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 5597c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5607c3eed5cSTejun Heo int pool_id) 5614d707b9fSOleg Nesterov { 56223657bb1STejun Heo /* 56323657bb1STejun Heo * The following wmb is paired with the implied mb in 56423657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 56523657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 56623657bb1STejun Heo * owner. 56723657bb1STejun Heo */ 56823657bb1STejun Heo smp_wmb(); 5697c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5704d707b9fSOleg Nesterov } 5714d707b9fSOleg Nesterov 5727a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 573365970a1SDavid Howells { 5747c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5757c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5767a22ad75STejun Heo } 5777a22ad75STejun Heo 5787a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5797a22ad75STejun Heo { 580e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5817a22ad75STejun Heo 582e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 583e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 584e120153dSTejun Heo else 585e120153dSTejun Heo return NULL; 5867a22ad75STejun Heo } 5877a22ad75STejun Heo 5887c3eed5cSTejun Heo /** 5897c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5907c3eed5cSTejun Heo * @work: the work item of interest 5917c3eed5cSTejun Heo * 5927c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 5937c3eed5cSTejun Heo */ 5947c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 5957a22ad75STejun Heo { 596e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5977c3eed5cSTejun Heo struct worker_pool *pool; 5987c3eed5cSTejun Heo int pool_id; 5997a22ad75STejun Heo 600e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 601e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 6027c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6037a22ad75STejun Heo 6047c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6057c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6067a22ad75STejun Heo return NULL; 6077a22ad75STejun Heo 6087c3eed5cSTejun Heo pool = worker_pool_by_id(pool_id); 6097c3eed5cSTejun Heo WARN_ON_ONCE(!pool); 6107c3eed5cSTejun Heo return pool; 6117c3eed5cSTejun Heo } 6127c3eed5cSTejun Heo 6137c3eed5cSTejun Heo /** 6147c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6157c3eed5cSTejun Heo * @work: the work item of interest 6167c3eed5cSTejun Heo * 6177c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6187c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6197c3eed5cSTejun Heo */ 6207c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6217c3eed5cSTejun Heo { 6227c3eed5cSTejun Heo struct worker_pool *pool = get_work_pool(work); 6237c3eed5cSTejun Heo 6247c3eed5cSTejun Heo return pool ? pool->id : WORK_OFFQ_POOL_NONE; 6257c3eed5cSTejun Heo } 6267c3eed5cSTejun Heo 627bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 628bbb68dfaSTejun Heo { 6297c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 630bbb68dfaSTejun Heo 6317c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6327c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 633bbb68dfaSTejun Heo } 634bbb68dfaSTejun Heo 635bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 636bbb68dfaSTejun Heo { 637bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 638bbb68dfaSTejun Heo 639bbb68dfaSTejun Heo return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING); 640bbb68dfaSTejun Heo } 641bbb68dfaSTejun Heo 642e22bee78STejun Heo /* 6433270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6443270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 645d565ed63STejun Heo * they're being called with pool->lock held. 646e22bee78STejun Heo */ 647e22bee78STejun Heo 64863d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 649649027d7STejun Heo { 6503270476aSTejun Heo return !atomic_read(get_pool_nr_running(pool)); 651649027d7STejun Heo } 652649027d7STejun Heo 653e22bee78STejun Heo /* 654e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 655e22bee78STejun Heo * running workers. 656974271c4STejun Heo * 657974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 658706026c2STejun Heo * function will always return %true for unbound pools as long as the 659974271c4STejun Heo * worklist isn't empty. 