11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4442f8570fSSasha Levin #include <linux/hashtable.h> 45e22bee78STejun Heo 46ea138446STejun Heo #include "workqueue_internal.h" 471da177e4SLinus Torvalds 48c8e55f36STejun Heo enum { 49bc2ae0f5STejun Heo /* 5024647570STejun Heo * worker_pool flags 51bc2ae0f5STejun Heo * 5224647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 53bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 54bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 55bc2ae0f5STejun Heo * is in effect. 56bc2ae0f5STejun Heo * 57bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 58bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 5924647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 60bc2ae0f5STejun Heo * 61bc2ae0f5STejun Heo * Note that DISASSOCIATED can be flipped only while holding 6224647570STejun Heo * assoc_mutex to avoid changing binding state while 6324647570STejun Heo * create_worker() is in progress. 64bc2ae0f5STejun Heo */ 6511ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 66552a37e9SLai Jiangshan POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */ 6724647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 6835b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 69db7bccf4STejun Heo 70c8e55f36STejun Heo /* worker flags */ 71c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 72c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 73c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 74e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 75fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 76f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 77e22bee78STejun Heo 785f7dabfdSLai Jiangshan WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | 79403c821dSTejun Heo WORKER_CPU_INTENSIVE, 80db7bccf4STejun Heo 81e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 824ce62e9eSTejun Heo 83c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 84db7bccf4STejun Heo 85e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 86e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 87e22bee78STejun Heo 883233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 893233cdbdSTejun Heo /* call for help after 10ms 903233cdbdSTejun Heo (min two ticks) */ 91e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 92e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds /* 95e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 96e22bee78STejun Heo * all cpus. Give -20. 97e22bee78STejun Heo */ 98e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 993270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 100c8e55f36STejun Heo }; 101c8e55f36STejun Heo 1021da177e4SLinus Torvalds /* 1034690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1044690c4abSTejun Heo * 105e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 106e41e704bSTejun Heo * everyone else. 1074690c4abSTejun Heo * 108e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 109e22bee78STejun Heo * only be modified and accessed from the local cpu. 110e22bee78STejun Heo * 111d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1124690c4abSTejun Heo * 113d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 114d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 115d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 116d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 117e22bee78STejun Heo * 11873f53c4aSTejun Heo * F: wq->flush_mutex protected. 11973f53c4aSTejun Heo * 1204690c4abSTejun Heo * W: workqueue_lock protected. 1214690c4abSTejun Heo */ 1224690c4abSTejun Heo 1232eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 124c34056a3STejun Heo 125bd7bdd43STejun Heo struct worker_pool { 126d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 127d84ff051STejun Heo int cpu; /* I: the associated cpu */ 1289daf9e67STejun Heo int id; /* I: pool ID */ 12911ebea50STejun Heo unsigned int flags; /* X: flags */ 130bd7bdd43STejun Heo 131bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 132bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 133ea1abd61SLai Jiangshan 134ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 135bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 136bd7bdd43STejun Heo 137bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 138bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 139bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 140bd7bdd43STejun Heo 141c9e7cf27STejun Heo /* workers are chained either in busy_hash or idle_list */ 142c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 143c9e7cf27STejun Heo /* L: hash of busy workers */ 144c9e7cf27STejun Heo 14524647570STejun Heo struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */ 146bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 147e19e397aSTejun Heo 148e19e397aSTejun Heo /* 149e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 150e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 151e19e397aSTejun Heo * cacheline. 152e19e397aSTejun Heo */ 153e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 1548b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1558b03ae3cSTejun Heo 1568b03ae3cSTejun Heo /* 157112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 158112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 159112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 160112202d9STejun Heo * number of flag bits. 1611da177e4SLinus Torvalds */ 162112202d9STejun Heo struct pool_workqueue { 163bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1644690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 16573f53c4aSTejun Heo int work_color; /* L: current color */ 16673f53c4aSTejun Heo int flush_color; /* L: flushing color */ 16773f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 16873f53c4aSTejun Heo /* L: nr of in_flight works */ 1691e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 170a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1711e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 17230cdf249STejun Heo struct list_head pwqs_node; /* I: node on wq->pwqs */ 173493a1724STejun Heo struct list_head mayday_node; /* W: node on wq->maydays */ 174e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 1751da177e4SLinus Torvalds 1761da177e4SLinus Torvalds /* 17773f53c4aSTejun Heo * Structure used to wait for workqueue flush. 17873f53c4aSTejun Heo */ 17973f53c4aSTejun Heo struct wq_flusher { 18073f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 18173f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 18273f53c4aSTejun Heo struct completion done; /* flush completion */ 18373f53c4aSTejun Heo }; 1841da177e4SLinus Torvalds 18573f53c4aSTejun Heo /* 1861da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 1871da177e4SLinus Torvalds * per-CPU workqueues: 1881da177e4SLinus Torvalds */ 1891da177e4SLinus Torvalds struct workqueue_struct { 1909c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 191420c0ddbSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ 19230cdf249STejun Heo struct list_head pwqs; /* I: all pwqs of this wq */ 1934690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 19473f53c4aSTejun Heo 19573f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 19673f53c4aSTejun Heo int work_color; /* F: current work color */ 19773f53c4aSTejun Heo int flush_color; /* F: current flush color */ 198112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 19973f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 20073f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 20173f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 20273f53c4aSTejun Heo 203493a1724STejun Heo struct list_head maydays; /* W: pwqs requesting rescue */ 204e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 205e22bee78STejun Heo 2069c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 207112202d9STejun Heo int saved_max_active; /* W: saved pwq max_active */ 2084e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2094e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2104e6045f1SJohannes Berg #endif 211b196be89STejun Heo char name[]; /* I: workqueue name */ 2121da177e4SLinus Torvalds }; 2131da177e4SLinus Torvalds 214e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 215e904e6c2STejun Heo 216d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 217d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 218044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2191aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 220044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 221d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 222044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 223f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 224044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 22524d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 226d320c038STejun Heo 22797bd2347STejun Heo #define CREATE_TRACE_POINTS 22897bd2347STejun Heo #include <trace/events/workqueue.h> 22997bd2347STejun Heo 23038db41d9STejun Heo #define for_each_std_worker_pool(pool, cpu) \ 231a60dc39cSTejun Heo for ((pool) = &std_worker_pools(cpu)[0]; \ 232a60dc39cSTejun Heo (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++) 2334ce62e9eSTejun Heo 234b67bfe0dSSasha Levin #define for_each_busy_worker(worker, i, pool) \ 235b67bfe0dSSasha Levin hash_for_each(pool->busy_hash, i, worker, hentry) 236db7bccf4STejun Heo 237706026c2STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 238f3421797STejun Heo unsigned int sw) 239f3421797STejun Heo { 240f3421797STejun Heo if (cpu < nr_cpu_ids) { 241f3421797STejun Heo if (sw & 1) { 242f3421797STejun Heo cpu = cpumask_next(cpu, mask); 243f3421797STejun Heo if (cpu < nr_cpu_ids) 244f3421797STejun Heo return cpu; 245f3421797STejun Heo } 246f3421797STejun Heo if (sw & 2) 247f3421797STejun Heo return WORK_CPU_UNBOUND; 248f3421797STejun Heo } 2496be19588SLai Jiangshan return WORK_CPU_END; 250f3421797STejun Heo } 251f3421797STejun Heo 25209884951STejun Heo /* 25309884951STejun Heo * CPU iterators 25409884951STejun Heo * 255706026c2STejun Heo * An extra cpu number is defined using an invalid cpu number 25609884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 257706026c2STejun Heo * specific CPU. The following iterators are similar to for_each_*_cpu() 258706026c2STejun Heo * iterators but also considers the unbound CPU. 25909884951STejun Heo * 260706026c2STejun Heo * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND 261706026c2STejun Heo * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND 26209884951STejun Heo */ 263706026c2STejun Heo #define for_each_wq_cpu(cpu) \ 264706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \ 2656be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 266706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3)) 267f3421797STejun Heo 268706026c2STejun Heo #define for_each_online_wq_cpu(cpu) \ 269706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \ 2706be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 271706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3)) 272f3421797STejun Heo 27349e3cf44STejun Heo /** 27417116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 27517116969STejun Heo * @pool: iteration cursor 27617116969STejun Heo * @id: integer used for iteration 27717116969STejun Heo */ 27817116969STejun Heo #define for_each_pool(pool, id) \ 27917116969STejun Heo idr_for_each_entry(&worker_pool_idr, pool, id) 28017116969STejun Heo 28117116969STejun Heo /** 28249e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 28349e3cf44STejun Heo * @pwq: iteration cursor 28449e3cf44STejun Heo * @wq: the target workqueue 28549e3cf44STejun Heo */ 28649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 28749e3cf44STejun Heo list_for_each_entry((pwq), &(wq)->pwqs, pwqs_node) 288f3421797STejun Heo 289dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 290dc186ad7SThomas Gleixner 291dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 292dc186ad7SThomas Gleixner 29399777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 29499777288SStanislaw Gruszka { 29599777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 29699777288SStanislaw Gruszka } 29799777288SStanislaw Gruszka 298dc186ad7SThomas Gleixner /* 299dc186ad7SThomas Gleixner * fixup_init is called when: 300dc186ad7SThomas Gleixner * - an active object is initialized 301dc186ad7SThomas Gleixner */ 302dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 303dc186ad7SThomas Gleixner { 304dc186ad7SThomas Gleixner struct work_struct *work = addr; 305dc186ad7SThomas Gleixner 306dc186ad7SThomas Gleixner switch (state) { 307dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 308dc186ad7SThomas Gleixner cancel_work_sync(work); 309dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 310dc186ad7SThomas Gleixner return 1; 311dc186ad7SThomas Gleixner default: 312dc186ad7SThomas Gleixner return 0; 313dc186ad7SThomas Gleixner } 314dc186ad7SThomas Gleixner } 315dc186ad7SThomas Gleixner 316dc186ad7SThomas Gleixner /* 317dc186ad7SThomas Gleixner * fixup_activate is called when: 318dc186ad7SThomas Gleixner * - an active object is activated 319dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 320dc186ad7SThomas Gleixner */ 321dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 322dc186ad7SThomas Gleixner { 323dc186ad7SThomas Gleixner struct work_struct *work = addr; 324dc186ad7SThomas Gleixner 325dc186ad7SThomas Gleixner switch (state) { 326dc186ad7SThomas Gleixner 327dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 328dc186ad7SThomas Gleixner /* 329dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 330dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 331dc186ad7SThomas Gleixner * is tracked in the object tracker. 332dc186ad7SThomas Gleixner */ 33322df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 334dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 335dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 336dc186ad7SThomas Gleixner return 0; 337dc186ad7SThomas Gleixner } 338dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 339dc186ad7SThomas Gleixner return 0; 340dc186ad7SThomas Gleixner 341dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 342dc186ad7SThomas Gleixner WARN_ON(1); 343dc186ad7SThomas Gleixner 344dc186ad7SThomas Gleixner default: 345dc186ad7SThomas Gleixner return 0; 346dc186ad7SThomas Gleixner } 347dc186ad7SThomas Gleixner } 348dc186ad7SThomas Gleixner 349dc186ad7SThomas Gleixner /* 350dc186ad7SThomas Gleixner * fixup_free is called when: 351dc186ad7SThomas Gleixner * - an active object is freed 352dc186ad7SThomas Gleixner */ 353dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 354dc186ad7SThomas Gleixner { 355dc186ad7SThomas Gleixner struct work_struct *work = addr; 356dc186ad7SThomas Gleixner 357dc186ad7SThomas Gleixner switch (state) { 358dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 359dc186ad7SThomas Gleixner cancel_work_sync(work); 360dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 361dc186ad7SThomas Gleixner return 1; 362dc186ad7SThomas Gleixner default: 363dc186ad7SThomas Gleixner return 0; 364dc186ad7SThomas Gleixner } 365dc186ad7SThomas Gleixner } 366dc186ad7SThomas Gleixner 367dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 368dc186ad7SThomas Gleixner .name = "work_struct", 36999777288SStanislaw Gruszka .debug_hint = work_debug_hint, 370dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 371dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 372dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 373dc186ad7SThomas Gleixner }; 374dc186ad7SThomas Gleixner 375dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 376dc186ad7SThomas Gleixner { 377dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 378dc186ad7SThomas Gleixner } 379dc186ad7SThomas Gleixner 380dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 381dc186ad7SThomas Gleixner { 382dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 383dc186ad7SThomas Gleixner } 384dc186ad7SThomas Gleixner 385dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 386dc186ad7SThomas Gleixner { 387dc186ad7SThomas Gleixner if (onstack) 388dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 389dc186ad7SThomas Gleixner else 390dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 391dc186ad7SThomas Gleixner } 392dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 393dc186ad7SThomas Gleixner 394dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 395dc186ad7SThomas Gleixner { 396dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 397dc186ad7SThomas Gleixner } 398dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 399dc186ad7SThomas Gleixner 400dc186ad7SThomas Gleixner #else 401dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 402dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 403dc186ad7SThomas Gleixner #endif 404dc186ad7SThomas Gleixner 40595402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 40695402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4071da177e4SLinus Torvalds static LIST_HEAD(workqueues); 408a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4091da177e4SLinus Torvalds 41014441960SOleg Nesterov /* 411e19e397aSTejun Heo * The CPU and unbound standard worker pools. The unbound ones have 412e19e397aSTejun Heo * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set. 41314441960SOleg Nesterov */ 414e19e397aSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 415a60dc39cSTejun Heo cpu_std_worker_pools); 416a60dc39cSTejun Heo static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS]; 417f3421797STejun Heo 4189daf9e67STejun Heo /* idr of all pools */ 4199daf9e67STejun Heo static DEFINE_MUTEX(worker_pool_idr_mutex); 4209daf9e67STejun Heo static DEFINE_IDR(worker_pool_idr); 4219daf9e67STejun Heo 422c34056a3STejun Heo static int worker_thread(void *__worker); 4231da177e4SLinus Torvalds 424a60dc39cSTejun Heo static struct worker_pool *std_worker_pools(int cpu) 4251da177e4SLinus Torvalds { 426f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 427a60dc39cSTejun Heo return per_cpu(cpu_std_worker_pools, cpu); 428f3421797STejun Heo else 429a60dc39cSTejun Heo return unbound_std_worker_pools; 4301da177e4SLinus Torvalds } 4311da177e4SLinus Torvalds 4324e8f0a60STejun Heo static int std_worker_pool_pri(struct worker_pool *pool) 4334e8f0a60STejun Heo { 434a60dc39cSTejun Heo return pool - std_worker_pools(pool->cpu); 4354e8f0a60STejun Heo } 4364e8f0a60STejun Heo 4379daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4389daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4399daf9e67STejun Heo { 4409daf9e67STejun Heo int ret; 4419daf9e67STejun Heo 4429daf9e67STejun Heo mutex_lock(&worker_pool_idr_mutex); 4439daf9e67STejun Heo idr_pre_get(&worker_pool_idr, GFP_KERNEL); 4449daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 4459daf9e67STejun Heo mutex_unlock(&worker_pool_idr_mutex); 4469daf9e67STejun Heo 4479daf9e67STejun Heo return ret; 4489daf9e67STejun Heo } 4499daf9e67STejun Heo 4507c3eed5cSTejun Heo /* 4517c3eed5cSTejun Heo * Lookup worker_pool by id. The idr currently is built during boot and 4527c3eed5cSTejun Heo * never modified. Don't worry about locking for now. 4537c3eed5cSTejun Heo */ 4547c3eed5cSTejun Heo static struct worker_pool *worker_pool_by_id(int pool_id) 4557c3eed5cSTejun Heo { 4567c3eed5cSTejun Heo return idr_find(&worker_pool_idr, pool_id); 4577c3eed5cSTejun Heo } 4587c3eed5cSTejun Heo 459d565ed63STejun Heo static struct worker_pool *get_std_worker_pool(int cpu, bool highpri) 460d565ed63STejun Heo { 461a60dc39cSTejun Heo struct worker_pool *pools = std_worker_pools(cpu); 462d565ed63STejun Heo 463a60dc39cSTejun Heo return &pools[highpri]; 464d565ed63STejun Heo } 465d565ed63STejun Heo 4667fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 467a848e3b6SOleg Nesterov { 4687fb98ea7STejun Heo return list_first_entry(&wq->pwqs, struct pool_workqueue, pwqs_node); 469f3421797STejun Heo } 470a848e3b6SOleg Nesterov 47173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 47273f53c4aSTejun Heo { 47373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 47473f53c4aSTejun Heo } 47573f53c4aSTejun Heo 47673f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 47773f53c4aSTejun Heo { 47873f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 47973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 48073f53c4aSTejun Heo } 48173f53c4aSTejun Heo 48273f53c4aSTejun Heo static int work_next_color(int color) 48373f53c4aSTejun Heo { 48473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 485b1f4ec17SOleg Nesterov } 486b1f4ec17SOleg Nesterov 4874594bf15SDavid Howells /* 488112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 489112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 4907c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 4917a22ad75STejun Heo * 492112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 493112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 494bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 495bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 4967a22ad75STejun Heo * 497112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 4987c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 499112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5007c3eed5cSTejun Heo * available only while the work item is queued. 