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 */ 127ec22ca5eSTejun Heo unsigned 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 */ 1720f900049STejun Heo }; 1731da177e4SLinus Torvalds 1741da177e4SLinus Torvalds /* 17573f53c4aSTejun Heo * Structure used to wait for workqueue flush. 17673f53c4aSTejun Heo */ 17773f53c4aSTejun Heo struct wq_flusher { 17873f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 17973f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 18073f53c4aSTejun Heo struct completion done; /* flush completion */ 18173f53c4aSTejun Heo }; 1821da177e4SLinus Torvalds 18373f53c4aSTejun Heo /* 184f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 185f2e005aaSTejun Heo * used to determine whether there's something to be done. 186f2e005aaSTejun Heo */ 187f2e005aaSTejun Heo #ifdef CONFIG_SMP 188f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 189f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 190f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 191f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 192f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 1939c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 194f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 195f2e005aaSTejun Heo #else 196f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 197f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 198f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 199f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 200f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 201f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 202f2e005aaSTejun Heo #endif 2031da177e4SLinus Torvalds 2041da177e4SLinus Torvalds /* 2051da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2061da177e4SLinus Torvalds * per-CPU workqueues: 2071da177e4SLinus Torvalds */ 2081da177e4SLinus Torvalds struct workqueue_struct { 2099c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 210bdbc5dd7STejun Heo union { 211112202d9STejun Heo struct pool_workqueue __percpu *pcpu; 212112202d9STejun Heo struct pool_workqueue *single; 213bdbc5dd7STejun Heo unsigned long v; 214112202d9STejun Heo } pool_wq; /* I: pwq's */ 2154690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 21673f53c4aSTejun Heo 21773f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 21873f53c4aSTejun Heo int work_color; /* F: current work color */ 21973f53c4aSTejun Heo int flush_color; /* F: current flush color */ 220112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 22173f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 22273f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 22373f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 22473f53c4aSTejun Heo 225f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 226e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 227e22bee78STejun Heo 2289c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 229112202d9STejun Heo int saved_max_active; /* W: saved pwq max_active */ 2304e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2314e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2324e6045f1SJohannes Berg #endif 233b196be89STejun Heo char name[]; /* I: workqueue name */ 2341da177e4SLinus Torvalds }; 2351da177e4SLinus Torvalds 236d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 237d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 238044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2391aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 240044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 241d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 242044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 243f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 244044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 24524d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 246d320c038STejun Heo 24797bd2347STejun Heo #define CREATE_TRACE_POINTS 24897bd2347STejun Heo #include <trace/events/workqueue.h> 24997bd2347STejun Heo 25038db41d9STejun Heo #define for_each_std_worker_pool(pool, cpu) \ 251a60dc39cSTejun Heo for ((pool) = &std_worker_pools(cpu)[0]; \ 252a60dc39cSTejun Heo (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++) 2534ce62e9eSTejun Heo 254b67bfe0dSSasha Levin #define for_each_busy_worker(worker, i, pool) \ 255b67bfe0dSSasha Levin hash_for_each(pool->busy_hash, i, worker, hentry) 256db7bccf4STejun Heo 257706026c2STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 258f3421797STejun Heo unsigned int sw) 259f3421797STejun Heo { 260f3421797STejun Heo if (cpu < nr_cpu_ids) { 261f3421797STejun Heo if (sw & 1) { 262f3421797STejun Heo cpu = cpumask_next(cpu, mask); 263f3421797STejun Heo if (cpu < nr_cpu_ids) 264f3421797STejun Heo return cpu; 265f3421797STejun Heo } 266f3421797STejun Heo if (sw & 2) 267f3421797STejun Heo return WORK_CPU_UNBOUND; 268f3421797STejun Heo } 2696be19588SLai Jiangshan return WORK_CPU_END; 270f3421797STejun Heo } 271f3421797STejun Heo 272112202d9STejun Heo static inline int __next_pwq_cpu(int cpu, const struct cpumask *mask, 273f3421797STejun Heo struct workqueue_struct *wq) 274f3421797STejun Heo { 275706026c2STejun Heo return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 276f3421797STejun Heo } 277f3421797STejun Heo 27809884951STejun Heo /* 27909884951STejun Heo * CPU iterators 28009884951STejun Heo * 281706026c2STejun Heo * An extra cpu number is defined using an invalid cpu number 28209884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 283706026c2STejun Heo * specific CPU. The following iterators are similar to for_each_*_cpu() 284706026c2STejun Heo * iterators but also considers the unbound CPU. 28509884951STejun Heo * 286706026c2STejun Heo * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND 287706026c2STejun Heo * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND 288112202d9STejun Heo * for_each_pwq_cpu() : possible CPUs for bound workqueues, 28909884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 29009884951STejun Heo */ 291706026c2STejun Heo #define for_each_wq_cpu(cpu) \ 292706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \ 2936be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 294706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3)) 295f3421797STejun Heo 296706026c2STejun Heo #define for_each_online_wq_cpu(cpu) \ 297706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \ 2986be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 299706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3)) 300f3421797STejun Heo 301112202d9STejun Heo #define for_each_pwq_cpu(cpu, wq) \ 302112202d9STejun Heo for ((cpu) = __next_pwq_cpu(-1, cpu_possible_mask, (wq)); \ 3036be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 304112202d9STejun Heo (cpu) = __next_pwq_cpu((cpu), cpu_possible_mask, (wq))) 305f3421797STejun Heo 306dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 307dc186ad7SThomas Gleixner 308dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 309dc186ad7SThomas Gleixner 31099777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 31199777288SStanislaw Gruszka { 31299777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 31399777288SStanislaw Gruszka } 31499777288SStanislaw Gruszka 315dc186ad7SThomas Gleixner /* 316dc186ad7SThomas Gleixner * fixup_init is called when: 317dc186ad7SThomas Gleixner * - an active object is initialized 318dc186ad7SThomas Gleixner */ 319dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 320dc186ad7SThomas Gleixner { 321dc186ad7SThomas Gleixner struct work_struct *work = addr; 322dc186ad7SThomas Gleixner 323dc186ad7SThomas Gleixner switch (state) { 324dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 325dc186ad7SThomas Gleixner cancel_work_sync(work); 326dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 327dc186ad7SThomas Gleixner return 1; 328dc186ad7SThomas Gleixner default: 329dc186ad7SThomas Gleixner return 0; 330dc186ad7SThomas Gleixner } 331dc186ad7SThomas Gleixner } 332dc186ad7SThomas Gleixner 333dc186ad7SThomas Gleixner /* 334dc186ad7SThomas Gleixner * fixup_activate is called when: 335dc186ad7SThomas Gleixner * - an active object is activated 336dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 337dc186ad7SThomas Gleixner */ 338dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 339dc186ad7SThomas Gleixner { 340dc186ad7SThomas Gleixner struct work_struct *work = addr; 341dc186ad7SThomas Gleixner 342dc186ad7SThomas Gleixner switch (state) { 343dc186ad7SThomas Gleixner 344dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 345dc186ad7SThomas Gleixner /* 346dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 347dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 348dc186ad7SThomas Gleixner * is tracked in the object tracker. 349dc186ad7SThomas Gleixner */ 35022df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 351dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 352dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 353dc186ad7SThomas Gleixner return 0; 354dc186ad7SThomas Gleixner } 355dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 356dc186ad7SThomas Gleixner return 0; 357dc186ad7SThomas Gleixner 358dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 359dc186ad7SThomas Gleixner WARN_ON(1); 360dc186ad7SThomas Gleixner 361dc186ad7SThomas Gleixner default: 362dc186ad7SThomas Gleixner return 0; 363dc186ad7SThomas Gleixner } 364dc186ad7SThomas Gleixner } 365dc186ad7SThomas Gleixner 366dc186ad7SThomas Gleixner /* 367dc186ad7SThomas Gleixner * fixup_free is called when: 368dc186ad7SThomas Gleixner * - an active object is freed 369dc186ad7SThomas Gleixner */ 370dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 371dc186ad7SThomas Gleixner { 372dc186ad7SThomas Gleixner struct work_struct *work = addr; 373dc186ad7SThomas Gleixner 374dc186ad7SThomas Gleixner switch (state) { 375dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 376dc186ad7SThomas Gleixner cancel_work_sync(work); 377dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 378dc186ad7SThomas Gleixner return 1; 379dc186ad7SThomas Gleixner default: 380dc186ad7SThomas Gleixner return 0; 381dc186ad7SThomas Gleixner } 382dc186ad7SThomas Gleixner } 383dc186ad7SThomas Gleixner 384dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 385dc186ad7SThomas Gleixner .name = "work_struct", 38699777288SStanislaw Gruszka .debug_hint = work_debug_hint, 387dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 388dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 389dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 390dc186ad7SThomas Gleixner }; 391dc186ad7SThomas Gleixner 392dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 393dc186ad7SThomas Gleixner { 394dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 395dc186ad7SThomas Gleixner } 396dc186ad7SThomas Gleixner 397dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 398dc186ad7SThomas Gleixner { 399dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 400dc186ad7SThomas Gleixner } 401dc186ad7SThomas Gleixner 402dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 403dc186ad7SThomas Gleixner { 404dc186ad7SThomas Gleixner if (onstack) 405dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 406dc186ad7SThomas Gleixner else 407dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 408dc186ad7SThomas Gleixner } 409dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 410dc186ad7SThomas Gleixner 411dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 412dc186ad7SThomas Gleixner { 413dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 414dc186ad7SThomas Gleixner } 415dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 416dc186ad7SThomas Gleixner 417dc186ad7SThomas Gleixner #else 418dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 419dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 420dc186ad7SThomas Gleixner #endif 421dc186ad7SThomas Gleixner 42295402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 42395402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4241da177e4SLinus Torvalds static LIST_HEAD(workqueues); 425a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4261da177e4SLinus Torvalds 42714441960SOleg Nesterov /* 428e19e397aSTejun Heo * The CPU and unbound standard worker pools. The unbound ones have 429e19e397aSTejun Heo * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set. 43014441960SOleg Nesterov */ 431e19e397aSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 432a60dc39cSTejun Heo cpu_std_worker_pools); 433a60dc39cSTejun Heo static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS]; 434f3421797STejun Heo 4359daf9e67STejun Heo /* idr of all pools */ 4369daf9e67STejun Heo static DEFINE_MUTEX(worker_pool_idr_mutex); 4379daf9e67STejun Heo static DEFINE_IDR(worker_pool_idr); 4389daf9e67STejun Heo 439c34056a3STejun Heo static int worker_thread(void *__worker); 4401da177e4SLinus Torvalds 441a60dc39cSTejun Heo static struct worker_pool *std_worker_pools(int cpu) 4421da177e4SLinus Torvalds { 443f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 444a60dc39cSTejun Heo return per_cpu(cpu_std_worker_pools, cpu); 445f3421797STejun Heo else 446a60dc39cSTejun Heo return unbound_std_worker_pools; 4471da177e4SLinus Torvalds } 4481da177e4SLinus Torvalds 4494e8f0a60STejun Heo static int std_worker_pool_pri(struct worker_pool *pool) 4504e8f0a60STejun Heo { 451a60dc39cSTejun Heo return pool - std_worker_pools(pool->cpu); 4524e8f0a60STejun Heo } 4534e8f0a60STejun Heo 4549daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4559daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4569daf9e67STejun Heo { 4579daf9e67STejun Heo int ret; 4589daf9e67STejun Heo 4599daf9e67STejun Heo mutex_lock(&worker_pool_idr_mutex); 4609daf9e67STejun Heo idr_pre_get(&worker_pool_idr, GFP_KERNEL); 4619daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 4629daf9e67STejun Heo mutex_unlock(&worker_pool_idr_mutex); 4639daf9e67STejun Heo 4649daf9e67STejun Heo return ret; 4659daf9e67STejun Heo } 4669daf9e67STejun Heo 4677c3eed5cSTejun Heo /* 4687c3eed5cSTejun Heo * Lookup worker_pool by id. The idr currently is built during boot and 4697c3eed5cSTejun Heo * never modified. Don't worry about locking for now. 4707c3eed5cSTejun Heo */ 4717c3eed5cSTejun Heo static struct worker_pool *worker_pool_by_id(int pool_id) 4727c3eed5cSTejun Heo { 4737c3eed5cSTejun Heo return idr_find(&worker_pool_idr, pool_id); 4747c3eed5cSTejun Heo } 4757c3eed5cSTejun Heo 476d565ed63STejun Heo static struct worker_pool *get_std_worker_pool(int cpu, bool highpri) 477d565ed63STejun Heo { 478a60dc39cSTejun Heo struct worker_pool *pools = std_worker_pools(cpu); 479d565ed63STejun Heo 480a60dc39cSTejun Heo return &pools[highpri]; 481d565ed63STejun Heo } 482d565ed63STejun Heo 483112202d9STejun Heo static struct pool_workqueue *get_pwq(unsigned int cpu, 4844690c4abSTejun Heo struct workqueue_struct *wq) 485a848e3b6SOleg Nesterov { 486f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 487e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 488112202d9STejun Heo return per_cpu_ptr(wq->pool_wq.pcpu, cpu); 489f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 490112202d9STejun Heo return wq->pool_wq.single; 491f3421797STejun Heo return NULL; 492f3421797STejun Heo } 493a848e3b6SOleg Nesterov 49473f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 49573f53c4aSTejun Heo { 49673f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 49773f53c4aSTejun Heo } 49873f53c4aSTejun Heo 49973f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 50073f53c4aSTejun Heo { 50173f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 50273f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 50373f53c4aSTejun Heo } 50473f53c4aSTejun Heo 50573f53c4aSTejun Heo static int work_next_color(int color) 50673f53c4aSTejun Heo { 50773f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 508b1f4ec17SOleg Nesterov } 509b1f4ec17SOleg Nesterov 5104594bf15SDavid Howells /* 511112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 512112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5137c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5147a22ad75STejun Heo * 515112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 516112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 517bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 518bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5197a22ad75STejun Heo * 520112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5217c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 522112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5237c3eed5cSTejun Heo * available only while the work item is queued. 524bbb68dfaSTejun Heo * 525bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 526bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 527bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 528bbb68dfaSTejun Heo * try to steal the PENDING bit. 5294594bf15SDavid Howells */ 5307a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5317a22ad75STejun Heo unsigned long flags) 5327a22ad75STejun Heo { 5336183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5347a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5357a22ad75STejun Heo } 5367a22ad75STejun Heo 537112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5384690c4abSTejun Heo unsigned long extra_flags) 539365970a1SDavid Howells { 540112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 541112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 542365970a1SDavid Howells } 543365970a1SDavid Howells 5444468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5454468a00fSLai Jiangshan int pool_id) 5464468a00fSLai Jiangshan { 5474468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5484468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5494468a00fSLai Jiangshan } 5504468a00fSLai Jiangshan 5517c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5527c3eed5cSTejun Heo int pool_id) 5534d707b9fSOleg Nesterov { 55423657bb1STejun Heo /* 55523657bb1STejun Heo * The following wmb is paired with the implied mb in 55623657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 55723657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 55823657bb1STejun Heo * owner. 55923657bb1STejun Heo */ 56023657bb1STejun Heo smp_wmb(); 5617c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5624d707b9fSOleg Nesterov } 5634d707b9fSOleg Nesterov 5647a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 565365970a1SDavid Howells { 5667c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5677c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5687a22ad75STejun Heo } 5697a22ad75STejun Heo 570112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5717a22ad75STejun Heo { 572e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5737a22ad75STejun Heo 574112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 575e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 576e120153dSTejun Heo else 577e120153dSTejun Heo return NULL; 5787a22ad75STejun Heo } 5797a22ad75STejun Heo 5807c3eed5cSTejun Heo /** 5817c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5827c3eed5cSTejun Heo * @work: the work item of interest 5837c3eed5cSTejun Heo * 5847c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 5857c3eed5cSTejun Heo */ 5867c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 5877a22ad75STejun Heo { 588e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5897c3eed5cSTejun Heo struct worker_pool *pool; 5907c3eed5cSTejun Heo int pool_id; 5917a22ad75STejun Heo 592112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 593112202d9STejun Heo return ((struct pool_workqueue *) 5947c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 5957a22ad75STejun Heo 5967c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 5977c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 5987a22ad75STejun Heo return NULL; 5997a22ad75STejun Heo 6007c3eed5cSTejun Heo pool = worker_pool_by_id(pool_id); 6017c3eed5cSTejun Heo WARN_ON_ONCE(!pool); 6027c3eed5cSTejun Heo return pool; 6037c3eed5cSTejun Heo } 6047c3eed5cSTejun Heo 6057c3eed5cSTejun Heo /** 6067c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6077c3eed5cSTejun Heo * @work: the work item of interest 6087c3eed5cSTejun Heo * 6097c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6107c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6117c3eed5cSTejun Heo */ 6127c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6137c3eed5cSTejun Heo { 61454d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6157c3eed5cSTejun Heo 616112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 617112202d9STejun Heo return ((struct pool_workqueue *) 61854d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 61954d5b7d0SLai Jiangshan 62054d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6217c3eed5cSTejun Heo } 6227c3eed5cSTejun Heo 623bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 624bbb68dfaSTejun Heo { 6257c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 626bbb68dfaSTejun Heo 6277c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6287c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 629bbb68dfaSTejun Heo } 630bbb68dfaSTejun Heo 631bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 632bbb68dfaSTejun Heo { 633bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 634bbb68dfaSTejun Heo 635112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 636bbb68dfaSTejun Heo } 637bbb68dfaSTejun Heo 638e22bee78STejun Heo /* 6393270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6403270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 641d565ed63STejun Heo * they're being called with pool->lock held. 642e22bee78STejun Heo */ 643e22bee78STejun Heo 64463d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 645649027d7STejun Heo { 646e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 647649027d7STejun Heo } 648649027d7STejun Heo 649e22bee78STejun Heo /* 650e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 651e22bee78STejun Heo * running workers. 