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 */ 1478b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1488b03ae3cSTejun Heo 1498b03ae3cSTejun Heo /* 150502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1510f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1520f900049STejun Heo * aligned at two's power of the number of flag bits. 1531da177e4SLinus Torvalds */ 1541da177e4SLinus Torvalds struct cpu_workqueue_struct { 155bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1564690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 15773f53c4aSTejun Heo int work_color; /* L: current color */ 15873f53c4aSTejun Heo int flush_color; /* L: flushing color */ 15973f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 16073f53c4aSTejun Heo /* L: nr of in_flight works */ 1611e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 162a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1631e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1640f900049STejun Heo }; 1651da177e4SLinus Torvalds 1661da177e4SLinus Torvalds /* 16773f53c4aSTejun Heo * Structure used to wait for workqueue flush. 16873f53c4aSTejun Heo */ 16973f53c4aSTejun Heo struct wq_flusher { 17073f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 17173f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 17273f53c4aSTejun Heo struct completion done; /* flush completion */ 17373f53c4aSTejun Heo }; 1741da177e4SLinus Torvalds 17573f53c4aSTejun Heo /* 176f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 177f2e005aaSTejun Heo * used to determine whether there's something to be done. 178f2e005aaSTejun Heo */ 179f2e005aaSTejun Heo #ifdef CONFIG_SMP 180f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 181f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 182f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 183f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 184f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 1859c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 186f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 187f2e005aaSTejun Heo #else 188f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 189f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 190f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 191f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 192f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 193f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 194f2e005aaSTejun Heo #endif 1951da177e4SLinus Torvalds 1961da177e4SLinus Torvalds /* 1971da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 1981da177e4SLinus Torvalds * per-CPU workqueues: 1991da177e4SLinus Torvalds */ 2001da177e4SLinus Torvalds struct workqueue_struct { 2019c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 202bdbc5dd7STejun Heo union { 203bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 204bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 205bdbc5dd7STejun Heo unsigned long v; 206bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2074690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 20873f53c4aSTejun Heo 20973f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 21073f53c4aSTejun Heo int work_color; /* F: current work color */ 21173f53c4aSTejun Heo int flush_color; /* F: current flush color */ 21273f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 21373f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 21473f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 21573f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 21673f53c4aSTejun Heo 217f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 218e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 219e22bee78STejun Heo 2209c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 221dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2224e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2234e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2244e6045f1SJohannes Berg #endif 225b196be89STejun Heo char name[]; /* I: workqueue name */ 2261da177e4SLinus Torvalds }; 2271da177e4SLinus Torvalds 228d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 229d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 230044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2311aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 232044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 233d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 234044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 235f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 236044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 23724d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 238d320c038STejun Heo 23997bd2347STejun Heo #define CREATE_TRACE_POINTS 24097bd2347STejun Heo #include <trace/events/workqueue.h> 24197bd2347STejun Heo 24238db41d9STejun Heo #define for_each_std_worker_pool(pool, cpu) \ 243a60dc39cSTejun Heo for ((pool) = &std_worker_pools(cpu)[0]; \ 244a60dc39cSTejun Heo (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++) 2454ce62e9eSTejun Heo 246c9e7cf27STejun Heo #define for_each_busy_worker(worker, i, pos, pool) \ 247c9e7cf27STejun Heo hash_for_each(pool->busy_hash, i, pos, worker, hentry) 248db7bccf4STejun Heo 249706026c2STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 250f3421797STejun Heo unsigned int sw) 251f3421797STejun Heo { 252f3421797STejun Heo if (cpu < nr_cpu_ids) { 253f3421797STejun Heo if (sw & 1) { 254f3421797STejun Heo cpu = cpumask_next(cpu, mask); 255f3421797STejun Heo if (cpu < nr_cpu_ids) 256f3421797STejun Heo return cpu; 257f3421797STejun Heo } 258f3421797STejun Heo if (sw & 2) 259f3421797STejun Heo return WORK_CPU_UNBOUND; 260f3421797STejun Heo } 2616be19588SLai Jiangshan return WORK_CPU_END; 262f3421797STejun Heo } 263f3421797STejun Heo 264706026c2STejun Heo static inline int __next_cwq_cpu(int cpu, const struct cpumask *mask, 265f3421797STejun Heo struct workqueue_struct *wq) 266f3421797STejun Heo { 267706026c2STejun Heo return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 268f3421797STejun Heo } 269f3421797STejun Heo 27009884951STejun Heo /* 27109884951STejun Heo * CPU iterators 27209884951STejun Heo * 273706026c2STejun Heo * An extra cpu number is defined using an invalid cpu number 27409884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 275706026c2STejun Heo * specific CPU. The following iterators are similar to for_each_*_cpu() 276706026c2STejun Heo * iterators but also considers the unbound CPU. 27709884951STejun Heo * 278706026c2STejun Heo * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND 279706026c2STejun Heo * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND 28009884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 28109884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 28209884951STejun Heo */ 283706026c2STejun Heo #define for_each_wq_cpu(cpu) \ 284706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \ 2856be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 286706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3)) 287f3421797STejun Heo 288706026c2STejun Heo #define for_each_online_wq_cpu(cpu) \ 289706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \ 2906be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 291706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3)) 292f3421797STejun Heo 293f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 294706026c2STejun Heo for ((cpu) = __next_cwq_cpu(-1, cpu_possible_mask, (wq)); \ 2956be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 296706026c2STejun Heo (cpu) = __next_cwq_cpu((cpu), cpu_possible_mask, (wq))) 297f3421797STejun Heo 298dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 299dc186ad7SThomas Gleixner 300dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 301dc186ad7SThomas Gleixner 30299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 30399777288SStanislaw Gruszka { 30499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 30599777288SStanislaw Gruszka } 30699777288SStanislaw Gruszka 307dc186ad7SThomas Gleixner /* 308dc186ad7SThomas Gleixner * fixup_init is called when: 309dc186ad7SThomas Gleixner * - an active object is initialized 310dc186ad7SThomas Gleixner */ 311dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 312dc186ad7SThomas Gleixner { 313dc186ad7SThomas Gleixner struct work_struct *work = addr; 314dc186ad7SThomas Gleixner 315dc186ad7SThomas Gleixner switch (state) { 316dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 317dc186ad7SThomas Gleixner cancel_work_sync(work); 318dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 319dc186ad7SThomas Gleixner return 1; 320dc186ad7SThomas Gleixner default: 321dc186ad7SThomas Gleixner return 0; 322dc186ad7SThomas Gleixner } 323dc186ad7SThomas Gleixner } 324dc186ad7SThomas Gleixner 325dc186ad7SThomas Gleixner /* 326dc186ad7SThomas Gleixner * fixup_activate is called when: 327dc186ad7SThomas Gleixner * - an active object is activated 328dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 329dc186ad7SThomas Gleixner */ 330dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 331dc186ad7SThomas Gleixner { 332dc186ad7SThomas Gleixner struct work_struct *work = addr; 333dc186ad7SThomas Gleixner 334dc186ad7SThomas Gleixner switch (state) { 335dc186ad7SThomas Gleixner 336dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 337dc186ad7SThomas Gleixner /* 338dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 339dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 340dc186ad7SThomas Gleixner * is tracked in the object tracker. 341dc186ad7SThomas Gleixner */ 34222df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 343dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 344dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 345dc186ad7SThomas Gleixner return 0; 346dc186ad7SThomas Gleixner } 347dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 348dc186ad7SThomas Gleixner return 0; 349dc186ad7SThomas Gleixner 350dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 351dc186ad7SThomas Gleixner WARN_ON(1); 352dc186ad7SThomas Gleixner 353dc186ad7SThomas Gleixner default: 354dc186ad7SThomas Gleixner return 0; 355dc186ad7SThomas Gleixner } 356dc186ad7SThomas Gleixner } 357dc186ad7SThomas Gleixner 358dc186ad7SThomas Gleixner /* 359dc186ad7SThomas Gleixner * fixup_free is called when: 360dc186ad7SThomas Gleixner * - an active object is freed 361dc186ad7SThomas Gleixner */ 362dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 363dc186ad7SThomas Gleixner { 364dc186ad7SThomas Gleixner struct work_struct *work = addr; 365dc186ad7SThomas Gleixner 366dc186ad7SThomas Gleixner switch (state) { 367dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 368dc186ad7SThomas Gleixner cancel_work_sync(work); 369dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 370dc186ad7SThomas Gleixner return 1; 371dc186ad7SThomas Gleixner default: 372dc186ad7SThomas Gleixner return 0; 373dc186ad7SThomas Gleixner } 374dc186ad7SThomas Gleixner } 375dc186ad7SThomas Gleixner 376dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 377dc186ad7SThomas Gleixner .name = "work_struct", 37899777288SStanislaw Gruszka .debug_hint = work_debug_hint, 379dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 380dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 381dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 382dc186ad7SThomas Gleixner }; 383dc186ad7SThomas Gleixner 384dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 385dc186ad7SThomas Gleixner { 386dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 387dc186ad7SThomas Gleixner } 388dc186ad7SThomas Gleixner 389dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 390dc186ad7SThomas Gleixner { 391dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 392dc186ad7SThomas Gleixner } 393dc186ad7SThomas Gleixner 394dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 395dc186ad7SThomas Gleixner { 396dc186ad7SThomas Gleixner if (onstack) 397dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 398dc186ad7SThomas Gleixner else 399dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 400dc186ad7SThomas Gleixner } 401dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 402dc186ad7SThomas Gleixner 403dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 404dc186ad7SThomas Gleixner { 405dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 406dc186ad7SThomas Gleixner } 407dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 408dc186ad7SThomas Gleixner 409dc186ad7SThomas Gleixner #else 410dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 411dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 412dc186ad7SThomas Gleixner #endif 413dc186ad7SThomas Gleixner 41495402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 41595402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4161da177e4SLinus Torvalds static LIST_HEAD(workqueues); 417a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4181da177e4SLinus Torvalds 41914441960SOleg Nesterov /* 420a60dc39cSTejun Heo * The CPU standard worker pools. nr_running is the only field which is 421a60dc39cSTejun Heo * expected to be used frequently by other cpus via try_to_wake_up(). Put 422a60dc39cSTejun Heo * it in a separate cacheline. 42314441960SOleg Nesterov */ 424a60dc39cSTejun Heo static DEFINE_PER_CPU(struct worker_pool [NR_STD_WORKER_POOLS], 425a60dc39cSTejun Heo cpu_std_worker_pools); 426e6e380edSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t [NR_STD_WORKER_POOLS], 427e6e380edSTejun Heo cpu_std_pool_nr_running); 428f756d5e2SNathan Lynch 429f3421797STejun Heo /* 430a60dc39cSTejun Heo * Standard worker pools and nr_running counter for unbound CPU. The pools 431a60dc39cSTejun Heo * have POOL_DISASSOCIATED set, and all workers have WORKER_UNBOUND set. 432f3421797STejun Heo */ 433a60dc39cSTejun Heo static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS]; 434e6e380edSTejun Heo static atomic_t unbound_std_pool_nr_running[NR_STD_WORKER_POOLS] = { 435e34cdddbSTejun Heo [0 ... NR_STD_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */ 4364ce62e9eSTejun Heo }; 437f3421797STejun Heo 4389daf9e67STejun Heo /* idr of all pools */ 4399daf9e67STejun Heo static DEFINE_MUTEX(worker_pool_idr_mutex); 4409daf9e67STejun Heo static DEFINE_IDR(worker_pool_idr); 4419daf9e67STejun Heo 442c34056a3STejun Heo static int worker_thread(void *__worker); 4431da177e4SLinus Torvalds 444a60dc39cSTejun Heo static struct worker_pool *std_worker_pools(int cpu) 4451da177e4SLinus Torvalds { 446f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 447a60dc39cSTejun Heo return per_cpu(cpu_std_worker_pools, cpu); 448f3421797STejun Heo else 449a60dc39cSTejun Heo return unbound_std_worker_pools; 4501da177e4SLinus Torvalds } 4511da177e4SLinus Torvalds 4524e8f0a60STejun Heo static int std_worker_pool_pri(struct worker_pool *pool) 4534e8f0a60STejun Heo { 454a60dc39cSTejun Heo return pool - std_worker_pools(pool->cpu); 4554e8f0a60STejun Heo } 4564e8f0a60STejun Heo 4579daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4589daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4599daf9e67STejun Heo { 4609daf9e67STejun Heo int ret; 4619daf9e67STejun Heo 4629daf9e67STejun Heo mutex_lock(&worker_pool_idr_mutex); 4639daf9e67STejun Heo idr_pre_get(&worker_pool_idr, GFP_KERNEL); 4649daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 4659daf9e67STejun Heo mutex_unlock(&worker_pool_idr_mutex); 4669daf9e67STejun Heo 4679daf9e67STejun Heo return ret; 4689daf9e67STejun Heo } 4699daf9e67STejun Heo 4707c3eed5cSTejun Heo /* 4717c3eed5cSTejun Heo * Lookup worker_pool by id. The idr currently is built during boot and 4727c3eed5cSTejun Heo * never modified. Don't worry about locking for now. 4737c3eed5cSTejun Heo */ 4747c3eed5cSTejun Heo static struct worker_pool *worker_pool_by_id(int pool_id) 4757c3eed5cSTejun Heo { 4767c3eed5cSTejun Heo return idr_find(&worker_pool_idr, pool_id); 4777c3eed5cSTejun Heo } 4787c3eed5cSTejun Heo 479d565ed63STejun Heo static struct worker_pool *get_std_worker_pool(int cpu, bool highpri) 480d565ed63STejun Heo { 481a60dc39cSTejun Heo struct worker_pool *pools = std_worker_pools(cpu); 482d565ed63STejun Heo 483a60dc39cSTejun Heo return &pools[highpri]; 484d565ed63STejun Heo } 485d565ed63STejun Heo 48663d95a91STejun Heo static atomic_t *get_pool_nr_running(struct worker_pool *pool) 487b1f4ec17SOleg Nesterov { 488ec22ca5eSTejun Heo int cpu = pool->cpu; 489e34cdddbSTejun Heo int idx = std_worker_pool_pri(pool); 49063d95a91STejun Heo 491f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 492e6e380edSTejun Heo return &per_cpu(cpu_std_pool_nr_running, cpu)[idx]; 493f3421797STejun Heo else 494e6e380edSTejun Heo return &unbound_std_pool_nr_running[idx]; 495b1f4ec17SOleg Nesterov } 496b1f4ec17SOleg Nesterov 4974690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 4984690c4abSTejun Heo struct workqueue_struct *wq) 499a848e3b6SOleg Nesterov { 500f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 501e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 502bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 503f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 504f3421797STejun Heo return wq->cpu_wq.single; 505f3421797STejun Heo return NULL; 506f3421797STejun Heo } 507a848e3b6SOleg Nesterov 50873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 50973f53c4aSTejun Heo { 51073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 51173f53c4aSTejun Heo } 51273f53c4aSTejun Heo 51373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 51473f53c4aSTejun Heo { 51573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 51673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 51773f53c4aSTejun Heo } 51873f53c4aSTejun Heo 51973f53c4aSTejun Heo static int work_next_color(int color) 52073f53c4aSTejun Heo { 52173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 5221da177e4SLinus Torvalds } 5231da177e4SLinus Torvalds 5244594bf15SDavid Howells /* 525b5490077STejun Heo * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data 526b5490077STejun Heo * contain the pointer to the queued cwq. Once execution starts, the flag 5277c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5287a22ad75STejun Heo * 5297c3eed5cSTejun Heo * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 5307c3eed5cSTejun Heo * and clear_work_data() can be used to set the cwq, pool or clear 531bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 532bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5337a22ad75STejun Heo * 5347c3eed5cSTejun Heo * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq 5357c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 5367c3eed5cSTejun Heo * queued anywhere after initialization until it is sync canceled. cwq is 5377c3eed5cSTejun Heo * available only while the work item is queued. 538bbb68dfaSTejun Heo * 539bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 540bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 541bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 542bbb68dfaSTejun Heo * try to steal the PENDING bit. 5434594bf15SDavid Howells */ 5447a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5457a22ad75STejun Heo unsigned long flags) 5467a22ad75STejun Heo { 5477a22ad75STejun Heo BUG_ON(!work_pending(work)); 5487a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5497a22ad75STejun Heo } 5507a22ad75STejun Heo 5517a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5524690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5534690c4abSTejun Heo unsigned long extra_flags) 554365970a1SDavid Howells { 5557a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 556e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 557365970a1SDavid Howells } 558365970a1SDavid Howells 5594468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5604468a00fSLai Jiangshan int pool_id) 5614468a00fSLai Jiangshan { 5624468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5634468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5644468a00fSLai Jiangshan } 5654468a00fSLai Jiangshan 5667c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5677c3eed5cSTejun Heo int pool_id) 5684d707b9fSOleg Nesterov { 56923657bb1STejun Heo /* 57023657bb1STejun Heo * The following wmb is paired with the implied mb in 57123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 57223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 57323657bb1STejun Heo * owner. 