501bbb68dfaSTejun Heo * 502bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 503bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 504bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 505bbb68dfaSTejun Heo * try to steal the PENDING bit. 5064594bf15SDavid Howells */ 5077a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5087a22ad75STejun Heo unsigned long flags) 5097a22ad75STejun Heo { 5106183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5117a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5127a22ad75STejun Heo } 5137a22ad75STejun Heo 514112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5154690c4abSTejun Heo unsigned long extra_flags) 516365970a1SDavid Howells { 517112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 518112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 519365970a1SDavid Howells } 520365970a1SDavid Howells 5214468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5224468a00fSLai Jiangshan int pool_id) 5234468a00fSLai Jiangshan { 5244468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5254468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5264468a00fSLai Jiangshan } 5274468a00fSLai Jiangshan 5287c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5297c3eed5cSTejun Heo int pool_id) 5304d707b9fSOleg Nesterov { 53123657bb1STejun Heo /* 53223657bb1STejun Heo * The following wmb is paired with the implied mb in 53323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 53423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 53523657bb1STejun Heo * owner. 53623657bb1STejun Heo */ 53723657bb1STejun Heo smp_wmb(); 5387c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5394d707b9fSOleg Nesterov } 5404d707b9fSOleg Nesterov 5417a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 542365970a1SDavid Howells { 5437c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5447c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5457a22ad75STejun Heo } 5467a22ad75STejun Heo 547112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5487a22ad75STejun Heo { 549e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5507a22ad75STejun Heo 551112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 552e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 553e120153dSTejun Heo else 554e120153dSTejun Heo return NULL; 5557a22ad75STejun Heo } 5567a22ad75STejun Heo 5577c3eed5cSTejun Heo /** 5587c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5597c3eed5cSTejun Heo * @work: the work item of interest 5607c3eed5cSTejun Heo * 5617c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 5627c3eed5cSTejun Heo */ 5637c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 5647a22ad75STejun Heo { 565e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5667c3eed5cSTejun Heo struct worker_pool *pool; 5677c3eed5cSTejun Heo int pool_id; 5687a22ad75STejun Heo 569112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 570112202d9STejun Heo return ((struct pool_workqueue *) 5717c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 5727a22ad75STejun Heo 5737c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 5747c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 5757a22ad75STejun Heo return NULL; 5767a22ad75STejun Heo 5777c3eed5cSTejun Heo pool = worker_pool_by_id(pool_id); 5787c3eed5cSTejun Heo WARN_ON_ONCE(!pool); 5797c3eed5cSTejun Heo return pool; 5807c3eed5cSTejun Heo } 5817c3eed5cSTejun Heo 5827c3eed5cSTejun Heo /** 5837c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 5847c3eed5cSTejun Heo * @work: the work item of interest 5857c3eed5cSTejun Heo * 5867c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 5877c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 5887c3eed5cSTejun Heo */ 5897c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 5907c3eed5cSTejun Heo { 59154d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 5927c3eed5cSTejun Heo 593112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 594112202d9STejun Heo return ((struct pool_workqueue *) 59554d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 59654d5b7d0SLai Jiangshan 59754d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 5987c3eed5cSTejun Heo } 5997c3eed5cSTejun Heo 600bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 601bbb68dfaSTejun Heo { 6027c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 603bbb68dfaSTejun Heo 6047c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6057c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 606bbb68dfaSTejun Heo } 607bbb68dfaSTejun Heo 608bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 609bbb68dfaSTejun Heo { 610bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 611bbb68dfaSTejun Heo 612112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 613bbb68dfaSTejun Heo } 614bbb68dfaSTejun Heo 615e22bee78STejun Heo /* 6163270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6173270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 618d565ed63STejun Heo * they're being called with pool->lock held. 619e22bee78STejun Heo */ 620e22bee78STejun Heo 62163d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 622649027d7STejun Heo { 623e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 624649027d7STejun Heo } 625649027d7STejun Heo 626e22bee78STejun Heo /* 627e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 628e22bee78STejun Heo * running workers. 629974271c4STejun Heo * 630974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 631706026c2STejun Heo * function will always return %true for unbound pools as long as the 632974271c4STejun Heo * worklist isn't empty. 633e22bee78STejun Heo */ 63463d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 635e22bee78STejun Heo { 63663d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 637e22bee78STejun Heo } 638e22bee78STejun Heo 639e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 64063d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 641e22bee78STejun Heo { 64263d95a91STejun Heo return pool->nr_idle; 643e22bee78STejun Heo } 644e22bee78STejun Heo 645e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 64663d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 647e22bee78STejun Heo { 648e19e397aSTejun Heo return !list_empty(&pool->worklist) && 649e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 650e22bee78STejun Heo } 651e22bee78STejun Heo 652e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 65363d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 654e22bee78STejun Heo { 65563d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 656e22bee78STejun Heo } 657e22bee78STejun Heo 658e22bee78STejun Heo /* Do I need to be the manager? */ 65963d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 660e22bee78STejun Heo { 66163d95a91STejun Heo return need_to_create_worker(pool) || 66211ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 663e22bee78STejun Heo } 664e22bee78STejun Heo 665e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 66663d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 667e22bee78STejun Heo { 668552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 66963d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 67063d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 671e22bee78STejun Heo 672ea1abd61SLai Jiangshan /* 673ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 674ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 675ea1abd61SLai Jiangshan */ 676ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 677ea1abd61SLai Jiangshan return false; 678ea1abd61SLai Jiangshan 679e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 680e22bee78STejun Heo } 681e22bee78STejun Heo 682e22bee78STejun Heo /* 683e22bee78STejun Heo * Wake up functions. 684e22bee78STejun Heo */ 685e22bee78STejun Heo 6867e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 68763d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 6887e11629dSTejun Heo { 68963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 6907e11629dSTejun Heo return NULL; 6917e11629dSTejun Heo 69263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 6937e11629dSTejun Heo } 6947e11629dSTejun Heo 6957e11629dSTejun Heo /** 6967e11629dSTejun Heo * wake_up_worker - wake up an idle worker 69763d95a91STejun Heo * @pool: worker pool to wake worker from 6987e11629dSTejun Heo * 69963d95a91STejun Heo * Wake up the first idle worker of @pool. 7007e11629dSTejun Heo * 7017e11629dSTejun Heo * CONTEXT: 702d565ed63STejun Heo * spin_lock_irq(pool->lock). 7037e11629dSTejun Heo */ 70463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7057e11629dSTejun Heo { 70663d95a91STejun Heo struct worker *worker = first_worker(pool); 7077e11629dSTejun Heo 7087e11629dSTejun Heo if (likely(worker)) 7097e11629dSTejun Heo wake_up_process(worker->task); 7107e11629dSTejun Heo } 7117e11629dSTejun Heo 7124690c4abSTejun Heo /** 713e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 714e22bee78STejun Heo * @task: task waking up 715e22bee78STejun Heo * @cpu: CPU @task is waking up to 716e22bee78STejun Heo * 717e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 718e22bee78STejun Heo * being awoken. 719e22bee78STejun Heo * 720e22bee78STejun Heo * CONTEXT: 721e22bee78STejun Heo * spin_lock_irq(rq->lock) 722e22bee78STejun Heo */ 723d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 724e22bee78STejun Heo { 725e22bee78STejun Heo struct worker *worker = kthread_data(task); 726e22bee78STejun Heo 72736576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 728ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 729e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 730e22bee78STejun Heo } 73136576000SJoonsoo Kim } 732e22bee78STejun Heo 733e22bee78STejun Heo /** 734e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 735e22bee78STejun Heo * @task: task going to sleep 736e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 737e22bee78STejun Heo * 738e22bee78STejun Heo * This function is called during schedule() when a busy worker is 739e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 740e22bee78STejun Heo * returning pointer to its task. 741e22bee78STejun Heo * 742e22bee78STejun Heo * CONTEXT: 743e22bee78STejun Heo * spin_lock_irq(rq->lock) 744e22bee78STejun Heo * 745e22bee78STejun Heo * RETURNS: 746e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 747e22bee78STejun Heo */ 748d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 749e22bee78STejun Heo { 750e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 751111c225aSTejun Heo struct worker_pool *pool; 752e22bee78STejun Heo 753111c225aSTejun Heo /* 754111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 755111c225aSTejun Heo * workers, also reach here, let's not access anything before 756111c225aSTejun Heo * checking NOT_RUNNING. 757111c225aSTejun Heo */ 7582d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 759e22bee78STejun Heo return NULL; 760e22bee78STejun Heo 761111c225aSTejun Heo pool = worker->pool; 762111c225aSTejun Heo 763e22bee78STejun Heo /* this can only happen on the local cpu */ 7646183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 7656183c009STejun Heo return NULL; 766e22bee78STejun Heo 767e22bee78STejun Heo /* 768e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 769e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 770e22bee78STejun Heo * Please read comment there. 771e22bee78STejun Heo * 772628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 773628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 774628c78e7STejun Heo * disabled, which in turn means that none else could be 775d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 776628c78e7STejun Heo * lock is safe. 777e22bee78STejun Heo */ 778e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 779e19e397aSTejun Heo !list_empty(&pool->worklist)) 78063d95a91STejun Heo to_wakeup = first_worker(pool); 781e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 782e22bee78STejun Heo } 783e22bee78STejun Heo 784e22bee78STejun Heo /** 785e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 786cb444766STejun Heo * @worker: self 787d302f017STejun Heo * @flags: flags to set 788d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 789d302f017STejun Heo * 790e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 791e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 792e22bee78STejun Heo * woken up. 793d302f017STejun Heo * 794cb444766STejun Heo * CONTEXT: 795d565ed63STejun Heo * spin_lock_irq(pool->lock) 796d302f017STejun Heo */ 797d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 798d302f017STejun Heo bool wakeup) 799d302f017STejun Heo { 800bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 801e22bee78STejun Heo 802cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 803cb444766STejun Heo 804e22bee78STejun Heo /* 805e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 806e22bee78STejun Heo * wake up an idle worker as necessary if requested by 807e22bee78STejun Heo * @wakeup. 808e22bee78STejun Heo */ 809e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 810e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 811e22bee78STejun Heo if (wakeup) { 812e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 813bd7bdd43STejun Heo !list_empty(&pool->worklist)) 81463d95a91STejun Heo wake_up_worker(pool); 815e22bee78STejun Heo } else 816e19e397aSTejun Heo atomic_dec(&pool->nr_running); 817e22bee78STejun Heo } 818e22bee78STejun Heo 819d302f017STejun Heo worker->flags |= flags; 820d302f017STejun Heo } 821d302f017STejun Heo 822d302f017STejun Heo /** 823e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 824cb444766STejun Heo * @worker: self 825d302f017STejun Heo * @flags: flags to clear 826d302f017STejun Heo * 827e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 828d302f017STejun Heo * 829cb444766STejun Heo * CONTEXT: 830d565ed63STejun Heo * spin_lock_irq(pool->lock) 831d302f017STejun Heo */ 832d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 833d302f017STejun Heo { 83463d95a91STejun Heo struct worker_pool *pool = worker->pool; 835e22bee78STejun Heo unsigned int oflags = worker->flags; 836e22bee78STejun Heo 837cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 838cb444766STejun Heo 839d302f017STejun Heo worker->flags &= ~flags; 840e22bee78STejun Heo 84142c025f3STejun Heo /* 84242c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 84342c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 84442c025f3STejun Heo * of multiple flags, not a single flag. 84542c025f3STejun Heo */ 846e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 847e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 848e19e397aSTejun Heo atomic_inc(&pool->nr_running); 849d302f017STejun Heo } 850d302f017STejun Heo 851d302f017STejun Heo /** 8528cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 853c9e7cf27STejun Heo * @pool: pool of interest 8548cca0eeaSTejun Heo * @work: work to find worker for 8558cca0eeaSTejun Heo * 856c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 857c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 858a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 859a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 860a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 861a2c1c57bSTejun Heo * being executed. 862a2c1c57bSTejun Heo * 863a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 864a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 865a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 866a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 867a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 868a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 869a2c1c57bSTejun Heo * 870a2c1c57bSTejun Heo * This function checks the work item address, work function and workqueue 871a2c1c57bSTejun Heo * to avoid false positives. Note that this isn't complete as one may 872a2c1c57bSTejun Heo * construct a work function which can introduce dependency onto itself 873a2c1c57bSTejun Heo * through a recycled work item. Well, if somebody wants to shoot oneself 874a2c1c57bSTejun Heo * in the foot that badly, there's only so much we can do, and if such 875a2c1c57bSTejun Heo * deadlock actually occurs, it should be easy to locate the culprit work 876a2c1c57bSTejun Heo * function. 8778cca0eeaSTejun Heo * 8788cca0eeaSTejun Heo * CONTEXT: 879d565ed63STejun Heo * spin_lock_irq(pool->lock). 8808cca0eeaSTejun Heo * 8818cca0eeaSTejun Heo * RETURNS: 8828cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8838cca0eeaSTejun Heo * otherwise. 8848cca0eeaSTejun Heo */ 885c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 8868cca0eeaSTejun Heo struct work_struct *work) 8878cca0eeaSTejun Heo { 88842f8570fSSasha Levin struct worker *worker; 88942f8570fSSasha Levin 890b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 891a2c1c57bSTejun Heo (unsigned long)work) 892a2c1c57bSTejun Heo if (worker->current_work == work && 893a2c1c57bSTejun Heo worker->current_func == work->func) 89442f8570fSSasha Levin return worker; 89542f8570fSSasha Levin 89642f8570fSSasha Levin return NULL; 8978cca0eeaSTejun Heo } 8988cca0eeaSTejun Heo 8998cca0eeaSTejun Heo /** 900bf4ede01STejun Heo * move_linked_works - move linked works to a list 901bf4ede01STejun Heo * @work: start of series of works to be scheduled 902bf4ede01STejun Heo * @head: target list to append @work to 903bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 904bf4ede01STejun Heo * 905bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 906bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 907bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 908bf4ede01STejun Heo * 909bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 910bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 911bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 912bf4ede01STejun Heo * 913bf4ede01STejun Heo * CONTEXT: 914d565ed63STejun Heo * spin_lock_irq(pool->lock). 915bf4ede01STejun Heo */ 916bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 917bf4ede01STejun Heo struct work_struct **nextp) 918bf4ede01STejun Heo { 919bf4ede01STejun Heo struct work_struct *n; 920bf4ede01STejun Heo 921bf4ede01STejun Heo /* 922bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 923bf4ede01STejun Heo * use NULL for list head. 924bf4ede01STejun Heo */ 925bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 926bf4ede01STejun Heo list_move_tail(&work->entry, head); 927bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 928bf4ede01STejun Heo break; 929bf4ede01STejun Heo } 930bf4ede01STejun Heo 931bf4ede01STejun Heo /* 932bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 933bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 934bf4ede01STejun Heo * needs to be updated. 935bf4ede01STejun Heo */ 936bf4ede01STejun Heo if (nextp) 937bf4ede01STejun Heo *nextp = n; 938bf4ede01STejun Heo } 939bf4ede01STejun Heo 940112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 941bf4ede01STejun Heo { 942112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 943bf4ede01STejun Heo 944bf4ede01STejun Heo trace_workqueue_activate_work(work); 945112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 946bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 947112202d9STejun Heo pwq->nr_active++; 948bf4ede01STejun Heo } 949bf4ede01STejun Heo 950112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 9513aa62497SLai Jiangshan { 952112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 9533aa62497SLai Jiangshan struct work_struct, entry); 9543aa62497SLai Jiangshan 955112202d9STejun Heo pwq_activate_delayed_work(work); 9563aa62497SLai Jiangshan } 9573aa62497SLai Jiangshan 958bf4ede01STejun Heo /** 959112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 960112202d9STejun Heo * @pwq: pwq of interest 961bf4ede01STejun Heo * @color: color of work which left the queue 962bf4ede01STejun Heo * 963bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 964112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 965bf4ede01STejun Heo * 966bf4ede01STejun Heo * CONTEXT: 967d565ed63STejun Heo * spin_lock_irq(pool->lock). 