660e22bee78STejun Heo */ 66163d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 662e22bee78STejun Heo { 66363d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 664e22bee78STejun Heo } 665e22bee78STejun Heo 666e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 66763d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 668e22bee78STejun Heo { 66963d95a91STejun Heo return pool->nr_idle; 670e22bee78STejun Heo } 671e22bee78STejun Heo 672e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 67363d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 674e22bee78STejun Heo { 67563d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 676e22bee78STejun Heo 6773270476aSTejun Heo return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1; 678e22bee78STejun Heo } 679e22bee78STejun Heo 680e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 68163d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 682e22bee78STejun Heo { 68363d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 684e22bee78STejun Heo } 685e22bee78STejun Heo 686e22bee78STejun Heo /* Do I need to be the manager? */ 68763d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 688e22bee78STejun Heo { 68963d95a91STejun Heo return need_to_create_worker(pool) || 69011ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 691e22bee78STejun Heo } 692e22bee78STejun Heo 693e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 69463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 695e22bee78STejun Heo { 696552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 69763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 69863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 699e22bee78STejun Heo 700ea1abd61SLai Jiangshan /* 701ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 702ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 703ea1abd61SLai Jiangshan */ 704ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 705ea1abd61SLai Jiangshan return false; 706ea1abd61SLai Jiangshan 707e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 708e22bee78STejun Heo } 709e22bee78STejun Heo 710e22bee78STejun Heo /* 711e22bee78STejun Heo * Wake up functions. 712e22bee78STejun Heo */ 713e22bee78STejun Heo 7147e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 71563d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7167e11629dSTejun Heo { 71763d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7187e11629dSTejun Heo return NULL; 7197e11629dSTejun Heo 72063d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7217e11629dSTejun Heo } 7227e11629dSTejun Heo 7237e11629dSTejun Heo /** 7247e11629dSTejun Heo * wake_up_worker - wake up an idle worker 72563d95a91STejun Heo * @pool: worker pool to wake worker from 7267e11629dSTejun Heo * 72763d95a91STejun Heo * Wake up the first idle worker of @pool. 7287e11629dSTejun Heo * 7297e11629dSTejun Heo * CONTEXT: 730d565ed63STejun Heo * spin_lock_irq(pool->lock). 7317e11629dSTejun Heo */ 73263d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7337e11629dSTejun Heo { 73463d95a91STejun Heo struct worker *worker = first_worker(pool); 7357e11629dSTejun Heo 7367e11629dSTejun Heo if (likely(worker)) 7377e11629dSTejun Heo wake_up_process(worker->task); 7387e11629dSTejun Heo } 7397e11629dSTejun Heo 7404690c4abSTejun Heo /** 741e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 742e22bee78STejun Heo * @task: task waking up 743e22bee78STejun Heo * @cpu: CPU @task is waking up to 744e22bee78STejun Heo * 745e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 746e22bee78STejun Heo * being awoken. 747e22bee78STejun Heo * 748e22bee78STejun Heo * CONTEXT: 749e22bee78STejun Heo * spin_lock_irq(rq->lock) 750e22bee78STejun Heo */ 751e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 752e22bee78STejun Heo { 753e22bee78STejun Heo struct worker *worker = kthread_data(task); 754e22bee78STejun Heo 75536576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 756ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 75763d95a91STejun Heo atomic_inc(get_pool_nr_running(worker->pool)); 758e22bee78STejun Heo } 75936576000SJoonsoo Kim } 760e22bee78STejun Heo 761e22bee78STejun Heo /** 762e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 763e22bee78STejun Heo * @task: task going to sleep 764e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 765e22bee78STejun Heo * 766e22bee78STejun Heo * This function is called during schedule() when a busy worker is 767e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 768e22bee78STejun Heo * returning pointer to its task. 769e22bee78STejun Heo * 770e22bee78STejun