57423657bb1STejun Heo */ 57523657bb1STejun Heo smp_wmb(); 5767c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5774d707b9fSOleg Nesterov } 5784d707b9fSOleg Nesterov 5797a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 580365970a1SDavid Howells { 5817c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5827c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5837a22ad75STejun Heo } 5847a22ad75STejun Heo 5857a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5867a22ad75STejun Heo { 587e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5887a22ad75STejun Heo 589e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 590e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 591e120153dSTejun Heo else 592e120153dSTejun Heo return NULL; 5937a22ad75STejun Heo } 5947a22ad75STejun Heo 5957c3eed5cSTejun Heo /** 5967c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5977c3eed5cSTejun Heo * @work: the work item of interest 5987c3eed5cSTejun Heo * 5997c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 6007c3eed5cSTejun Heo */ 6017c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6027a22ad75STejun Heo { 603e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6047c3eed5cSTejun Heo struct worker_pool *pool; 6057c3eed5cSTejun Heo int pool_id; 6067a22ad75STejun Heo 607e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 608e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 6097c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6107a22ad75STejun Heo 6117c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6127c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6137a22ad75STejun Heo return NULL; 6147a22ad75STejun Heo 6157c3eed5cSTejun Heo pool = worker_pool_by_id(pool_id); 6167c3eed5cSTejun Heo WARN_ON_ONCE(!pool); 6177c3eed5cSTejun Heo return pool; 6187c3eed5cSTejun Heo } 6197c3eed5cSTejun Heo 6207c3eed5cSTejun Heo /** 6217c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6227c3eed5cSTejun Heo * @work: the work item of interest 6237c3eed5cSTejun Heo * 6247c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6257c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6267c3eed5cSTejun Heo */ 6277c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6287c3eed5cSTejun Heo { 6297c3eed5cSTejun Heo struct worker_pool *pool = get_work_pool(work); 6307c3eed5cSTejun Heo 6317c3eed5cSTejun Heo return pool ? pool->id : WORK_OFFQ_POOL_NONE; 6327c3eed5cSTejun Heo } 6337c3eed5cSTejun Heo 634bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 635bbb68dfaSTejun Heo { 6367c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 637bbb68dfaSTejun Heo 6387c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6397c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 640bbb68dfaSTejun Heo } 641bbb68dfaSTejun Heo 642bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 643bbb68dfaSTejun Heo { 644bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 645bbb68dfaSTejun Heo 646bbb68dfaSTejun Heo return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING); 647bbb68dfaSTejun Heo } 648bbb68dfaSTejun Heo 649e22bee78STejun Heo /* 6503270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6513270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 652d565ed63STejun Heo * they're being called with pool->lock held. 653e22bee78STejun Heo */ 654e22bee78STejun Heo 65563d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 656649027d7STejun Heo { 6573270476aSTejun Heo return !atomic_read(get_pool_nr_running(pool)); 658649027d7STejun Heo } 659649027d7STejun Heo 660e22bee78STejun Heo /* 661e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 662e22bee78STejun Heo * running workers. 663974271c4STejun Heo * 664974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 665706026c2STejun Heo * function will always return %true for unbound pools as long as the 666974271c4STejun Heo * worklist isn't empty. 667e22bee78STejun Heo */ 66863d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 669e22bee78STejun Heo { 67063d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 671e22bee78STejun Heo } 672e22bee78STejun Heo 673e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 67463d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 675e22bee78STejun Heo { 67663d95a91STejun Heo return pool->nr_idle; 677e22bee78STejun Heo } 678e22bee78STejun Heo 679e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 68063d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 681e22bee78STejun Heo { 68263d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 683e22bee78STejun Heo 6843270476aSTejun Heo return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1; 685e22bee78STejun Heo } 686e22bee78STejun Heo 687e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 68863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 689e22bee78STejun Heo { 69063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 691e22bee78STejun Heo } 692e22bee78STejun Heo 693e22bee78STejun Heo /* Do I need to be the manager? */ 69463d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 695e22bee78STejun Heo { 69663d95a91STejun Heo return need_to_create_worker(pool) || 69711ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 698e22bee78STejun Heo } 699e22bee78STejun Heo 700e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 70163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 702e22bee78STejun Heo { 703552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 70463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 70563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 706e22bee78STejun Heo 707ea1abd61SLai Jiangshan /* 708ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 709ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 710ea1abd61SLai Jiangshan */ 711ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 712ea1abd61SLai Jiangshan return false; 713ea1abd61SLai Jiangshan 714e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 715e22bee78STejun Heo } 716e22bee78STejun Heo 717e22bee78STejun Heo /* 718e22bee78STejun Heo * Wake up functions. 719e22bee78STejun Heo */ 720e22bee78STejun Heo 7217e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 72263d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7237e11629dSTejun Heo { 72463d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7257e11629dSTejun Heo return NULL; 7267e11629dSTejun Heo 72763d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7287e11629dSTejun Heo } 7297e11629dSTejun Heo 7307e11629dSTejun Heo /** 7317e11629dSTejun Heo * wake_up_worker - wake up an idle worker 73263d95a91STejun Heo * @pool: worker pool to wake worker from 7337e11629dSTejun Heo * 73463d95a91STejun Heo * Wake up the first idle worker of @pool. 7357e11629dSTejun Heo * 7367e11629dSTejun Heo * CONTEXT: 737d565ed63STejun Heo * spin_lock_irq(pool->lock). 7387e11629dSTejun Heo */ 73963d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7407e11629dSTejun Heo { 74163d95a91STejun Heo struct worker *worker = first_worker(pool); 7427e11629dSTejun Heo 7437e11629dSTejun Heo if (likely(worker)) 7447e11629dSTejun Heo wake_up_process(worker->task); 7457e11629dSTejun Heo } 7467e11629dSTejun Heo 7474690c4abSTejun Heo /** 748e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 749e22bee78STejun Heo * @task: task waking up 750e22bee78STejun Heo * @cpu: CPU @task is waking up to 751e22bee78STejun Heo * 752e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 753e22bee78STejun Heo * being awoken. 754e22bee78STejun Heo * 755e22bee78STejun Heo * CONTEXT: 756e22bee78STejun Heo * spin_lock_irq(rq->lock) 757e22bee78STejun Heo */ 758e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 759e22bee78STejun Heo { 760e22bee78STejun Heo struct worker *worker = kthread_data(task); 761e22bee78STejun Heo 76236576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 763ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 76463d95a91STejun Heo atomic_inc(get_pool_nr_running(worker->pool)); 765e22bee78STejun Heo } 76636576000SJoonsoo Kim } 767e22bee78STejun Heo 768e22bee78STejun Heo /** 769e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 770e22bee78STejun Heo * @task: task going to sleep 771e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 772e22bee78STejun Heo * 773e22bee78STejun Heo * This function is called during schedule() when a busy worker is 774e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 775e22bee78STejun Heo * returning pointer to its task. 776e22bee78STejun Heo * 777e22bee78STejun Heo * CONTEXT: 778e22bee78STejun Heo * spin_lock_irq(rq->lock) 779e22bee78STejun Heo * 780e22bee78STejun Heo * RETURNS: 781e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 782e22bee78STejun Heo */ 783e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 784e22bee78STejun Heo unsigned int cpu) 785e22bee78STejun Heo { 786e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 787111c225aSTejun Heo struct worker_pool *pool; 788111c225aSTejun Heo atomic_t *nr_running; 789e22bee78STejun Heo 790111c225aSTejun Heo /* 791111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 792111c225aSTejun Heo * workers, also reach here, let's not access anything before 793111c225aSTejun Heo * checking NOT_RUNNING. 794111c225aSTejun Heo */ 7952d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 796e22bee78STejun Heo return NULL; 797e22bee78STejun Heo 798111c225aSTejun Heo pool = worker->pool; 799111c225aSTejun Heo nr_running = get_pool_nr_running(pool); 800111c225aSTejun Heo 801e22bee78STejun Heo /* this can only happen on the local cpu */ 802e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 803e22bee78STejun Heo 804e22bee78STejun Heo /* 805e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 806e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 807e22bee78STejun Heo * Please read comment there. 808e22bee78STejun Heo * 809628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 810628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 811628c78e7STejun Heo * disabled, which in turn means that none else could be 812d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 813628c78e7STejun Heo * lock is safe. 814e22bee78STejun Heo */ 815bd7bdd43STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist)) 81663d95a91STejun Heo to_wakeup = first_worker(pool); 817e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 818e22bee78STejun Heo } 819e22bee78STejun Heo 820e22bee78STejun Heo /** 821e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 822cb444766STejun Heo * @worker: self 823d302f017STejun Heo * @flags: flags to set 824d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 825d302f017STejun Heo * 826e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 827e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 828e22bee78STejun Heo * woken up. 829d302f017STejun Heo * 830cb444766STejun Heo * CONTEXT: 831d565ed63STejun Heo * spin_lock_irq(pool->lock) 832d302f017STejun Heo */ 833d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 834d302f017STejun Heo bool wakeup) 835d302f017STejun Heo { 836bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 837e22bee78STejun Heo 838cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 839cb444766STejun Heo 840e22bee78STejun Heo /* 841e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 842e22bee78STejun Heo * wake up an idle worker as necessary if requested by 843e22bee78STejun Heo * @wakeup. 844e22bee78STejun Heo */ 845e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 846e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 84763d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 848e22bee78STejun Heo 849e22bee78STejun Heo if (wakeup) { 850e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 851bd7bdd43STejun Heo !list_empty(&pool->worklist)) 85263d95a91STejun Heo wake_up_worker(pool); 853e22bee78STejun Heo } else 854e22bee78STejun Heo atomic_dec(nr_running); 855e22bee78STejun Heo } 856e22bee78STejun Heo 857d302f017STejun Heo worker->flags |= flags; 858d302f017STejun Heo } 859d302f017STejun Heo 860d302f017STejun Heo /** 861e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 862cb444766STejun Heo * @worker: self 863d302f017STejun Heo * @flags: flags to clear 864d302f017STejun Heo * 865e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 866d302f017STejun Heo * 867cb444766STejun Heo * CONTEXT: 868d565ed63STejun Heo * spin_lock_irq(pool->lock) 869d302f017STejun Heo */ 870d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 871d302f017STejun Heo { 87263d95a91STejun Heo struct worker_pool *pool = worker->pool; 873e22bee78STejun Heo unsigned int oflags = worker->flags; 874e22bee78STejun Heo 875cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 876cb444766STejun Heo 877d302f017STejun Heo worker->flags &= ~flags; 878e22bee78STejun Heo 87942c025f3STejun Heo /* 88042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 88142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 88242c025f3STejun Heo * of multiple flags, not a single flag. 88342c025f3STejun Heo */ 884e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 885e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 88663d95a91STejun Heo atomic_inc(get_pool_nr_running(pool)); 887d302f017STejun Heo } 888d302f017STejun Heo 889d302f017STejun Heo /** 8908cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 891c9e7cf27STejun Heo * @pool: pool of interest 8928cca0eeaSTejun Heo * @work: work to find worker for 8938cca0eeaSTejun Heo * 894c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 895c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 896a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 897a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 898a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 899a2c1c57bSTejun Heo * being executed. 900a2c1c57bSTejun Heo * 901a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 902a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 903a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 904a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 905a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 906a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 907a2c1c57bSTejun Heo * 908a2c1c57bSTejun Heo * This function checks the work item address, work function and workqueue 909a2c1c57bSTejun Heo * to avoid false positives. Note that this isn't complete as one may 910a2c1c57bSTejun Heo * construct a work function which can introduce dependency onto itself 911a2c1c57bSTejun Heo * through a recycled work item. Well, if somebody wants to shoot oneself 912a2c1c57bSTejun Heo * in the foot that badly, there's only so much we can do, and if such 913a2c1c57bSTejun Heo * deadlock actually occurs, it should be easy to locate the culprit work 914a2c1c57bSTejun Heo * function. 9158cca0eeaSTejun Heo * 9168cca0eeaSTejun Heo * CONTEXT: 917d565ed63STejun Heo * spin_lock_irq(pool->lock). 9188cca0eeaSTejun Heo * 9198cca0eeaSTejun Heo * RETURNS: 9208cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9218cca0eeaSTejun Heo * otherwise. 9228cca0eeaSTejun Heo */ 923c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9248cca0eeaSTejun Heo struct work_struct *work) 9258cca0eeaSTejun Heo { 92642f8570fSSasha Levin struct worker *worker; 92742f8570fSSasha Levin struct hlist_node *tmp; 92842f8570fSSasha Levin 929c9e7cf27STejun Heo hash_for_each_possible(pool->busy_hash, worker, tmp, hentry, 930a2c1c57bSTejun Heo (unsigned long)work) 931a2c1c57bSTejun Heo if (worker->current_work == work && 932a2c1c57bSTejun Heo worker->current_func == work->func) 93342f8570fSSasha Levin return worker; 93442f8570fSSasha Levin 93542f8570fSSasha Levin return NULL; 9368cca0eeaSTejun Heo } 9378cca0eeaSTejun Heo 9388cca0eeaSTejun Heo /** 939bf4ede01STejun Heo * move_linked_works - move linked works to a list 940bf4ede01STejun Heo * @work: start of series of works to be scheduled 941bf4ede01STejun Heo * @head: target list to append @work to 942bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 943bf4ede01STejun Heo * 944bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 945bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 946bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 947bf4ede01STejun Heo * 948bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 949bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 950bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 951bf4ede01STejun Heo * 952bf4ede01STejun Heo * CONTEXT: 953d565ed63STejun Heo * spin_lock_irq(pool->lock). 954bf4ede01STejun Heo */ 955bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 956bf4ede01STejun Heo struct work_struct **nextp) 957bf4ede01STejun Heo { 958bf4ede01STejun Heo struct work_struct *n; 959bf4ede01STejun Heo 960bf4ede01STejun Heo /* 961bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 962bf4ede01STejun Heo * use NULL for list head. 963bf4ede01STejun Heo */ 964bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 965bf4ede01STejun Heo list_move_tail(&work->entry, head); 966bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 967bf4ede01STejun Heo break; 968bf4ede01STejun Heo } 969bf4ede01STejun Heo 970bf4ede01STejun Heo /* 971bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 972bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 973bf4ede01STejun Heo * needs to be updated. 974bf4ede01STejun Heo */ 975bf4ede01STejun Heo if (nextp) 976bf4ede01STejun Heo *nextp = n; 977bf4ede01STejun Heo } 978bf4ede01STejun Heo 9793aa62497SLai Jiangshan static void cwq_activate_delayed_work(struct work_struct *work) 980bf4ede01STejun Heo { 9813aa62497SLai Jiangshan struct cpu_workqueue_struct *cwq = get_work_cwq(work); 982bf4ede01STejun Heo 983bf4ede01STejun Heo trace_workqueue_activate_work(work); 984bf4ede01STejun Heo move_linked_works(work, &cwq->pool->worklist, NULL); 985bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 986bf4ede01STejun Heo cwq->nr_active++; 987bf4ede01STejun Heo } 988bf4ede01STejun Heo 9893aa62497SLai Jiangshan static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 9903aa62497SLai Jiangshan { 9913aa62497SLai Jiangshan struct work_struct *work = list_first_entry(&cwq->delayed_works, 9923aa62497SLai Jiangshan struct work_struct, entry); 9933aa62497SLai Jiangshan 9943aa62497SLai Jiangshan cwq_activate_delayed_work(work); 9953aa62497SLai Jiangshan } 9963aa62497SLai Jiangshan 997bf4ede01STejun Heo /** 998bf4ede01STejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 999bf4ede01STejun Heo * @cwq: cwq of interest 1000bf4ede01STejun Heo * @color: color of work which left the queue 1001bf4ede01STejun Heo * 1002bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1003bf4ede01STejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 1004bf4ede01STejun Heo * 1005bf4ede01STejun Heo * CONTEXT: 1006d565ed63STejun Heo * spin_lock_irq(pool->lock). 