /** 1447 * wq_worker_sleeping - a worker is going to sleep 1448 * @task: task going to sleep 1449 * 1450 * This function is called from schedule() when a busy worker is 1451 * going to sleep. 1452 */ 1453 void wq_worker_sleeping(struct task_struct *task) 1454 { 1455 struct worker *worker = kthread_data(task); 1456 struct worker_pool *pool; 1457 1458 /* 1459 * Rescuers, which may not have all the fields set up like normal 1460 * workers, also reach here, let's not access anything before 1461 * checking NOT_RUNNING. 1462 */ 1463 if (worker->flags & WORKER_NOT_RUNNING) 1464 return; 1465 1466 pool = worker->pool; 1467 1468 /* Return if preempted before wq_worker_running() was reached */ 1469 if (READ_ONCE(worker->sleeping)) 1470 return; 1471 1472 WRITE_ONCE(worker->sleeping, 1); 1473 raw_spin_lock_irq(&pool->lock); 1474 1475 /* 1476 * Recheck in case unbind_workers() preempted us. We don't 1477 * want to decrement nr_running after the worker is unbound 1478 * and nr_running has been reset. 1479 */ 1480 if (worker->flags & WORKER_NOT_RUNNING) { 1481 raw_spin_unlock_irq(&pool->lock); 1482 return; 1483 } 1484 1485 pool->nr_running--; 1486 if (kick_pool(pool)) 1487 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1488 1489 raw_spin_unlock_irq(&pool->lock); 1490 } 1491 1492 /** 1493 * wq_worker_tick - a scheduler tick occurred while a kworker is running 1494 * @task: task currently running 1495 * 1496 * Called from sched_tick(). We're in the IRQ context and the current 1497 * worker's fields which follow the 'K' locking rule can be accessed safely. 1498 */ 1499 void wq_worker_tick(struct task_struct *task) 1500 { 1501 struct worker *worker = kthread_data(task); 1502 struct pool_workqueue *pwq = worker->current_pwq; 1503 struct worker_pool *pool = worker->pool; 1504 1505 if (!pwq) 1506 return; 1507 1508 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 1509 1510 if (!wq_cpu_intensive_thresh_us) 1511 return; 1512 1513 /* 1514 * If the current worker is concurrency managed and hogged the CPU for 1515 * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1516 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1517 * 1518 * Set @worker->sleeping means that @worker is in the process of 1519 * switching out voluntarily and won't be contributing to 1520 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1521 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1522 * double decrements. The task is releasing the CPU anyway. Let's skip. 1523 * We probably want to make this prettier in the future. 1524 */ 1525 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1526 worker->task->se.sum_exec_runtime - worker->current_at < 1527 wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1528 return; 1529 1530 raw_spin_lock(&pool->lock); 1531 1532 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 1533 wq_cpu_intensive_report(worker->current_func); 1534 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1535 1536 if (kick_pool(pool)) 1537 pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1538 1539 raw_spin_unlock(&pool->lock); 1540 } 1541 1542 /** 1543 * wq_worker_last_func - retrieve worker's last work function 1544 * @task: Task to retrieve last work function of. 1545 * 1546 * Determine the last function a worker executed. This is called from 1547 * the scheduler to get a worker's last known identity. 1548 * 1549 * CONTEXT: 1550 * raw_spin_lock_irq(rq->lock) 1551 * 1552 * This function is called during schedule() when a kworker is going 1553 * to sleep. It's used by psi to identify aggregation workers during 1554 * dequeuing, to allow periodic aggregation to shut-off when that 1555 * worker is the last task in the system or cgroup to go to sleep. 1556 * 1557 * As this function doesn't involve any workqueue-related locking, it 1558 * only returns stable values when called from inside the scheduler's 1559 * queuing and dequeuing paths, when @task, which must be a kworker, 1560 * is guaranteed to not be processing any works. 1561 * 1562 * Return: 1563 * The last work function %current executed as a worker, NULL if it 1564 * hasn't executed any work yet. 1565 */ 1566 work_func_t wq_worker_last_func(struct task_struct *task) 1567 { 1568 struct worker *worker = kthread_data(task); 1569 1570 return worker->last_func; 1571 } 1572 1573 /** 1574 * wq_node_nr_active - Determine wq_node_nr_active to use 1575 * @wq: workqueue of interest 1576 * @node: NUMA node, can be %NUMA_NO_NODE 1577 * 1578 * Determine wq_node_nr_active to use for @wq on @node. Returns: 1579 * 1580 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active. 1581 * 1582 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE. 1583 * 1584 * - Otherwise, node_nr_active[@node]. 1585 */ 1586 static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq, 1587 int node) 1588 { 1589 if (!(wq->flags & WQ_UNBOUND)) 1590 return NULL; 1591 1592 if (node == NUMA_NO_NODE) 1593 node = nr_node_ids; 1594 1595 return wq->node_nr_active[node]; 1596 } 1597 1598 /** 1599 * wq_update_node_max_active - Update per-node max_actives to use 1600 * @wq: workqueue to update 1601 * @off_cpu: CPU that's going down, -1 if a CPU is not going down 1602 * 1603 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is 1604 * distributed among nodes according to the proportions of numbers of online 1605 * cpus. The result is always between @wq->min_active and max_active. 1606 */ 1607 static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu) 1608 { 1609 struct cpumask *effective = unbound_effective_cpumask(wq); 1610 int min_active = READ_ONCE(wq->min_active); 1611 int max_active = READ_ONCE(wq->max_active); 1612 int total_cpus, node; 1613 1614 lockdep_assert_held(&wq->mutex); 1615 1616 if (!wq_topo_initialized) 1617 return; 1618 1619 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective)) 1620 off_cpu = -1; 1621 1622 total_cpus = cpumask_weight_and(effective, cpu_online_mask); 1623 if (off_cpu >= 0) 1624 total_cpus--; 1625 1626 /* If all CPUs of the wq get offline, use the default values */ 1627 if (unlikely(!total_cpus)) { 1628 for_each_node(node) 1629 wq_node_nr_active(wq, node)->max = min_active; 1630 1631 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; 1632 return; 1633 } 1634 1635 for_each_node(node) { 1636 int node_cpus; 1637 1638 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node)); 1639 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node) 1640 node_cpus--; 1641 1642 wq_node_nr_active(wq, node)->max = 1643 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus), 1644 min_active, max_active); 1645 } 1646 1647 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active; 1648 } 1649 1650 /** 1651 * get_pwq - get an extra reference on the specified pool_workqueue 1652 * @pwq: pool_workqueue to get 1653 * 1654 * Obtain an extra reference on @pwq. The caller should guarantee that 1655 * @pwq has positive refcnt and be holding the matching pool->lock. 1656 */ 1657 static void get_pwq(struct pool_workqueue *pwq) 1658 { 1659 lockdep_assert_held(&pwq->pool->lock); 1660 WARN_ON_ONCE(pwq->refcnt <= 0); 1661 pwq->refcnt++; 1662 } 1663 1664 /** 1665 * put_pwq - put a pool_workqueue reference 1666 * @pwq: pool_workqueue to put 1667 * 1668 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 1669 * destruction. The caller should be holding the matching pool->lock. 1670 */ 1671 static void put_pwq(struct pool_workqueue *pwq) 1672 { 1673 lockdep_assert_held(&pwq->pool->lock); 1674 if (likely(--pwq->refcnt)) 1675 return; 1676 /* 1677 * @pwq can't be released under pool->lock, bounce to a dedicated 1678 * kthread_worker to avoid A-A deadlocks. 1679 */ 1680 kthread_queue_work(pwq_release_worker, &pwq->release_work); 1681 } 1682 1683 /** 1684 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1685 * @pwq: pool_workqueue to put (can be %NULL) 1686 * 1687 * put_pwq() with locking. This function also allows %NULL @pwq. 1688 */ 1689 static void put_pwq_unlocked(struct pool_workqueue *pwq) 1690 { 1691 if (pwq) { 1692 /* 1693 * As both pwqs and pools are RCU protected, the 1694 * following lock operations are safe. 1695 */ 1696 raw_spin_lock_irq(&pwq->pool->lock); 1697 put_pwq(pwq); 1698 raw_spin_unlock_irq(&pwq->pool->lock); 1699 } 1700 } 1701 1702 static bool pwq_is_empty(struct pool_workqueue *pwq) 1703 { 1704 return !pwq->nr_active && list_empty(&pwq->inactive_works); 1705 } 1706 1707 static void __pwq_activate_work(struct pool_workqueue *pwq, 1708 struct work_struct *work) 1709 { 1710 unsigned long *wdb = work_data_bits(work); 1711 1712 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE)); 1713 trace_workqueue_activate_work(work); 1714 if (list_empty(&pwq->pool->worklist)) 1715 pwq->pool->last_progress_ts = jiffies; 1716 move_linked_works(work, &pwq->pool->worklist, NULL); 1717 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb); 1718 } 1719 1720 static bool tryinc_node_nr_active(struct wq_node_nr_active *nna) 1721 { 1722 int max = READ_ONCE(nna->max); 1723 int old = atomic_read(&nna->nr); 1724 1725 do { 1726 if (old >= max) 1727 return false; 1728 } while (!atomic_try_cmpxchg_relaxed(&nna->nr, &old, old + 1)); 1729 1730 return true; 1731 } 1732 1733 /** 1734 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq 1735 * @pwq: pool_workqueue of interest 1736 * @fill: max_active may have increased, try to increase concurrency level 1737 * 1738 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is 1739 * successfully obtained. %false otherwise. 1740 */ 1741 static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill) 1742 { 1743 struct workqueue_struct *wq = pwq->wq; 1744 struct worker_pool *pool = pwq->pool; 1745 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node); 1746 bool obtained = false; 1747 1748 lockdep_assert_held(&pool->lock); 1749 1750 if (!nna) { 1751 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */ 1752 obtained = pwq->nr_active < READ_ONCE(wq->max_active); 1753 goto out; 1754 } 1755 1756 if (unlikely(pwq->plugged)) 1757 return false; 1758 1759 /* 1760 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is 1761 * already waiting on $nna, pwq_dec_nr_active() will maintain the 1762 * concurrency level. Don't jump the line. 1763 * 1764 * We need to ignore the pending test after max_active has increased as 1765 * pwq_dec_nr_active() can only maintain the concurrency level but not 1766 * increase it. This is indicated by @fill. 1767 */ 1768 if (!list_empty(&pwq->pending_node) && likely(!fill)) 1769 goto out; 1770 1771 obtained = tryinc_node_nr_active(nna); 1772 if (obtained) 1773 goto out; 1774 1775 /* 1776 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs 1777 * and try again. The smp_mb() is paired with the implied memory barrier 1778 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either 1779 * we see the decremented $nna->nr or they see non-empty 1780 * $nna->pending_pwqs. 1781 */ 1782 raw_spin_lock(&nna->lock); 1783 1784 if (list_empty(&pwq->pending_node)) 1785 list_add_tail(&pwq->pending_node, &nna->pending_pwqs); 1786 else if (likely(!fill)) 1787 goto out_unlock; 1788 1789 smp_mb(); 1790 1791 obtained = tryinc_node_nr_active(nna); 1792 1793 /* 1794 * If @fill, @pwq might have already been pending. Being spuriously 1795 * pending in cold paths doesn't affect anything. Let's leave it be. 1796 */ 1797 if (obtained && likely(!fill)) 1798 list_del_init(&pwq->pending_node); 1799 1800 out_unlock: 1801 raw_spin_unlock(&nna->lock); 1802 out: 1803 if (obtained) 1804 pwq->nr_active++; 1805 return obtained; 1806 } 1807 1808 /** 1809 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq 1810 * @pwq: pool_workqueue of interest 1811 * @fill: max_active may have increased, try to increase concurrency level 1812 * 1813 * Activate the first inactive work item of @pwq if available and allowed by 1814 * max_active limit. 1815 * 1816 * Returns %true if an inactive work item has been activated. %false if no 1817 * inactive work item is found or max_active limit is reached. 1818 */ 1819 static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill) 1820 { 1821 struct work_struct *work = 1822 list_first_entry_or_null(&pwq->inactive_works, 1823 struct work_struct, entry); 1824 1825 if (work && pwq_tryinc_nr_active(pwq, fill)) { 1826 __pwq_activate_work(pwq, work); 1827 return true; 1828 } else { 1829 return false; 1830 } 1831 } 1832 1833 /** 1834 * unplug_oldest_pwq - unplug the oldest pool_workqueue 1835 * @wq: workqueue_struct where its oldest pwq is to be unplugged 1836 * 1837 * This function should only be called for ordered workqueues where only the 1838 * oldest pwq is unplugged, the others are plugged to suspend execution to 1839 * ensure proper work item ordering:: 1840 * 1841 * dfl_pwq --------------+ [P] - plugged 1842 * | 1843 * v 1844 * pwqs -> A -> B [P] -> C [P] (newest) 1845 * | | | 1846 * 1 3 5 1847 * | | | 1848 * 2 4 6 1849 * 1850 * When the oldest pwq is drained and removed, this function should be called 1851 * to unplug the next oldest one to start its work item execution. Note that 1852 * pwq's are linked into wq->pwqs with the oldest first, so the first one in 1853 * the list is the oldest. 1854 */ 1855 static void unplug_oldest_pwq(struct workqueue_struct *wq) 1856 { 1857 struct pool_workqueue *pwq; 1858 1859 lockdep_assert_held(&wq->mutex); 1860 1861 /* Caller should make sure that pwqs isn't empty before calling */ 1862 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue, 1863 pwqs_node); 1864 raw_spin_lock_irq(&pwq->pool->lock); 1865 if (pwq->plugged) { 1866 pwq->plugged = false; 1867 if (pwq_activate_first_inactive(pwq, true)) { 1868 /* 1869 * While plugged, queueing skips activation which 1870 * includes bumping the nr_active count and adding the 1871 * pwq to nna->pending_pwqs if the count can't be 1872 * obtained. We need to restore both for the pwq being 1873 * unplugged. The first call activates the first 1874 * inactive work item and the second, if there are more 1875 * inactive, puts the pwq on pending_pwqs. 1876 */ 1877 pwq_activate_first_inactive(pwq, false); 1878 1879 kick_pool(pwq->pool); 1880 } 1881 } 1882 raw_spin_unlock_irq(&pwq->pool->lock); 1883 } 1884 1885 /** 1886 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active 1887 * @nna: wq_node_nr_active to activate a pending pwq for 1888 * @caller_pool: worker_pool the caller is locking 1889 * 1890 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked. 1891 * @caller_pool may be unlocked and relocked to lock other worker_pools. 1892 */ 1893 static void node_activate_pending_pwq(struct wq_node_nr_active *nna, 1894 struct worker_pool *caller_pool) 1895 { 1896 struct worker_pool *locked_pool = caller_pool; 1897 struct pool_workqueue *pwq; 1898 struct work_struct *work; 1899 1900 lockdep_assert_held(&caller_pool->lock); 1901 1902 raw_spin_lock(&nna->lock); 1903 retry: 1904 pwq = list_first_entry_or_null(&nna->pending_pwqs, 1905 struct pool_workqueue, pending_node); 1906 if (!pwq) 1907 goto out_unlock; 1908 1909 /* 1910 * If @pwq is for a different pool than @locked_pool, we need to lock 1911 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock 1912 * / lock dance. For that, we also need to release @nna->lock as it's 1913 * nested inside pool locks. 1914 */ 1915 if (pwq->pool != locked_pool) { 1916 raw_spin_unlock(&locked_pool->lock); 1917 locked_pool = pwq->pool; 1918 if (!raw_spin_trylock(&locked_pool->lock)) { 1919 raw_spin_unlock(&nna->lock); 1920 raw_spin_lock(&locked_pool->lock); 1921 raw_spin_lock(&nna->lock); 1922 goto retry; 1923 } 1924 } 1925 1926 /* 1927 * $pwq may not have any inactive work items due to e.g. cancellations. 1928 * Drop it from pending_pwqs and see if there's another one. 1929 */ 1930 work = list_first_entry_or_null(&pwq->inactive_works, 1931 struct work_struct, entry); 1932 if (!work) { 1933 list_del_init(&pwq->pending_node); 1934 goto retry; 1935 } 1936 1937 /* 1938 * Acquire an nr_active count and activate the inactive work item. If 1939 * $pwq still has inactive work items, rotate it to the end of the 1940 * pending_pwqs so that we round-robin through them. This means that 1941 * inactive work items are not activated in queueing order which is fine 1942 * given that there has never been any ordering across different pwqs. 1943 */ 1944 if (likely(tryinc_node_nr_active(nna))) { 1945 pwq->nr_active++; 1946 __pwq_activate_work(pwq, work); 1947 1948 if (list_empty(&pwq->inactive_works)) 1949 list_del_init(&pwq->pending_node); 1950 else 1951 list_move_tail(&pwq->pending_node, &nna->pending_pwqs); 1952 1953 /* if activating a foreign pool, make sure it's running */ 1954 if (pwq->pool != caller_pool) 1955 kick_pool(pwq->pool); 1956 } 1957 1958 out_unlock: 1959 raw_spin_unlock(&nna->lock); 1960 if (locked_pool != caller_pool) { 1961 raw_spin_unlock(&locked_pool->lock); 1962 raw_spin_lock(&caller_pool->lock); 1963 } 1964 } 1965 1966 /** 1967 * pwq_dec_nr_active - Retire an active count 1968 * @pwq: pool_workqueue of interest 1969 * 1970 * Decrement @pwq's nr_active and try to activate the first inactive work item. 1971 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock. 1972 */ 1973 static void pwq_dec_nr_active(struct pool_workqueue *pwq) 1974 { 1975 struct worker_pool *pool = pwq->pool; 1976 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node); 1977 1978 lockdep_assert_held(&pool->lock); 1979 1980 /* 1981 * @pwq->nr_active should be decremented for both percpu and unbound 1982 * workqueues. 1983 */ 1984 pwq->nr_active--; 1985 1986 /* 1987 * For a percpu workqueue, it's simple. Just need to kick the first 1988 * inactive work item on @pwq itself. 1989 */ 1990 if (!nna) { 1991 pwq_activate_first_inactive(pwq, false); 1992 return; 1993 } 1994 1995 /* 1996 * If @pwq is for an unbound workqueue, it's more complicated because 1997 * multiple pwqs and pools may be sharing the nr_active count. When a 1998 * pwq needs to wait for an nr_active count, it puts itself on 1999 * $nna->pending_pwqs. The following atomic_dec_return()'s implied 2000 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to 2001 * guarantee that either we see non-empty pending_pwqs or they see 2002 * decremented $nna->nr. 2003 * 2004 * $nna->max may change as CPUs come online/offline and @pwq->wq's 2005 * max_active gets updated. However, it is guaranteed to be equal to or 2006 * larger than @pwq->wq->min_active which is above zero unless freezing. 2007 * This maintains the forward progress guarantee. 2008 */ 2009 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max)) 2010 return; 2011 2012 if (!list_empty(&nna->pending_pwqs)) 2013 node_activate_pending_pwq(nna, pool); 2014 } 2015 2016 /** 2017 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 2018 * @pwq: pwq of interest 2019 * @work_data: work_data of work which left the queue 2020 * 2021 * A work either has completed or is removed from pending queue, 2022 * decrement nr_in_flight of its pwq and handle workqueue flushing. 2023 * 2024 * NOTE: 2025 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock 2026 * and thus should be called after all other state updates for the in-flight 2027 * work item is complete. 2028 * 2029 * CONTEXT: 2030 * raw_spin_lock_irq(pool->lock). 2031 */ 2032 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 2033 { 2034 int color = get_work_color(work_data); 2035 2036 if (!(work_data & WORK_STRUCT_INACTIVE)) 2037 pwq_dec_nr_active(pwq); 2038 2039 pwq->nr_in_flight[color]--; 2040 2041 /* is flush in progress and are we at the flushing tip? */ 2042 if (likely(pwq->flush_color != color)) 2043 goto out_put; 2044 2045 /* are there still in-flight works? */ 2046 if (pwq->nr_in_flight[color]) 2047 goto out_put; 2048 2049 /* this pwq is done, clear flush_color */ 2050 pwq->flush_color = -1; 2051 2052 /* 2053 * If this was the last pwq, wake up the first flusher. It 2054 * will handle the rest. 2055 */ 2056 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 2057 complete(&pwq->wq->first_flusher->done); 2058 out_put: 2059 put_pwq(pwq); 2060 } 2061 2062 /** 2063 * try_to_grab_pending - steal work item from worklist and disable irq 2064 * @work: work item to steal 2065 * @cflags: %WORK_CANCEL_ flags 2066 * @irq_flags: place to store irq state 2067 * 2068 * Try to grab PENDING bit of @work. This function can handle @work in any 2069 * stable state - idle, on timer or on worklist. 2070 * 2071 * Return: 2072 * 2073 * ======== ================================================================ 2074 * 1 if @work was pending and we successfully stole PENDING 2075 * 0 if @work was idle and we claimed PENDING 2076 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 2077 * ======== ================================================================ 2078 * 2079 * Note: 2080 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 2081 * interrupted while holding PENDING and @work off queue, irq must be 2082 * disabled on entry. This, combined with delayed_work->timer being 2083 * irqsafe, ensures that we return -EAGAIN for finite short period of time. 2084 * 2085 * On successful return, >= 0, irq is disabled and the caller is 2086 * responsible for releasing it using local_irq_restore(*@irq_flags). 2087 * 2088 * This function is safe to call from any context including IRQ handler. 2089 */ 2090 static int try_to_grab_pending(struct work_struct *work, u32 cflags, 2091 unsigned long *irq_flags) 2092 { 2093 struct worker_pool *pool; 2094 struct pool_workqueue *pwq; 2095 2096 local_irq_save(*irq_flags); 2097 2098 /* try to steal the timer if it exists */ 2099 if (cflags & WORK_CANCEL_DELAYED) { 2100 struct delayed_work *dwork = to_delayed_work(work); 2101 2102 /* 2103 * dwork->timer is irqsafe. If timer_delete() fails, it's 2104 * guaranteed that the timer is not queued anywhere and not 2105 * running on the local CPU. 2106 */ 2107 if (likely(timer_delete(&dwork->timer))) 2108 return 1; 2109 } 2110 2111 /* try to claim PENDING the normal way */ 2112 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 2113 return 0; 2114 2115 rcu_read_lock(); 2116 /* 2117 * The queueing is in progress, or it is already queued. Try to 2118 * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 2119 */ 2120 pool = get_work_pool(work); 2121 if (!pool) 2122 goto fail; 2123 2124 raw_spin_lock(&pool->lock); 2125 /* 2126 * work->data is guaranteed to point to pwq only while the work 2127 * item is queued on pwq->wq, and both updating work->data to point 2128 * to pwq on queueing and to pool on dequeueing are done under 2129 * pwq->pool->lock. This in turn guarantees that, if work->data 2130 * points to pwq which is associated with a locked pool, the work 2131 * item is currently queued on that pool. 2132 */ 2133 pwq = get_work_pwq(work); 2134 if (pwq && pwq->pool == pool) { 2135 unsigned long work_data = *work_data_bits(work); 2136 2137 debug_work_deactivate(work); 2138 2139 /* 2140 * A cancelable inactive work item must be in the 2141 * pwq->inactive_works since a queued barrier can't be 2142 * canceled (see the comments in insert_wq_barrier()). 2143 * 2144 * An inactive work item cannot be deleted directly because 2145 * it might have linked barrier work items which, if left 2146 * on the inactive_works list, will confuse pwq->nr_active 2147 * management later on and cause stall. Move the linked 2148 * barrier work items to the worklist when deleting the grabbed 2149 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that 2150 * it doesn't participate in nr_active management in later 2151 * pwq_dec_nr_in_flight(). 2152 */ 2153 if (work_data & WORK_STRUCT_INACTIVE) 2154 move_linked_works(work, &pwq->pool->worklist, NULL); 2155 2156 list_del_init(&work->entry); 2157 2158 /* 2159 * work->data points to pwq iff queued. Let's point to pool. As 2160 * this destroys work->data needed by the next step, stash it. 2161 */ 2162 set_work_pool_and_keep_pending(work, pool->id, 2163 pool_offq_flags(pool)); 2164 2165 /* must be the last step, see the function comment */ 2166 pwq_dec_nr_in_flight(pwq, work_data); 2167 2168 raw_spin_unlock(&pool->lock); 2169 rcu_read_unlock(); 2170 return 1; 2171 } 2172 raw_spin_unlock(&pool->lock); 2173 fail: 2174 rcu_read_unlock(); 2175 local_irq_restore(*irq_flags); 2176 return -EAGAIN; 2177 } 2178 2179 /** 2180 * work_grab_pending - steal work item from worklist and disable irq 2181 * @work: work item to steal 2182 * @cflags: %WORK_CANCEL_ flags 2183 * @irq_flags: place to store IRQ state 2184 * 2185 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer 2186 * or on worklist. 2187 * 2188 * Can be called from any context. IRQ is disabled on return with IRQ state 2189 * stored in *@irq_flags. The caller is responsible for re-enabling it using 2190 * local_irq_restore(). 2191 * 2192 * Returns %true if @work was pending. %false if idle. 2193 */ 2194 static bool work_grab_pending(struct work_struct *work, u32 cflags, 2195 unsigned long *irq_flags) 2196 { 2197 int ret; 2198 2199 while (true) { 2200 ret = try_to_grab_pending(work, cflags, irq_flags); 2201 if (ret >= 0) 2202 return ret; 2203 cpu_relax(); 2204 } 2205 } 2206 2207 /** 2208 * insert_work - insert a work into a pool 2209 * @pwq: pwq @work belongs to 2210 * @work: work to insert 2211 * @head: insertion point 2212 * @extra_flags: extra WORK_STRUCT_* flags to set 2213 * 2214 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 2215 * work_struct flags. 2216 * 2217 * CONTEXT: 2218 * raw_spin_lock_irq(pool->lock). 2219 */ 2220 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 2221 struct list_head *head, unsigned int extra_flags) 2222 { 2223 debug_work_activate(work); 2224 2225 /* record the work call stack in order to print it in KASAN reports */ 2226 kasan_record_aux_stack(work); 2227 2228 /* we own @work, set data and link */ 2229 set_work_pwq(work, pwq, extra_flags); 2230 list_add_tail(&work->entry, head); 2231 get_pwq(pwq); 2232 } 2233 2234 /* 2235 * Test whether @work is being queued from another work executing on the 2236 * same workqueue. 2237 */ 2238 static bool is_chained_work(struct workqueue_struct *wq) 2239 { 2240 struct worker *worker; 2241 2242 worker = current_wq_worker(); 2243 /* 2244 * Return %true iff I'm a worker executing a work item on @wq. If 2245 * I'm @worker, it's safe to dereference it without locking. 2246 */ 2247 return worker && worker->current_pwq->wq == wq; 2248 } 2249 2250 /* 2251 * When queueing an unbound work item to a wq, prefer local CPU if allowed 2252 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 2253 * avoid perturbing sensitive tasks. 2254 */ 2255 static int wq_select_unbound_cpu(int cpu) 2256 { 2257 int new_cpu; 2258 2259 if (likely(!wq_debug_force_rr_cpu)) { 2260 if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 2261 return cpu; 2262 } else { 2263 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 2264 } 2265 2266 new_cpu = __this_cpu_read(wq_rr_cpu_last); 2267 new_cpu = cpumask_next_and_wrap(new_cpu, wq_unbound_cpumask, cpu_online_mask); 2268 if (unlikely(new_cpu >= nr_cpu_ids)) 2269 return cpu; 2270 __this_cpu_write(wq_rr_cpu_last, new_cpu); 2271 2272 return new_cpu; 2273 } 2274 2275 static void __queue_work(int cpu, struct workqueue_struct *wq, 2276 struct work_struct *work) 2277 { 2278 struct pool_workqueue *pwq; 2279 struct worker_pool *last_pool, *pool; 2280 unsigned int work_flags; 2281 unsigned int req_cpu = cpu; 2282 2283 /* 2284 * NOTE: Check whether the used workqueue is deprecated and warn 2285 */ 2286 if (unlikely(wq->flags & __WQ_DEPRECATED)) 2287 pr_warn_once("workqueue: work func %ps enqueued on deprecated workqueue. " 2288 "Use system_{percpu|dfl}_wq instead.