Heo * CONTEXT: 771e22bee78STejun Heo * spin_lock_irq(rq->lock) 772e22bee78STejun Heo * 773e22bee78STejun Heo * RETURNS: 774e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 775e22bee78STejun Heo */ 776e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 777e22bee78STejun Heo unsigned int cpu) 778e22bee78STejun Heo { 779e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 780111c225aSTejun Heo struct worker_pool *pool; 781111c225aSTejun Heo atomic_t *nr_running; 782e22bee78STejun Heo 783111c225aSTejun Heo /* 784111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 785111c225aSTejun Heo * workers, also reach here, let's not access anything before 786111c225aSTejun Heo * checking NOT_RUNNING. 787111c225aSTejun Heo */ 7882d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 789e22bee78STejun Heo return NULL; 790e22bee78STejun Heo 791111c225aSTejun Heo pool = worker->pool; 792111c225aSTejun Heo nr_running = get_pool_nr_running(pool); 793111c225aSTejun Heo 794e22bee78STejun Heo /* this can only happen on the local cpu */ 795e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 796e22bee78STejun Heo 797e22bee78STejun Heo /* 798e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 799e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 800e22bee78STejun Heo * Please read comment there. 801e22bee78STejun Heo * 802628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 803628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 804628c78e7STejun Heo * disabled, which in turn means that none else could be 805d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 806628c78e7STejun Heo * lock is safe. 807e22bee78STejun Heo */ 808bd7bdd43STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist)) 80963d95a91STejun Heo to_wakeup = first_worker(pool); 810e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 811e22bee78STejun Heo } 812e22bee78STejun Heo 813e22bee78STejun Heo /** 814e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 815cb444766STejun Heo * @worker: self 816d302f017STejun Heo * @flags: flags to set 817d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 818d302f017STejun Heo * 819e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 820e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 821e22bee78STejun Heo * woken up. 822d302f017STejun Heo * 823cb444766STejun Heo * CONTEXT: 824d565ed63STejun Heo * spin_lock_irq(pool->lock) 825d302f017STejun Heo */ 826d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 827d302f017STejun Heo bool wakeup) 828d302f017STejun Heo { 829bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 830e22bee78STejun Heo 831cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 832cb444766STejun Heo 833e22bee78STejun Heo /* 834e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 835e22bee78STejun Heo * wake up an idle worker as necessary if requested by 836e22bee78STejun Heo * @wakeup. 837e22bee78STejun Heo */ 838e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 839e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 84063d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 841e22bee78STejun Heo 842e22bee78STejun Heo if (wakeup) { 843e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 844bd7bdd43STejun Heo !list_empty(&pool->worklist)) 84563d95a91STejun Heo wake_up_worker(pool); 846e22bee78STejun Heo } else 847e22bee78STejun Heo atomic_dec(nr_running); 848e22bee78STejun Heo } 849e22bee78STejun Heo 850d302f017STejun Heo worker->flags |= flags; 851d302f017STejun Heo } 852d302f017STejun Heo 853d302f017STejun Heo /** 854e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 855cb444766STejun Heo * @worker: self 856d302f017STejun Heo * @flags: flags to clear 857d302f017STejun Heo * 858e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 859d302f017STejun Heo * 860cb444766STejun Heo * CONTEXT: 861d565ed63STejun Heo * spin_lock_irq(pool->lock) 862d302f017STejun Heo */ 863d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 864d302f017STejun Heo { 86563d95a91STejun Heo struct worker_pool *pool = worker->pool; 866e22bee78STejun Heo unsigned int oflags = worker->flags; 867e22bee78STejun Heo 868cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 869cb444766STejun Heo 870d302f017STejun Heo worker->flags &= ~flags; 871e22bee78STejun Heo 87242c025f3STejun Heo /* 87342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 87442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 87542c025f3STejun