1007bf4ede01STejun Heo */ 1008b3f9f405SLai Jiangshan static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color) 1009bf4ede01STejun Heo { 1010bf4ede01STejun Heo /* ignore uncolored works */ 1011bf4ede01STejun Heo if (color == WORK_NO_COLOR) 1012bf4ede01STejun Heo return; 1013bf4ede01STejun Heo 1014bf4ede01STejun Heo cwq->nr_in_flight[color]--; 1015bf4ede01STejun Heo 1016bf4ede01STejun Heo cwq->nr_active--; 1017bf4ede01STejun Heo if (!list_empty(&cwq->delayed_works)) { 1018bf4ede01STejun Heo /* one down, submit a delayed one */ 1019bf4ede01STejun Heo if (cwq->nr_active < cwq->max_active) 1020bf4ede01STejun Heo cwq_activate_first_delayed(cwq); 1021bf4ede01STejun Heo } 1022bf4ede01STejun Heo 1023bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1024bf4ede01STejun Heo if (likely(cwq->flush_color != color)) 1025bf4ede01STejun Heo return; 1026bf4ede01STejun Heo 1027bf4ede01STejun Heo /* are there still in-flight works? */ 1028bf4ede01STejun Heo if (cwq->nr_in_flight[color]) 1029bf4ede01STejun Heo return; 1030bf4ede01STejun Heo 1031bf4ede01STejun Heo /* this cwq is done, clear flush_color */ 1032bf4ede01STejun Heo cwq->flush_color = -1; 1033bf4ede01STejun Heo 1034bf4ede01STejun Heo /* 1035bf4ede01STejun Heo * If this was the last cwq, wake up the first flusher. It 1036bf4ede01STejun Heo * will handle the rest. 1037bf4ede01STejun Heo */ 1038bf4ede01STejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 1039bf4ede01STejun Heo complete(&cwq->wq->first_flusher->done); 1040bf4ede01STejun Heo } 1041bf4ede01STejun Heo 104236e227d2STejun Heo /** 1043bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 104436e227d2STejun Heo * @work: work item to steal 104536e227d2STejun Heo * @is_dwork: @work is a delayed_work 1046bbb68dfaSTejun Heo * @flags: place to store irq state 104736e227d2STejun Heo * 104836e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 104936e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 105036e227d2STejun Heo * 105136e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 105236e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 105336e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1054bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1055bbb68dfaSTejun Heo * for arbitrarily long 105636e227d2STejun Heo * 1057bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1058e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1059e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1060e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1061bbb68dfaSTejun Heo * 1062bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1063bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1064bbb68dfaSTejun Heo * 1065e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1066bf4ede01STejun Heo */ 1067bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1068bbb68dfaSTejun Heo unsigned long *flags) 1069bf4ede01STejun Heo { 1070d565ed63STejun Heo struct worker_pool *pool; 1071bf4ede01STejun Heo 1072bbb68dfaSTejun Heo local_irq_save(*flags); 1073bbb68dfaSTejun Heo 107436e227d2STejun Heo /* try to steal the timer if it exists */ 107536e227d2STejun Heo if (is_dwork) { 107636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 107736e227d2STejun Heo 1078e0aecdd8STejun Heo /* 1079e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1080e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1081e0aecdd8STejun Heo * running on the local CPU. 1082e0aecdd8STejun Heo */ 108336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 108436e227d2STejun Heo return 1; 108536e227d2STejun Heo } 108636e227d2STejun Heo 108736e227d2STejun Heo /* try to claim PENDING the normal way */ 1088bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1089bf4ede01STejun Heo return 0; 1090bf4ede01STejun Heo 1091bf4ede01STejun Heo /* 1092bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1093bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1094bf4ede01STejun Heo */ 1095d565ed63STejun Heo pool = get_work_pool(work); 1096d565ed63STejun Heo if (!pool) 1097bbb68dfaSTejun Heo goto fail; 1098bf4ede01STejun Heo 1099d565ed63STejun Heo spin_lock(&pool->lock); 1100bf4ede01STejun Heo if (!list_empty(&work->entry)) { 1101bf4ede01STejun Heo /* 1102d565ed63STejun Heo * This work is queued, but perhaps we locked the wrong 1103d565ed63STejun Heo * pool. In that case we must see the new value after 1104d565ed63STejun Heo * rmb(), see insert_work()->wmb(). 1105bf4ede01STejun Heo */ 1106bf4ede01STejun Heo smp_rmb(); 1107d565ed63STejun Heo if (pool == get_work_pool(work)) { 1108bf4ede01STejun Heo debug_work_deactivate(work); 11093aa62497SLai Jiangshan 11103aa62497SLai Jiangshan /* 11113aa62497SLai Jiangshan * A delayed work item cannot be grabbed directly 11123aa62497SLai Jiangshan * because it might have linked NO_COLOR work items 11133aa62497SLai Jiangshan * which, if left on the delayed_list, will confuse 11143aa62497SLai Jiangshan * cwq->nr_active management later on and cause 11153aa62497SLai Jiangshan * stall. Make sure the work item is activated 11163aa62497SLai Jiangshan * before grabbing. 11173aa62497SLai Jiangshan */ 11183aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 11193aa62497SLai Jiangshan cwq_activate_delayed_work(work); 11203aa62497SLai Jiangshan 1121bf4ede01STejun Heo list_del_init(&work->entry); 1122bf4ede01STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 1123b3f9f405SLai Jiangshan get_work_color(work)); 112436e227d2STejun Heo 11254468a00fSLai Jiangshan /* work->data points to cwq iff queued, point to pool */ 11264468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11274468a00fSLai Jiangshan 1128d565ed63STejun Heo spin_unlock(&pool->lock); 112936e227d2STejun Heo return 1; 1130bf4ede01STejun Heo } 1131bf4ede01STejun Heo } 1132d565ed63STejun Heo spin_unlock(&pool->lock); 1133bbb68dfaSTejun Heo fail: 1134bbb68dfaSTejun Heo local_irq_restore(*flags); 1135bbb68dfaSTejun Heo if (work_is_canceling(work)) 1136bbb68dfaSTejun Heo return -ENOENT; 1137bbb68dfaSTejun Heo cpu_relax(); 113836e227d2STejun Heo return -EAGAIN; 1139bf4ede01STejun Heo } 1140bf4ede01STejun Heo 1141bf4ede01STejun Heo /** 1142706026c2STejun Heo * insert_work - insert a work into a pool 11434690c4abSTejun Heo * @cwq: cwq @work belongs to 11444690c4abSTejun Heo * @work: work to insert 11454690c4abSTejun Heo * @head: insertion point 11464690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11474690c4abSTejun Heo * 1148706026c2STejun Heo * Insert @work which belongs to @cwq after @head. @extra_flags is or'd to 1149706026c2STejun Heo * work_struct flags. 11504690c4abSTejun Heo * 11514690c4abSTejun Heo * CONTEXT: 1152d565ed63STejun Heo * spin_lock_irq(pool->lock). 1153365970a1SDavid Howells */ 1154b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 11554690c4abSTejun Heo struct work_struct *work, struct list_head *head, 11564690c4abSTejun Heo unsigned int extra_flags) 1157b89deed3SOleg Nesterov { 115863d95a91STejun Heo struct worker_pool *pool = cwq->pool; 1159e1d8aa9fSFrederic Weisbecker 11604690c4abSTejun Heo /* we own @work, set data and link */ 11617a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 11624690c4abSTejun Heo 11636e84d644SOleg Nesterov /* 11646e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 11656e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 11666e84d644SOleg Nesterov */ 11676e84d644SOleg Nesterov smp_wmb(); 11684690c4abSTejun Heo 11691a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1170e22bee78STejun Heo 1171e22bee78STejun Heo /* 1172e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1173e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1174e22bee78STejun Heo * lying around lazily while there are works to be processed. 1175e22bee78STejun Heo */ 1176e22bee78STejun Heo smp_mb(); 1177e22bee78STejun Heo 117863d95a91STejun Heo if (__need_more_worker(pool)) 117963d95a91STejun Heo wake_up_worker(pool); 1180b89deed3SOleg Nesterov } 1181b89deed3SOleg Nesterov 1182c8efcc25STejun Heo /* 1183c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 1184c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 1185c8efcc25STejun Heo * cold paths. 1186c8efcc25STejun Heo */ 1187c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1188c8efcc25STejun Heo { 1189c8efcc25STejun Heo unsigned long flags; 1190c8efcc25STejun Heo unsigned int cpu; 1191c8efcc25STejun Heo 1192706026c2STejun Heo for_each_wq_cpu(cpu) { 1193c9e7cf27STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 1194c9e7cf27STejun Heo struct worker_pool *pool = cwq->pool; 1195c8efcc25STejun Heo struct worker *worker; 1196c8efcc25STejun Heo struct hlist_node *pos; 1197c8efcc25STejun Heo int i; 1198c8efcc25STejun Heo 1199d565ed63STejun Heo spin_lock_irqsave(&pool->lock, flags); 1200c9e7cf27STejun Heo for_each_busy_worker(worker, i, pos, pool) { 1201c8efcc25STejun Heo if (worker->task != current) 1202c8efcc25STejun Heo continue; 1203d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 1204c8efcc25STejun Heo /* 1205c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 1206c8efcc25STejun Heo * is headed to the same workqueue. 1207c8efcc25STejun Heo */ 1208c8efcc25STejun Heo return worker->current_cwq->wq == wq; 1209c8efcc25STejun Heo } 1210d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 1211c8efcc25STejun Heo } 1212c8efcc25STejun Heo return false; 1213c8efcc25STejun Heo } 1214c8efcc25STejun Heo 12154690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 12161da177e4SLinus Torvalds struct work_struct *work) 12171da177e4SLinus Torvalds { 1218d565ed63STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 1219d565ed63STejun Heo struct worker_pool *pool; 1220502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 12211e19ffc6STejun Heo struct list_head *worklist; 12228a2e8e5dSTejun Heo unsigned int work_flags; 1223b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12248930cabaSTejun Heo 12258930cabaSTejun Heo /* 12268930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12278930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12288930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12298930cabaSTejun Heo * happen with IRQ disabled. 12308930cabaSTejun Heo */ 12318930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12321da177e4SLinus Torvalds 1233dc186ad7SThomas Gleixner debug_work_activate(work); 12341e19ffc6STejun Heo 1235c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 12369c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1237c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1238e41e704bSTejun Heo return; 1239e41e704bSTejun Heo 1240d565ed63STejun Heo /* determine pool to use */ 1241c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1242c9e7cf27STejun Heo struct worker_pool *last_pool; 1243c7fc77f7STejun Heo 124457469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1245f3421797STejun Heo cpu = raw_smp_processor_id(); 1246f3421797STejun Heo 124718aa9effSTejun Heo /* 1248dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1249dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1250dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1251dbf2576eSTejun Heo * non-reentrancy. 125218aa9effSTejun Heo */ 1253d565ed63STejun Heo pool = get_std_worker_pool(cpu, highpri); 1254c9e7cf27STejun Heo last_pool = get_work_pool(work); 1255dbf2576eSTejun Heo 1256d565ed63STejun Heo if (last_pool && last_pool != pool) { 125718aa9effSTejun Heo struct worker *worker; 125818aa9effSTejun Heo 1259d565ed63STejun Heo spin_lock(&last_pool->lock); 126018aa9effSTejun Heo 1261c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 126218aa9effSTejun Heo 126318aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 1264d565ed63STejun Heo pool = last_pool; 126518aa9effSTejun Heo else { 126618aa9effSTejun Heo /* meh... not running there, queue here */ 1267d565ed63STejun Heo spin_unlock(&last_pool->lock); 1268d565ed63STejun Heo spin_lock(&pool->lock); 126918aa9effSTejun Heo } 12708930cabaSTejun Heo } else { 1271d565ed63STejun Heo spin_lock(&pool->lock); 12728930cabaSTejun Heo } 1273f3421797STejun Heo } else { 1274d565ed63STejun Heo pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri); 1275d565ed63STejun Heo spin_lock(&pool->lock); 1276502ca9d8STejun Heo } 1277502ca9d8STejun Heo 1278d565ed63STejun Heo /* pool determined, get cwq and queue */ 1279d565ed63STejun Heo cwq = get_cwq(pool->cpu, wq); 1280b75cac93SJoonsoo Kim trace_workqueue_queue_work(req_cpu, cwq, work); 1281502ca9d8STejun Heo 1282f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1283d565ed63STejun Heo spin_unlock(&pool->lock); 1284f5b2552bSDan Carpenter return; 1285f5b2552bSDan Carpenter } 12861e19ffc6STejun Heo 128773f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 12888a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 12891e19ffc6STejun Heo 12901e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1291cdadf009STejun Heo trace_workqueue_activate_work(work); 12921e19ffc6STejun Heo cwq->nr_active++; 12933270476aSTejun Heo worklist = &cwq->pool->worklist; 12948a2e8e5dSTejun Heo } else { 12958a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 12961e19ffc6STejun Heo worklist = &cwq->delayed_works; 12978a2e8e5dSTejun Heo } 12981e19ffc6STejun Heo 12998a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 13001e19ffc6STejun Heo 1301d565ed63STejun Heo spin_unlock(&pool->lock); 13021da177e4SLinus Torvalds } 13031da177e4SLinus Torvalds 13040fcb78c2SRolf Eike Beer /** 1305c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1306c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1307c1a220e7SZhang Rui * @wq: workqueue to use 1308c1a220e7SZhang Rui * @work: work to queue 1309c1a220e7SZhang Rui * 1310d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1311c1a220e7SZhang Rui * 1312c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1313c1a220e7SZhang Rui * can't go away. 1314c1a220e7SZhang Rui */ 1315d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1316d4283e93STejun Heo struct work_struct *work) 1317c1a220e7SZhang Rui { 1318d4283e93STejun Heo bool ret = false; 13198930cabaSTejun Heo unsigned long flags; 13208930cabaSTejun Heo 13218930cabaSTejun Heo local_irq_save(flags); 1322c1a220e7SZhang Rui 132322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13244690c4abSTejun Heo __queue_work(cpu, wq, work); 1325d4283e93STejun Heo ret = true; 1326c1a220e7SZhang Rui } 13278930cabaSTejun Heo 13288930cabaSTejun Heo local_irq_restore(flags); 1329c1a220e7SZhang Rui return ret; 1330c1a220e7SZhang Rui } 1331c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1332c1a220e7SZhang Rui 13330a13c00eSTejun Heo /** 13340a13c00eSTejun Heo * queue_work - queue work on a workqueue 13350a13c00eSTejun Heo * @wq: workqueue to use 13360a13c00eSTejun Heo * @work: work to queue 13370a13c00eSTejun Heo * 1338d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 13390a13c00eSTejun Heo * 13400a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 13410a13c00eSTejun Heo * it can be processed by another CPU. 13420a13c00eSTejun Heo */ 1343d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13440a13c00eSTejun Heo { 134557469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13460a13c00eSTejun Heo } 13470a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13480a13c00eSTejun Heo 1349d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13501da177e4SLinus Torvalds { 135152bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13521da177e4SLinus Torvalds 1353e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 135460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13551da177e4SLinus Torvalds } 1356d8e794dfSTejun Heo EXPORT_SYMBOL_GPL(delayed_work_timer_fn); 13571da177e4SLinus Torvalds 13587beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 135952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13601da177e4SLinus Torvalds { 13617beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13627beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13631da177e4SLinus Torvalds 13647beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13657beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1366fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1367fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13687beb2edfSTejun Heo 13698852aac2STejun Heo /* 13708852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13718852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13728852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13738852aac2STejun Heo * on that there's no such delay when @delay is 0. 13748852aac2STejun Heo */ 13758852aac2STejun Heo if (!delay) { 13768852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13778852aac2STejun Heo return; 13788852aac2STejun Heo } 13798852aac2STejun Heo 13807beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13817beb2edfSTejun Heo 138260c057bcSLai Jiangshan dwork->wq = wq; 13831265057fSTejun Heo dwork->cpu = cpu; 13847beb2edfSTejun Heo timer->expires = jiffies + delay; 13857beb2edfSTejun Heo 13867beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13877beb2edfSTejun Heo add_timer_on(timer, cpu); 13887beb2edfSTejun Heo else 13897beb2edfSTejun Heo add_timer(timer); 13907beb2edfSTejun Heo } 13911da177e4SLinus Torvalds 13920fcb78c2SRolf Eike Beer /** 13930fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13940fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13950fcb78c2SRolf Eike Beer * @wq: workqueue to use 1396af9997e4SRandy Dunlap * @dwork: work to queue 13970fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13980fcb78c2SRolf Eike Beer * 1399715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1400715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1401715f1300STejun Heo * execution. 14020fcb78c2SRolf Eike Beer */ 1403d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 140452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14057a6bc1cdSVenkatesh Pallipadi { 140652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1407d4283e93STejun Heo bool ret = false; 14088930cabaSTejun Heo unsigned long flags; 14098930cabaSTejun Heo 14108930cabaSTejun Heo /* read the comment in __queue_work() */ 14118930cabaSTejun Heo local_irq_save(flags); 14127a6bc1cdSVenkatesh Pallipadi 141322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14147beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1415d4283e93STejun Heo ret = true; 14167a6bc1cdSVenkatesh Pallipadi } 14178930cabaSTejun Heo 14188930cabaSTejun Heo local_irq_restore(flags); 14197a6bc1cdSVenkatesh Pallipadi return ret; 14207a6bc1cdSVenkatesh Pallipadi } 1421ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14221da177e4SLinus Torvalds 1423c8e55f36STejun Heo /** 14240a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 14250a13c00eSTejun Heo * @wq: workqueue to use 14260a13c00eSTejun Heo * @dwork: delayable work to queue 14270a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 14280a13c00eSTejun Heo * 1429715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 14300a13c00eSTejun Heo */ 1431d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 14320a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 14330a13c00eSTejun Heo { 143457469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14350a13c00eSTejun Heo } 14360a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 14370a13c00eSTejun Heo 14380a13c00eSTejun Heo /** 14398376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14408376fe22STejun Heo * @cpu: CPU number to execute work on 14418376fe22STejun Heo * @wq: workqueue to use 14428376fe22STejun Heo * @dwork: work to queue 14438376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14448376fe22STejun Heo * 14458376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14468376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14478376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14488376fe22STejun Heo * current state. 14498376fe22STejun Heo * 14508376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14518376fe22STejun Heo * pending and its timer was modified. 14528376fe22STejun Heo * 1453e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14548376fe22STejun Heo * See try_to_grab_pending() for details. 14558376fe22STejun Heo */ 14568376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14578376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14588376fe22STejun Heo { 14598376fe22STejun Heo unsigned long flags; 14608376fe22STejun Heo int ret; 14618376fe22STejun Heo 14628376fe22STejun Heo do { 14638376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14648376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14658376fe22STejun Heo 14668376fe22STejun Heo if (likely(ret >= 0)) { 14678376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14688376fe22STejun Heo local_irq_restore(flags); 14698376fe22STejun Heo } 14708376fe22STejun Heo 14718376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14728376fe22STejun Heo return ret; 14738376fe22STejun Heo } 14748376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14758376fe22STejun Heo 14768376fe22STejun Heo /** 14778376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14788376fe22STejun Heo * @wq: workqueue to use 14798376fe22STejun Heo * @dwork: work to queue 14808376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14818376fe22STejun Heo * 14828376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14838376fe22STejun Heo */ 14848376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14858376fe22STejun Heo unsigned long delay) 14868376fe22STejun Heo { 14878376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14888376fe22STejun Heo } 14898376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14908376fe22STejun Heo 14918376fe22STejun Heo /** 1492c8e55f36STejun Heo * worker_enter_idle - enter idle state 1493c8e55f36STejun Heo * @worker: worker which is entering idle state 1494c8e55f36STejun Heo * 1495c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1496c8e55f36STejun Heo * necessary. 1497c8e55f36STejun Heo * 1498c8e55f36STejun Heo * LOCKING: 1499d565ed63STejun Heo * spin_lock_irq(pool->lock). 1500c8e55f36STejun Heo */ 1501c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15021da177e4SLinus Torvalds { 1503bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1504c8e55f36STejun Heo 1505c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1506c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1507c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1508c8e55f36STejun Heo 1509cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1510cb444766STejun Heo worker->flags |= WORKER_IDLE; 1511bd7bdd43STejun Heo pool->nr_idle++; 1512e22bee78STejun Heo worker->last_active = jiffies; 1513c8e55f36STejun Heo 1514c8e55f36STejun Heo /* idle_list is LIFO */ 1515bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1516db7bccf4STejun Heo 151763d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1518628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1519cb444766STejun Heo 1520544ecf31STejun Heo /* 1521706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1522d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1523628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1524628c78e7STejun Heo * unbind is not in progress. 