\n", 2289 work->func); 2290 2291 /* 2292 * While a work item is PENDING && off queue, a task trying to 2293 * steal the PENDING will busy-loop waiting for it to either get 2294 * queued or lose PENDING. Grabbing PENDING and queueing should 2295 * happen with IRQ disabled. 2296 */ 2297 lockdep_assert_irqs_disabled(); 2298 2299 /* 2300 * For a draining wq, only works from the same workqueue are 2301 * allowed. The __WQ_DESTROYING helps to spot the issue that 2302 * queues a new work item to a wq after destroy_workqueue(wq). 2303 */ 2304 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 2305 WARN_ONCE(!is_chained_work(wq), "workqueue: cannot queue %ps on wq %s\n", 2306 work->func, wq->name))) { 2307 struct work_offq_data offqd; 2308 2309 /* 2310 * State on entry: PENDING is set, work is off-queue (no 2311 * insert_work() has run). 2312 * 2313 * Returning without clearing PENDING would leave the work 2314 * in a weird state (PENDING=1, PWQ=0, entry empty) 2315 */ 2316 work_offqd_unpack(&offqd, *work_data_bits(work)); 2317 set_work_pool_and_clear_pending(work, offqd.pool_id, 2318 work_offqd_pack_flags(&offqd)); 2319 return; 2320 } 2321 rcu_read_lock(); 2322 retry: 2323 /* pwq which will be used unless @work is executing elsewhere */ 2324 if (req_cpu == WORK_CPU_UNBOUND) { 2325 if (wq->flags & WQ_UNBOUND) 2326 cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 2327 else 2328 cpu = raw_smp_processor_id(); 2329 } 2330 2331 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 2332 pool = pwq->pool; 2333 2334 /* 2335 * If @work was previously on a different pool, it might still be 2336 * running there, in which case the work needs to be queued on that 2337 * pool to guarantee non-reentrancy. 2338 * 2339 * For ordered workqueue, work items must be queued on the newest pwq 2340 * for accurate order management. Guaranteed order also guarantees 2341 * non-reentrancy. See the comments above unplug_oldest_pwq(). 2342 */ 2343 last_pool = get_work_pool(work); 2344 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) { 2345 struct worker *worker; 2346 2347 raw_spin_lock(&last_pool->lock); 2348 2349 worker = find_worker_executing_work(last_pool, work); 2350 2351 if (worker && worker->current_pwq->wq == wq) { 2352 pwq = worker->current_pwq; 2353 pool = pwq->pool; 2354 WARN_ON_ONCE(pool != last_pool); 2355 } else { 2356 /* meh... not running there, queue here */ 2357 raw_spin_unlock(&last_pool->lock); 2358 raw_spin_lock(&pool->lock); 2359 } 2360 } else { 2361 raw_spin_lock(&pool->lock); 2362 } 2363 2364 /* 2365 * pwq is determined and locked. For unbound pools, we could have raced 2366 * with pwq release and it could already be dead. If its refcnt is zero, 2367 * repeat pwq selection. Note that unbound pwqs never die without 2368 * another pwq replacing it in cpu_pwq or while work items are executing 2369 * on it, so the retrying is guaranteed to make forward-progress. 2370 */ 2371 if (unlikely(!pwq->refcnt)) { 2372 if (wq->flags & WQ_UNBOUND) { 2373 raw_spin_unlock(&pool->lock); 2374 cpu_relax(); 2375 goto retry; 2376 } 2377 /* oops */ 2378 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 2379 wq->name, cpu); 2380 } 2381 2382 /* pwq determined, queue */ 2383 trace_workqueue_queue_work(req_cpu, pwq, work); 2384 2385 if (WARN_ON(!list_empty(&work->entry))) 2386 goto out; 2387 2388 pwq->nr_in_flight[pwq->work_color]++; 2389 work_flags = work_color_to_flags(pwq->work_color); 2390 2391 /* 2392 * Limit the number of concurrently active work items to max_active. 2393 * @work must also queue behind existing inactive work items to maintain 2394 * ordering when max_active changes. See wq_adjust_max_active(). 2395 */ 2396 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) { 2397 if (list_empty(&pool->worklist)) 2398 pool->last_progress_ts = jiffies; 2399 2400 trace_workqueue_activate_work(work); 2401 insert_work(pwq, work, &pool->worklist, work_flags); 2402 kick_pool(pool); 2403 } else { 2404 work_flags |= WORK_STRUCT_INACTIVE; 2405 insert_work(pwq, work, &pwq->inactive_works, work_flags); 2406 } 2407 2408 out: 2409 raw_spin_unlock(&pool->lock); 2410 rcu_read_unlock(); 2411 } 2412 2413 static bool clear_pending_if_disabled(struct work_struct *work) 2414 { 2415 unsigned long data = *work_data_bits(work); 2416 struct work_offq_data offqd; 2417 2418 if (likely((data & WORK_STRUCT_PWQ) || 2419 !(data & WORK_OFFQ_DISABLE_MASK))) 2420 return false; 2421 2422 work_offqd_unpack(&offqd, data); 2423 set_work_pool_and_clear_pending(work, offqd.pool_id, 2424 work_offqd_pack_flags(&offqd)); 2425 return true; 2426 } 2427 2428 /** 2429 * queue_work_on - queue work on specific cpu 2430 * @cpu: CPU number to execute work on 2431 * @wq: workqueue to use 2432 * @work: work to queue 2433 * 2434 * We queue the work to a specific CPU, the caller must ensure it 2435 * can't go away. Callers that fail to ensure that the specified 2436 * CPU cannot go away will execute on a randomly chosen CPU. 2437 * But note well that callers specifying a CPU that never has been 2438 * online will get a splat. 2439 * 2440 * Return: %false if @work was already on a queue, %true otherwise. 2441 */ 2442 bool queue_work_on(int cpu, struct workqueue_struct *wq, 2443 struct work_struct *work) 2444 { 2445 bool ret = false; 2446 unsigned long irq_flags; 2447 2448 local_irq_save(irq_flags); 2449 2450 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2451 !clear_pending_if_disabled(work)) { 2452 __queue_work(cpu, wq, work); 2453 ret = true; 2454 } 2455 2456 local_irq_restore(irq_flags); 2457 return ret; 2458 } 2459 EXPORT_SYMBOL(queue_work_on); 2460 2461 /** 2462 * select_numa_node_cpu - Select a CPU based on NUMA node 2463 * @node: NUMA node ID that we want to select a CPU from 2464 * 2465 * This function will attempt to find a "random" cpu available on a given 2466 * node. If there are no CPUs available on the given node it will return 2467 * WORK_CPU_UNBOUND indicating that we should just schedule to any 2468 * available CPU if we need to schedule this work. 2469 */ 2470 static int select_numa_node_cpu(int node) 2471 { 2472 int cpu; 2473 2474 /* Delay binding to CPU if node is not valid or online */ 2475 if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 2476 return WORK_CPU_UNBOUND; 2477 2478 /* Use local node/cpu if we are already there */ 2479 cpu = raw_smp_processor_id(); 2480 if (node == cpu_to_node(cpu)) 2481 return cpu; 2482 2483 /* Use "random" otherwise know as "first" online CPU of node */ 2484 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 2485 2486 /* If CPU is valid return that, otherwise just defer */ 2487 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 2488 } 2489 2490 /** 2491 * queue_work_node - queue work on a "random" cpu for a given NUMA node 2492 * @node: NUMA node that we are targeting the work for 2493 * @wq: workqueue to use 2494 * @work: work to queue 2495 * 2496 * We queue the work to a "random" CPU within a given NUMA node. The basic 2497 * idea here is to provide a way to somehow associate work with a given 2498 * NUMA node. 2499 * 2500 * This function will only make a best effort attempt at getting this onto 2501 * the right NUMA node. If no node is requested or the requested node is 2502 * offline then we just fall back to standard queue_work behavior. 2503 * 2504 * Currently the "random" CPU ends up being the first available CPU in the 2505 * intersection of cpu_online_mask and the cpumask of the node, unless we 2506 * are running on the node. In that case we just use the current CPU. 2507 * 2508 * Return: %false if @work was already on a queue, %true otherwise. 2509 */ 2510 bool queue_work_node(int node, struct workqueue_struct *wq, 2511 struct work_struct *work) 2512 { 2513 unsigned long irq_flags; 2514 bool ret = false; 2515 2516 /* 2517 * This current implementation is specific to unbound workqueues. 2518 * Specifically we only return the first available CPU for a given 2519 * node instead of cycling through individual CPUs within the node. 2520 * 2521 * If this is used with a per-cpu workqueue then the logic in 2522 * workqueue_select_cpu_near would need to be updated to allow for 2523 * some round robin type logic. 2524 */ 2525 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 2526 2527 local_irq_save(irq_flags); 2528 2529 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2530 !clear_pending_if_disabled(work)) { 2531 int cpu = select_numa_node_cpu(node); 2532 2533 __queue_work(cpu, wq, work); 2534 ret = true; 2535 } 2536 2537 local_irq_restore(irq_flags); 2538 return ret; 2539 } 2540 EXPORT_SYMBOL_GPL(queue_work_node); 2541 2542 void delayed_work_timer_fn(struct timer_list *t) 2543 { 2544 struct delayed_work *dwork = timer_container_of(dwork, t, timer); 2545 2546 /* should have been called from irqsafe timer with irq already off */ 2547 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 2548 } 2549 EXPORT_SYMBOL(delayed_work_timer_fn); 2550 2551 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 2552 struct delayed_work *dwork, unsigned long delay) 2553 { 2554 struct timer_list *timer = &dwork->timer; 2555 struct work_struct *work = &dwork->work; 2556 2557 WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 2558 WARN_ON_ONCE(timer_pending(timer)); 2559 WARN_ON_ONCE(!list_empty(&work->entry)); 2560 2561 /* 2562 * If @delay is 0, queue @dwork->work immediately. This is for 2563 * both optimization and correctness. The earliest @timer can 2564 * expire is on the closest next tick and delayed_work users depend 2565 * on that there's no such delay when @delay is 0. 2566 */ 2567 if (!delay) { 2568 __queue_work(cpu, wq, &dwork->work); 2569 return; 2570 } 2571 2572 WARN_ON_ONCE(cpu != WORK_CPU_UNBOUND && !cpu_online(cpu)); 2573 dwork->wq = wq; 2574 dwork->cpu = cpu; 2575 timer->expires = jiffies + delay; 2576 2577 if (housekeeping_enabled(HK_TYPE_TIMER)) { 2578 /* If the current cpu is a housekeeping cpu, use it. */ 2579 cpu = smp_processor_id(); 2580 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER)) 2581 cpu = housekeeping_any_cpu(HK_TYPE_TIMER); 2582 add_timer_on(timer, cpu); 2583 } else { 2584 if (likely(cpu == WORK_CPU_UNBOUND)) 2585 add_timer_global(timer); 2586 else 2587 add_timer_on(timer, cpu); 2588 } 2589 } 2590 2591 /** 2592 * queue_delayed_work_on - queue work on specific CPU after delay 2593 * @cpu: CPU number to execute work on 2594 * @wq: workqueue to use 2595 * @dwork: work to queue 2596 * @delay: number of jiffies to wait before queueing 2597 * 2598 * We queue the delayed_work to a specific CPU, for non-zero delays the 2599 * caller must ensure it is online and can't go away. Callers that fail 2600 * to ensure this, may get @dwork->timer queued to an offlined CPU and 2601 * this will prevent queueing of @dwork->work unless the offlined CPU 2602 * becomes online again. 2603 * 2604 * Return: %false if @work was already on a queue, %true otherwise. If 2605 * @delay is zero and @dwork is idle, it will be scheduled for immediate 2606 * execution. 2607 */ 2608 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 2609 struct delayed_work *dwork, unsigned long delay) 2610 { 2611 struct work_struct *work = &dwork->work; 2612 bool ret = false; 2613 unsigned long irq_flags; 2614 2615 /* read the comment in __queue_work() */ 2616 local_irq_save(irq_flags); 2617 2618 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2619 !clear_pending_if_disabled(work)) { 2620 __queue_delayed_work(cpu, wq, dwork, delay); 2621 ret = true; 2622 } 2623 2624 local_irq_restore(irq_flags); 2625 return ret; 2626 } 2627 EXPORT_SYMBOL(queue_delayed_work_on); 2628 2629 /** 2630 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 2631 * @cpu: CPU number to execute work on 2632 * @wq: workqueue to use 2633 * @dwork: work to queue 2634 * @delay: number of jiffies to wait before queueing 2635 * 2636 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 2637 * modify @dwork's timer so that it expires after @delay. If @delay is 2638 * zero, @work is guaranteed to be scheduled immediately regardless of its 2639 * current state. 2640 * 2641 * Return: %false if @dwork was idle and queued, %true if @dwork was 2642 * pending and its timer was modified. 2643 * 2644 * This function is safe to call from any context including IRQ handler. 2645 * See try_to_grab_pending() for details. 2646 */ 2647 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 2648 struct delayed_work *dwork, unsigned long delay) 2649 { 2650 unsigned long irq_flags; 2651 bool ret; 2652 2653 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags); 2654 2655 if (!clear_pending_if_disabled(&dwork->work)) 2656 __queue_delayed_work(cpu, wq, dwork, delay); 2657 2658 local_irq_restore(irq_flags); 2659 return ret; 2660 } 2661 EXPORT_SYMBOL_GPL(mod_delayed_work_on); 2662 2663 static void rcu_work_rcufn(struct rcu_head *rcu) 2664 { 2665 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 2666 2667 /* read the comment in __queue_work() */ 2668 local_irq_disable(); 2669 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 2670 local_irq_enable(); 2671 } 2672 2673 /** 2674 * queue_rcu_work - queue work after a RCU grace period 2675 * @wq: workqueue to use 2676 * @rwork: work to queue 2677 * 2678 * Return: %false if @rwork was already pending, %true otherwise. Note 2679 * that a full RCU grace period is guaranteed only after a %true return. 2680 * While @rwork is guaranteed to be executed after a %false return, the 2681 * execution may happen before a full RCU grace period has passed. 2682 */ 2683 bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 2684 { 2685 struct work_struct *work = &rwork->work; 2686 2687 /* 2688 * rcu_work can't be canceled or disabled. Warn if the user reached 2689 * inside @rwork and disabled the inner work. 2690 */ 2691 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) && 2692 !WARN_ON_ONCE(clear_pending_if_disabled(work))) { 2693 rwork->wq = wq; 2694 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 2695 return true; 2696 } 2697 2698 return false; 2699 } 2700 EXPORT_SYMBOL(queue_rcu_work); 2701 2702 static struct worker *alloc_worker(int node) 2703 { 2704 struct worker *worker; 2705 2706 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2707 if (worker) { 2708 INIT_LIST_HEAD(&worker->entry); 2709 INIT_LIST_HEAD(&worker->scheduled); 2710 INIT_LIST_HEAD(&worker->node); 2711 /* on creation a worker is in !idle && prep state */ 2712 worker->flags = WORKER_PREP; 2713 } 2714 return worker; 2715 } 2716 2717 static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2718 { 2719 if (pool->cpu < 0 && pool->attrs->affn_strict) 2720 return pool->attrs->__pod_cpumask; 2721 else 2722 return pool->attrs->cpumask; 2723 } 2724 2725 /** 2726 * worker_attach_to_pool() - attach a worker to a pool 2727 * @worker: worker to be attached 2728 * @pool: the target pool 2729 * 2730 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 2731 * cpu-binding of @worker are kept coordinated with the pool across 2732 * cpu-[un]hotplugs. 2733 */ 2734 static void worker_attach_to_pool(struct worker *worker, 2735 struct worker_pool *pool) 2736 { 2737 mutex_lock(&wq_pool_attach_mutex); 2738 2739 /* 2740 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable 2741 * across this function. See the comments above the flag definition for 2742 * details. BH workers are, while per-CPU, always DISASSOCIATED. 2743 */ 2744 if (pool->flags & POOL_DISASSOCIATED) { 2745 worker->flags |= WORKER_UNBOUND; 2746 } else { 2747 WARN_ON_ONCE(pool->flags & POOL_BH); 2748 kthread_set_per_cpu(worker->task, pool->cpu); 2749 } 2750 2751 if (worker->rescue_wq) 2752 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2753 2754 list_add_tail(&worker->node, &pool->workers); 2755 worker->pool = pool; 2756 2757 mutex_unlock(&wq_pool_attach_mutex); 2758 } 2759 2760 static void unbind_worker(struct worker *worker) 2761 { 2762 lockdep_assert_held(&wq_pool_attach_mutex); 2763 2764 kthread_set_per_cpu(worker->task, -1); 2765 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2766 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2767 else 2768 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2769 } 2770 2771 2772 static void detach_worker(struct worker *worker) 2773 { 2774 lockdep_assert_held(&wq_pool_attach_mutex); 2775 2776 unbind_worker(worker); 2777 list_del(&worker->node); 2778 } 2779 2780 /** 2781 * worker_detach_from_pool() - detach a worker from its pool 2782 * @worker: worker which is attached to its pool 2783 * 2784 * Undo the attaching which had been done in worker_attach_to_pool(). The 2785 * caller worker shouldn't access to the pool after detached except it has 2786 * other reference to the pool. 2787 */ 2788 static void worker_detach_from_pool(struct worker *worker) 2789 { 2790 struct worker_pool *pool = worker->pool; 2791 2792 /* there is one permanent BH worker per CPU which should never detach */ 2793 WARN_ON_ONCE(pool->flags & POOL_BH); 2794 2795 mutex_lock(&wq_pool_attach_mutex); 2796 detach_worker(worker); 2797 worker->pool = NULL; 2798 mutex_unlock(&wq_pool_attach_mutex); 2799 2800 /* clear leftover flags without pool->lock after it is detached */ 2801 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2802 } 2803 2804 static int format_worker_id(char *buf, size_t size, struct worker *worker, 2805 struct worker_pool *pool) 2806 { 2807 if (worker->rescue_wq) 2808 return scnprintf(buf, size, "kworker/R-%s", 2809 worker->rescue_wq->name); 2810 2811 if (pool) { 2812 if (pool->cpu >= 0) 2813 return scnprintf(buf, size, "kworker/%d:%d%s", 2814 pool->cpu, worker->id, 2815 pool->attrs->nice < 0 ? "H" : ""); 2816 else 2817 return scnprintf(buf, size, "kworker/u%d:%d", 2818 pool->id, worker->id); 2819 } else { 2820 return scnprintf(buf, size, "kworker/dying"); 2821 } 2822 } 2823 2824 /** 2825 * create_worker - create a new workqueue worker 2826 * @pool: pool the new worker will belong to 2827 * 2828 * Create and start a new worker which is attached to @pool. 2829 * 2830 * CONTEXT: 2831 * Might sleep. Does GFP_KERNEL allocations. 2832 * 2833 * Return: 2834 * Pointer to the newly created worker. 2835 */ 2836 static struct worker *create_worker(struct worker_pool *pool) 2837 { 2838 struct worker *worker; 2839 int id; 2840 2841 /* ID is needed to determine kthread name */ 2842 id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 2843 if (id < 0) { 2844 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 2845 ERR_PTR(id)); 2846 return NULL; 2847 } 2848 2849 worker = alloc_worker(pool->node); 2850 if (!worker) { 2851 pr_err_once("workqueue: Failed to allocate a worker\n"); 2852 goto fail; 2853 } 2854 2855 worker->id = id; 2856 2857 if (!(pool->flags & POOL_BH)) { 2858 char id_buf[WORKER_ID_LEN]; 2859 2860 format_worker_id(id_buf, sizeof(id_buf), worker, pool); 2861 worker->task = kthread_create_on_node(worker_thread, worker, 2862 pool->node, "%s", id_buf); 2863 if (IS_ERR(worker->task)) { 2864 if (PTR_ERR(worker->task) == -EINTR) { 2865 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n", 2866 id_buf); 2867 } else { 2868 pr_err_once("workqueue: Failed to create a worker thread: %pe", 2869 worker->task); 2870 } 2871 goto fail; 2872 } 2873 2874 set_user_nice(worker->task, pool->attrs->nice); 2875 kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 2876 } 2877 2878 /* successful, attach the worker to the pool */ 2879 worker_attach_to_pool(worker, pool); 2880 2881 /* start the newly created worker */ 2882 raw_spin_lock_irq(&pool->lock); 2883 2884 worker->pool->nr_workers++; 2885 worker_enter_idle(worker); 2886 2887 /* 2888 * @worker is waiting on a completion in kthread() and will trigger hung 2889 * check if not woken up soon. As kick_pool() is noop if @pool is empty, 2890 * wake it up explicitly. 2891 */ 2892 if (worker->task) 2893 wake_up_process(worker->task); 2894 2895 raw_spin_unlock_irq(&pool->lock); 2896 2897 return worker; 2898 2899 fail: 2900 ida_free(&pool->worker_ida, id); 2901 kfree(worker); 2902 return NULL; 2903 } 2904 2905 static void detach_dying_workers(struct list_head *cull_list) 2906 { 2907 struct worker *worker; 2908 2909 list_for_each_entry(worker, cull_list, entry) 2910 detach_worker(worker); 2911 } 2912 2913 static void reap_dying_workers(struct list_head *cull_list) 2914 { 2915 struct worker *worker, *tmp; 2916 2917 list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2918 list_del_init(&worker->entry); 2919 kthread_stop_put(worker->task); 2920 kfree(worker); 2921 } 2922 } 2923 2924 /** 2925 * set_worker_dying - Tag a worker for destruction 2926 * @worker: worker to be destroyed 2927 * @list: transfer worker away from its pool->idle_list and into list 2928 * 2929 * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2930 * should be idle. 2931 * 2932 * CONTEXT: 2933 * raw_spin_lock_irq(pool->lock). 2934 */ 2935 static void set_worker_dying(struct worker *worker, struct list_head *list) 2936 { 2937 struct worker_pool *pool = worker->pool; 2938 2939 lockdep_assert_held(&pool->lock); 2940 lockdep_assert_held(&wq_pool_attach_mutex); 2941 2942 /* sanity check frenzy */ 2943 if (WARN_ON(worker->current_work) || 2944 WARN_ON(!list_empty(&worker->scheduled)) || 2945 WARN_ON(!(worker->flags & WORKER_IDLE))) 2946 return; 2947 2948 pool->nr_workers--; 2949 pool->nr_idle--; 2950 2951 worker->flags |= WORKER_DIE; 2952 2953 list_move(&worker->entry, list); 2954 2955 /* get an extra task struct reference for later kthread_stop_put() */ 2956 get_task_struct(worker->task); 2957 } 2958 2959 /** 2960 * idle_worker_timeout - check if some idle workers can now be deleted. 2961 * @t: The pool's idle_timer that just expired 2962 * 2963 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 2964 * worker_leave_idle(), as a worker flicking between idle and active while its 2965 * pool is at the too_many_workers() tipping point would cause too much timer 2966 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 2967 * it expire and re-evaluate things from there. 2968 */ 2969 static void idle_worker_timeout(struct timer_list *t) 2970 { 2971 struct worker_pool *pool = timer_container_of(pool, t, idle_timer); 2972 bool do_cull = false; 2973 2974 if (work_pending(&pool->idle_cull_work)) 2975 return; 2976 2977 raw_spin_lock_irq(&pool->lock); 2978 2979 if (too_many_workers(pool)) { 2980 struct worker *worker; 2981 unsigned long expires; 2982 2983 /* idle_list is kept in LIFO order, check the last one */ 2984 worker = list_last_entry(&pool->idle_list, struct worker, entry); 2985 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2986 do_cull = !time_before(jiffies, expires); 2987 2988 if (!do_cull) 2989 mod_timer(&pool->idle_timer, expires); 2990 } 2991 raw_spin_unlock_irq(&pool->lock); 2992 2993 if (do_cull) 2994 queue_work(system_dfl_wq, &pool->idle_cull_work); 2995 } 2996 2997 /** 2998 * idle_cull_fn - cull workers that have been idle for too long. 2999 * @work: the pool's work for handling these idle workers 3000 * 3001 * This goes through a pool's idle workers and gets rid of those that have been 3002 * idle for at least IDLE_WORKER_TIMEOUT seconds. 3003 * 3004 * We don't want to disturb isolated CPUs because of a pcpu kworker being 3005 * culled, so this also resets worker affinity. This requires a sleepable 3006 * context, hence the split between timer callback and work item. 3007 */ 3008 static void idle_cull_fn(struct work_struct *work) 3009 { 3010 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 3011 LIST_HEAD(cull_list); 3012 3013 /* 3014 * Grabbing wq_pool_attach_mutex here ensures an already-running worker 3015 * cannot proceed beyong set_pf_worker() in its self-destruct path. 3016 * This is required as a previously-preempted worker could run after 3017 * set_worker_dying() has happened but before detach_dying_workers() did. 3018 */ 3019 mutex_lock(&wq_pool_attach_mutex); 3020 raw_spin_lock_irq(&pool->lock); 3021 3022 while (too_many_workers(pool)) { 3023 struct worker *worker; 3024 unsigned long expires; 3025 3026 worker = list_last_entry(&pool->idle_list, struct worker, entry); 3027 expires = worker->last_active + IDLE_WORKER_TIMEOUT; 3028 3029 if (time_before(jiffies, expires)) { 3030 mod_timer(&pool->idle_timer, expires); 3031 break; 3032 } 3033 3034 set_worker_dying(worker, &cull_list); 3035 } 3036 3037 raw_spin_unlock_irq(&pool->lock); 3038 detach_dying_workers(&cull_list); 3039 mutex_unlock(&wq_pool_attach_mutex); 3040 3041 reap_dying_workers(&cull_list); 3042 } 3043 3044 static void send_mayday(struct pool_workqueue *pwq) 3045 { 3046 struct workqueue_struct *wq = pwq->wq; 3047 3048 lockdep_assert_held(&wq_mayday_lock); 3049 3050 if (!wq->rescuer) 3051 return; 3052 3053 /* mayday mayday mayday */ 3054 if (list_empty(&pwq->mayday_node)) { 3055 /* 3056 * If @pwq is for an unbound wq, its base ref may be put at 3057 * any time due to an attribute change. Pin @pwq until the 3058 * rescuer is done with it. 3059 */ 3060 get_pwq(pwq); 3061 list_add_tail(&pwq->mayday_node, &wq->maydays); 3062 wake_up_process(wq->rescuer->task); 3063 pwq->stats[PWQ_STAT_MAYDAY]++; 3064 } 3065 } 3066 3067 static void pool_mayday_timeout(struct timer_list *t) 3068 { 3069 struct worker_pool *pool = timer_container_of(pool, t, mayday_timer); 3070 struct work_struct *work; 3071 3072 raw_spin_lock_irq(&pool->lock); 3073 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 3074 3075 if (need_to_create_worker(pool)) { 3076 /* 3077 * We've been trying to create a new worker but 3078 * haven't been successful. We might be hitting an 3079 * allocation deadlock. Send distress signals to 3080 * rescuers. 3081 */ 3082 list_for_each_entry(work, &pool->worklist, entry) 3083 send_mayday(get_work_pwq(work)); 3084 } 3085 3086 raw_spin_unlock(&wq_mayday_lock); 3087 raw_spin_unlock_irq(&pool->lock); 3088 3089 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 3090 } 3091 3092 /** 3093 * maybe_create_worker - create a new worker if necessary 3094 * @pool: pool to create a new worker for 3095 * 3096 * Create a new worker for @pool if necessary. @pool is guaranteed to 3097 * have at least one idle worker on return from this function. If 3098 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 3099 * sent to all rescuers with works scheduled on @pool to resolve 3100 * possible allocation deadlock. 3101 * 3102 * On return, need_to_create_worker() is guaranteed to be %false and 3103 * may_start_working() %true. 3104 * 3105 * LOCKING: 3106 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3107 * multiple times. Does GFP_KERNEL allocations. Called only from 3108 * manager. 3109 */ 3110 static void maybe_create_worker(struct worker_pool *pool) 3111 __releases(&pool->lock) 3112 __acquires(&pool->lock) 3113 { 3114 restart: 3115 raw_spin_unlock_irq(&pool->lock); 3116 3117 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 3118 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 3119 3120 while (true) { 3121 if (create_worker(pool) || !need_to_create_worker(pool)) 3122 break; 3123 3124 schedule_timeout_interruptible(CREATE_COOLDOWN); 3125 3126 if (!need_to_create_worker(pool)) 3127 break; 3128 } 3129 3130 timer_delete_sync(&pool->mayday_timer); 3131 raw_spin_lock_irq(&pool->lock); 3132 /* 3133 * This is necessary even after a new worker was just successfully 3134 * created as @pool->lock was dropped and the new worker might have 3135 * already become busy. 3136 */ 3137 if (need_to_create_worker(pool)) 3138 goto restart; 3139 } 3140 3141 #ifdef CONFIG_PREEMPT_RT 3142 static void worker_lock_callback(struct worker_pool *pool) 3143 { 3144 spin_lock(&pool->cb_lock); 3145 } 3146 3147 static void worker_unlock_callback(struct worker_pool *pool) 3148 { 3149 spin_unlock(&pool->cb_lock); 3150 } 3151 3152 static void workqueue_callback_cancel_wait_running(struct worker_pool *pool) 3153 { 3154 spin_lock(&pool->cb_lock); 3155 spin_unlock(&pool->cb_lock); 3156 } 3157 3158 #else 3159 3160 static void worker_lock_callback(struct worker_pool *pool) { } 3161 static void worker_unlock_callback(struct worker_pool *pool) { } 3162 static void workqueue_callback_cancel_wait_running(struct worker_pool *pool) { } 3163 3164 #endif 3165 3166 /** 3167 * manage_workers - manage worker pool 3168 * @worker: self 3169 * 3170 * Assume the manager role and manage the worker pool @worker belongs 3171 * to. At any given time, there can be only zero or one manager per 3172 * pool. The exclusion is handled automatically by this function. 