652974271c4STejun Heo * 653974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 654706026c2STejun Heo * function will always return %true for unbound pools as long as the 655974271c4STejun Heo * worklist isn't empty. 656e22bee78STejun Heo */ 65763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 658e22bee78STejun Heo { 65963d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 660e22bee78STejun Heo } 661e22bee78STejun Heo 662e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 66363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 664e22bee78STejun Heo { 66563d95a91STejun Heo return pool->nr_idle; 666e22bee78STejun Heo } 667e22bee78STejun Heo 668e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 66963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 670e22bee78STejun Heo { 671e19e397aSTejun Heo return !list_empty(&pool->worklist) && 672e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 673e22bee78STejun Heo } 674e22bee78STejun Heo 675e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 67663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 677e22bee78STejun Heo { 67863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 679e22bee78STejun Heo } 680e22bee78STejun Heo 681e22bee78STejun Heo /* Do I need to be the manager? */ 68263d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 683e22bee78STejun Heo { 68463d95a91STejun Heo return need_to_create_worker(pool) || 68511ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 686e22bee78STejun Heo } 687e22bee78STejun Heo 688e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 68963d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 690e22bee78STejun Heo { 691552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 69263d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 69363d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 694e22bee78STejun Heo 695ea1abd61SLai Jiangshan /* 696ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 697ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 698ea1abd61SLai Jiangshan */ 699ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 700ea1abd61SLai Jiangshan return false; 701ea1abd61SLai Jiangshan 702e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 703e22bee78STejun Heo } 704e22bee78STejun Heo 705e22bee78STejun Heo /* 706e22bee78STejun Heo * Wake up functions. 707e22bee78STejun Heo */ 708e22bee78STejun Heo 7097e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 71063d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7117e11629dSTejun Heo { 71263d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7137e11629dSTejun Heo return NULL; 7147e11629dSTejun Heo 71563d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7167e11629dSTejun Heo } 7177e11629dSTejun Heo 7187e11629dSTejun Heo /** 7197e11629dSTejun Heo * wake_up_worker - wake up an idle worker 72063d95a91STejun Heo * @pool: worker pool to wake worker from 7217e11629dSTejun Heo * 72263d95a91STejun Heo * Wake up the first idle worker of @pool. 7237e11629dSTejun Heo * 7247e11629dSTejun Heo * CONTEXT: 725d565ed63STejun Heo * spin_lock_irq(pool->lock). 7267e11629dSTejun Heo */ 72763d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7287e11629dSTejun Heo { 72963d95a91STejun Heo struct worker *worker = first_worker(pool); 7307e11629dSTejun Heo 7317e11629dSTejun Heo if (likely(worker)) 7327e11629dSTejun Heo wake_up_process(worker->task); 7337e11629dSTejun Heo } 7347e11629dSTejun Heo 7354690c4abSTejun Heo /** 736e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 737e22bee78STejun Heo * @task: task waking up 738e22bee78STejun Heo * @cpu: CPU @task is waking up to 739e22bee78STejun Heo * 740e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 741e22bee78STejun Heo * being awoken. 742e22bee78STejun Heo * 743e22bee78STejun Heo * CONTEXT: 744e22bee78STejun Heo * spin_lock_irq(rq->lock) 745e22bee78STejun Heo */ 746e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 747e22bee78STejun Heo { 748e22bee78STejun Heo struct worker *worker = kthread_data(task); 749e22bee78STejun Heo 75036576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 751ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 752e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 753e22bee78STejun Heo } 75436576000SJoonsoo Kim } 755e22bee78STejun Heo 756e22bee78STejun Heo /** 757e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 758e22bee78STejun Heo * @task: task going to sleep 759e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 760e22bee78STejun Heo * 761e22bee78STejun Heo * This function is called during schedule() when a busy worker is 762e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 763e22bee78STejun Heo * returning pointer to its task. 764e22bee78STejun Heo * 765e22bee78STejun Heo * CONTEXT: 766e22bee78STejun Heo * spin_lock_irq(rq->lock) 767e22bee78STejun Heo * 768e22bee78STejun Heo * RETURNS: 769e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 770e22bee78STejun Heo */ 771e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 772e22bee78STejun Heo unsigned int cpu) 773e22bee78STejun Heo { 774e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 775111c225aSTejun Heo struct worker_pool *pool; 776e22bee78STejun Heo 777111c225aSTejun Heo /* 778111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 779111c225aSTejun Heo * workers, also reach here, let's not access anything before 780111c225aSTejun Heo * checking NOT_RUNNING. 781111c225aSTejun Heo */ 7822d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 783e22bee78STejun Heo return NULL; 784e22bee78STejun Heo 785111c225aSTejun Heo pool = worker->pool; 786111c225aSTejun Heo 787e22bee78STejun Heo /* this can only happen on the local cpu */ 7886183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 7896183c009STejun Heo return NULL; 790e22bee78STejun Heo 791e22bee78STejun Heo /* 792e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 793e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 794e22bee78STejun Heo * Please read comment there. 795e22bee78STejun Heo * 796628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 797628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 798628c78e7STejun Heo * disabled, which in turn means that none else could be 799d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 800628c78e7STejun Heo * lock is safe. 801e22bee78STejun Heo */ 802e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 803e19e397aSTejun Heo !list_empty(&pool->worklist)) 80463d95a91STejun Heo to_wakeup = first_worker(pool); 805e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 806e22bee78STejun Heo } 807e22bee78STejun Heo 808e22bee78STejun Heo /** 809e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 810cb444766STejun Heo * @worker: self 811d302f017STejun Heo * @flags: flags to set 812d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 813d302f017STejun Heo * 814e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 815e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 816e22bee78STejun Heo * woken up. 817d302f017STejun Heo * 818cb444766STejun Heo * CONTEXT: 819d565ed63STejun Heo * spin_lock_irq(pool->lock) 820d302f017STejun Heo */ 821d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 822d302f017STejun Heo bool wakeup) 823d302f017STejun Heo { 824bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 825e22bee78STejun Heo 826cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 827cb444766STejun Heo 828e22bee78STejun Heo /* 829e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 830e22bee78STejun Heo * wake up an idle worker as necessary if requested by 831e22bee78STejun Heo * @wakeup. 832e22bee78STejun Heo */ 833e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 834e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 835e22bee78STejun Heo if (wakeup) { 836e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 837bd7bdd43STejun Heo !list_empty(&pool->worklist)) 83863d95a91STejun Heo wake_up_worker(pool); 839e22bee78STejun Heo } else 840e19e397aSTejun Heo atomic_dec(&pool->nr_running); 841e22bee78STejun Heo } 842e22bee78STejun Heo 843d302f017STejun Heo worker->flags |= flags; 844d302f017STejun Heo } 845d302f017STejun Heo 846d302f017STejun Heo /** 847e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 848cb444766STejun Heo * @worker: self 849d302f017STejun Heo * @flags: flags to clear 850d302f017STejun Heo * 851e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 852d302f017STejun Heo * 853cb444766STejun Heo * CONTEXT: 854d565ed63STejun Heo * spin_lock_irq(pool->lock) 855d302f017STejun Heo */ 856d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 857d302f017STejun Heo { 85863d95a91STejun Heo struct worker_pool *pool = worker->pool; 859e22bee78STejun Heo unsigned int oflags = worker->flags; 860e22bee78STejun Heo 861cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 862cb444766STejun Heo 863d302f017STejun Heo worker->flags &= ~flags; 864e22bee78STejun Heo 86542c025f3STejun Heo /* 86642c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 86742c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 86842c025f3STejun Heo * of multiple flags, not a single flag. 86942c025f3STejun Heo */ 870e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 871e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 872e19e397aSTejun Heo atomic_inc(&pool->nr_running); 873d302f017STejun Heo } 874d302f017STejun Heo 875d302f017STejun Heo /** 8768cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 877c9e7cf27STejun Heo * @pool: pool of interest 8788cca0eeaSTejun Heo * @work: work to find worker for 8798cca0eeaSTejun Heo * 880c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 881c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 882a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 883a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 884a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 885a2c1c57bSTejun Heo * being executed. 886a2c1c57bSTejun Heo * 887a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 888a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 889a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 890a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 891a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 892a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 893a2c1c57bSTejun Heo * 894a2c1c57bSTejun Heo * This function checks the work item address, work function and workqueue 895a2c1c57bSTejun Heo * to avoid false positives. Note that this isn't complete as one may 896a2c1c57bSTejun Heo * construct a work function which can introduce dependency onto itself 897a2c1c57bSTejun Heo * through a recycled work item. Well, if somebody wants to shoot oneself 898a2c1c57bSTejun Heo * in the foot that badly, there's only so much we can do, and if such 899a2c1c57bSTejun Heo * deadlock actually occurs, it should be easy to locate the culprit work 900a2c1c57bSTejun Heo * function. 9018cca0eeaSTejun Heo * 9028cca0eeaSTejun Heo * CONTEXT: 903d565ed63STejun Heo * spin_lock_irq(pool->lock). 9048cca0eeaSTejun Heo * 9058cca0eeaSTejun Heo * RETURNS: 9068cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9078cca0eeaSTejun Heo * otherwise. 9088cca0eeaSTejun Heo */ 909c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9108cca0eeaSTejun Heo struct work_struct *work) 9118cca0eeaSTejun Heo { 91242f8570fSSasha Levin struct worker *worker; 91342f8570fSSasha Levin 914b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 915a2c1c57bSTejun Heo (unsigned long)work) 916a2c1c57bSTejun Heo if (worker->current_work == work && 917a2c1c57bSTejun Heo worker->current_func == work->func) 91842f8570fSSasha Levin return worker; 91942f8570fSSasha Levin 92042f8570fSSasha Levin return NULL; 9218cca0eeaSTejun Heo } 9228cca0eeaSTejun Heo 9238cca0eeaSTejun Heo /** 924bf4ede01STejun Heo * move_linked_works - move linked works to a list 925bf4ede01STejun Heo * @work: start of series of works to be scheduled 926bf4ede01STejun Heo * @head: target list to append @work to 927bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 928bf4ede01STejun Heo * 929bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 930bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 931bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 932bf4ede01STejun Heo * 933bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 934bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 935bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 936bf4ede01STejun Heo * 937bf4ede01STejun Heo * CONTEXT: 938d565ed63STejun Heo * spin_lock_irq(pool->lock). 939bf4ede01STejun Heo */ 940bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 941bf4ede01STejun Heo struct work_struct **nextp) 942bf4ede01STejun Heo { 943bf4ede01STejun Heo struct work_struct *n; 944bf4ede01STejun Heo 945bf4ede01STejun Heo /* 946bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 947bf4ede01STejun Heo * use NULL for list head. 948bf4ede01STejun Heo */ 949bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 950bf4ede01STejun Heo list_move_tail(&work->entry, head); 951bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 952bf4ede01STejun Heo break; 953bf4ede01STejun Heo } 954bf4ede01STejun Heo 955bf4ede01STejun Heo /* 956bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 957bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 958bf4ede01STejun Heo * needs to be updated. 959bf4ede01STejun Heo */ 960bf4ede01STejun Heo if (nextp) 961bf4ede01STejun Heo *nextp = n; 962bf4ede01STejun Heo } 963bf4ede01STejun Heo 964112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 965bf4ede01STejun Heo { 966112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 967bf4ede01STejun Heo 968bf4ede01STejun Heo trace_workqueue_activate_work(work); 969112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 970bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 971112202d9STejun Heo pwq->nr_active++; 972bf4ede01STejun Heo } 973bf4ede01STejun Heo 974112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 9753aa62497SLai Jiangshan { 976112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 9773aa62497SLai Jiangshan struct work_struct, entry); 9783aa62497SLai Jiangshan 979112202d9STejun Heo pwq_activate_delayed_work(work); 9803aa62497SLai Jiangshan } 9813aa62497SLai Jiangshan 982bf4ede01STejun Heo /** 983112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 984112202d9STejun Heo * @pwq: pwq of interest 985bf4ede01STejun Heo * @color: color of work which left the queue 986bf4ede01STejun Heo * 987bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 988112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 989bf4ede01STejun Heo * 990bf4ede01STejun Heo * CONTEXT: 991d565ed63STejun Heo * spin_lock_irq(pool->lock). 992bf4ede01STejun Heo */ 993112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 994bf4ede01STejun Heo { 995bf4ede01STejun Heo /* ignore uncolored works */ 996bf4ede01STejun Heo if (color == WORK_NO_COLOR) 997bf4ede01STejun Heo return; 998bf4ede01STejun Heo 999112202d9STejun Heo pwq->nr_in_flight[color]--; 1000bf4ede01STejun Heo 1001112202d9STejun Heo pwq->nr_active--; 1002112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1003bf4ede01STejun Heo /* one down, submit a delayed one */ 1004112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1005112202d9STejun Heo pwq_activate_first_delayed(pwq); 1006bf4ede01STejun Heo } 1007bf4ede01STejun Heo 1008bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1009112202d9STejun Heo if (likely(pwq->flush_color != color)) 1010bf4ede01STejun Heo return; 1011bf4ede01STejun Heo 1012bf4ede01STejun Heo /* are there still in-flight works? */ 1013112202d9STejun Heo if (pwq->nr_in_flight[color]) 1014bf4ede01STejun Heo return; 1015bf4ede01STejun Heo 1016112202d9STejun Heo /* this pwq is done, clear flush_color */ 1017112202d9STejun Heo pwq->flush_color = -1; 1018bf4ede01STejun Heo 1019bf4ede01STejun Heo /* 1020112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1021bf4ede01STejun Heo * will handle the rest. 1022bf4ede01STejun Heo */ 1023112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1024112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 1025bf4ede01STejun Heo } 1026bf4ede01STejun Heo 102736e227d2STejun Heo /** 1028bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 102936e227d2STejun Heo * @work: work item to steal 103036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1031bbb68dfaSTejun Heo * @flags: place to store irq state 103236e227d2STejun Heo * 103336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 103436e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 103536e227d2STejun Heo * 103636e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 103736e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 103836e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1039bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1040bbb68dfaSTejun Heo * for arbitrarily long 104136e227d2STejun Heo * 1042bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1043e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1044e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1045e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1046bbb68dfaSTejun Heo * 1047bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1048bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1049bbb68dfaSTejun Heo * 1050e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1051bf4ede01STejun Heo */ 1052bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1053bbb68dfaSTejun Heo unsigned long *flags) 1054bf4ede01STejun Heo { 1055d565ed63STejun Heo struct worker_pool *pool; 1056112202d9STejun Heo struct pool_workqueue *pwq; 1057bf4ede01STejun Heo 1058bbb68dfaSTejun Heo local_irq_save(*flags); 1059bbb68dfaSTejun Heo 106036e227d2STejun Heo /* try to steal the timer if it exists */ 106136e227d2STejun Heo if (is_dwork) { 106236e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 106336e227d2STejun Heo 1064e0aecdd8STejun Heo /* 1065e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1066e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1067e0aecdd8STejun Heo * running on the local CPU. 1068e0aecdd8STejun Heo */ 106936e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 107036e227d2STejun Heo return 1; 107136e227d2STejun Heo } 107236e227d2STejun Heo 107336e227d2STejun Heo /* try to claim PENDING the normal way */ 1074bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1075bf4ede01STejun Heo return 0; 1076bf4ede01STejun Heo 1077bf4ede01STejun Heo /* 1078bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1079bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1080bf4ede01STejun Heo */ 1081d565ed63STejun Heo pool = get_work_pool(work); 1082d565ed63STejun Heo if (!pool) 1083bbb68dfaSTejun Heo goto fail; 1084bf4ede01STejun Heo 1085d565ed63STejun Heo spin_lock(&pool->lock); 1086bf4ede01STejun Heo /* 1087112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1088112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1089112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1090112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1091112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 10920b3dae68SLai Jiangshan * item is currently queued on that pool. 1093bf4ede01STejun Heo */ 1094112202d9STejun Heo pwq = get_work_pwq(work); 1095112202d9STejun Heo if (pwq && pwq->pool == pool) { 1096bf4ede01STejun Heo debug_work_deactivate(work); 10973aa62497SLai Jiangshan 10983aa62497SLai Jiangshan /* 109916062836STejun Heo * A delayed work item cannot be grabbed directly because 110016062836STejun Heo * it might have linked NO_COLOR work items which, if left 1101112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 110216062836STejun Heo * management later on and cause stall. Make sure the work 110316062836STejun Heo * item is activated before grabbing. 