1525544ecf31STejun Heo */ 152624647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1527bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 152863d95a91STejun Heo atomic_read(get_pool_nr_running(pool))); 1529c8e55f36STejun Heo } 1530c8e55f36STejun Heo 1531c8e55f36STejun Heo /** 1532c8e55f36STejun Heo * worker_leave_idle - leave idle state 1533c8e55f36STejun Heo * @worker: worker which is leaving idle state 1534c8e55f36STejun Heo * 1535c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1536c8e55f36STejun Heo * 1537c8e55f36STejun Heo * LOCKING: 1538d565ed63STejun Heo * spin_lock_irq(pool->lock). 1539c8e55f36STejun Heo */ 1540c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1541c8e55f36STejun Heo { 1542bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1543c8e55f36STejun Heo 1544c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1545d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1546bd7bdd43STejun Heo pool->nr_idle--; 1547c8e55f36STejun Heo list_del_init(&worker->entry); 1548c8e55f36STejun Heo } 1549c8e55f36STejun Heo 1550e22bee78STejun Heo /** 1551706026c2STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool 1552e22bee78STejun Heo * @worker: self 1553e22bee78STejun Heo * 1554e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1555e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1556e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1557e22bee78STejun Heo * guaranteed to execute on the cpu. 1558e22bee78STejun Heo * 1559e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1560e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1561e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1562e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1563706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1564e22bee78STejun Heo * [dis]associated in the meantime. 1565e22bee78STejun Heo * 1566706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 156724647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1568f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1569f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1570f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1571e22bee78STejun Heo * 1572e22bee78STejun Heo * CONTEXT: 1573d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1574e22bee78STejun Heo * held. 1575e22bee78STejun Heo * 1576e22bee78STejun Heo * RETURNS: 1577706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1578e22bee78STejun Heo * bound), %false if offline. 1579e22bee78STejun Heo */ 1580e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1581d565ed63STejun Heo __acquires(&pool->lock) 1582e22bee78STejun Heo { 158324647570STejun Heo struct worker_pool *pool = worker->pool; 1584e22bee78STejun Heo struct task_struct *task = worker->task; 1585e22bee78STejun Heo 1586e22bee78STejun Heo while (true) { 1587e22bee78STejun Heo /* 1588e22bee78STejun Heo * The following call may fail, succeed or succeed 1589e22bee78STejun Heo * without actually migrating the task to the cpu if 1590e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 159124647570STejun Heo * against POOL_DISASSOCIATED. 1592e22bee78STejun Heo */ 159324647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 1594ec22ca5eSTejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu)); 1595e22bee78STejun Heo 1596d565ed63STejun Heo spin_lock_irq(&pool->lock); 159724647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1598e22bee78STejun Heo return false; 1599ec22ca5eSTejun Heo if (task_cpu(task) == pool->cpu && 1600e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1601ec22ca5eSTejun Heo get_cpu_mask(pool->cpu))) 1602e22bee78STejun Heo return true; 1603d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1604e22bee78STejun Heo 16055035b20fSTejun Heo /* 16065035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16075035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16085035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16095035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16105035b20fSTejun Heo */ 1611e22bee78STejun Heo cpu_relax(); 16125035b20fSTejun Heo cond_resched(); 1613e22bee78STejun Heo } 1614e22bee78STejun Heo } 1615e22bee78STejun Heo 1616e22bee78STejun Heo /* 1617ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 16185f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 161925511a47STejun Heo */ 162025511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 162125511a47STejun Heo { 16225f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 16235f7dabfdSLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 16245f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 162525511a47STejun Heo 1626ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1627ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1628d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 162925511a47STejun Heo } 163025511a47STejun Heo 163125511a47STejun Heo /* 163225511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1633403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1634403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1635403c821dSTejun Heo * executed twice without intervening cpu down. 1636e22bee78STejun Heo */ 163725511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1638e22bee78STejun Heo { 1639e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1640e22bee78STejun Heo 1641eab6d828SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 1642eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1643e22bee78STejun Heo 1644d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1645e22bee78STejun Heo } 1646e22bee78STejun Heo 164725511a47STejun Heo /** 164894cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 164994cf58bbSTejun Heo * @pool: pool of interest 165025511a47STejun Heo * 165194cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 165225511a47STejun Heo * is different for idle and busy ones. 165325511a47STejun Heo * 1654ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1655ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1656ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1657ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 165825511a47STejun Heo * 1659ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1660ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1661ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1662ea1abd61SLai Jiangshan * rebind. 166325511a47STejun Heo * 1664ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1665ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1666ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1667ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 166825511a47STejun Heo */ 166994cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 167025511a47STejun Heo { 1671ea1abd61SLai Jiangshan struct worker *worker, *n; 167225511a47STejun Heo struct hlist_node *pos; 167325511a47STejun Heo int i; 167425511a47STejun Heo 1675b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 1676d565ed63STejun Heo lockdep_assert_held(&pool->lock); 167725511a47STejun Heo 16785f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1679ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1680ea1abd61SLai Jiangshan /* 168194cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 168294cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1683ea1abd61SLai Jiangshan */ 1684ea1abd61SLai Jiangshan list_del_init(&worker->entry); 168525511a47STejun Heo 168625511a47STejun Heo /* 168794cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 168894cf58bbSTejun Heo * idle_worker_rebind(). 168925511a47STejun Heo */ 169025511a47STejun Heo wake_up_process(worker->task); 169125511a47STejun Heo } 169225511a47STejun Heo 1693ea1abd61SLai Jiangshan /* rebind busy workers */ 1694c9e7cf27STejun Heo for_each_busy_worker(worker, i, pos, pool) { 169525511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1696e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 169725511a47STejun Heo 169825511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 169925511a47STejun Heo work_data_bits(rebind_work))) 170025511a47STejun Heo continue; 170125511a47STejun Heo 170225511a47STejun Heo debug_work_activate(rebind_work); 170390beca5dSTejun Heo 170490beca5dSTejun Heo /* 170594cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 170694cf58bbSTejun Heo * and @cwq->pool consistent for sanity. 170790beca5dSTejun Heo */ 1708e34cdddbSTejun Heo if (std_worker_pool_pri(worker->pool)) 1709e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1710e2b6a6d5SJoonsoo Kim else 1711e2b6a6d5SJoonsoo Kim wq = system_wq; 1712ec58815aSTejun Heo 1713ec22ca5eSTejun Heo insert_work(get_cwq(pool->cpu, wq), rebind_work, 171425511a47STejun Heo worker->scheduled.next, 171525511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1716ec58815aSTejun Heo } 171725511a47STejun Heo } 171825511a47STejun Heo 1719c34056a3STejun Heo static struct worker *alloc_worker(void) 1720c34056a3STejun Heo { 1721c34056a3STejun Heo struct worker *worker; 1722c34056a3STejun Heo 1723c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1724c8e55f36STejun Heo if (worker) { 1725c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1726affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 172725511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1728e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1729e22bee78STejun Heo worker->flags = WORKER_PREP; 1730c8e55f36STejun Heo } 1731c34056a3STejun Heo return worker; 1732c34056a3STejun Heo } 1733c34056a3STejun Heo 1734c34056a3STejun Heo /** 1735c34056a3STejun Heo * create_worker - create a new workqueue worker 173663d95a91STejun Heo * @pool: pool the new worker will belong to 1737c34056a3STejun Heo * 173863d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1739c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1740c34056a3STejun Heo * destroy_worker(). 1741c34056a3STejun Heo * 1742c34056a3STejun Heo * CONTEXT: 1743c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1744c34056a3STejun Heo * 1745c34056a3STejun Heo * RETURNS: 1746c34056a3STejun Heo * Pointer to the newly created worker. 1747c34056a3STejun Heo */ 1748bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1749c34056a3STejun Heo { 1750e34cdddbSTejun Heo const char *pri = std_worker_pool_pri(pool) ? "H" : ""; 1751c34056a3STejun Heo struct worker *worker = NULL; 1752f3421797STejun Heo int id = -1; 1753c34056a3STejun Heo 1754d565ed63STejun Heo spin_lock_irq(&pool->lock); 1755bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1756d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1757bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1758c34056a3STejun Heo goto fail; 1759d565ed63STejun Heo spin_lock_irq(&pool->lock); 1760c34056a3STejun Heo } 1761d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1762c34056a3STejun Heo 1763c34056a3STejun Heo worker = alloc_worker(); 1764c34056a3STejun Heo if (!worker) 1765c34056a3STejun Heo goto fail; 1766c34056a3STejun Heo 1767bd7bdd43STejun Heo worker->pool = pool; 1768c34056a3STejun Heo worker->id = id; 1769c34056a3STejun Heo 1770ec22ca5eSTejun Heo if (pool->cpu != WORK_CPU_UNBOUND) 177194dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1772ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1773ec22ca5eSTejun Heo "kworker/%u:%d%s", pool->cpu, id, pri); 1774f3421797STejun Heo else 1775f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 17763270476aSTejun Heo "kworker/u:%d%s", id, pri); 1777c34056a3STejun Heo if (IS_ERR(worker->task)) 1778c34056a3STejun Heo goto fail; 1779c34056a3STejun Heo 1780e34cdddbSTejun Heo if (std_worker_pool_pri(pool)) 17813270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 17823270476aSTejun Heo 1783db7bccf4STejun Heo /* 1784bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 178524647570STejun Heo * %POOL_DISASSOCIATED. The caller is responsible for ensuring the 1786bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1787bc2ae0f5STejun Heo * above the flag definition for details. 1788bc2ae0f5STejun Heo * 1789bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1790bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1791db7bccf4STejun Heo */ 179224647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) { 1793ec22ca5eSTejun Heo kthread_bind(worker->task, pool->cpu); 1794bc2ae0f5STejun Heo } else { 1795db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1796f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1797f3421797STejun Heo } 1798c34056a3STejun Heo 1799c34056a3STejun Heo return worker; 1800c34056a3STejun Heo fail: 1801c34056a3STejun Heo if (id >= 0) { 1802d565ed63STejun Heo spin_lock_irq(&pool->lock); 1803bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1804d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1805c34056a3STejun Heo } 1806c34056a3STejun Heo kfree(worker); 1807c34056a3STejun Heo return NULL; 1808c34056a3STejun Heo } 1809c34056a3STejun Heo 1810c34056a3STejun Heo /** 1811c34056a3STejun Heo * start_worker - start a newly created worker 1812c34056a3STejun Heo * @worker: worker to start 1813c34056a3STejun Heo * 1814706026c2STejun Heo * Make the pool aware of @worker and start it. 1815c34056a3STejun Heo * 1816c34056a3STejun Heo * CONTEXT: 1817d565ed63STejun Heo * spin_lock_irq(pool->lock). 1818c34056a3STejun Heo */ 1819c34056a3STejun Heo static void start_worker(struct worker *worker) 1820c34056a3STejun Heo { 1821cb444766STejun Heo worker->flags |= WORKER_STARTED; 1822bd7bdd43STejun Heo worker->pool->nr_workers++; 1823c8e55f36STejun Heo worker_enter_idle(worker); 1824c34056a3STejun Heo wake_up_process(worker->task); 1825c34056a3STejun Heo } 1826c34056a3STejun Heo 1827c34056a3STejun Heo /** 1828c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1829c34056a3STejun Heo * @worker: worker to be destroyed 1830c34056a3STejun Heo * 1831706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1832c8e55f36STejun Heo * 1833c8e55f36STejun Heo * CONTEXT: 1834d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1835c34056a3STejun Heo */ 1836c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1837c34056a3STejun Heo { 1838bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1839c34056a3STejun Heo int id = worker->id; 1840c34056a3STejun Heo 1841c34056a3STejun Heo /* sanity check frenzy */ 1842c34056a3STejun Heo BUG_ON(worker->current_work); 1843affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1844c34056a3STejun Heo 1845c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1846bd7bdd43STejun Heo pool->nr_workers--; 1847c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1848bd7bdd43STejun Heo pool->nr_idle--; 1849c8e55f36STejun Heo 1850c8e55f36STejun Heo list_del_init(&worker->entry); 1851cb444766STejun Heo worker->flags |= WORKER_DIE; 1852c8e55f36STejun Heo 1853d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1854c8e55f36STejun Heo 1855c34056a3STejun Heo kthread_stop(worker->task); 1856c34056a3STejun Heo kfree(worker); 1857c34056a3STejun Heo 1858d565ed63STejun Heo spin_lock_irq(&pool->lock); 1859bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1860c34056a3STejun Heo } 1861c34056a3STejun Heo 186263d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1863e22bee78STejun Heo { 186463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1865e22bee78STejun Heo 1866d565ed63STejun Heo spin_lock_irq(&pool->lock); 1867e22bee78STejun Heo 186863d95a91STejun Heo if (too_many_workers(pool)) { 1869e22bee78STejun Heo struct worker *worker; 1870e22bee78STejun Heo unsigned long expires; 1871e22bee78STejun Heo 1872e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 187363d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1874e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1875e22bee78STejun Heo 1876e22bee78STejun Heo if (time_before(jiffies, expires)) 187763d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1878e22bee78STejun Heo else { 1879e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 188011ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 188163d95a91STejun Heo wake_up_worker(pool); 1882e22bee78STejun Heo } 1883e22bee78STejun Heo } 1884e22bee78STejun Heo 1885d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1886e22bee78STejun Heo } 1887e22bee78STejun Heo 1888e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1889e22bee78STejun Heo { 1890e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1891e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1892f3421797STejun Heo unsigned int cpu; 1893e22bee78STejun Heo 1894e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1895e22bee78STejun Heo return false; 1896e22bee78STejun Heo 1897e22bee78STejun Heo /* mayday mayday mayday */ 1898ec22ca5eSTejun Heo cpu = cwq->pool->cpu; 1899f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1900f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1901f3421797STejun Heo cpu = 0; 1902f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1903e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1904e22bee78STejun Heo return true; 1905e22bee78STejun Heo } 1906e22bee78STejun Heo 1907706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1908e22bee78STejun Heo { 190963d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1910e22bee78STejun Heo struct work_struct *work; 1911e22bee78STejun Heo 1912d565ed63STejun Heo spin_lock_irq(&pool->lock); 1913e22bee78STejun Heo 191463d95a91STejun Heo if (need_to_create_worker(pool)) { 1915e22bee78STejun Heo /* 1916e22bee78STejun Heo * We've been trying to create a new worker but 1917e22bee78STejun Heo * haven't been successful. We might be hitting an 1918e22bee78STejun Heo * allocation deadlock. Send distress signals to 1919e22bee78STejun Heo * rescuers. 1920e22bee78STejun Heo */ 192163d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1922e22bee78STejun Heo send_mayday(work); 1923e22bee78STejun Heo } 1924e22bee78STejun Heo 1925d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1926e22bee78STejun Heo 192763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1928e22bee78STejun Heo } 1929e22bee78STejun Heo 1930e22bee78STejun Heo /** 1931e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 193263d95a91STejun Heo * @pool: pool to create a new worker for 1933e22bee78STejun Heo * 193463d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1935e22bee78STejun Heo * have at least one idle worker on return from this function. If 1936e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 193763d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1938e22bee78STejun Heo * possible allocation deadlock. 1939e22bee78STejun Heo * 1940e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1941e22bee78STejun Heo * may_start_working() true. 1942e22bee78STejun Heo * 1943e22bee78STejun Heo * LOCKING: 1944d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1945e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1946e22bee78STejun Heo * manager. 1947e22bee78STejun Heo * 1948e22bee78STejun Heo * RETURNS: 1949d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1950e22bee78STejun Heo * otherwise. 1951e22bee78STejun Heo */ 195263d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1953d565ed63STejun Heo __releases(&pool->lock) 1954d565ed63STejun Heo __acquires(&pool->lock) 1955e22bee78STejun Heo { 195663d95a91STejun Heo if (!need_to_create_worker(pool)) 1957e22bee78STejun Heo return false; 1958e22bee78STejun Heo restart: 1959d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19609f9c2364STejun Heo 1961e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 196263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1963e22bee78STejun Heo 1964e22bee78STejun Heo while (true) { 1965e22bee78STejun Heo struct worker *worker; 1966e22bee78STejun Heo 1967bc2ae0f5STejun Heo worker = create_worker(pool); 1968e22bee78STejun Heo if (worker) { 196963d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1970d565ed63STejun Heo spin_lock_irq(&pool->lock); 1971e22bee78STejun Heo start_worker(worker); 197263d95a91STejun Heo BUG_ON(need_to_create_worker(pool)); 1973e22bee78STejun Heo return true; 1974e22bee78STejun Heo } 1975e22bee78STejun Heo 197663d95a91STejun Heo if (!need_to_create_worker(pool)) 1977e22bee78STejun Heo break; 1978e22bee78STejun Heo 1979e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1980e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19819f9c2364STejun Heo 198263d95a91STejun Heo if (!need_to_create_worker(pool)) 1983e22bee78STejun Heo break; 1984e22bee78STejun Heo } 1985e22bee78STejun Heo 198663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1987d565ed63STejun Heo spin_lock_irq(&pool->lock); 198863d95a91STejun Heo if (need_to_create_worker(pool)) 1989e22bee78STejun Heo goto restart; 1990e22bee78STejun Heo return true; 1991e22bee78STejun Heo } 1992e22bee78STejun Heo 1993e22bee78STejun Heo /** 1994e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 199563d95a91STejun Heo * @pool: pool to destroy workers for 1996e22bee78STejun Heo * 199763d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1998e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1999e22bee78STejun Heo * 2000e22bee78STejun Heo * LOCKING: 2001d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2002e22bee78STejun Heo * multiple times. Called only from manager. 