3173 * 3174 * The caller can safely start processing works on false return. On 3175 * true return, it's guaranteed that need_to_create_worker() is false 3176 * and may_start_working() is true. 3177 * 3178 * CONTEXT: 3179 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3180 * multiple times. Does GFP_KERNEL allocations. 3181 * 3182 * Return: 3183 * %false if the pool doesn't need management and the caller can safely 3184 * start processing works, %true if management function was performed and 3185 * the conditions that the caller verified before calling the function may 3186 * no longer be true. 3187 */ 3188 static bool manage_workers(struct worker *worker) 3189 { 3190 struct worker_pool *pool = worker->pool; 3191 3192 if (pool->flags & POOL_MANAGER_ACTIVE) 3193 return false; 3194 3195 pool->flags |= POOL_MANAGER_ACTIVE; 3196 pool->manager = worker; 3197 3198 maybe_create_worker(pool); 3199 3200 pool->manager = NULL; 3201 pool->flags &= ~POOL_MANAGER_ACTIVE; 3202 rcuwait_wake_up(&manager_wait); 3203 return true; 3204 } 3205 3206 /** 3207 * process_one_work - process single work 3208 * @worker: self 3209 * @work: work to process 3210 * 3211 * Process @work. This function contains all the logics necessary to 3212 * process a single work including synchronization against and 3213 * interaction with other workers on the same cpu, queueing and 3214 * flushing. As long as context requirement is met, any worker can 3215 * call this function to process a work. 3216 * 3217 * CONTEXT: 3218 * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 3219 */ 3220 static void process_one_work(struct worker *worker, struct work_struct *work) 3221 __releases(&pool->lock) 3222 __acquires(&pool->lock) 3223 { 3224 struct pool_workqueue *pwq = get_work_pwq(work); 3225 struct worker_pool *pool = worker->pool; 3226 unsigned long work_data; 3227 int lockdep_start_depth, rcu_start_depth; 3228 bool bh_draining = pool->flags & POOL_BH_DRAINING; 3229 #ifdef CONFIG_LOCKDEP 3230 /* 3231 * It is permissible to free the struct work_struct from 3232 * inside the function that is called from it, this we need to 3233 * take into account for lockdep too. To avoid bogus "held 3234 * lock freed" warnings as well as problems when looking into 3235 * work->lockdep_map, make a copy and use that here. 3236 */ 3237 struct lockdep_map lockdep_map; 3238 3239 lockdep_copy_map(&lockdep_map, &work->lockdep_map); 3240 #endif 3241 /* ensure we're on the correct CPU */ 3242 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 3243 raw_smp_processor_id() != pool->cpu); 3244 3245 /* claim and dequeue */ 3246 debug_work_deactivate(work); 3247 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 3248 worker->current_work = work; 3249 worker->current_func = work->func; 3250 worker->current_pwq = pwq; 3251 if (worker->task) 3252 worker->current_at = worker->task->se.sum_exec_runtime; 3253 worker->current_start = jiffies; 3254 work_data = *work_data_bits(work); 3255 worker->current_color = get_work_color(work_data); 3256 3257 /* 3258 * Record wq name for cmdline and debug reporting, may get 3259 * overridden through set_worker_desc(). 3260 */ 3261 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 3262 3263 list_del_init(&work->entry); 3264 3265 /* 3266 * CPU intensive works don't participate in concurrency management. 3267 * They're the scheduler's responsibility. This takes @worker out 3268 * of concurrency management and the next code block will chain 3269 * execution of the pending work items. 3270 */ 3271 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 3272 worker_set_flags(worker, WORKER_CPU_INTENSIVE); 3273 3274 /* 3275 * Kick @pool if necessary. It's always noop for per-cpu worker pools 3276 * since nr_running would always be >= 1 at this point. This is used to 3277 * chain execution of the pending work items for WORKER_NOT_RUNNING 3278 * workers such as the UNBOUND and CPU_INTENSIVE ones. 3279 */ 3280 kick_pool(pool); 3281 3282 /* 3283 * Record the last pool and clear PENDING which should be the last 3284 * update to @work. Also, do this inside @pool->lock so that 3285 * PENDING and queued state changes happen together while IRQ is 3286 * disabled. 3287 */ 3288 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool)); 3289 3290 pwq->stats[PWQ_STAT_STARTED]++; 3291 raw_spin_unlock_irq(&pool->lock); 3292 3293 rcu_start_depth = rcu_preempt_depth(); 3294 lockdep_start_depth = lockdep_depth(current); 3295 /* see drain_dead_softirq_workfn() */ 3296 if (!bh_draining) 3297 lock_map_acquire(pwq->wq->lockdep_map); 3298 lock_map_acquire(&lockdep_map); 3299 /* 3300 * Strictly speaking we should mark the invariant state without holding 3301 * any locks, that is, before these two lock_map_acquire()'s. 3302 * 3303 * However, that would result in: 3304 * 3305 * A(W1) 3306 * WFC(C) 3307 * A(W1) 3308 * C(C) 3309 * 3310 * Which would create W1->C->W1 dependencies, even though there is no 3311 * actual deadlock possible. There are two solutions, using a 3312 * read-recursive acquire on the work(queue) 'locks', but this will then 3313 * hit the lockdep limitation on recursive locks, or simply discard 3314 * these locks. 3315 * 3316 * AFAICT there is no possible deadlock scenario between the 3317 * flush_work() and complete() primitives (except for single-threaded 3318 * workqueues), so hiding them isn't a problem. 3319 */ 3320 lockdep_invariant_state(true); 3321 trace_workqueue_execute_start(work); 3322 worker->current_func(work); 3323 /* 3324 * While we must be careful to not use "work" after this, the trace 3325 * point will only record its address. 3326 */ 3327 trace_workqueue_execute_end(work, worker->current_func); 3328 3329 lock_map_release(&lockdep_map); 3330 if (!bh_draining) 3331 lock_map_release(pwq->wq->lockdep_map); 3332 3333 if (unlikely((worker->task && in_atomic()) || 3334 lockdep_depth(current) != lockdep_start_depth || 3335 rcu_preempt_depth() != rcu_start_depth)) { 3336 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n" 3337 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n", 3338 current->comm, task_pid_nr(current), preempt_count(), 3339 lockdep_start_depth, lockdep_depth(current), 3340 rcu_start_depth, rcu_preempt_depth(), 3341 worker->current_func); 3342 debug_show_held_locks(current); 3343 dump_stack(); 3344 } 3345 3346 /* 3347 * The following prevents a kworker from hogging CPU on !PREEMPTION 3348 * kernels, where a requeueing work item waiting for something to 3349 * happen could deadlock with stop_machine as such work item could 3350 * indefinitely requeue itself while all other CPUs are trapped in 3351 * stop_machine. At the same time, report a quiescent RCU state so 3352 * the same condition doesn't freeze RCU. 3353 */ 3354 if (worker->task) 3355 cond_resched(); 3356 3357 raw_spin_lock_irq(&pool->lock); 3358 3359 pwq->stats[PWQ_STAT_COMPLETED]++; 3360 3361 /* 3362 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 3363 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 3364 * wq_cpu_intensive_thresh_us. Clear it. 3365 */ 3366 worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 3367 3368 /* tag the worker for identification in schedule() */ 3369 worker->last_func = worker->current_func; 3370 3371 /* we're done with it, release */ 3372 hash_del(&worker->hentry); 3373 worker->current_work = NULL; 3374 worker->current_func = NULL; 3375 worker->current_pwq = NULL; 3376 worker->current_color = INT_MAX; 3377 3378 /* must be the last step, see the function comment */ 3379 pwq_dec_nr_in_flight(pwq, work_data); 3380 } 3381 3382 /** 3383 * process_scheduled_works - process scheduled works 3384 * @worker: self 3385 * 3386 * Process all scheduled works. Please note that the scheduled list 3387 * may change while processing a work, so this function repeatedly 3388 * fetches a work from the top and executes it. 3389 * 3390 * CONTEXT: 3391 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 3392 * multiple times. 3393 */ 3394 static void process_scheduled_works(struct worker *worker) 3395 { 3396 struct work_struct *work; 3397 bool first = true; 3398 3399 while ((work = list_first_entry_or_null(&worker->scheduled, 3400 struct work_struct, entry))) { 3401 if (first) { 3402 worker->pool->last_progress_ts = jiffies; 3403 first = false; 3404 } 3405 process_one_work(worker, work); 3406 } 3407 } 3408 3409 static void set_pf_worker(bool val) 3410 { 3411 mutex_lock(&wq_pool_attach_mutex); 3412 if (val) 3413 current->flags |= PF_WQ_WORKER; 3414 else 3415 current->flags &= ~PF_WQ_WORKER; 3416 mutex_unlock(&wq_pool_attach_mutex); 3417 } 3418 3419 /** 3420 * worker_thread - the worker thread function 3421 * @__worker: self 3422 * 3423 * The worker thread function. All workers belong to a worker_pool - 3424 * either a per-cpu one or dynamic unbound one. These workers process all 3425 * work items regardless of their specific target workqueue. The only 3426 * exception is work items which belong to workqueues with a rescuer which 3427 * will be explained in rescuer_thread(). 3428 * 3429 * Return: 0 3430 */ 3431 static int worker_thread(void *__worker) 3432 { 3433 struct worker *worker = __worker; 3434 struct worker_pool *pool = worker->pool; 3435 3436 /* tell the scheduler that this is a workqueue worker */ 3437 set_pf_worker(true); 3438 woke_up: 3439 raw_spin_lock_irq(&pool->lock); 3440 3441 /* am I supposed to die? */ 3442 if (unlikely(worker->flags & WORKER_DIE)) { 3443 raw_spin_unlock_irq(&pool->lock); 3444 set_pf_worker(false); 3445 /* 3446 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool 3447 * shouldn't be accessed, reset it to NULL in case otherwise. 3448 */ 3449 worker->pool = NULL; 3450 ida_free(&pool->worker_ida, worker->id); 3451 return 0; 3452 } 3453 3454 worker_leave_idle(worker); 3455 recheck: 3456 /* no more worker necessary? */ 3457 if (!need_more_worker(pool)) 3458 goto sleep; 3459 3460 /* do we need to manage? */ 3461 if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 3462 goto recheck; 3463 3464 /* 3465 * ->scheduled list can only be filled while a worker is 3466 * preparing to process a work or actually processing it. 3467 * Make sure nobody diddled with it while I was sleeping. 3468 */ 3469 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3470 3471 /* 3472 * Finish PREP stage. We're guaranteed to have at least one idle 3473 * worker or that someone else has already assumed the manager 3474 * role. This is where @worker starts participating in concurrency 3475 * management if applicable and concurrency management is restored 3476 * after being rebound. See rebind_workers() for details. 3477 */ 3478 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3479 3480 do { 3481 struct work_struct *work = 3482 list_first_entry(&pool->worklist, 3483 struct work_struct, entry); 3484 3485 if (assign_work(work, worker, NULL)) 3486 process_scheduled_works(worker); 3487 } while (keep_working(pool)); 3488 3489 worker_set_flags(worker, WORKER_PREP); 3490 sleep: 3491 /* 3492 * pool->lock is held and there's no work to process and no need to 3493 * manage, sleep. Workers are woken up only while holding 3494 * pool->lock or from local cpu, so setting the current state 3495 * before releasing pool->lock is enough to prevent losing any 3496 * event. 3497 */ 3498 worker_enter_idle(worker); 3499 __set_current_state(TASK_IDLE); 3500 raw_spin_unlock_irq(&pool->lock); 3501 schedule(); 3502 goto woke_up; 3503 } 3504 3505 static bool assign_rescuer_work(struct pool_workqueue *pwq, struct worker *rescuer) 3506 { 3507 struct worker_pool *pool = pwq->pool; 3508 struct work_struct *cursor = &pwq->mayday_cursor; 3509 struct work_struct *work, *n; 3510 3511 /* have work items to rescue? */ 3512 if (!pwq->nr_active) 3513 return false; 3514 3515 /* need rescue? */ 3516 if (!need_to_create_worker(pool)) { 3517 /* 3518 * The pool has idle workers and doesn't need the rescuer, so it 3519 * could simply return false here. 3520 * 3521 * However, the memory pressure might not be fully relieved. 3522 * In PERCPU pool with concurrency enabled, having idle workers 3523 * does not necessarily mean memory pressure is gone; it may 3524 * simply mean regular workers have woken up, completed their 3525 * work, and gone idle again due to concurrency limits. 3526 * 3527 * In this case, those working workers may later sleep again, 3528 * the pool may run out of idle workers, and it will have to 3529 * allocate new ones and wait for the timer to send mayday, 3530 * causing unnecessary delay - especially if memory pressure 3531 * was never resolved throughout. 3532 * 3533 * Do more work if memory pressure is still on to reduce 3534 * relapse, using (pool->flags & POOL_MANAGER_ACTIVE), though 3535 * not precisely, unless there are other PWQs needing help. 3536 */ 3537 if (!(pool->flags & POOL_MANAGER_ACTIVE) || 3538 !list_empty(&pwq->wq->maydays)) 3539 return false; 3540 } 3541 3542 /* search from the start or cursor if available */ 3543 if (list_empty(&cursor->entry)) 3544 work = list_first_entry(&pool->worklist, struct work_struct, entry); 3545 else 3546 work = list_next_entry(cursor, entry); 3547 3548 /* find the next work item to rescue */ 3549 list_for_each_entry_safe_from(work, n, &pool->worklist, entry) { 3550 if (get_work_pwq(work) == pwq && assign_work(work, rescuer, &n)) { 3551 pwq->stats[PWQ_STAT_RESCUED]++; 3552 /* put the cursor for next search */ 3553 list_move_tail(&cursor->entry, &n->entry); 3554 return true; 3555 } 3556 } 3557 3558 return false; 3559 } 3560 3561 /** 3562 * rescuer_thread - the rescuer thread function 3563 * @__rescuer: self 3564 * 3565 * Workqueue rescuer thread function. There's one rescuer for each 3566 * workqueue which has WQ_MEM_RECLAIM set. 3567 * 3568 * Regular work processing on a pool may block trying to create a new 3569 * worker which uses GFP_KERNEL allocation which has slight chance of 3570 * developing into deadlock if some works currently on the same queue 3571 * need to be processed to satisfy the GFP_KERNEL allocation. This is 3572 * the problem rescuer solves. 3573 * 3574 * When such condition is possible, the pool summons rescuers of all 3575 * workqueues which have works queued on the pool and let them process 3576 * those works so that forward progress can be guaranteed. 3577 * 3578 * This should happen rarely. 3579 * 3580 * Return: 0 3581 */ 3582 static int rescuer_thread(void *__rescuer) 3583 { 3584 struct worker *rescuer = __rescuer; 3585 struct workqueue_struct *wq = rescuer->rescue_wq; 3586 bool should_stop; 3587 3588 set_user_nice(current, RESCUER_NICE_LEVEL); 3589 3590 /* 3591 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 3592 * doesn't participate in concurrency management. 3593 */ 3594 set_pf_worker(true); 3595 repeat: 3596 set_current_state(TASK_IDLE); 3597 3598 /* 3599 * By the time the rescuer is requested to stop, the workqueue 3600 * shouldn't have any work pending, but @wq->maydays may still have 3601 * pwq(s) queued. This can happen by non-rescuer workers consuming 3602 * all the work items before the rescuer got to them. Go through 3603 * @wq->maydays processing before acting on should_stop so that the 3604 * list is always empty on exit. 3605 */ 3606 should_stop = kthread_should_stop(); 3607 3608 /* see whether any pwq is asking for help */ 3609 raw_spin_lock_irq(&wq_mayday_lock); 3610 3611 while (!list_empty(&wq->maydays)) { 3612 struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 3613 struct pool_workqueue, mayday_node); 3614 struct worker_pool *pool = pwq->pool; 3615 unsigned int count = 0; 3616 3617 __set_current_state(TASK_RUNNING); 3618 list_del_init(&pwq->mayday_node); 3619 3620 raw_spin_unlock_irq(&wq_mayday_lock); 3621 3622 worker_attach_to_pool(rescuer, pool); 3623 3624 raw_spin_lock_irq(&pool->lock); 3625 3626 WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 3627 3628 while (assign_rescuer_work(pwq, rescuer)) { 3629 process_scheduled_works(rescuer); 3630 3631 /* 3632 * If the per-turn work item limit is reached and other 3633 * PWQs are in mayday, requeue mayday for this PWQ and 3634 * let the rescuer handle the other PWQs first. 3635 */ 3636 if (++count > RESCUER_BATCH && !list_empty(&pwq->wq->maydays) && 3637 pwq->nr_active && need_to_create_worker(pool)) { 3638 raw_spin_lock(&wq_mayday_lock); 3639 send_mayday(pwq); 3640 raw_spin_unlock(&wq_mayday_lock); 3641 break; 3642 } 3643 } 3644 3645 /* The cursor can not be left behind without the rescuer watching it. */ 3646 if (!list_empty(&pwq->mayday_cursor.entry) && list_empty(&pwq->mayday_node)) 3647 list_del_init(&pwq->mayday_cursor.entry); 3648 3649 /* 3650 * Leave this pool. Notify regular workers; otherwise, we end up 3651 * with 0 concurrency and stalling the execution. 3652 */ 3653 kick_pool(pool); 3654 3655 raw_spin_unlock_irq(&pool->lock); 3656 3657 worker_detach_from_pool(rescuer); 3658 3659 /* 3660 * Put the reference grabbed by send_mayday(). @pool might 3661 * go away any time after it. 3662 */ 3663 put_pwq_unlocked(pwq); 3664 3665 raw_spin_lock_irq(&wq_mayday_lock); 3666 } 3667 3668 raw_spin_unlock_irq(&wq_mayday_lock); 3669 3670 if (should_stop) { 3671 __set_current_state(TASK_RUNNING); 3672 set_pf_worker(false); 3673 return 0; 3674 } 3675 3676 /* rescuers should never participate in concurrency management */ 3677 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 3678 schedule(); 3679 goto repeat; 3680 } 3681 3682 static void bh_worker(struct worker *worker) 3683 { 3684 struct worker_pool *pool = worker->pool; 3685 int nr_restarts = BH_WORKER_RESTARTS; 3686 unsigned long end = jiffies + BH_WORKER_JIFFIES; 3687 3688 worker_lock_callback(pool); 3689 raw_spin_lock_irq(&pool->lock); 3690 worker_leave_idle(worker); 3691 3692 /* 3693 * This function follows the structure of worker_thread(). See there for 3694 * explanations on each step. 3695 */ 3696 if (!need_more_worker(pool)) 3697 goto done; 3698 3699 WARN_ON_ONCE(!list_empty(&worker->scheduled)); 3700 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 3701 3702 do { 3703 struct work_struct *work = 3704 list_first_entry(&pool->worklist, 3705 struct work_struct, entry); 3706 3707 if (assign_work(work, worker, NULL)) 3708 process_scheduled_works(worker); 3709 } while (keep_working(pool) && 3710 --nr_restarts && time_before(jiffies, end)); 3711 3712 worker_set_flags(worker, WORKER_PREP); 3713 done: 3714 worker_enter_idle(worker); 3715 kick_pool(pool); 3716 raw_spin_unlock_irq(&pool->lock); 3717 worker_unlock_callback(pool); 3718 } 3719 3720 /* 3721 * TODO: Convert all tasklet users to workqueue and use softirq directly. 3722 * 3723 * This is currently called from tasklet[_hi]action() and thus is also called 3724 * whenever there are tasklets to run. Let's do an early exit if there's nothing 3725 * queued. Once conversion from tasklet is complete, the need_more_worker() test 3726 * can be dropped. 3727 * 3728 * After full conversion, we'll add worker->softirq_action, directly use the 3729 * softirq action and obtain the worker pointer from the softirq_action pointer. 3730 */ 3731 void workqueue_softirq_action(bool highpri) 3732 { 3733 struct worker_pool *pool = 3734 &per_cpu(bh_worker_pools, smp_processor_id())[highpri]; 3735 if (need_more_worker(pool)) 3736 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3737 } 3738 3739 struct wq_drain_dead_softirq_work { 3740 struct work_struct work; 3741 struct worker_pool *pool; 3742 struct completion done; 3743 }; 3744 3745 static void drain_dead_softirq_workfn(struct work_struct *work) 3746 { 3747 struct wq_drain_dead_softirq_work *dead_work = 3748 container_of(work, struct wq_drain_dead_softirq_work, work); 3749 struct worker_pool *pool = dead_work->pool; 3750 bool repeat; 3751 3752 /* 3753 * @pool's CPU is dead and we want to execute its still pending work 3754 * items from this BH work item which is running on a different CPU. As 3755 * its CPU is dead, @pool can't be kicked and, as work execution path 3756 * will be nested, a lockdep annotation needs to be suppressed. Mark 3757 * @pool with %POOL_BH_DRAINING for the special treatments. 3758 */ 3759 raw_spin_lock_irq(&pool->lock); 3760 pool->flags |= POOL_BH_DRAINING; 3761 raw_spin_unlock_irq(&pool->lock); 3762 3763 bh_worker(list_first_entry(&pool->workers, struct worker, node)); 3764 3765 raw_spin_lock_irq(&pool->lock); 3766 pool->flags &= ~POOL_BH_DRAINING; 3767 repeat = need_more_worker(pool); 3768 raw_spin_unlock_irq(&pool->lock); 3769 3770 /* 3771 * bh_worker() might hit consecutive execution limit and bail. If there 3772 * still are pending work items, reschedule self and return so that we 3773 * don't hog this CPU's BH. 3774 */ 3775 if (repeat) { 3776 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3777 queue_work(system_bh_highpri_wq, work); 3778 else 3779 queue_work(system_bh_wq, work); 3780 } else { 3781 complete(&dead_work->done); 3782 } 3783 } 3784 3785 /* 3786 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's 3787 * possible to allocate dead_work per CPU and avoid flushing. However, then we 3788 * have to worry about draining overlapping with CPU coming back online or 3789 * nesting (one CPU's dead_work queued on another CPU which is also dead and so 3790 * on). Let's keep it simple and drain them synchronously. These are BH work 3791 * items which shouldn't be requeued on the same pool. Shouldn't take long. 3792 */ 3793 void workqueue_softirq_dead(unsigned int cpu) 3794 { 3795 int i; 3796 3797 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 3798 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i]; 3799 struct wq_drain_dead_softirq_work dead_work; 3800 3801 if (!need_more_worker(pool)) 3802 continue; 3803 3804 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn); 3805 dead_work.pool = pool; 3806 init_completion(&dead_work.done); 3807 3808 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL) 3809 queue_work(system_bh_highpri_wq, &dead_work.work); 3810 else 3811 queue_work(system_bh_wq, &dead_work.work); 3812 3813 wait_for_completion(&dead_work.done); 3814 destroy_work_on_stack(&dead_work.work); 3815 } 3816 } 3817 3818 /** 3819 * check_flush_dependency - check for flush dependency sanity 3820 * @target_wq: workqueue being flushed 3821 * @target_work: work item being flushed (NULL for workqueue flushes) 3822 * @from_cancel: are we called from the work cancel path 3823 * 3824 * %current is trying to flush the whole @target_wq or @target_work on it. 3825 * If this is not the cancel path (which implies work being flushed is either 3826 * already running, or will not be at all), check if @target_wq doesn't have 3827 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running 3828 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward- 3829 * progress guarantee leading to a deadlock. 3830 */ 3831 static void check_flush_dependency(struct workqueue_struct *target_wq, 3832 struct work_struct *target_work, 3833 bool from_cancel) 3834 { 3835 work_func_t target_func; 3836 struct worker *worker; 3837 3838 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM) 3839 return; 3840 3841 worker = current_wq_worker(); 3842 target_func = target_work ? target_work->func : NULL; 3843 3844 WARN_ONCE(current->flags & PF_MEMALLOC, 3845 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 3846 current->pid, current->comm, target_wq->name, target_func); 3847 WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 3848 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 3849 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 3850 worker->current_pwq->wq->name, worker->current_func, 3851 target_wq->name, target_func); 3852 } 3853 3854 struct wq_barrier { 3855 struct work_struct work; 3856 struct completion done; 3857 struct task_struct *task; /* purely informational */ 3858 }; 3859 3860 static void wq_barrier_func(struct work_struct *work) 3861 { 3862 struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 3863 complete(&barr->done); 3864 } 3865 3866 /** 3867 * insert_wq_barrier - insert a barrier work 3868 * @pwq: pwq to insert barrier into 3869 * @barr: wq_barrier to insert 3870 * @target: target work to attach @barr to 3871 * @worker: worker currently executing @target, NULL if @target is not executing 3872 * 3873 * @barr is linked to @target such that @barr is completed only after 3874 * @target finishes execution. Please note that the ordering 3875 * guarantee is observed only with respect to @target and on the local 3876 * cpu. 3877 * 3878 * Currently, a queued barrier can't be canceled. This is because 3879 * try_to_grab_pending() can't determine whether the work to be 3880 * grabbed is at the head of the queue and thus can't clear LINKED 3881 * flag of the previous work while there must be a valid next work 3882 * after a work with LINKED flag set. 3883 * 3884 * Note that when @worker is non-NULL, @target may be modified 3885 * underneath us, so we can't reliably determine pwq from @target. 3886 * 3887 * CONTEXT: 3888 * raw_spin_lock_irq(pool->lock). 3889 */ 3890 static void insert_wq_barrier(struct pool_workqueue *pwq, 3891 struct wq_barrier *barr, 3892 struct work_struct *target, struct worker *worker) 3893 { 3894 static __maybe_unused struct lock_class_key bh_key, thr_key; 3895 unsigned int work_flags = 0; 3896 unsigned int work_color; 3897 struct list_head *head; 3898 3899 /* 3900 * debugobject calls are safe here even with pool->lock locked 3901 * as we know for sure that this will not trigger any of the 3902 * checks and call back into the fixup functions where we 3903 * might deadlock. 3904 * 3905 * BH and threaded workqueues need separate lockdep keys to avoid 3906 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} 3907 * usage". 3908 */ 3909 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func, 3910 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key); 3911 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 3912 3913 init_completion_map(&barr->done, &target->lockdep_map); 3914 3915 barr->task = current; 3916 3917 /* The barrier work item does not participate in nr_active. */ 3918 work_flags |= WORK_STRUCT_INACTIVE; 3919 3920 /* 3921 * If @target is currently being executed, schedule the 3922 * barrier to the worker; otherwise, put it after @target. 3923 */ 3924 if (worker) { 3925 head = worker->scheduled.next; 3926 work_color = worker->current_color; 3927 } else { 3928 unsigned long *bits = work_data_bits(target); 3929 3930 head = target->entry.next; 3931 /* there can already be other linked works, inherit and set */ 3932 work_flags |= *bits & WORK_STRUCT_LINKED; 3933 work_color = get_work_color(*bits); 3934 __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3935 } 3936 3937 pwq->nr_in_flight[work_color]++; 3938 work_flags |= work_color_to_flags(work_color); 3939 3940 insert_work(pwq, &barr->work, head, work_flags); 3941 } 3942 3943 /** 3944 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 3945 * @wq: workqueue being flushed 3946 * @flush_color: new flush color, < 0 for no-op 3947 * @work_color: new work color, < 0 for no-op 3948 * 3949 * Prepare pwqs for workqueue flushing. 