11043aa62497SLai Jiangshan */ 11053aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1106112202d9STejun Heo pwq_activate_delayed_work(work); 11073aa62497SLai Jiangshan 1108bf4ede01STejun Heo list_del_init(&work->entry); 1109112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 111036e227d2STejun Heo 1111112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11124468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11134468a00fSLai Jiangshan 1114d565ed63STejun Heo spin_unlock(&pool->lock); 111536e227d2STejun Heo return 1; 1116bf4ede01STejun Heo } 1117d565ed63STejun Heo spin_unlock(&pool->lock); 1118bbb68dfaSTejun Heo fail: 1119bbb68dfaSTejun Heo local_irq_restore(*flags); 1120bbb68dfaSTejun Heo if (work_is_canceling(work)) 1121bbb68dfaSTejun Heo return -ENOENT; 1122bbb68dfaSTejun Heo cpu_relax(); 112336e227d2STejun Heo return -EAGAIN; 1124bf4ede01STejun Heo } 1125bf4ede01STejun Heo 1126bf4ede01STejun Heo /** 1127706026c2STejun Heo * insert_work - insert a work into a pool 1128112202d9STejun Heo * @pwq: pwq @work belongs to 11294690c4abSTejun Heo * @work: work to insert 11304690c4abSTejun Heo * @head: insertion point 11314690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11324690c4abSTejun Heo * 1133112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1134706026c2STejun Heo * work_struct flags. 11354690c4abSTejun Heo * 11364690c4abSTejun Heo * CONTEXT: 1137d565ed63STejun Heo * spin_lock_irq(pool->lock). 1138365970a1SDavid Howells */ 1139112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1140112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1141b89deed3SOleg Nesterov { 1142112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1143e1d8aa9fSFrederic Weisbecker 11444690c4abSTejun Heo /* we own @work, set data and link */ 1145112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 11461a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1147e22bee78STejun Heo 1148e22bee78STejun Heo /* 1149e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1150e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1151e22bee78STejun Heo * lying around lazily while there are works to be processed. 1152e22bee78STejun Heo */ 1153e22bee78STejun Heo smp_mb(); 1154e22bee78STejun Heo 115563d95a91STejun Heo if (__need_more_worker(pool)) 115663d95a91STejun Heo wake_up_worker(pool); 1157b89deed3SOleg Nesterov } 1158b89deed3SOleg Nesterov 1159c8efcc25STejun Heo /* 1160c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 11618d03ecfeSTejun Heo * same workqueue. 1162c8efcc25STejun Heo */ 1163c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1164c8efcc25STejun Heo { 1165c8efcc25STejun Heo struct worker *worker; 1166c8efcc25STejun Heo 11678d03ecfeSTejun Heo worker = current_wq_worker(); 1168c8efcc25STejun Heo /* 11698d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 11708d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1171c8efcc25STejun Heo */ 1172112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1173c8efcc25STejun Heo } 1174c8efcc25STejun Heo 11754690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 11761da177e4SLinus Torvalds struct work_struct *work) 11771da177e4SLinus Torvalds { 1178112202d9STejun Heo struct pool_workqueue *pwq; 11791e19ffc6STejun Heo struct list_head *worklist; 11808a2e8e5dSTejun Heo unsigned int work_flags; 1181b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 11828930cabaSTejun Heo 11838930cabaSTejun Heo /* 11848930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 11858930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 11868930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 11878930cabaSTejun Heo * happen with IRQ disabled. 11888930cabaSTejun Heo */ 11898930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 11901da177e4SLinus Torvalds 1191dc186ad7SThomas Gleixner debug_work_activate(work); 11921e19ffc6STejun Heo 1193c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 11949c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1195c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1196e41e704bSTejun Heo return; 1197e41e704bSTejun Heo 1198112202d9STejun Heo /* determine the pwq to use */ 1199c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1200c9e7cf27STejun Heo struct worker_pool *last_pool; 1201c7fc77f7STejun Heo 120257469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1203f3421797STejun Heo cpu = raw_smp_processor_id(); 1204f3421797STejun Heo 120518aa9effSTejun Heo /* 1206dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1207dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1208dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1209dbf2576eSTejun Heo * non-reentrancy. 121018aa9effSTejun Heo */ 1211112202d9STejun Heo pwq = get_pwq(cpu, wq); 1212c9e7cf27STejun Heo last_pool = get_work_pool(work); 1213dbf2576eSTejun Heo 1214112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 121518aa9effSTejun Heo struct worker *worker; 121618aa9effSTejun Heo 1217d565ed63STejun Heo spin_lock(&last_pool->lock); 121818aa9effSTejun Heo 1219c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 122018aa9effSTejun Heo 1221112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1222112202d9STejun Heo pwq = get_pwq(last_pool->cpu, wq); 12238594fadeSLai Jiangshan } else { 122418aa9effSTejun Heo /* meh... not running there, queue here */ 1225d565ed63STejun Heo spin_unlock(&last_pool->lock); 1226112202d9STejun Heo spin_lock(&pwq->pool->lock); 122718aa9effSTejun Heo } 12288930cabaSTejun Heo } else { 1229112202d9STejun Heo spin_lock(&pwq->pool->lock); 12308930cabaSTejun Heo } 1231f3421797STejun Heo } else { 1232112202d9STejun Heo pwq = get_pwq(WORK_CPU_UNBOUND, wq); 1233112202d9STejun Heo spin_lock(&pwq->pool->lock); 1234502ca9d8STejun Heo } 1235502ca9d8STejun Heo 1236112202d9STejun Heo /* pwq determined, queue */ 1237112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1238502ca9d8STejun Heo 1239f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1240112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1241f5b2552bSDan Carpenter return; 1242f5b2552bSDan Carpenter } 12431e19ffc6STejun Heo 1244112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1245112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 12461e19ffc6STejun Heo 1247112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1248cdadf009STejun Heo trace_workqueue_activate_work(work); 1249112202d9STejun Heo pwq->nr_active++; 1250112202d9STejun Heo worklist = &pwq->pool->worklist; 12518a2e8e5dSTejun Heo } else { 12528a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1253112202d9STejun Heo worklist = &pwq->delayed_works; 12548a2e8e5dSTejun Heo } 12551e19ffc6STejun Heo 1256112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 12571e19ffc6STejun Heo 1258112202d9STejun Heo spin_unlock(&pwq->pool->lock); 12591da177e4SLinus Torvalds } 12601da177e4SLinus Torvalds 12610fcb78c2SRolf Eike Beer /** 1262c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1263c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1264c1a220e7SZhang Rui * @wq: workqueue to use 1265c1a220e7SZhang Rui * @work: work to queue 1266c1a220e7SZhang Rui * 1267d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1268c1a220e7SZhang Rui * 1269c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1270c1a220e7SZhang Rui * can't go away. 1271c1a220e7SZhang Rui */ 1272d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1273d4283e93STejun Heo struct work_struct *work) 1274c1a220e7SZhang Rui { 1275d4283e93STejun Heo bool ret = false; 12768930cabaSTejun Heo unsigned long flags; 12778930cabaSTejun Heo 12788930cabaSTejun Heo local_irq_save(flags); 1279c1a220e7SZhang Rui 128022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 12814690c4abSTejun Heo __queue_work(cpu, wq, work); 1282d4283e93STejun Heo ret = true; 1283c1a220e7SZhang Rui } 12848930cabaSTejun Heo 12858930cabaSTejun Heo local_irq_restore(flags); 1286c1a220e7SZhang Rui return ret; 1287c1a220e7SZhang Rui } 1288c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1289c1a220e7SZhang Rui 12900a13c00eSTejun Heo /** 12910a13c00eSTejun Heo * queue_work - queue work on a workqueue 12920a13c00eSTejun Heo * @wq: workqueue to use 12930a13c00eSTejun Heo * @work: work to queue 12940a13c00eSTejun Heo * 1295d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 12960a13c00eSTejun Heo * 12970a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 12980a13c00eSTejun Heo * it can be processed by another CPU. 12990a13c00eSTejun Heo */ 1300d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13010a13c00eSTejun Heo { 130257469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13030a13c00eSTejun Heo } 13040a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13050a13c00eSTejun Heo 1306d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13071da177e4SLinus Torvalds { 130852bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13091da177e4SLinus Torvalds 1310e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 131160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13121da177e4SLinus Torvalds } 13131438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13141da177e4SLinus Torvalds 13157beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 131652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13171da177e4SLinus Torvalds { 13187beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13197beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13201da177e4SLinus Torvalds 13217beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13227beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1323fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1324fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13257beb2edfSTejun Heo 13268852aac2STejun Heo /* 13278852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13288852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13298852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13308852aac2STejun Heo * on that there's no such delay when @delay is 0. 13318852aac2STejun Heo */ 13328852aac2STejun Heo if (!delay) { 13338852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13348852aac2STejun Heo return; 13358852aac2STejun Heo } 13368852aac2STejun Heo 13377beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13387beb2edfSTejun Heo 133960c057bcSLai Jiangshan dwork->wq = wq; 13401265057fSTejun Heo dwork->cpu = cpu; 13417beb2edfSTejun Heo timer->expires = jiffies + delay; 13427beb2edfSTejun Heo 13437beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13447beb2edfSTejun Heo add_timer_on(timer, cpu); 13457beb2edfSTejun Heo else 13467beb2edfSTejun Heo add_timer(timer); 13477beb2edfSTejun Heo } 13481da177e4SLinus Torvalds 13490fcb78c2SRolf Eike Beer /** 13500fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13510fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13520fcb78c2SRolf Eike Beer * @wq: workqueue to use 1353af9997e4SRandy Dunlap * @dwork: work to queue 13540fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13550fcb78c2SRolf Eike Beer * 1356715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1357715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1358715f1300STejun Heo * execution. 13590fcb78c2SRolf Eike Beer */ 1360d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 136152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13627a6bc1cdSVenkatesh Pallipadi { 136352bad64dSDavid Howells struct work_struct *work = &dwork->work; 1364d4283e93STejun Heo bool ret = false; 13658930cabaSTejun Heo unsigned long flags; 13668930cabaSTejun Heo 13678930cabaSTejun Heo /* read the comment in __queue_work() */ 13688930cabaSTejun Heo local_irq_save(flags); 13697a6bc1cdSVenkatesh Pallipadi 137022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13717beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1372d4283e93STejun Heo ret = true; 13737a6bc1cdSVenkatesh Pallipadi } 13748930cabaSTejun Heo 13758930cabaSTejun Heo local_irq_restore(flags); 13767a6bc1cdSVenkatesh Pallipadi return ret; 13777a6bc1cdSVenkatesh Pallipadi } 1378ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 13791da177e4SLinus Torvalds 1380c8e55f36STejun Heo /** 13810a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 13820a13c00eSTejun Heo * @wq: workqueue to use 13830a13c00eSTejun Heo * @dwork: delayable work to queue 13840a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 13850a13c00eSTejun Heo * 1386715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 13870a13c00eSTejun Heo */ 1388d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 13890a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 13900a13c00eSTejun Heo { 139157469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 13920a13c00eSTejun Heo } 13930a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 13940a13c00eSTejun Heo 13950a13c00eSTejun Heo /** 13968376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 13978376fe22STejun Heo * @cpu: CPU number to execute work on 13988376fe22STejun Heo * @wq: workqueue to use 13998376fe22STejun Heo * @dwork: work to queue 14008376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14018376fe22STejun Heo * 14028376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14038376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14048376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14058376fe22STejun Heo * current state. 14068376fe22STejun Heo * 14078376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14088376fe22STejun Heo * pending and its timer was modified. 14098376fe22STejun Heo * 1410e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14118376fe22STejun Heo * See try_to_grab_pending() for details. 14128376fe22STejun Heo */ 14138376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14148376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14158376fe22STejun Heo { 14168376fe22STejun Heo unsigned long flags; 14178376fe22STejun Heo int ret; 14188376fe22STejun Heo 14198376fe22STejun Heo do { 14208376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14218376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14228376fe22STejun Heo 14238376fe22STejun Heo if (likely(ret >= 0)) { 14248376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14258376fe22STejun Heo local_irq_restore(flags); 14268376fe22STejun Heo } 14278376fe22STejun Heo 14288376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14298376fe22STejun Heo return ret; 14308376fe22STejun Heo } 14318376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14328376fe22STejun Heo 14338376fe22STejun Heo /** 14348376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14358376fe22STejun Heo * @wq: workqueue to use 14368376fe22STejun Heo * @dwork: work to queue 14378376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14388376fe22STejun Heo * 14398376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14408376fe22STejun Heo */ 14418376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14428376fe22STejun Heo unsigned long delay) 14438376fe22STejun Heo { 14448376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14458376fe22STejun Heo } 14468376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14478376fe22STejun Heo 14488376fe22STejun Heo /** 1449c8e55f36STejun Heo * worker_enter_idle - enter idle state 1450c8e55f36STejun Heo * @worker: worker which is entering idle state 1451c8e55f36STejun Heo * 1452c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1453c8e55f36STejun Heo * necessary. 1454c8e55f36STejun Heo * 1455c8e55f36STejun Heo * LOCKING: 1456d565ed63STejun Heo * spin_lock_irq(pool->lock). 1457c8e55f36STejun Heo */ 1458c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14591da177e4SLinus Torvalds { 1460bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1461c8e55f36STejun Heo 14626183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 14636183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 14646183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 14656183c009STejun Heo return; 1466c8e55f36STejun Heo 1467cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1468cb444766STejun Heo worker->flags |= WORKER_IDLE; 1469bd7bdd43STejun Heo pool->nr_idle++; 1470e22bee78STejun Heo worker->last_active = jiffies; 1471c8e55f36STejun Heo 1472c8e55f36STejun Heo /* idle_list is LIFO */ 1473bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1474db7bccf4STejun Heo 147563d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1476628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1477cb444766STejun Heo 1478544ecf31STejun Heo /* 1479706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1480d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1481628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1482628c78e7STejun Heo * unbind is not in progress. 1483544ecf31STejun Heo */ 148424647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1485bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1486e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1487c8e55f36STejun Heo } 1488c8e55f36STejun Heo 1489c8e55f36STejun Heo /** 1490c8e55f36STejun Heo * worker_leave_idle - leave idle state 1491c8e55f36STejun Heo * @worker: worker which is leaving idle state 1492c8e55f36STejun Heo * 1493c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1494c8e55f36STejun Heo * 1495c8e55f36STejun Heo * LOCKING: 1496d565ed63STejun Heo * spin_lock_irq(pool->lock). 1497c8e55f36STejun Heo */ 1498c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1499c8e55f36STejun Heo { 1500bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1501c8e55f36STejun Heo 15026183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15036183c009STejun Heo return; 1504d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1505bd7bdd43STejun Heo pool->nr_idle--; 1506c8e55f36STejun Heo list_del_init(&worker->entry); 1507c8e55f36STejun Heo } 1508c8e55f36STejun Heo 1509e22bee78STejun Heo /** 1510f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1511f36dc67bSLai Jiangshan * @pool: target worker_pool 1512f36dc67bSLai Jiangshan * 1513f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1514e22bee78STejun Heo * 1515e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1516e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1517e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1518e22bee78STejun Heo * guaranteed to execute on the cpu. 1519e22bee78STejun Heo * 1520f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1521e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1522e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1523e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1524706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1525e22bee78STejun Heo * [dis]associated in the meantime. 