2003e22bee78STejun Heo * 2004e22bee78STejun Heo * RETURNS: 2005d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 2006e22bee78STejun Heo * otherwise. 2007e22bee78STejun Heo */ 200863d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2009e22bee78STejun Heo { 2010e22bee78STejun Heo bool ret = false; 2011e22bee78STejun Heo 201263d95a91STejun Heo while (too_many_workers(pool)) { 2013e22bee78STejun Heo struct worker *worker; 2014e22bee78STejun Heo unsigned long expires; 2015e22bee78STejun Heo 201663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2017e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2018e22bee78STejun Heo 2019e22bee78STejun Heo if (time_before(jiffies, expires)) { 202063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2021e22bee78STejun Heo break; 2022e22bee78STejun Heo } 2023e22bee78STejun Heo 2024e22bee78STejun Heo destroy_worker(worker); 2025e22bee78STejun Heo ret = true; 2026e22bee78STejun Heo } 2027e22bee78STejun Heo 2028e22bee78STejun Heo return ret; 2029e22bee78STejun Heo } 2030e22bee78STejun Heo 2031e22bee78STejun Heo /** 2032e22bee78STejun Heo * manage_workers - manage worker pool 2033e22bee78STejun Heo * @worker: self 2034e22bee78STejun Heo * 2035706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2036e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2037706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2038e22bee78STejun Heo * 2039e22bee78STejun Heo * The caller can safely start processing works on false return. On 2040e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2041e22bee78STejun Heo * and may_start_working() is true. 2042e22bee78STejun Heo * 2043e22bee78STejun Heo * CONTEXT: 2044d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2045e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2046e22bee78STejun Heo * 2047e22bee78STejun Heo * RETURNS: 2048d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2049d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2050e22bee78STejun Heo */ 2051e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2052e22bee78STejun Heo { 205363d95a91STejun Heo struct worker_pool *pool = worker->pool; 2054e22bee78STejun Heo bool ret = false; 2055e22bee78STejun Heo 2056ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2057e22bee78STejun Heo return ret; 2058e22bee78STejun Heo 2059552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2060ee378aa4SLai Jiangshan 2061ee378aa4SLai Jiangshan /* 2062ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2063ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2064ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2065ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2066ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2067b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2068ee378aa4SLai Jiangshan * manager against CPU hotplug. 2069ee378aa4SLai Jiangshan * 2070b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2071d565ed63STejun Heo * progress. trylock first without dropping @pool->lock. 2072ee378aa4SLai Jiangshan */ 2073b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2074d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2075b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2076ee378aa4SLai Jiangshan /* 2077ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2078b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2079ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2080706026c2STejun Heo * @pool's state and ours could have deviated. 2081ee378aa4SLai Jiangshan * 2082b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2083ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2084706026c2STejun Heo * on @pool's current state. Try it and adjust 2085ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2086ee378aa4SLai Jiangshan */ 2087ee378aa4SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 2088ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2089ee378aa4SLai Jiangshan else 2090ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2091ee378aa4SLai Jiangshan 2092ee378aa4SLai Jiangshan ret = true; 2093ee378aa4SLai Jiangshan } 2094ee378aa4SLai Jiangshan 209511ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2096e22bee78STejun Heo 2097e22bee78STejun Heo /* 2098e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2099e22bee78STejun Heo * on return. 2100e22bee78STejun Heo */ 210163d95a91STejun Heo ret |= maybe_destroy_workers(pool); 210263d95a91STejun Heo ret |= maybe_create_worker(pool); 2103e22bee78STejun Heo 2104552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2105b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2106e22bee78STejun Heo return ret; 2107e22bee78STejun Heo } 2108e22bee78STejun Heo 2109a62428c0STejun Heo /** 2110a62428c0STejun Heo * process_one_work - process single work 2111c34056a3STejun Heo * @worker: self 2112a62428c0STejun Heo * @work: work to process 2113a62428c0STejun Heo * 2114a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2115a62428c0STejun Heo * process a single work including synchronization against and 2116a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2117a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2118a62428c0STejun Heo * call this function to process a work. 2119a62428c0STejun Heo * 2120a62428c0STejun Heo * CONTEXT: 2121d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2122a62428c0STejun Heo */ 2123c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2124d565ed63STejun Heo __releases(&pool->lock) 2125d565ed63STejun Heo __acquires(&pool->lock) 21261da177e4SLinus Torvalds { 21277e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 2128bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2129fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 213073f53c4aSTejun Heo int work_color; 21317e11629dSTejun Heo struct worker *collision; 21324e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21334e6045f1SJohannes Berg /* 2134a62428c0STejun Heo * It is permissible to free the struct work_struct from 2135a62428c0STejun Heo * inside the function that is called from it, this we need to 2136a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2137a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2138a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21394e6045f1SJohannes Berg */ 21404d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21414d82a1deSPeter Zijlstra 21424d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21434e6045f1SJohannes Berg #endif 21446fec10a1STejun Heo /* 21456fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21466fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 214724647570STejun Heo * unbound or a disassociated pool. 21486fec10a1STejun Heo */ 21495f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 215024647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2151ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 215225511a47STejun Heo 21537e11629dSTejun Heo /* 21547e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21557e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21567e11629dSTejun Heo * already processing the work. If so, defer the work to the 21577e11629dSTejun Heo * currently executing one. 21587e11629dSTejun Heo */ 2159c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21607e11629dSTejun Heo if (unlikely(collision)) { 21617e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21627e11629dSTejun Heo return; 21637e11629dSTejun Heo } 21641da177e4SLinus Torvalds 21658930cabaSTejun Heo /* claim and dequeue */ 21661da177e4SLinus Torvalds debug_work_deactivate(work); 2167c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2168c34056a3STejun Heo worker->current_work = work; 2169a2c1c57bSTejun Heo worker->current_func = work->func; 21708cca0eeaSTejun Heo worker->current_cwq = cwq; 217173f53c4aSTejun Heo work_color = get_work_color(work); 21727a22ad75STejun Heo 2173a62428c0STejun Heo list_del_init(&work->entry); 2174a62428c0STejun Heo 2175649027d7STejun Heo /* 2176fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2177fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2178fb0e7bebSTejun Heo */ 2179fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2180fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2181fb0e7bebSTejun Heo 2182974271c4STejun Heo /* 2183d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2184974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2185974271c4STejun Heo */ 218663d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 218763d95a91STejun Heo wake_up_worker(pool); 2188974271c4STejun Heo 21898930cabaSTejun Heo /* 21907c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2191d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 219223657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 219323657bb1STejun Heo * disabled. 21948930cabaSTejun Heo */ 21957c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21961da177e4SLinus Torvalds 2197d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2198365970a1SDavid Howells 2199e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 22003295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2201e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2202a2c1c57bSTejun Heo worker->current_func(work); 2203e36c886aSArjan van de Ven /* 2204e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2205e36c886aSArjan van de Ven * point will only record its address. 2206e36c886aSArjan van de Ven */ 2207e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22083295f0efSIngo Molnar lock_map_release(&lockdep_map); 22093295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 22101da177e4SLinus Torvalds 2211d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2212044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2213044c782cSValentin Ilie " last function: %pf\n", 2214a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2215a2c1c57bSTejun Heo worker->current_func); 2216d5abe669SPeter Zijlstra debug_show_held_locks(current); 2217d5abe669SPeter Zijlstra dump_stack(); 2218d5abe669SPeter Zijlstra } 2219d5abe669SPeter Zijlstra 2220d565ed63STejun Heo spin_lock_irq(&pool->lock); 2221a62428c0STejun Heo 2222fb0e7bebSTejun Heo /* clear cpu intensive status */ 2223fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2224fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2225fb0e7bebSTejun Heo 2226a62428c0STejun Heo /* we're done with it, release */ 222742f8570fSSasha Levin hash_del(&worker->hentry); 2228c34056a3STejun Heo worker->current_work = NULL; 2229a2c1c57bSTejun Heo worker->current_func = NULL; 22308cca0eeaSTejun Heo worker->current_cwq = NULL; 2231b3f9f405SLai Jiangshan cwq_dec_nr_in_flight(cwq, work_color); 22321da177e4SLinus Torvalds } 22331da177e4SLinus Torvalds 2234affee4b2STejun Heo /** 2235affee4b2STejun Heo * process_scheduled_works - process scheduled works 2236affee4b2STejun Heo * @worker: self 2237affee4b2STejun Heo * 2238affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2239affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2240affee4b2STejun Heo * fetches a work from the top and executes it. 2241affee4b2STejun Heo * 2242affee4b2STejun Heo * CONTEXT: 2243d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2244affee4b2STejun Heo * multiple times. 2245affee4b2STejun Heo */ 2246affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22471da177e4SLinus Torvalds { 2248affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2249affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2250a62428c0STejun Heo struct work_struct, entry); 2251c34056a3STejun Heo process_one_work(worker, work); 2252a62428c0STejun Heo } 22531da177e4SLinus Torvalds } 22541da177e4SLinus Torvalds 22554690c4abSTejun Heo /** 22564690c4abSTejun Heo * worker_thread - the worker thread function 2257c34056a3STejun Heo * @__worker: self 22584690c4abSTejun Heo * 2259706026c2STejun Heo * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools 2260706026c2STejun Heo * of these per each cpu. These workers process all works regardless of 2261e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2262e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2263e22bee78STejun Heo * rescuer_thread(). 22644690c4abSTejun Heo */ 2265c34056a3STejun Heo static int worker_thread(void *__worker) 22661da177e4SLinus Torvalds { 2267c34056a3STejun Heo struct worker *worker = __worker; 2268bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22691da177e4SLinus Torvalds 2270e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2271e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2272c8e55f36STejun Heo woke_up: 2273d565ed63STejun Heo spin_lock_irq(&pool->lock); 2274affee4b2STejun Heo 22755f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22765f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2277d565ed63STejun Heo spin_unlock_irq(&pool->lock); 227825511a47STejun Heo 22795f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 228025511a47STejun Heo if (worker->flags & WORKER_DIE) { 2281e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2282c8e55f36STejun Heo return 0; 2283c8e55f36STejun Heo } 2284c8e55f36STejun Heo 22855f7dabfdSLai Jiangshan /* otherwise, rebind */ 228625511a47STejun Heo idle_worker_rebind(worker); 228725511a47STejun Heo goto woke_up; 228825511a47STejun Heo } 228925511a47STejun Heo 2290c8e55f36STejun Heo worker_leave_idle(worker); 2291db7bccf4STejun Heo recheck: 2292e22bee78STejun Heo /* no more worker necessary? */ 229363d95a91STejun Heo if (!need_more_worker(pool)) 2294e22bee78STejun Heo goto sleep; 2295e22bee78STejun Heo 2296e22bee78STejun Heo /* do we need to manage? */ 229763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2298e22bee78STejun Heo goto recheck; 2299e22bee78STejun Heo 2300c8e55f36STejun Heo /* 2301c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2302c8e55f36STejun Heo * preparing to process a work or actually processing it. 2303c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2304c8e55f36STejun Heo */ 2305c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 2306c8e55f36STejun Heo 2307e22bee78STejun Heo /* 2308e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2309e22bee78STejun Heo * at least one idle worker or that someone else has already 2310e22bee78STejun Heo * assumed the manager role. 2311e22bee78STejun Heo */ 2312e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2313e22bee78STejun Heo 2314e22bee78STejun Heo do { 2315affee4b2STejun Heo struct work_struct *work = 2316bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2317affee4b2STejun Heo struct work_struct, entry); 2318affee4b2STejun Heo 2319c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2320affee4b2STejun Heo /* optimization path, not strictly necessary */ 2321affee4b2STejun Heo process_one_work(worker, work); 2322affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2323affee4b2STejun Heo process_scheduled_works(worker); 2324affee4b2STejun Heo } else { 2325c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2326affee4b2STejun Heo process_scheduled_works(worker); 2327affee4b2STejun Heo } 232863d95a91STejun Heo } while (keep_working(pool)); 2329affee4b2STejun Heo 2330e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2331d313dd85STejun Heo sleep: 233263d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2333e22bee78STejun Heo goto recheck; 2334d313dd85STejun Heo 2335c8e55f36STejun Heo /* 2336d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2337d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2338d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2339d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2340d565ed63STejun Heo * event. 2341c8e55f36STejun Heo */ 2342c8e55f36STejun Heo worker_enter_idle(worker); 2343c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2344d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23451da177e4SLinus Torvalds schedule(); 2346c8e55f36STejun Heo goto woke_up; 23471da177e4SLinus Torvalds } 23481da177e4SLinus Torvalds 2349e22bee78STejun Heo /** 2350e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2351111c225aSTejun Heo * @__rescuer: self 2352e22bee78STejun Heo * 2353e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2354e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2355e22bee78STejun Heo * 2356706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2357e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2358e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2359e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2360e22bee78STejun Heo * the problem rescuer solves. 2361e22bee78STejun Heo * 2362706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2363706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2364e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2365e22bee78STejun Heo * 2366e22bee78STejun Heo * This should happen rarely. 2367e22bee78STejun Heo */ 2368111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2369e22bee78STejun Heo { 2370111c225aSTejun Heo struct worker *rescuer = __rescuer; 2371111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2372e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2373f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2374e22bee78STejun Heo unsigned int cpu; 2375e22bee78STejun Heo 2376e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2377111c225aSTejun Heo 2378111c225aSTejun Heo /* 2379111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2380111c225aSTejun Heo * doesn't participate in concurrency management. 2381111c225aSTejun Heo */ 2382111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2383e22bee78STejun Heo repeat: 2384e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23851da177e4SLinus Torvalds 2386412d32e6SMike Galbraith if (kthread_should_stop()) { 2387412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2388111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2389e22bee78STejun Heo return 0; 2390412d32e6SMike Galbraith } 23911da177e4SLinus Torvalds 2392f3421797STejun Heo /* 2393f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2394f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2395f3421797STejun Heo */ 2396f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2397f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2398f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2399bd7bdd43STejun Heo struct worker_pool *pool = cwq->pool; 2400e22bee78STejun Heo struct work_struct *work, *n; 2401e22bee78STejun Heo 2402e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2403f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2404e22bee78STejun Heo 2405e22bee78STejun Heo /* migrate to the target cpu if possible */ 2406bd7bdd43STejun Heo rescuer->pool = pool; 2407e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2408e22bee78STejun Heo 2409e22bee78STejun Heo /* 2410e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2411e22bee78STejun Heo * process'em. 2412e22bee78STejun Heo */ 2413e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2414bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2415e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2416e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2417e22bee78STejun Heo 2418e22bee78STejun Heo process_scheduled_works(rescuer); 24197576958aSTejun Heo 24207576958aSTejun Heo /* 2421d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24227576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24237576958aSTejun Heo * and stalling the execution. 