3950 * 3951 * If @flush_color is non-negative, flush_color on all pwqs should be 3952 * -1. If no pwq has in-flight commands at the specified color, all 3953 * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3954 * has in flight commands, its pwq->flush_color is set to 3955 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 3956 * wakeup logic is armed and %true is returned. 3957 * 3958 * The caller should have initialized @wq->first_flusher prior to 3959 * calling this function with non-negative @flush_color. If 3960 * @flush_color is negative, no flush color update is done and %false 3961 * is returned. 3962 * 3963 * If @work_color is non-negative, all pwqs should have the same 3964 * work_color which is previous to @work_color and all will be 3965 * advanced to @work_color. 3966 * 3967 * CONTEXT: 3968 * mutex_lock(wq->mutex). 3969 * 3970 * Return: 3971 * %true if @flush_color >= 0 and there's something to flush. %false 3972 * otherwise. 3973 */ 3974 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 3975 int flush_color, int work_color) 3976 { 3977 bool wait = false; 3978 struct pool_workqueue *pwq; 3979 struct worker_pool *current_pool = NULL; 3980 3981 if (flush_color >= 0) { 3982 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3983 atomic_set(&wq->nr_pwqs_to_flush, 1); 3984 } 3985 3986 /* 3987 * For unbound workqueue, pwqs will map to only a few pools. 3988 * Most of the time, pwqs within the same pool will be linked 3989 * sequentially to wq->pwqs by cpu index. So in the majority 3990 * of pwq iters, the pool is the same, only doing lock/unlock 3991 * if the pool has changed. This can largely reduce expensive 3992 * lock operations. 3993 */ 3994 for_each_pwq(pwq, wq) { 3995 if (current_pool != pwq->pool) { 3996 if (likely(current_pool)) 3997 raw_spin_unlock_irq(¤t_pool->lock); 3998 current_pool = pwq->pool; 3999 raw_spin_lock_irq(¤t_pool->lock); 4000 } 4001 4002 if (flush_color >= 0) { 4003 WARN_ON_ONCE(pwq->flush_color != -1); 4004 4005 if (pwq->nr_in_flight[flush_color]) { 4006 pwq->flush_color = flush_color; 4007 atomic_inc(&wq->nr_pwqs_to_flush); 4008 wait = true; 4009 } 4010 } 4011 4012 if (work_color >= 0) { 4013 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 4014 pwq->work_color = work_color; 4015 } 4016 4017 } 4018 4019 if (current_pool) 4020 raw_spin_unlock_irq(¤t_pool->lock); 4021 4022 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 4023 complete(&wq->first_flusher->done); 4024 4025 return wait; 4026 } 4027 4028 static void touch_wq_lockdep_map(struct workqueue_struct *wq) 4029 { 4030 #ifdef CONFIG_LOCKDEP 4031 if (unlikely(!wq->lockdep_map)) 4032 return; 4033 4034 if (wq->flags & WQ_BH) 4035 local_bh_disable(); 4036 4037 lock_map_acquire(wq->lockdep_map); 4038 lock_map_release(wq->lockdep_map); 4039 4040 if (wq->flags & WQ_BH) 4041 local_bh_enable(); 4042 #endif 4043 } 4044 4045 static void touch_work_lockdep_map(struct work_struct *work, 4046 struct workqueue_struct *wq) 4047 { 4048 #ifdef CONFIG_LOCKDEP 4049 if (wq->flags & WQ_BH) 4050 local_bh_disable(); 4051 4052 lock_map_acquire(&work->lockdep_map); 4053 lock_map_release(&work->lockdep_map); 4054 4055 if (wq->flags & WQ_BH) 4056 local_bh_enable(); 4057 #endif 4058 } 4059 4060 /** 4061 * __flush_workqueue - ensure that any scheduled work has run to completion. 4062 * @wq: workqueue to flush 4063 * 4064 * This function sleeps until all work items which were queued on entry 4065 * have finished execution, but it is not livelocked by new incoming ones. 4066 */ 4067 void __flush_workqueue(struct workqueue_struct *wq) 4068 { 4069 struct wq_flusher this_flusher = { 4070 .list = LIST_HEAD_INIT(this_flusher.list), 4071 .flush_color = -1, 4072 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)), 4073 }; 4074 int next_color; 4075 4076 if (WARN_ON(!wq_online)) 4077 return; 4078 4079 touch_wq_lockdep_map(wq); 4080 4081 mutex_lock(&wq->mutex); 4082 4083 /* 4084 * Start-to-wait phase 4085 */ 4086 next_color = work_next_color(wq->work_color); 4087 4088 if (next_color != wq->flush_color) { 4089 /* 4090 * Color space is not full. The current work_color 4091 * becomes our flush_color and work_color is advanced 4092 * by one. 4093 */ 4094 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 4095 this_flusher.flush_color = wq->work_color; 4096 wq->work_color = next_color; 4097 4098 if (!wq->first_flusher) { 4099 /* no flush in progress, become the first flusher */ 4100 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 4101 4102 wq->first_flusher = &this_flusher; 4103 4104 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 4105 wq->work_color)) { 4106 /* nothing to flush, done */ 4107 wq->flush_color = next_color; 4108 wq->first_flusher = NULL; 4109 goto out_unlock; 4110 } 4111 } else { 4112 /* wait in queue */ 4113 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 4114 list_add_tail(&this_flusher.list, &wq->flusher_queue); 4115 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 4116 } 4117 } else { 4118 /* 4119 * Oops, color space is full, wait on overflow queue. 4120 * The next flush completion will assign us 4121 * flush_color and transfer to flusher_queue. 4122 */ 4123 list_add_tail(&this_flusher.list, &wq->flusher_overflow); 4124 } 4125 4126 check_flush_dependency(wq, NULL, false); 4127 4128 mutex_unlock(&wq->mutex); 4129 4130 wait_for_completion(&this_flusher.done); 4131 4132 /* 4133 * Wake-up-and-cascade phase 4134 * 4135 * First flushers are responsible for cascading flushes and 4136 * handling overflow. Non-first flushers can simply return. 4137 */ 4138 if (READ_ONCE(wq->first_flusher) != &this_flusher) 4139 return; 4140 4141 mutex_lock(&wq->mutex); 4142 4143 /* we might have raced, check again with mutex held */ 4144 if (wq->first_flusher != &this_flusher) 4145 goto out_unlock; 4146 4147 WRITE_ONCE(wq->first_flusher, NULL); 4148 4149 WARN_ON_ONCE(!list_empty(&this_flusher.list)); 4150 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 4151 4152 while (true) { 4153 struct wq_flusher *next, *tmp; 4154 4155 /* complete all the flushers sharing the current flush color */ 4156 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 4157 if (next->flush_color != wq->flush_color) 4158 break; 4159 list_del_init(&next->list); 4160 complete(&next->done); 4161 } 4162 4163 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 4164 wq->flush_color != work_next_color(wq->work_color)); 4165 4166 /* this flush_color is finished, advance by one */ 4167 wq->flush_color = work_next_color(wq->flush_color); 4168 4169 /* one color has been freed, handle overflow queue */ 4170 if (!list_empty(&wq->flusher_overflow)) { 4171 /* 4172 * Assign the same color to all overflowed 4173 * flushers, advance work_color and append to 4174 * flusher_queue. This is the start-to-wait 4175 * phase for these overflowed flushers. 4176 */ 4177 list_for_each_entry(tmp, &wq->flusher_overflow, list) 4178 tmp->flush_color = wq->work_color; 4179 4180 wq->work_color = work_next_color(wq->work_color); 4181 4182 list_splice_tail_init(&wq->flusher_overflow, 4183 &wq->flusher_queue); 4184 flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 4185 } 4186 4187 if (list_empty(&wq->flusher_queue)) { 4188 WARN_ON_ONCE(wq->flush_color != wq->work_color); 4189 break; 4190 } 4191 4192 /* 4193 * Need to flush more colors. Make the next flusher 4194 * the new first flusher and arm pwqs. 4195 */ 4196 WARN_ON_ONCE(wq->flush_color == wq->work_color); 4197 WARN_ON_ONCE(wq->flush_color != next->flush_color); 4198 4199 list_del_init(&next->list); 4200 wq->first_flusher = next; 4201 4202 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 4203 break; 4204 4205 /* 4206 * Meh... this color is already done, clear first 4207 * flusher and repeat cascading. 4208 */ 4209 wq->first_flusher = NULL; 4210 } 4211 4212 out_unlock: 4213 mutex_unlock(&wq->mutex); 4214 } 4215 EXPORT_SYMBOL(__flush_workqueue); 4216 4217 /** 4218 * drain_workqueue - drain a workqueue 4219 * @wq: workqueue to drain 4220 * 4221 * Wait until the workqueue becomes empty. While draining is in progress, 4222 * only chain queueing is allowed. IOW, only currently pending or running 4223 * work items on @wq can queue further work items on it. @wq is flushed 4224 * repeatedly until it becomes empty. The number of flushing is determined 4225 * by the depth of chaining and should be relatively short. Whine if it 4226 * takes too long. 4227 */ 4228 void drain_workqueue(struct workqueue_struct *wq) 4229 { 4230 unsigned int flush_cnt = 0; 4231 struct pool_workqueue *pwq; 4232 4233 /* 4234 * __queue_work() needs to test whether there are drainers, is much 4235 * hotter than drain_workqueue() and already looks at @wq->flags. 4236 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 4237 */ 4238 mutex_lock(&wq->mutex); 4239 if (!wq->nr_drainers++) 4240 wq->flags |= __WQ_DRAINING; 4241 mutex_unlock(&wq->mutex); 4242 reflush: 4243 __flush_workqueue(wq); 4244 4245 mutex_lock(&wq->mutex); 4246 4247 for_each_pwq(pwq, wq) { 4248 bool drained; 4249 4250 raw_spin_lock_irq(&pwq->pool->lock); 4251 drained = pwq_is_empty(pwq); 4252 raw_spin_unlock_irq(&pwq->pool->lock); 4253 4254 if (drained) 4255 continue; 4256 4257 if (++flush_cnt == 10 || 4258 (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 4259 pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 4260 wq->name, __func__, flush_cnt); 4261 4262 mutex_unlock(&wq->mutex); 4263 goto reflush; 4264 } 4265 4266 if (!--wq->nr_drainers) 4267 wq->flags &= ~__WQ_DRAINING; 4268 mutex_unlock(&wq->mutex); 4269 } 4270 EXPORT_SYMBOL_GPL(drain_workqueue); 4271 4272 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 4273 bool from_cancel) 4274 { 4275 struct worker *worker = NULL; 4276 struct worker_pool *pool; 4277 struct pool_workqueue *pwq; 4278 struct workqueue_struct *wq; 4279 4280 rcu_read_lock(); 4281 pool = get_work_pool(work); 4282 if (!pool) { 4283 rcu_read_unlock(); 4284 return false; 4285 } 4286 4287 raw_spin_lock_irq(&pool->lock); 4288 /* see the comment in try_to_grab_pending() with the same code */ 4289 pwq = get_work_pwq(work); 4290 if (pwq) { 4291 if (unlikely(pwq->pool != pool)) 4292 goto already_gone; 4293 } else { 4294 worker = find_worker_executing_work(pool, work); 4295 if (!worker) 4296 goto already_gone; 4297 pwq = worker->current_pwq; 4298 } 4299 4300 wq = pwq->wq; 4301 check_flush_dependency(wq, work, from_cancel); 4302 4303 insert_wq_barrier(pwq, barr, work, worker); 4304 raw_spin_unlock_irq(&pool->lock); 4305 4306 touch_work_lockdep_map(work, wq); 4307 4308 /* 4309 * Force a lock recursion deadlock when using flush_work() inside a 4310 * single-threaded or rescuer equipped workqueue. 4311 * 4312 * For single threaded workqueues the deadlock happens when the work 4313 * is after the work issuing the flush_work(). For rescuer equipped 4314 * workqueues the deadlock happens when the rescuer stalls, blocking 4315 * forward progress. 4316 */ 4317 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer)) 4318 touch_wq_lockdep_map(wq); 4319 4320 rcu_read_unlock(); 4321 return true; 4322 already_gone: 4323 raw_spin_unlock_irq(&pool->lock); 4324 rcu_read_unlock(); 4325 return false; 4326 } 4327 4328 static bool __flush_work(struct work_struct *work, bool from_cancel) 4329 { 4330 struct wq_barrier barr; 4331 4332 if (WARN_ON(!wq_online)) 4333 return false; 4334 4335 if (WARN_ON(!work->func)) 4336 return false; 4337 4338 if (!start_flush_work(work, &barr, from_cancel)) 4339 return false; 4340 4341 /* 4342 * start_flush_work() returned %true. If @from_cancel is set, we know 4343 * that @work must have been executing during start_flush_work() and 4344 * can't currently be queued. Its data must contain OFFQ bits. If @work 4345 * was queued on a BH workqueue, we also know that it was running in the 4346 * BH context and thus can be busy-waited. 4347 */ 4348 if (from_cancel) { 4349 unsigned long data = *work_data_bits(work); 4350 4351 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) && 4352 (data & WORK_OFFQ_BH)) { 4353 /* 4354 * On RT, prevent a live lock when %current preempted 4355 * soft interrupt processing by blocking on lock which 4356 * is owned by the thread invoking the callback. 4357 */ 4358 while (!try_wait_for_completion(&barr.done)) { 4359 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 4360 struct worker_pool *pool; 4361 4362 guard(rcu)(); 4363 pool = get_work_pool(work); 4364 if (pool) 4365 workqueue_callback_cancel_wait_running(pool); 4366 } else { 4367 cpu_relax(); 4368 } 4369 } 4370 goto out_destroy; 4371 } 4372 } 4373 4374 wait_for_completion(&barr.done); 4375 4376 out_destroy: 4377 destroy_work_on_stack(&barr.work); 4378 return true; 4379 } 4380 4381 /** 4382 * flush_work - wait for a work to finish executing the last queueing instance 4383 * @work: the work to flush 4384 * 4385 * Wait until @work has finished execution. @work is guaranteed to be idle 4386 * on return if it hasn't been requeued since flush started. 4387 * 4388 * Return: 4389 * %true if flush_work() waited for the work to finish execution, 4390 * %false if it was already idle. 4391 */ 4392 bool flush_work(struct work_struct *work) 4393 { 4394 might_sleep(); 4395 return __flush_work(work, false); 4396 } 4397 EXPORT_SYMBOL_GPL(flush_work); 4398 4399 /** 4400 * flush_delayed_work - wait for a dwork to finish executing the last queueing 4401 * @dwork: the delayed work to flush 4402 * 4403 * Delayed timer is cancelled and the pending work is queued for 4404 * immediate execution. Like flush_work(), this function only 4405 * considers the last queueing instance of @dwork. 4406 * 4407 * Return: 4408 * %true if flush_work() waited for the work to finish execution, 4409 * %false if it was already idle. 4410 */ 4411 bool flush_delayed_work(struct delayed_work *dwork) 4412 { 4413 local_irq_disable(); 4414 if (timer_delete_sync(&dwork->timer)) 4415 __queue_work(dwork->cpu, dwork->wq, &dwork->work); 4416 local_irq_enable(); 4417 return flush_work(&dwork->work); 4418 } 4419 EXPORT_SYMBOL(flush_delayed_work); 4420 4421 /** 4422 * flush_rcu_work - wait for a rwork to finish executing the last queueing 4423 * @rwork: the rcu work to flush 4424 * 4425 * Return: 4426 * %true if flush_rcu_work() waited for the work to finish execution, 4427 * %false if it was already idle. 4428 */ 4429 bool flush_rcu_work(struct rcu_work *rwork) 4430 { 4431 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 4432 rcu_barrier(); 4433 flush_work(&rwork->work); 4434 return true; 4435 } else { 4436 return flush_work(&rwork->work); 4437 } 4438 } 4439 EXPORT_SYMBOL(flush_rcu_work); 4440 4441 static void work_offqd_disable(struct work_offq_data *offqd) 4442 { 4443 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1; 4444 4445 if (likely(offqd->disable < max)) 4446 offqd->disable++; 4447 else 4448 WARN_ONCE(true, "workqueue: work disable count overflowed\n"); 4449 } 4450 4451 static void work_offqd_enable(struct work_offq_data *offqd) 4452 { 4453 if (likely(offqd->disable > 0)) 4454 offqd->disable--; 4455 else 4456 WARN_ONCE(true, "workqueue: work disable count underflowed\n"); 4457 } 4458 4459 static bool __cancel_work(struct work_struct *work, u32 cflags) 4460 { 4461 struct work_offq_data offqd; 4462 unsigned long irq_flags; 4463 int ret; 4464 4465 ret = work_grab_pending(work, cflags, &irq_flags); 4466 4467 work_offqd_unpack(&offqd, *work_data_bits(work)); 4468 4469 if (cflags & WORK_CANCEL_DISABLE) 4470 work_offqd_disable(&offqd); 4471 4472 set_work_pool_and_clear_pending(work, offqd.pool_id, 4473 work_offqd_pack_flags(&offqd)); 4474 local_irq_restore(irq_flags); 4475 return ret; 4476 } 4477 4478 static bool __cancel_work_sync(struct work_struct *work, u32 cflags) 4479 { 4480 bool ret; 4481 4482 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE); 4483 4484 if (*work_data_bits(work) & WORK_OFFQ_BH) 4485 WARN_ON_ONCE(in_hardirq()); 4486 else 4487 might_sleep(); 4488 4489 /* 4490 * Skip __flush_work() during early boot when we know that @work isn't 4491 * executing. This allows canceling during early boot. 4492 */ 4493 if (wq_online) 4494 __flush_work(work, true); 4495 4496 if (!(cflags & WORK_CANCEL_DISABLE)) 4497 enable_work(work); 4498 4499 return ret; 4500 } 4501 4502 /* 4503 * See cancel_delayed_work() 4504 */ 4505 bool cancel_work(struct work_struct *work) 4506 { 4507 return __cancel_work(work, 0); 4508 } 4509 EXPORT_SYMBOL(cancel_work); 4510 4511 /** 4512 * cancel_work_sync - cancel a work and wait for it to finish 4513 * @work: the work to cancel 4514 * 4515 * Cancel @work and wait for its execution to finish. This function can be used 4516 * even if the work re-queues itself or migrates to another workqueue. On return 4517 * from this function, @work is guaranteed to be not pending or executing on any 4518 * CPU as long as there aren't racing enqueues. 4519 * 4520 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's. 4521 * Use cancel_delayed_work_sync() instead. 4522 * 4523 * Must be called from a sleepable context if @work was last queued on a non-BH 4524 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4525 * if @work was last queued on a BH workqueue. 4526 * 4527 * Returns %true if @work was pending, %false otherwise. 4528 */ 4529 bool cancel_work_sync(struct work_struct *work) 4530 { 4531 return __cancel_work_sync(work, 0); 4532 } 4533 EXPORT_SYMBOL_GPL(cancel_work_sync); 4534 4535 /** 4536 * cancel_delayed_work - cancel a delayed work 4537 * @dwork: delayed_work to cancel 4538 * 4539 * Kill off a pending delayed_work. 4540 * 4541 * Return: %true if @dwork was pending and canceled; %false if it wasn't 4542 * pending. 4543 * 4544 * Note: 4545 * The work callback function may still be running on return, unless 4546 * it returns %true and the work doesn't re-arm itself. Explicitly flush or 4547 * use cancel_delayed_work_sync() to wait on it. 4548 * 4549 * This function is safe to call from any context including IRQ handler. 4550 */ 4551 bool cancel_delayed_work(struct delayed_work *dwork) 4552 { 4553 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED); 4554 } 4555 EXPORT_SYMBOL(cancel_delayed_work); 4556 4557 /** 4558 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 4559 * @dwork: the delayed work cancel 4560 * 4561 * This is cancel_work_sync() for delayed works. 4562 * 4563 * Return: 4564 * %true if @dwork was pending, %false otherwise. 4565 */ 4566 bool cancel_delayed_work_sync(struct delayed_work *dwork) 4567 { 4568 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED); 4569 } 4570 EXPORT_SYMBOL(cancel_delayed_work_sync); 4571 4572 /** 4573 * disable_work - Disable and cancel a work item 4574 * @work: work item to disable 4575 * 4576 * Disable @work by incrementing its disable count and cancel it if currently 4577 * pending. As long as the disable count is non-zero, any attempt to queue @work 4578 * will fail and return %false. The maximum supported disable depth is 2 to the 4579 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536. 4580 * 4581 * Can be called from any context. Returns %true if @work was pending, %false 4582 * otherwise. 4583 */ 4584 bool disable_work(struct work_struct *work) 4585 { 4586 return __cancel_work(work, WORK_CANCEL_DISABLE); 4587 } 4588 EXPORT_SYMBOL_GPL(disable_work); 4589 4590 /** 4591 * disable_work_sync - Disable, cancel and drain a work item 4592 * @work: work item to disable 4593 * 4594 * Similar to disable_work() but also wait for @work to finish if currently 4595 * executing. 4596 * 4597 * Must be called from a sleepable context if @work was last queued on a non-BH 4598 * workqueue. Can also be called from non-hardirq atomic contexts including BH 4599 * if @work was last queued on a BH workqueue. 4600 * 4601 * Returns %true if @work was pending, %false otherwise. 4602 */ 4603 bool disable_work_sync(struct work_struct *work) 4604 { 4605 return __cancel_work_sync(work, WORK_CANCEL_DISABLE); 4606 } 4607 EXPORT_SYMBOL_GPL(disable_work_sync); 4608 4609 /** 4610 * enable_work - Enable a work item 4611 * @work: work item to enable 4612 * 4613 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can 4614 * only be queued if its disable count is 0. 4615 * 4616 * Can be called from any context. Returns %true if the disable count reached 0. 4617 * Otherwise, %false. 4618 */ 4619 bool enable_work(struct work_struct *work) 4620 { 4621 struct work_offq_data offqd; 4622 unsigned long irq_flags; 4623 4624 work_grab_pending(work, 0, &irq_flags); 4625 4626 work_offqd_unpack(&offqd, *work_data_bits(work)); 4627 work_offqd_enable(&offqd); 4628 set_work_pool_and_clear_pending(work, offqd.pool_id, 4629 work_offqd_pack_flags(&offqd)); 4630 local_irq_restore(irq_flags); 4631 4632 return !offqd.disable; 4633 } 4634 EXPORT_SYMBOL_GPL(enable_work); 4635 4636 /** 4637 * disable_delayed_work - Disable and cancel a delayed work item 4638 * @dwork: delayed work item to disable 4639 * 4640 * disable_work() for delayed work items. 4641 */ 4642 bool disable_delayed_work(struct delayed_work *dwork) 4643 { 4644 return __cancel_work(&dwork->work, 4645 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4646 } 4647 EXPORT_SYMBOL_GPL(disable_delayed_work); 4648 4649 /** 4650 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item 4651 * @dwork: delayed work item to disable 4652 * 4653 * disable_work_sync() for delayed work items. 4654 */ 4655 bool disable_delayed_work_sync(struct delayed_work *dwork) 4656 { 4657 return __cancel_work_sync(&dwork->work, 4658 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE); 4659 } 4660 EXPORT_SYMBOL_GPL(disable_delayed_work_sync); 4661 4662 /** 4663 * enable_delayed_work - Enable a delayed work item 4664 * @dwork: delayed work item to enable 4665 * 4666 * enable_work() for delayed work items. 4667 */ 4668 bool enable_delayed_work(struct delayed_work *dwork) 4669 { 4670 return enable_work(&dwork->work); 4671 } 4672 EXPORT_SYMBOL_GPL(enable_delayed_work); 4673 4674 /** 4675 * schedule_on_each_cpu - execute a function synchronously on each online CPU 4676 * @func: the function to call 4677 * 4678 * schedule_on_each_cpu() executes @func on each online CPU using the 4679 * system workqueue and blocks until all CPUs have completed. 4680 * schedule_on_each_cpu() is very slow. 4681 * 4682 * Return: 4683 * 0 on success, -errno on failure. 4684 */ 4685 int schedule_on_each_cpu(work_func_t func) 4686 { 4687 int cpu; 4688 struct work_struct __percpu *works; 4689 4690 works = alloc_percpu(struct work_struct); 4691 if (!works) 4692 return -ENOMEM; 4693 4694 cpus_read_lock(); 4695 4696 for_each_online_cpu(cpu) { 4697 struct work_struct *work = per_cpu_ptr(works, cpu); 4698 4699 INIT_WORK(work, func); 4700 schedule_work_on(cpu, work); 4701 } 4702 4703 for_each_online_cpu(cpu) 4704 flush_work(per_cpu_ptr(works, cpu)); 4705 4706 cpus_read_unlock(); 4707 free_percpu(works); 4708 return 0; 4709 } 4710 4711 /** 4712 * execute_in_process_context - reliably execute the routine with user context 4713 * @fn: the function to execute 4714 * @ew: guaranteed storage for the execute work structure (must 4715 * be available when the work executes) 4716 * 4717 * Executes the function immediately if process context is available, 4718 * otherwise schedules the function for delayed execution. 4719 * 4720 * Return: 0 - function was executed 4721 * 1 - function was scheduled for execution 4722 */ 4723 int execute_in_process_context(work_func_t fn, struct execute_work *ew) 4724 { 4725 if (!in_interrupt()) { 4726 fn(&ew->work); 4727 return 0; 4728 } 4729 4730 INIT_WORK(&ew->work, fn); 4731 schedule_work(&ew->work); 4732 4733 return 1; 4734 } 4735 EXPORT_SYMBOL_GPL(execute_in_process_context); 4736 4737 /** 4738 * free_workqueue_attrs - free a workqueue_attrs 4739 * @attrs: workqueue_attrs to free 4740 * 4741 * Undo alloc_workqueue_attrs(). 4742 */ 4743 void free_workqueue_attrs(struct workqueue_attrs *attrs) 4744 { 4745 if (attrs) { 4746 free_cpumask_var(attrs->cpumask); 4747 free_cpumask_var(attrs->__pod_cpumask); 4748 kfree(attrs); 4749 } 4750 } 4751 4752 /** 4753 * alloc_workqueue_attrs - allocate a workqueue_attrs 4754 * 4755 * Allocate a new workqueue_attrs, initialize with default settings and 4756 * return it. 4757 * 4758 * Return: The allocated new workqueue_attr on success. %NULL on failure. 4759 */ 4760 struct workqueue_attrs *alloc_workqueue_attrs_noprof(void) 4761 { 4762 struct workqueue_attrs *attrs; 4763 4764 attrs = kzalloc_obj(*attrs); 4765 if (!attrs) 4766 goto fail; 4767 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 4768 goto fail; 4769 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 4770 goto fail; 4771 4772 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4773 attrs->affn_scope = WQ_AFFN_DFL; 4774 return attrs; 4775 fail: 4776 free_workqueue_attrs(attrs); 4777 return NULL; 4778 } 4779 4780 static void copy_workqueue_attrs(struct workqueue_attrs *to, 4781 const struct workqueue_attrs *from) 4782 { 4783 to->nice = from->nice; 4784 cpumask_copy(to->cpumask, from->cpumask); 4785 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 4786 to->affn_strict = from->affn_strict; 4787 4788 /* 4789 * Unlike hash and equality test, copying shouldn't ignore wq-only 4790 * fields as copying is used for both pool and wq attrs. Instead, 4791 * get_unbound_pool() explicitly clears the fields. 4792 */ 4793 to->affn_scope = from->affn_scope; 4794 to->ordered = from->ordered; 4795 } 4796 4797 /* 4798 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 4799 * comments in 'struct workqueue_attrs' definition. 4800 */ 4801 static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 4802 { 4803 attrs->affn_scope = WQ_AFFN_NR_TYPES; 4804 attrs->ordered = false; 4805 if (attrs->affn_strict) 4806 cpumask_copy(attrs->cpumask, cpu_possible_mask); 4807 } 4808 4809 /* hash value of the content of @attr */ 4810 static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 4811 { 4812 u32 hash = 0; 4813 4814 hash = jhash_1word(attrs->nice, hash); 4815 hash = jhash_1word(attrs->affn_strict, hash); 4816 hash = jhash(cpumask_bits(attrs->__pod_cpumask), 4817 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4818 if (!attrs->affn_strict) 4819 hash = jhash(cpumask_bits(attrs->cpumask), 4820 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 4821 return hash; 4822 } 4823 4824 /* content equality test */ 4825 static bool wqattrs_equal(const struct workqueue_attrs *a, 4826 const struct workqueue_attrs *b) 4827 { 4828 if (a->nice != b->nice) 4829 return false; 4830 if (a->affn_strict != b->affn_strict) 4831 return false; 4832 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 4833 return false; 4834 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask)) 4835 return false; 4836 return true; 4837 } 4838 4839 /* Update @attrs with actually available CPUs */ 4840 static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 4841 const cpumask_t *unbound_cpumask) 4842 { 4843 /* 4844 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 4845 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 4846 * @unbound_cpumask. 