Heo * of multiple flags, not a single flag. 87642c025f3STejun Heo */ 877e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 878e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 87963d95a91STejun Heo atomic_inc(get_pool_nr_running(pool)); 880d302f017STejun Heo } 881d302f017STejun Heo 882d302f017STejun Heo /** 8838cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 884c9e7cf27STejun Heo * @pool: pool of interest 8858cca0eeaSTejun Heo * @work: work to find worker for 8868cca0eeaSTejun Heo * 887c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 888c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 889a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 890a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 891a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 892a2c1c57bSTejun Heo * being executed. 893a2c1c57bSTejun Heo * 894a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 895a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 896a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 897a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 898a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 899a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 900a2c1c57bSTejun Heo * 901a2c1c57bSTejun Heo * This function checks the work item address, work function and workqueue 902a2c1c57bSTejun Heo * to avoid false positives. Note that this isn't complete as one may 903a2c1c57bSTejun Heo * construct a work function which can introduce dependency onto itself 904a2c1c57bSTejun Heo * through a recycled work item. Well, if somebody wants to shoot oneself 905a2c1c57bSTejun Heo * in the foot that badly, there's only so much we can do, and if such 906a2c1c57bSTejun Heo * deadlock actually occurs, it should be easy to locate the culprit work 907a2c1c57bSTejun Heo * function. 9088cca0eeaSTejun Heo * 9098cca0eeaSTejun Heo * CONTEXT: 910d565ed63STejun Heo * spin_lock_irq(pool->lock). 9118cca0eeaSTejun Heo * 9128cca0eeaSTejun Heo * RETURNS: 9138cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9148cca0eeaSTejun Heo * otherwise. 9158cca0eeaSTejun Heo */ 916c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9178cca0eeaSTejun Heo struct work_struct *work) 9188cca0eeaSTejun Heo { 91942f8570fSSasha Levin struct worker *worker; 92042f8570fSSasha Levin struct hlist_node *tmp; 92142f8570fSSasha Levin 922c9e7cf27STejun Heo hash_for_each_possible(pool->busy_hash, worker, tmp, hentry, 923a2c1c57bSTejun Heo (unsigned long)work) 924a2c1c57bSTejun Heo if (worker->current_work == work && 925a2c1c57bSTejun Heo worker->current_func == work->func) 92642f8570fSSasha Levin return worker; 92742f8570fSSasha Levin 92842f8570fSSasha Levin return NULL; 9298cca0eeaSTejun Heo } 9308cca0eeaSTejun Heo 9318cca0eeaSTejun Heo /** 932bf4ede01STejun Heo * move_linked_works - move linked works to a list 933bf4ede01STejun Heo * @work: start of series of works to be scheduled 934bf4ede01STejun Heo * @head: target list to append @work to 935bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 936bf4ede01STejun Heo * 937bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 938bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 939bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 940bf4ede01STejun Heo * 941bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 942bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 943bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 944bf4ede01STejun Heo * 945bf4ede01STejun Heo * CONTEXT: 946d565ed63STejun Heo * spin_lock_irq(pool->lock). 947bf4ede01STejun Heo */ 948bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 949bf4ede01STejun Heo struct work_struct **nextp) 950bf4ede01STejun Heo { 951bf4ede01STejun Heo struct work_struct *n; 952bf4ede01STejun Heo 953bf4ede01STejun Heo /* 954bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 955bf4ede01STejun Heo * use NULL for list head. 956bf4ede01STejun Heo */ 957bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 958bf4ede01STejun Heo list_move_tail(&work->entry, head); 959bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 960bf4ede01STejun Heo break; 961bf4ede01STejun Heo } 962bf4ede01STejun Heo 963bf4ede01STejun Heo /* 964bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 965bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 966bf4ede01STejun Heo * needs to be updated. 