24247576958aSTejun Heo */ 242563d95a91STejun Heo if (keep_working(pool)) 242663d95a91STejun Heo wake_up_worker(pool); 24277576958aSTejun Heo 2428d565ed63STejun Heo spin_unlock_irq(&pool->lock); 24291da177e4SLinus Torvalds } 24301da177e4SLinus Torvalds 2431111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2432111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2433e22bee78STejun Heo schedule(); 2434e22bee78STejun Heo goto repeat; 24351da177e4SLinus Torvalds } 24361da177e4SLinus Torvalds 2437fc2e4d70SOleg Nesterov struct wq_barrier { 2438fc2e4d70SOleg Nesterov struct work_struct work; 2439fc2e4d70SOleg Nesterov struct completion done; 2440fc2e4d70SOleg Nesterov }; 2441fc2e4d70SOleg Nesterov 2442fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2443fc2e4d70SOleg Nesterov { 2444fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2445fc2e4d70SOleg Nesterov complete(&barr->done); 2446fc2e4d70SOleg Nesterov } 2447fc2e4d70SOleg Nesterov 24484690c4abSTejun Heo /** 24494690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 24504690c4abSTejun Heo * @cwq: cwq to insert barrier into 24514690c4abSTejun Heo * @barr: wq_barrier to insert 2452affee4b2STejun Heo * @target: target work to attach @barr to 2453affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24544690c4abSTejun Heo * 2455affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2456affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2457affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2458affee4b2STejun Heo * cpu. 2459affee4b2STejun Heo * 2460affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2461affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2462affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2463affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2464affee4b2STejun Heo * after a work with LINKED flag set. 2465affee4b2STejun Heo * 2466affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2467affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 24684690c4abSTejun Heo * 24694690c4abSTejun Heo * CONTEXT: 2470d565ed63STejun Heo * spin_lock_irq(pool->lock). 24714690c4abSTejun Heo */ 247283c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2473affee4b2STejun Heo struct wq_barrier *barr, 2474affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2475fc2e4d70SOleg Nesterov { 2476affee4b2STejun Heo struct list_head *head; 2477affee4b2STejun Heo unsigned int linked = 0; 2478affee4b2STejun Heo 2479dc186ad7SThomas Gleixner /* 2480d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2481dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2482dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2483dc186ad7SThomas Gleixner * might deadlock. 2484dc186ad7SThomas Gleixner */ 2485ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 248622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2487fc2e4d70SOleg Nesterov init_completion(&barr->done); 248883c22520SOleg Nesterov 2489affee4b2STejun Heo /* 2490affee4b2STejun Heo * If @target is currently being executed, schedule the 2491affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2492affee4b2STejun Heo */ 2493affee4b2STejun Heo if (worker) 2494affee4b2STejun Heo head = worker->scheduled.next; 2495affee4b2STejun Heo else { 2496affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2497affee4b2STejun Heo 2498affee4b2STejun Heo head = target->entry.next; 2499affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2500affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2501affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2502affee4b2STejun Heo } 2503affee4b2STejun Heo 2504dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2505affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2506affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2507fc2e4d70SOleg Nesterov } 2508fc2e4d70SOleg Nesterov 250973f53c4aSTejun Heo /** 251073f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 251173f53c4aSTejun Heo * @wq: workqueue being flushed 251273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 251373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 251473f53c4aSTejun Heo * 251573f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 251673f53c4aSTejun Heo * 251773f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 251873f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 251973f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 252073f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 252173f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 252273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 252373f53c4aSTejun Heo * 252473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 252573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 252673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 252773f53c4aSTejun Heo * is returned. 252873f53c4aSTejun Heo * 252973f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 253073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 253173f53c4aSTejun Heo * advanced to @work_color. 253273f53c4aSTejun Heo * 253373f53c4aSTejun Heo * CONTEXT: 253473f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 253573f53c4aSTejun Heo * 253673f53c4aSTejun Heo * RETURNS: 253773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 253873f53c4aSTejun Heo * otherwise. 253973f53c4aSTejun Heo */ 254073f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 254173f53c4aSTejun Heo int flush_color, int work_color) 25421da177e4SLinus Torvalds { 254373f53c4aSTejun Heo bool wait = false; 254473f53c4aSTejun Heo unsigned int cpu; 25451da177e4SLinus Torvalds 254673f53c4aSTejun Heo if (flush_color >= 0) { 254773f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 254873f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2549dc186ad7SThomas Gleixner } 255014441960SOleg Nesterov 2551f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 255273f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2553d565ed63STejun Heo struct worker_pool *pool = cwq->pool; 25541da177e4SLinus Torvalds 2555d565ed63STejun Heo spin_lock_irq(&pool->lock); 255673f53c4aSTejun Heo 255773f53c4aSTejun Heo if (flush_color >= 0) { 255873f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 255973f53c4aSTejun Heo 256073f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 256173f53c4aSTejun Heo cwq->flush_color = flush_color; 256273f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 256373f53c4aSTejun Heo wait = true; 25641da177e4SLinus Torvalds } 256573f53c4aSTejun Heo } 256673f53c4aSTejun Heo 256773f53c4aSTejun Heo if (work_color >= 0) { 256873f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 256973f53c4aSTejun Heo cwq->work_color = work_color; 257073f53c4aSTejun Heo } 257173f53c4aSTejun Heo 2572d565ed63STejun Heo spin_unlock_irq(&pool->lock); 25731da177e4SLinus Torvalds } 25741da177e4SLinus Torvalds 257573f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 257673f53c4aSTejun Heo complete(&wq->first_flusher->done); 257773f53c4aSTejun Heo 257873f53c4aSTejun Heo return wait; 257983c22520SOleg Nesterov } 25801da177e4SLinus Torvalds 25810fcb78c2SRolf Eike Beer /** 25821da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25830fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25841da177e4SLinus Torvalds * 25851da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 25861da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 25871da177e4SLinus Torvalds * 2588fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2589fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 25901da177e4SLinus Torvalds */ 25917ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25921da177e4SLinus Torvalds { 259373f53c4aSTejun Heo struct wq_flusher this_flusher = { 259473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 259573f53c4aSTejun Heo .flush_color = -1, 259673f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 259773f53c4aSTejun Heo }; 259873f53c4aSTejun Heo int next_color; 2599b1f4ec17SOleg Nesterov 26003295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26013295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 260273f53c4aSTejun Heo 260373f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 260473f53c4aSTejun Heo 260573f53c4aSTejun Heo /* 260673f53c4aSTejun Heo * Start-to-wait phase 260773f53c4aSTejun Heo */ 260873f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 260973f53c4aSTejun Heo 261073f53c4aSTejun Heo if (next_color != wq->flush_color) { 261173f53c4aSTejun Heo /* 261273f53c4aSTejun Heo * Color space is not full. The current work_color 261373f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 261473f53c4aSTejun Heo * by one. 261573f53c4aSTejun Heo */ 261673f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 261773f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 261873f53c4aSTejun Heo wq->work_color = next_color; 261973f53c4aSTejun Heo 262073f53c4aSTejun Heo if (!wq->first_flusher) { 262173f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 262273f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 262373f53c4aSTejun Heo 262473f53c4aSTejun Heo wq->first_flusher = &this_flusher; 262573f53c4aSTejun Heo 262673f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 262773f53c4aSTejun Heo wq->work_color)) { 262873f53c4aSTejun Heo /* nothing to flush, done */ 262973f53c4aSTejun Heo wq->flush_color = next_color; 263073f53c4aSTejun Heo wq->first_flusher = NULL; 263173f53c4aSTejun Heo goto out_unlock; 263273f53c4aSTejun Heo } 263373f53c4aSTejun Heo } else { 263473f53c4aSTejun Heo /* wait in queue */ 263573f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 263673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 263773f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 263873f53c4aSTejun Heo } 263973f53c4aSTejun Heo } else { 264073f53c4aSTejun Heo /* 264173f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 264273f53c4aSTejun Heo * The next flush completion will assign us 264373f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 264473f53c4aSTejun Heo */ 264573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 264673f53c4aSTejun Heo } 264773f53c4aSTejun Heo 264873f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 264973f53c4aSTejun Heo 265073f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 265173f53c4aSTejun Heo 265273f53c4aSTejun Heo /* 265373f53c4aSTejun Heo * Wake-up-and-cascade phase 265473f53c4aSTejun Heo * 265573f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 265673f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 265773f53c4aSTejun Heo */ 265873f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 265973f53c4aSTejun Heo return; 266073f53c4aSTejun Heo 266173f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 266273f53c4aSTejun Heo 26634ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26644ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26654ce48b37STejun Heo goto out_unlock; 26664ce48b37STejun Heo 266773f53c4aSTejun Heo wq->first_flusher = NULL; 266873f53c4aSTejun Heo 266973f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 267073f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 267173f53c4aSTejun Heo 267273f53c4aSTejun Heo while (true) { 267373f53c4aSTejun Heo struct wq_flusher *next, *tmp; 267473f53c4aSTejun Heo 267573f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 267673f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 267773f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 267873f53c4aSTejun Heo break; 267973f53c4aSTejun Heo list_del_init(&next->list); 268073f53c4aSTejun Heo complete(&next->done); 268173f53c4aSTejun Heo } 268273f53c4aSTejun Heo 268373f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 268473f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 268573f53c4aSTejun Heo 268673f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 268773f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 268873f53c4aSTejun Heo 268973f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 269073f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 269173f53c4aSTejun Heo /* 269273f53c4aSTejun Heo * Assign the same color to all overflowed 269373f53c4aSTejun Heo * flushers, advance work_color and append to 269473f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 269573f53c4aSTejun Heo * phase for these overflowed flushers. 269673f53c4aSTejun Heo */ 269773f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 269873f53c4aSTejun Heo tmp->flush_color = wq->work_color; 269973f53c4aSTejun Heo 270073f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 270373f53c4aSTejun Heo &wq->flusher_queue); 270473f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 270573f53c4aSTejun Heo } 270673f53c4aSTejun Heo 270773f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 270873f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 270973f53c4aSTejun Heo break; 271073f53c4aSTejun Heo } 271173f53c4aSTejun Heo 271273f53c4aSTejun Heo /* 271373f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 271473f53c4aSTejun Heo * the new first flusher and arm cwqs. 271573f53c4aSTejun Heo */ 271673f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 271773f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 271873f53c4aSTejun Heo 271973f53c4aSTejun Heo list_del_init(&next->list); 272073f53c4aSTejun Heo wq->first_flusher = next; 272173f53c4aSTejun Heo 272273f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 272373f53c4aSTejun Heo break; 272473f53c4aSTejun Heo 272573f53c4aSTejun Heo /* 272673f53c4aSTejun Heo * Meh... this color is already done, clear first 272773f53c4aSTejun Heo * flusher and repeat cascading. 272873f53c4aSTejun Heo */ 272973f53c4aSTejun Heo wq->first_flusher = NULL; 273073f53c4aSTejun Heo } 273173f53c4aSTejun Heo 273273f53c4aSTejun Heo out_unlock: 273373f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27341da177e4SLinus Torvalds } 2735ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27361da177e4SLinus Torvalds 27379c5a2ba7STejun Heo /** 27389c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27399c5a2ba7STejun Heo * @wq: workqueue to drain 27409c5a2ba7STejun Heo * 27419c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27429c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27439c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27449c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27459c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27469c5a2ba7STejun Heo * takes too long. 27479c5a2ba7STejun Heo */ 27489c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27499c5a2ba7STejun Heo { 27509c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27519c5a2ba7STejun Heo unsigned int cpu; 27529c5a2ba7STejun Heo 27539c5a2ba7STejun Heo /* 27549c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27559c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27569c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27579c5a2ba7STejun Heo */ 27589c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27599c5a2ba7STejun Heo if (!wq->nr_drainers++) 27609c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 27619c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27629c5a2ba7STejun Heo reflush: 27639c5a2ba7STejun Heo flush_workqueue(wq); 27649c5a2ba7STejun Heo 27659c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 27669c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2767fa2563e4SThomas Tuttle bool drained; 27689c5a2ba7STejun Heo 2769d565ed63STejun Heo spin_lock_irq(&cwq->pool->lock); 2770fa2563e4SThomas Tuttle drained = !cwq->nr_active && list_empty(&cwq->delayed_works); 2771d565ed63STejun Heo spin_unlock_irq(&cwq->pool->lock); 2772fa2563e4SThomas Tuttle 2773fa2563e4SThomas Tuttle if (drained) 27749c5a2ba7STejun Heo continue; 27759c5a2ba7STejun Heo 27769c5a2ba7STejun Heo if (++flush_cnt == 10 || 27779c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2778044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 27799c5a2ba7STejun Heo wq->name, flush_cnt); 27809c5a2ba7STejun Heo goto reflush; 27819c5a2ba7STejun Heo } 27829c5a2ba7STejun Heo 27839c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27849c5a2ba7STejun Heo if (!--wq->nr_drainers) 27859c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 27869c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27879c5a2ba7STejun Heo } 27889c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27899c5a2ba7STejun Heo 2790606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2791baf59022STejun Heo { 2792baf59022STejun Heo struct worker *worker = NULL; 2793c9e7cf27STejun Heo struct worker_pool *pool; 2794baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2795baf59022STejun Heo 2796baf59022STejun Heo might_sleep(); 2797c9e7cf27STejun Heo pool = get_work_pool(work); 2798c9e7cf27STejun Heo if (!pool) 2799baf59022STejun Heo return false; 2800baf59022STejun Heo 2801d565ed63STejun Heo spin_lock_irq(&pool->lock); 2802baf59022STejun Heo if (!list_empty(&work->entry)) { 2803baf59022STejun Heo /* 2804baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2805d565ed63STejun Heo * If it was re-queued to a different pool under us, we 2806baf59022STejun Heo * are not going to wait. 2807baf59022STejun Heo */ 2808baf59022STejun Heo smp_rmb(); 2809baf59022STejun Heo cwq = get_work_cwq(work); 2810d565ed63STejun Heo if (unlikely(!cwq || pool != cwq->pool)) 2811baf59022STejun Heo goto already_gone; 2812606a5020STejun Heo } else { 2813c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2814baf59022STejun Heo if (!worker) 2815baf59022STejun Heo goto already_gone; 2816baf59022STejun Heo cwq = worker->current_cwq; 2817606a5020STejun Heo } 2818baf59022STejun Heo 2819baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2820d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2821baf59022STejun Heo 2822e159489bSTejun Heo /* 2823e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2824e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2825e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2826e159489bSTejun Heo * access. 2827e159489bSTejun Heo */ 2828e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2829baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2830e159489bSTejun Heo else 2831e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2832baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2833e159489bSTejun Heo 2834baf59022STejun Heo return true; 2835baf59022STejun Heo already_gone: 2836d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2837baf59022STejun Heo return false; 2838baf59022STejun Heo } 2839baf59022STejun Heo 2840db700897SOleg Nesterov /** 2841401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2842401a8d04STejun Heo * @work: the work to flush 2843db700897SOleg Nesterov * 2844606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2845606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2846401a8d04STejun Heo * 2847401a8d04STejun Heo * RETURNS: 2848401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2849401a8d04STejun Heo * %false if it was already idle. 2850db700897SOleg Nesterov */ 2851401a8d04STejun Heo bool flush_work(struct work_struct *work) 2852db700897SOleg Nesterov { 2853db700897SOleg Nesterov struct wq_barrier barr; 2854db700897SOleg Nesterov 28550976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28560976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28570976dfc1SStephen Boyd 2858606a5020STejun Heo if (start_flush_work(work, &barr)) { 2859db700897SOleg Nesterov wait_for_completion(&barr.done); 2860dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2861401a8d04STejun Heo return true; 2862606a5020STejun Heo } else { 2863401a8d04STejun Heo return false; 2864db700897SOleg Nesterov } 2865606a5020STejun Heo } 2866db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2867db700897SOleg Nesterov 286836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2869401a8d04STejun Heo { 2870bbb68dfaSTejun Heo unsigned long flags; 28711f1f642eSOleg Nesterov int ret; 28721f1f642eSOleg Nesterov 28731f1f642eSOleg Nesterov do { 2874bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2875bbb68dfaSTejun Heo /* 2876bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2877bbb68dfaSTejun Heo * would be waiting for before retrying. 2878bbb68dfaSTejun Heo */ 2879bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2880606a5020STejun Heo flush_work(work); 28811f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28821f1f642eSOleg Nesterov 2883bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2884bbb68dfaSTejun Heo mark_work_canceling(work); 2885bbb68dfaSTejun Heo local_irq_restore(flags); 2886bbb68dfaSTejun Heo 2887606a5020STejun Heo flush_work(work); 28887a22ad75STejun Heo clear_work_data(work); 28891f1f642eSOleg Nesterov return ret; 28901f1f642eSOleg Nesterov } 28911f1f642eSOleg Nesterov 28926e84d644SOleg Nesterov /** 2893401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2894401a8d04STejun Heo * @work: the work to cancel 28956e84d644SOleg Nesterov * 2896401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2897401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2898401a8d04STejun Heo * another workqueue. On return from this function, @work is 2899401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29001f1f642eSOleg Nesterov * 2901401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2902401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29036e84d644SOleg Nesterov * 2904401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29056e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2906401a8d04STejun Heo * 2907401a8d04STejun Heo * RETURNS: 2908401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29096e84d644SOleg Nesterov */ 2910401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29116e84d644SOleg Nesterov { 291236e227d2STejun Heo return __cancel_work_timer(work, false); 2913b89deed3SOleg Nesterov } 291428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2915b89deed3SOleg Nesterov 29166e84d644SOleg Nesterov /** 2917401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2918401a8d04STejun Heo * @dwork: the delayed work to flush 29196e84d644SOleg Nesterov * 2920401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2921401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2922401a8d04STejun Heo * considers the last queueing instance of @dwork. 