4847 */ 4848 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 4849 if (unlikely(cpumask_empty(attrs->cpumask))) 4850 cpumask_copy(attrs->cpumask, unbound_cpumask); 4851 } 4852 4853 /* find wq_pod_type to use for @attrs */ 4854 static const struct wq_pod_type * 4855 wqattrs_pod_type(const struct workqueue_attrs *attrs) 4856 { 4857 enum wq_affn_scope scope; 4858 struct wq_pod_type *pt; 4859 4860 /* to synchronize access to wq_affn_dfl */ 4861 lockdep_assert_held(&wq_pool_mutex); 4862 4863 if (attrs->affn_scope == WQ_AFFN_DFL) 4864 scope = wq_affn_dfl; 4865 else 4866 scope = attrs->affn_scope; 4867 4868 pt = &wq_pod_types[scope]; 4869 4870 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 4871 likely(pt->nr_pods)) 4872 return pt; 4873 4874 /* 4875 * Before workqueue_init_topology(), only SYSTEM is available which is 4876 * initialized in workqueue_init_early(). 4877 */ 4878 pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 4879 BUG_ON(!pt->nr_pods); 4880 return pt; 4881 } 4882 4883 /** 4884 * init_worker_pool - initialize a newly zalloc'd worker_pool 4885 * @pool: worker_pool to initialize 4886 * 4887 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 4888 * 4889 * Return: 0 on success, -errno on failure. Even on failure, all fields 4890 * inside @pool proper are initialized and put_unbound_pool() can be called 4891 * on @pool safely to release it. 4892 */ 4893 static int init_worker_pool(struct worker_pool *pool) 4894 { 4895 raw_spin_lock_init(&pool->lock); 4896 pool->id = -1; 4897 pool->cpu = -1; 4898 pool->node = NUMA_NO_NODE; 4899 pool->flags |= POOL_DISASSOCIATED; 4900 pool->last_progress_ts = jiffies; 4901 INIT_LIST_HEAD(&pool->worklist); 4902 INIT_LIST_HEAD(&pool->idle_list); 4903 hash_init(pool->busy_hash); 4904 4905 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 4906 INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 4907 4908 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 4909 4910 INIT_LIST_HEAD(&pool->workers); 4911 4912 ida_init(&pool->worker_ida); 4913 INIT_HLIST_NODE(&pool->hash_node); 4914 pool->refcnt = 1; 4915 #ifdef CONFIG_PREEMPT_RT 4916 spin_lock_init(&pool->cb_lock); 4917 #endif 4918 4919 /* shouldn't fail above this point */ 4920 pool->attrs = alloc_workqueue_attrs(); 4921 if (!pool->attrs) 4922 return -ENOMEM; 4923 4924 wqattrs_clear_for_pool(pool->attrs); 4925 4926 return 0; 4927 } 4928 4929 #ifdef CONFIG_LOCKDEP 4930 static void wq_init_lockdep(struct workqueue_struct *wq) 4931 { 4932 char *lock_name; 4933 4934 lockdep_register_key(&wq->key); 4935 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 4936 if (!lock_name) 4937 lock_name = wq->name; 4938 4939 wq->lock_name = lock_name; 4940 wq->lockdep_map = &wq->__lockdep_map; 4941 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0); 4942 } 4943 4944 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4945 { 4946 if (wq->lockdep_map != &wq->__lockdep_map) 4947 return; 4948 4949 lockdep_unregister_key(&wq->key); 4950 } 4951 4952 static void wq_free_lockdep(struct workqueue_struct *wq) 4953 { 4954 if (wq->lockdep_map != &wq->__lockdep_map) 4955 return; 4956 4957 if (wq->lock_name != wq->name) 4958 kfree(wq->lock_name); 4959 } 4960 #else 4961 static void wq_init_lockdep(struct workqueue_struct *wq) 4962 { 4963 } 4964 4965 static void wq_unregister_lockdep(struct workqueue_struct *wq) 4966 { 4967 } 4968 4969 static void wq_free_lockdep(struct workqueue_struct *wq) 4970 { 4971 } 4972 #endif 4973 4974 static void free_node_nr_active(struct wq_node_nr_active **nna_ar) 4975 { 4976 int node; 4977 4978 for_each_node(node) { 4979 kfree(nna_ar[node]); 4980 nna_ar[node] = NULL; 4981 } 4982 4983 kfree(nna_ar[nr_node_ids]); 4984 nna_ar[nr_node_ids] = NULL; 4985 } 4986 4987 static void init_node_nr_active(struct wq_node_nr_active *nna) 4988 { 4989 nna->max = WQ_DFL_MIN_ACTIVE; 4990 atomic_set(&nna->nr, 0); 4991 raw_spin_lock_init(&nna->lock); 4992 INIT_LIST_HEAD(&nna->pending_pwqs); 4993 } 4994 4995 /* 4996 * Each node's nr_active counter will be accessed mostly from its own node and 4997 * should be allocated in the node. 4998 */ 4999 static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar) 5000 { 5001 struct wq_node_nr_active *nna; 5002 int node; 5003 5004 for_each_node(node) { 5005 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node); 5006 if (!nna) 5007 goto err_free; 5008 init_node_nr_active(nna); 5009 nna_ar[node] = nna; 5010 } 5011 5012 /* [nr_node_ids] is used as the fallback */ 5013 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE); 5014 if (!nna) 5015 goto err_free; 5016 init_node_nr_active(nna); 5017 nna_ar[nr_node_ids] = nna; 5018 5019 return 0; 5020 5021 err_free: 5022 free_node_nr_active(nna_ar); 5023 return -ENOMEM; 5024 } 5025 5026 static void rcu_free_wq(struct rcu_head *rcu) 5027 { 5028 struct workqueue_struct *wq = 5029 container_of(rcu, struct workqueue_struct, rcu); 5030 5031 if (wq->flags & WQ_UNBOUND) 5032 free_node_nr_active(wq->node_nr_active); 5033 5034 wq_free_lockdep(wq); 5035 free_percpu(wq->cpu_pwq); 5036 free_workqueue_attrs(wq->unbound_attrs); 5037 kfree(wq); 5038 } 5039 5040 static void rcu_free_pool(struct rcu_head *rcu) 5041 { 5042 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 5043 5044 ida_destroy(&pool->worker_ida); 5045 free_workqueue_attrs(pool->attrs); 5046 kfree(pool); 5047 } 5048 5049 /** 5050 * put_unbound_pool - put a worker_pool 5051 * @pool: worker_pool to put 5052 * 5053 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 5054 * safe manner. get_unbound_pool() calls this function on its failure path 5055 * and this function should be able to release pools which went through, 5056 * successfully or not, init_worker_pool(). 5057 * 5058 * Should be called with wq_pool_mutex held. 5059 */ 5060 static void put_unbound_pool(struct worker_pool *pool) 5061 { 5062 struct worker *worker; 5063 LIST_HEAD(cull_list); 5064 5065 lockdep_assert_held(&wq_pool_mutex); 5066 5067 if (--pool->refcnt) 5068 return; 5069 5070 /* sanity checks */ 5071 if (WARN_ON(!(pool->cpu < 0)) || 5072 WARN_ON(!list_empty(&pool->worklist))) 5073 return; 5074 5075 /* release id and unhash */ 5076 if (pool->id >= 0) 5077 idr_remove(&worker_pool_idr, pool->id); 5078 hash_del(&pool->hash_node); 5079 5080 /* 5081 * Become the manager and destroy all workers. This prevents 5082 * @pool's workers from blocking on attach_mutex. We're the last 5083 * manager and @pool gets freed with the flag set. 5084 * 5085 * Having a concurrent manager is quite unlikely to happen as we can 5086 * only get here with 5087 * pwq->refcnt == pool->refcnt == 0 5088 * which implies no work queued to the pool, which implies no worker can 5089 * become the manager. However a worker could have taken the role of 5090 * manager before the refcnts dropped to 0, since maybe_create_worker() 5091 * drops pool->lock 5092 */ 5093 while (true) { 5094 rcuwait_wait_event(&manager_wait, 5095 !(pool->flags & POOL_MANAGER_ACTIVE), 5096 TASK_UNINTERRUPTIBLE); 5097 5098 mutex_lock(&wq_pool_attach_mutex); 5099 raw_spin_lock_irq(&pool->lock); 5100 if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 5101 pool->flags |= POOL_MANAGER_ACTIVE; 5102 break; 5103 } 5104 raw_spin_unlock_irq(&pool->lock); 5105 mutex_unlock(&wq_pool_attach_mutex); 5106 } 5107 5108 while ((worker = first_idle_worker(pool))) 5109 set_worker_dying(worker, &cull_list); 5110 WARN_ON(pool->nr_workers || pool->nr_idle); 5111 raw_spin_unlock_irq(&pool->lock); 5112 5113 detach_dying_workers(&cull_list); 5114 5115 mutex_unlock(&wq_pool_attach_mutex); 5116 5117 reap_dying_workers(&cull_list); 5118 5119 /* shut down the timers */ 5120 timer_delete_sync(&pool->idle_timer); 5121 cancel_work_sync(&pool->idle_cull_work); 5122 timer_delete_sync(&pool->mayday_timer); 5123 5124 /* RCU protected to allow dereferences from get_work_pool() */ 5125 call_rcu(&pool->rcu, rcu_free_pool); 5126 } 5127 5128 /** 5129 * get_unbound_pool - get a worker_pool with the specified attributes 5130 * @attrs: the attributes of the worker_pool to get 5131 * 5132 * Obtain a worker_pool which has the same attributes as @attrs, bump the 5133 * reference count and return it. If there already is a matching 5134 * worker_pool, it will be used; otherwise, this function attempts to 5135 * create a new one. 5136 * 5137 * Should be called with wq_pool_mutex held. 5138 * 5139 * Return: On success, a worker_pool with the same attributes as @attrs. 5140 * On failure, %NULL. 5141 */ 5142 static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 5143 { 5144 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 5145 u32 hash = wqattrs_hash(attrs); 5146 struct worker_pool *pool; 5147 int pod, node = NUMA_NO_NODE; 5148 5149 lockdep_assert_held(&wq_pool_mutex); 5150 5151 /* do we already have a matching pool? */ 5152 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 5153 if (wqattrs_equal(pool->attrs, attrs)) { 5154 pool->refcnt++; 5155 return pool; 5156 } 5157 } 5158 5159 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 5160 for (pod = 0; pod < pt->nr_pods; pod++) { 5161 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 5162 node = pt->pod_node[pod]; 5163 break; 5164 } 5165 } 5166 5167 /* nope, create a new one */ 5168 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 5169 if (!pool || init_worker_pool(pool) < 0) 5170 goto fail; 5171 5172 pool->node = node; 5173 copy_workqueue_attrs(pool->attrs, attrs); 5174 wqattrs_clear_for_pool(pool->attrs); 5175 5176 if (worker_pool_assign_id(pool) < 0) 5177 goto fail; 5178 5179 /* create and start the initial worker */ 5180 if (wq_online && !create_worker(pool)) 5181 goto fail; 5182 5183 /* install */ 5184 hash_add(unbound_pool_hash, &pool->hash_node, hash); 5185 5186 return pool; 5187 fail: 5188 if (pool) 5189 put_unbound_pool(pool); 5190 return NULL; 5191 } 5192 5193 /* 5194 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 5195 * refcnt and needs to be destroyed. 5196 */ 5197 static void pwq_release_workfn(struct kthread_work *work) 5198 { 5199 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 5200 release_work); 5201 struct workqueue_struct *wq = pwq->wq; 5202 struct worker_pool *pool = pwq->pool; 5203 bool is_last = false; 5204 5205 /* 5206 * When @pwq is not linked, it doesn't hold any reference to the 5207 * @wq, and @wq is invalid to access. 5208 */ 5209 if (!list_empty(&pwq->pwqs_node)) { 5210 mutex_lock(&wq->mutex); 5211 list_del_rcu(&pwq->pwqs_node); 5212 is_last = list_empty(&wq->pwqs); 5213 5214 /* 5215 * For ordered workqueue with a plugged dfl_pwq, restart it now. 5216 */ 5217 if (!is_last && (wq->flags & __WQ_ORDERED)) 5218 unplug_oldest_pwq(wq); 5219 5220 mutex_unlock(&wq->mutex); 5221 } 5222 5223 if (wq->flags & WQ_UNBOUND) { 5224 mutex_lock(&wq_pool_mutex); 5225 put_unbound_pool(pool); 5226 mutex_unlock(&wq_pool_mutex); 5227 } 5228 5229 if (!list_empty(&pwq->pending_node)) { 5230 struct wq_node_nr_active *nna = 5231 wq_node_nr_active(pwq->wq, pwq->pool->node); 5232 5233 raw_spin_lock_irq(&nna->lock); 5234 list_del_init(&pwq->pending_node); 5235 raw_spin_unlock_irq(&nna->lock); 5236 } 5237 5238 kfree_rcu(pwq, rcu); 5239 5240 /* 5241 * If we're the last pwq going away, @wq is already dead and no one 5242 * is gonna access it anymore. Schedule RCU free. 5243 */ 5244 if (is_last) { 5245 wq_unregister_lockdep(wq); 5246 call_rcu(&wq->rcu, rcu_free_wq); 5247 } 5248 } 5249 5250 /* initialize newly allocated @pwq which is associated with @wq and @pool */ 5251 static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 5252 struct worker_pool *pool) 5253 { 5254 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK); 5255 5256 memset(pwq, 0, sizeof(*pwq)); 5257 5258 pwq->pool = pool; 5259 pwq->wq = wq; 5260 pwq->flush_color = -1; 5261 pwq->refcnt = 1; 5262 INIT_LIST_HEAD(&pwq->inactive_works); 5263 INIT_LIST_HEAD(&pwq->pending_node); 5264 INIT_LIST_HEAD(&pwq->pwqs_node); 5265 INIT_LIST_HEAD(&pwq->mayday_node); 5266 kthread_init_work(&pwq->release_work, pwq_release_workfn); 5267 5268 /* 5269 * Set the dummy cursor work with valid function and get_work_pwq(). 5270 * 5271 * The cursor work should only be in the pwq->pool->worklist, and 5272 * should not be treated as a processable work item. 5273 * 5274 * WORK_STRUCT_PENDING and WORK_STRUCT_INACTIVE just make it less 5275 * surprise for kernel debugging tools and reviewers. 5276 */ 5277 INIT_WORK(&pwq->mayday_cursor, mayday_cursor_func); 5278 atomic_long_set(&pwq->mayday_cursor.data, (unsigned long)pwq | 5279 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | WORK_STRUCT_INACTIVE); 5280 } 5281 5282 /* sync @pwq with the current state of its associated wq and link it */ 5283 static void link_pwq(struct pool_workqueue *pwq) 5284 { 5285 struct workqueue_struct *wq = pwq->wq; 5286 5287 lockdep_assert_held(&wq->mutex); 5288 5289 /* may be called multiple times, ignore if already linked */ 5290 if (!list_empty(&pwq->pwqs_node)) 5291 return; 5292 5293 /* set the matching work_color */ 5294 pwq->work_color = wq->work_color; 5295 5296 /* link in @pwq */ 5297 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs); 5298 } 5299 5300 /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 5301 static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 5302 const struct workqueue_attrs *attrs) 5303 { 5304 struct worker_pool *pool; 5305 struct pool_workqueue *pwq; 5306 5307 lockdep_assert_held(&wq_pool_mutex); 5308 5309 pool = get_unbound_pool(attrs); 5310 if (!pool) 5311 return NULL; 5312 5313 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 5314 if (!pwq) { 5315 put_unbound_pool(pool); 5316 return NULL; 5317 } 5318 5319 init_pwq(pwq, wq, pool); 5320 return pwq; 5321 } 5322 5323 /** 5324 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 5325 * @attrs: the wq_attrs of the default pwq of the target workqueue 5326 * @cpu: the target CPU 5327 * 5328 * Calculate the cpumask a workqueue with @attrs should use on @pod. 5329 * The result is stored in @attrs->__pod_cpumask. 5330 * 5331 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 5332 * and @pod has online CPUs requested by @attrs, the returned cpumask is the 5333 * intersection of the possible CPUs of @pod and @attrs->cpumask. 5334 * 5335 * The caller is responsible for ensuring that the cpumask of @pod stays stable. 5336 */ 5337 static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu) 5338 { 5339 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 5340 int pod = pt->cpu_pod[cpu]; 5341 5342 /* calculate possible CPUs in @pod that @attrs wants */ 5343 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 5344 /* does @pod have any online CPUs @attrs wants? */ 5345 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) { 5346 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 5347 return; 5348 } 5349 } 5350 5351 /* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */ 5352 static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 5353 int cpu, struct pool_workqueue *pwq) 5354 { 5355 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu); 5356 struct pool_workqueue *old_pwq; 5357 5358 lockdep_assert_held(&wq_pool_mutex); 5359 lockdep_assert_held(&wq->mutex); 5360 5361 /* link_pwq() can handle duplicate calls */ 5362 link_pwq(pwq); 5363 5364 old_pwq = rcu_access_pointer(*slot); 5365 rcu_assign_pointer(*slot, pwq); 5366 return old_pwq; 5367 } 5368 5369 /* context to store the prepared attrs & pwqs before applying */ 5370 struct apply_wqattrs_ctx { 5371 struct workqueue_struct *wq; /* target workqueue */ 5372 struct workqueue_attrs *attrs; /* attrs to apply */ 5373 struct list_head list; /* queued for batching commit */ 5374 struct pool_workqueue *dfl_pwq; 5375 struct pool_workqueue *pwq_tbl[]; 5376 }; 5377 5378 /* free the resources after success or abort */ 5379 static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 5380 { 5381 if (ctx) { 5382 int cpu; 5383 5384 for_each_possible_cpu(cpu) 5385 put_pwq_unlocked(ctx->pwq_tbl[cpu]); 5386 put_pwq_unlocked(ctx->dfl_pwq); 5387 5388 free_workqueue_attrs(ctx->attrs); 5389 5390 kfree(ctx); 5391 } 5392 } 5393 5394 /* allocate the attrs and pwqs for later installation */ 5395 static struct apply_wqattrs_ctx * 5396 apply_wqattrs_prepare(struct workqueue_struct *wq, 5397 const struct workqueue_attrs *attrs, 5398 const cpumask_var_t unbound_cpumask) 5399 { 5400 struct apply_wqattrs_ctx *ctx; 5401 struct workqueue_attrs *new_attrs; 5402 int cpu; 5403 5404 lockdep_assert_held(&wq_pool_mutex); 5405 5406 if (WARN_ON(attrs->affn_scope < 0 || 5407 attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 5408 return ERR_PTR(-EINVAL); 5409 5410 ctx = kzalloc_flex(*ctx, pwq_tbl, nr_cpu_ids); 5411 5412 new_attrs = alloc_workqueue_attrs(); 5413 if (!ctx || !new_attrs) 5414 goto out_free; 5415 5416 /* 5417 * If something goes wrong during CPU up/down, we'll fall back to 5418 * the default pwq covering whole @attrs->cpumask. Always create 5419 * it even if we don't use it immediately. 5420 */ 5421 copy_workqueue_attrs(new_attrs, attrs); 5422 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 5423 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5424 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 5425 if (!ctx->dfl_pwq) 5426 goto out_free; 5427 5428 for_each_possible_cpu(cpu) { 5429 if (new_attrs->ordered) { 5430 ctx->dfl_pwq->refcnt++; 5431 ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 5432 } else { 5433 wq_calc_pod_cpumask(new_attrs, cpu); 5434 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 5435 if (!ctx->pwq_tbl[cpu]) 5436 goto out_free; 5437 } 5438 } 5439 5440 /* save the user configured attrs and sanitize it. */ 5441 copy_workqueue_attrs(new_attrs, attrs); 5442 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 5443 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 5444 ctx->attrs = new_attrs; 5445 5446 /* 5447 * For initialized ordered workqueues, there should only be one pwq 5448 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution 5449 * of newly queued work items until execution of older work items in 5450 * the old pwq's have completed. 5451 */ 5452 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)) 5453 ctx->dfl_pwq->plugged = true; 5454 5455 ctx->wq = wq; 5456 return ctx; 5457 5458 out_free: 5459 free_workqueue_attrs(new_attrs); 5460 apply_wqattrs_cleanup(ctx); 5461 return ERR_PTR(-ENOMEM); 5462 } 5463 5464 /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 5465 static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 5466 { 5467 int cpu; 5468 5469 /* all pwqs have been created successfully, let's install'em */ 5470 mutex_lock(&ctx->wq->mutex); 5471 5472 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 5473 5474 /* save the previous pwqs and install the new ones */ 5475 for_each_possible_cpu(cpu) 5476 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 5477 ctx->pwq_tbl[cpu]); 5478 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq); 5479 5480 /* update node_nr_active->max */ 5481 wq_update_node_max_active(ctx->wq, -1); 5482 5483 mutex_unlock(&ctx->wq->mutex); 5484 } 5485 5486 static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 5487 const struct workqueue_attrs *attrs) 5488 { 5489 struct apply_wqattrs_ctx *ctx; 5490 5491 /* only unbound workqueues can change attributes */ 5492 if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 5493 return -EINVAL; 5494 5495 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 5496 if (IS_ERR(ctx)) 5497 return PTR_ERR(ctx); 5498 5499 /* the ctx has been prepared successfully, let's commit it */ 5500 apply_wqattrs_commit(ctx); 5501 apply_wqattrs_cleanup(ctx); 5502 5503 return 0; 5504 } 5505 5506 /** 5507 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 5508 * @wq: the target workqueue 5509 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 5510 * 5511 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 5512 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 5513 * work items are affine to the pod it was issued on. Older pwqs are released as 5514 * in-flight work items finish. Note that a work item which repeatedly requeues 5515 * itself back-to-back will stay on its current pwq. 5516 * 5517 * Performs GFP_KERNEL allocations. 5518 * 5519 * Return: 0 on success and -errno on failure. 5520 */ 5521 int apply_workqueue_attrs(struct workqueue_struct *wq, 5522 const struct workqueue_attrs *attrs) 5523 { 5524 int ret; 5525 5526 mutex_lock(&wq_pool_mutex); 5527 ret = apply_workqueue_attrs_locked(wq, attrs); 5528 mutex_unlock(&wq_pool_mutex); 5529 5530 return ret; 5531 } 5532 5533 /** 5534 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug 5535 * @wq: the target workqueue 5536 * @cpu: the CPU to update the pwq slot for 5537 * 5538 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 5539 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged. 5540 * 5541 * 5542 * If pod affinity can't be adjusted due to memory allocation failure, it falls 5543 * back to @wq->dfl_pwq which may not be optimal but is always correct. 5544 * 5545 * Note that when the last allowed CPU of a pod goes offline for a workqueue 5546 * with a cpumask spanning multiple pods, the workers which were already 5547 * executing the work items for the workqueue will lose their CPU affinity and 5548 * may execute on any CPU. This is similar to how per-cpu workqueues behave on 5549 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 5550 * responsibility to flush the work item from CPU_DOWN_PREPARE. 5551 */ 5552 static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu) 5553 { 5554 struct pool_workqueue *old_pwq = NULL, *pwq; 5555 struct workqueue_attrs *target_attrs; 5556 5557 lockdep_assert_held(&wq_pool_mutex); 5558 5559 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 5560 return; 5561 5562 /* 5563 * We don't wanna alloc/free wq_attrs for each wq for each CPU. 5564 * Let's use a preallocated one. The following buf is protected by 5565 * CPU hotplug exclusion. 5566 */ 5567 target_attrs = unbound_wq_update_pwq_attrs_buf; 5568 5569 copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 5570 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 5571 5572 /* nothing to do if the target cpumask matches the current pwq */ 5573 wq_calc_pod_cpumask(target_attrs, cpu); 5574 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs)) 5575 return; 5576 5577 /* create a new pwq */ 5578 pwq = alloc_unbound_pwq(wq, target_attrs); 5579 if (!pwq) { 5580 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 5581 wq->name); 5582 goto use_dfl_pwq; 5583 } 5584 5585 /* Install the new pwq. */ 5586 mutex_lock(&wq->mutex); 5587 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5588 goto out_unlock; 5589 5590 use_dfl_pwq: 5591 mutex_lock(&wq->mutex); 5592 pwq = unbound_pwq(wq, -1); 5593 raw_spin_lock_irq(&pwq->pool->lock); 5594 get_pwq(pwq); 5595 raw_spin_unlock_irq(&pwq->pool->lock); 5596 old_pwq = install_unbound_pwq(wq, cpu, pwq); 5597 out_unlock: 5598 mutex_unlock(&wq->mutex); 5599 put_pwq_unlocked(old_pwq); 5600 } 5601 5602 static int alloc_and_link_pwqs(struct workqueue_struct *wq) 5603 { 5604 bool highpri = wq->flags & WQ_HIGHPRI; 5605 int cpu, ret; 5606 5607 lockdep_assert_held(&wq_pool_mutex); 5608 5609 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 5610 if (!wq->cpu_pwq) 5611 goto enomem; 5612 5613 if (!(wq->flags & WQ_UNBOUND)) { 5614 struct worker_pool __percpu *pools; 5615 5616 if (wq->flags & WQ_BH) 5617 pools = bh_worker_pools; 5618 else 5619 pools = cpu_worker_pools; 5620 5621 for_each_possible_cpu(cpu) { 5622 struct pool_workqueue **pwq_p; 5623 struct worker_pool *pool; 5624 5625 pool = &(per_cpu_ptr(pools, cpu)[highpri]); 5626 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu); 5627 5628 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 5629 pool->node); 5630 if (!*pwq_p) 5631 goto enomem; 5632 5633 init_pwq(*pwq_p, wq, pool); 5634 5635 mutex_lock(&wq->mutex); 5636 link_pwq(*pwq_p); 5637 mutex_unlock(&wq->mutex); 5638 } 5639 return 0; 5640 } 5641 5642 if (wq->flags & __WQ_ORDERED) { 5643 struct pool_workqueue *dfl_pwq; 5644 5645 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]); 5646 /* there should only be single pwq for ordering guarantee */ 5647 dfl_pwq = rcu_access_pointer(wq->dfl_pwq); 5648 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node || 5649 wq->pwqs.prev != &dfl_pwq->pwqs_node), 5650 "ordering guarantee broken for workqueue %s\n", wq->name); 5651 } else { 5652 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]); 5653 } 5654 5655 if (ret) 5656 goto enomem; 5657 return 0; 5658 5659 enomem: 5660 if (wq->cpu_pwq) { 5661 for_each_possible_cpu(cpu) { 5662 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 5663 5664 if (pwq) { 5665 /* 5666 * Unlink pwq from wq->pwqs since link_pwq() 5667 * may have already added it. wq->mutex is not 5668 * needed as the wq has not been published yet. 5669 */ 5670 if (!list_empty(&pwq->pwqs_node)) 5671 list_del_rcu(&pwq->pwqs_node); 5672 kmem_cache_free(pwq_cache, pwq); 5673 } 5674 } 5675 free_percpu(wq->cpu_pwq); 5676 wq->cpu_pwq = NULL; 5677 } 5678 return -ENOMEM; 5679 } 5680 5681 static int wq_clamp_max_active(int max_active, unsigned int flags, 5682 const char *name) 5683 { 5684 if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 5685 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 5686 max_active, name, 1, WQ_MAX_ACTIVE); 5687 5688 return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 5689 } 5690 5691 /* 5692 * Workqueues which may be used during memory reclaim should have a rescuer 5693 * to guarantee forward progress. 5694 */ 5695 static int init_rescuer(struct workqueue_struct *wq) 5696 { 5697 struct worker *rescuer; 5698 char id_buf[WORKER_ID_LEN]; 5699 int ret; 5700 5701 lockdep_assert_held(&wq_pool_mutex); 5702 5703 if (!(wq->flags & WQ_MEM_RECLAIM)) 5704 return 0; 5705 5706 rescuer = alloc_worker(NUMA_NO_NODE); 5707 if (!rescuer) { 5708 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 5709 wq->name); 5710 return -ENOMEM; 5711 } 5712 5713 rescuer->rescue_wq = wq; 5714 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL); 5715 5716 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf); 5717 if (IS_ERR(rescuer->task)) { 5718 ret = PTR_ERR(rescuer->task); 5719 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 5720 wq->name, ERR_PTR(ret)); 5721 kfree(rescuer); 5722 return ret; 5723 } 5724 5725 wq->rescuer = rescuer; 5726 5727 /* initial cpumask is consistent with the detached rescuer and unbind_worker() */ 5728 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 5729 kthread_bind_mask(rescuer->task, wq_unbound_cpumask); 5730 else 5731 kthread_bind_mask(rescuer->task, cpu_possible_mask); 5732 5733 wake_up_process(rescuer->task); 5734 5735 return 0; 5736 } 5737 5738 /** 5739 * wq_adjust_max_active - update a wq's max_active to the current setting 5740 * @wq: target workqueue 5741 * 5742 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and 5743 * activate inactive work items accordingly. If @wq is freezing, clear 5744 * @wq->max_active to zero. 5745 */ 5746 static void wq_adjust_max_active(struct workqueue_struct *wq) 5747 { 5748 bool activated; 5749 int new_max, new_min; 5750 5751 lockdep_assert_held(&wq->mutex); 5752 5753 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) { 5754 new_max = 0; 5755 new_min = 0; 5756 } else { 5757 new_max = wq->saved_max_active; 5758 new_min = wq->saved_min_active; 5759 } 5760 5761 if (wq->max_active == new_max && wq->min_active == new_min) 5762 return; 5763 5764 /* 5765 * Update @wq->max/min_active and then kick inactive work items if more 5766 * active work items are allowed. This doesn't break work item ordering 5767 * because new work items are always queued behind existing inactive 5768 * work items if there are any. 5769 */ 5770 WRITE_ONCE(wq->max_active, new_max); 5771 WRITE_ONCE(wq->min_active, new_min); 5772 5773 if (wq->flags & WQ_UNBOUND) 5774 wq_update_node_max_active(wq, -1); 5775 5776 if (new_max == 0) 5777 return; 5778 5779 /* 5780 * Round-robin through pwq's activating the first inactive work item 5781 * until max_active is filled. 