1526e22bee78STejun Heo * 1527706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 152824647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1529f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1530f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1531f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1532e22bee78STejun Heo * 1533e22bee78STejun Heo * CONTEXT: 1534d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1535e22bee78STejun Heo * held. 1536e22bee78STejun Heo * 1537e22bee78STejun Heo * RETURNS: 1538706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1539e22bee78STejun Heo * bound), %false if offline. 1540e22bee78STejun Heo */ 1541f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1542d565ed63STejun Heo __acquires(&pool->lock) 1543e22bee78STejun Heo { 1544e22bee78STejun Heo while (true) { 1545e22bee78STejun Heo /* 1546e22bee78STejun Heo * The following call may fail, succeed or succeed 1547e22bee78STejun Heo * without actually migrating the task to the cpu if 1548e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 154924647570STejun Heo * against POOL_DISASSOCIATED. 1550e22bee78STejun Heo */ 155124647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 1552f5faa077SLai Jiangshan set_cpus_allowed_ptr(current, get_cpu_mask(pool->cpu)); 1553e22bee78STejun Heo 1554d565ed63STejun Heo spin_lock_irq(&pool->lock); 155524647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1556e22bee78STejun Heo return false; 1557f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 1558e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1559ec22ca5eSTejun Heo get_cpu_mask(pool->cpu))) 1560e22bee78STejun Heo return true; 1561d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1562e22bee78STejun Heo 15635035b20fSTejun Heo /* 15645035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 15655035b20fSTejun Heo * and retry migration. cond_resched() is required here; 15665035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 15675035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 15685035b20fSTejun Heo */ 1569e22bee78STejun Heo cpu_relax(); 15705035b20fSTejun Heo cond_resched(); 1571e22bee78STejun Heo } 1572e22bee78STejun Heo } 1573e22bee78STejun Heo 1574e22bee78STejun Heo /* 1575ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 15765f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 157725511a47STejun Heo */ 157825511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 157925511a47STejun Heo { 15805f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 1581f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 15825f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 158325511a47STejun Heo 1584ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1585ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1586d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 158725511a47STejun Heo } 158825511a47STejun Heo 158925511a47STejun Heo /* 159025511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1591403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1592403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1593403c821dSTejun Heo * executed twice without intervening cpu down. 1594e22bee78STejun Heo */ 159525511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1596e22bee78STejun Heo { 1597e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1598e22bee78STejun Heo 1599f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 1600eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1601e22bee78STejun Heo 1602d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1603e22bee78STejun Heo } 1604e22bee78STejun Heo 160525511a47STejun Heo /** 160694cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 160794cf58bbSTejun Heo * @pool: pool of interest 160825511a47STejun Heo * 160994cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 161025511a47STejun Heo * is different for idle and busy ones. 161125511a47STejun Heo * 1612ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1613ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1614ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1615ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 161625511a47STejun Heo * 1617ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1618ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1619ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1620ea1abd61SLai Jiangshan * rebind. 162125511a47STejun Heo * 1622ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1623ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1624ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1625ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 162625511a47STejun Heo */ 162794cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 162825511a47STejun Heo { 1629ea1abd61SLai Jiangshan struct worker *worker, *n; 163025511a47STejun Heo int i; 163125511a47STejun Heo 1632b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 1633d565ed63STejun Heo lockdep_assert_held(&pool->lock); 163425511a47STejun Heo 16355f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1636ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1637ea1abd61SLai Jiangshan /* 163894cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 163994cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1640ea1abd61SLai Jiangshan */ 1641ea1abd61SLai Jiangshan list_del_init(&worker->entry); 164225511a47STejun Heo 164325511a47STejun Heo /* 164494cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 164594cf58bbSTejun Heo * idle_worker_rebind(). 164625511a47STejun Heo */ 164725511a47STejun Heo wake_up_process(worker->task); 164825511a47STejun Heo } 164925511a47STejun Heo 1650ea1abd61SLai Jiangshan /* rebind busy workers */ 1651b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) { 165225511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1653e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 165425511a47STejun Heo 165525511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 165625511a47STejun Heo work_data_bits(rebind_work))) 165725511a47STejun Heo continue; 165825511a47STejun Heo 165925511a47STejun Heo debug_work_activate(rebind_work); 166090beca5dSTejun Heo 166190beca5dSTejun Heo /* 166294cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 1663112202d9STejun Heo * and @pwq->pool consistent for sanity. 166490beca5dSTejun Heo */ 1665e34cdddbSTejun Heo if (std_worker_pool_pri(worker->pool)) 1666e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1667e2b6a6d5SJoonsoo Kim else 1668e2b6a6d5SJoonsoo Kim wq = system_wq; 1669ec58815aSTejun Heo 1670112202d9STejun Heo insert_work(get_pwq(pool->cpu, wq), rebind_work, 167125511a47STejun Heo worker->scheduled.next, 167225511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1673ec58815aSTejun Heo } 167425511a47STejun Heo } 167525511a47STejun Heo 1676c34056a3STejun Heo static struct worker *alloc_worker(void) 1677c34056a3STejun Heo { 1678c34056a3STejun Heo struct worker *worker; 1679c34056a3STejun Heo 1680c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1681c8e55f36STejun Heo if (worker) { 1682c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1683affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 168425511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1685e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1686e22bee78STejun Heo worker->flags = WORKER_PREP; 1687c8e55f36STejun Heo } 1688c34056a3STejun Heo return worker; 1689c34056a3STejun Heo } 1690c34056a3STejun Heo 1691c34056a3STejun Heo /** 1692c34056a3STejun Heo * create_worker - create a new workqueue worker 169363d95a91STejun Heo * @pool: pool the new worker will belong to 1694c34056a3STejun Heo * 169563d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1696c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1697c34056a3STejun Heo * destroy_worker(). 1698c34056a3STejun Heo * 1699c34056a3STejun Heo * CONTEXT: 1700c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1701c34056a3STejun Heo * 1702c34056a3STejun Heo * RETURNS: 1703c34056a3STejun Heo * Pointer to the newly created worker. 1704c34056a3STejun Heo */ 1705bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1706c34056a3STejun Heo { 1707e34cdddbSTejun Heo const char *pri = std_worker_pool_pri(pool) ? "H" : ""; 1708c34056a3STejun Heo struct worker *worker = NULL; 1709f3421797STejun Heo int id = -1; 1710c34056a3STejun Heo 1711d565ed63STejun Heo spin_lock_irq(&pool->lock); 1712bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1713d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1714bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1715c34056a3STejun Heo goto fail; 1716d565ed63STejun Heo spin_lock_irq(&pool->lock); 1717c34056a3STejun Heo } 1718d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1719c34056a3STejun Heo 1720c34056a3STejun Heo worker = alloc_worker(); 1721c34056a3STejun Heo if (!worker) 1722c34056a3STejun Heo goto fail; 1723c34056a3STejun Heo 1724bd7bdd43STejun Heo worker->pool = pool; 1725c34056a3STejun Heo worker->id = id; 1726c34056a3STejun Heo 1727ec22ca5eSTejun Heo if (pool->cpu != WORK_CPU_UNBOUND) 172894dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1729ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1730ec22ca5eSTejun Heo "kworker/%u:%d%s", pool->cpu, id, pri); 1731f3421797STejun Heo else 1732f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 17333270476aSTejun Heo "kworker/u:%d%s", id, pri); 1734c34056a3STejun Heo if (IS_ERR(worker->task)) 1735c34056a3STejun Heo goto fail; 1736c34056a3STejun Heo 1737e34cdddbSTejun Heo if (std_worker_pool_pri(pool)) 17383270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 17393270476aSTejun Heo 1740db7bccf4STejun Heo /* 1741bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 174224647570STejun Heo * %POOL_DISASSOCIATED. The caller is responsible for ensuring the 1743bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1744bc2ae0f5STejun Heo * above the flag definition for details. 1745bc2ae0f5STejun Heo * 1746bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1747bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1748db7bccf4STejun Heo */ 174924647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) { 1750ec22ca5eSTejun Heo kthread_bind(worker->task, pool->cpu); 1751bc2ae0f5STejun Heo } else { 1752db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1753f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1754f3421797STejun Heo } 1755c34056a3STejun Heo 1756c34056a3STejun Heo return worker; 1757c34056a3STejun Heo fail: 1758c34056a3STejun Heo if (id >= 0) { 1759d565ed63STejun Heo spin_lock_irq(&pool->lock); 1760bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1761d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1762c34056a3STejun Heo } 1763c34056a3STejun Heo kfree(worker); 1764c34056a3STejun Heo return NULL; 1765c34056a3STejun Heo } 1766c34056a3STejun Heo 1767c34056a3STejun Heo /** 1768c34056a3STejun Heo * start_worker - start a newly created worker 1769c34056a3STejun Heo * @worker: worker to start 1770c34056a3STejun Heo * 1771706026c2STejun Heo * Make the pool aware of @worker and start it. 1772c34056a3STejun Heo * 1773c34056a3STejun Heo * CONTEXT: 1774d565ed63STejun Heo * spin_lock_irq(pool->lock). 1775c34056a3STejun Heo */ 1776c34056a3STejun Heo static void start_worker(struct worker *worker) 1777c34056a3STejun Heo { 1778cb444766STejun Heo worker->flags |= WORKER_STARTED; 1779bd7bdd43STejun Heo worker->pool->nr_workers++; 1780c8e55f36STejun Heo worker_enter_idle(worker); 1781c34056a3STejun Heo wake_up_process(worker->task); 1782c34056a3STejun Heo } 1783c34056a3STejun Heo 1784c34056a3STejun Heo /** 1785c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1786c34056a3STejun Heo * @worker: worker to be destroyed 1787c34056a3STejun Heo * 1788706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1789c8e55f36STejun Heo * 1790c8e55f36STejun Heo * CONTEXT: 1791d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1792c34056a3STejun Heo */ 1793c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1794c34056a3STejun Heo { 1795bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1796c34056a3STejun Heo int id = worker->id; 1797c34056a3STejun Heo 1798c34056a3STejun Heo /* sanity check frenzy */ 17996183c009STejun Heo if (WARN_ON(worker->current_work) || 18006183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18016183c009STejun Heo return; 1802c34056a3STejun Heo 1803c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1804bd7bdd43STejun Heo pool->nr_workers--; 1805c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1806bd7bdd43STejun Heo pool->nr_idle--; 1807c8e55f36STejun Heo 1808c8e55f36STejun Heo list_del_init(&worker->entry); 1809cb444766STejun Heo worker->flags |= WORKER_DIE; 1810c8e55f36STejun Heo 1811d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1812c8e55f36STejun Heo 1813c34056a3STejun Heo kthread_stop(worker->task); 1814c34056a3STejun Heo kfree(worker); 1815c34056a3STejun Heo 1816d565ed63STejun Heo spin_lock_irq(&pool->lock); 1817bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1818c34056a3STejun Heo } 1819c34056a3STejun Heo 182063d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1821e22bee78STejun Heo { 182263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1823e22bee78STejun Heo 1824d565ed63STejun Heo spin_lock_irq(&pool->lock); 1825e22bee78STejun Heo 182663d95a91STejun Heo if (too_many_workers(pool)) { 1827e22bee78STejun Heo struct worker *worker; 1828e22bee78STejun Heo unsigned long expires; 1829e22bee78STejun Heo 1830e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 183163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1832e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1833e22bee78STejun Heo 1834e22bee78STejun Heo if (time_before(jiffies, expires)) 183563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1836e22bee78STejun Heo else { 1837e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 183811ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 183963d95a91STejun Heo wake_up_worker(pool); 1840e22bee78STejun Heo } 1841e22bee78STejun Heo } 1842e22bee78STejun Heo 1843d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1844e22bee78STejun Heo } 1845e22bee78STejun Heo 1846e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1847e22bee78STejun Heo { 1848112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1849112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1850f3421797STejun Heo unsigned int cpu; 1851e22bee78STejun Heo 1852e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1853e22bee78STejun Heo return false; 1854e22bee78STejun Heo 1855e22bee78STejun Heo /* mayday mayday mayday */ 1856112202d9STejun Heo cpu = pwq->pool->cpu; 1857f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1858f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1859f3421797STejun Heo cpu = 0; 1860f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1861e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1862e22bee78STejun Heo return true; 1863e22bee78STejun Heo } 1864e22bee78STejun Heo 1865706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1866e22bee78STejun Heo { 186763d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1868e22bee78STejun Heo struct work_struct *work; 1869e22bee78STejun Heo 1870d565ed63STejun Heo spin_lock_irq(&pool->lock); 1871e22bee78STejun Heo 187263d95a91STejun Heo if (need_to_create_worker(pool)) { 1873e22bee78STejun Heo /* 1874e22bee78STejun Heo * We've been trying to create a new worker but 1875e22bee78STejun Heo * haven't been successful. We might be hitting an 1876e22bee78STejun Heo * allocation deadlock. Send distress signals to 1877e22bee78STejun Heo * rescuers. 1878e22bee78STejun Heo */ 187963d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1880e22bee78STejun Heo send_mayday(work); 1881e22bee78STejun Heo } 1882e22bee78STejun Heo 1883d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1884e22bee78STejun Heo 188563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1886e22bee78STejun Heo } 1887e22bee78STejun Heo 1888e22bee78STejun Heo /** 1889e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 189063d95a91STejun Heo * @pool: pool to create a new worker for 1891e22bee78STejun Heo * 189263d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1893e22bee78STejun Heo * have at least one idle worker on return from this function. If 1894e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 189563d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1896e22bee78STejun Heo * possible allocation deadlock. 1897e22bee78STejun Heo * 1898e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1899e22bee78STejun Heo * may_start_working() true. 1900e22bee78STejun Heo * 1901e22bee78STejun Heo * LOCKING: 1902d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1903e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1904e22bee78STejun Heo * manager. 1905e22bee78STejun Heo * 1906e22bee78STejun Heo * RETURNS: 1907d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1908e22bee78STejun Heo * otherwise. 1909e22bee78STejun Heo */ 191063d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1911d565ed63STejun Heo __releases(&pool->lock) 1912d565ed63STejun Heo __acquires(&pool->lock) 1913e22bee78STejun Heo { 191463d95a91STejun Heo if (!need_to_create_worker(pool)) 1915e22bee78STejun Heo return false; 1916e22bee78STejun Heo restart: 1917d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19189f9c2364STejun Heo 1919e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 192063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1921e22bee78STejun Heo 1922e22bee78STejun Heo while (true) { 1923e22bee78STejun Heo struct worker *worker; 1924e22bee78STejun Heo 1925bc2ae0f5STejun Heo worker = create_worker(pool); 1926e22bee78STejun Heo if (worker) { 192763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1928d565ed63STejun Heo spin_lock_irq(&pool->lock); 1929e22bee78STejun Heo start_worker(worker); 19306183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19316183c009STejun Heo goto restart; 1932e22bee78STejun Heo return true; 1933e22bee78STejun Heo } 1934e22bee78STejun Heo 193563d95a91STejun Heo if (!need_to_create_worker(pool)) 1936e22bee78STejun Heo break; 1937e22bee78STejun Heo 1938e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1939e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19409f9c2364STejun Heo 194163d95a91STejun Heo if (!need_to_create_worker(pool)) 1942e22bee78STejun Heo break; 1943e22bee78STejun Heo } 1944e22bee78STejun Heo 194563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1946d565ed63STejun Heo spin_lock_irq(&pool->lock); 194763d95a91STejun Heo if (need_to_create_worker(pool)) 1948e22bee78STejun Heo goto restart; 1949e22bee78STejun Heo return true; 1950e22bee78STejun Heo } 1951e22bee78STejun Heo 1952e22bee78STejun Heo /** 1953e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 195463d95a91STejun Heo * @pool: pool to destroy workers for 1955e22bee78STejun Heo * 195663d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1957e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1958e22bee78STejun Heo * 1959e22bee78STejun Heo * LOCKING: 1960d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1961e22bee78STejun Heo * multiple times. Called only from manager. 1962e22bee78STejun Heo * 1963e22bee78STejun Heo * RETURNS: 1964d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1965e22bee78STejun Heo * otherwise. 