968bf4ede01STejun Heo */ 969112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 970bf4ede01STejun Heo { 971bf4ede01STejun Heo /* ignore uncolored works */ 972bf4ede01STejun Heo if (color == WORK_NO_COLOR) 973bf4ede01STejun Heo return; 974bf4ede01STejun Heo 975112202d9STejun Heo pwq->nr_in_flight[color]--; 976bf4ede01STejun Heo 977112202d9STejun Heo pwq->nr_active--; 978112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 979bf4ede01STejun Heo /* one down, submit a delayed one */ 980112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 981112202d9STejun Heo pwq_activate_first_delayed(pwq); 982bf4ede01STejun Heo } 983bf4ede01STejun Heo 984bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 985112202d9STejun Heo if (likely(pwq->flush_color != color)) 986bf4ede01STejun Heo return; 987bf4ede01STejun Heo 988bf4ede01STejun Heo /* are there still in-flight works? */ 989112202d9STejun Heo if (pwq->nr_in_flight[color]) 990bf4ede01STejun Heo return; 991bf4ede01STejun Heo 992112202d9STejun Heo /* this pwq is done, clear flush_color */ 993112202d9STejun Heo pwq->flush_color = -1; 994bf4ede01STejun Heo 995bf4ede01STejun Heo /* 996112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 997bf4ede01STejun Heo * will handle the rest. 998bf4ede01STejun Heo */ 999112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1000112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 1001bf4ede01STejun Heo } 1002bf4ede01STejun Heo 100336e227d2STejun Heo /** 1004bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 100536e227d2STejun Heo * @work: work item to steal 100636e227d2STejun Heo * @is_dwork: @work is a delayed_work 1007bbb68dfaSTejun Heo * @flags: place to store irq state 100836e227d2STejun Heo * 100936e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 101036e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 101136e227d2STejun Heo * 101236e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 101336e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 101436e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1015bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1016bbb68dfaSTejun Heo * for arbitrarily long 101736e227d2STejun Heo * 1018bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1019e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1020e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1021e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1022bbb68dfaSTejun Heo * 1023bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1024bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1025bbb68dfaSTejun Heo * 1026e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1027bf4ede01STejun Heo */ 1028bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1029bbb68dfaSTejun Heo unsigned long *flags) 1030bf4ede01STejun Heo { 1031d565ed63STejun Heo struct worker_pool *pool; 1032112202d9STejun Heo struct pool_workqueue *pwq; 1033bf4ede01STejun Heo 1034bbb68dfaSTejun Heo local_irq_save(*flags); 1035bbb68dfaSTejun Heo 103636e227d2STejun Heo /* try to steal the timer if it exists */ 103736e227d2STejun Heo if (is_dwork) { 103836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 103936e227d2STejun Heo 1040e0aecdd8STejun Heo /* 1041e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1042e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1043e0aecdd8STejun Heo * running on the local CPU. 1044e0aecdd8STejun Heo */ 104536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 104636e227d2STejun Heo return 1; 104736e227d2STejun Heo } 104836e227d2STejun Heo 104936e227d2STejun Heo /* try to claim PENDING the normal way */ 1050bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1051bf4ede01STejun Heo return 0; 1052bf4ede01STejun Heo 1053bf4ede01STejun Heo /* 1054bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1055bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1056bf4ede01STejun Heo */ 1057d565ed63STejun Heo pool = get_work_pool(work); 1058d565ed63STejun Heo if (!pool) 1059bbb68dfaSTejun Heo goto fail; 1060bf4ede01STejun Heo 1061d565ed63STejun Heo spin_lock(&pool->lock); 1062bf4ede01STejun Heo /* 1063112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1064112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1065112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1066112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1067112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 10680b3dae68SLai Jiangshan * item is currently queued on that pool. 1069bf4ede01STejun Heo */ 1070112202d9STejun Heo pwq = get_work_pwq(work); 1071112202d9STejun Heo if (pwq && pwq->pool == pool) { 1072bf4ede01STejun Heo debug_work_deactivate(work); 10733aa62497SLai Jiangshan 10743aa62497SLai Jiangshan /* 107516062836STejun Heo * A delayed work item cannot be grabbed directly because 107616062836STejun Heo * it might have linked NO_COLOR work items which, if left 1077112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 107816062836STejun Heo * management later on and cause stall. Make sure the work 107916062836STejun Heo * item is activated before grabbing. 10803aa62497SLai Jiangshan */ 10813aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1082112202d9STejun Heo pwq_activate_delayed_work(work); 10833aa62497SLai Jiangshan 1084bf4ede01STejun Heo list_del_init(&work->entry); 1085112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 108636e227d2STejun Heo 1087112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 10884468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 10894468a00fSLai Jiangshan 1090d565ed63STejun Heo spin_unlock(&pool->lock); 109136e227d2STejun Heo return 1; 1092bf4ede01STejun Heo } 1093d565ed63STejun Heo spin_unlock(&pool->lock); 1094bbb68dfaSTejun Heo fail: 1095bbb68dfaSTejun Heo local_irq_restore(*flags); 1096bbb68dfaSTejun Heo if (work_is_canceling(work)) 1097bbb68dfaSTejun Heo return -ENOENT; 1098bbb68dfaSTejun Heo cpu_relax(); 109936e227d2STejun Heo return -EAGAIN; 1100bf4ede01STejun Heo } 1101bf4ede01STejun Heo 1102bf4ede01STejun Heo /** 1103706026c2STejun Heo * insert_work - insert a work into a pool 1104112202d9STejun Heo * @pwq: pwq @work belongs to 11054690c4abSTejun Heo * @work: work to insert 11064690c4abSTejun Heo * @head: insertion point 11074690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11084690c4abSTejun Heo * 1109112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1110706026c2STejun Heo * work_struct flags. 11114690c4abSTejun Heo * 11124690c4abSTejun Heo * CONTEXT: 1113d565ed63STejun Heo * spin_lock_irq(pool->lock). 1114365970a1SDavid Howells */ 1115112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1116112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1117b89deed3SOleg Nesterov { 1118112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1119e1d8aa9fSFrederic Weisbecker 11204690c4abSTejun Heo /* we own @work, set data and link */ 1121112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 11221a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1123e22bee78STejun Heo 1124e22bee78STejun Heo /* 1125e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1126e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1127e22bee78STejun Heo * lying around lazily while there are works to be processed. 1128e22bee78STejun Heo */ 1129e22bee78STejun Heo smp_mb(); 1130e22bee78STejun Heo 113163d95a91STejun Heo if (__need_more_worker(pool)) 113263d95a91STejun Heo wake_up_worker(pool); 1133b89deed3SOleg Nesterov } 1134b89deed3SOleg Nesterov 1135c8efcc25STejun Heo /* 1136c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 11378d03ecfeSTejun Heo * same workqueue. 1138c8efcc25STejun Heo */ 1139c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1140c8efcc25STejun Heo { 1141c8efcc25STejun Heo struct worker *worker; 1142c8efcc25STejun Heo 11438d03ecfeSTejun Heo worker = current_wq_worker(); 1144c8efcc25STejun Heo /* 11458d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 11468d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1147c8efcc25STejun Heo */ 1148112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1149c8efcc25STejun Heo } 1150c8efcc25STejun Heo 1151d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 11521da177e4SLinus Torvalds struct work_struct *work) 11531da177e4SLinus Torvalds { 1154112202d9STejun Heo struct pool_workqueue *pwq; 11551e19ffc6STejun Heo struct list_head *worklist; 11568a2e8e5dSTejun Heo unsigned int work_flags; 1157b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 11588930cabaSTejun Heo 11598930cabaSTejun Heo /* 11608930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 11618930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 11628930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 11638930cabaSTejun Heo * happen with IRQ disabled. 11648930cabaSTejun Heo */ 11658930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 11661da177e4SLinus Torvalds 1167dc186ad7SThomas Gleixner debug_work_activate(work); 11681e19ffc6STejun Heo 1169c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 11709c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1171c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1172e41e704bSTejun Heo return; 1173e41e704bSTejun Heo 1174112202d9STejun Heo /* determine the pwq to use */ 1175c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1176c9e7cf27STejun Heo struct worker_pool *last_pool; 1177c7fc77f7STejun Heo 117857469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1179f3421797STejun Heo cpu = raw_smp_processor_id(); 1180f3421797STejun Heo 118118aa9effSTejun Heo /* 1182dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1183dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1184dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1185dbf2576eSTejun Heo * non-reentrancy. 118618aa9effSTejun Heo */ 11877fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1188c9e7cf27STejun Heo last_pool = get_work_pool(work); 1189dbf2576eSTejun Heo 1190112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 119118aa9effSTejun Heo struct worker *worker; 119218aa9effSTejun Heo 1193d565ed63STejun Heo spin_lock(&last_pool->lock); 119418aa9effSTejun Heo 1195c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 119618aa9effSTejun Heo 1197112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 11987fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, last_pool->cpu); 11998594fadeSLai Jiangshan } else { 120018aa9effSTejun Heo /* meh... not running there, queue here */ 1201d565ed63STejun Heo spin_unlock(&last_pool->lock); 1202112202d9STejun Heo spin_lock(&pwq->pool->lock); 120318aa9effSTejun Heo } 12048930cabaSTejun Heo } else { 1205112202d9STejun Heo spin_lock(&pwq->pool->lock); 12068930cabaSTejun Heo } 1207f3421797STejun Heo } else { 12087fb98ea7STejun Heo pwq = first_pwq(wq); 1209112202d9STejun Heo spin_lock(&pwq->pool->lock); 1210502ca9d8STejun Heo } 1211502ca9d8STejun Heo 1212112202d9STejun Heo /* pwq determined, queue */ 1213112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1214502ca9d8STejun Heo 1215f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1216112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1217f5b2552bSDan Carpenter return; 1218f5b2552bSDan Carpenter } 12191e19ffc6STejun Heo 1220112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1221112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 12221e19ffc6STejun Heo 1223112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1224cdadf009STejun Heo trace_workqueue_activate_work(work); 1225112202d9STejun Heo pwq->nr_active++; 1226112202d9STejun Heo worklist = &pwq->pool->worklist; 12278a2e8e5dSTejun Heo } else { 12288a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1229112202d9STejun Heo worklist = &pwq->delayed_works; 12308a2e8e5dSTejun Heo } 12311e19ffc6STejun Heo 1232112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 12331e19ffc6STejun Heo 1234112202d9STejun Heo spin_unlock(&pwq->pool->lock); 12351da177e4SLinus Torvalds } 12361da177e4SLinus Torvalds 12370fcb78c2SRolf Eike Beer /** 1238c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1239c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1240c1a220e7SZhang Rui * @wq: workqueue to use 1241c1a220e7SZhang Rui * @work: work to queue 1242c1a220e7SZhang Rui * 1243d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1244c1a220e7SZhang Rui * 1245c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1246c1a220e7SZhang Rui * can't go away. 1247c1a220e7SZhang Rui */ 1248d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1249d4283e93STejun Heo struct work_struct *work) 1250c1a220e7SZhang Rui { 1251d4283e93STejun Heo bool ret = false; 12528930cabaSTejun Heo unsigned long flags; 12538930cabaSTejun Heo 12548930cabaSTejun Heo local_irq_save(flags); 1255c1a220e7SZhang Rui 125622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 12574690c4abSTejun Heo __queue_work(cpu, wq, work); 1258d4283e93STejun Heo ret = true; 1259c1a220e7SZhang Rui } 12608930cabaSTejun Heo 12618930cabaSTejun Heo local_irq_restore(flags); 1262c1a220e7SZhang Rui return ret; 1263c1a220e7SZhang Rui } 1264c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1265c1a220e7SZhang Rui 12660a13c00eSTejun Heo /** 12670a13c00eSTejun Heo * queue_work - queue work on a workqueue 12680a13c00eSTejun Heo * @wq: workqueue to use 12690a13c00eSTejun Heo * @work: work to queue 12700a13c00eSTejun Heo * 1271d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 12720a13c00eSTejun Heo * 12730a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 12740a13c00eSTejun Heo * it can be processed by another CPU. 12750a13c00eSTejun Heo */ 1276d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 12770a13c00eSTejun Heo { 127857469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 12790a13c00eSTejun Heo } 12800a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 12810a13c00eSTejun Heo 1282d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 12831da177e4SLinus Torvalds { 128452bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 12851da177e4SLinus Torvalds 1286e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 128760c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 12881da177e4SLinus Torvalds } 12891438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 12901da177e4SLinus Torvalds 12917beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 129252bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 12931da177e4SLinus Torvalds { 12947beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 12957beb2edfSTejun Heo struct work_struct *work = &dwork->work; 12961da177e4SLinus Torvalds 12977beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 12987beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1299fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1300fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13017beb2edfSTejun Heo 13028852aac2STejun Heo /* 13038852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13048852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13058852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13068852aac2STejun Heo * on that there's no such delay when @delay is 0. 13078852aac2STejun Heo */ 13088852aac2STejun Heo if (!delay) { 13098852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13108852aac2STejun Heo return; 13118852aac2STejun Heo } 13128852aac2STejun Heo 13137beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13147beb2edfSTejun Heo 131560c057bcSLai Jiangshan dwork->wq = wq; 13161265057fSTejun Heo dwork->cpu = cpu; 13177beb2edfSTejun Heo timer->expires = jiffies + delay; 13187beb2edfSTejun Heo 13197beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13207beb2edfSTejun Heo add_timer_on(timer, cpu); 13217beb2edfSTejun Heo else 13227beb2edfSTejun Heo add_timer(timer); 13237beb2edfSTejun Heo } 13241da177e4SLinus Torvalds 13250fcb78c2SRolf Eike Beer /** 13260fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13270fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13280fcb78c2SRolf Eike Beer * @wq: workqueue to use 1329af9997e4SRandy Dunlap * @dwork: work to queue 13300fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13310fcb78c2SRolf Eike Beer * 1332715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1333715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1334715f1300STejun Heo * execution. 13350fcb78c2SRolf Eike Beer */ 1336d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 133752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13387a6bc1cdSVenkatesh Pallipadi { 133952bad64dSDavid Howells struct work_struct *work = &dwork->work; 1340d4283e93STejun Heo bool ret = false; 13418930cabaSTejun Heo unsigned long flags; 13428930cabaSTejun Heo 13438930cabaSTejun Heo /* read the comment in __queue_work() */ 13448930cabaSTejun Heo local_irq_save(flags); 13457a6bc1cdSVenkatesh Pallipadi 134622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13477beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1348d4283e93STejun Heo ret = true; 13497a6bc1cdSVenkatesh Pallipadi } 13508930cabaSTejun Heo 13518930cabaSTejun Heo local_irq_restore(flags); 13527a6bc1cdSVenkatesh Pallipadi return ret; 13537a6bc1cdSVenkatesh Pallipadi } 1354ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 13551da177e4SLinus Torvalds 1356c8e55f36STejun Heo /** 13570a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 13580a13c00eSTejun Heo * @wq: workqueue to use 13590a13c00eSTejun Heo * @dwork: delayable work to queue 13600a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 13610a13c00eSTejun Heo * 1362715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 13630a13c00eSTejun Heo */ 1364d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 13650a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 13660a13c00eSTejun Heo { 136757469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 13680a13c00eSTejun Heo } 13690a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 13700a13c00eSTejun Heo 13710a13c00eSTejun Heo /** 13728376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 13738376fe22STejun Heo * @cpu: CPU number to execute work on 13748376fe22STejun Heo * @wq: workqueue to use 13758376fe22STejun Heo * @dwork: work to queue 13768376fe22STejun Heo * @delay: number of jiffies to wait before queueing 13778376fe22STejun Heo * 13788376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 13798376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 13808376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 13818376fe22STejun Heo * current state. 13828376fe22STejun Heo * 13838376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 13848376fe22STejun Heo * pending and its timer was modified. 13858376fe22STejun Heo * 1386e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 13878376fe22STejun Heo * See try_to_grab_pending() for details. 13888376fe22STejun Heo */ 13898376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 13908376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 13918376fe22STejun Heo { 13928376fe22STejun Heo unsigned long flags; 13938376fe22STejun Heo int ret; 13948376fe22STejun Heo 13958376fe22STejun Heo do { 13968376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 13978376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 13988376fe22STejun Heo 13998376fe22STejun Heo if (likely(ret >= 0)) { 14008376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14018376fe22STejun Heo local_irq_restore(flags); 14028376fe22STejun Heo } 14038376fe22STejun Heo 14048376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14058376fe22STejun Heo return ret; 14068376fe22STejun Heo } 14078376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14088376fe22STejun Heo 14098376fe22STejun Heo /** 14108376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14118376fe22STejun Heo * @wq: workqueue to use 14128376fe22STejun Heo * @dwork: work to queue 14138376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14148376fe22STejun Heo * 14158376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14168376fe22STejun Heo */ 14178376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14188376fe22STejun Heo unsigned long delay) 14198376fe22STejun Heo { 14208376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14218376fe22STejun Heo } 14228376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14238376fe22STejun Heo 14248376fe22STejun Heo /** 1425c8e55f36STejun Heo * worker_enter_idle - enter idle state 1426c8e55f36STejun Heo * @worker: worker which is entering idle state 1427c8e55f36STejun Heo * 1428c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1429c8e55f36STejun Heo * necessary. 1430c8e55f36STejun Heo * 1431c8e55f36STejun Heo * LOCKING: 1432d565ed63STejun Heo * spin_lock_irq(pool->lock). 