5782 */ 5783 do { 5784 struct pool_workqueue *pwq; 5785 5786 activated = false; 5787 for_each_pwq(pwq, wq) { 5788 unsigned long irq_flags; 5789 5790 /* can be called during early boot w/ irq disabled */ 5791 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 5792 if (pwq_activate_first_inactive(pwq, true)) { 5793 activated = true; 5794 kick_pool(pwq->pool); 5795 } 5796 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 5797 } 5798 } while (activated); 5799 } 5800 5801 __printf(1, 0) 5802 static struct workqueue_struct *__alloc_workqueue(const char *fmt, 5803 unsigned int flags, 5804 int max_active, va_list args) 5805 { 5806 struct workqueue_struct *wq; 5807 size_t wq_size; 5808 int name_len; 5809 5810 if (flags & WQ_BH) { 5811 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS)) 5812 return NULL; 5813 if (WARN_ON_ONCE(max_active)) 5814 return NULL; 5815 } 5816 5817 /* see the comment above the definition of WQ_POWER_EFFICIENT */ 5818 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 5819 flags = (flags & ~WQ_PERCPU) | WQ_UNBOUND; 5820 5821 /* allocate wq and format name */ 5822 if (flags & WQ_UNBOUND) 5823 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1); 5824 else 5825 wq_size = sizeof(*wq); 5826 5827 wq = kzalloc_noprof(wq_size, GFP_KERNEL); 5828 if (!wq) 5829 return NULL; 5830 5831 if (flags & WQ_UNBOUND) { 5832 wq->unbound_attrs = alloc_workqueue_attrs_noprof(); 5833 if (!wq->unbound_attrs) 5834 goto err_free_wq; 5835 } 5836 5837 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args); 5838 5839 if (name_len >= WQ_NAME_LEN) 5840 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n", 5841 wq->name); 5842 5843 /* 5844 * One among WQ_PERCPU and WQ_UNBOUND must be set, but not both. 5845 * - If neither is set, default to WQ_PERCPU 5846 * - If both are set, default to WQ_UNBOUND 5847 * 5848 * This code can be removed after workqueue are unbound by default 5849 */ 5850 if (unlikely(!(flags & (WQ_UNBOUND | WQ_PERCPU)))) { 5851 WARN_ONCE(1, "workqueue: %s is using neither WQ_PERCPU or WQ_UNBOUND. " 5852 "Setting WQ_PERCPU.\n", wq->name); 5853 flags |= WQ_PERCPU; 5854 } else if (unlikely((flags & WQ_PERCPU) && (flags & WQ_UNBOUND))) { 5855 WARN_ONCE(1, "workqueue: %s uses both WQ_PERCPU and WQ_UNBOUND. " 5856 "Dropped WQ_PERCPU, keeping WQ_UNBOUND.\n", wq->name); 5857 flags &= ~WQ_PERCPU; 5858 } 5859 5860 if (flags & WQ_BH) { 5861 /* 5862 * BH workqueues always share a single execution context per CPU 5863 * and don't impose any max_active limit. 5864 */ 5865 max_active = INT_MAX; 5866 } else { 5867 max_active = max_active ?: WQ_DFL_ACTIVE; 5868 max_active = wq_clamp_max_active(max_active, flags, wq->name); 5869 } 5870 5871 /* init wq */ 5872 wq->flags = flags; 5873 wq->max_active = max_active; 5874 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE); 5875 wq->saved_max_active = wq->max_active; 5876 wq->saved_min_active = wq->min_active; 5877 mutex_init(&wq->mutex); 5878 atomic_set(&wq->nr_pwqs_to_flush, 0); 5879 INIT_LIST_HEAD(&wq->pwqs); 5880 INIT_LIST_HEAD(&wq->flusher_queue); 5881 INIT_LIST_HEAD(&wq->flusher_overflow); 5882 INIT_LIST_HEAD(&wq->maydays); 5883 5884 INIT_LIST_HEAD(&wq->list); 5885 5886 if (flags & WQ_UNBOUND) { 5887 if (alloc_node_nr_active(wq->node_nr_active) < 0) 5888 goto err_free_wq; 5889 } 5890 5891 /* 5892 * wq_pool_mutex protects the workqueues list, allocations of PWQs, 5893 * and the global freeze state. 5894 */ 5895 mutex_lock(&wq_pool_mutex); 5896 5897 if (alloc_and_link_pwqs(wq) < 0) 5898 goto err_unlock_free_node_nr_active; 5899 5900 mutex_lock(&wq->mutex); 5901 wq_adjust_max_active(wq); 5902 mutex_unlock(&wq->mutex); 5903 5904 list_add_tail_rcu(&wq->list, &workqueues); 5905 5906 if (wq_online && init_rescuer(wq) < 0) 5907 goto err_unlock_destroy; 5908 5909 mutex_unlock(&wq_pool_mutex); 5910 5911 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 5912 goto err_destroy; 5913 5914 return wq; 5915 5916 err_unlock_free_node_nr_active: 5917 mutex_unlock(&wq_pool_mutex); 5918 /* 5919 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work, 5920 * flushing the pwq_release_worker ensures that the pwq_release_workfn() 5921 * completes before calling kfree(wq). 5922 */ 5923 if (wq->flags & WQ_UNBOUND) { 5924 kthread_flush_worker(pwq_release_worker); 5925 free_node_nr_active(wq->node_nr_active); 5926 } 5927 err_free_wq: 5928 free_workqueue_attrs(wq->unbound_attrs); 5929 kfree(wq); 5930 return NULL; 5931 err_unlock_destroy: 5932 mutex_unlock(&wq_pool_mutex); 5933 err_destroy: 5934 destroy_workqueue(wq); 5935 return NULL; 5936 } 5937 5938 __printf(1, 0) 5939 static struct workqueue_struct *alloc_workqueue_va(const char *fmt, 5940 unsigned int flags, 5941 int max_active, 5942 va_list args) 5943 { 5944 struct workqueue_struct *wq; 5945 5946 wq = __alloc_workqueue(fmt, flags, max_active, args); 5947 if (wq) 5948 wq_init_lockdep(wq); 5949 5950 return wq; 5951 } 5952 5953 __printf(1, 4) 5954 struct workqueue_struct *alloc_workqueue_noprof(const char *fmt, 5955 unsigned int flags, 5956 int max_active, ...) 5957 { 5958 struct workqueue_struct *wq; 5959 va_list args; 5960 5961 va_start(args, max_active); 5962 wq = alloc_workqueue_va(fmt, flags, max_active, args); 5963 va_end(args); 5964 5965 return wq; 5966 } 5967 EXPORT_SYMBOL_GPL(alloc_workqueue_noprof); 5968 5969 static void devm_workqueue_release(void *res) 5970 { 5971 destroy_workqueue(res); 5972 } 5973 5974 __printf(2, 5) struct workqueue_struct * 5975 devm_alloc_workqueue_noprof(struct device *dev, const char *fmt, 5976 unsigned int flags, int max_active, ...) 5977 { 5978 struct workqueue_struct *wq; 5979 va_list args; 5980 int ret; 5981 5982 va_start(args, max_active); 5983 wq = alloc_workqueue_va(fmt, flags, max_active, args); 5984 va_end(args); 5985 if (!wq) 5986 return NULL; 5987 5988 ret = devm_add_action_or_reset(dev, devm_workqueue_release, wq); 5989 if (ret) 5990 return NULL; 5991 5992 return wq; 5993 } 5994 EXPORT_SYMBOL_GPL(devm_alloc_workqueue_noprof); 5995 5996 #ifdef CONFIG_LOCKDEP 5997 __printf(1, 5) 5998 struct workqueue_struct * 5999 alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags, 6000 int max_active, struct lockdep_map *lockdep_map, ...) 6001 { 6002 struct workqueue_struct *wq; 6003 va_list args; 6004 6005 va_start(args, lockdep_map); 6006 wq = __alloc_workqueue(fmt, flags, max_active, args); 6007 va_end(args); 6008 if (!wq) 6009 return NULL; 6010 6011 wq->lockdep_map = lockdep_map; 6012 6013 return wq; 6014 } 6015 EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map); 6016 #endif 6017 6018 static bool pwq_busy(struct pool_workqueue *pwq) 6019 { 6020 int i; 6021 6022 for (i = 0; i < WORK_NR_COLORS; i++) 6023 if (pwq->nr_in_flight[i]) 6024 return true; 6025 6026 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1)) 6027 return true; 6028 if (!pwq_is_empty(pwq)) 6029 return true; 6030 6031 return false; 6032 } 6033 6034 /** 6035 * destroy_workqueue - safely terminate a workqueue 6036 * @wq: target workqueue 6037 * 6038 * Safely destroy a workqueue. All work currently pending will be done first. 6039 * 6040 * This function does NOT guarantee that non-pending work that has been 6041 * submitted with queue_delayed_work() and similar functions will be done 6042 * before destroying the workqueue. The fundamental problem is that, currently, 6043 * the workqueue has no way of accessing non-pending delayed_work. delayed_work 6044 * is only linked on the timer-side. All delayed_work must, therefore, be 6045 * canceled before calling this function. 6046 * 6047 * TODO: It would be better if the problem described above wouldn't exist and 6048 * destroy_workqueue() would cleanly cancel all pending and non-pending 6049 * delayed_work. 6050 */ 6051 void destroy_workqueue(struct workqueue_struct *wq) 6052 { 6053 struct pool_workqueue *pwq; 6054 int cpu; 6055 6056 /* 6057 * Remove it from sysfs first so that sanity check failure doesn't 6058 * lead to sysfs name conflicts. 6059 */ 6060 workqueue_sysfs_unregister(wq); 6061 6062 /* mark the workqueue destruction is in progress */ 6063 mutex_lock(&wq->mutex); 6064 wq->flags |= __WQ_DESTROYING; 6065 mutex_unlock(&wq->mutex); 6066 6067 /* drain it before proceeding with destruction */ 6068 drain_workqueue(wq); 6069 6070 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 6071 if (wq->rescuer) { 6072 /* rescuer will empty maydays list before exiting */ 6073 kthread_stop(wq->rescuer->task); 6074 kfree(wq->rescuer); 6075 wq->rescuer = NULL; 6076 } 6077 6078 /* 6079 * Sanity checks - grab all the locks so that we wait for all 6080 * in-flight operations which may do put_pwq(). 6081 */ 6082 mutex_lock(&wq_pool_mutex); 6083 mutex_lock(&wq->mutex); 6084 for_each_pwq(pwq, wq) { 6085 raw_spin_lock_irq(&pwq->pool->lock); 6086 if (WARN_ON(pwq_busy(pwq))) { 6087 pr_warn("%s: %s has the following busy pwq\n", 6088 __func__, wq->name); 6089 show_pwq(pwq); 6090 raw_spin_unlock_irq(&pwq->pool->lock); 6091 mutex_unlock(&wq->mutex); 6092 mutex_unlock(&wq_pool_mutex); 6093 show_one_workqueue(wq); 6094 return; 6095 } 6096 raw_spin_unlock_irq(&pwq->pool->lock); 6097 } 6098 mutex_unlock(&wq->mutex); 6099 6100 /* 6101 * wq list is used to freeze wq, remove from list after 6102 * flushing is complete in case freeze races us. 6103 */ 6104 list_del_rcu(&wq->list); 6105 mutex_unlock(&wq_pool_mutex); 6106 6107 /* 6108 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 6109 * to put the base refs. @wq will be auto-destroyed from the last 6110 * pwq_put. RCU read lock prevents @wq from going away from under us. 6111 */ 6112 rcu_read_lock(); 6113 6114 for_each_possible_cpu(cpu) { 6115 put_pwq_unlocked(unbound_pwq(wq, cpu)); 6116 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL); 6117 } 6118 6119 put_pwq_unlocked(unbound_pwq(wq, -1)); 6120 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL); 6121 6122 rcu_read_unlock(); 6123 } 6124 EXPORT_SYMBOL_GPL(destroy_workqueue); 6125 6126 /** 6127 * workqueue_set_max_active - adjust max_active of a workqueue 6128 * @wq: target workqueue 6129 * @max_active: new max_active value. 6130 * 6131 * Set max_active of @wq to @max_active. See the alloc_workqueue() function 6132 * comment. 6133 * 6134 * CONTEXT: 6135 * Don't call from IRQ context. 6136 */ 6137 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 6138 { 6139 /* max_active doesn't mean anything for BH workqueues */ 6140 if (WARN_ON(wq->flags & WQ_BH)) 6141 return; 6142 /* disallow meddling with max_active for ordered workqueues */ 6143 if (WARN_ON(wq->flags & __WQ_ORDERED)) 6144 return; 6145 6146 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 6147 6148 mutex_lock(&wq->mutex); 6149 6150 wq->saved_max_active = max_active; 6151 if (wq->flags & WQ_UNBOUND) 6152 wq->saved_min_active = min(wq->saved_min_active, max_active); 6153 6154 wq_adjust_max_active(wq); 6155 6156 mutex_unlock(&wq->mutex); 6157 } 6158 EXPORT_SYMBOL_GPL(workqueue_set_max_active); 6159 6160 /** 6161 * workqueue_set_min_active - adjust min_active of an unbound workqueue 6162 * @wq: target unbound workqueue 6163 * @min_active: new min_active value 6164 * 6165 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an 6166 * unbound workqueue is not guaranteed to be able to process max_active 6167 * interdependent work items. Instead, an unbound workqueue is guaranteed to be 6168 * able to process min_active number of interdependent work items which is 6169 * %WQ_DFL_MIN_ACTIVE by default. 6170 * 6171 * Use this function to adjust the min_active value between 0 and the current 6172 * max_active. 6173 */ 6174 void workqueue_set_min_active(struct workqueue_struct *wq, int min_active) 6175 { 6176 /* min_active is only meaningful for non-ordered unbound workqueues */ 6177 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) != 6178 WQ_UNBOUND)) 6179 return; 6180 6181 mutex_lock(&wq->mutex); 6182 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active); 6183 wq_adjust_max_active(wq); 6184 mutex_unlock(&wq->mutex); 6185 } 6186 6187 /** 6188 * current_work - retrieve %current task's work struct 6189 * 6190 * Determine if %current task is a workqueue worker and what it's working on. 6191 * Useful to find out the context that the %current task is running in. 6192 * 6193 * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 6194 */ 6195 struct work_struct *current_work(void) 6196 { 6197 struct worker *worker = current_wq_worker(); 6198 6199 return worker ? worker->current_work : NULL; 6200 } 6201 EXPORT_SYMBOL(current_work); 6202 6203 /** 6204 * current_is_workqueue_rescuer - is %current workqueue rescuer? 6205 * 6206 * Determine whether %current is a workqueue rescuer. Can be used from 6207 * work functions to determine whether it's being run off the rescuer task. 6208 * 6209 * Return: %true if %current is a workqueue rescuer. %false otherwise. 6210 */ 6211 bool current_is_workqueue_rescuer(void) 6212 { 6213 struct worker *worker = current_wq_worker(); 6214 6215 return worker && worker->rescue_wq; 6216 } 6217 6218 /** 6219 * workqueue_congested - test whether a workqueue is congested 6220 * @cpu: CPU in question 6221 * @wq: target workqueue 6222 * 6223 * Test whether @wq's cpu workqueue for @cpu is congested. There is 6224 * no synchronization around this function and the test result is 6225 * unreliable and only useful as advisory hints or for debugging. 6226 * 6227 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 6228 * 6229 * With the exception of ordered workqueues, all workqueues have per-cpu 6230 * pool_workqueues, each with its own congested state. A workqueue being 6231 * congested on one CPU doesn't mean that the workqueue is contested on any 6232 * other CPUs. 6233 * 6234 * Return: 6235 * %true if congested, %false otherwise. 6236 */ 6237 bool workqueue_congested(int cpu, struct workqueue_struct *wq) 6238 { 6239 struct pool_workqueue *pwq; 6240 bool ret; 6241 6242 preempt_disable(); 6243 6244 if (cpu == WORK_CPU_UNBOUND) 6245 cpu = smp_processor_id(); 6246 6247 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 6248 ret = !list_empty(&pwq->inactive_works); 6249 6250 preempt_enable(); 6251 6252 return ret; 6253 } 6254 EXPORT_SYMBOL_GPL(workqueue_congested); 6255 6256 /** 6257 * work_busy - test whether a work is currently pending or running 6258 * @work: the work to be tested 6259 * 6260 * Test whether @work is currently pending or running. There is no 6261 * synchronization around this function and the test result is 6262 * unreliable and only useful as advisory hints or for debugging. 6263 * 6264 * Return: 6265 * OR'd bitmask of WORK_BUSY_* bits. 6266 */ 6267 unsigned int work_busy(struct work_struct *work) 6268 { 6269 struct worker_pool *pool; 6270 unsigned long irq_flags; 6271 unsigned int ret = 0; 6272 6273 if (work_pending(work)) 6274 ret |= WORK_BUSY_PENDING; 6275 6276 rcu_read_lock(); 6277 pool = get_work_pool(work); 6278 if (pool) { 6279 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6280 if (find_worker_executing_work(pool, work)) 6281 ret |= WORK_BUSY_RUNNING; 6282 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6283 } 6284 rcu_read_unlock(); 6285 6286 return ret; 6287 } 6288 EXPORT_SYMBOL_GPL(work_busy); 6289 6290 /** 6291 * set_worker_desc - set description for the current work item 6292 * @fmt: printf-style format string 6293 * @...: arguments for the format string 6294 * 6295 * This function can be called by a running work function to describe what 6296 * the work item is about. If the worker task gets dumped, this 6297 * information will be printed out together to help debugging. The 6298 * description can be at most WORKER_DESC_LEN including the trailing '\0'. 6299 */ 6300 void set_worker_desc(const char *fmt, ...) 6301 { 6302 struct worker *worker = current_wq_worker(); 6303 va_list args; 6304 6305 if (worker) { 6306 va_start(args, fmt); 6307 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 6308 va_end(args); 6309 } 6310 } 6311 EXPORT_SYMBOL_GPL(set_worker_desc); 6312 6313 /** 6314 * print_worker_info - print out worker information and description 6315 * @log_lvl: the log level to use when printing 6316 * @task: target task 6317 * 6318 * If @task is a worker and currently executing a work item, print out the 6319 * name of the workqueue being serviced and worker description set with 6320 * set_worker_desc() by the currently executing work item. 6321 * 6322 * This function can be safely called on any task as long as the 6323 * task_struct itself is accessible. While safe, this function isn't 6324 * synchronized and may print out mixups or garbages of limited length. 6325 */ 6326 void print_worker_info(const char *log_lvl, struct task_struct *task) 6327 { 6328 work_func_t fn = NULL; 6329 char name[WQ_NAME_LEN] = { }; 6330 char desc[WORKER_DESC_LEN] = { }; 6331 struct pool_workqueue *pwq = NULL; 6332 struct workqueue_struct *wq = NULL; 6333 struct worker *worker; 6334 6335 if (!(task->flags & PF_WQ_WORKER)) 6336 return; 6337 6338 /* 6339 * This function is called without any synchronization and @task 6340 * could be in any state. Be careful with dereferences. 6341 */ 6342 worker = kthread_probe_data(task); 6343 6344 /* 6345 * Carefully copy the associated workqueue's workfn, name and desc. 6346 * Keep the original last '\0' in case the original is garbage. 6347 */ 6348 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 6349 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 6350 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 6351 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 6352 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 6353 6354 if (fn || name[0] || desc[0]) { 6355 printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 6356 if (strcmp(name, desc)) 6357 pr_cont(" (%s)", desc); 6358 pr_cont("\n"); 6359 } 6360 } 6361 6362 static void pr_cont_pool_info(struct worker_pool *pool) 6363 { 6364 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 6365 if (pool->node != NUMA_NO_NODE) 6366 pr_cont(" node=%d", pool->node); 6367 pr_cont(" flags=0x%x", pool->flags); 6368 if (pool->flags & POOL_BH) 6369 pr_cont(" bh%s", 6370 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6371 else 6372 pr_cont(" nice=%d", pool->attrs->nice); 6373 } 6374 6375 static void pr_cont_worker_id(struct worker *worker) 6376 { 6377 struct worker_pool *pool = worker->pool; 6378 6379 if (pool->flags & POOL_BH) 6380 pr_cont("bh%s", 6381 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : ""); 6382 else 6383 pr_cont("%d%s", task_pid_nr(worker->task), 6384 worker->rescue_wq ? "(RESCUER)" : ""); 6385 } 6386 6387 struct pr_cont_work_struct { 6388 bool comma; 6389 work_func_t func; 6390 long ctr; 6391 }; 6392 6393 static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 6394 { 6395 if (!pcwsp->ctr) 6396 goto out_record; 6397 if (func == pcwsp->func) { 6398 pcwsp->ctr++; 6399 return; 6400 } 6401 if (pcwsp->ctr == 1) 6402 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 6403 else 6404 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 6405 pcwsp->ctr = 0; 6406 out_record: 6407 if ((long)func == -1L) 6408 return; 6409 pcwsp->comma = comma; 6410 pcwsp->func = func; 6411 pcwsp->ctr = 1; 6412 } 6413 6414 static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 6415 { 6416 if (work->func == wq_barrier_func) { 6417 struct wq_barrier *barr; 6418 6419 barr = container_of(work, struct wq_barrier, work); 6420 6421 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6422 pr_cont("%s BAR(%d)", comma ? "," : "", 6423 task_pid_nr(barr->task)); 6424 } else { 6425 if (!comma) 6426 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 6427 pr_cont_work_flush(comma, work->func, pcwsp); 6428 } 6429 } 6430 6431 static void show_pwq(struct pool_workqueue *pwq) 6432 { 6433 struct pr_cont_work_struct pcws = { .ctr = 0, }; 6434 struct worker_pool *pool = pwq->pool; 6435 struct work_struct *work; 6436 struct worker *worker; 6437 bool has_in_flight = false, has_pending = false; 6438 int bkt; 6439 6440 pr_info(" pwq %d:", pool->id); 6441 pr_cont_pool_info(pool); 6442 6443 pr_cont(" active=%d refcnt=%d%s\n", 6444 pwq->nr_active, pwq->refcnt, 6445 !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 6446 6447 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6448 if (worker->current_pwq == pwq) { 6449 has_in_flight = true; 6450 break; 6451 } 6452 } 6453 if (has_in_flight) { 6454 bool comma = false; 6455 6456 pr_info(" in-flight:"); 6457 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6458 if (worker->current_pwq != pwq) 6459 continue; 6460 6461 pr_cont(" %s", comma ? "," : ""); 6462 pr_cont_worker_id(worker); 6463 pr_cont(":%ps", worker->current_func); 6464 pr_cont(" for %us", 6465 jiffies_to_msecs(jiffies - worker->current_start) / 1000); 6466 list_for_each_entry(work, &worker->scheduled, entry) 6467 pr_cont_work(false, work, &pcws); 6468 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6469 comma = true; 6470 } 6471 pr_cont("\n"); 6472 } 6473 6474 list_for_each_entry(work, &pool->worklist, entry) { 6475 if (get_work_pwq(work) == pwq) { 6476 has_pending = true; 6477 break; 6478 } 6479 } 6480 if (has_pending) { 6481 bool comma = false; 6482 6483 pr_info(" pending:"); 6484 list_for_each_entry(work, &pool->worklist, entry) { 6485 if (get_work_pwq(work) != pwq) 6486 continue; 6487 6488 pr_cont_work(comma, work, &pcws); 6489 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6490 } 6491 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6492 pr_cont("\n"); 6493 } 6494 6495 if (!list_empty(&pwq->inactive_works)) { 6496 bool comma = false; 6497 6498 pr_info(" inactive:"); 6499 list_for_each_entry(work, &pwq->inactive_works, entry) { 6500 pr_cont_work(comma, work, &pcws); 6501 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 6502 } 6503 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 6504 pr_cont("\n"); 6505 } 6506 } 6507 6508 /** 6509 * show_one_workqueue - dump state of specified workqueue 6510 * @wq: workqueue whose state will be printed 6511 */ 6512 void show_one_workqueue(struct workqueue_struct *wq) 6513 { 6514 struct pool_workqueue *pwq; 6515 bool idle = true; 6516 unsigned long irq_flags; 6517 6518 for_each_pwq(pwq, wq) { 6519 if (!pwq_is_empty(pwq)) { 6520 idle = false; 6521 break; 6522 } 6523 } 6524 if (idle) /* Nothing to print for idle workqueue */ 6525 return; 6526 6527 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 6528 6529 for_each_pwq(pwq, wq) { 6530 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags); 6531 if (!pwq_is_empty(pwq)) { 6532 /* 6533 * Defer printing to avoid deadlocks in console 6534 * drivers that queue work while holding locks 6535 * also taken in their write paths. 6536 */ 6537 printk_deferred_enter(); 6538 show_pwq(pwq); 6539 printk_deferred_exit(); 6540 } 6541 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags); 6542 /* 6543 * We could be printing a lot from atomic context, e.g. 6544 * sysrq-t -> show_all_workqueues(). Avoid triggering 6545 * hard lockup. 6546 */ 6547 touch_nmi_watchdog(); 6548 } 6549 6550 } 6551 6552 /** 6553 * show_one_worker_pool - dump state of specified worker pool 6554 * @pool: worker pool whose state will be printed 6555 */ 6556 static void show_one_worker_pool(struct worker_pool *pool) 6557 { 6558 struct worker *worker; 6559 bool first = true; 6560 unsigned long irq_flags; 6561 unsigned long hung = 0; 6562 6563 raw_spin_lock_irqsave(&pool->lock, irq_flags); 6564 if (pool->nr_workers == pool->nr_idle) 6565 goto next_pool; 6566 6567 /* How long the first pending work is waiting for a worker. */ 6568 if (!list_empty(&pool->worklist)) 6569 hung = jiffies_to_msecs(jiffies - pool->last_progress_ts) / 1000; 6570 6571 /* 6572 * Defer printing to avoid deadlocks in console drivers that 6573 * queue work while holding locks also taken in their write 6574 * paths. 6575 */ 6576 printk_deferred_enter(); 6577 pr_info("pool %d:", pool->id); 6578 pr_cont_pool_info(pool); 6579 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 6580 if (pool->manager) 6581 pr_cont(" manager: %d", 6582 task_pid_nr(pool->manager->task)); 6583 list_for_each_entry(worker, &pool->idle_list, entry) { 6584 pr_cont(" %s", first ? "idle: " : ""); 6585 pr_cont_worker_id(worker); 6586 first = false; 6587 } 6588 pr_cont("\n"); 6589 printk_deferred_exit(); 6590 next_pool: 6591 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 6592 /* 6593 * We could be printing a lot from atomic context, e.g. 6594 * sysrq-t -> show_all_workqueues(). Avoid triggering 6595 * hard lockup. 6596 */ 6597 touch_nmi_watchdog(); 6598 6599 } 6600 6601 /** 6602 * show_all_workqueues - dump workqueue state 6603 * 6604 * Called from a sysrq handler and prints out all busy workqueues and pools. 6605 */ 6606 void show_all_workqueues(void) 6607 { 6608 struct workqueue_struct *wq; 6609 struct worker_pool *pool; 6610 int pi; 6611 6612 rcu_read_lock(); 6613 6614 pr_info("Showing busy workqueues and worker pools:\n"); 6615 6616 list_for_each_entry_rcu(wq, &workqueues, list) 6617 show_one_workqueue(wq); 6618 6619 for_each_pool(pool, pi) 6620 show_one_worker_pool(pool); 6621 6622 rcu_read_unlock(); 6623 } 6624 6625 /** 6626 * show_freezable_workqueues - dump freezable workqueue state 6627 * 6628 * Called from try_to_freeze_tasks() and prints out all freezable workqueues 6629 * still busy. 6630 */ 6631 void show_freezable_workqueues(void) 6632 { 6633 struct workqueue_struct *wq; 6634 6635 rcu_read_lock(); 6636 6637 pr_info("Showing freezable workqueues that are still busy:\n"); 6638 6639 list_for_each_entry_rcu(wq, &workqueues, list) { 6640 if (!(wq->flags & WQ_FREEZABLE)) 6641 continue; 6642 show_one_workqueue(wq); 6643 } 6644 6645 rcu_read_unlock(); 6646 } 6647 6648 /* used to show worker information through /proc/PID/{comm,stat,status} */ 6649 void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 6650 { 6651 /* stabilize PF_WQ_WORKER and worker pool association */ 6652 mutex_lock(&wq_pool_attach_mutex); 6653 6654 if (task->flags & PF_WQ_WORKER) { 6655 struct worker *worker = kthread_data(task); 6656 struct worker_pool *pool = worker->pool; 6657 int off; 6658 6659 off = format_worker_id(buf, size, worker, pool); 6660 6661 if (pool) { 6662 raw_spin_lock_irq(&pool->lock); 6663 /* 6664 * ->desc tracks information (wq name or 6665 * set_worker_desc()) for the latest execution. If 6666 * current, prepend '+', otherwise '-'. 6667 */ 6668 if (worker->desc[0] != '\0') { 6669 if (worker->current_work) 6670 scnprintf(buf + off, size - off, "+%s", 6671 worker->desc); 6672 else 6673 scnprintf(buf + off, size - off, "-%s", 6674 worker->desc); 6675 } 6676 raw_spin_unlock_irq(&pool->lock); 6677 } 6678 } else { 6679 strscpy(buf, task->comm, size); 6680 } 6681 6682 mutex_unlock(&wq_pool_attach_mutex); 6683 } 6684 6685 #ifdef CONFIG_SMP 6686 6687 /* 6688 * CPU hotplug. 6689 * 6690 * There are two challenges in supporting CPU hotplug. Firstly, there 6691 * are a lot of assumptions on strong associations among work, pwq and 6692 * pool which make migrating pending and scheduled works very 6693 * difficult to implement without impacting hot paths. Secondly, 6694 * worker pools serve mix of short, long and very long running works making 6695 * blocked draining impractical. 6696 * 6697 * This is solved by allowing the pools to be disassociated from the CPU 6698 * running as an unbound one and allowing it to be reattached later if the 6699 * cpu comes back online. 