967bf4ede01STejun Heo */ 968bf4ede01STejun Heo if (nextp) 969bf4ede01STejun Heo *nextp = n; 970bf4ede01STejun Heo } 971bf4ede01STejun Heo 9723aa62497SLai Jiangshan static void cwq_activate_delayed_work(struct work_struct *work) 973bf4ede01STejun Heo { 9743aa62497SLai Jiangshan struct cpu_workqueue_struct *cwq = get_work_cwq(work); 975bf4ede01STejun Heo 976bf4ede01STejun Heo trace_workqueue_activate_work(work); 977bf4ede01STejun Heo move_linked_works(work, &cwq->pool->worklist, NULL); 978bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 979bf4ede01STejun Heo cwq->nr_active++; 980bf4ede01STejun Heo } 981bf4ede01STejun Heo 9823aa62497SLai Jiangshan static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 9833aa62497SLai Jiangshan { 9843aa62497SLai Jiangshan struct work_struct *work = list_first_entry(&cwq->delayed_works, 9853aa62497SLai Jiangshan struct work_struct, entry); 9863aa62497SLai Jiangshan 9873aa62497SLai Jiangshan cwq_activate_delayed_work(work); 9883aa62497SLai Jiangshan } 9893aa62497SLai Jiangshan 990bf4ede01STejun Heo /** 991bf4ede01STejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 992bf4ede01STejun Heo * @cwq: cwq of interest 993bf4ede01STejun Heo * @color: color of work which left the queue 994bf4ede01STejun Heo * 995bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 996bf4ede01STejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 997bf4ede01STejun Heo * 998bf4ede01STejun Heo * CONTEXT: 999d565ed63STejun Heo * spin_lock_irq(pool->lock). 1000bf4ede01STejun Heo */ 1001b3f9f405SLai Jiangshan static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color) 1002bf4ede01STejun Heo { 1003bf4ede01STejun Heo /* ignore uncolored works */ 1004bf4ede01STejun Heo if (color == WORK_NO_COLOR) 1005bf4ede01STejun Heo return; 1006bf4ede01STejun Heo 1007bf4ede01STejun Heo cwq->nr_in_flight[color]--; 1008bf4ede01STejun Heo 1009bf4ede01STejun Heo cwq->nr_active--; 1010bf4ede01STejun Heo if (!list_empty(&cwq->delayed_works)) { 1011bf4ede01STejun Heo /* one down, submit a delayed one */ 1012bf4ede01STejun Heo if (cwq->nr_active < cwq->max_active) 1013bf4ede01STejun Heo cwq_activate_first_delayed(cwq); 1014bf4ede01STejun Heo } 1015bf4ede01STejun Heo 1016bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1017bf4ede01STejun Heo if (likely(cwq->flush_color != color)) 1018bf4ede01STejun Heo return; 1019bf4ede01STejun Heo 1020bf4ede01STejun Heo /* are there still in-flight works? */ 1021bf4ede01STejun Heo if (cwq->nr_in_flight[color]) 1022bf4ede01STejun Heo return; 1023bf4ede01STejun Heo 1024bf4ede01STejun Heo /* this cwq is done, clear flush_color */ 1025bf4ede01STejun Heo cwq->flush_color = -1; 1026bf4ede01STejun Heo 1027bf4ede01STejun Heo /* 1028bf4ede01STejun Heo * If this was the last cwq, wake up the first flusher. It 1029bf4ede01STejun Heo * will handle the rest. 1030bf4ede01STejun Heo */ 1031bf4ede01STejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 1032bf4ede01STejun Heo complete(&cwq->wq->first_flusher->done); 1033bf4ede01STejun Heo } 1034bf4ede01STejun Heo 103536e227d2STejun Heo /** 1036bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 103736e227d2STejun Heo * @work: work item to steal 103836e227d2STejun Heo * @is_dwork: @work is a delayed_work 1039bbb68dfaSTejun Heo * @flags: place to store irq state 104036e227d2STejun Heo * 104136e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 104236e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 104336e227d2STejun Heo * 104436e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 104536e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 104636e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1047bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1048bbb68dfaSTejun Heo * for arbitrarily long 104936e227d2STejun Heo * 1050bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1051e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1052e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1053e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1054bbb68dfaSTejun Heo * 1055bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1056bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1057bbb68dfaSTejun Heo * 1058e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1059bf4ede01STejun Heo */ 1060bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1061bbb68dfaSTejun Heo unsigned long *flags) 1062bf4ede01STejun Heo { 1063d565ed63STejun Heo struct worker_pool *pool; 1064bf4ede01STejun Heo 1065bbb68dfaSTejun Heo local_irq_save(*flags); 1066bbb68dfaSTejun