29231f1f642eSOleg Nesterov * 2924401a8d04STejun Heo * RETURNS: 2925401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2926401a8d04STejun Heo * %false if it was already idle. 29276e84d644SOleg Nesterov */ 2928401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2929401a8d04STejun Heo { 29308930cabaSTejun Heo local_irq_disable(); 2931401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 293260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29338930cabaSTejun Heo local_irq_enable(); 2934401a8d04STejun Heo return flush_work(&dwork->work); 2935401a8d04STejun Heo } 2936401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2937401a8d04STejun Heo 2938401a8d04STejun Heo /** 293957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 294057b30ae7STejun Heo * @dwork: delayed_work to cancel 294109383498STejun Heo * 294257b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 294357b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 294457b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 294557b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 294657b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 294709383498STejun Heo * 294857b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 294909383498STejun Heo */ 295057b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 295109383498STejun Heo { 295257b30ae7STejun Heo unsigned long flags; 295357b30ae7STejun Heo int ret; 295457b30ae7STejun Heo 295557b30ae7STejun Heo do { 295657b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 295757b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 295857b30ae7STejun Heo 295957b30ae7STejun Heo if (unlikely(ret < 0)) 296057b30ae7STejun Heo return false; 296157b30ae7STejun Heo 29627c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29637c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 296457b30ae7STejun Heo local_irq_restore(flags); 2965c0158ca6SDan Magenheimer return ret; 296609383498STejun Heo } 296757b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 296809383498STejun Heo 296909383498STejun Heo /** 2970401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2971401a8d04STejun Heo * @dwork: the delayed work cancel 2972401a8d04STejun Heo * 2973401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2974401a8d04STejun Heo * 2975401a8d04STejun Heo * RETURNS: 2976401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2977401a8d04STejun Heo */ 2978401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29796e84d644SOleg Nesterov { 298036e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29816e84d644SOleg Nesterov } 2982f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29831da177e4SLinus Torvalds 29840fcb78c2SRolf Eike Beer /** 2985c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2986c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2987c1a220e7SZhang Rui * @work: job to be done 2988c1a220e7SZhang Rui * 2989c1a220e7SZhang Rui * This puts a job on a specific cpu 2990c1a220e7SZhang Rui */ 2991d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 2992c1a220e7SZhang Rui { 2993d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2994c1a220e7SZhang Rui } 2995c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2996c1a220e7SZhang Rui 29970fcb78c2SRolf Eike Beer /** 2998ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 29991da177e4SLinus Torvalds * @work: job to be done 30001da177e4SLinus Torvalds * 3001d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 3002d4283e93STejun Heo * %true otherwise. 300352bad64dSDavid Howells * 30040fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 30050fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 30060fcb78c2SRolf Eike Beer * workqueue otherwise. 30071da177e4SLinus Torvalds */ 3008d4283e93STejun Heo bool schedule_work(struct work_struct *work) 30091da177e4SLinus Torvalds { 30100fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 30111da177e4SLinus Torvalds } 30120fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 30131da177e4SLinus Torvalds 30140fcb78c2SRolf Eike Beer /** 30150fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 30160fcb78c2SRolf Eike Beer * @cpu: cpu to use 30170fcb78c2SRolf Eike Beer * @dwork: job to be done 30180fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 30190fcb78c2SRolf Eike Beer * 30200fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 30210fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 30220fcb78c2SRolf Eike Beer */ 3023d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 3024d4283e93STejun Heo unsigned long delay) 30251da177e4SLinus Torvalds { 3026d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 30271da177e4SLinus Torvalds } 3028ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 30291da177e4SLinus Torvalds 3030b6136773SAndrew Morton /** 30310a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 30320a13c00eSTejun Heo * @dwork: job to be done 30330a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 30340a13c00eSTejun Heo * 30350a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 30360a13c00eSTejun Heo * workqueue. 30370a13c00eSTejun Heo */ 3038d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 30390a13c00eSTejun Heo { 30400a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30410a13c00eSTejun Heo } 30420a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30430a13c00eSTejun Heo 30440a13c00eSTejun Heo /** 304531ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3046b6136773SAndrew Morton * @func: the function to call 3047b6136773SAndrew Morton * 304831ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 304931ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3050b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 305131ddd871STejun Heo * 305231ddd871STejun Heo * RETURNS: 305331ddd871STejun Heo * 0 on success, -errno on failure. 3054b6136773SAndrew Morton */ 305565f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 305615316ba8SChristoph Lameter { 305715316ba8SChristoph Lameter int cpu; 305838f51568SNamhyung Kim struct work_struct __percpu *works; 305915316ba8SChristoph Lameter 3060b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3061b6136773SAndrew Morton if (!works) 306215316ba8SChristoph Lameter return -ENOMEM; 3063b6136773SAndrew Morton 306495402b38SGautham R Shenoy get_online_cpus(); 306593981800STejun Heo 306615316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30679bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30689bfb1839SIngo Molnar 30699bfb1839SIngo Molnar INIT_WORK(work, func); 30708de6d308SOleg Nesterov schedule_work_on(cpu, work); 307115316ba8SChristoph Lameter } 307293981800STejun Heo 307393981800STejun Heo for_each_online_cpu(cpu) 30748616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 307593981800STejun Heo 307695402b38SGautham R Shenoy put_online_cpus(); 3077b6136773SAndrew Morton free_percpu(works); 307815316ba8SChristoph Lameter return 0; 307915316ba8SChristoph Lameter } 308015316ba8SChristoph Lameter 3081eef6a7d5SAlan Stern /** 3082eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3083eef6a7d5SAlan Stern * 3084eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3085eef6a7d5SAlan Stern * completion. 3086eef6a7d5SAlan Stern * 3087eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3088eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3089eef6a7d5SAlan Stern * will lead to deadlock: 3090eef6a7d5SAlan Stern * 3091eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3092eef6a7d5SAlan Stern * a lock held by your code or its caller. 3093eef6a7d5SAlan Stern * 3094eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3095eef6a7d5SAlan Stern * 3096eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3097eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3098eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3099eef6a7d5SAlan Stern * 3100eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3101eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3102eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3103eef6a7d5SAlan Stern * cancel_work_sync() instead. 3104eef6a7d5SAlan Stern */ 31051da177e4SLinus Torvalds void flush_scheduled_work(void) 31061da177e4SLinus Torvalds { 3107d320c038STejun Heo flush_workqueue(system_wq); 31081da177e4SLinus Torvalds } 3109ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 31101da177e4SLinus Torvalds 31111da177e4SLinus Torvalds /** 31121fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31131fa44ecaSJames Bottomley * @fn: the function to execute 31141fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31151fa44ecaSJames Bottomley * be available when the work executes) 31161fa44ecaSJames Bottomley * 31171fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31181fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31191fa44ecaSJames Bottomley * 31201fa44ecaSJames Bottomley * Returns: 0 - function was executed 31211fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31221fa44ecaSJames Bottomley */ 312365f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31241fa44ecaSJames Bottomley { 31251fa44ecaSJames Bottomley if (!in_interrupt()) { 312665f27f38SDavid Howells fn(&ew->work); 31271fa44ecaSJames Bottomley return 0; 31281fa44ecaSJames Bottomley } 31291fa44ecaSJames Bottomley 313065f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31311fa44ecaSJames Bottomley schedule_work(&ew->work); 31321fa44ecaSJames Bottomley 31331fa44ecaSJames Bottomley return 1; 31341fa44ecaSJames Bottomley } 31351fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31361fa44ecaSJames Bottomley 31371da177e4SLinus Torvalds int keventd_up(void) 31381da177e4SLinus Torvalds { 3139d320c038STejun Heo return system_wq != NULL; 31401da177e4SLinus Torvalds } 31411da177e4SLinus Torvalds 3142bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 31431da177e4SLinus Torvalds { 31443af24433SOleg Nesterov /* 31450f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 31460f900049STejun Heo * Make sure that the alignment isn't lower than that of 31470f900049STejun Heo * unsigned long long. 31483af24433SOleg Nesterov */ 31490f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 31500f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 31510f900049STejun Heo __alignof__(unsigned long long)); 31523af24433SOleg Nesterov 3153e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3154f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 3155931ac77eSTejun Heo else { 31560f900049STejun Heo void *ptr; 3157e1d8aa9fSFrederic Weisbecker 31580f900049STejun Heo /* 3159f3421797STejun Heo * Allocate enough room to align cwq and put an extra 3160f3421797STejun Heo * pointer at the end pointing back to the originally 3161f3421797STejun Heo * allocated pointer which will be used for free. 31620f900049STejun Heo */ 3163bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3164bdbc5dd7STejun Heo if (ptr) { 3165bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 3166bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 3167bdbc5dd7STejun Heo } 31683af24433SOleg Nesterov } 31693af24433SOleg Nesterov 31700415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3171bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 3172bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 31730f900049STejun Heo } 31740f900049STejun Heo 3175bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 317606ba38a9SOleg Nesterov { 3177e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3178bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 3179f3421797STejun Heo else if (wq->cpu_wq.single) { 3180f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 3181f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 318206ba38a9SOleg Nesterov } 318306ba38a9SOleg Nesterov } 318406ba38a9SOleg Nesterov 3185f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3186f3421797STejun Heo const char *name) 3187b71ab8c2STejun Heo { 3188f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3189f3421797STejun Heo 3190f3421797STejun Heo if (max_active < 1 || max_active > lim) 3191044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3192f3421797STejun Heo max_active, name, 1, lim); 3193b71ab8c2STejun Heo 3194f3421797STejun Heo return clamp_val(max_active, 1, lim); 3195b71ab8c2STejun Heo } 3196b71ab8c2STejun Heo 3197b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 319897e37d7bSTejun Heo unsigned int flags, 31991e19ffc6STejun Heo int max_active, 3200eb13ba87SJohannes Berg struct lock_class_key *key, 3201b196be89STejun Heo const char *lock_name, ...) 32023af24433SOleg Nesterov { 3203b196be89STejun Heo va_list args, args1; 32043af24433SOleg Nesterov struct workqueue_struct *wq; 3205c34056a3STejun Heo unsigned int cpu; 3206b196be89STejun Heo size_t namelen; 3207b196be89STejun Heo 3208b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3209b196be89STejun Heo va_start(args, lock_name); 3210b196be89STejun Heo va_copy(args1, args); 3211b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3212b196be89STejun Heo 3213b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3214b196be89STejun Heo if (!wq) 3215b196be89STejun Heo goto err; 3216b196be89STejun Heo 3217b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3218b196be89STejun Heo va_end(args); 3219b196be89STejun Heo va_end(args1); 32203af24433SOleg Nesterov 3221f3421797STejun Heo /* 32226370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 32236370a6adSTejun Heo * have a rescuer to guarantee forward progress. 32246370a6adSTejun Heo */ 32256370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 32266370a6adSTejun Heo flags |= WQ_RESCUER; 32276370a6adSTejun Heo 3228d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3229b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 32303af24433SOleg Nesterov 3231b196be89STejun Heo /* init wq */ 323297e37d7bSTejun Heo wq->flags = flags; 3233a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 323473f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 323573f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 323673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 323773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 32383af24433SOleg Nesterov 3239eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3240cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 32413af24433SOleg Nesterov 3242bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 3243bdbc5dd7STejun Heo goto err; 3244bdbc5dd7STejun Heo 3245f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 32461537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 32471537663fSTejun Heo 32480f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 3249a60dc39cSTejun Heo cwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI); 3250c34056a3STejun Heo cwq->wq = wq; 325173f53c4aSTejun Heo cwq->flush_color = -1; 32521e19ffc6STejun Heo cwq->max_active = max_active; 32531e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3254e22bee78STejun Heo } 32551537663fSTejun Heo 3256e22bee78STejun Heo if (flags & WQ_RESCUER) { 3257e22bee78STejun Heo struct worker *rescuer; 3258e22bee78STejun Heo 3259f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3260e22bee78STejun Heo goto err; 3261e22bee78STejun Heo 3262e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3263e22bee78STejun Heo if (!rescuer) 3264e22bee78STejun Heo goto err; 3265e22bee78STejun Heo 3266111c225aSTejun Heo rescuer->rescue_wq = wq; 3267111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3268b196be89STejun Heo wq->name); 3269e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3270e22bee78STejun Heo goto err; 3271e22bee78STejun Heo 3272e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3273e22bee78STejun Heo wake_up_process(rescuer->task); 32743af24433SOleg Nesterov } 32751537663fSTejun Heo 32763af24433SOleg Nesterov /* 3277a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3278a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3279a0a1a5fdSTejun Heo * workqueue to workqueues list. 32803af24433SOleg Nesterov */ 32813af24433SOleg Nesterov spin_lock(&workqueue_lock); 3282a0a1a5fdSTejun Heo 328358a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3284f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3285a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3286a0a1a5fdSTejun Heo 32873af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3288a0a1a5fdSTejun Heo 32893af24433SOleg Nesterov spin_unlock(&workqueue_lock); 32903af24433SOleg Nesterov 32913af24433SOleg Nesterov return wq; 32924690c4abSTejun Heo err: 32934690c4abSTejun Heo if (wq) { 3294bdbc5dd7STejun Heo free_cwqs(wq); 3295f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3296e22bee78STejun Heo kfree(wq->rescuer); 32974690c4abSTejun Heo kfree(wq); 32983af24433SOleg Nesterov } 32994690c4abSTejun Heo return NULL; 33001da177e4SLinus Torvalds } 3301d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 33021da177e4SLinus Torvalds 33033af24433SOleg Nesterov /** 33043af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 33053af24433SOleg Nesterov * @wq: target workqueue 33063af24433SOleg Nesterov * 33073af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 33083af24433SOleg Nesterov */ 33093af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 33103af24433SOleg Nesterov { 3311c8e55f36STejun Heo unsigned int cpu; 33123af24433SOleg Nesterov 33139c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 33149c5a2ba7STejun Heo drain_workqueue(wq); 3315c8efcc25STejun Heo 3316a0a1a5fdSTejun Heo /* 3317a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3318a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3319a0a1a5fdSTejun Heo */ 332095402b38SGautham R Shenoy spin_lock(&workqueue_lock); 33213af24433SOleg Nesterov list_del(&wq->list); 332295402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 33233af24433SOleg Nesterov 3324e22bee78STejun Heo /* sanity check */ 3325f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 332673f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 332773f53c4aSTejun Heo int i; 33283af24433SOleg Nesterov 332973f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 333073f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 33311e19ffc6STejun Heo BUG_ON(cwq->nr_active); 33321e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 333373f53c4aSTejun Heo } 33341537663fSTejun Heo 3335e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3336e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3337f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 33388d9df9f0SXiaotian Feng kfree(wq->rescuer); 3339e22bee78STejun Heo } 3340e22bee78STejun Heo 3341bdbc5dd7STejun Heo free_cwqs(wq); 33423af24433SOleg Nesterov kfree(wq); 33433af24433SOleg Nesterov } 33443af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33453af24433SOleg Nesterov 3346dcd989cbSTejun Heo /** 33479f4bd4cdSLai Jiangshan * cwq_set_max_active - adjust max_active of a cwq 33489f4bd4cdSLai Jiangshan * @cwq: target cpu_workqueue_struct 33499f4bd4cdSLai Jiangshan * @max_active: new max_active value. 33509f4bd4cdSLai Jiangshan * 33519f4bd4cdSLai Jiangshan * Set @cwq->max_active to @max_active and activate delayed works if 33529f4bd4cdSLai Jiangshan * increased. 33539f4bd4cdSLai Jiangshan * 33549f4bd4cdSLai Jiangshan * CONTEXT: 3355d565ed63STejun Heo * spin_lock_irq(pool->lock). 33569f4bd4cdSLai Jiangshan */ 33579f4bd4cdSLai Jiangshan static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active) 33589f4bd4cdSLai Jiangshan { 33599f4bd4cdSLai Jiangshan cwq->max_active = max_active; 33609f4bd4cdSLai Jiangshan 33619f4bd4cdSLai Jiangshan while (!list_empty(&cwq->delayed_works) && 33629f4bd4cdSLai Jiangshan cwq->nr_active < cwq->max_active) 33639f4bd4cdSLai Jiangshan cwq_activate_first_delayed(cwq); 33649f4bd4cdSLai Jiangshan } 33659f4bd4cdSLai Jiangshan 33669f4bd4cdSLai Jiangshan /** 3367dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3368dcd989cbSTejun Heo * @wq: target workqueue 3369dcd989cbSTejun Heo * @max_active: new max_active value. 