1433c8e55f36STejun Heo */ 1434c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14351da177e4SLinus Torvalds { 1436bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1437c8e55f36STejun Heo 14386183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 14396183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 14406183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 14416183c009STejun Heo return; 1442c8e55f36STejun Heo 1443cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1444cb444766STejun Heo worker->flags |= WORKER_IDLE; 1445bd7bdd43STejun Heo pool->nr_idle++; 1446e22bee78STejun Heo worker->last_active = jiffies; 1447c8e55f36STejun Heo 1448c8e55f36STejun Heo /* idle_list is LIFO */ 1449bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1450db7bccf4STejun Heo 145163d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1452628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1453cb444766STejun Heo 1454544ecf31STejun Heo /* 1455706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1456d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1457628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1458628c78e7STejun Heo * unbind is not in progress. 1459544ecf31STejun Heo */ 146024647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1461bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1462e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1463c8e55f36STejun Heo } 1464c8e55f36STejun Heo 1465c8e55f36STejun Heo /** 1466c8e55f36STejun Heo * worker_leave_idle - leave idle state 1467c8e55f36STejun Heo * @worker: worker which is leaving idle state 1468c8e55f36STejun Heo * 1469c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1470c8e55f36STejun Heo * 1471c8e55f36STejun Heo * LOCKING: 1472d565ed63STejun Heo * spin_lock_irq(pool->lock). 1473c8e55f36STejun Heo */ 1474c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1475c8e55f36STejun Heo { 1476bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1477c8e55f36STejun Heo 14786183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 14796183c009STejun Heo return; 1480d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1481bd7bdd43STejun Heo pool->nr_idle--; 1482c8e55f36STejun Heo list_del_init(&worker->entry); 1483c8e55f36STejun Heo } 1484c8e55f36STejun Heo 1485e22bee78STejun Heo /** 1486f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1487f36dc67bSLai Jiangshan * @pool: target worker_pool 1488f36dc67bSLai Jiangshan * 1489f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1490e22bee78STejun Heo * 1491e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1492e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1493e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1494e22bee78STejun Heo * guaranteed to execute on the cpu. 1495e22bee78STejun Heo * 1496f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1497e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1498e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1499e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1500706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1501e22bee78STejun Heo * [dis]associated in the meantime. 1502e22bee78STejun Heo * 1503706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 150424647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1505f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1506f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1507f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1508e22bee78STejun Heo * 1509e22bee78STejun Heo * CONTEXT: 1510d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1511e22bee78STejun Heo * held. 1512e22bee78STejun Heo * 1513e22bee78STejun Heo * RETURNS: 1514706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1515e22bee78STejun Heo * bound), %false if offline. 1516e22bee78STejun Heo */ 1517f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1518d565ed63STejun Heo __acquires(&pool->lock) 1519e22bee78STejun Heo { 1520e22bee78STejun Heo while (true) { 1521e22bee78STejun Heo /* 1522e22bee78STejun Heo * The following call may fail, succeed or succeed 1523e22bee78STejun Heo * without actually migrating the task to the cpu if 1524e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 152524647570STejun Heo * against POOL_DISASSOCIATED. 1526e22bee78STejun Heo */ 152724647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 1528f5faa077SLai Jiangshan set_cpus_allowed_ptr(current, get_cpu_mask(pool->cpu)); 1529e22bee78STejun Heo 1530d565ed63STejun Heo spin_lock_irq(&pool->lock); 153124647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1532e22bee78STejun Heo return false; 1533f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 1534e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1535ec22ca5eSTejun Heo get_cpu_mask(pool->cpu))) 1536e22bee78STejun Heo return true; 1537d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1538e22bee78STejun Heo 15395035b20fSTejun Heo /* 15405035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 15415035b20fSTejun Heo * and retry migration. cond_resched() is required here; 15425035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 15435035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 15445035b20fSTejun Heo */ 1545e22bee78STejun Heo cpu_relax(); 15465035b20fSTejun Heo cond_resched(); 1547e22bee78STejun Heo } 1548e22bee78STejun Heo } 1549e22bee78STejun Heo 1550e22bee78STejun Heo /* 1551ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 15525f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 155325511a47STejun Heo */ 155425511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 155525511a47STejun Heo { 15565f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 1557f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 15585f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 155925511a47STejun Heo 1560ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1561ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1562d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 156325511a47STejun Heo } 156425511a47STejun Heo 156525511a47STejun Heo /* 156625511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1567403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1568403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1569403c821dSTejun Heo * executed twice without intervening cpu down. 1570e22bee78STejun Heo */ 157125511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1572e22bee78STejun Heo { 1573e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1574e22bee78STejun Heo 1575f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 1576eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1577e22bee78STejun Heo 1578d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1579e22bee78STejun Heo } 1580e22bee78STejun Heo 158125511a47STejun Heo /** 158294cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 158394cf58bbSTejun Heo * @pool: pool of interest 158425511a47STejun Heo * 158594cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 158625511a47STejun Heo * is different for idle and busy ones. 158725511a47STejun Heo * 1588ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1589ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1590ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1591ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 159225511a47STejun Heo * 1593ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1594ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1595ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1596ea1abd61SLai Jiangshan * rebind. 159725511a47STejun Heo * 1598ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1599ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1600ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1601ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 160225511a47STejun Heo */ 160394cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 160425511a47STejun Heo { 1605ea1abd61SLai Jiangshan struct worker *worker, *n; 160625511a47STejun Heo int i; 160725511a47STejun Heo 1608b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 1609d565ed63STejun Heo lockdep_assert_held(&pool->lock); 161025511a47STejun Heo 16115f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1612ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1613ea1abd61SLai Jiangshan /* 161494cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 161594cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1616ea1abd61SLai Jiangshan */ 1617ea1abd61SLai Jiangshan list_del_init(&worker->entry); 161825511a47STejun Heo 161925511a47STejun Heo /* 162094cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 162194cf58bbSTejun Heo * idle_worker_rebind(). 162225511a47STejun Heo */ 162325511a47STejun Heo wake_up_process(worker->task); 162425511a47STejun Heo } 162525511a47STejun Heo 1626ea1abd61SLai Jiangshan /* rebind busy workers */ 1627b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) { 162825511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1629e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 163025511a47STejun Heo 163125511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 163225511a47STejun Heo work_data_bits(rebind_work))) 163325511a47STejun Heo continue; 163425511a47STejun Heo 163525511a47STejun Heo debug_work_activate(rebind_work); 163690beca5dSTejun Heo 163790beca5dSTejun Heo /* 163894cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 1639112202d9STejun Heo * and @pwq->pool consistent for sanity. 164090beca5dSTejun Heo */ 1641e34cdddbSTejun Heo if (std_worker_pool_pri(worker->pool)) 1642e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1643e2b6a6d5SJoonsoo Kim else 1644e2b6a6d5SJoonsoo Kim wq = system_wq; 1645ec58815aSTejun Heo 16467fb98ea7STejun Heo insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work, 164725511a47STejun Heo worker->scheduled.next, 164825511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1649ec58815aSTejun Heo } 165025511a47STejun Heo } 165125511a47STejun Heo 1652c34056a3STejun Heo static struct worker *alloc_worker(void) 1653c34056a3STejun Heo { 1654c34056a3STejun Heo struct worker *worker; 1655c34056a3STejun Heo 1656c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1657c8e55f36STejun Heo if (worker) { 1658c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1659affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 166025511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1661e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1662e22bee78STejun Heo worker->flags = WORKER_PREP; 1663c8e55f36STejun Heo } 1664c34056a3STejun Heo return worker; 1665c34056a3STejun Heo } 1666c34056a3STejun Heo 1667c34056a3STejun Heo /** 1668c34056a3STejun Heo * create_worker - create a new workqueue worker 166963d95a91STejun Heo * @pool: pool the new worker will belong to 1670c34056a3STejun Heo * 167163d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1672c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1673c34056a3STejun Heo * destroy_worker(). 1674c34056a3STejun Heo * 1675c34056a3STejun Heo * CONTEXT: 1676c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1677c34056a3STejun Heo * 1678c34056a3STejun Heo * RETURNS: 1679c34056a3STejun Heo * Pointer to the newly created worker. 1680c34056a3STejun Heo */ 1681bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1682c34056a3STejun Heo { 1683e34cdddbSTejun Heo const char *pri = std_worker_pool_pri(pool) ? "H" : ""; 1684c34056a3STejun Heo struct worker *worker = NULL; 1685f3421797STejun Heo int id = -1; 1686c34056a3STejun Heo 1687d565ed63STejun Heo spin_lock_irq(&pool->lock); 1688bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1689d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1690bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1691c34056a3STejun Heo goto fail; 1692d565ed63STejun Heo spin_lock_irq(&pool->lock); 1693c34056a3STejun Heo } 1694d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1695c34056a3STejun Heo 1696c34056a3STejun Heo worker = alloc_worker(); 1697c34056a3STejun Heo if (!worker) 1698c34056a3STejun Heo goto fail; 1699c34056a3STejun Heo 1700bd7bdd43STejun Heo worker->pool = pool; 1701c34056a3STejun Heo worker->id = id; 1702c34056a3STejun Heo 1703ec22ca5eSTejun Heo if (pool->cpu != WORK_CPU_UNBOUND) 170494dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1705ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1706d84ff051STejun Heo "kworker/%d:%d%s", pool->cpu, id, pri); 1707f3421797STejun Heo else 1708f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 17093270476aSTejun Heo "kworker/u:%d%s", id, pri); 1710c34056a3STejun Heo if (IS_ERR(worker->task)) 1711c34056a3STejun Heo goto fail; 1712c34056a3STejun Heo 1713e34cdddbSTejun Heo if (std_worker_pool_pri(pool)) 17143270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 17153270476aSTejun Heo 1716db7bccf4STejun Heo /* 1717bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 171824647570STejun Heo * %POOL_DISASSOCIATED. The caller is responsible for ensuring the 1719bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1720bc2ae0f5STejun Heo * above the flag definition for details. 1721bc2ae0f5STejun Heo * 1722bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1723bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1724db7bccf4STejun Heo */ 172524647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) { 1726ec22ca5eSTejun Heo kthread_bind(worker->task, pool->cpu); 1727bc2ae0f5STejun Heo } else { 1728db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1729f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1730f3421797STejun Heo } 1731c34056a3STejun Heo 1732c34056a3STejun Heo return worker; 1733c34056a3STejun Heo fail: 1734c34056a3STejun Heo if (id >= 0) { 1735d565ed63STejun Heo spin_lock_irq(&pool->lock); 1736bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1737d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1738c34056a3STejun Heo } 1739c34056a3STejun Heo kfree(worker); 1740c34056a3STejun Heo return NULL; 1741c34056a3STejun Heo } 1742c34056a3STejun Heo 1743c34056a3STejun Heo /** 1744c34056a3STejun Heo * start_worker - start a newly created worker 1745c34056a3STejun Heo * @worker: worker to start 1746c34056a3STejun Heo * 1747706026c2STejun Heo * Make the pool aware of @worker and start it. 1748c34056a3STejun Heo * 1749c34056a3STejun Heo * CONTEXT: 1750d565ed63STejun Heo * spin_lock_irq(pool->lock). 1751c34056a3STejun Heo */ 1752c34056a3STejun Heo static void start_worker(struct worker *worker) 1753c34056a3STejun Heo { 1754cb444766STejun Heo worker->flags |= WORKER_STARTED; 1755bd7bdd43STejun Heo worker->pool->nr_workers++; 1756c8e55f36STejun Heo worker_enter_idle(worker); 1757c34056a3STejun Heo wake_up_process(worker->task); 1758c34056a3STejun Heo } 1759c34056a3STejun Heo 1760c34056a3STejun Heo /** 1761c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1762c34056a3STejun Heo * @worker: worker to be destroyed 1763c34056a3STejun Heo * 1764706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1765c8e55f36STejun Heo * 1766c8e55f36STejun Heo * CONTEXT: 1767d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1768c34056a3STejun Heo */ 1769c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1770c34056a3STejun Heo { 1771bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1772c34056a3STejun Heo int id = worker->id; 1773c34056a3STejun Heo 1774c34056a3STejun Heo /* sanity check frenzy */ 17756183c009STejun Heo if (WARN_ON(worker->current_work) || 17766183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 17776183c009STejun Heo return; 1778c34056a3STejun Heo 1779c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1780bd7bdd43STejun Heo pool->nr_workers--; 1781c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1782bd7bdd43STejun Heo pool->nr_idle--; 1783c8e55f36STejun Heo 1784c8e55f36STejun Heo list_del_init(&worker->entry); 1785cb444766STejun Heo worker->flags |= WORKER_DIE; 1786c8e55f36STejun Heo 1787d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1788c8e55f36STejun Heo 1789c34056a3STejun Heo kthread_stop(worker->task); 1790c34056a3STejun Heo kfree(worker); 1791c34056a3STejun Heo 1792d565ed63STejun Heo spin_lock_irq(&pool->lock); 1793bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1794c34056a3STejun Heo } 1795c34056a3STejun Heo 179663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1797e22bee78STejun Heo { 179863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1799e22bee78STejun Heo 1800d565ed63STejun Heo spin_lock_irq(&pool->lock); 1801e22bee78STejun Heo 180263d95a91STejun Heo if (too_many_workers(pool)) { 1803e22bee78STejun Heo struct worker *worker; 1804e22bee78STejun Heo unsigned long expires; 1805e22bee78STejun Heo 1806e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 180763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1808e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1809e22bee78STejun Heo 1810e22bee78STejun Heo if (time_before(jiffies, expires)) 181163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1812e22bee78STejun Heo else { 1813e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 181411ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 181563d95a91STejun Heo wake_up_worker(pool); 1816e22bee78STejun Heo } 1817e22bee78STejun Heo } 1818e22bee78STejun Heo 1819d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1820e22bee78STejun Heo } 1821e22bee78STejun Heo 1822493a1724STejun Heo static void send_mayday(struct work_struct *work) 1823e22bee78STejun Heo { 1824112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1825112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1826493a1724STejun Heo 1827493a1724STejun Heo lockdep_assert_held(&workqueue_lock); 1828e22bee78STejun Heo 1829e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1830493a1724STejun Heo return; 1831e22bee78STejun Heo 1832e22bee78STejun Heo /* mayday mayday mayday */ 1833493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1834493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1835e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1836493a1724STejun Heo } 1837e22bee78STejun Heo } 1838e22bee78STejun Heo 1839706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1840e22bee78STejun Heo { 184163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1842e22bee78STejun Heo struct work_struct *work; 1843e22bee78STejun Heo 1844493a1724STejun Heo spin_lock_irq(&workqueue_lock); /* for wq->maydays */ 1845493a1724STejun Heo spin_lock(&pool->lock); 1846e22bee78STejun Heo 184763d95a91STejun Heo if (need_to_create_worker(pool)) { 1848e22bee78STejun Heo /* 1849e22bee78STejun Heo * We've been trying to create a new worker but 1850e22bee78STejun Heo * haven't been successful. We might be hitting an 1851e22bee78STejun Heo * allocation deadlock. Send distress signals to 1852e22bee78STejun Heo * rescuers. 1853e22bee78STejun Heo */ 185463d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1855e22bee78STejun Heo send_mayday(work); 1856e22bee78STejun Heo } 1857e22bee78STejun Heo 1858493a1724STejun Heo spin_unlock(&pool->lock); 1859493a1724STejun Heo spin_unlock_irq(&workqueue_lock); 1860e22bee78STejun Heo 186163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1862e22bee78STejun Heo } 1863e22bee78STejun Heo 1864e22bee78STejun Heo /** 1865e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 186663d95a91STejun Heo * @pool: pool to create a new worker for 1867e22bee78STejun Heo * 186863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1869e22bee78STejun Heo * have at least one idle worker on return from this function. If 1870e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 187163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1872e22bee78STejun Heo * possible allocation deadlock. 1873e22bee78STejun Heo * 1874e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1875e22bee78STejun Heo * may_start_working() true. 1876e22bee78STejun Heo * 1877e22bee78STejun Heo * LOCKING: 1878d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1879e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1880e22bee78STejun Heo * manager. 1881e22bee78STejun Heo * 1882e22bee78STejun Heo * RETURNS: 1883d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1884e22bee78STejun Heo * otherwise. 