6700 */ 6701 6702 static void unbind_workers(int cpu) 6703 { 6704 struct worker_pool *pool; 6705 struct worker *worker; 6706 6707 for_each_cpu_worker_pool(pool, cpu) { 6708 mutex_lock(&wq_pool_attach_mutex); 6709 raw_spin_lock_irq(&pool->lock); 6710 6711 /* 6712 * We've blocked all attach/detach operations. Make all workers 6713 * unbound and set DISASSOCIATED. Before this, all workers 6714 * must be on the cpu. After this, they may become diasporas. 6715 * And the preemption disabled section in their sched callbacks 6716 * are guaranteed to see WORKER_UNBOUND since the code here 6717 * is on the same cpu. 6718 */ 6719 for_each_pool_worker(worker, pool) 6720 worker->flags |= WORKER_UNBOUND; 6721 6722 pool->flags |= POOL_DISASSOCIATED; 6723 6724 /* 6725 * The handling of nr_running in sched callbacks are disabled 6726 * now. Zap nr_running. After this, nr_running stays zero and 6727 * need_more_worker() and keep_working() are always true as 6728 * long as the worklist is not empty. This pool now behaves as 6729 * an unbound (in terms of concurrency management) pool which 6730 * are served by workers tied to the pool. 6731 */ 6732 pool->nr_running = 0; 6733 6734 /* 6735 * With concurrency management just turned off, a busy 6736 * worker blocking could lead to lengthy stalls. Kick off 6737 * unbound chain execution of currently pending work items. 6738 */ 6739 kick_pool(pool); 6740 6741 raw_spin_unlock_irq(&pool->lock); 6742 6743 for_each_pool_worker(worker, pool) 6744 unbind_worker(worker); 6745 6746 mutex_unlock(&wq_pool_attach_mutex); 6747 } 6748 } 6749 6750 /** 6751 * rebind_workers - rebind all workers of a pool to the associated CPU 6752 * @pool: pool of interest 6753 * 6754 * @pool->cpu is coming online. Rebind all workers to the CPU. 6755 */ 6756 static void rebind_workers(struct worker_pool *pool) 6757 { 6758 struct worker *worker; 6759 6760 lockdep_assert_held(&wq_pool_attach_mutex); 6761 6762 /* 6763 * Restore CPU affinity of all workers. As all idle workers should 6764 * be on the run-queue of the associated CPU before any local 6765 * wake-ups for concurrency management happen, restore CPU affinity 6766 * of all workers first and then clear UNBOUND. As we're called 6767 * from CPU_ONLINE, the following shouldn't fail. 6768 */ 6769 for_each_pool_worker(worker, pool) { 6770 kthread_set_per_cpu(worker->task, pool->cpu); 6771 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 6772 pool_allowed_cpus(pool)) < 0); 6773 } 6774 6775 raw_spin_lock_irq(&pool->lock); 6776 6777 pool->flags &= ~POOL_DISASSOCIATED; 6778 6779 for_each_pool_worker(worker, pool) { 6780 unsigned int worker_flags = worker->flags; 6781 6782 /* 6783 * We want to clear UNBOUND but can't directly call 6784 * worker_clr_flags() or adjust nr_running. Atomically 6785 * replace UNBOUND with another NOT_RUNNING flag REBOUND. 6786 * @worker will clear REBOUND using worker_clr_flags() when 6787 * it initiates the next execution cycle thus restoring 6788 * concurrency management. Note that when or whether 6789 * @worker clears REBOUND doesn't affect correctness. 6790 * 6791 * WRITE_ONCE() is necessary because @worker->flags may be 6792 * tested without holding any lock in 6793 * wq_worker_running(). Without it, NOT_RUNNING test may 6794 * fail incorrectly leading to premature concurrency 6795 * management operations. 6796 */ 6797 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 6798 worker_flags |= WORKER_REBOUND; 6799 worker_flags &= ~WORKER_UNBOUND; 6800 WRITE_ONCE(worker->flags, worker_flags); 6801 } 6802 6803 raw_spin_unlock_irq(&pool->lock); 6804 } 6805 6806 /** 6807 * restore_unbound_workers_cpumask - restore cpumask of unbound workers 6808 * @pool: unbound pool of interest 6809 * @cpu: the CPU which is coming up 6810 * 6811 * An unbound pool may end up with a cpumask which doesn't have any online 6812 * CPUs. When a worker of such pool get scheduled, the scheduler resets 6813 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 6814 * online CPU before, cpus_allowed of all its workers should be restored. 6815 */ 6816 static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 6817 { 6818 static cpumask_t cpumask; 6819 struct worker *worker; 6820 6821 lockdep_assert_held(&wq_pool_attach_mutex); 6822 6823 /* is @cpu allowed for @pool? */ 6824 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 6825 return; 6826 6827 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 6828 6829 /* as we're called from CPU_ONLINE, the following shouldn't fail */ 6830 for_each_pool_worker(worker, pool) 6831 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 6832 } 6833 6834 int workqueue_prepare_cpu(unsigned int cpu) 6835 { 6836 struct worker_pool *pool; 6837 6838 for_each_cpu_worker_pool(pool, cpu) { 6839 if (pool->nr_workers) 6840 continue; 6841 if (!create_worker(pool)) 6842 return -ENOMEM; 6843 } 6844 return 0; 6845 } 6846 6847 int workqueue_online_cpu(unsigned int cpu) 6848 { 6849 struct worker_pool *pool; 6850 struct workqueue_struct *wq; 6851 int pi; 6852 6853 mutex_lock(&wq_pool_mutex); 6854 6855 cpumask_set_cpu(cpu, wq_online_cpumask); 6856 6857 for_each_pool(pool, pi) { 6858 /* BH pools aren't affected by hotplug */ 6859 if (pool->flags & POOL_BH) 6860 continue; 6861 6862 mutex_lock(&wq_pool_attach_mutex); 6863 if (pool->cpu == cpu) 6864 rebind_workers(pool); 6865 else if (pool->cpu < 0) 6866 restore_unbound_workers_cpumask(pool, cpu); 6867 mutex_unlock(&wq_pool_attach_mutex); 6868 } 6869 6870 /* update pod affinity of unbound workqueues */ 6871 list_for_each_entry(wq, &workqueues, list) { 6872 struct workqueue_attrs *attrs = wq->unbound_attrs; 6873 6874 if (attrs) { 6875 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6876 int tcpu; 6877 6878 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6879 unbound_wq_update_pwq(wq, tcpu); 6880 6881 mutex_lock(&wq->mutex); 6882 wq_update_node_max_active(wq, -1); 6883 mutex_unlock(&wq->mutex); 6884 } 6885 } 6886 6887 mutex_unlock(&wq_pool_mutex); 6888 return 0; 6889 } 6890 6891 int workqueue_offline_cpu(unsigned int cpu) 6892 { 6893 struct workqueue_struct *wq; 6894 6895 /* unbinding per-cpu workers should happen on the local CPU */ 6896 if (WARN_ON(cpu != smp_processor_id())) 6897 return -1; 6898 6899 unbind_workers(cpu); 6900 6901 /* update pod affinity of unbound workqueues */ 6902 mutex_lock(&wq_pool_mutex); 6903 6904 cpumask_clear_cpu(cpu, wq_online_cpumask); 6905 6906 list_for_each_entry(wq, &workqueues, list) { 6907 struct workqueue_attrs *attrs = wq->unbound_attrs; 6908 6909 if (attrs) { 6910 const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 6911 int tcpu; 6912 6913 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 6914 unbound_wq_update_pwq(wq, tcpu); 6915 6916 mutex_lock(&wq->mutex); 6917 wq_update_node_max_active(wq, cpu); 6918 mutex_unlock(&wq->mutex); 6919 } 6920 } 6921 mutex_unlock(&wq_pool_mutex); 6922 6923 return 0; 6924 } 6925 6926 struct work_for_cpu { 6927 struct work_struct work; 6928 long (*fn)(void *); 6929 void *arg; 6930 long ret; 6931 }; 6932 6933 static void work_for_cpu_fn(struct work_struct *work) 6934 { 6935 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 6936 6937 wfc->ret = wfc->fn(wfc->arg); 6938 } 6939 6940 /** 6941 * work_on_cpu_key - run a function in thread context on a particular cpu 6942 * @cpu: the cpu to run on 6943 * @fn: the function to run 6944 * @arg: the function arg 6945 * @key: The lock class key for lock debugging purposes 6946 * 6947 * It is up to the caller to ensure that the cpu doesn't go offline. 6948 * The caller must not hold any locks which would prevent @fn from completing. 6949 * 6950 * Return: The value @fn returns. 6951 */ 6952 long work_on_cpu_key(int cpu, long (*fn)(void *), 6953 void *arg, struct lock_class_key *key) 6954 { 6955 struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 6956 6957 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key); 6958 schedule_work_on(cpu, &wfc.work); 6959 flush_work(&wfc.work); 6960 destroy_work_on_stack(&wfc.work); 6961 return wfc.ret; 6962 } 6963 EXPORT_SYMBOL_GPL(work_on_cpu_key); 6964 #endif /* CONFIG_SMP */ 6965 6966 #ifdef CONFIG_FREEZER 6967 6968 /** 6969 * freeze_workqueues_begin - begin freezing workqueues 6970 * 6971 * Start freezing workqueues. After this function returns, all freezable 6972 * workqueues will queue new works to their inactive_works list instead of 6973 * pool->worklist. 6974 * 6975 * CONTEXT: 6976 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 6977 */ 6978 void freeze_workqueues_begin(void) 6979 { 6980 struct workqueue_struct *wq; 6981 6982 mutex_lock(&wq_pool_mutex); 6983 6984 WARN_ON_ONCE(workqueue_freezing); 6985 workqueue_freezing = true; 6986 6987 list_for_each_entry(wq, &workqueues, list) { 6988 mutex_lock(&wq->mutex); 6989 wq_adjust_max_active(wq); 6990 mutex_unlock(&wq->mutex); 6991 } 6992 6993 mutex_unlock(&wq_pool_mutex); 6994 } 6995 6996 /** 6997 * freeze_workqueues_busy - are freezable workqueues still busy? 6998 * 6999 * Check whether freezing is complete. This function must be called 7000 * between freeze_workqueues_begin() and thaw_workqueues(). 7001 * 7002 * CONTEXT: 7003 * Grabs and releases wq_pool_mutex. 7004 * 7005 * Return: 7006 * %true if some freezable workqueues are still busy. %false if freezing 7007 * is complete. 7008 */ 7009 bool freeze_workqueues_busy(void) 7010 { 7011 bool busy = false; 7012 struct workqueue_struct *wq; 7013 struct pool_workqueue *pwq; 7014 7015 mutex_lock(&wq_pool_mutex); 7016 7017 WARN_ON_ONCE(!workqueue_freezing); 7018 7019 list_for_each_entry(wq, &workqueues, list) { 7020 if (!(wq->flags & WQ_FREEZABLE)) 7021 continue; 7022 /* 7023 * nr_active is monotonically decreasing. It's safe 7024 * to peek without lock. 7025 */ 7026 rcu_read_lock(); 7027 for_each_pwq(pwq, wq) { 7028 WARN_ON_ONCE(pwq->nr_active < 0); 7029 if (pwq->nr_active) { 7030 busy = true; 7031 rcu_read_unlock(); 7032 goto out_unlock; 7033 } 7034 } 7035 rcu_read_unlock(); 7036 } 7037 out_unlock: 7038 mutex_unlock(&wq_pool_mutex); 7039 return busy; 7040 } 7041 7042 /** 7043 * thaw_workqueues - thaw workqueues 7044 * 7045 * Thaw workqueues. Normal queueing is restored and all collected 7046 * frozen works are transferred to their respective pool worklists. 7047 * 7048 * CONTEXT: 7049 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 7050 */ 7051 void thaw_workqueues(void) 7052 { 7053 struct workqueue_struct *wq; 7054 7055 mutex_lock(&wq_pool_mutex); 7056 7057 if (!workqueue_freezing) 7058 goto out_unlock; 7059 7060 workqueue_freezing = false; 7061 7062 /* restore max_active and repopulate worklist */ 7063 list_for_each_entry(wq, &workqueues, list) { 7064 mutex_lock(&wq->mutex); 7065 wq_adjust_max_active(wq); 7066 mutex_unlock(&wq->mutex); 7067 } 7068 7069 out_unlock: 7070 mutex_unlock(&wq_pool_mutex); 7071 } 7072 #endif /* CONFIG_FREEZER */ 7073 7074 static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 7075 { 7076 LIST_HEAD(ctxs); 7077 int ret = 0; 7078 struct workqueue_struct *wq; 7079 struct apply_wqattrs_ctx *ctx, *n; 7080 7081 lockdep_assert_held(&wq_pool_mutex); 7082 7083 list_for_each_entry(wq, &workqueues, list) { 7084 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING)) 7085 continue; 7086 7087 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 7088 if (IS_ERR(ctx)) { 7089 ret = PTR_ERR(ctx); 7090 break; 7091 } 7092 7093 list_add_tail(&ctx->list, &ctxs); 7094 } 7095 7096 list_for_each_entry_safe(ctx, n, &ctxs, list) { 7097 if (!ret) 7098 apply_wqattrs_commit(ctx); 7099 apply_wqattrs_cleanup(ctx); 7100 } 7101 7102 if (!ret) { 7103 int cpu; 7104 struct worker_pool *pool; 7105 struct worker *worker; 7106 7107 mutex_lock(&wq_pool_attach_mutex); 7108 cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 7109 /* rescuer needs to respect cpumask changes when it is not attached */ 7110 list_for_each_entry(wq, &workqueues, list) { 7111 if (wq->rescuer && !wq->rescuer->pool) 7112 unbind_worker(wq->rescuer); 7113 } 7114 /* DISASSOCIATED worker needs to respect wq_unbound_cpumask */ 7115 for_each_possible_cpu(cpu) { 7116 for_each_cpu_worker_pool(pool, cpu) { 7117 if (!(pool->flags & POOL_DISASSOCIATED)) 7118 continue; 7119 for_each_pool_worker(worker, pool) 7120 unbind_worker(worker); 7121 } 7122 } 7123 mutex_unlock(&wq_pool_attach_mutex); 7124 } 7125 return ret; 7126 } 7127 7128 /** 7129 * workqueue_unbound_housekeeping_update - Propagate housekeeping cpumask update 7130 * @hk: the new housekeeping cpumask 7131 * 7132 * Update the unbound workqueue cpumask on top of the new housekeeping cpumask such 7133 * that the effective unbound affinity is the intersection of the new housekeeping 7134 * with the requested affinity set via nohz_full=/isolcpus= or sysfs. 7135 * 7136 * Return: 0 on success and -errno on failure. 7137 */ 7138 int workqueue_unbound_housekeeping_update(const struct cpumask *hk) 7139 { 7140 cpumask_var_t cpumask; 7141 int ret = 0; 7142 7143 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 7144 return -ENOMEM; 7145 7146 mutex_lock(&wq_pool_mutex); 7147 7148 /* 7149 * If the operation fails, it will fall back to 7150 * wq_requested_unbound_cpumask which is initially set to 7151 * HK_TYPE_DOMAIN house keeping mask and rewritten 7152 * by any subsequent write to workqueue/cpumask sysfs file. 7153 */ 7154 if (!cpumask_and(cpumask, wq_requested_unbound_cpumask, hk)) 7155 cpumask_copy(cpumask, wq_requested_unbound_cpumask); 7156 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 7157 ret = workqueue_apply_unbound_cpumask(cpumask); 7158 7159 /* Save the current isolated cpumask & export it via sysfs */ 7160 if (!ret) 7161 cpumask_andnot(wq_isolated_cpumask, cpu_possible_mask, hk); 7162 7163 mutex_unlock(&wq_pool_mutex); 7164 free_cpumask_var(cpumask); 7165 return ret; 7166 } 7167 7168 static int parse_affn_scope(const char *val) 7169 { 7170 return sysfs_match_string(wq_affn_names, val); 7171 } 7172 7173 static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 7174 { 7175 struct workqueue_struct *wq; 7176 int affn, cpu; 7177 7178 affn = parse_affn_scope(val); 7179 if (affn < 0) 7180 return affn; 7181 if (affn == WQ_AFFN_DFL) 7182 return -EINVAL; 7183 7184 cpus_read_lock(); 7185 mutex_lock(&wq_pool_mutex); 7186 7187 wq_affn_dfl = affn; 7188 7189 list_for_each_entry(wq, &workqueues, list) { 7190 for_each_online_cpu(cpu) 7191 unbound_wq_update_pwq(wq, cpu); 7192 } 7193 7194 mutex_unlock(&wq_pool_mutex); 7195 cpus_read_unlock(); 7196 7197 return 0; 7198 } 7199 7200 static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 7201 { 7202 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 7203 } 7204 7205 static const struct kernel_param_ops wq_affn_dfl_ops = { 7206 .set = wq_affn_dfl_set, 7207 .get = wq_affn_dfl_get, 7208 }; 7209 7210 module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 7211 7212 #ifdef CONFIG_SYSFS 7213 /* 7214 * Workqueues with WQ_SYSFS flag set is visible to userland via 7215 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 7216 * following attributes. 7217 * 7218 * per_cpu RO bool : whether the workqueue is per-cpu or unbound 7219 * max_active RW int : maximum number of in-flight work items 7220 * 7221 * Unbound workqueues have the following extra attributes. 7222 * 7223 * nice RW int : nice value of the workers 7224 * cpumask RW mask : bitmask of allowed CPUs for the workers 7225 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 7226 * affinity_strict RW bool : worker CPU affinity is strict 7227 */ 7228 struct wq_device { 7229 struct workqueue_struct *wq; 7230 struct device dev; 7231 }; 7232 7233 static struct workqueue_struct *dev_to_wq(struct device *dev) 7234 { 7235 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7236 7237 return wq_dev->wq; 7238 } 7239 7240 static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 7241 char *buf) 7242 { 7243 struct workqueue_struct *wq = dev_to_wq(dev); 7244 7245 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 7246 } 7247 static DEVICE_ATTR_RO(per_cpu); 7248 7249 static ssize_t max_active_show(struct device *dev, 7250 struct device_attribute *attr, char *buf) 7251 { 7252 struct workqueue_struct *wq = dev_to_wq(dev); 7253 7254 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 7255 } 7256 7257 static ssize_t max_active_store(struct device *dev, 7258 struct device_attribute *attr, const char *buf, 7259 size_t count) 7260 { 7261 struct workqueue_struct *wq = dev_to_wq(dev); 7262 int val; 7263 7264 if (sscanf(buf, "%d", &val) != 1 || val <= 0) 7265 return -EINVAL; 7266 7267 workqueue_set_max_active(wq, val); 7268 return count; 7269 } 7270 static DEVICE_ATTR_RW(max_active); 7271 7272 static struct attribute *wq_sysfs_attrs[] = { 7273 &dev_attr_per_cpu.attr, 7274 &dev_attr_max_active.attr, 7275 NULL, 7276 }; 7277 7278 static umode_t wq_sysfs_is_visible(struct kobject *kobj, struct attribute *a, int n) 7279 { 7280 struct device *dev = kobj_to_dev(kobj); 7281 struct workqueue_struct *wq = dev_to_wq(dev); 7282 7283 /* 7284 * Adjusting max_active breaks ordering guarantee. Changing it has no 7285 * effect on BH worker. Limit max_active to RO in such case. 7286 */ 7287 if (wq->flags & (WQ_BH | __WQ_ORDERED)) 7288 return 0444; 7289 return a->mode; 7290 } 7291 7292 static const struct attribute_group wq_sysfs_group = { 7293 .is_visible = wq_sysfs_is_visible, 7294 .attrs = wq_sysfs_attrs, 7295 }; 7296 __ATTRIBUTE_GROUPS(wq_sysfs); 7297 7298 static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 7299 char *buf) 7300 { 7301 struct workqueue_struct *wq = dev_to_wq(dev); 7302 int written; 7303 7304 mutex_lock(&wq->mutex); 7305 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 7306 mutex_unlock(&wq->mutex); 7307 7308 return written; 7309 } 7310 7311 /* prepare workqueue_attrs for sysfs store operations */ 7312 static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 7313 { 7314 struct workqueue_attrs *attrs; 7315 7316 lockdep_assert_held(&wq_pool_mutex); 7317 7318 attrs = alloc_workqueue_attrs(); 7319 if (!attrs) 7320 return NULL; 7321 7322 copy_workqueue_attrs(attrs, wq->unbound_attrs); 7323 return attrs; 7324 } 7325 7326 static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 7327 const char *buf, size_t count) 7328 { 7329 struct workqueue_struct *wq = dev_to_wq(dev); 7330 struct workqueue_attrs *attrs; 7331 int ret = -ENOMEM; 7332 7333 mutex_lock(&wq_pool_mutex); 7334 7335 attrs = wq_sysfs_prep_attrs(wq); 7336 if (!attrs) 7337 goto out_unlock; 7338 7339 if (sscanf(buf, "%d", &attrs->nice) == 1 && 7340 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 7341 ret = apply_workqueue_attrs_locked(wq, attrs); 7342 else 7343 ret = -EINVAL; 7344 7345 out_unlock: 7346 mutex_unlock(&wq_pool_mutex); 7347 free_workqueue_attrs(attrs); 7348 return ret ?: count; 7349 } 7350 7351 static ssize_t wq_cpumask_show(struct device *dev, 7352 struct device_attribute *attr, char *buf) 7353 { 7354 struct workqueue_struct *wq = dev_to_wq(dev); 7355 int written; 7356 7357 mutex_lock(&wq->mutex); 7358 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 7359 cpumask_pr_args(wq->unbound_attrs->cpumask)); 7360 mutex_unlock(&wq->mutex); 7361 return written; 7362 } 7363 7364 static ssize_t wq_cpumask_store(struct device *dev, 7365 struct device_attribute *attr, 7366 const char *buf, size_t count) 7367 { 7368 struct workqueue_struct *wq = dev_to_wq(dev); 7369 struct workqueue_attrs *attrs; 7370 int ret = -ENOMEM; 7371 7372 mutex_lock(&wq_pool_mutex); 7373 7374 attrs = wq_sysfs_prep_attrs(wq); 7375 if (!attrs) 7376 goto out_unlock; 7377 7378 ret = cpumask_parse(buf, attrs->cpumask); 7379 if (!ret) 7380 ret = apply_workqueue_attrs_locked(wq, attrs); 7381 7382 out_unlock: 7383 mutex_unlock(&wq_pool_mutex); 7384 free_workqueue_attrs(attrs); 7385 return ret ?: count; 7386 } 7387 7388 static ssize_t wq_affn_scope_show(struct device *dev, 7389 struct device_attribute *attr, char *buf) 7390 { 7391 struct workqueue_struct *wq = dev_to_wq(dev); 7392 int written; 7393 7394 mutex_lock(&wq->mutex); 7395 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 7396 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 7397 wq_affn_names[WQ_AFFN_DFL], 7398 wq_affn_names[wq_affn_dfl]); 7399 else 7400 written = scnprintf(buf, PAGE_SIZE, "%s\n", 7401 wq_affn_names[wq->unbound_attrs->affn_scope]); 7402 mutex_unlock(&wq->mutex); 7403 7404 return written; 7405 } 7406 7407 static ssize_t wq_affn_scope_store(struct device *dev, 7408 struct device_attribute *attr, 7409 const char *buf, size_t count) 7410 { 7411 struct workqueue_struct *wq = dev_to_wq(dev); 7412 struct workqueue_attrs *attrs; 7413 int affn, ret = -ENOMEM; 7414 7415 affn = parse_affn_scope(buf); 7416 if (affn < 0) 7417 return affn; 7418 7419 mutex_lock(&wq_pool_mutex); 7420 attrs = wq_sysfs_prep_attrs(wq); 7421 if (attrs) { 7422 attrs->affn_scope = affn; 7423 ret = apply_workqueue_attrs_locked(wq, attrs); 7424 } 7425 mutex_unlock(&wq_pool_mutex); 7426 free_workqueue_attrs(attrs); 7427 return ret ?: count; 7428 } 7429 7430 static ssize_t wq_affinity_strict_show(struct device *dev, 7431 struct device_attribute *attr, char *buf) 7432 { 7433 struct workqueue_struct *wq = dev_to_wq(dev); 7434 7435 return scnprintf(buf, PAGE_SIZE, "%d\n", 7436 wq->unbound_attrs->affn_strict); 7437 } 7438 7439 static ssize_t wq_affinity_strict_store(struct device *dev, 7440 struct device_attribute *attr, 7441 const char *buf, size_t count) 7442 { 7443 struct workqueue_struct *wq = dev_to_wq(dev); 7444 struct workqueue_attrs *attrs; 7445 int v, ret = -ENOMEM; 7446 7447 if (sscanf(buf, "%d", &v) != 1) 7448 return -EINVAL; 7449 7450 mutex_lock(&wq_pool_mutex); 7451 attrs = wq_sysfs_prep_attrs(wq); 7452 if (attrs) { 7453 attrs->affn_strict = (bool)v; 7454 ret = apply_workqueue_attrs_locked(wq, attrs); 7455 } 7456 mutex_unlock(&wq_pool_mutex); 7457 free_workqueue_attrs(attrs); 7458 return ret ?: count; 7459 } 7460 7461 static struct device_attribute wq_sysfs_unbound_attrs[] = { 7462 __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 7463 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 7464 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 7465 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 7466 __ATTR_NULL, 7467 }; 7468 7469 static const struct bus_type wq_subsys = { 7470 .name = "workqueue", 7471 .dev_groups = wq_sysfs_groups, 7472 }; 7473 7474 /** 7475 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 7476 * @cpumask: the cpumask to set 7477 * 7478 * The low-level workqueues cpumask is a global cpumask that limits 7479 * the affinity of all unbound workqueues. This function check the @cpumask 7480 * and apply it to all unbound workqueues and updates all pwqs of them. 7481 * 7482 * Return: 0 - Success 7483 * -EINVAL - Invalid @cpumask 7484 * -ENOMEM - Failed to allocate memory for attrs or pwqs. 7485 */ 7486 static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 7487 { 7488 int ret = -EINVAL; 7489 7490 /* 7491 * Not excluding isolated cpus on purpose. 7492 * If the user wishes to include them, we allow that. 7493 */ 7494 cpumask_and(cpumask, cpumask, cpu_possible_mask); 7495 if (!cpumask_empty(cpumask)) { 7496 ret = 0; 7497 mutex_lock(&wq_pool_mutex); 7498 if (!cpumask_equal(cpumask, wq_unbound_cpumask)) 7499 ret = workqueue_apply_unbound_cpumask(cpumask); 7500 if (!ret) 7501 cpumask_copy(wq_requested_unbound_cpumask, cpumask); 7502 mutex_unlock(&wq_pool_mutex); 7503 } 7504 7505 return ret; 7506 } 7507 7508 static ssize_t __wq_cpumask_show(struct device *dev, 7509 struct device_attribute *attr, char *buf, cpumask_var_t mask) 7510 { 7511 int written; 7512 7513 mutex_lock(&wq_pool_mutex); 7514 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask)); 7515 mutex_unlock(&wq_pool_mutex); 7516 7517 return written; 7518 } 7519 7520 static ssize_t cpumask_requested_show(struct device *dev, 7521 struct device_attribute *attr, char *buf) 7522 { 7523 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask); 7524 } 7525 static DEVICE_ATTR_RO(cpumask_requested); 7526 7527 static ssize_t cpumask_isolated_show(struct device *dev, 7528 struct device_attribute *attr, char *buf) 7529 { 7530 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask); 7531 } 7532 static DEVICE_ATTR_RO(cpumask_isolated); 7533 7534 static ssize_t cpumask_show(struct device *dev, 7535 struct device_attribute *attr, char *buf) 7536 { 7537 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask); 7538 } 7539 7540 static ssize_t cpumask_store(struct device *dev, 7541 struct device_attribute *attr, const char *buf, size_t count) 7542 { 7543 cpumask_var_t cpumask; 7544 int ret; 7545 7546 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 7547 return -ENOMEM; 7548 7549 ret = cpumask_parse(buf, cpumask); 7550 if (!ret) 7551 ret = workqueue_set_unbound_cpumask(cpumask); 7552 7553 free_cpumask_var(cpumask); 7554 return ret ? ret : count; 7555 } 7556 static DEVICE_ATTR_RW(cpumask); 7557 7558 static struct attribute *wq_sysfs_cpumask_attrs[] = { 7559 &dev_attr_cpumask.attr, 7560 &dev_attr_cpumask_requested.attr, 7561 &dev_attr_cpumask_isolated.attr, 7562 NULL, 7563 }; 7564 ATTRIBUTE_GROUPS(wq_sysfs_cpumask); 7565 7566 static int __init wq_sysfs_init(void) 7567 { 7568 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups); 7569 } 7570 core_initcall(wq_sysfs_init); 7571 7572 static void wq_device_release(struct device *dev) 7573 { 7574 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 7575 7576 kfree(wq_dev); 7577 } 7578 7579 /** 7580 * workqueue_sysfs_register - make a workqueue visible in sysfs 7581 * @wq: the workqueue to register 7582 * 7583 * Expose @wq in sysfs under /sys/bus/workqueue/devices. 7584 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 7585 * which is the preferred method. 7586 * 7587 * Workqueue user should use this function directly iff it wants to apply 7588 * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 7589 * apply_workqueue_attrs() may race against userland updating the 7590 * attributes. 7591 * 7592 * Return: 0 on success, -errno on failure. 7593 */ 7594 int workqueue_sysfs_register(struct workqueue_struct *wq) 7595 { 7596 struct wq_device *wq_dev; 7597 int ret; 7598 7599 wq->wq_dev = wq_dev = kzalloc_obj(*wq_dev); 7600 if (!wq_dev) 7601 return -ENOMEM; 7602 7603 wq_dev->wq = wq; 7604 wq_dev->dev.bus = &wq_subsys; 7605 wq_dev->dev.release = wq_device_release; 7606 dev_set_name(&wq_dev->dev, "%s", wq->name); 7607 7608 /* 7609 * unbound_attrs are created separately. Suppress uevent until 7610 * everything is ready. 7611 */ 7612 dev_set_uevent_suppress(&wq_dev->dev, true); 7613 7614 ret = device_register(&wq_dev->dev); 7615 if (ret) { 7616 put_device(&wq_dev->dev); 7617 wq->wq_dev = NULL; 7618 return ret; 7619 } 7620 7621 if (wq->flags & WQ_UNBOUND) { 7622 struct device_attribute *attr; 7623 7624 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 7625 ret = device_create_file(&wq_dev->dev, attr); 7626 if (ret) { 7627 device_unregister(&wq_dev->dev); 7628 wq->wq_dev = NULL; 7629 return ret; 7630 } 7631 } 7632 } 7633 7634 dev_set_uevent_suppress(&wq_dev->dev, false); 7635 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 7636 return 0; 7637 } 7638 7639 /** 7640 * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 7641 * @wq: the workqueue to unregister 7642 * 7643 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 7644 */ 7645 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 7646 { 7647 struct wq_device *wq_dev = wq->wq_dev; 7648 7649 if (!wq->wq_dev) 7650 return; 7651 7652 wq->wq_dev = NULL; 7653 device_unregister(&wq_dev->dev); 7654 } 7655 #else /* CONFIG_SYSFS */ 7656 static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 7657 #endif /* CONFIG_SYSFS */ 7658 7659 /* 7660 * Workqueue watchdog. 