1966e22bee78STejun Heo */ 196763d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1968e22bee78STejun Heo { 1969e22bee78STejun Heo bool ret = false; 1970e22bee78STejun Heo 197163d95a91STejun Heo while (too_many_workers(pool)) { 1972e22bee78STejun Heo struct worker *worker; 1973e22bee78STejun Heo unsigned long expires; 1974e22bee78STejun Heo 197563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1976e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1977e22bee78STejun Heo 1978e22bee78STejun Heo if (time_before(jiffies, expires)) { 197963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1980e22bee78STejun Heo break; 1981e22bee78STejun Heo } 1982e22bee78STejun Heo 1983e22bee78STejun Heo destroy_worker(worker); 1984e22bee78STejun Heo ret = true; 1985e22bee78STejun Heo } 1986e22bee78STejun Heo 1987e22bee78STejun Heo return ret; 1988e22bee78STejun Heo } 1989e22bee78STejun Heo 1990e22bee78STejun Heo /** 1991e22bee78STejun Heo * manage_workers - manage worker pool 1992e22bee78STejun Heo * @worker: self 1993e22bee78STejun Heo * 1994706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1995e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1996706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1997e22bee78STejun Heo * 1998e22bee78STejun Heo * The caller can safely start processing works on false return. On 1999e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2000e22bee78STejun Heo * and may_start_working() is true. 2001e22bee78STejun Heo * 2002e22bee78STejun Heo * CONTEXT: 2003d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2004e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2005e22bee78STejun Heo * 2006e22bee78STejun Heo * RETURNS: 2007d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2008d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2009e22bee78STejun Heo */ 2010e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2011e22bee78STejun Heo { 201263d95a91STejun Heo struct worker_pool *pool = worker->pool; 2013e22bee78STejun Heo bool ret = false; 2014e22bee78STejun Heo 2015ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2016e22bee78STejun Heo return ret; 2017e22bee78STejun Heo 2018552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2019ee378aa4SLai Jiangshan 2020ee378aa4SLai Jiangshan /* 2021ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2022ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2023ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2024ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2025ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2026b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2027ee378aa4SLai Jiangshan * manager against CPU hotplug. 2028ee378aa4SLai Jiangshan * 2029b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2030d565ed63STejun Heo * progress. trylock first without dropping @pool->lock. 2031ee378aa4SLai Jiangshan */ 2032b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2033d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2034b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2035ee378aa4SLai Jiangshan /* 2036ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2037b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2038ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2039706026c2STejun Heo * @pool's state and ours could have deviated. 2040ee378aa4SLai Jiangshan * 2041b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2042ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2043706026c2STejun Heo * on @pool's current state. Try it and adjust 2044ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2045ee378aa4SLai Jiangshan */ 2046f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(pool)) 2047ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2048ee378aa4SLai Jiangshan else 2049ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2050ee378aa4SLai Jiangshan 2051ee378aa4SLai Jiangshan ret = true; 2052ee378aa4SLai Jiangshan } 2053ee378aa4SLai Jiangshan 205411ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2055e22bee78STejun Heo 2056e22bee78STejun Heo /* 2057e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2058e22bee78STejun Heo * on return. 2059e22bee78STejun Heo */ 206063d95a91STejun Heo ret |= maybe_destroy_workers(pool); 206163d95a91STejun Heo ret |= maybe_create_worker(pool); 2062e22bee78STejun Heo 2063552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2064b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2065e22bee78STejun Heo return ret; 2066e22bee78STejun Heo } 2067e22bee78STejun Heo 2068a62428c0STejun Heo /** 2069a62428c0STejun Heo * process_one_work - process single work 2070c34056a3STejun Heo * @worker: self 2071a62428c0STejun Heo * @work: work to process 2072a62428c0STejun Heo * 2073a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2074a62428c0STejun Heo * process a single work including synchronization against and 2075a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2076a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2077a62428c0STejun Heo * call this function to process a work. 2078a62428c0STejun Heo * 2079a62428c0STejun Heo * CONTEXT: 2080d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2081a62428c0STejun Heo */ 2082c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2083d565ed63STejun Heo __releases(&pool->lock) 2084d565ed63STejun Heo __acquires(&pool->lock) 20851da177e4SLinus Torvalds { 2086112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2087bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2088112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 208973f53c4aSTejun Heo int work_color; 20907e11629dSTejun Heo struct worker *collision; 20914e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20924e6045f1SJohannes Berg /* 2093a62428c0STejun Heo * It is permissible to free the struct work_struct from 2094a62428c0STejun Heo * inside the function that is called from it, this we need to 2095a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2096a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2097a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20984e6045f1SJohannes Berg */ 20994d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21004d82a1deSPeter Zijlstra 21014d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21024e6045f1SJohannes Berg #endif 21036fec10a1STejun Heo /* 21046fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21056fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 210624647570STejun Heo * unbound or a disassociated pool. 21076fec10a1STejun Heo */ 21085f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 210924647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2110ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 211125511a47STejun Heo 21127e11629dSTejun Heo /* 21137e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21147e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21157e11629dSTejun Heo * already processing the work. If so, defer the work to the 21167e11629dSTejun Heo * currently executing one. 21177e11629dSTejun Heo */ 2118c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21197e11629dSTejun Heo if (unlikely(collision)) { 21207e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21217e11629dSTejun Heo return; 21227e11629dSTejun Heo } 21231da177e4SLinus Torvalds 21248930cabaSTejun Heo /* claim and dequeue */ 21251da177e4SLinus Torvalds debug_work_deactivate(work); 2126c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2127c34056a3STejun Heo worker->current_work = work; 2128a2c1c57bSTejun Heo worker->current_func = work->func; 2129112202d9STejun Heo worker->current_pwq = pwq; 213073f53c4aSTejun Heo work_color = get_work_color(work); 21317a22ad75STejun Heo 2132a62428c0STejun Heo list_del_init(&work->entry); 2133a62428c0STejun Heo 2134649027d7STejun Heo /* 2135fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2136fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2137fb0e7bebSTejun Heo */ 2138fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2139fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2140fb0e7bebSTejun Heo 2141974271c4STejun Heo /* 2142d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2143974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2144974271c4STejun Heo */ 214563d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 214663d95a91STejun Heo wake_up_worker(pool); 2147974271c4STejun Heo 21488930cabaSTejun Heo /* 21497c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2150d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 215123657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 215223657bb1STejun Heo * disabled. 21538930cabaSTejun Heo */ 21547c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21551da177e4SLinus Torvalds 2156d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2157365970a1SDavid Howells 2158112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21593295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2160e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2161a2c1c57bSTejun Heo worker->current_func(work); 2162e36c886aSArjan van de Ven /* 2163e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2164e36c886aSArjan van de Ven * point will only record its address. 2165e36c886aSArjan van de Ven */ 2166e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21673295f0efSIngo Molnar lock_map_release(&lockdep_map); 2168112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21691da177e4SLinus Torvalds 2170d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2171044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2172044c782cSValentin Ilie " last function: %pf\n", 2173a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2174a2c1c57bSTejun Heo worker->current_func); 2175d5abe669SPeter Zijlstra debug_show_held_locks(current); 2176d5abe669SPeter Zijlstra dump_stack(); 2177d5abe669SPeter Zijlstra } 2178d5abe669SPeter Zijlstra 2179d565ed63STejun Heo spin_lock_irq(&pool->lock); 2180a62428c0STejun Heo 2181fb0e7bebSTejun Heo /* clear cpu intensive status */ 2182fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2183fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2184fb0e7bebSTejun Heo 2185a62428c0STejun Heo /* we're done with it, release */ 218642f8570fSSasha Levin hash_del(&worker->hentry); 2187c34056a3STejun Heo worker->current_work = NULL; 2188a2c1c57bSTejun Heo worker->current_func = NULL; 2189112202d9STejun Heo worker->current_pwq = NULL; 2190112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21911da177e4SLinus Torvalds } 21921da177e4SLinus Torvalds 2193affee4b2STejun Heo /** 2194affee4b2STejun Heo * process_scheduled_works - process scheduled works 2195affee4b2STejun Heo * @worker: self 2196affee4b2STejun Heo * 2197affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2198affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2199affee4b2STejun Heo * fetches a work from the top and executes it. 2200affee4b2STejun Heo * 2201affee4b2STejun Heo * CONTEXT: 2202d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2203affee4b2STejun Heo * multiple times. 2204affee4b2STejun Heo */ 2205affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22061da177e4SLinus Torvalds { 2207affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2208affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2209a62428c0STejun Heo struct work_struct, entry); 2210c34056a3STejun Heo process_one_work(worker, work); 2211a62428c0STejun Heo } 22121da177e4SLinus Torvalds } 22131da177e4SLinus Torvalds 22144690c4abSTejun Heo /** 22154690c4abSTejun Heo * worker_thread - the worker thread function 2216c34056a3STejun Heo * @__worker: self 22174690c4abSTejun Heo * 2218706026c2STejun Heo * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools 2219706026c2STejun Heo * of these per each cpu. These workers process all works regardless of 2220e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2221e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2222e22bee78STejun Heo * rescuer_thread(). 22234690c4abSTejun Heo */ 2224c34056a3STejun Heo static int worker_thread(void *__worker) 22251da177e4SLinus Torvalds { 2226c34056a3STejun Heo struct worker *worker = __worker; 2227bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22281da177e4SLinus Torvalds 2229e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2230e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2231c8e55f36STejun Heo woke_up: 2232d565ed63STejun Heo spin_lock_irq(&pool->lock); 2233affee4b2STejun Heo 22345f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22355f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2236d565ed63STejun Heo spin_unlock_irq(&pool->lock); 223725511a47STejun Heo 22385f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 223925511a47STejun Heo if (worker->flags & WORKER_DIE) { 2240e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2241c8e55f36STejun Heo return 0; 2242c8e55f36STejun Heo } 2243c8e55f36STejun Heo 22445f7dabfdSLai Jiangshan /* otherwise, rebind */ 224525511a47STejun Heo idle_worker_rebind(worker); 224625511a47STejun Heo goto woke_up; 224725511a47STejun Heo } 224825511a47STejun Heo 2249c8e55f36STejun Heo worker_leave_idle(worker); 2250db7bccf4STejun Heo recheck: 2251e22bee78STejun Heo /* no more worker necessary? */ 225263d95a91STejun Heo if (!need_more_worker(pool)) 2253e22bee78STejun Heo goto sleep; 2254e22bee78STejun Heo 2255e22bee78STejun Heo /* do we need to manage? */ 225663d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2257e22bee78STejun Heo goto recheck; 2258e22bee78STejun Heo 2259c8e55f36STejun Heo /* 2260c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2261c8e55f36STejun Heo * preparing to process a work or actually processing it. 2262c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2263c8e55f36STejun Heo */ 22646183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2265c8e55f36STejun Heo 2266e22bee78STejun Heo /* 2267e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2268e22bee78STejun Heo * at least one idle worker or that someone else has already 2269e22bee78STejun Heo * assumed the manager role. 2270e22bee78STejun Heo */ 2271e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2272e22bee78STejun Heo 2273e22bee78STejun Heo do { 2274affee4b2STejun Heo struct work_struct *work = 2275bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2276affee4b2STejun Heo struct work_struct, entry); 2277affee4b2STejun Heo 2278c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2279affee4b2STejun Heo /* optimization path, not strictly necessary */ 2280affee4b2STejun Heo process_one_work(worker, work); 2281affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2282affee4b2STejun Heo process_scheduled_works(worker); 2283affee4b2STejun Heo } else { 2284c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2285affee4b2STejun Heo process_scheduled_works(worker); 2286affee4b2STejun Heo } 228763d95a91STejun Heo } while (keep_working(pool)); 2288affee4b2STejun Heo 2289e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2290d313dd85STejun Heo sleep: 229163d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2292e22bee78STejun Heo goto recheck; 2293d313dd85STejun Heo 2294c8e55f36STejun Heo /* 2295d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2296d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2297d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2298d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2299d565ed63STejun Heo * event. 2300c8e55f36STejun Heo */ 2301c8e55f36STejun Heo worker_enter_idle(worker); 2302c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2303d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23041da177e4SLinus Torvalds schedule(); 2305c8e55f36STejun Heo goto woke_up; 23061da177e4SLinus Torvalds } 23071da177e4SLinus Torvalds 2308e22bee78STejun Heo /** 2309e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2310111c225aSTejun Heo * @__rescuer: self 2311e22bee78STejun Heo * 2312e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2313e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2314e22bee78STejun Heo * 2315706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2316e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2317e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2318e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2319e22bee78STejun Heo * the problem rescuer solves. 2320e22bee78STejun Heo * 2321706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2322706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2323e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2324e22bee78STejun Heo * 2325e22bee78STejun Heo * This should happen rarely. 2326e22bee78STejun Heo */ 2327111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2328e22bee78STejun Heo { 2329111c225aSTejun Heo struct worker *rescuer = __rescuer; 2330111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2331e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2332f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2333e22bee78STejun Heo unsigned int cpu; 2334e22bee78STejun Heo 2335e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2336111c225aSTejun Heo 2337111c225aSTejun Heo /* 2338111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2339111c225aSTejun Heo * doesn't participate in concurrency management. 2340111c225aSTejun Heo */ 2341111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2342e22bee78STejun Heo repeat: 2343e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23441da177e4SLinus Torvalds 2345412d32e6SMike Galbraith if (kthread_should_stop()) { 2346412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2347111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2348e22bee78STejun Heo return 0; 2349412d32e6SMike Galbraith } 23501da177e4SLinus Torvalds 2351f3421797STejun Heo /* 2352f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2353f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2354f3421797STejun Heo */ 2355f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2356f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2357112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(tcpu, wq); 2358112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2359e22bee78STejun Heo struct work_struct *work, *n; 2360e22bee78STejun Heo 2361e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2362f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2363e22bee78STejun Heo 2364e22bee78STejun Heo /* migrate to the target cpu if possible */ 2365f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2366b3104104SLai Jiangshan rescuer->pool = pool; 2367e22bee78STejun Heo 2368e22bee78STejun Heo /* 2369e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2370e22bee78STejun Heo * process'em. 2371e22bee78STejun Heo */ 23726183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2373bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2374112202d9STejun Heo if (get_work_pwq(work) == pwq) 2375e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2376e22bee78STejun Heo 2377e22bee78STejun Heo process_scheduled_works(rescuer); 23787576958aSTejun Heo 23797576958aSTejun Heo /* 2380d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23817576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23827576958aSTejun Heo * and stalling the execution. 23837576958aSTejun Heo */ 238463d95a91STejun Heo if (keep_working(pool)) 238563d95a91STejun Heo wake_up_worker(pool); 23867576958aSTejun Heo 2387b3104104SLai Jiangshan rescuer->pool = NULL; 2388d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23891da177e4SLinus Torvalds } 23901da177e4SLinus Torvalds 2391111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2392111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2393e22bee78STejun Heo schedule(); 2394e22bee78STejun Heo goto repeat; 23951da177e4SLinus Torvalds } 23961da177e4SLinus Torvalds 2397fc2e4d70SOleg Nesterov struct wq_barrier { 2398fc2e4d70SOleg Nesterov struct work_struct work; 2399fc2e4d70SOleg Nesterov struct completion done; 2400fc2e4d70SOleg Nesterov }; 2401fc2e4d70SOleg Nesterov 2402fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2403fc2e4d70SOleg Nesterov { 2404fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2405fc2e4d70SOleg Nesterov complete(&barr->done); 2406fc2e4d70SOleg Nesterov } 2407fc2e4d70SOleg Nesterov 24084690c4abSTejun Heo /** 24094690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2410112202d9STejun Heo * @pwq: pwq to insert barrier into 24114690c4abSTejun Heo * @barr: wq_barrier to insert 2412affee4b2STejun Heo * @target: target work to attach @barr to 2413affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24144690c4abSTejun Heo * 2415affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2416affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2417affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2418affee4b2STejun Heo * cpu. 2419affee4b2STejun Heo * 2420affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2421affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2422affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2423affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2424affee4b2STejun Heo * after a work with LINKED flag set. 2425affee4b2STejun Heo * 2426affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2427112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24284690c4abSTejun Heo * 24294690c4abSTejun Heo * CONTEXT: 2430d565ed63STejun Heo * spin_lock_irq(pool->lock). 