Heo 106736e227d2STejun Heo /* try to steal the timer if it exists */ 106836e227d2STejun Heo if (is_dwork) { 106936e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 107036e227d2STejun Heo 1071e0aecdd8STejun Heo /* 1072e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1073e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1074e0aecdd8STejun Heo * running on the local CPU. 1075e0aecdd8STejun Heo */ 107636e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 107736e227d2STejun Heo return 1; 107836e227d2STejun Heo } 107936e227d2STejun Heo 108036e227d2STejun Heo /* try to claim PENDING the normal way */ 1081bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1082bf4ede01STejun Heo return 0; 1083bf4ede01STejun Heo 1084bf4ede01STejun Heo /* 1085bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1086bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1087bf4ede01STejun Heo */ 1088d565ed63STejun Heo pool = get_work_pool(work); 1089d565ed63STejun Heo if (!pool) 1090bbb68dfaSTejun Heo goto fail; 1091bf4ede01STejun Heo 1092d565ed63STejun Heo spin_lock(&pool->lock); 1093bf4ede01STejun Heo if (!list_empty(&work->entry)) { 1094bf4ede01STejun Heo /* 1095d565ed63STejun Heo * This work is queued, but perhaps we locked the wrong 1096d565ed63STejun Heo * pool. In that case we must see the new value after 1097d565ed63STejun Heo * rmb(), see insert_work()->wmb(). 1098bf4ede01STejun Heo */ 1099bf4ede01STejun Heo smp_rmb(); 1100d565ed63STejun Heo if (pool == get_work_pool(work)) { 1101bf4ede01STejun Heo debug_work_deactivate(work); 11023aa62497SLai Jiangshan 11033aa62497SLai Jiangshan /* 11043aa62497SLai Jiangshan * A delayed work item cannot be grabbed directly 11053aa62497SLai Jiangshan * because it might have linked NO_COLOR work items 11063aa62497SLai Jiangshan * which, if left on the delayed_list, will confuse 11073aa62497SLai Jiangshan * cwq->nr_active management later on and cause 11083aa62497SLai Jiangshan * stall. Make sure the work item is activated 11093aa62497SLai Jiangshan * before grabbing. 11103aa62497SLai Jiangshan */ 11113aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 11123aa62497SLai Jiangshan cwq_activate_delayed_work(work); 11133aa62497SLai Jiangshan 1114bf4ede01STejun Heo list_del_init(&work->entry); 1115bf4ede01STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 1116b3f9f405SLai Jiangshan get_work_color(work)); 111736e227d2STejun Heo 1118d565ed63STejun Heo spin_unlock(&pool->lock); 111936e227d2STejun Heo return 1; 1120bf4ede01STejun Heo } 1121bf4ede01STejun Heo } 1122d565ed63STejun Heo spin_unlock(&pool->lock); 1123bbb68dfaSTejun Heo fail: 1124bbb68dfaSTejun Heo local_irq_restore(*flags); 1125bbb68dfaSTejun Heo if (work_is_canceling(work)) 1126bbb68dfaSTejun Heo return -ENOENT; 1127bbb68dfaSTejun Heo cpu_relax(); 112836e227d2STejun Heo return -EAGAIN; 1129bf4ede01STejun Heo } 1130bf4ede01STejun Heo 1131bf4ede01STejun Heo /** 1132706026c2STejun Heo * insert_work - insert a work into a pool 11334690c4abSTejun Heo * @cwq: cwq @work belongs to 11344690c4abSTejun Heo * @work: work to insert 11354690c4abSTejun Heo * @head: insertion point 11364690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11374690c4abSTejun Heo * 1138706026c2STejun Heo * Insert @work which belongs to @cwq after @head. @extra_flags is or'd to 1139706026c2STejun Heo * work_struct flags. 11404690c4abSTejun Heo * 11414690c4abSTejun Heo * CONTEXT: 1142d565ed63STejun Heo * spin_lock_irq(pool->lock). 1143365970a1SDavid Howells */ 1144b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 11454690c4abSTejun Heo struct work_struct *work, struct list_head *head, 11464690c4abSTejun Heo unsigned int extra_flags) 1147b89deed3SOleg Nesterov { 114863d95a91STejun Heo struct worker_pool *pool = cwq->pool; 1149e1d8aa9fSFrederic Weisbecker 11504690c4abSTejun Heo /* we own @work, set data and link */ 11517a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 11524690c4abSTejun Heo 11536e84d644SOleg Nesterov /* 11546e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 11556e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 11566e84d644SOleg Nesterov */ 11576e84d644SOleg Nesterov smp_wmb(); 11584690c4abSTejun Heo 11591a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1160e22bee78STejun Heo 1161e22bee78STejun Heo /* 1162e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1163e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1164e22bee78STejun Heo * lying around lazily while there are works to be processed. 