7661 * 7662 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 7663 * flush dependency, a concurrency managed work item which stays RUNNING 7664 * indefinitely. Workqueue stalls can be very difficult to debug as the 7665 * usual warning mechanisms don't trigger and internal workqueue state is 7666 * largely opaque. 7667 * 7668 * Workqueue watchdog monitors all worker pools periodically and dumps 7669 * state if some pools failed to make forward progress for a while where 7670 * forward progress is defined as the first item on ->worklist changing. 7671 * 7672 * This mechanism is controlled through the kernel parameter 7673 * "workqueue.watchdog_thresh" which can be updated at runtime through the 7674 * corresponding sysfs parameter file. 7675 */ 7676 #ifdef CONFIG_WQ_WATCHDOG 7677 7678 static unsigned long wq_watchdog_thresh = 30; 7679 static struct timer_list wq_watchdog_timer; 7680 7681 static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 7682 static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 7683 7684 static unsigned int wq_panic_on_stall = CONFIG_BOOTPARAM_WQ_STALL_PANIC; 7685 module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644); 7686 7687 static unsigned int wq_panic_on_stall_time; 7688 module_param_named(panic_on_stall_time, wq_panic_on_stall_time, uint, 0644); 7689 MODULE_PARM_DESC(panic_on_stall_time, "Panic if stall exceeds this many seconds (0=disabled)"); 7690 7691 /* 7692 * Show workers that might prevent the processing of pending work items. 7693 * A busy worker that is not running on the CPU (e.g. sleeping in 7694 * wait_event_idle() with PF_WQ_WORKER cleared) can stall the pool just as 7695 * effectively as a CPU-bound one, so dump every in-flight worker. 7696 */ 7697 static void show_cpu_pool_busy_workers(struct worker_pool *pool) 7698 { 7699 struct worker *worker; 7700 unsigned long irq_flags; 7701 int bkt; 7702 7703 raw_spin_lock_irqsave(&pool->lock, irq_flags); 7704 7705 hash_for_each(pool->busy_hash, bkt, worker, hentry) { 7706 /* 7707 * Defer printing to avoid deadlocks in console 7708 * drivers that queue work while holding locks 7709 * also taken in their write paths. 7710 */ 7711 printk_deferred_enter(); 7712 7713 pr_info("pool %d:\n", pool->id); 7714 sched_show_task(worker->task); 7715 7716 printk_deferred_exit(); 7717 } 7718 7719 raw_spin_unlock_irqrestore(&pool->lock, irq_flags); 7720 } 7721 7722 static void show_cpu_pools_busy_workers(void) 7723 { 7724 struct worker_pool *pool; 7725 int pi; 7726 7727 pr_info("Showing backtraces of busy workers in stalled worker pools:\n"); 7728 7729 rcu_read_lock(); 7730 7731 for_each_pool(pool, pi) { 7732 if (pool->cpu_stall) 7733 show_cpu_pool_busy_workers(pool); 7734 7735 } 7736 7737 rcu_read_unlock(); 7738 } 7739 7740 /* 7741 * It triggers a panic in two scenarios: when the total number of stalls 7742 * exceeds a threshold, and when a stall lasts longer than 7743 * wq_panic_on_stall_time 7744 */ 7745 static void panic_on_wq_watchdog(unsigned int stall_time_sec) 7746 { 7747 static unsigned int wq_stall; 7748 7749 if (wq_panic_on_stall) { 7750 wq_stall++; 7751 if (wq_stall >= wq_panic_on_stall) 7752 panic("workqueue: %u stall(s) exceeded threshold %u\n", 7753 wq_stall, wq_panic_on_stall); 7754 } 7755 7756 if (wq_panic_on_stall_time && stall_time_sec >= wq_panic_on_stall_time) 7757 panic("workqueue: stall lasted %us, exceeding threshold %us\n", 7758 stall_time_sec, wq_panic_on_stall_time); 7759 } 7760 7761 static void wq_watchdog_reset_touched(void) 7762 { 7763 int cpu; 7764 7765 wq_watchdog_touched = jiffies; 7766 for_each_possible_cpu(cpu) 7767 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 7768 } 7769 7770 static void wq_watchdog_timer_fn(struct timer_list *unused) 7771 { 7772 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7773 unsigned int max_stall_time = 0; 7774 bool lockup_detected = false; 7775 bool cpu_pool_stall = false; 7776 unsigned long now = jiffies; 7777 struct worker_pool *pool; 7778 unsigned int stall_time; 7779 int pi; 7780 7781 if (!thresh) 7782 return; 7783 7784 for_each_pool(pool, pi) { 7785 unsigned long pool_ts, touched, ts; 7786 7787 pool->cpu_stall = false; 7788 if (list_empty(&pool->worklist)) 7789 continue; 7790 7791 /* 7792 * If a virtual machine is stopped by the host it can look to 7793 * the watchdog like a stall. 7794 */ 7795 kvm_check_and_clear_guest_paused(); 7796 7797 /* get the latest of pool and touched timestamps */ 7798 if (pool->cpu >= 0) 7799 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 7800 else 7801 touched = READ_ONCE(wq_watchdog_touched); 7802 pool_ts = READ_ONCE(pool->last_progress_ts); 7803 7804 if (time_after(pool_ts, touched)) 7805 ts = pool_ts; 7806 else 7807 ts = touched; 7808 7809 /* 7810 * Did we stall? 7811 * 7812 * Do a lockless check first to do not disturb the system. 7813 * 7814 * Prevent false positives by double checking the timestamp 7815 * under pool->lock. The lock makes sure that the check reads 7816 * an updated pool->last_progress_ts when this CPU saw 7817 * an already updated pool->worklist above. It seems better 7818 * than adding another barrier into __queue_work() which 7819 * is a hotter path. 7820 */ 7821 if (time_after(now, ts + thresh)) { 7822 scoped_guard(raw_spinlock_irqsave, &pool->lock) { 7823 pool_ts = pool->last_progress_ts; 7824 if (time_after(pool_ts, touched)) 7825 ts = pool_ts; 7826 else 7827 ts = touched; 7828 } 7829 if (!time_after(now, ts + thresh)) 7830 continue; 7831 7832 lockup_detected = true; 7833 stall_time = jiffies_to_msecs(now - pool_ts) / 1000; 7834 max_stall_time = max(max_stall_time, stall_time); 7835 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) { 7836 pool->cpu_stall = true; 7837 cpu_pool_stall = true; 7838 } 7839 pr_emerg("BUG: workqueue lockup - pool"); 7840 pr_cont_pool_info(pool); 7841 pr_cont(" stuck for %us!\n", stall_time); 7842 } 7843 } 7844 7845 if (lockup_detected) 7846 show_all_workqueues(); 7847 7848 if (cpu_pool_stall) 7849 show_cpu_pools_busy_workers(); 7850 7851 if (lockup_detected) 7852 panic_on_wq_watchdog(max_stall_time); 7853 7854 wq_watchdog_reset_touched(); 7855 mod_timer(&wq_watchdog_timer, jiffies + thresh); 7856 } 7857 7858 notrace void wq_watchdog_touch(int cpu) 7859 { 7860 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 7861 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched); 7862 unsigned long now = jiffies; 7863 7864 if (cpu >= 0) 7865 per_cpu(wq_watchdog_touched_cpu, cpu) = now; 7866 else 7867 WARN_ONCE(1, "%s should be called with valid CPU", __func__); 7868 7869 /* Don't unnecessarily store to global cacheline */ 7870 if (time_after(now, touch_ts + thresh / 4)) 7871 WRITE_ONCE(wq_watchdog_touched, jiffies); 7872 } 7873 7874 static void wq_watchdog_set_thresh(unsigned long thresh) 7875 { 7876 wq_watchdog_thresh = 0; 7877 timer_delete_sync(&wq_watchdog_timer); 7878 7879 if (thresh) { 7880 wq_watchdog_thresh = thresh; 7881 wq_watchdog_reset_touched(); 7882 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 7883 } 7884 } 7885 7886 static int wq_watchdog_param_set_thresh(const char *val, 7887 const struct kernel_param *kp) 7888 { 7889 unsigned long thresh; 7890 int ret; 7891 7892 ret = kstrtoul(val, 0, &thresh); 7893 if (ret) 7894 return ret; 7895 7896 if (system_percpu_wq) 7897 wq_watchdog_set_thresh(thresh); 7898 else 7899 wq_watchdog_thresh = thresh; 7900 7901 return 0; 7902 } 7903 7904 static const struct kernel_param_ops wq_watchdog_thresh_ops = { 7905 .set = wq_watchdog_param_set_thresh, 7906 .get = param_get_ulong, 7907 }; 7908 7909 module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 7910 0644); 7911 7912 static void wq_watchdog_init(void) 7913 { 7914 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 7915 wq_watchdog_set_thresh(wq_watchdog_thresh); 7916 } 7917 7918 #else /* CONFIG_WQ_WATCHDOG */ 7919 7920 static inline void wq_watchdog_init(void) { } 7921 7922 #endif /* CONFIG_WQ_WATCHDOG */ 7923 7924 static void bh_pool_kick_normal(struct irq_work *irq_work) 7925 { 7926 raise_softirq_irqoff(TASKLET_SOFTIRQ); 7927 } 7928 7929 static void bh_pool_kick_highpri(struct irq_work *irq_work) 7930 { 7931 raise_softirq_irqoff(HI_SOFTIRQ); 7932 } 7933 7934 static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask) 7935 { 7936 if (!cpumask_intersects(wq_unbound_cpumask, mask)) { 7937 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n", 7938 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask)); 7939 return; 7940 } 7941 7942 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask); 7943 } 7944 7945 static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice) 7946 { 7947 BUG_ON(init_worker_pool(pool)); 7948 pool->cpu = cpu; 7949 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 7950 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 7951 pool->attrs->nice = nice; 7952 pool->attrs->affn_strict = true; 7953 pool->node = cpu_to_node(cpu); 7954 7955 /* alloc pool ID */ 7956 mutex_lock(&wq_pool_mutex); 7957 BUG_ON(worker_pool_assign_id(pool)); 7958 mutex_unlock(&wq_pool_mutex); 7959 } 7960 7961 /** 7962 * workqueue_init_early - early init for workqueue subsystem 7963 * 7964 * This is the first step of three-staged workqueue subsystem initialization and 7965 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 7966 * up. It sets up all the data structures and system workqueues and allows early 7967 * boot code to create workqueues and queue/cancel work items. Actual work item 7968 * execution starts only after kthreads can be created and scheduled right 7969 * before early initcalls. 7970 */ 7971 void __init workqueue_init_early(void) 7972 { 7973 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 7974 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 7975 void (*irq_work_fns[NR_STD_WORKER_POOLS])(struct irq_work *) = 7976 { bh_pool_kick_normal, bh_pool_kick_highpri }; 7977 int i, cpu; 7978 7979 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 7980 7981 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL)); 7982 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 7983 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL)); 7984 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL)); 7985 7986 cpumask_copy(wq_online_cpumask, cpu_online_mask); 7987 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask); 7988 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN)); 7989 if (!cpumask_empty(&wq_cmdline_cpumask)) 7990 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask); 7991 7992 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask); 7993 cpumask_andnot(wq_isolated_cpumask, cpu_possible_mask, 7994 housekeeping_cpumask(HK_TYPE_DOMAIN)); 7995 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 7996 7997 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs(); 7998 BUG_ON(!unbound_wq_update_pwq_attrs_buf); 7999 8000 /* 8001 * If nohz_full is enabled, set power efficient workqueue as unbound. 8002 * This allows workqueue items to be moved to HK CPUs. 8003 */ 8004 if (housekeeping_enabled(HK_TYPE_TICK)) 8005 wq_power_efficient = true; 8006 8007 /* initialize WQ_AFFN_SYSTEM pods */ 8008 pt->pod_cpus = kzalloc_objs(pt->pod_cpus[0], 1); 8009 pt->pod_node = kzalloc_objs(pt->pod_node[0], 1); 8010 pt->cpu_pod = kzalloc_objs(pt->cpu_pod[0], nr_cpu_ids); 8011 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 8012 8013 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 8014 8015 pt->nr_pods = 1; 8016 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 8017 pt->pod_node[0] = NUMA_NO_NODE; 8018 pt->cpu_pod[0] = 0; 8019 8020 /* initialize BH and CPU pools */ 8021 for_each_possible_cpu(cpu) { 8022 struct worker_pool *pool; 8023 8024 i = 0; 8025 for_each_bh_worker_pool(pool, cpu) { 8026 init_cpu_worker_pool(pool, cpu, std_nice[i]); 8027 pool->flags |= POOL_BH; 8028 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]); 8029 i++; 8030 } 8031 8032 i = 0; 8033 for_each_cpu_worker_pool(pool, cpu) 8034 init_cpu_worker_pool(pool, cpu, std_nice[i++]); 8035 } 8036 8037 /* create default unbound and ordered wq attrs */ 8038 for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 8039 struct workqueue_attrs *attrs; 8040 8041 BUG_ON(!(attrs = alloc_workqueue_attrs())); 8042 attrs->nice = std_nice[i]; 8043 unbound_std_wq_attrs[i] = attrs; 8044 8045 /* 8046 * An ordered wq should have only one pwq as ordering is 8047 * guaranteed by max_active which is enforced by pwqs. 8048 */ 8049 BUG_ON(!(attrs = alloc_workqueue_attrs())); 8050 attrs->nice = std_nice[i]; 8051 attrs->ordered = true; 8052 ordered_wq_attrs[i] = attrs; 8053 } 8054 8055 system_wq = alloc_workqueue("events", WQ_PERCPU | __WQ_DEPRECATED, 0); 8056 system_percpu_wq = alloc_workqueue("events", WQ_PERCPU, 0); 8057 system_highpri_wq = alloc_workqueue("events_highpri", 8058 WQ_HIGHPRI | WQ_PERCPU, 0); 8059 system_long_wq = alloc_workqueue("events_long", WQ_PERCPU, 0); 8060 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND | __WQ_DEPRECATED, WQ_MAX_ACTIVE); 8061 system_dfl_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, WQ_MAX_ACTIVE); 8062 system_freezable_wq = alloc_workqueue("events_freezable", 8063 WQ_FREEZABLE | WQ_PERCPU, 0); 8064 system_power_efficient_wq = alloc_workqueue("events_power_efficient", 8065 WQ_POWER_EFFICIENT | WQ_PERCPU, 0); 8066 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient", 8067 WQ_FREEZABLE | WQ_POWER_EFFICIENT | WQ_PERCPU, 0); 8068 system_bh_wq = alloc_workqueue("events_bh", WQ_BH | WQ_PERCPU, 0); 8069 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri", 8070 WQ_BH | WQ_HIGHPRI | WQ_PERCPU, 0); 8071 system_dfl_long_wq = alloc_workqueue("events_dfl_long", WQ_UNBOUND, WQ_MAX_ACTIVE); 8072 BUG_ON(!system_wq || !system_percpu_wq|| !system_highpri_wq || !system_long_wq || 8073 !system_unbound_wq || !system_freezable_wq || !system_dfl_wq || 8074 !system_power_efficient_wq || 8075 !system_freezable_power_efficient_wq || 8076 !system_bh_wq || !system_bh_highpri_wq || !system_dfl_long_wq); 8077 } 8078 8079 static void __init wq_cpu_intensive_thresh_init(void) 8080 { 8081 unsigned long thresh; 8082 unsigned long bogo; 8083 8084 pwq_release_worker = kthread_run_worker(0, "pool_workqueue_release"); 8085 BUG_ON(IS_ERR(pwq_release_worker)); 8086 8087 /* if the user set it to a specific value, keep it */ 8088 if (wq_cpu_intensive_thresh_us != ULONG_MAX) 8089 return; 8090 8091 /* 8092 * The default of 10ms is derived from the fact that most modern (as of 8093 * 2023) processors can do a lot in 10ms and that it's just below what 8094 * most consider human-perceivable. However, the kernel also runs on a 8095 * lot slower CPUs including microcontrollers where the threshold is way 8096 * too low. 8097 * 8098 * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 8099 * This is by no means accurate but it doesn't have to be. The mechanism 8100 * is still useful even when the threshold is fully scaled up. Also, as 8101 * the reports would usually be applicable to everyone, some machines 8102 * operating on longer thresholds won't significantly diminish their 8103 * usefulness. 8104 */ 8105 thresh = 10 * USEC_PER_MSEC; 8106 8107 /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 8108 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 8109 if (bogo < 4000) 8110 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 8111 8112 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 8113 loops_per_jiffy, bogo, thresh); 8114 8115 wq_cpu_intensive_thresh_us = thresh; 8116 } 8117 8118 /** 8119 * workqueue_init - bring workqueue subsystem fully online 8120 * 8121 * This is the second step of three-staged workqueue subsystem initialization 8122 * and invoked as soon as kthreads can be created and scheduled. Workqueues have 8123 * been created and work items queued on them, but there are no kworkers 8124 * executing the work items yet. Populate the worker pools with the initial 8125 * workers and enable future kworker creations. 8126 */ 8127 void __init workqueue_init(void) 8128 { 8129 struct workqueue_struct *wq; 8130 struct worker_pool *pool; 8131 int cpu, bkt; 8132 8133 wq_cpu_intensive_thresh_init(); 8134 8135 mutex_lock(&wq_pool_mutex); 8136 8137 /* 8138 * Per-cpu pools created earlier could be missing node hint. Fix them 8139 * up. Also, create a rescuer for workqueues that requested it. 8140 */ 8141 for_each_possible_cpu(cpu) { 8142 for_each_bh_worker_pool(pool, cpu) 8143 pool->node = cpu_to_node(cpu); 8144 for_each_cpu_worker_pool(pool, cpu) 8145 pool->node = cpu_to_node(cpu); 8146 } 8147 8148 list_for_each_entry(wq, &workqueues, list) { 8149 WARN(init_rescuer(wq), 8150 "workqueue: failed to create early rescuer for %s", 8151 wq->name); 8152 } 8153 8154 mutex_unlock(&wq_pool_mutex); 8155 8156 /* 8157 * Create the initial workers. A BH pool has one pseudo worker that 8158 * represents the shared BH execution context and thus doesn't get 8159 * affected by hotplug events. Create the BH pseudo workers for all 8160 * possible CPUs here. 8161 */ 8162 for_each_possible_cpu(cpu) 8163 for_each_bh_worker_pool(pool, cpu) 8164 BUG_ON(!create_worker(pool)); 8165 8166 for_each_online_cpu(cpu) { 8167 for_each_cpu_worker_pool(pool, cpu) { 8168 pool->flags &= ~POOL_DISASSOCIATED; 8169 BUG_ON(!create_worker(pool)); 8170 } 8171 } 8172 8173 hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 8174 BUG_ON(!create_worker(pool)); 8175 8176 wq_online = true; 8177 wq_watchdog_init(); 8178 } 8179 8180 /* 8181 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 8182 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 8183 * and consecutive pod ID. The rest of @pt is initialized accordingly. 8184 */ 8185 static void __init init_pod_type(struct wq_pod_type *pt, 8186 bool (*cpus_share_pod)(int, int)) 8187 { 8188 int cur, pre, cpu, pod; 8189 8190 pt->nr_pods = 0; 8191 8192 /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 8193 pt->cpu_pod = kzalloc_objs(pt->cpu_pod[0], nr_cpu_ids); 8194 BUG_ON(!pt->cpu_pod); 8195 8196 for_each_possible_cpu(cur) { 8197 for_each_possible_cpu(pre) { 8198 if (pre >= cur) { 8199 pt->cpu_pod[cur] = pt->nr_pods++; 8200 break; 8201 } 8202 if (cpus_share_pod(cur, pre)) { 8203 pt->cpu_pod[cur] = pt->cpu_pod[pre]; 8204 break; 8205 } 8206 } 8207 } 8208 8209 /* init the rest to match @pt->cpu_pod[] */ 8210 pt->pod_cpus = kzalloc_objs(pt->pod_cpus[0], pt->nr_pods); 8211 pt->pod_node = kzalloc_objs(pt->pod_node[0], pt->nr_pods); 8212 BUG_ON(!pt->pod_cpus || !pt->pod_node); 8213 8214 for (pod = 0; pod < pt->nr_pods; pod++) 8215 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 8216 8217 for_each_possible_cpu(cpu) { 8218 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 8219 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 8220 } 8221 } 8222 8223 static bool __init cpus_dont_share(int cpu0, int cpu1) 8224 { 8225 return false; 8226 } 8227 8228 static bool __init cpus_share_smt(int cpu0, int cpu1) 8229 { 8230 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 8231 } 8232 8233 static bool __init cpus_share_numa(int cpu0, int cpu1) 8234 { 8235 return cpu_to_node(cpu0) == cpu_to_node(cpu1); 8236 } 8237 8238 /* Maps each CPU to its shard index within the LLC pod it belongs to */ 8239 static int cpu_shard_id[NR_CPUS] __initdata; 8240 8241 /** 8242 * llc_count_cores - count distinct cores (SMT groups) within an LLC pod 8243 * @pod_cpus: the cpumask of CPUs in the LLC pod 8244 * @smt_pods: the SMT pod type, used to identify sibling groups 8245 * 8246 * A core is represented by the lowest-numbered CPU in its SMT group. Returns 8247 * the number of distinct cores found in @pod_cpus. 8248 */ 8249 static int __init llc_count_cores(const struct cpumask *pod_cpus, 8250 struct wq_pod_type *smt_pods) 8251 { 8252 const struct cpumask *sibling_cpus; 8253 int nr_cores = 0, c; 8254 8255 /* 8256 * Count distinct cores by only counting the first CPU in each 8257 * SMT sibling group. 8258 */ 8259 for_each_cpu(c, pod_cpus) { 8260 sibling_cpus = smt_pods->pod_cpus[smt_pods->cpu_pod[c]]; 8261 if (cpumask_first(sibling_cpus) == c) 8262 nr_cores++; 8263 } 8264 8265 return nr_cores; 8266 } 8267 8268 /* 8269 * llc_shard_size - number of cores in a given shard 8270 * 8271 * Cores are spread as evenly as possible. The first @nr_large_shards shards are 8272 * "large shards" with (cores_per_shard + 1) cores; the rest are "default 8273 * shards" with cores_per_shard cores. 8274 */ 8275 static int __init llc_shard_size(int shard_id, int cores_per_shard, int nr_large_shards) 8276 { 8277 /* The first @nr_large_shards shards are large shards */ 8278 if (shard_id < nr_large_shards) 8279 return cores_per_shard + 1; 8280 8281 /* The remaining shards are default shards */ 8282 return cores_per_shard; 8283 } 8284 8285 /* 8286 * llc_calc_shard_layout - compute the shard layout for an LLC pod 8287 * @nr_cores: number of distinct cores in the LLC pod 8288 * 8289 * Chooses the number of shards that keeps average shard size closest to 8290 * wq_cache_shard_size. Returns a struct describing the total number of shards, 8291 * the base size of each, and how many are large shards. 8292 */ 8293 static struct llc_shard_layout __init llc_calc_shard_layout(int nr_cores) 8294 { 8295 struct llc_shard_layout layout; 8296 8297 /* Ensure at least one shard; pick the count closest to the target size */ 8298 layout.nr_shards = max(1, DIV_ROUND_CLOSEST(nr_cores, wq_cache_shard_size)); 8299 layout.cores_per_shard = nr_cores / layout.nr_shards; 8300 layout.nr_large_shards = nr_cores % layout.nr_shards; 8301 8302 return layout; 8303 } 8304 8305 /* 8306 * llc_shard_is_full - check whether a shard has reached its core capacity 8307 * @cores_in_shard: number of cores already assigned to this shard 8308 * @shard_id: index of the shard being checked 8309 * @layout: the shard layout computed by llc_calc_shard_layout() 8310 * 8311 * Returns true if @cores_in_shard equals the expected size for @shard_id. 8312 */ 8313 static bool __init llc_shard_is_full(int cores_in_shard, int shard_id, 8314 const struct llc_shard_layout *layout) 8315 { 8316 return cores_in_shard == llc_shard_size(shard_id, layout->cores_per_shard, 8317 layout->nr_large_shards); 8318 } 8319 8320 /** 8321 * llc_populate_cpu_shard_id - populate cpu_shard_id[] for each CPU in an LLC pod 8322 * @pod_cpus: the cpumask of CPUs in the LLC pod 8323 * @smt_pods: the SMT pod type, used to identify sibling groups 8324 * @nr_cores: number of distinct cores in @pod_cpus (from llc_count_cores()) 8325 * 8326 * Walks @pod_cpus in order. At each SMT group leader, advances to the next 8327 * shard once the current shard is full. Results are written to cpu_shard_id[]. 8328 */ 8329 static void __init llc_populate_cpu_shard_id(const struct cpumask *pod_cpus, 8330 struct wq_pod_type *smt_pods, 8331 int nr_cores) 8332 { 8333 struct llc_shard_layout layout = llc_calc_shard_layout(nr_cores); 8334 const struct cpumask *sibling_cpus; 8335 /* Count the number of cores in the current shard_id */ 8336 int cores_in_shard = 0; 8337 unsigned int leader; 8338 /* This is a cursor for the shards. Go from zero to nr_shards - 1*/ 8339 int shard_id = 0; 8340 int c; 8341 8342 /* Iterate at every CPU for a given LLC pod, and assign it a shard */ 8343 for_each_cpu(c, pod_cpus) { 8344 sibling_cpus = smt_pods->pod_cpus[smt_pods->cpu_pod[c]]; 8345 if (cpumask_first(sibling_cpus) == c) { 8346 /* This is the CPU leader for the siblings */ 8347 if (llc_shard_is_full(cores_in_shard, shard_id, &layout)) { 8348 shard_id++; 8349 cores_in_shard = 0; 8350 } 8351 cores_in_shard++; 8352 cpu_shard_id[c] = shard_id; 8353 } else { 8354 /* 8355 * The siblings' shard MUST be the same as the leader. 8356 * never split threads in the same core. 8357 */ 8358 leader = cpumask_first(sibling_cpus); 8359 8360 /* 8361 * This check silences a Warray-bounds warning on UP 8362 * configs where NR_CPUS=1 makes cpu_shard_id[] 8363 * a single-element array, and the compiler can't 8364 * prove the index is always 0. 8365 */ 8366 if (WARN_ON_ONCE(leader >= nr_cpu_ids)) 8367 continue; 8368 cpu_shard_id[c] = cpu_shard_id[leader]; 8369 } 8370 } 8371 8372 WARN_ON_ONCE(shard_id != (layout.nr_shards - 1)); 8373 } 8374 8375 /** 8376 * precompute_cache_shard_ids - assign each CPU its shard index within its LLC 8377 * 8378 * Iterates over all LLC pods. For each pod, counts distinct cores then assigns 8379 * shard indices to all CPUs in the pod. Must be called after WQ_AFFN_CACHE and 8380 * WQ_AFFN_SMT have been initialized. 8381 */ 8382 static void __init precompute_cache_shard_ids(void) 8383 { 8384 struct wq_pod_type *llc_pods = &wq_pod_types[WQ_AFFN_CACHE]; 8385 struct wq_pod_type *smt_pods = &wq_pod_types[WQ_AFFN_SMT]; 8386 const struct cpumask *cpus_sharing_llc; 8387 int nr_cores; 8388 int pod; 8389 8390 if (!wq_cache_shard_size) { 8391 pr_warn("workqueue: cache_shard_size must be > 0, setting to 1\n"); 8392 wq_cache_shard_size = 1; 8393 } 8394 8395 for (pod = 0; pod < llc_pods->nr_pods; pod++) { 8396 cpus_sharing_llc = llc_pods->pod_cpus[pod]; 8397 8398 /* Number of cores in this given LLC */ 8399 nr_cores = llc_count_cores(cpus_sharing_llc, smt_pods); 8400 llc_populate_cpu_shard_id(cpus_sharing_llc, smt_pods, nr_cores); 8401 } 8402 } 8403 8404 /* 8405 * cpus_share_cache_shard - test whether two CPUs belong to the same cache shard 8406 * 8407 * Two CPUs share a cache shard if they are in the same LLC and have the same 8408 * shard index. Used as the pod affinity callback for WQ_AFFN_CACHE_SHARD. 8409 */ 8410 static bool __init cpus_share_cache_shard(int cpu0, int cpu1) 8411 { 8412 if (!cpus_share_cache(cpu0, cpu1)) 8413 return false; 8414 8415 return cpu_shard_id[cpu0] == cpu_shard_id[cpu1]; 8416 } 8417 8418 /** 8419 * workqueue_init_topology - initialize CPU pods for unbound workqueues 8420 * 8421 * This is the third step of three-staged workqueue subsystem initialization and 8422 * invoked after SMP and topology information are fully initialized. It 8423 * initializes the unbound CPU pods accordingly. 8424 */ 8425 void __init workqueue_init_topology(void) 8426 { 8427 struct workqueue_struct *wq; 8428 int cpu; 8429 8430 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 8431 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 8432 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 8433 precompute_cache_shard_ids(); 8434 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE_SHARD], cpus_share_cache_shard); 8435 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 8436 8437 wq_topo_initialized = true; 8438 8439 mutex_lock(&wq_pool_mutex); 8440 8441 /* 8442 * Workqueues allocated earlier would have all CPUs sharing the default 8443 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue 8444 * and CPU combinations to apply per-pod sharing. 8445 */ 8446 list_for_each_entry(wq, &workqueues, list) { 8447 for_each_online_cpu(cpu) 8448 unbound_wq_update_pwq(wq, cpu); 8449 if (wq->flags & WQ_UNBOUND) { 8450 mutex_lock(&wq->mutex); 8451 wq_update_node_max_active(wq, -1); 8452 mutex_unlock(&wq->mutex); 8453 } 8454 } 8455 8456 mutex_unlock(&wq_pool_mutex); 8457 } 8458 8459 void __warn_flushing_systemwide_wq(void) 8460 { 8461 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 8462 dump_stack(); 8463 } 8464 EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 8465 8466 static int __init workqueue_unbound_cpus_setup(char *str) 8467 { 8468 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 8469 cpumask_clear(&wq_cmdline_cpumask); 8470 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 8471 } 8472 8473 return 1; 8474 } 8475 __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 8476