24314690c4abSTejun Heo */ 2432112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2433affee4b2STejun Heo struct wq_barrier *barr, 2434affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2435fc2e4d70SOleg Nesterov { 2436affee4b2STejun Heo struct list_head *head; 2437affee4b2STejun Heo unsigned int linked = 0; 2438affee4b2STejun Heo 2439dc186ad7SThomas Gleixner /* 2440d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2441dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2442dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2443dc186ad7SThomas Gleixner * might deadlock. 2444dc186ad7SThomas Gleixner */ 2445ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 244622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2447fc2e4d70SOleg Nesterov init_completion(&barr->done); 244883c22520SOleg Nesterov 2449affee4b2STejun Heo /* 2450affee4b2STejun Heo * If @target is currently being executed, schedule the 2451affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2452affee4b2STejun Heo */ 2453affee4b2STejun Heo if (worker) 2454affee4b2STejun Heo head = worker->scheduled.next; 2455affee4b2STejun Heo else { 2456affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2457affee4b2STejun Heo 2458affee4b2STejun Heo head = target->entry.next; 2459affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2460affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2461affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2462affee4b2STejun Heo } 2463affee4b2STejun Heo 2464dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2465112202d9STejun Heo insert_work(pwq, &barr->work, head, 2466affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2467fc2e4d70SOleg Nesterov } 2468fc2e4d70SOleg Nesterov 246973f53c4aSTejun Heo /** 2470112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 247173f53c4aSTejun Heo * @wq: workqueue being flushed 247273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 247373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 247473f53c4aSTejun Heo * 2475112202d9STejun Heo * Prepare pwqs for workqueue flushing. 247673f53c4aSTejun Heo * 2477112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2478112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2479112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2480112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2481112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 248273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 248373f53c4aSTejun Heo * 248473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 248573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 248673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 248773f53c4aSTejun Heo * is returned. 248873f53c4aSTejun Heo * 2489112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 249073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 249173f53c4aSTejun Heo * advanced to @work_color. 249273f53c4aSTejun Heo * 249373f53c4aSTejun Heo * CONTEXT: 249473f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 249573f53c4aSTejun Heo * 249673f53c4aSTejun Heo * RETURNS: 249773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 249873f53c4aSTejun Heo * otherwise. 249973f53c4aSTejun Heo */ 2500112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 250173f53c4aSTejun Heo int flush_color, int work_color) 25021da177e4SLinus Torvalds { 250373f53c4aSTejun Heo bool wait = false; 250473f53c4aSTejun Heo unsigned int cpu; 25051da177e4SLinus Torvalds 250673f53c4aSTejun Heo if (flush_color >= 0) { 25076183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2508112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2509dc186ad7SThomas Gleixner } 251014441960SOleg Nesterov 2511112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 2512112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 2513112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25141da177e4SLinus Torvalds 2515d565ed63STejun Heo spin_lock_irq(&pool->lock); 251673f53c4aSTejun Heo 251773f53c4aSTejun Heo if (flush_color >= 0) { 25186183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 251973f53c4aSTejun Heo 2520112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2521112202d9STejun Heo pwq->flush_color = flush_color; 2522112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 252373f53c4aSTejun Heo wait = true; 25241da177e4SLinus Torvalds } 252573f53c4aSTejun Heo } 252673f53c4aSTejun Heo 252773f53c4aSTejun Heo if (work_color >= 0) { 25286183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2529112202d9STejun Heo pwq->work_color = work_color; 253073f53c4aSTejun Heo } 253173f53c4aSTejun Heo 2532d565ed63STejun Heo spin_unlock_irq(&pool->lock); 25331da177e4SLinus Torvalds } 25341da177e4SLinus Torvalds 2535112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 253673f53c4aSTejun Heo complete(&wq->first_flusher->done); 253773f53c4aSTejun Heo 253873f53c4aSTejun Heo return wait; 253983c22520SOleg Nesterov } 25401da177e4SLinus Torvalds 25410fcb78c2SRolf Eike Beer /** 25421da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25430fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25441da177e4SLinus Torvalds * 25451da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 25461da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 25471da177e4SLinus Torvalds * 2548fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2549fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 25501da177e4SLinus Torvalds */ 25517ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25521da177e4SLinus Torvalds { 255373f53c4aSTejun Heo struct wq_flusher this_flusher = { 255473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 255573f53c4aSTejun Heo .flush_color = -1, 255673f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 255773f53c4aSTejun Heo }; 255873f53c4aSTejun Heo int next_color; 2559b1f4ec17SOleg Nesterov 25603295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25613295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 256273f53c4aSTejun Heo 256373f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 256473f53c4aSTejun Heo 256573f53c4aSTejun Heo /* 256673f53c4aSTejun Heo * Start-to-wait phase 256773f53c4aSTejun Heo */ 256873f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 256973f53c4aSTejun Heo 257073f53c4aSTejun Heo if (next_color != wq->flush_color) { 257173f53c4aSTejun Heo /* 257273f53c4aSTejun Heo * Color space is not full. The current work_color 257373f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 257473f53c4aSTejun Heo * by one. 257573f53c4aSTejun Heo */ 25766183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 257773f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 257873f53c4aSTejun Heo wq->work_color = next_color; 257973f53c4aSTejun Heo 258073f53c4aSTejun Heo if (!wq->first_flusher) { 258173f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25826183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 258373f53c4aSTejun Heo 258473f53c4aSTejun Heo wq->first_flusher = &this_flusher; 258573f53c4aSTejun Heo 2586112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 258773f53c4aSTejun Heo wq->work_color)) { 258873f53c4aSTejun Heo /* nothing to flush, done */ 258973f53c4aSTejun Heo wq->flush_color = next_color; 259073f53c4aSTejun Heo wq->first_flusher = NULL; 259173f53c4aSTejun Heo goto out_unlock; 259273f53c4aSTejun Heo } 259373f53c4aSTejun Heo } else { 259473f53c4aSTejun Heo /* wait in queue */ 25956183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 259673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2597112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 259873f53c4aSTejun Heo } 259973f53c4aSTejun Heo } else { 260073f53c4aSTejun Heo /* 260173f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 260273f53c4aSTejun Heo * The next flush completion will assign us 260373f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 260473f53c4aSTejun Heo */ 260573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 260673f53c4aSTejun Heo } 260773f53c4aSTejun Heo 260873f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 260973f53c4aSTejun Heo 261073f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 261173f53c4aSTejun Heo 261273f53c4aSTejun Heo /* 261373f53c4aSTejun Heo * Wake-up-and-cascade phase 261473f53c4aSTejun Heo * 261573f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 261673f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 261773f53c4aSTejun Heo */ 261873f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 261973f53c4aSTejun Heo return; 262073f53c4aSTejun Heo 262173f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 262273f53c4aSTejun Heo 26234ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26244ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26254ce48b37STejun Heo goto out_unlock; 26264ce48b37STejun Heo 262773f53c4aSTejun Heo wq->first_flusher = NULL; 262873f53c4aSTejun Heo 26296183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26306183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 263173f53c4aSTejun Heo 263273f53c4aSTejun Heo while (true) { 263373f53c4aSTejun Heo struct wq_flusher *next, *tmp; 263473f53c4aSTejun Heo 263573f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 263673f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 263773f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 263873f53c4aSTejun Heo break; 263973f53c4aSTejun Heo list_del_init(&next->list); 264073f53c4aSTejun Heo complete(&next->done); 264173f53c4aSTejun Heo } 264273f53c4aSTejun Heo 26436183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 264473f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 264573f53c4aSTejun Heo 264673f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 264773f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 264873f53c4aSTejun Heo 264973f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 265073f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 265173f53c4aSTejun Heo /* 265273f53c4aSTejun Heo * Assign the same color to all overflowed 265373f53c4aSTejun Heo * flushers, advance work_color and append to 265473f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 265573f53c4aSTejun Heo * phase for these overflowed flushers. 265673f53c4aSTejun Heo */ 265773f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 265873f53c4aSTejun Heo tmp->flush_color = wq->work_color; 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 266173f53c4aSTejun Heo 266273f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 266373f53c4aSTejun Heo &wq->flusher_queue); 2664112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 266573f53c4aSTejun Heo } 266673f53c4aSTejun Heo 266773f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26686183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 266973f53c4aSTejun Heo break; 267073f53c4aSTejun Heo } 267173f53c4aSTejun Heo 267273f53c4aSTejun Heo /* 267373f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2674112202d9STejun Heo * the new first flusher and arm pwqs. 267573f53c4aSTejun Heo */ 26766183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26776183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 267873f53c4aSTejun Heo 267973f53c4aSTejun Heo list_del_init(&next->list); 268073f53c4aSTejun Heo wq->first_flusher = next; 268173f53c4aSTejun Heo 2682112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 268373f53c4aSTejun Heo break; 268473f53c4aSTejun Heo 268573f53c4aSTejun Heo /* 268673f53c4aSTejun Heo * Meh... this color is already done, clear first 268773f53c4aSTejun Heo * flusher and repeat cascading. 268873f53c4aSTejun Heo */ 268973f53c4aSTejun Heo wq->first_flusher = NULL; 269073f53c4aSTejun Heo } 269173f53c4aSTejun Heo 269273f53c4aSTejun Heo out_unlock: 269373f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 26941da177e4SLinus Torvalds } 2695ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26961da177e4SLinus Torvalds 26979c5a2ba7STejun Heo /** 26989c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26999c5a2ba7STejun Heo * @wq: workqueue to drain 27009c5a2ba7STejun Heo * 27019c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27029c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27039c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27049c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27059c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27069c5a2ba7STejun Heo * takes too long. 27079c5a2ba7STejun Heo */ 27089c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27099c5a2ba7STejun Heo { 27109c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27119c5a2ba7STejun Heo unsigned int cpu; 27129c5a2ba7STejun Heo 27139c5a2ba7STejun Heo /* 27149c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27159c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27169c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27179c5a2ba7STejun Heo */ 27189c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27199c5a2ba7STejun Heo if (!wq->nr_drainers++) 27209c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 27219c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27229c5a2ba7STejun Heo reflush: 27239c5a2ba7STejun Heo flush_workqueue(wq); 27249c5a2ba7STejun Heo 2725112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 2726112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 2727fa2563e4SThomas Tuttle bool drained; 27289c5a2ba7STejun Heo 2729112202d9STejun Heo spin_lock_irq(&pwq->pool->lock); 2730112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2731112202d9STejun Heo spin_unlock_irq(&pwq->pool->lock); 2732fa2563e4SThomas Tuttle 2733fa2563e4SThomas Tuttle if (drained) 27349c5a2ba7STejun Heo continue; 27359c5a2ba7STejun Heo 27369c5a2ba7STejun Heo if (++flush_cnt == 10 || 27379c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2738044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 27399c5a2ba7STejun Heo wq->name, flush_cnt); 27409c5a2ba7STejun Heo goto reflush; 27419c5a2ba7STejun Heo } 27429c5a2ba7STejun Heo 27439c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27449c5a2ba7STejun Heo if (!--wq->nr_drainers) 27459c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 27469c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27479c5a2ba7STejun Heo } 27489c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27499c5a2ba7STejun Heo 2750606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2751baf59022STejun Heo { 2752baf59022STejun Heo struct worker *worker = NULL; 2753c9e7cf27STejun Heo struct worker_pool *pool; 2754112202d9STejun Heo struct pool_workqueue *pwq; 2755baf59022STejun Heo 2756baf59022STejun Heo might_sleep(); 2757c9e7cf27STejun Heo pool = get_work_pool(work); 2758c9e7cf27STejun Heo if (!pool) 2759baf59022STejun Heo return false; 2760baf59022STejun Heo 2761d565ed63STejun Heo spin_lock_irq(&pool->lock); 27620b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2763112202d9STejun Heo pwq = get_work_pwq(work); 2764112202d9STejun Heo if (pwq) { 2765112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2766baf59022STejun Heo goto already_gone; 2767606a5020STejun Heo } else { 2768c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2769baf59022STejun Heo if (!worker) 2770baf59022STejun Heo goto already_gone; 2771112202d9STejun Heo pwq = worker->current_pwq; 2772606a5020STejun Heo } 2773baf59022STejun Heo 2774112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2775d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2776baf59022STejun Heo 2777e159489bSTejun Heo /* 2778e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2779e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2780e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2781e159489bSTejun Heo * access. 2782e159489bSTejun Heo */ 2783112202d9STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER) 2784112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2785e159489bSTejun Heo else 2786112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2787112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2788e159489bSTejun Heo 2789baf59022STejun Heo return true; 2790baf59022STejun Heo already_gone: 2791d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2792baf59022STejun Heo return false; 2793baf59022STejun Heo } 2794baf59022STejun Heo 2795db700897SOleg Nesterov /** 2796401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2797401a8d04STejun Heo * @work: the work to flush 2798db700897SOleg Nesterov * 2799606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2800606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2801401a8d04STejun Heo * 2802401a8d04STejun Heo * RETURNS: 2803401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2804401a8d04STejun Heo * %false if it was already idle. 2805db700897SOleg Nesterov */ 2806401a8d04STejun Heo bool flush_work(struct work_struct *work) 2807db700897SOleg Nesterov { 2808db700897SOleg Nesterov struct wq_barrier barr; 2809db700897SOleg Nesterov 28100976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28110976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28120976dfc1SStephen Boyd 2813606a5020STejun Heo if (start_flush_work(work, &barr)) { 2814db700897SOleg Nesterov wait_for_completion(&barr.done); 2815dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2816401a8d04STejun Heo return true; 2817606a5020STejun Heo } else { 2818401a8d04STejun Heo return false; 2819db700897SOleg Nesterov } 2820606a5020STejun Heo } 2821db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2822db700897SOleg Nesterov 282336e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2824401a8d04STejun Heo { 2825bbb68dfaSTejun Heo unsigned long flags; 28261f1f642eSOleg Nesterov int ret; 28271f1f642eSOleg Nesterov 28281f1f642eSOleg Nesterov do { 2829bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2830bbb68dfaSTejun Heo /* 2831bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2832bbb68dfaSTejun Heo * would be waiting for before retrying. 2833bbb68dfaSTejun Heo */ 2834bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2835606a5020STejun Heo flush_work(work); 28361f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28371f1f642eSOleg Nesterov 2838bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2839bbb68dfaSTejun Heo mark_work_canceling(work); 2840bbb68dfaSTejun Heo local_irq_restore(flags); 2841bbb68dfaSTejun Heo 2842606a5020STejun Heo flush_work(work); 28437a22ad75STejun Heo clear_work_data(work); 28441f1f642eSOleg Nesterov return ret; 28451f1f642eSOleg Nesterov } 28461f1f642eSOleg Nesterov 28476e84d644SOleg Nesterov /** 2848401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2849401a8d04STejun Heo * @work: the work to cancel 28506e84d644SOleg Nesterov * 2851401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2852401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2853401a8d04STejun Heo * another workqueue. On return from this function, @work is 2854401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28551f1f642eSOleg Nesterov * 2856401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2857401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28586e84d644SOleg Nesterov * 2859401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28606e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2861401a8d04STejun Heo * 2862401a8d04STejun Heo * RETURNS: 2863401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28646e84d644SOleg Nesterov */ 2865401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28666e84d644SOleg Nesterov { 286736e227d2STejun Heo return __cancel_work_timer(work, false); 2868b89deed3SOleg Nesterov } 286928e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2870b89deed3SOleg Nesterov 28716e84d644SOleg Nesterov /** 2872401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2873401a8d04STejun Heo * @dwork: the delayed work to flush 28746e84d644SOleg Nesterov * 2875401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2876401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2877401a8d04STejun Heo * considers the last queueing instance of @dwork. 