1165e22bee78STejun Heo */ 1166e22bee78STejun Heo smp_mb(); 1167e22bee78STejun Heo 116863d95a91STejun Heo if (__need_more_worker(pool)) 116963d95a91STejun Heo wake_up_worker(pool); 1170b89deed3SOleg Nesterov } 1171b89deed3SOleg Nesterov 1172c8efcc25STejun Heo /* 1173c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 1174c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 1175c8efcc25STejun Heo * cold paths. 1176c8efcc25STejun Heo */ 1177c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1178c8efcc25STejun Heo { 1179c8efcc25STejun Heo unsigned long flags; 1180c8efcc25STejun Heo unsigned int cpu; 1181c8efcc25STejun Heo 1182706026c2STejun Heo for_each_wq_cpu(cpu) { 1183c9e7cf27STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 1184c9e7cf27STejun Heo struct worker_pool *pool = cwq->pool; 1185c8efcc25STejun Heo struct worker *worker; 1186c8efcc25STejun Heo struct hlist_node *pos; 1187c8efcc25STejun Heo int i; 1188c8efcc25STejun Heo 1189d565ed63STejun Heo spin_lock_irqsave(&pool->lock, flags); 1190c9e7cf27STejun Heo for_each_busy_worker(worker, i, pos, pool) { 1191c8efcc25STejun Heo if (worker->task != current) 1192c8efcc25STejun Heo continue; 1193d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 1194c8efcc25STejun Heo /* 1195c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 1196c8efcc25STejun Heo * is headed to the same workqueue. 1197c8efcc25STejun Heo */ 1198c8efcc25STejun Heo return worker->current_cwq->wq == wq; 1199c8efcc25STejun Heo } 1200d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 1201c8efcc25STejun Heo } 1202c8efcc25STejun Heo return false; 1203c8efcc25STejun Heo } 1204c8efcc25STejun Heo 12054690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 12061da177e4SLinus Torvalds struct work_struct *work) 12071da177e4SLinus Torvalds { 1208d565ed63STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 1209d565ed63STejun Heo struct worker_pool *pool; 1210502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 12111e19ffc6STejun Heo struct list_head *worklist; 12128a2e8e5dSTejun Heo unsigned int work_flags; 1213b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12148930cabaSTejun Heo 12158930cabaSTejun Heo /* 12168930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12178930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12188930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12198930cabaSTejun Heo * happen with IRQ disabled. 12208930cabaSTejun Heo */ 12218930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12221da177e4SLinus Torvalds 1223dc186ad7SThomas Gleixner debug_work_activate(work); 12241e19ffc6STejun Heo 1225c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 12269c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1227c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1228e41e704bSTejun Heo return; 1229e41e704bSTejun Heo 1230d565ed63STejun Heo /* determine pool to use */ 1231c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1232c9e7cf27STejun Heo struct worker_pool *last_pool; 1233c7fc77f7STejun Heo 123457469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1235f3421797STejun Heo cpu = raw_smp_processor_id(); 1236f3421797STejun Heo 123718aa9effSTejun Heo /* 1238dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1239dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1240dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1241dbf2576eSTejun Heo * non-reentrancy. 124218aa9effSTejun Heo */ 1243d565ed63STejun Heo pool = get_std_worker_pool(cpu, highpri); 1244c9e7cf27STejun Heo last_pool = get_work_pool(work); 1245dbf2576eSTejun Heo 1246d565ed63STejun Heo if (last_pool && last_pool != pool) { 124718aa9effSTejun Heo struct worker *worker; 124818aa9effSTejun Heo 1249d565ed63STejun Heo spin_lock(&last_pool->lock); 125018aa9effSTejun Heo 1251c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 125218aa9effSTejun Heo 125318aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 1254d565ed63STejun Heo pool = last_pool; 125518aa9effSTejun Heo else { 125618aa9effSTejun Heo /* meh... not running there, queue here */ 1257d565ed63STejun Heo spin_unlock(&last_pool->lock); 1258d565ed63STejun Heo spin_lock(&pool->lock); 125918aa9effSTejun Heo } 12608930cabaSTejun Heo } else { 1261d565ed63STejun Heo spin_lock(&pool->lock); 12628930cabaSTejun Heo } 1263f3421797STejun Heo } else { 1264d565ed63STejun Heo pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri); 