1885e22bee78STejun Heo */ 188663d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1887d565ed63STejun Heo __releases(&pool->lock) 1888d565ed63STejun Heo __acquires(&pool->lock) 1889e22bee78STejun Heo { 189063d95a91STejun Heo if (!need_to_create_worker(pool)) 1891e22bee78STejun Heo return false; 1892e22bee78STejun Heo restart: 1893d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18949f9c2364STejun Heo 1895e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 189663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1897e22bee78STejun Heo 1898e22bee78STejun Heo while (true) { 1899e22bee78STejun Heo struct worker *worker; 1900e22bee78STejun Heo 1901bc2ae0f5STejun Heo worker = create_worker(pool); 1902e22bee78STejun Heo if (worker) { 190363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1904d565ed63STejun Heo spin_lock_irq(&pool->lock); 1905e22bee78STejun Heo start_worker(worker); 19066183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19076183c009STejun Heo goto restart; 1908e22bee78STejun Heo return true; 1909e22bee78STejun Heo } 1910e22bee78STejun Heo 191163d95a91STejun Heo if (!need_to_create_worker(pool)) 1912e22bee78STejun Heo break; 1913e22bee78STejun Heo 1914e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1915e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19169f9c2364STejun Heo 191763d95a91STejun Heo if (!need_to_create_worker(pool)) 1918e22bee78STejun Heo break; 1919e22bee78STejun Heo } 1920e22bee78STejun Heo 192163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1922d565ed63STejun Heo spin_lock_irq(&pool->lock); 192363d95a91STejun Heo if (need_to_create_worker(pool)) 1924e22bee78STejun Heo goto restart; 1925e22bee78STejun Heo return true; 1926e22bee78STejun Heo } 1927e22bee78STejun Heo 1928e22bee78STejun Heo /** 1929e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 193063d95a91STejun Heo * @pool: pool to destroy workers for 1931e22bee78STejun Heo * 193263d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1933e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1934e22bee78STejun Heo * 1935e22bee78STejun Heo * LOCKING: 1936d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1937e22bee78STejun Heo * multiple times. Called only from manager. 1938e22bee78STejun Heo * 1939e22bee78STejun Heo * RETURNS: 1940d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1941e22bee78STejun Heo * otherwise. 1942e22bee78STejun Heo */ 194363d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1944e22bee78STejun Heo { 1945e22bee78STejun Heo bool ret = false; 1946e22bee78STejun Heo 194763d95a91STejun Heo while (too_many_workers(pool)) { 1948e22bee78STejun Heo struct worker *worker; 1949e22bee78STejun Heo unsigned long expires; 1950e22bee78STejun Heo 195163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1952e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1953e22bee78STejun Heo 1954e22bee78STejun Heo if (time_before(jiffies, expires)) { 195563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1956e22bee78STejun Heo break; 1957e22bee78STejun Heo } 1958e22bee78STejun Heo 1959e22bee78STejun Heo destroy_worker(worker); 1960e22bee78STejun Heo ret = true; 1961e22bee78STejun Heo } 1962e22bee78STejun Heo 1963e22bee78STejun Heo return ret; 1964e22bee78STejun Heo } 1965e22bee78STejun Heo 1966e22bee78STejun Heo /** 1967e22bee78STejun Heo * manage_workers - manage worker pool 1968e22bee78STejun Heo * @worker: self 1969e22bee78STejun Heo * 1970706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1971e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1972706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1973e22bee78STejun Heo * 1974e22bee78STejun Heo * The caller can safely start processing works on false return. On 1975e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1976e22bee78STejun Heo * and may_start_working() is true. 1977e22bee78STejun Heo * 1978e22bee78STejun Heo * CONTEXT: 1979d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1980e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1981e22bee78STejun Heo * 1982e22bee78STejun Heo * RETURNS: 1983d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1984d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 1985e22bee78STejun Heo */ 1986e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1987e22bee78STejun Heo { 198863d95a91STejun Heo struct worker_pool *pool = worker->pool; 1989e22bee78STejun Heo bool ret = false; 1990e22bee78STejun Heo 1991ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 1992e22bee78STejun Heo return ret; 1993e22bee78STejun Heo 1994552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 1995ee378aa4SLai Jiangshan 1996ee378aa4SLai Jiangshan /* 1997ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 1998ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 1999ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2000ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2001ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2002b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2003ee378aa4SLai Jiangshan * manager against CPU hotplug. 2004ee378aa4SLai Jiangshan * 2005b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2006d565ed63STejun Heo * progress. trylock first without dropping @pool->lock. 2007ee378aa4SLai Jiangshan */ 2008b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2009d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2010b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2011ee378aa4SLai Jiangshan /* 2012ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2013b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2014ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2015706026c2STejun Heo * @pool's state and ours could have deviated. 2016ee378aa4SLai Jiangshan * 2017b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2018ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2019706026c2STejun Heo * on @pool's current state. Try it and adjust 2020ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2021ee378aa4SLai Jiangshan */ 2022f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(pool)) 2023ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2024ee378aa4SLai Jiangshan else 2025ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2026ee378aa4SLai Jiangshan 2027ee378aa4SLai Jiangshan ret = true; 2028ee378aa4SLai Jiangshan } 2029ee378aa4SLai Jiangshan 203011ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2031e22bee78STejun Heo 2032e22bee78STejun Heo /* 2033e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2034e22bee78STejun Heo * on return. 2035e22bee78STejun Heo */ 203663d95a91STejun Heo ret |= maybe_destroy_workers(pool); 203763d95a91STejun Heo ret |= maybe_create_worker(pool); 2038e22bee78STejun Heo 2039552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2040b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2041e22bee78STejun Heo return ret; 2042e22bee78STejun Heo } 2043e22bee78STejun Heo 2044a62428c0STejun Heo /** 2045a62428c0STejun Heo * process_one_work - process single work 2046c34056a3STejun Heo * @worker: self 2047a62428c0STejun Heo * @work: work to process 2048a62428c0STejun Heo * 2049a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2050a62428c0STejun Heo * process a single work including synchronization against and 2051a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2052a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2053a62428c0STejun Heo * call this function to process a work. 2054a62428c0STejun Heo * 2055a62428c0STejun Heo * CONTEXT: 2056d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2057a62428c0STejun Heo */ 2058c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2059d565ed63STejun Heo __releases(&pool->lock) 2060d565ed63STejun Heo __acquires(&pool->lock) 20611da177e4SLinus Torvalds { 2062112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2063bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2064112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 206573f53c4aSTejun Heo int work_color; 20667e11629dSTejun Heo struct worker *collision; 20674e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20684e6045f1SJohannes Berg /* 2069a62428c0STejun Heo * It is permissible to free the struct work_struct from 2070a62428c0STejun Heo * inside the function that is called from it, this we need to 2071a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2072a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2073a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20744e6045f1SJohannes Berg */ 20754d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20764d82a1deSPeter Zijlstra 20774d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20784e6045f1SJohannes Berg #endif 20796fec10a1STejun Heo /* 20806fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20816fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 208224647570STejun Heo * unbound or a disassociated pool. 20836fec10a1STejun Heo */ 20845f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 208524647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2086ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 208725511a47STejun Heo 20887e11629dSTejun Heo /* 20897e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20907e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20917e11629dSTejun Heo * already processing the work. If so, defer the work to the 20927e11629dSTejun Heo * currently executing one. 20937e11629dSTejun Heo */ 2094c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20957e11629dSTejun Heo if (unlikely(collision)) { 20967e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20977e11629dSTejun Heo return; 20987e11629dSTejun Heo } 20991da177e4SLinus Torvalds 21008930cabaSTejun Heo /* claim and dequeue */ 21011da177e4SLinus Torvalds debug_work_deactivate(work); 2102c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2103c34056a3STejun Heo worker->current_work = work; 2104a2c1c57bSTejun Heo worker->current_func = work->func; 2105112202d9STejun Heo worker->current_pwq = pwq; 210673f53c4aSTejun Heo work_color = get_work_color(work); 21077a22ad75STejun Heo 2108a62428c0STejun Heo list_del_init(&work->entry); 2109a62428c0STejun Heo 2110649027d7STejun Heo /* 2111fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2112fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2113fb0e7bebSTejun Heo */ 2114fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2115fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2116fb0e7bebSTejun Heo 2117974271c4STejun Heo /* 2118d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2119974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2120974271c4STejun Heo */ 212163d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 212263d95a91STejun Heo wake_up_worker(pool); 2123974271c4STejun Heo 21248930cabaSTejun Heo /* 21257c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2126d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 212723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 212823657bb1STejun Heo * disabled. 21298930cabaSTejun Heo */ 21307c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21311da177e4SLinus Torvalds 2132d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2133365970a1SDavid Howells 2134112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21353295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2136e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2137a2c1c57bSTejun Heo worker->current_func(work); 2138e36c886aSArjan van de Ven /* 2139e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2140e36c886aSArjan van de Ven * point will only record its address. 2141e36c886aSArjan van de Ven */ 2142e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21433295f0efSIngo Molnar lock_map_release(&lockdep_map); 2144112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21451da177e4SLinus Torvalds 2146d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2147044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2148044c782cSValentin Ilie " last function: %pf\n", 2149a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2150a2c1c57bSTejun Heo worker->current_func); 2151d5abe669SPeter Zijlstra debug_show_held_locks(current); 2152d5abe669SPeter Zijlstra dump_stack(); 2153d5abe669SPeter Zijlstra } 2154d5abe669SPeter Zijlstra 2155d565ed63STejun Heo spin_lock_irq(&pool->lock); 2156a62428c0STejun Heo 2157fb0e7bebSTejun Heo /* clear cpu intensive status */ 2158fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2159fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2160fb0e7bebSTejun Heo 2161a62428c0STejun Heo /* we're done with it, release */ 216242f8570fSSasha Levin hash_del(&worker->hentry); 2163c34056a3STejun Heo worker->current_work = NULL; 2164a2c1c57bSTejun Heo worker->current_func = NULL; 2165112202d9STejun Heo worker->current_pwq = NULL; 2166112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21671da177e4SLinus Torvalds } 21681da177e4SLinus Torvalds 2169affee4b2STejun Heo /** 2170affee4b2STejun Heo * process_scheduled_works - process scheduled works 2171affee4b2STejun Heo * @worker: self 2172affee4b2STejun Heo * 2173affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2174affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2175affee4b2STejun Heo * fetches a work from the top and executes it. 2176affee4b2STejun Heo * 2177affee4b2STejun Heo * CONTEXT: 2178d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2179affee4b2STejun Heo * multiple times. 2180affee4b2STejun Heo */ 2181affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21821da177e4SLinus Torvalds { 2183affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2184affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2185a62428c0STejun Heo struct work_struct, entry); 2186c34056a3STejun Heo process_one_work(worker, work); 2187a62428c0STejun Heo } 21881da177e4SLinus Torvalds } 21891da177e4SLinus Torvalds 21904690c4abSTejun Heo /** 21914690c4abSTejun Heo * worker_thread - the worker thread function 2192c34056a3STejun Heo * @__worker: self 21934690c4abSTejun Heo * 2194706026c2STejun Heo * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools 2195706026c2STejun Heo * of these per each cpu. These workers process all works regardless of 2196e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2197e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2198e22bee78STejun Heo * rescuer_thread(). 21994690c4abSTejun Heo */ 2200c34056a3STejun Heo static int worker_thread(void *__worker) 22011da177e4SLinus Torvalds { 2202c34056a3STejun Heo struct worker *worker = __worker; 2203bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22041da177e4SLinus Torvalds 2205e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2206e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2207c8e55f36STejun Heo woke_up: 2208d565ed63STejun Heo spin_lock_irq(&pool->lock); 2209affee4b2STejun Heo 22105f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22115f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2212d565ed63STejun Heo spin_unlock_irq(&pool->lock); 221325511a47STejun Heo 22145f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 221525511a47STejun Heo if (worker->flags & WORKER_DIE) { 2216e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2217c8e55f36STejun Heo return 0; 2218c8e55f36STejun Heo } 2219c8e55f36STejun Heo 22205f7dabfdSLai Jiangshan /* otherwise, rebind */ 222125511a47STejun Heo idle_worker_rebind(worker); 222225511a47STejun Heo goto woke_up; 222325511a47STejun Heo } 222425511a47STejun Heo 2225c8e55f36STejun Heo worker_leave_idle(worker); 2226db7bccf4STejun Heo recheck: 2227e22bee78STejun Heo /* no more worker necessary? */ 222863d95a91STejun Heo if (!need_more_worker(pool)) 2229e22bee78STejun Heo goto sleep; 2230e22bee78STejun Heo 2231e22bee78STejun Heo /* do we need to manage? */ 223263d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2233e22bee78STejun Heo goto recheck; 2234e22bee78STejun Heo 2235c8e55f36STejun Heo /* 2236c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2237c8e55f36STejun Heo * preparing to process a work or actually processing it. 2238c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2239c8e55f36STejun Heo */ 22406183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2241c8e55f36STejun Heo 2242e22bee78STejun Heo /* 2243e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2244e22bee78STejun Heo * at least one idle worker or that someone else has already 2245e22bee78STejun Heo * assumed the manager role. 2246e22bee78STejun Heo */ 2247e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2248e22bee78STejun Heo 2249e22bee78STejun Heo do { 2250affee4b2STejun Heo struct work_struct *work = 2251bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2252affee4b2STejun Heo struct work_struct, entry); 2253affee4b2STejun Heo 2254c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2255affee4b2STejun Heo /* optimization path, not strictly necessary */ 2256affee4b2STejun Heo process_one_work(worker, work); 2257affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2258affee4b2STejun Heo process_scheduled_works(worker); 2259affee4b2STejun Heo } else { 2260c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2261affee4b2STejun Heo process_scheduled_works(worker); 2262affee4b2STejun Heo } 226363d95a91STejun Heo } while (keep_working(pool)); 2264affee4b2STejun Heo 2265e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2266d313dd85STejun Heo sleep: 226763d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2268e22bee78STejun Heo goto recheck; 2269d313dd85STejun Heo 2270c8e55f36STejun Heo /* 2271d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2272d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2273d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2274d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2275d565ed63STejun Heo * event. 2276c8e55f36STejun Heo */ 2277c8e55f36STejun Heo worker_enter_idle(worker); 2278c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2279d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22801da177e4SLinus Torvalds schedule(); 2281c8e55f36STejun Heo goto woke_up; 22821da177e4SLinus Torvalds } 22831da177e4SLinus Torvalds 2284e22bee78STejun Heo /** 2285e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2286111c225aSTejun Heo * @__rescuer: self 2287e22bee78STejun Heo * 2288e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2289e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2290e22bee78STejun Heo * 2291706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2292e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2293e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2294e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2295e22bee78STejun Heo * the problem rescuer solves. 2296e22bee78STejun Heo * 2297706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2298706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2299e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2300e22bee78STejun Heo * 2301e22bee78STejun Heo * This should happen rarely. 2302e22bee78STejun Heo */ 2303111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2304e22bee78STejun Heo { 2305111c225aSTejun Heo struct worker *rescuer = __rescuer; 2306111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2307e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2308e22bee78STejun Heo 2309e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2310111c225aSTejun Heo 2311111c225aSTejun Heo /* 2312111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2313111c225aSTejun Heo * doesn't participate in concurrency management. 2314111c225aSTejun Heo */ 2315111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2316e22bee78STejun Heo repeat: 2317e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23181da177e4SLinus Torvalds 2319412d32e6SMike Galbraith if (kthread_should_stop()) { 2320412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2321111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2322e22bee78STejun Heo return 0; 2323412d32e6SMike Galbraith } 23241da177e4SLinus Torvalds 2325493a1724STejun Heo /* see whether any pwq is asking for help */ 2326493a1724STejun Heo spin_lock_irq(&workqueue_lock); 2327493a1724STejun Heo 2328493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2329493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2330493a1724STejun Heo struct pool_workqueue, mayday_node); 2331112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2332e22bee78STejun Heo struct work_struct *work, *n; 2333e22bee78STejun Heo 2334e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2335493a1724STejun Heo list_del_init(&pwq->mayday_node); 2336493a1724STejun Heo 2337493a1724STejun Heo spin_unlock_irq(&workqueue_lock); 2338e22bee78STejun Heo 2339e22bee78STejun Heo /* migrate to the target cpu if possible */ 2340f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2341b3104104SLai Jiangshan rescuer->pool = pool; 2342e22bee78STejun Heo 2343e22bee78STejun Heo /* 2344e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2345e22bee78STejun Heo * process'em. 2346e22bee78STejun Heo */ 23476183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2348bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2349112202d9STejun Heo if (get_work_pwq(work) == pwq) 2350e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2351e22bee78STejun Heo 2352e22bee78STejun Heo process_scheduled_works(rescuer); 23537576958aSTejun Heo 23547576958aSTejun Heo /* 2355d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23567576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23577576958aSTejun Heo * and stalling the execution. 