3370dcd989cbSTejun Heo * 3371dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3372dcd989cbSTejun Heo * 3373dcd989cbSTejun Heo * CONTEXT: 3374dcd989cbSTejun Heo * Don't call from IRQ context. 3375dcd989cbSTejun Heo */ 3376dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3377dcd989cbSTejun Heo { 3378dcd989cbSTejun Heo unsigned int cpu; 3379dcd989cbSTejun Heo 3380f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3381dcd989cbSTejun Heo 3382dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3383dcd989cbSTejun Heo 3384dcd989cbSTejun Heo wq->saved_max_active = max_active; 3385dcd989cbSTejun Heo 3386f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 338735b6bb63STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 338835b6bb63STejun Heo struct worker_pool *pool = cwq->pool; 3389dcd989cbSTejun Heo 3390d565ed63STejun Heo spin_lock_irq(&pool->lock); 3391dcd989cbSTejun Heo 339258a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 339335b6bb63STejun Heo !(pool->flags & POOL_FREEZING)) 339435b6bb63STejun Heo cwq_set_max_active(cwq, max_active); 3395dcd989cbSTejun Heo 3396d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3397dcd989cbSTejun Heo } 3398dcd989cbSTejun Heo 3399dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3400dcd989cbSTejun Heo } 3401dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3402dcd989cbSTejun Heo 3403dcd989cbSTejun Heo /** 3404dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3405dcd989cbSTejun Heo * @cpu: CPU in question 3406dcd989cbSTejun Heo * @wq: target workqueue 3407dcd989cbSTejun Heo * 3408dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3409dcd989cbSTejun Heo * no synchronization around this function and the test result is 3410dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3411dcd989cbSTejun Heo * 3412dcd989cbSTejun Heo * RETURNS: 3413dcd989cbSTejun Heo * %true if congested, %false otherwise. 3414dcd989cbSTejun Heo */ 3415dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3416dcd989cbSTejun Heo { 3417dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3418dcd989cbSTejun Heo 3419dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3420dcd989cbSTejun Heo } 3421dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3422dcd989cbSTejun Heo 3423dcd989cbSTejun Heo /** 3424dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3425dcd989cbSTejun Heo * @work: the work to be tested 3426dcd989cbSTejun Heo * 3427dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3428dcd989cbSTejun Heo * synchronization around this function and the test result is 3429dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3430dcd989cbSTejun Heo * 3431dcd989cbSTejun Heo * RETURNS: 3432dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3433dcd989cbSTejun Heo */ 3434dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3435dcd989cbSTejun Heo { 3436c9e7cf27STejun Heo struct worker_pool *pool = get_work_pool(work); 3437dcd989cbSTejun Heo unsigned long flags; 3438dcd989cbSTejun Heo unsigned int ret = 0; 3439dcd989cbSTejun Heo 3440dcd989cbSTejun Heo if (work_pending(work)) 3441dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3442038366c5SLai Jiangshan 3443038366c5SLai Jiangshan if (pool) { 3444038366c5SLai Jiangshan spin_lock_irqsave(&pool->lock, flags); 3445c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3446dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3447d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 3448038366c5SLai Jiangshan } 3449dcd989cbSTejun Heo 3450dcd989cbSTejun Heo return ret; 3451dcd989cbSTejun Heo } 3452dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3453dcd989cbSTejun Heo 3454db7bccf4STejun Heo /* 3455db7bccf4STejun Heo * CPU hotplug. 3456db7bccf4STejun Heo * 3457e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3458e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3459706026c2STejun Heo * pool which make migrating pending and scheduled works very 3460e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 346194cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 3462e22bee78STejun Heo * blocked draining impractical. 3463e22bee78STejun Heo * 346424647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 3465628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3466628c78e7STejun Heo * cpu comes back online. 3467db7bccf4STejun Heo */ 3468db7bccf4STejun Heo 3469706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 3470db7bccf4STejun Heo { 347138db41d9STejun Heo int cpu = smp_processor_id(); 34724ce62e9eSTejun Heo struct worker_pool *pool; 3473db7bccf4STejun Heo struct worker *worker; 3474db7bccf4STejun Heo struct hlist_node *pos; 3475db7bccf4STejun Heo int i; 3476db7bccf4STejun Heo 347738db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 347838db41d9STejun Heo BUG_ON(cpu != smp_processor_id()); 3479db7bccf4STejun Heo 348094cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 348194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 3482e22bee78STejun Heo 3483f2d5a0eeSTejun Heo /* 348494cf58bbSTejun Heo * We've claimed all manager positions. Make all workers 348594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 348694cf58bbSTejun Heo * except for the ones which are still executing works from 348794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 348894cf58bbSTejun Heo * this, they may become diasporas. 3489f2d5a0eeSTejun Heo */ 34904ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3491403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3492db7bccf4STejun Heo 3493c9e7cf27STejun Heo for_each_busy_worker(worker, i, pos, pool) 3494403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3495db7bccf4STejun Heo 349624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 3497f2d5a0eeSTejun Heo 349894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 349994cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 350094cf58bbSTejun Heo } 3501e22bee78STejun Heo 3502e22bee78STejun Heo /* 3503628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3504628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3505628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3506628c78e7STejun Heo */ 3507628c78e7STejun Heo schedule(); 3508628c78e7STejun Heo 3509628c78e7STejun Heo /* 3510628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3511628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 351238db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 351338db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 351438db41d9STejun Heo * pools which are served by workers tied to the CPU. 3515628c78e7STejun Heo * 3516628c78e7STejun Heo * On return from this function, the current worker would trigger 3517628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3518628c78e7STejun Heo * didn't already. 3519e22bee78STejun Heo */ 352038db41d9STejun Heo for_each_std_worker_pool(pool, cpu) 35214ce62e9eSTejun Heo atomic_set(get_pool_nr_running(pool), 0); 3522db7bccf4STejun Heo } 3523db7bccf4STejun Heo 35248db25e78STejun Heo /* 35258db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 35268db25e78STejun Heo * This will be registered high priority CPU notifier. 35278db25e78STejun Heo */ 35289fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 35291da177e4SLinus Torvalds unsigned long action, 35301da177e4SLinus Torvalds void *hcpu) 35311da177e4SLinus Torvalds { 35323af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 35334ce62e9eSTejun Heo struct worker_pool *pool; 35341da177e4SLinus Torvalds 35358db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 35363af24433SOleg Nesterov case CPU_UP_PREPARE: 353738db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 35383ce63377STejun Heo struct worker *worker; 35393ce63377STejun Heo 35403ce63377STejun Heo if (pool->nr_workers) 35413ce63377STejun Heo continue; 35423ce63377STejun Heo 35433ce63377STejun Heo worker = create_worker(pool); 35443ce63377STejun Heo if (!worker) 35453ce63377STejun Heo return NOTIFY_BAD; 35463ce63377STejun Heo 3547d565ed63STejun Heo spin_lock_irq(&pool->lock); 35483ce63377STejun Heo start_worker(worker); 3549d565ed63STejun Heo spin_unlock_irq(&pool->lock); 35503af24433SOleg Nesterov } 35511da177e4SLinus Torvalds break; 35521da177e4SLinus Torvalds 355365758202STejun Heo case CPU_DOWN_FAILED: 355465758202STejun Heo case CPU_ONLINE: 355538db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 355694cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 355794cf58bbSTejun Heo spin_lock_irq(&pool->lock); 355894cf58bbSTejun Heo 355924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 356094cf58bbSTejun Heo rebind_workers(pool); 356194cf58bbSTejun Heo 356294cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 356394cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 356494cf58bbSTejun Heo } 35658db25e78STejun Heo break; 356665758202STejun Heo } 356765758202STejun Heo return NOTIFY_OK; 356865758202STejun Heo } 356965758202STejun Heo 357065758202STejun Heo /* 357165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 357265758202STejun Heo * This will be registered as low priority CPU notifier. 357365758202STejun Heo */ 35749fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 357565758202STejun Heo unsigned long action, 357665758202STejun Heo void *hcpu) 357765758202STejun Heo { 35788db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 35798db25e78STejun Heo struct work_struct unbind_work; 35808db25e78STejun Heo 358165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 358265758202STejun Heo case CPU_DOWN_PREPARE: 35838db25e78STejun Heo /* unbinding should happen on the local CPU */ 3584706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 35857635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 35868db25e78STejun Heo flush_work(&unbind_work); 35878db25e78STejun Heo break; 358865758202STejun Heo } 358965758202STejun Heo return NOTIFY_OK; 359065758202STejun Heo } 359165758202STejun Heo 35922d3854a3SRusty Russell #ifdef CONFIG_SMP 35938ccad40dSRusty Russell 35942d3854a3SRusty Russell struct work_for_cpu { 3595ed48ece2STejun Heo struct work_struct work; 35962d3854a3SRusty Russell long (*fn)(void *); 35972d3854a3SRusty Russell void *arg; 35982d3854a3SRusty Russell long ret; 35992d3854a3SRusty Russell }; 36002d3854a3SRusty Russell 3601ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 36022d3854a3SRusty Russell { 3603ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3604ed48ece2STejun Heo 36052d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 36062d3854a3SRusty Russell } 36072d3854a3SRusty Russell 36082d3854a3SRusty Russell /** 36092d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 36102d3854a3SRusty Russell * @cpu: the cpu to run on 36112d3854a3SRusty Russell * @fn: the function to run 36122d3854a3SRusty Russell * @arg: the function arg 36132d3854a3SRusty Russell * 361431ad9081SRusty Russell * This will return the value @fn returns. 361531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 36166b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 36172d3854a3SRusty Russell */ 36182d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 36192d3854a3SRusty Russell { 3620ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 36212d3854a3SRusty Russell 3622ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3623ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3624ed48ece2STejun Heo flush_work(&wfc.work); 36252d3854a3SRusty Russell return wfc.ret; 36262d3854a3SRusty Russell } 36272d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 36282d3854a3SRusty Russell #endif /* CONFIG_SMP */ 36292d3854a3SRusty Russell 3630a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3631e7577c50SRusty Russell 3632a0a1a5fdSTejun Heo /** 3633a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3634a0a1a5fdSTejun Heo * 363558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 363658a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 3637706026c2STejun Heo * pool->worklist. 3638a0a1a5fdSTejun Heo * 3639a0a1a5fdSTejun Heo * CONTEXT: 3640d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3641a0a1a5fdSTejun Heo */ 3642a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3643a0a1a5fdSTejun Heo { 3644a0a1a5fdSTejun Heo unsigned int cpu; 3645a0a1a5fdSTejun Heo 3646a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3647a0a1a5fdSTejun Heo 3648a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3649a0a1a5fdSTejun Heo workqueue_freezing = true; 3650a0a1a5fdSTejun Heo 3651706026c2STejun Heo for_each_wq_cpu(cpu) { 365235b6bb63STejun Heo struct worker_pool *pool; 3653bdbc5dd7STejun Heo struct workqueue_struct *wq; 36548b03ae3cSTejun Heo 365538db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3656a1056305STejun Heo spin_lock_irq(&pool->lock); 3657d565ed63STejun Heo 365835b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 365935b6bb63STejun Heo pool->flags |= POOL_FREEZING; 3660db7bccf4STejun Heo 3661a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3662a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3663a0a1a5fdSTejun Heo 3664a1056305STejun Heo if (cwq && cwq->pool == pool && 3665a1056305STejun Heo (wq->flags & WQ_FREEZABLE)) 3666a0a1a5fdSTejun Heo cwq->max_active = 0; 36671da177e4SLinus Torvalds } 36688b03ae3cSTejun Heo 3669a1056305STejun Heo spin_unlock_irq(&pool->lock); 3670a1056305STejun Heo } 3671a0a1a5fdSTejun Heo } 3672a0a1a5fdSTejun Heo 3673a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3674a0a1a5fdSTejun Heo } 3675a0a1a5fdSTejun Heo 3676a0a1a5fdSTejun Heo /** 367758a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3678a0a1a5fdSTejun Heo * 3679a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3680a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3681a0a1a5fdSTejun Heo * 3682a0a1a5fdSTejun Heo * CONTEXT: 3683a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3684a0a1a5fdSTejun Heo * 3685a0a1a5fdSTejun Heo * RETURNS: 368658a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 368758a69cb4STejun Heo * is complete. 3688a0a1a5fdSTejun Heo */ 3689a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3690a0a1a5fdSTejun Heo { 3691a0a1a5fdSTejun Heo unsigned int cpu; 3692a0a1a5fdSTejun Heo bool busy = false; 3693a0a1a5fdSTejun Heo 3694a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3695a0a1a5fdSTejun Heo 3696a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3697a0a1a5fdSTejun Heo 3698706026c2STejun Heo for_each_wq_cpu(cpu) { 3699bdbc5dd7STejun Heo struct workqueue_struct *wq; 3700a0a1a5fdSTejun Heo /* 3701a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3702a0a1a5fdSTejun Heo * to peek without lock. 3703a0a1a5fdSTejun Heo */ 3704a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3705a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3706a0a1a5fdSTejun Heo 370758a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3708a0a1a5fdSTejun Heo continue; 3709a0a1a5fdSTejun Heo 3710a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3711a0a1a5fdSTejun Heo if (cwq->nr_active) { 3712a0a1a5fdSTejun Heo busy = true; 3713a0a1a5fdSTejun Heo goto out_unlock; 3714a0a1a5fdSTejun Heo } 3715a0a1a5fdSTejun Heo } 3716a0a1a5fdSTejun Heo } 3717a0a1a5fdSTejun Heo out_unlock: 3718a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3719a0a1a5fdSTejun Heo return busy; 3720a0a1a5fdSTejun Heo } 3721a0a1a5fdSTejun Heo 3722a0a1a5fdSTejun Heo /** 3723a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3724a0a1a5fdSTejun Heo * 3725a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 3726706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 3727a0a1a5fdSTejun Heo * 3728a0a1a5fdSTejun Heo * CONTEXT: 3729d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3730a0a1a5fdSTejun Heo */ 3731a0a1a5fdSTejun Heo void thaw_workqueues(void) 3732a0a1a5fdSTejun Heo { 3733a0a1a5fdSTejun Heo unsigned int cpu; 3734a0a1a5fdSTejun Heo 3735a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3736a0a1a5fdSTejun Heo 3737a0a1a5fdSTejun Heo if (!workqueue_freezing) 3738a0a1a5fdSTejun Heo goto out_unlock; 3739a0a1a5fdSTejun Heo 3740706026c2STejun Heo for_each_wq_cpu(cpu) { 37414ce62e9eSTejun Heo struct worker_pool *pool; 3742bdbc5dd7STejun Heo struct workqueue_struct *wq; 37438b03ae3cSTejun Heo 374438db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3745a1056305STejun Heo spin_lock_irq(&pool->lock); 3746d565ed63STejun Heo 374735b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 374835b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 3749db7bccf4STejun Heo 3750a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3751a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3752a0a1a5fdSTejun Heo 3753a1056305STejun Heo if (!cwq || cwq->pool != pool || 3754a1056305STejun Heo !(wq->flags & WQ_FREEZABLE)) 3755a0a1a5fdSTejun Heo continue; 3756a0a1a5fdSTejun Heo 3757a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 37589f4bd4cdSLai Jiangshan cwq_set_max_active(cwq, wq->saved_max_active); 3759a0a1a5fdSTejun Heo } 37608b03ae3cSTejun Heo 37614ce62e9eSTejun Heo wake_up_worker(pool); 3762a1056305STejun Heo 3763a1056305STejun Heo spin_unlock_irq(&pool->lock); 3764d565ed63STejun Heo } 3765a0a1a5fdSTejun Heo } 3766a0a1a5fdSTejun Heo 3767a0a1a5fdSTejun Heo workqueue_freezing = false; 3768a0a1a5fdSTejun Heo out_unlock: 3769a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3770a0a1a5fdSTejun Heo } 3771a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3772a0a1a5fdSTejun Heo 37736ee0578bSSuresh Siddha static int __init init_workqueues(void) 37741da177e4SLinus Torvalds { 3775c34056a3STejun Heo unsigned int cpu; 3776c34056a3STejun Heo 37777c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 37787c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 37796be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 3780b5490077STejun Heo 378165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3782a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 37838b03ae3cSTejun Heo 3784706026c2STejun Heo /* initialize CPU pools */ 3785706026c2STejun Heo for_each_wq_cpu(cpu) { 37864ce62e9eSTejun Heo struct worker_pool *pool; 37878b03ae3cSTejun Heo 378838db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3789d565ed63STejun Heo spin_lock_init(&pool->lock); 3790ec22ca5eSTejun Heo pool->cpu = cpu; 379124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 37924ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37934ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3794c9e7cf27STejun Heo hash_init(pool->busy_hash); 3795e22bee78STejun Heo 37964ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 37974ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 37984ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3799e22bee78STejun Heo 3800706026c2STejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 38014ce62e9eSTejun Heo (unsigned long)pool); 38024ce62e9eSTejun Heo 3803b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 38044ce62e9eSTejun Heo ida_init(&pool->worker_ida); 38059daf9e67STejun Heo 38069daf9e67STejun Heo /* alloc pool ID */ 38079daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 38084ce62e9eSTejun Heo } 38098b03ae3cSTejun Heo } 38108b03ae3cSTejun Heo 3811e22bee78STejun Heo /* create the initial worker */ 3812706026c2STejun Heo for_each_online_wq_cpu(cpu) { 38134ce62e9eSTejun Heo struct worker_pool *pool; 3814e22bee78STejun Heo 381538db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 38164ce62e9eSTejun Heo struct worker *worker; 38174ce62e9eSTejun Heo 381824647570STejun Heo if (cpu != WORK_CPU_UNBOUND) 381924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 382024647570STejun Heo 3821bc2ae0f5STejun Heo worker = create_worker(pool); 3822e22bee78STejun Heo BUG_ON(!worker); 3823d565ed63STejun Heo spin_lock_irq(&pool->lock); 3824e22bee78STejun Heo start_worker(worker); 3825d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3826e22bee78STejun Heo } 38274ce62e9eSTejun Heo } 3828e22bee78STejun Heo 3829d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 38301aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3831d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3832f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3833f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 383424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 383524d51addSTejun Heo WQ_FREEZABLE, 0); 38361aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3837ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 38386ee0578bSSuresh Siddha return 0; 38391da177e4SLinus Torvalds } 38406ee0578bSSuresh Siddha early_initcall(init_workqueues); 3841