28781f1f642eSOleg Nesterov * 2879401a8d04STejun Heo * RETURNS: 2880401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2881401a8d04STejun Heo * %false if it was already idle. 28826e84d644SOleg Nesterov */ 2883401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2884401a8d04STejun Heo { 28858930cabaSTejun Heo local_irq_disable(); 2886401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 288760c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28888930cabaSTejun Heo local_irq_enable(); 2889401a8d04STejun Heo return flush_work(&dwork->work); 2890401a8d04STejun Heo } 2891401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2892401a8d04STejun Heo 2893401a8d04STejun Heo /** 289457b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 289557b30ae7STejun Heo * @dwork: delayed_work to cancel 289609383498STejun Heo * 289757b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 289857b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 289957b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 290057b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 290157b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 290209383498STejun Heo * 290357b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 290409383498STejun Heo */ 290557b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 290609383498STejun Heo { 290757b30ae7STejun Heo unsigned long flags; 290857b30ae7STejun Heo int ret; 290957b30ae7STejun Heo 291057b30ae7STejun Heo do { 291157b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 291257b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 291357b30ae7STejun Heo 291457b30ae7STejun Heo if (unlikely(ret < 0)) 291557b30ae7STejun Heo return false; 291657b30ae7STejun Heo 29177c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29187c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 291957b30ae7STejun Heo local_irq_restore(flags); 2920c0158ca6SDan Magenheimer return ret; 292109383498STejun Heo } 292257b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 292309383498STejun Heo 292409383498STejun Heo /** 2925401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2926401a8d04STejun Heo * @dwork: the delayed work cancel 2927401a8d04STejun Heo * 2928401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2929401a8d04STejun Heo * 2930401a8d04STejun Heo * RETURNS: 2931401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2932401a8d04STejun Heo */ 2933401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29346e84d644SOleg Nesterov { 293536e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29366e84d644SOleg Nesterov } 2937f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29381da177e4SLinus Torvalds 29390fcb78c2SRolf Eike Beer /** 2940c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2941c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2942c1a220e7SZhang Rui * @work: job to be done 2943c1a220e7SZhang Rui * 2944c1a220e7SZhang Rui * This puts a job on a specific cpu 2945c1a220e7SZhang Rui */ 2946d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 2947c1a220e7SZhang Rui { 2948d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2949c1a220e7SZhang Rui } 2950c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2951c1a220e7SZhang Rui 29520fcb78c2SRolf Eike Beer /** 2953ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 29541da177e4SLinus Torvalds * @work: job to be done 29551da177e4SLinus Torvalds * 2956d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 2957d4283e93STejun Heo * %true otherwise. 295852bad64dSDavid Howells * 29590fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 29600fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 29610fcb78c2SRolf Eike Beer * workqueue otherwise. 29621da177e4SLinus Torvalds */ 2963d4283e93STejun Heo bool schedule_work(struct work_struct *work) 29641da177e4SLinus Torvalds { 29650fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 29661da177e4SLinus Torvalds } 29670fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 29681da177e4SLinus Torvalds 29690fcb78c2SRolf Eike Beer /** 29700fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 29710fcb78c2SRolf Eike Beer * @cpu: cpu to use 29720fcb78c2SRolf Eike Beer * @dwork: job to be done 29730fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 29740fcb78c2SRolf Eike Beer * 29750fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 29760fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 29770fcb78c2SRolf Eike Beer */ 2978d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 2979d4283e93STejun Heo unsigned long delay) 29801da177e4SLinus Torvalds { 2981d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 29821da177e4SLinus Torvalds } 2983ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 29841da177e4SLinus Torvalds 2985b6136773SAndrew Morton /** 29860a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 29870a13c00eSTejun Heo * @dwork: job to be done 29880a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 29890a13c00eSTejun Heo * 29900a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 29910a13c00eSTejun Heo * workqueue. 29920a13c00eSTejun Heo */ 2993d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 29940a13c00eSTejun Heo { 29950a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 29960a13c00eSTejun Heo } 29970a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 29980a13c00eSTejun Heo 29990a13c00eSTejun Heo /** 300031ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3001b6136773SAndrew Morton * @func: the function to call 3002b6136773SAndrew Morton * 300331ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 300431ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3005b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 300631ddd871STejun Heo * 300731ddd871STejun Heo * RETURNS: 300831ddd871STejun Heo * 0 on success, -errno on failure. 3009b6136773SAndrew Morton */ 301065f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 301115316ba8SChristoph Lameter { 301215316ba8SChristoph Lameter int cpu; 301338f51568SNamhyung Kim struct work_struct __percpu *works; 301415316ba8SChristoph Lameter 3015b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3016b6136773SAndrew Morton if (!works) 301715316ba8SChristoph Lameter return -ENOMEM; 3018b6136773SAndrew Morton 301995402b38SGautham R Shenoy get_online_cpus(); 302093981800STejun Heo 302115316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30229bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30239bfb1839SIngo Molnar 30249bfb1839SIngo Molnar INIT_WORK(work, func); 30258de6d308SOleg Nesterov schedule_work_on(cpu, work); 302615316ba8SChristoph Lameter } 302793981800STejun Heo 302893981800STejun Heo for_each_online_cpu(cpu) 30298616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 303093981800STejun Heo 303195402b38SGautham R Shenoy put_online_cpus(); 3032b6136773SAndrew Morton free_percpu(works); 303315316ba8SChristoph Lameter return 0; 303415316ba8SChristoph Lameter } 303515316ba8SChristoph Lameter 3036eef6a7d5SAlan Stern /** 3037eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3038eef6a7d5SAlan Stern * 3039eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3040eef6a7d5SAlan Stern * completion. 3041eef6a7d5SAlan Stern * 3042eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3043eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3044eef6a7d5SAlan Stern * will lead to deadlock: 3045eef6a7d5SAlan Stern * 3046eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3047eef6a7d5SAlan Stern * a lock held by your code or its caller. 3048eef6a7d5SAlan Stern * 3049eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3050eef6a7d5SAlan Stern * 3051eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3052eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3053eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3054eef6a7d5SAlan Stern * 3055eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3056eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3057eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3058eef6a7d5SAlan Stern * cancel_work_sync() instead. 3059eef6a7d5SAlan Stern */ 30601da177e4SLinus Torvalds void flush_scheduled_work(void) 30611da177e4SLinus Torvalds { 3062d320c038STejun Heo flush_workqueue(system_wq); 30631da177e4SLinus Torvalds } 3064ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30651da177e4SLinus Torvalds 30661da177e4SLinus Torvalds /** 30671fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30681fa44ecaSJames Bottomley * @fn: the function to execute 30691fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30701fa44ecaSJames Bottomley * be available when the work executes) 30711fa44ecaSJames Bottomley * 30721fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30731fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30741fa44ecaSJames Bottomley * 30751fa44ecaSJames Bottomley * Returns: 0 - function was executed 30761fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30771fa44ecaSJames Bottomley */ 307865f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30791fa44ecaSJames Bottomley { 30801fa44ecaSJames Bottomley if (!in_interrupt()) { 308165f27f38SDavid Howells fn(&ew->work); 30821fa44ecaSJames Bottomley return 0; 30831fa44ecaSJames Bottomley } 30841fa44ecaSJames Bottomley 308565f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30861fa44ecaSJames Bottomley schedule_work(&ew->work); 30871fa44ecaSJames Bottomley 30881fa44ecaSJames Bottomley return 1; 30891fa44ecaSJames Bottomley } 30901fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30911fa44ecaSJames Bottomley 30921da177e4SLinus Torvalds int keventd_up(void) 30931da177e4SLinus Torvalds { 3094d320c038STejun Heo return system_wq != NULL; 30951da177e4SLinus Torvalds } 30961da177e4SLinus Torvalds 3097112202d9STejun Heo static int alloc_pwqs(struct workqueue_struct *wq) 30981da177e4SLinus Torvalds { 30993af24433SOleg Nesterov /* 3100112202d9STejun Heo * pwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 31010f900049STejun Heo * Make sure that the alignment isn't lower than that of 31020f900049STejun Heo * unsigned long long. 31033af24433SOleg Nesterov */ 3104112202d9STejun Heo const size_t size = sizeof(struct pool_workqueue); 31050f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 31060f900049STejun Heo __alignof__(unsigned long long)); 31073af24433SOleg Nesterov 3108e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3109112202d9STejun Heo wq->pool_wq.pcpu = __alloc_percpu(size, align); 3110931ac77eSTejun Heo else { 31110f900049STejun Heo void *ptr; 3112e1d8aa9fSFrederic Weisbecker 31130f900049STejun Heo /* 3114112202d9STejun Heo * Allocate enough room to align pwq and put an extra 3115f3421797STejun Heo * pointer at the end pointing back to the originally 3116f3421797STejun Heo * allocated pointer which will be used for free. 31170f900049STejun Heo */ 3118bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3119bdbc5dd7STejun Heo if (ptr) { 3120112202d9STejun Heo wq->pool_wq.single = PTR_ALIGN(ptr, align); 3121112202d9STejun Heo *(void **)(wq->pool_wq.single + 1) = ptr; 3122bdbc5dd7STejun Heo } 31233af24433SOleg Nesterov } 31243af24433SOleg Nesterov 31250415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3126112202d9STejun Heo BUG_ON(!IS_ALIGNED(wq->pool_wq.v, align)); 3127112202d9STejun Heo return wq->pool_wq.v ? 0 : -ENOMEM; 31280f900049STejun Heo } 31290f900049STejun Heo 3130112202d9STejun Heo static void free_pwqs(struct workqueue_struct *wq) 313106ba38a9SOleg Nesterov { 3132e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3133112202d9STejun Heo free_percpu(wq->pool_wq.pcpu); 3134112202d9STejun Heo else if (wq->pool_wq.single) { 3135112202d9STejun Heo /* the pointer to free is stored right after the pwq */ 3136112202d9STejun Heo kfree(*(void **)(wq->pool_wq.single + 1)); 313706ba38a9SOleg Nesterov } 313806ba38a9SOleg Nesterov } 313906ba38a9SOleg Nesterov 3140f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3141f3421797STejun Heo const char *name) 3142b71ab8c2STejun Heo { 3143f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3144f3421797STejun Heo 3145f3421797STejun Heo if (max_active < 1 || max_active > lim) 3146044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3147f3421797STejun Heo max_active, name, 1, lim); 3148b71ab8c2STejun Heo 3149f3421797STejun Heo return clamp_val(max_active, 1, lim); 3150b71ab8c2STejun Heo } 3151b71ab8c2STejun Heo 3152b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 315397e37d7bSTejun Heo unsigned int flags, 31541e19ffc6STejun Heo int max_active, 3155eb13ba87SJohannes Berg struct lock_class_key *key, 3156b196be89STejun Heo const char *lock_name, ...) 31573af24433SOleg Nesterov { 3158b196be89STejun Heo va_list args, args1; 31593af24433SOleg Nesterov struct workqueue_struct *wq; 3160c34056a3STejun Heo unsigned int cpu; 3161b196be89STejun Heo size_t namelen; 3162b196be89STejun Heo 3163b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3164b196be89STejun Heo va_start(args, lock_name); 3165b196be89STejun Heo va_copy(args1, args); 3166b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3167b196be89STejun Heo 3168b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3169b196be89STejun Heo if (!wq) 3170b196be89STejun Heo goto err; 3171b196be89STejun Heo 3172b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3173b196be89STejun Heo va_end(args); 3174b196be89STejun Heo va_end(args1); 31753af24433SOleg Nesterov 3176f3421797STejun Heo /* 31776370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 31786370a6adSTejun Heo * have a rescuer to guarantee forward progress. 31796370a6adSTejun Heo */ 31806370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 31816370a6adSTejun Heo flags |= WQ_RESCUER; 31826370a6adSTejun Heo 3183d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3184b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 31853af24433SOleg Nesterov 3186b196be89STejun Heo /* init wq */ 318797e37d7bSTejun Heo wq->flags = flags; 3188a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 318973f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3190112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 319173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 319273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 31933af24433SOleg Nesterov 3194eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3195cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 31963af24433SOleg Nesterov 3197112202d9STejun Heo if (alloc_pwqs(wq) < 0) 3198bdbc5dd7STejun Heo goto err; 3199bdbc5dd7STejun Heo 3200112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 3201112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 32021537663fSTejun Heo 3203112202d9STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3204112202d9STejun Heo pwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI); 3205112202d9STejun Heo pwq->wq = wq; 3206112202d9STejun Heo pwq->flush_color = -1; 3207112202d9STejun Heo pwq->max_active = max_active; 3208112202d9STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3209e22bee78STejun Heo } 32101537663fSTejun Heo 3211e22bee78STejun Heo if (flags & WQ_RESCUER) { 3212e22bee78STejun Heo struct worker *rescuer; 3213e22bee78STejun Heo 3214f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3215e22bee78STejun Heo goto err; 3216e22bee78STejun Heo 3217e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3218e22bee78STejun Heo if (!rescuer) 3219e22bee78STejun Heo goto err; 3220e22bee78STejun Heo 3221111c225aSTejun Heo rescuer->rescue_wq = wq; 3222111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3223b196be89STejun Heo wq->name); 3224e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3225e22bee78STejun Heo goto err; 3226e22bee78STejun Heo 3227e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3228e22bee78STejun Heo wake_up_process(rescuer->task); 32293af24433SOleg Nesterov } 32301537663fSTejun Heo 32313af24433SOleg Nesterov /* 3232a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3233a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3234a0a1a5fdSTejun Heo * workqueue to workqueues list. 32353af24433SOleg Nesterov */ 32363af24433SOleg Nesterov spin_lock(&workqueue_lock); 3237a0a1a5fdSTejun Heo 323858a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3239112202d9STejun Heo for_each_pwq_cpu(cpu, wq) 3240112202d9STejun Heo get_pwq(cpu, wq)->max_active = 0; 3241a0a1a5fdSTejun Heo 32423af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3243a0a1a5fdSTejun Heo 32443af24433SOleg Nesterov spin_unlock(&workqueue_lock); 32453af24433SOleg Nesterov 32463af24433SOleg Nesterov return wq; 32474690c4abSTejun Heo err: 32484690c4abSTejun Heo if (wq) { 3249112202d9STejun Heo free_pwqs(wq); 3250f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3251e22bee78STejun Heo kfree(wq->rescuer); 32524690c4abSTejun Heo kfree(wq); 32533af24433SOleg Nesterov } 32544690c4abSTejun Heo return NULL; 32551da177e4SLinus Torvalds } 3256d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 32571da177e4SLinus Torvalds 32583af24433SOleg Nesterov /** 32593af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 32603af24433SOleg Nesterov * @wq: target workqueue 32613af24433SOleg Nesterov * 32623af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 32633af24433SOleg Nesterov */ 32643af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 32653af24433SOleg Nesterov { 3266c8e55f36STejun Heo unsigned int cpu; 32673af24433SOleg Nesterov 32689c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 32699c5a2ba7STejun Heo drain_workqueue(wq); 3270c8efcc25STejun Heo 32716183c009STejun Heo /* sanity checks */ 32726183c009STejun Heo for_each_pwq_cpu(cpu, wq) { 32736183c009STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 32746183c009STejun Heo int i; 32756183c009STejun Heo 32766183c009STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 32776183c009STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) 32786183c009STejun Heo return; 32796183c009STejun Heo if (WARN_ON(pwq->nr_active) || 32806183c009STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) 32816183c009STejun Heo return; 32826183c009STejun Heo } 32836183c009STejun Heo 3284a0a1a5fdSTejun Heo /* 3285a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3286a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3287a0a1a5fdSTejun Heo */ 328895402b38SGautham R Shenoy spin_lock(&workqueue_lock); 32893af24433SOleg Nesterov list_del(&wq->list); 329095402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 32913af24433SOleg Nesterov 3292e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3293e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3294f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 32958d9df9f0SXiaotian Feng kfree(wq->rescuer); 3296e22bee78STejun Heo } 3297e22bee78STejun Heo 3298112202d9STejun Heo free_pwqs(wq); 32993af24433SOleg Nesterov kfree(wq); 33003af24433SOleg Nesterov } 33013af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33023af24433SOleg Nesterov 3303dcd989cbSTejun Heo /** 3304112202d9STejun Heo * pwq_set_max_active - adjust max_active of a pwq 3305112202d9STejun Heo * @pwq: target pool_workqueue 33069f4bd4cdSLai Jiangshan * @max_active: new max_active value. 33079f4bd4cdSLai Jiangshan * 3308112202d9STejun Heo * Set @pwq->max_active to @max_active and activate delayed works if 33099f4bd4cdSLai Jiangshan * increased. 33109f4bd4cdSLai Jiangshan * 33119f4bd4cdSLai Jiangshan * CONTEXT: 3312d565ed63STejun Heo * spin_lock_irq(pool->lock). 33139f4bd4cdSLai Jiangshan */ 3314112202d9STejun Heo static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active) 33159f4bd4cdSLai Jiangshan { 3316112202d9STejun Heo pwq->max_active = max_active; 33179f4bd4cdSLai Jiangshan 3318112202d9STejun Heo while (!list_empty(&pwq->delayed_works) && 3319112202d9STejun Heo pwq->nr_active < pwq->max_active) 3320112202d9STejun Heo pwq_activate_first_delayed(pwq); 33219f4bd4cdSLai Jiangshan } 33229f4bd4cdSLai Jiangshan 33239f4bd4cdSLai Jiangshan /** 3324dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3325dcd989cbSTejun Heo * @wq: target workqueue 3326dcd989cbSTejun Heo * @max_active: new max_active value. 3327dcd989cbSTejun Heo * 3328dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3329dcd989cbSTejun Heo * 3330dcd989cbSTejun Heo * CONTEXT: 3331dcd989cbSTejun Heo * Don't call from IRQ context. 