23587576958aSTejun Heo */ 235963d95a91STejun Heo if (keep_working(pool)) 236063d95a91STejun Heo wake_up_worker(pool); 23617576958aSTejun Heo 2362b3104104SLai Jiangshan rescuer->pool = NULL; 2363493a1724STejun Heo spin_unlock(&pool->lock); 2364493a1724STejun Heo spin_lock(&workqueue_lock); 23651da177e4SLinus Torvalds } 23661da177e4SLinus Torvalds 2367493a1724STejun Heo spin_unlock_irq(&workqueue_lock); 2368493a1724STejun Heo 2369111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2370111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2371e22bee78STejun Heo schedule(); 2372e22bee78STejun Heo goto repeat; 23731da177e4SLinus Torvalds } 23741da177e4SLinus Torvalds 2375fc2e4d70SOleg Nesterov struct wq_barrier { 2376fc2e4d70SOleg Nesterov struct work_struct work; 2377fc2e4d70SOleg Nesterov struct completion done; 2378fc2e4d70SOleg Nesterov }; 2379fc2e4d70SOleg Nesterov 2380fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2381fc2e4d70SOleg Nesterov { 2382fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2383fc2e4d70SOleg Nesterov complete(&barr->done); 2384fc2e4d70SOleg Nesterov } 2385fc2e4d70SOleg Nesterov 23864690c4abSTejun Heo /** 23874690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2388112202d9STejun Heo * @pwq: pwq to insert barrier into 23894690c4abSTejun Heo * @barr: wq_barrier to insert 2390affee4b2STejun Heo * @target: target work to attach @barr to 2391affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23924690c4abSTejun Heo * 2393affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2394affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2395affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2396affee4b2STejun Heo * cpu. 2397affee4b2STejun Heo * 2398affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2399affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2400affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2401affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2402affee4b2STejun Heo * after a work with LINKED flag set. 2403affee4b2STejun Heo * 2404affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2405112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24064690c4abSTejun Heo * 24074690c4abSTejun Heo * CONTEXT: 2408d565ed63STejun Heo * spin_lock_irq(pool->lock). 24094690c4abSTejun Heo */ 2410112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2411affee4b2STejun Heo struct wq_barrier *barr, 2412affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2413fc2e4d70SOleg Nesterov { 2414affee4b2STejun Heo struct list_head *head; 2415affee4b2STejun Heo unsigned int linked = 0; 2416affee4b2STejun Heo 2417dc186ad7SThomas Gleixner /* 2418d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2419dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2420dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2421dc186ad7SThomas Gleixner * might deadlock. 2422dc186ad7SThomas Gleixner */ 2423ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 242422df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2425fc2e4d70SOleg Nesterov init_completion(&barr->done); 242683c22520SOleg Nesterov 2427affee4b2STejun Heo /* 2428affee4b2STejun Heo * If @target is currently being executed, schedule the 2429affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2430affee4b2STejun Heo */ 2431affee4b2STejun Heo if (worker) 2432affee4b2STejun Heo head = worker->scheduled.next; 2433affee4b2STejun Heo else { 2434affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2435affee4b2STejun Heo 2436affee4b2STejun Heo head = target->entry.next; 2437affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2438affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2439affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2440affee4b2STejun Heo } 2441affee4b2STejun Heo 2442dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2443112202d9STejun Heo insert_work(pwq, &barr->work, head, 2444affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2445fc2e4d70SOleg Nesterov } 2446fc2e4d70SOleg Nesterov 244773f53c4aSTejun Heo /** 2448112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 244973f53c4aSTejun Heo * @wq: workqueue being flushed 245073f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 245173f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 245273f53c4aSTejun Heo * 2453112202d9STejun Heo * Prepare pwqs for workqueue flushing. 245473f53c4aSTejun Heo * 2455112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2456112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2457112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2458112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2459112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 246073f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 246173f53c4aSTejun Heo * 246273f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 246373f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 246473f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 246573f53c4aSTejun Heo * is returned. 246673f53c4aSTejun Heo * 2467112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 246873f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 246973f53c4aSTejun Heo * advanced to @work_color. 247073f53c4aSTejun Heo * 247173f53c4aSTejun Heo * CONTEXT: 247273f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 247373f53c4aSTejun Heo * 247473f53c4aSTejun Heo * RETURNS: 247573f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 247673f53c4aSTejun Heo * otherwise. 247773f53c4aSTejun Heo */ 2478112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 247973f53c4aSTejun Heo int flush_color, int work_color) 24801da177e4SLinus Torvalds { 248173f53c4aSTejun Heo bool wait = false; 248249e3cf44STejun Heo struct pool_workqueue *pwq; 24831da177e4SLinus Torvalds 248473f53c4aSTejun Heo if (flush_color >= 0) { 24856183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2486112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2487dc186ad7SThomas Gleixner } 248814441960SOleg Nesterov 248949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2490112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24911da177e4SLinus Torvalds 2492d565ed63STejun Heo spin_lock_irq(&pool->lock); 249373f53c4aSTejun Heo 249473f53c4aSTejun Heo if (flush_color >= 0) { 24956183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 249673f53c4aSTejun Heo 2497112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2498112202d9STejun Heo pwq->flush_color = flush_color; 2499112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 250073f53c4aSTejun Heo wait = true; 25011da177e4SLinus Torvalds } 250273f53c4aSTejun Heo } 250373f53c4aSTejun Heo 250473f53c4aSTejun Heo if (work_color >= 0) { 25056183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2506112202d9STejun Heo pwq->work_color = work_color; 250773f53c4aSTejun Heo } 250873f53c4aSTejun Heo 2509d565ed63STejun Heo spin_unlock_irq(&pool->lock); 25101da177e4SLinus Torvalds } 25111da177e4SLinus Torvalds 2512112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 251373f53c4aSTejun Heo complete(&wq->first_flusher->done); 251473f53c4aSTejun Heo 251573f53c4aSTejun Heo return wait; 251683c22520SOleg Nesterov } 25171da177e4SLinus Torvalds 25180fcb78c2SRolf Eike Beer /** 25191da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25200fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25211da177e4SLinus Torvalds * 25221da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 25231da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 25241da177e4SLinus Torvalds * 2525fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2526fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 25271da177e4SLinus Torvalds */ 25287ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25291da177e4SLinus Torvalds { 253073f53c4aSTejun Heo struct wq_flusher this_flusher = { 253173f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 253273f53c4aSTejun Heo .flush_color = -1, 253373f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 253473f53c4aSTejun Heo }; 253573f53c4aSTejun Heo int next_color; 2536b1f4ec17SOleg Nesterov 25373295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25383295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 253973f53c4aSTejun Heo 254073f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 254173f53c4aSTejun Heo 254273f53c4aSTejun Heo /* 254373f53c4aSTejun Heo * Start-to-wait phase 254473f53c4aSTejun Heo */ 254573f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 254673f53c4aSTejun Heo 254773f53c4aSTejun Heo if (next_color != wq->flush_color) { 254873f53c4aSTejun Heo /* 254973f53c4aSTejun Heo * Color space is not full. The current work_color 255073f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 255173f53c4aSTejun Heo * by one. 255273f53c4aSTejun Heo */ 25536183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 255473f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 255573f53c4aSTejun Heo wq->work_color = next_color; 255673f53c4aSTejun Heo 255773f53c4aSTejun Heo if (!wq->first_flusher) { 255873f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25596183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 256073f53c4aSTejun Heo 256173f53c4aSTejun Heo wq->first_flusher = &this_flusher; 256273f53c4aSTejun Heo 2563112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 256473f53c4aSTejun Heo wq->work_color)) { 256573f53c4aSTejun Heo /* nothing to flush, done */ 256673f53c4aSTejun Heo wq->flush_color = next_color; 256773f53c4aSTejun Heo wq->first_flusher = NULL; 256873f53c4aSTejun Heo goto out_unlock; 256973f53c4aSTejun Heo } 257073f53c4aSTejun Heo } else { 257173f53c4aSTejun Heo /* wait in queue */ 25726183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 257373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2574112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 257573f53c4aSTejun Heo } 257673f53c4aSTejun Heo } else { 257773f53c4aSTejun Heo /* 257873f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 257973f53c4aSTejun Heo * The next flush completion will assign us 258073f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 258173f53c4aSTejun Heo */ 258273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 258373f53c4aSTejun Heo } 258473f53c4aSTejun Heo 258573f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 258873f53c4aSTejun Heo 258973f53c4aSTejun Heo /* 259073f53c4aSTejun Heo * Wake-up-and-cascade phase 259173f53c4aSTejun Heo * 259273f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 259373f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 259473f53c4aSTejun Heo */ 259573f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 259673f53c4aSTejun Heo return; 259773f53c4aSTejun Heo 259873f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 259973f53c4aSTejun Heo 26004ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26014ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26024ce48b37STejun Heo goto out_unlock; 26034ce48b37STejun Heo 260473f53c4aSTejun Heo wq->first_flusher = NULL; 260573f53c4aSTejun Heo 26066183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26076183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 260873f53c4aSTejun Heo 260973f53c4aSTejun Heo while (true) { 261073f53c4aSTejun Heo struct wq_flusher *next, *tmp; 261173f53c4aSTejun Heo 261273f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 261373f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 261473f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 261573f53c4aSTejun Heo break; 261673f53c4aSTejun Heo list_del_init(&next->list); 261773f53c4aSTejun Heo complete(&next->done); 261873f53c4aSTejun Heo } 261973f53c4aSTejun Heo 26206183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 262173f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 262273f53c4aSTejun Heo 262373f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 262473f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 262573f53c4aSTejun Heo 262673f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 262773f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 262873f53c4aSTejun Heo /* 262973f53c4aSTejun Heo * Assign the same color to all overflowed 263073f53c4aSTejun Heo * flushers, advance work_color and append to 263173f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 263273f53c4aSTejun Heo * phase for these overflowed flushers. 263373f53c4aSTejun Heo */ 263473f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 263573f53c4aSTejun Heo tmp->flush_color = wq->work_color; 263673f53c4aSTejun Heo 263773f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 263873f53c4aSTejun Heo 263973f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 264073f53c4aSTejun Heo &wq->flusher_queue); 2641112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 264273f53c4aSTejun Heo } 264373f53c4aSTejun Heo 264473f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26456183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 264673f53c4aSTejun Heo break; 264773f53c4aSTejun Heo } 264873f53c4aSTejun Heo 264973f53c4aSTejun Heo /* 265073f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2651112202d9STejun Heo * the new first flusher and arm pwqs. 265273f53c4aSTejun Heo */ 26536183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26546183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 265573f53c4aSTejun Heo 265673f53c4aSTejun Heo list_del_init(&next->list); 265773f53c4aSTejun Heo wq->first_flusher = next; 265873f53c4aSTejun Heo 2659112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 266073f53c4aSTejun Heo break; 266173f53c4aSTejun Heo 266273f53c4aSTejun Heo /* 266373f53c4aSTejun Heo * Meh... this color is already done, clear first 266473f53c4aSTejun Heo * flusher and repeat cascading. 266573f53c4aSTejun Heo */ 266673f53c4aSTejun Heo wq->first_flusher = NULL; 266773f53c4aSTejun Heo } 266873f53c4aSTejun Heo 266973f53c4aSTejun Heo out_unlock: 267073f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 26711da177e4SLinus Torvalds } 2672ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26731da177e4SLinus Torvalds 26749c5a2ba7STejun Heo /** 26759c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26769c5a2ba7STejun Heo * @wq: workqueue to drain 26779c5a2ba7STejun Heo * 26789c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26799c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26809c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26819c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26829c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26839c5a2ba7STejun Heo * takes too long. 26849c5a2ba7STejun Heo */ 26859c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26869c5a2ba7STejun Heo { 26879c5a2ba7STejun Heo unsigned int flush_cnt = 0; 268849e3cf44STejun Heo struct pool_workqueue *pwq; 26899c5a2ba7STejun Heo 26909c5a2ba7STejun Heo /* 26919c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26929c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 26939c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 26949c5a2ba7STejun Heo */ 2695e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 26969c5a2ba7STejun Heo if (!wq->nr_drainers++) 26979c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 2698e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 26999c5a2ba7STejun Heo reflush: 27009c5a2ba7STejun Heo flush_workqueue(wq); 27019c5a2ba7STejun Heo 270249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2703fa2563e4SThomas Tuttle bool drained; 27049c5a2ba7STejun Heo 2705112202d9STejun Heo spin_lock_irq(&pwq->pool->lock); 2706112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2707112202d9STejun Heo spin_unlock_irq(&pwq->pool->lock); 2708fa2563e4SThomas Tuttle 2709fa2563e4SThomas Tuttle if (drained) 27109c5a2ba7STejun Heo continue; 27119c5a2ba7STejun Heo 27129c5a2ba7STejun Heo if (++flush_cnt == 10 || 27139c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2714044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 27159c5a2ba7STejun Heo wq->name, flush_cnt); 27169c5a2ba7STejun Heo goto reflush; 27179c5a2ba7STejun Heo } 27189c5a2ba7STejun Heo 2719e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 27209c5a2ba7STejun Heo if (!--wq->nr_drainers) 27219c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 2722e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 27239c5a2ba7STejun Heo } 27249c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27259c5a2ba7STejun Heo 2726606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2727baf59022STejun Heo { 2728baf59022STejun Heo struct worker *worker = NULL; 2729c9e7cf27STejun Heo struct worker_pool *pool; 2730112202d9STejun Heo struct pool_workqueue *pwq; 2731baf59022STejun Heo 2732baf59022STejun Heo might_sleep(); 2733c9e7cf27STejun Heo pool = get_work_pool(work); 2734c9e7cf27STejun Heo if (!pool) 2735baf59022STejun Heo return false; 2736baf59022STejun Heo 2737d565ed63STejun Heo spin_lock_irq(&pool->lock); 27380b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2739112202d9STejun Heo pwq = get_work_pwq(work); 2740112202d9STejun Heo if (pwq) { 2741112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2742baf59022STejun Heo goto already_gone; 2743606a5020STejun Heo } else { 2744c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2745baf59022STejun Heo if (!worker) 2746baf59022STejun Heo goto already_gone; 2747112202d9STejun Heo pwq = worker->current_pwq; 2748606a5020STejun Heo } 2749baf59022STejun Heo 2750112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2751d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2752baf59022STejun Heo 2753e159489bSTejun Heo /* 2754e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2755e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2756e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2757e159489bSTejun Heo * access. 2758e159489bSTejun Heo */ 2759112202d9STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER) 2760112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2761e159489bSTejun Heo else 2762112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2763112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2764e159489bSTejun Heo 2765baf59022STejun Heo return true; 2766baf59022STejun Heo already_gone: 2767d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2768baf59022STejun Heo return false; 2769baf59022STejun Heo } 2770baf59022STejun Heo 2771db700897SOleg Nesterov /** 2772401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2773401a8d04STejun Heo * @work: the work to flush 2774db700897SOleg Nesterov * 2775606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2776606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2777401a8d04STejun Heo * 2778401a8d04STejun Heo * RETURNS: 2779401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2780401a8d04STejun Heo * %false if it was already idle. 2781db700897SOleg Nesterov */ 2782401a8d04STejun Heo bool flush_work(struct work_struct *work) 2783db700897SOleg Nesterov { 2784db700897SOleg Nesterov struct wq_barrier barr; 2785db700897SOleg Nesterov 27860976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27870976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27880976dfc1SStephen Boyd 2789606a5020STejun Heo if (start_flush_work(work, &barr)) { 2790db700897SOleg Nesterov wait_for_completion(&barr.done); 2791dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2792401a8d04STejun Heo return true; 2793606a5020STejun Heo } else { 2794401a8d04STejun Heo return false; 2795db700897SOleg Nesterov } 2796606a5020STejun Heo } 2797db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2798db700897SOleg Nesterov 279936e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2800401a8d04STejun Heo { 2801bbb68dfaSTejun Heo unsigned long flags; 28021f1f642eSOleg Nesterov int ret; 28031f1f642eSOleg Nesterov 28041f1f642eSOleg Nesterov do { 2805bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2806bbb68dfaSTejun Heo /* 2807bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2808bbb68dfaSTejun Heo * would be waiting for before retrying. 2809bbb68dfaSTejun Heo */ 2810bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2811606a5020STejun Heo flush_work(work); 28121f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28131f1f642eSOleg Nesterov 2814bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2815bbb68dfaSTejun Heo mark_work_canceling(work); 2816bbb68dfaSTejun Heo local_irq_restore(flags); 2817bbb68dfaSTejun Heo 2818606a5020STejun Heo flush_work(work); 28197a22ad75STejun Heo clear_work_data(work); 28201f1f642eSOleg Nesterov return ret; 28211f1f642eSOleg Nesterov } 28221f1f642eSOleg Nesterov 28236e84d644SOleg Nesterov /** 2824401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2825401a8d04STejun Heo * @work: the work to cancel 28266e84d644SOleg Nesterov * 2827401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2828401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2829401a8d04STejun Heo * another workqueue. On return from this function, @work is 2830401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28311f1f642eSOleg Nesterov * 2832401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2833401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28346e84d644SOleg Nesterov * 2835401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28366e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2837401a8d04STejun Heo * 2838401a8d04STejun Heo * RETURNS: 2839401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28406e84d644SOleg Nesterov */ 2841401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28426e84d644SOleg Nesterov { 284336e227d2STejun Heo return __cancel_work_timer(work, false); 2844b89deed3SOleg Nesterov } 284528e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2846b89deed3SOleg Nesterov 28476e84d644SOleg Nesterov /** 2848401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2849401a8d04STejun Heo * @dwork: the delayed work to flush 28506e84d644SOleg Nesterov * 2851401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2852401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2853401a8d04STejun Heo * considers the last queueing instance of @dwork. 28541f1f642eSOleg Nesterov * 2855401a8d04STejun Heo * RETURNS: 2856401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2857401a8d04STejun Heo * %false if it was already idle. 28586e84d644SOleg Nesterov */ 2859401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2860401a8d04STejun Heo { 28618930cabaSTejun Heo local_irq_disable(); 2862401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 286360c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28648930cabaSTejun Heo local_irq_enable(); 2865401a8d04STejun Heo return flush_work(&dwork->work); 2866401a8d04STejun Heo } 2867401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2868401a8d04STejun Heo 2869401a8d04STejun Heo /** 287057b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 287157b30ae7STejun Heo * @dwork: delayed_work to cancel 287209383498STejun Heo * 287357b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 287457b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 287557b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 287657b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 287757b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 287809383498STejun Heo * 287957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 288009383498STejun Heo */ 288157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 288209383498STejun Heo { 288357b30ae7STejun Heo unsigned long flags; 288457b30ae7STejun Heo int ret; 288557b30ae7STejun Heo 288657b30ae7STejun Heo do { 288757b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 288857b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 288957b30ae7STejun Heo 289057b30ae7STejun Heo if (unlikely(ret < 0)) 289157b30ae7STejun Heo return false; 289257b30ae7STejun Heo 28937c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28947c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289557b30ae7STejun Heo local_irq_restore(flags); 2896c0158ca6SDan Magenheimer return ret; 289709383498STejun Heo } 289857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 289909383498STejun Heo 290009383498STejun Heo /** 2901401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2902401a8d04STejun Heo * @dwork: the delayed work cancel 2903401a8d04STejun Heo * 2904401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2905401a8d04STejun Heo * 2906401a8d04STejun Heo * RETURNS: 2907401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2908401a8d04STejun Heo */ 2909401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29106e84d644SOleg Nesterov { 291136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29126e84d644SOleg Nesterov } 2913f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29141da177e4SLinus Torvalds 29150fcb78c2SRolf Eike Beer /** 2916c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2917c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2918c1a220e7SZhang Rui * @work: job to be done 2919c1a220e7SZhang Rui * 2920c1a220e7SZhang Rui * This puts a job on a specific cpu 2921c1a220e7SZhang Rui */ 2922d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 2923c1a220e7SZhang Rui { 2924d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2925c1a220e7SZhang Rui } 2926c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2927c1a220e7SZhang Rui 29280fcb78c2SRolf Eike Beer /** 2929ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 29301da177e4SLinus Torvalds * @work: job to be done 29311da177e4SLinus Torvalds * 2932d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 2933d4283e93STejun Heo * %true otherwise. 293452bad64dSDavid Howells * 29350fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 29360fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 29370fcb78c2SRolf Eike Beer * workqueue otherwise. 29381da177e4SLinus Torvalds */ 2939d4283e93STejun Heo bool schedule_work(struct work_struct *work) 29401da177e4SLinus Torvalds { 29410fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 29421da177e4SLinus Torvalds } 29430fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 29441da177e4SLinus Torvalds 29450fcb78c2SRolf Eike Beer /** 29460fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 29470fcb78c2SRolf Eike Beer * @cpu: cpu to use 29480fcb78c2SRolf Eike Beer * @dwork: job to be done 29490fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 29500fcb78c2SRolf Eike Beer * 29510fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 29520fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 29530fcb78c2SRolf Eike Beer */ 2954d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 2955d4283e93STejun Heo unsigned long delay) 29561da177e4SLinus Torvalds { 2957d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 29581da177e4SLinus Torvalds } 2959ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 29601da177e4SLinus Torvalds 2961b6136773SAndrew Morton /** 29620a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 29630a13c00eSTejun Heo * @dwork: job to be done 29640a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 29650a13c00eSTejun Heo * 29660a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 29670a13c00eSTejun Heo * workqueue. 29680a13c00eSTejun Heo */ 2969d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 29700a13c00eSTejun Heo { 29710a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 29720a13c00eSTejun Heo } 29730a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 29740a13c00eSTejun Heo 29750a13c00eSTejun Heo /** 297631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2977b6136773SAndrew Morton * @func: the function to call 2978b6136773SAndrew Morton * 297931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 298031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2981b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 298231ddd871STejun Heo * 298331ddd871STejun Heo * RETURNS: 298431ddd871STejun Heo * 0 on success, -errno on failure. 2985b6136773SAndrew Morton */ 298665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 298715316ba8SChristoph Lameter { 298815316ba8SChristoph Lameter int cpu; 298938f51568SNamhyung Kim struct work_struct __percpu *works; 299015316ba8SChristoph Lameter 2991b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2992b6136773SAndrew Morton if (!works) 299315316ba8SChristoph Lameter return -ENOMEM; 2994b6136773SAndrew Morton 299595402b38SGautham R Shenoy get_online_cpus(); 299693981800STejun Heo 299715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29989bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29999bfb1839SIngo Molnar 30009bfb1839SIngo Molnar INIT_WORK(work, func); 30018de6d308SOleg Nesterov schedule_work_on(cpu, work); 300215316ba8SChristoph Lameter } 300393981800STejun Heo 300493981800STejun Heo for_each_online_cpu(cpu) 30058616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 300693981800STejun Heo 300795402b38SGautham R Shenoy put_online_cpus(); 3008b6136773SAndrew Morton free_percpu(works); 300915316ba8SChristoph Lameter return 0; 301015316ba8SChristoph Lameter } 301115316ba8SChristoph Lameter 3012eef6a7d5SAlan Stern /** 3013eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3014eef6a7d5SAlan Stern * 3015eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3016eef6a7d5SAlan Stern * completion. 3017eef6a7d5SAlan Stern * 3018eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3019eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3020eef6a7d5SAlan Stern * will lead to deadlock: 3021eef6a7d5SAlan Stern * 3022eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3023eef6a7d5SAlan Stern * a lock held by your code or its caller. 3024eef6a7d5SAlan Stern * 3025eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3026eef6a7d5SAlan Stern * 3027eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3028eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3029eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3030eef6a7d5SAlan Stern * 3031eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3032eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3033eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3034eef6a7d5SAlan Stern * cancel_work_sync() instead. 3035eef6a7d5SAlan Stern */ 30361da177e4SLinus Torvalds void flush_scheduled_work(void) 30371da177e4SLinus Torvalds { 3038d320c038STejun Heo flush_workqueue(system_wq); 30391da177e4SLinus Torvalds } 3040ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30411da177e4SLinus Torvalds 30421da177e4SLinus Torvalds /** 30431fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30441fa44ecaSJames Bottomley * @fn: the function to execute 30451fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30461fa44ecaSJames Bottomley * be available when the work executes) 30471fa44ecaSJames Bottomley * 30481fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30491fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30501fa44ecaSJames Bottomley * 30511fa44ecaSJames Bottomley * Returns: 0 - function was executed 30521fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30531fa44ecaSJames Bottomley */ 305465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30551fa44ecaSJames Bottomley { 30561fa44ecaSJames Bottomley if (!in_interrupt()) { 305765f27f38SDavid Howells fn(&ew->work); 30581fa44ecaSJames Bottomley return 0; 30591fa44ecaSJames Bottomley } 30601fa44ecaSJames Bottomley 306165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30621fa44ecaSJames Bottomley schedule_work(&ew->work); 30631fa44ecaSJames Bottomley 30641fa44ecaSJames Bottomley return 1; 30651fa44ecaSJames Bottomley } 30661fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30671fa44ecaSJames Bottomley 30681da177e4SLinus Torvalds int keventd_up(void) 30691da177e4SLinus Torvalds { 3070d320c038STejun Heo return system_wq != NULL; 30711da177e4SLinus Torvalds } 30721da177e4SLinus Torvalds 307330cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 30741da177e4SLinus Torvalds { 307549e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 307630cdf249STejun Heo int cpu; 3077e1d8aa9fSFrederic Weisbecker 307830cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3079420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3080420c0ddbSTejun Heo if (!wq->cpu_pwqs) 308130cdf249STejun Heo return -ENOMEM; 308230cdf249STejun Heo 308330cdf249STejun Heo for_each_possible_cpu(cpu) { 30847fb98ea7STejun Heo struct pool_workqueue *pwq = 30857fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 308630cdf249STejun Heo 308749e3cf44STejun Heo pwq->pool = get_std_worker_pool(cpu, highpri); 308830cdf249STejun Heo list_add_tail(&pwq->pwqs_node, &wq->pwqs); 308930cdf249STejun Heo } 309030cdf249STejun Heo } else { 309130cdf249STejun Heo struct pool_workqueue *pwq; 309230cdf249STejun Heo 309330cdf249STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 309430cdf249STejun Heo if (!pwq) 309530cdf249STejun Heo return -ENOMEM; 309630cdf249STejun Heo 309749e3cf44STejun Heo pwq->pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri); 309830cdf249STejun Heo list_add_tail(&pwq->pwqs_node, &wq->pwqs); 309930cdf249STejun Heo } 310030cdf249STejun Heo 310130cdf249STejun Heo return 0; 31020f900049STejun Heo } 31030f900049STejun Heo 3104112202d9STejun Heo static void free_pwqs(struct workqueue_struct *wq) 310506ba38a9SOleg Nesterov { 3106e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3107420c0ddbSTejun Heo free_percpu(wq->cpu_pwqs); 3108420c0ddbSTejun Heo else if (!list_empty(&wq->pwqs)) 3109420c0ddbSTejun Heo kmem_cache_free(pwq_cache, list_first_entry(&wq->pwqs, 3110420c0ddbSTejun Heo struct pool_workqueue, pwqs_node)); 311106ba38a9SOleg Nesterov } 311206ba38a9SOleg Nesterov 3113f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3114f3421797STejun Heo const char *name) 3115b71ab8c2STejun Heo { 3116f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3117f3421797STejun Heo 3118f3421797STejun Heo if (max_active < 1 || max_active > lim) 3119044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3120f3421797STejun Heo max_active, name, 1, lim); 3121b71ab8c2STejun Heo 3122f3421797STejun Heo return clamp_val(max_active, 1, lim); 3123b71ab8c2STejun Heo } 3124b71ab8c2STejun Heo 3125b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 312697e37d7bSTejun Heo unsigned int flags, 31271e19ffc6STejun Heo int max_active, 3128eb13ba87SJohannes Berg struct lock_class_key *key, 3129b196be89STejun Heo const char *lock_name, ...) 31303af24433SOleg Nesterov { 3131b196be89STejun Heo va_list args, args1; 31323af24433SOleg Nesterov struct workqueue_struct *wq; 313349e3cf44STejun Heo struct pool_workqueue *pwq; 3134b196be89STejun Heo size_t namelen; 3135b196be89STejun Heo 3136b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3137b196be89STejun Heo va_start(args, lock_name); 3138b196be89STejun Heo va_copy(args1, args); 3139b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3140b196be89STejun Heo 3141b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3142b196be89STejun Heo if (!wq) 3143b196be89STejun Heo goto err; 3144b196be89STejun Heo 3145b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3146b196be89STejun Heo va_end(args); 3147b196be89STejun Heo va_end(args1); 31483af24433SOleg Nesterov 3149f3421797STejun Heo /* 31506370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 31516370a6adSTejun Heo * have a rescuer to guarantee forward progress. 31526370a6adSTejun Heo */ 31536370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 31546370a6adSTejun Heo flags |= WQ_RESCUER; 31556370a6adSTejun Heo 3156d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3157b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 31583af24433SOleg Nesterov 3159b196be89STejun Heo /* init wq */ 316097e37d7bSTejun Heo wq->flags = flags; 3161a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 316273f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3163112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 316430cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 316573f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 316673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3167493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 31683af24433SOleg Nesterov 3169eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3170cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 31713af24433SOleg Nesterov 317230cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3173bdbc5dd7STejun Heo goto err; 3174bdbc5dd7STejun Heo 317549e3cf44STejun Heo for_each_pwq(pwq, wq) { 3176112202d9STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3177112202d9STejun Heo pwq->wq = wq; 3178112202d9STejun Heo pwq->flush_color = -1; 3179112202d9STejun Heo pwq->max_active = max_active; 3180112202d9STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3181493a1724STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 3182e22bee78STejun Heo } 31831537663fSTejun Heo 3184e22bee78STejun Heo if (flags & WQ_RESCUER) { 3185e22bee78STejun Heo struct worker *rescuer; 3186e22bee78STejun Heo 3187e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3188e22bee78STejun Heo if (!rescuer) 3189e22bee78STejun Heo goto err; 3190e22bee78STejun Heo 3191111c225aSTejun Heo rescuer->rescue_wq = wq; 3192111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3193b196be89STejun Heo wq->name); 3194e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3195e22bee78STejun Heo goto err; 3196e22bee78STejun Heo 3197e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3198e22bee78STejun Heo wake_up_process(rescuer->task); 31993af24433SOleg Nesterov } 32001537663fSTejun Heo 32013af24433SOleg Nesterov /* 3202a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3203a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3204a0a1a5fdSTejun Heo * workqueue to workqueues list. 32053af24433SOleg Nesterov */ 3206e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3207a0a1a5fdSTejun Heo 320858a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 320949e3cf44STejun Heo for_each_pwq(pwq, wq) 321049e3cf44STejun Heo pwq->max_active = 0; 3211a0a1a5fdSTejun Heo 32123af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3213a0a1a5fdSTejun Heo 3214e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 32153af24433SOleg Nesterov 32163af24433SOleg Nesterov return wq; 32174690c4abSTejun Heo err: 32184690c4abSTejun Heo if (wq) { 3219112202d9STejun Heo free_pwqs(wq); 3220e22bee78STejun Heo kfree(wq->rescuer); 32214690c4abSTejun Heo kfree(wq); 32223af24433SOleg Nesterov } 32234690c4abSTejun Heo return NULL; 32241da177e4SLinus Torvalds } 3225d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 32261da177e4SLinus Torvalds 32273af24433SOleg Nesterov /** 32283af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 32293af24433SOleg Nesterov * @wq: target workqueue 32303af24433SOleg Nesterov * 32313af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 32323af24433SOleg Nesterov */ 32333af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 32343af24433SOleg Nesterov { 323549e3cf44STejun Heo struct pool_workqueue *pwq; 32363af24433SOleg Nesterov 32379c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 32389c5a2ba7STejun Heo drain_workqueue(wq); 3239c8efcc25STejun Heo 32406183c009STejun Heo /* sanity checks */ 324149e3cf44STejun Heo for_each_pwq(pwq, wq) { 32426183c009STejun Heo int i; 32436183c009STejun Heo 32446183c009STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 32456183c009STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) 32466183c009STejun Heo return; 32476183c009STejun Heo if (WARN_ON(pwq->nr_active) || 32486183c009STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) 32496183c009STejun Heo return; 32506183c009STejun Heo } 32516183c009STejun Heo 3252a0a1a5fdSTejun Heo /* 3253a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3254a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3255a0a1a5fdSTejun Heo */ 3256e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 32573af24433SOleg Nesterov list_del(&wq->list); 3258e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 32593af24433SOleg Nesterov 3260e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3261e22bee78STejun Heo kthread_stop(wq->rescuer->task); 32628d9df9f0SXiaotian Feng kfree(wq->rescuer); 3263e22bee78STejun Heo } 3264e22bee78STejun Heo 3265112202d9STejun Heo free_pwqs(wq); 32663af24433SOleg Nesterov kfree(wq); 32673af24433SOleg Nesterov } 32683af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 32693af24433SOleg Nesterov 3270dcd989cbSTejun Heo /** 3271112202d9STejun Heo * pwq_set_max_active - adjust max_active of a pwq 3272112202d9STejun Heo * @pwq: target pool_workqueue 32739f4bd4cdSLai Jiangshan * @max_active: new max_active value. 32749f4bd4cdSLai Jiangshan * 3275112202d9STejun Heo * Set @pwq->max_active to @max_active and activate delayed works if 32769f4bd4cdSLai Jiangshan * increased. 32779f4bd4cdSLai Jiangshan * 32789f4bd4cdSLai Jiangshan * CONTEXT: 3279d565ed63STejun Heo * spin_lock_irq(pool->lock). 32809f4bd4cdSLai Jiangshan */ 3281112202d9STejun Heo static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active) 32829f4bd4cdSLai Jiangshan { 3283112202d9STejun Heo pwq->max_active = max_active; 32849f4bd4cdSLai Jiangshan 3285112202d9STejun Heo while (!list_empty(&pwq->delayed_works) && 3286112202d9STejun Heo pwq->nr_active < pwq->max_active) 3287112202d9STejun Heo pwq_activate_first_delayed(pwq); 32889f4bd4cdSLai Jiangshan } 32899f4bd4cdSLai Jiangshan 32909f4bd4cdSLai Jiangshan /** 3291dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3292dcd989cbSTejun Heo * @wq: target workqueue 3293dcd989cbSTejun Heo * @max_active: new max_active value. 3294dcd989cbSTejun Heo * 3295dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3296dcd989cbSTejun Heo * 3297dcd989cbSTejun Heo * CONTEXT: 3298dcd989cbSTejun Heo * Don't call from IRQ context. 