3332dcd989cbSTejun Heo */ 3333dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3334dcd989cbSTejun Heo { 3335dcd989cbSTejun Heo unsigned int cpu; 3336dcd989cbSTejun Heo 3337f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3338dcd989cbSTejun Heo 3339dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3340dcd989cbSTejun Heo 3341dcd989cbSTejun Heo wq->saved_max_active = max_active; 3342dcd989cbSTejun Heo 3343112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 3344112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3345112202d9STejun Heo struct worker_pool *pool = pwq->pool; 3346dcd989cbSTejun Heo 3347d565ed63STejun Heo spin_lock_irq(&pool->lock); 3348dcd989cbSTejun Heo 334958a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 335035b6bb63STejun Heo !(pool->flags & POOL_FREEZING)) 3351112202d9STejun Heo pwq_set_max_active(pwq, max_active); 3352dcd989cbSTejun Heo 3353d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3354dcd989cbSTejun Heo } 3355dcd989cbSTejun Heo 3356dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3357dcd989cbSTejun Heo } 3358dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3359dcd989cbSTejun Heo 3360dcd989cbSTejun Heo /** 3361dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3362dcd989cbSTejun Heo * @cpu: CPU in question 3363dcd989cbSTejun Heo * @wq: target workqueue 3364dcd989cbSTejun Heo * 3365dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3366dcd989cbSTejun Heo * no synchronization around this function and the test result is 3367dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3368dcd989cbSTejun Heo * 3369dcd989cbSTejun Heo * RETURNS: 3370dcd989cbSTejun Heo * %true if congested, %false otherwise. 3371dcd989cbSTejun Heo */ 3372dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3373dcd989cbSTejun Heo { 3374112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3375dcd989cbSTejun Heo 3376112202d9STejun Heo return !list_empty(&pwq->delayed_works); 3377dcd989cbSTejun Heo } 3378dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3379dcd989cbSTejun Heo 3380dcd989cbSTejun Heo /** 3381dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3382dcd989cbSTejun Heo * @work: the work to be tested 3383dcd989cbSTejun Heo * 3384dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3385dcd989cbSTejun Heo * synchronization around this function and the test result is 3386dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3387dcd989cbSTejun Heo * 3388dcd989cbSTejun Heo * RETURNS: 3389dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3390dcd989cbSTejun Heo */ 3391dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3392dcd989cbSTejun Heo { 3393c9e7cf27STejun Heo struct worker_pool *pool = get_work_pool(work); 3394dcd989cbSTejun Heo unsigned long flags; 3395dcd989cbSTejun Heo unsigned int ret = 0; 3396dcd989cbSTejun Heo 3397dcd989cbSTejun Heo if (work_pending(work)) 3398dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3399038366c5SLai Jiangshan 3400038366c5SLai Jiangshan if (pool) { 3401038366c5SLai Jiangshan spin_lock_irqsave(&pool->lock, flags); 3402c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3403dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3404d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 3405038366c5SLai Jiangshan } 3406dcd989cbSTejun Heo 3407dcd989cbSTejun Heo return ret; 3408dcd989cbSTejun Heo } 3409dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3410dcd989cbSTejun Heo 3411db7bccf4STejun Heo /* 3412db7bccf4STejun Heo * CPU hotplug. 3413db7bccf4STejun Heo * 3414e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3415112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 3416706026c2STejun Heo * pool which make migrating pending and scheduled works very 3417e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 341894cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 3419e22bee78STejun Heo * blocked draining impractical. 3420e22bee78STejun Heo * 342124647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 3422628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3423628c78e7STejun Heo * cpu comes back online. 3424db7bccf4STejun Heo */ 3425db7bccf4STejun Heo 3426706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 3427db7bccf4STejun Heo { 342838db41d9STejun Heo int cpu = smp_processor_id(); 34294ce62e9eSTejun Heo struct worker_pool *pool; 3430db7bccf4STejun Heo struct worker *worker; 3431db7bccf4STejun Heo int i; 3432db7bccf4STejun Heo 343338db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 34346183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 3435db7bccf4STejun Heo 343694cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 343794cf58bbSTejun Heo spin_lock_irq(&pool->lock); 3438e22bee78STejun Heo 3439f2d5a0eeSTejun Heo /* 344094cf58bbSTejun Heo * We've claimed all manager positions. Make all workers 344194cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 344294cf58bbSTejun Heo * except for the ones which are still executing works from 344394cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 344494cf58bbSTejun Heo * this, they may become diasporas. 3445f2d5a0eeSTejun Heo */ 34464ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3447403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3448db7bccf4STejun Heo 3449b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) 3450403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3451db7bccf4STejun Heo 345224647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 3453f2d5a0eeSTejun Heo 345494cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 345594cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 345694cf58bbSTejun Heo } 3457e22bee78STejun Heo 3458e22bee78STejun Heo /* 3459628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3460628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3461628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3462628c78e7STejun Heo */ 3463628c78e7STejun Heo schedule(); 3464628c78e7STejun Heo 3465628c78e7STejun Heo /* 3466628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3467628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 346838db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 346938db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 347038db41d9STejun Heo * pools which are served by workers tied to the CPU. 3471628c78e7STejun Heo * 3472628c78e7STejun Heo * On return from this function, the current worker would trigger 3473628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3474628c78e7STejun Heo * didn't already. 3475e22bee78STejun Heo */ 347638db41d9STejun Heo for_each_std_worker_pool(pool, cpu) 3477e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 3478db7bccf4STejun Heo } 3479db7bccf4STejun Heo 34808db25e78STejun Heo /* 34818db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 34828db25e78STejun Heo * This will be registered high priority CPU notifier. 34838db25e78STejun Heo */ 34849fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 34851da177e4SLinus Torvalds unsigned long action, 34861da177e4SLinus Torvalds void *hcpu) 34871da177e4SLinus Torvalds { 34883af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 34894ce62e9eSTejun Heo struct worker_pool *pool; 34901da177e4SLinus Torvalds 34918db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 34923af24433SOleg Nesterov case CPU_UP_PREPARE: 349338db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 34943ce63377STejun Heo struct worker *worker; 34953ce63377STejun Heo 34963ce63377STejun Heo if (pool->nr_workers) 34973ce63377STejun Heo continue; 34983ce63377STejun Heo 34993ce63377STejun Heo worker = create_worker(pool); 35003ce63377STejun Heo if (!worker) 35013ce63377STejun Heo return NOTIFY_BAD; 35023ce63377STejun Heo 3503d565ed63STejun Heo spin_lock_irq(&pool->lock); 35043ce63377STejun Heo start_worker(worker); 3505d565ed63STejun Heo spin_unlock_irq(&pool->lock); 35063af24433SOleg Nesterov } 35071da177e4SLinus Torvalds break; 35081da177e4SLinus Torvalds 350965758202STejun Heo case CPU_DOWN_FAILED: 351065758202STejun Heo case CPU_ONLINE: 351138db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 351294cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 351394cf58bbSTejun Heo spin_lock_irq(&pool->lock); 351494cf58bbSTejun Heo 351524647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 351694cf58bbSTejun Heo rebind_workers(pool); 351794cf58bbSTejun Heo 351894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 351994cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 352094cf58bbSTejun Heo } 35218db25e78STejun Heo break; 352265758202STejun Heo } 352365758202STejun Heo return NOTIFY_OK; 352465758202STejun Heo } 352565758202STejun Heo 352665758202STejun Heo /* 352765758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 352865758202STejun Heo * This will be registered as low priority CPU notifier. 352965758202STejun Heo */ 35309fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 353165758202STejun Heo unsigned long action, 353265758202STejun Heo void *hcpu) 353365758202STejun Heo { 35348db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 35358db25e78STejun Heo struct work_struct unbind_work; 35368db25e78STejun Heo 353765758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 353865758202STejun Heo case CPU_DOWN_PREPARE: 35398db25e78STejun Heo /* unbinding should happen on the local CPU */ 3540706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 35417635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 35428db25e78STejun Heo flush_work(&unbind_work); 35438db25e78STejun Heo break; 354465758202STejun Heo } 354565758202STejun Heo return NOTIFY_OK; 354665758202STejun Heo } 354765758202STejun Heo 35482d3854a3SRusty Russell #ifdef CONFIG_SMP 35498ccad40dSRusty Russell 35502d3854a3SRusty Russell struct work_for_cpu { 3551ed48ece2STejun Heo struct work_struct work; 35522d3854a3SRusty Russell long (*fn)(void *); 35532d3854a3SRusty Russell void *arg; 35542d3854a3SRusty Russell long ret; 35552d3854a3SRusty Russell }; 35562d3854a3SRusty Russell 3557ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 35582d3854a3SRusty Russell { 3559ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3560ed48ece2STejun Heo 35612d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 35622d3854a3SRusty Russell } 35632d3854a3SRusty Russell 35642d3854a3SRusty Russell /** 35652d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 35662d3854a3SRusty Russell * @cpu: the cpu to run on 35672d3854a3SRusty Russell * @fn: the function to run 35682d3854a3SRusty Russell * @arg: the function arg 35692d3854a3SRusty Russell * 357031ad9081SRusty Russell * This will return the value @fn returns. 357131ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 35726b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 35732d3854a3SRusty Russell */ 35742d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 35752d3854a3SRusty Russell { 3576ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 35772d3854a3SRusty Russell 3578ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3579ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3580ed48ece2STejun Heo flush_work(&wfc.work); 35812d3854a3SRusty Russell return wfc.ret; 35822d3854a3SRusty Russell } 35832d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 35842d3854a3SRusty Russell #endif /* CONFIG_SMP */ 35852d3854a3SRusty Russell 3586a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3587e7577c50SRusty Russell 3588a0a1a5fdSTejun Heo /** 3589a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3590a0a1a5fdSTejun Heo * 359158a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 359258a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 3593706026c2STejun Heo * pool->worklist. 3594a0a1a5fdSTejun Heo * 3595a0a1a5fdSTejun Heo * CONTEXT: 3596d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3597a0a1a5fdSTejun Heo */ 3598a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3599a0a1a5fdSTejun Heo { 3600a0a1a5fdSTejun Heo unsigned int cpu; 3601a0a1a5fdSTejun Heo 3602a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3603a0a1a5fdSTejun Heo 36046183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 3605a0a1a5fdSTejun Heo workqueue_freezing = true; 3606a0a1a5fdSTejun Heo 3607706026c2STejun Heo for_each_wq_cpu(cpu) { 360835b6bb63STejun Heo struct worker_pool *pool; 3609bdbc5dd7STejun Heo struct workqueue_struct *wq; 36108b03ae3cSTejun Heo 361138db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3612a1056305STejun Heo spin_lock_irq(&pool->lock); 3613d565ed63STejun Heo 361435b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 361535b6bb63STejun Heo pool->flags |= POOL_FREEZING; 3616db7bccf4STejun Heo 3617a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3618112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3619a0a1a5fdSTejun Heo 3620112202d9STejun Heo if (pwq && pwq->pool == pool && 3621a1056305STejun Heo (wq->flags & WQ_FREEZABLE)) 3622112202d9STejun Heo pwq->max_active = 0; 36231da177e4SLinus Torvalds } 36248b03ae3cSTejun Heo 3625a1056305STejun Heo spin_unlock_irq(&pool->lock); 3626a1056305STejun Heo } 3627a0a1a5fdSTejun Heo } 3628a0a1a5fdSTejun Heo 3629a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3630a0a1a5fdSTejun Heo } 3631a0a1a5fdSTejun Heo 3632a0a1a5fdSTejun Heo /** 363358a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3634a0a1a5fdSTejun Heo * 3635a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3636a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3637a0a1a5fdSTejun Heo * 3638a0a1a5fdSTejun Heo * CONTEXT: 3639a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3640a0a1a5fdSTejun Heo * 3641a0a1a5fdSTejun Heo * RETURNS: 364258a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 364358a69cb4STejun Heo * is complete. 3644a0a1a5fdSTejun Heo */ 3645a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3646a0a1a5fdSTejun Heo { 3647a0a1a5fdSTejun Heo unsigned int cpu; 3648a0a1a5fdSTejun Heo bool busy = false; 3649a0a1a5fdSTejun Heo 3650a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3651a0a1a5fdSTejun Heo 36526183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 3653a0a1a5fdSTejun Heo 3654706026c2STejun Heo for_each_wq_cpu(cpu) { 3655bdbc5dd7STejun Heo struct workqueue_struct *wq; 3656a0a1a5fdSTejun Heo /* 3657a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3658a0a1a5fdSTejun Heo * to peek without lock. 3659a0a1a5fdSTejun Heo */ 3660a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3661112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3662a0a1a5fdSTejun Heo 3663112202d9STejun Heo if (!pwq || !(wq->flags & WQ_FREEZABLE)) 3664a0a1a5fdSTejun Heo continue; 3665a0a1a5fdSTejun Heo 36666183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 3667112202d9STejun Heo if (pwq->nr_active) { 3668a0a1a5fdSTejun Heo busy = true; 3669a0a1a5fdSTejun Heo goto out_unlock; 3670a0a1a5fdSTejun Heo } 3671a0a1a5fdSTejun Heo } 3672a0a1a5fdSTejun Heo } 3673a0a1a5fdSTejun Heo out_unlock: 3674a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3675a0a1a5fdSTejun Heo return busy; 3676a0a1a5fdSTejun Heo } 3677a0a1a5fdSTejun Heo 3678a0a1a5fdSTejun Heo /** 3679a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3680a0a1a5fdSTejun Heo * 3681a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 3682706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 3683a0a1a5fdSTejun Heo * 3684a0a1a5fdSTejun Heo * CONTEXT: 3685d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3686a0a1a5fdSTejun Heo */ 3687a0a1a5fdSTejun Heo void thaw_workqueues(void) 3688a0a1a5fdSTejun Heo { 3689a0a1a5fdSTejun Heo unsigned int cpu; 3690a0a1a5fdSTejun Heo 3691a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3692a0a1a5fdSTejun Heo 3693a0a1a5fdSTejun Heo if (!workqueue_freezing) 3694a0a1a5fdSTejun Heo goto out_unlock; 3695a0a1a5fdSTejun Heo 3696706026c2STejun Heo for_each_wq_cpu(cpu) { 36974ce62e9eSTejun Heo struct worker_pool *pool; 3698bdbc5dd7STejun Heo struct workqueue_struct *wq; 36998b03ae3cSTejun Heo 370038db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3701a1056305STejun Heo spin_lock_irq(&pool->lock); 3702d565ed63STejun Heo 370335b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 370435b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 3705db7bccf4STejun Heo 3706a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3707112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3708a0a1a5fdSTejun Heo 3709112202d9STejun Heo if (!pwq || pwq->pool != pool || 3710a1056305STejun Heo !(wq->flags & WQ_FREEZABLE)) 3711a0a1a5fdSTejun Heo continue; 3712a0a1a5fdSTejun Heo 3713a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3714112202d9STejun Heo pwq_set_max_active(pwq, wq->saved_max_active); 3715a0a1a5fdSTejun Heo } 37168b03ae3cSTejun Heo 37174ce62e9eSTejun Heo wake_up_worker(pool); 3718a1056305STejun Heo 3719a1056305STejun Heo spin_unlock_irq(&pool->lock); 3720d565ed63STejun Heo } 3721a0a1a5fdSTejun Heo } 3722a0a1a5fdSTejun Heo 3723a0a1a5fdSTejun Heo workqueue_freezing = false; 3724a0a1a5fdSTejun Heo out_unlock: 3725a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3726a0a1a5fdSTejun Heo } 3727a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3728a0a1a5fdSTejun Heo 37296ee0578bSSuresh Siddha static int __init init_workqueues(void) 37301da177e4SLinus Torvalds { 3731c34056a3STejun Heo unsigned int cpu; 3732c34056a3STejun Heo 37337c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 37347c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 37356be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 3736b5490077STejun Heo 373765758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3738a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 37398b03ae3cSTejun Heo 3740706026c2STejun Heo /* initialize CPU pools */ 3741706026c2STejun Heo for_each_wq_cpu(cpu) { 37424ce62e9eSTejun Heo struct worker_pool *pool; 37438b03ae3cSTejun Heo 374438db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3745d565ed63STejun Heo spin_lock_init(&pool->lock); 3746ec22ca5eSTejun Heo pool->cpu = cpu; 374724647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 37484ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37494ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3750c9e7cf27STejun Heo hash_init(pool->busy_hash); 3751e22bee78STejun Heo 37524ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 37534ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 37544ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3755e22bee78STejun Heo 3756706026c2STejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 37574ce62e9eSTejun Heo (unsigned long)pool); 37584ce62e9eSTejun Heo 3759b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 37604ce62e9eSTejun Heo ida_init(&pool->worker_ida); 37619daf9e67STejun Heo 37629daf9e67STejun Heo /* alloc pool ID */ 37639daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 37644ce62e9eSTejun Heo } 37658b03ae3cSTejun Heo } 37668b03ae3cSTejun Heo 3767e22bee78STejun Heo /* create the initial worker */ 3768706026c2STejun Heo for_each_online_wq_cpu(cpu) { 37694ce62e9eSTejun Heo struct worker_pool *pool; 3770e22bee78STejun Heo 377138db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 37724ce62e9eSTejun Heo struct worker *worker; 37734ce62e9eSTejun Heo 377424647570STejun Heo if (cpu != WORK_CPU_UNBOUND) 377524647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 377624647570STejun Heo 3777bc2ae0f5STejun Heo worker = create_worker(pool); 3778e22bee78STejun Heo BUG_ON(!worker); 3779d565ed63STejun Heo spin_lock_irq(&pool->lock); 3780e22bee78STejun Heo start_worker(worker); 3781d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3782e22bee78STejun Heo } 37834ce62e9eSTejun Heo } 3784e22bee78STejun Heo 3785d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 37861aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3787d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3788f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3789f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 379024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 379124d51addSTejun Heo WQ_FREEZABLE, 0); 37921aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3793ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 37946ee0578bSSuresh Siddha return 0; 37951da177e4SLinus Torvalds } 37966ee0578bSSuresh Siddha early_initcall(init_workqueues); 3797