3299dcd989cbSTejun Heo */ 3300dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3301dcd989cbSTejun Heo { 330249e3cf44STejun Heo struct pool_workqueue *pwq; 3303dcd989cbSTejun Heo 3304f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3305dcd989cbSTejun Heo 3306e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3307dcd989cbSTejun Heo 3308dcd989cbSTejun Heo wq->saved_max_active = max_active; 3309dcd989cbSTejun Heo 331049e3cf44STejun Heo for_each_pwq(pwq, wq) { 3311112202d9STejun Heo struct worker_pool *pool = pwq->pool; 3312dcd989cbSTejun Heo 3313e98d5b16STejun Heo spin_lock(&pool->lock); 3314dcd989cbSTejun Heo 331558a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 331635b6bb63STejun Heo !(pool->flags & POOL_FREEZING)) 3317112202d9STejun Heo pwq_set_max_active(pwq, max_active); 3318dcd989cbSTejun Heo 3319e98d5b16STejun Heo spin_unlock(&pool->lock); 3320dcd989cbSTejun Heo } 3321dcd989cbSTejun Heo 3322e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3323dcd989cbSTejun Heo } 3324dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3325dcd989cbSTejun Heo 3326dcd989cbSTejun Heo /** 3327dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3328dcd989cbSTejun Heo * @cpu: CPU in question 3329dcd989cbSTejun Heo * @wq: target workqueue 3330dcd989cbSTejun Heo * 3331dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3332dcd989cbSTejun Heo * no synchronization around this function and the test result is 3333dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3334dcd989cbSTejun Heo * 3335dcd989cbSTejun Heo * RETURNS: 3336dcd989cbSTejun Heo * %true if congested, %false otherwise. 3337dcd989cbSTejun Heo */ 3338d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 3339dcd989cbSTejun Heo { 33407fb98ea7STejun Heo struct pool_workqueue *pwq; 33417fb98ea7STejun Heo 33427fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 33437fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 33447fb98ea7STejun Heo else 33457fb98ea7STejun Heo pwq = first_pwq(wq); 3346dcd989cbSTejun Heo 3347112202d9STejun Heo return !list_empty(&pwq->delayed_works); 3348dcd989cbSTejun Heo } 3349dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3350dcd989cbSTejun Heo 3351dcd989cbSTejun Heo /** 3352dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3353dcd989cbSTejun Heo * @work: the work to be tested 3354dcd989cbSTejun Heo * 3355dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3356dcd989cbSTejun Heo * synchronization around this function and the test result is 3357dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3358dcd989cbSTejun Heo * 3359dcd989cbSTejun Heo * RETURNS: 3360dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3361dcd989cbSTejun Heo */ 3362dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3363dcd989cbSTejun Heo { 3364c9e7cf27STejun Heo struct worker_pool *pool = get_work_pool(work); 3365dcd989cbSTejun Heo unsigned long flags; 3366dcd989cbSTejun Heo unsigned int ret = 0; 3367dcd989cbSTejun Heo 3368dcd989cbSTejun Heo if (work_pending(work)) 3369dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3370038366c5SLai Jiangshan 3371038366c5SLai Jiangshan if (pool) { 3372038366c5SLai Jiangshan spin_lock_irqsave(&pool->lock, flags); 3373c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3374dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3375d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 3376038366c5SLai Jiangshan } 3377dcd989cbSTejun Heo 3378dcd989cbSTejun Heo return ret; 3379dcd989cbSTejun Heo } 3380dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3381dcd989cbSTejun Heo 3382db7bccf4STejun Heo /* 3383db7bccf4STejun Heo * CPU hotplug. 3384db7bccf4STejun Heo * 3385e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3386112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 3387706026c2STejun Heo * pool which make migrating pending and scheduled works very 3388e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 338994cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 3390e22bee78STejun Heo * blocked draining impractical. 3391e22bee78STejun Heo * 339224647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 3393628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3394628c78e7STejun Heo * cpu comes back online. 3395db7bccf4STejun Heo */ 3396db7bccf4STejun Heo 3397706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 3398db7bccf4STejun Heo { 339938db41d9STejun Heo int cpu = smp_processor_id(); 34004ce62e9eSTejun Heo struct worker_pool *pool; 3401db7bccf4STejun Heo struct worker *worker; 3402db7bccf4STejun Heo int i; 3403db7bccf4STejun Heo 340438db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 34056183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 3406db7bccf4STejun Heo 340794cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 340894cf58bbSTejun Heo spin_lock_irq(&pool->lock); 3409e22bee78STejun Heo 3410f2d5a0eeSTejun Heo /* 341194cf58bbSTejun Heo * We've claimed all manager positions. Make all workers 341294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 341394cf58bbSTejun Heo * except for the ones which are still executing works from 341494cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 341594cf58bbSTejun Heo * this, they may become diasporas. 3416f2d5a0eeSTejun Heo */ 34174ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3418403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3419db7bccf4STejun Heo 3420b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) 3421403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3422db7bccf4STejun Heo 342324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 3424f2d5a0eeSTejun Heo 342594cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 342694cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 342794cf58bbSTejun Heo } 3428e22bee78STejun Heo 3429e22bee78STejun Heo /* 3430628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3431628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3432628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3433628c78e7STejun Heo */ 3434628c78e7STejun Heo schedule(); 3435628c78e7STejun Heo 3436628c78e7STejun Heo /* 3437628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3438628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 343938db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 344038db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 344138db41d9STejun Heo * pools which are served by workers tied to the CPU. 3442628c78e7STejun Heo * 3443628c78e7STejun Heo * On return from this function, the current worker would trigger 3444628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3445628c78e7STejun Heo * didn't already. 3446e22bee78STejun Heo */ 344738db41d9STejun Heo for_each_std_worker_pool(pool, cpu) 3448e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 3449db7bccf4STejun Heo } 3450db7bccf4STejun Heo 34518db25e78STejun Heo /* 34528db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 34538db25e78STejun Heo * This will be registered high priority CPU notifier. 34548db25e78STejun Heo */ 34559fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 34561da177e4SLinus Torvalds unsigned long action, 34571da177e4SLinus Torvalds void *hcpu) 34581da177e4SLinus Torvalds { 3459d84ff051STejun Heo int cpu = (unsigned long)hcpu; 34604ce62e9eSTejun Heo struct worker_pool *pool; 34611da177e4SLinus Torvalds 34628db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 34633af24433SOleg Nesterov case CPU_UP_PREPARE: 346438db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 34653ce63377STejun Heo struct worker *worker; 34663ce63377STejun Heo 34673ce63377STejun Heo if (pool->nr_workers) 34683ce63377STejun Heo continue; 34693ce63377STejun Heo 34703ce63377STejun Heo worker = create_worker(pool); 34713ce63377STejun Heo if (!worker) 34723ce63377STejun Heo return NOTIFY_BAD; 34733ce63377STejun Heo 3474d565ed63STejun Heo spin_lock_irq(&pool->lock); 34753ce63377STejun Heo start_worker(worker); 3476d565ed63STejun Heo spin_unlock_irq(&pool->lock); 34773af24433SOleg Nesterov } 34781da177e4SLinus Torvalds break; 34791da177e4SLinus Torvalds 348065758202STejun Heo case CPU_DOWN_FAILED: 348165758202STejun Heo case CPU_ONLINE: 348238db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 348394cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 348494cf58bbSTejun Heo spin_lock_irq(&pool->lock); 348594cf58bbSTejun Heo 348624647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 348794cf58bbSTejun Heo rebind_workers(pool); 348894cf58bbSTejun Heo 348994cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 349094cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 349194cf58bbSTejun Heo } 34928db25e78STejun Heo break; 349365758202STejun Heo } 349465758202STejun Heo return NOTIFY_OK; 349565758202STejun Heo } 349665758202STejun Heo 349765758202STejun Heo /* 349865758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 349965758202STejun Heo * This will be registered as low priority CPU notifier. 350065758202STejun Heo */ 35019fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 350265758202STejun Heo unsigned long action, 350365758202STejun Heo void *hcpu) 350465758202STejun Heo { 3505d84ff051STejun Heo int cpu = (unsigned long)hcpu; 35068db25e78STejun Heo struct work_struct unbind_work; 35078db25e78STejun Heo 350865758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 350965758202STejun Heo case CPU_DOWN_PREPARE: 35108db25e78STejun Heo /* unbinding should happen on the local CPU */ 3511706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 35127635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 35138db25e78STejun Heo flush_work(&unbind_work); 35148db25e78STejun Heo break; 351565758202STejun Heo } 351665758202STejun Heo return NOTIFY_OK; 351765758202STejun Heo } 351865758202STejun Heo 35192d3854a3SRusty Russell #ifdef CONFIG_SMP 35208ccad40dSRusty Russell 35212d3854a3SRusty Russell struct work_for_cpu { 3522ed48ece2STejun Heo struct work_struct work; 35232d3854a3SRusty Russell long (*fn)(void *); 35242d3854a3SRusty Russell void *arg; 35252d3854a3SRusty Russell long ret; 35262d3854a3SRusty Russell }; 35272d3854a3SRusty Russell 3528ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 35292d3854a3SRusty Russell { 3530ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3531ed48ece2STejun Heo 35322d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 35332d3854a3SRusty Russell } 35342d3854a3SRusty Russell 35352d3854a3SRusty Russell /** 35362d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 35372d3854a3SRusty Russell * @cpu: the cpu to run on 35382d3854a3SRusty Russell * @fn: the function to run 35392d3854a3SRusty Russell * @arg: the function arg 35402d3854a3SRusty Russell * 354131ad9081SRusty Russell * This will return the value @fn returns. 354231ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 35436b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 35442d3854a3SRusty Russell */ 3545d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 35462d3854a3SRusty Russell { 3547ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 35482d3854a3SRusty Russell 3549ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3550ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3551ed48ece2STejun Heo flush_work(&wfc.work); 35522d3854a3SRusty Russell return wfc.ret; 35532d3854a3SRusty Russell } 35542d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 35552d3854a3SRusty Russell #endif /* CONFIG_SMP */ 35562d3854a3SRusty Russell 3557a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3558e7577c50SRusty Russell 3559a0a1a5fdSTejun Heo /** 3560a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3561a0a1a5fdSTejun Heo * 356258a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 356358a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 3564706026c2STejun Heo * pool->worklist. 3565a0a1a5fdSTejun Heo * 3566a0a1a5fdSTejun Heo * CONTEXT: 3567d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3568a0a1a5fdSTejun Heo */ 3569a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3570a0a1a5fdSTejun Heo { 357117116969STejun Heo struct worker_pool *pool; 357224b8a847STejun Heo struct workqueue_struct *wq; 357324b8a847STejun Heo struct pool_workqueue *pwq; 357417116969STejun Heo int id; 3575a0a1a5fdSTejun Heo 3576e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3577a0a1a5fdSTejun Heo 35786183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 3579a0a1a5fdSTejun Heo workqueue_freezing = true; 3580a0a1a5fdSTejun Heo 358124b8a847STejun Heo /* set FREEZING */ 358217116969STejun Heo for_each_pool(pool, id) { 3583e98d5b16STejun Heo spin_lock(&pool->lock); 358435b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 358535b6bb63STejun Heo pool->flags |= POOL_FREEZING; 358624b8a847STejun Heo spin_unlock(&pool->lock); 35871da177e4SLinus Torvalds } 35888b03ae3cSTejun Heo 358924b8a847STejun Heo /* suppress further executions by setting max_active to zero */ 359024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 359124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 359224b8a847STejun Heo continue; 359324b8a847STejun Heo 359424b8a847STejun Heo for_each_pwq(pwq, wq) { 359524b8a847STejun Heo spin_lock(&pwq->pool->lock); 359624b8a847STejun Heo pwq->max_active = 0; 359724b8a847STejun Heo spin_unlock(&pwq->pool->lock); 359824b8a847STejun Heo } 3599a1056305STejun Heo } 3600a0a1a5fdSTejun Heo 3601e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3602a0a1a5fdSTejun Heo } 3603a0a1a5fdSTejun Heo 3604a0a1a5fdSTejun Heo /** 360558a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3606a0a1a5fdSTejun Heo * 3607a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3608a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3609a0a1a5fdSTejun Heo * 3610a0a1a5fdSTejun Heo * CONTEXT: 3611a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3612a0a1a5fdSTejun Heo * 3613a0a1a5fdSTejun Heo * RETURNS: 361458a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 361558a69cb4STejun Heo * is complete. 3616a0a1a5fdSTejun Heo */ 3617a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3618a0a1a5fdSTejun Heo { 3619a0a1a5fdSTejun Heo bool busy = false; 362024b8a847STejun Heo struct workqueue_struct *wq; 362124b8a847STejun Heo struct pool_workqueue *pwq; 3622a0a1a5fdSTejun Heo 3623e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3624a0a1a5fdSTejun Heo 36256183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 3626a0a1a5fdSTejun Heo 362724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 362824b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 362924b8a847STejun Heo continue; 3630a0a1a5fdSTejun Heo /* 3631a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3632a0a1a5fdSTejun Heo * to peek without lock. 3633a0a1a5fdSTejun Heo */ 363424b8a847STejun Heo for_each_pwq(pwq, wq) { 36356183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 3636112202d9STejun Heo if (pwq->nr_active) { 3637a0a1a5fdSTejun Heo busy = true; 3638a0a1a5fdSTejun Heo goto out_unlock; 3639a0a1a5fdSTejun Heo } 3640a0a1a5fdSTejun Heo } 3641a0a1a5fdSTejun Heo } 3642a0a1a5fdSTejun Heo out_unlock: 3643e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3644a0a1a5fdSTejun Heo return busy; 3645a0a1a5fdSTejun Heo } 3646a0a1a5fdSTejun Heo 3647a0a1a5fdSTejun Heo /** 3648a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3649a0a1a5fdSTejun Heo * 3650a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 3651706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 3652a0a1a5fdSTejun Heo * 3653a0a1a5fdSTejun Heo * CONTEXT: 3654d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3655a0a1a5fdSTejun Heo */ 3656a0a1a5fdSTejun Heo void thaw_workqueues(void) 3657a0a1a5fdSTejun Heo { 365824b8a847STejun Heo struct workqueue_struct *wq; 365924b8a847STejun Heo struct pool_workqueue *pwq; 366024b8a847STejun Heo struct worker_pool *pool; 366124b8a847STejun Heo int id; 3662a0a1a5fdSTejun Heo 3663e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3664a0a1a5fdSTejun Heo 3665a0a1a5fdSTejun Heo if (!workqueue_freezing) 3666a0a1a5fdSTejun Heo goto out_unlock; 3667a0a1a5fdSTejun Heo 366824b8a847STejun Heo /* clear FREEZING */ 366924b8a847STejun Heo for_each_pool(pool, id) { 3670e98d5b16STejun Heo spin_lock(&pool->lock); 367135b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 367235b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 3673e98d5b16STejun Heo spin_unlock(&pool->lock); 3674d565ed63STejun Heo } 367524b8a847STejun Heo 367624b8a847STejun Heo /* restore max_active and repopulate worklist */ 367724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 367824b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 367924b8a847STejun Heo continue; 368024b8a847STejun Heo 368124b8a847STejun Heo for_each_pwq(pwq, wq) { 368224b8a847STejun Heo spin_lock(&pwq->pool->lock); 368324b8a847STejun Heo pwq_set_max_active(pwq, wq->saved_max_active); 368424b8a847STejun Heo spin_unlock(&pwq->pool->lock); 368524b8a847STejun Heo } 368624b8a847STejun Heo } 368724b8a847STejun Heo 368824b8a847STejun Heo /* kick workers */ 368924b8a847STejun Heo for_each_pool(pool, id) { 369024b8a847STejun Heo spin_lock(&pool->lock); 369124b8a847STejun Heo wake_up_worker(pool); 369224b8a847STejun Heo spin_unlock(&pool->lock); 3693a0a1a5fdSTejun Heo } 3694a0a1a5fdSTejun Heo 3695a0a1a5fdSTejun Heo workqueue_freezing = false; 3696a0a1a5fdSTejun Heo out_unlock: 3697e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3698a0a1a5fdSTejun Heo } 3699a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3700a0a1a5fdSTejun Heo 37016ee0578bSSuresh Siddha static int __init init_workqueues(void) 37021da177e4SLinus Torvalds { 3703d84ff051STejun Heo int cpu; 3704c34056a3STejun Heo 37057c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 37067c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 37076be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 3708b5490077STejun Heo 3709e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 3710e904e6c2STejun Heo 3711e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 3712e904e6c2STejun Heo 371365758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3714a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 37158b03ae3cSTejun Heo 3716706026c2STejun Heo /* initialize CPU pools */ 3717706026c2STejun Heo for_each_wq_cpu(cpu) { 37184ce62e9eSTejun Heo struct worker_pool *pool; 37198b03ae3cSTejun Heo 372038db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3721d565ed63STejun Heo spin_lock_init(&pool->lock); 3722ec22ca5eSTejun Heo pool->cpu = cpu; 372324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 37244ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37254ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3726c9e7cf27STejun Heo hash_init(pool->busy_hash); 3727e22bee78STejun Heo 37284ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 37294ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 37304ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3731e22bee78STejun Heo 3732706026c2STejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 37334ce62e9eSTejun Heo (unsigned long)pool); 37344ce62e9eSTejun Heo 3735b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 37364ce62e9eSTejun Heo ida_init(&pool->worker_ida); 37379daf9e67STejun Heo 37389daf9e67STejun Heo /* alloc pool ID */ 37399daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 37404ce62e9eSTejun Heo } 37418b03ae3cSTejun Heo } 37428b03ae3cSTejun Heo 3743e22bee78STejun Heo /* create the initial worker */ 3744706026c2STejun Heo for_each_online_wq_cpu(cpu) { 37454ce62e9eSTejun Heo struct worker_pool *pool; 3746e22bee78STejun Heo 374738db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 37484ce62e9eSTejun Heo struct worker *worker; 37494ce62e9eSTejun Heo 375024647570STejun Heo if (cpu != WORK_CPU_UNBOUND) 375124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 375224647570STejun Heo 3753bc2ae0f5STejun Heo worker = create_worker(pool); 3754e22bee78STejun Heo BUG_ON(!worker); 3755d565ed63STejun Heo spin_lock_irq(&pool->lock); 3756e22bee78STejun Heo start_worker(worker); 3757d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3758e22bee78STejun Heo } 37594ce62e9eSTejun Heo } 3760e22bee78STejun Heo 3761d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 37621aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3763d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3764f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3765f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 376624d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 376724d51addSTejun Heo WQ_FREEZABLE, 0); 37681aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3769ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 37706ee0578bSSuresh Siddha return 0; 37711da177e4SLinus Torvalds } 37726ee0578bSSuresh Siddha early_initcall(init_workqueues); 3773