1265d565ed63STejun Heo spin_lock(&pool->lock); 1266502ca9d8STejun Heo } 1267502ca9d8STejun Heo 1268d565ed63STejun Heo /* pool determined, get cwq and queue */ 1269d565ed63STejun Heo cwq = get_cwq(pool->cpu, wq); 1270b75cac93SJoonsoo Kim trace_workqueue_queue_work(req_cpu, cwq, work); 1271502ca9d8STejun Heo 1272f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1273d565ed63STejun Heo spin_unlock(&pool->lock); 1274f5b2552bSDan Carpenter return; 1275f5b2552bSDan Carpenter } 12761e19ffc6STejun Heo 127773f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 12788a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 12791e19ffc6STejun Heo 12801e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1281cdadf009STejun Heo trace_workqueue_activate_work(work); 12821e19ffc6STejun Heo cwq->nr_active++; 12833270476aSTejun Heo worklist = &cwq->pool->worklist; 12848a2e8e5dSTejun Heo } else { 12858a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 12861e19ffc6STejun Heo worklist = &cwq->delayed_works; 12878a2e8e5dSTejun Heo } 12881e19ffc6STejun Heo 12898a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 12901e19ffc6STejun Heo 1291d565ed63STejun Heo spin_unlock(&pool->lock); 12921da177e4SLinus Torvalds } 12931da177e4SLinus Torvalds 12940fcb78c2SRolf Eike Beer /** 1295c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1296c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1297c1a220e7SZhang Rui * @wq: workqueue to use 1298c1a220e7SZhang Rui * @work: work to queue 1299c1a220e7SZhang Rui * 1300d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1301c1a220e7SZhang Rui * 1302c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1303c1a220e7SZhang Rui * can't go away. 1304c1a220e7SZhang Rui */ 1305d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1306d4283e93STejun Heo struct work_struct *work) 1307c1a220e7SZhang Rui { 1308d4283e93STejun Heo bool ret = false; 13098930cabaSTejun Heo unsigned long flags; 13108930cabaSTejun Heo 13118930cabaSTejun Heo local_irq_save(flags); 1312c1a220e7SZhang Rui 131322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13144690c4abSTejun Heo __queue_work(cpu, wq, work); 1315d4283e93STejun Heo ret = true; 1316c1a220e7SZhang Rui } 13178930cabaSTejun Heo 13188930cabaSTejun Heo local_irq_restore(flags); 1319c1a220e7SZhang Rui return ret; 1320c1a220e7SZhang Rui } 1321c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1322c1a220e7SZhang Rui 13230a13c00eSTejun Heo /** 13240a13c00eSTejun Heo * queue_work - queue work on a workqueue 13250a13c00eSTejun Heo * @wq: workqueue to use 13260a13c00eSTejun Heo * @work: work to queue 13270a13c00eSTejun Heo * 1328d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 13290a13c00eSTejun Heo * 13300a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 13310a13c00eSTejun Heo * it can be processed by another CPU. 13320a13c00eSTejun Heo */ 1333d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13340a13c00eSTejun Heo { 133557469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13360a13c00eSTejun Heo } 13370a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13380a13c00eSTejun Heo 1339d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13401da177e4SLinus Torvalds { 134152bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13421da177e4SLinus Torvalds 1343e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 134460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13451da177e4SLinus Torvalds } 1346d8e794dfSTejun Heo EXPORT_SYMBOL_GPL(delayed_work_timer_fn); 13471da177e4SLinus Torvalds 13487beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 134952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13501da177e4SLinus Torvalds { 13517beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13527beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13531da177e4SLinus Torvalds 13547beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13557beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1356fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1357fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13587beb2edfSTejun Heo 13598852aac2STejun Heo /* 13608852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13618852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13628852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13638852aac2STejun Heo * on that there's no such delay when @delay is 0. 13648852aac2STejun Heo */ 13658852aac2STejun Heo if (!delay) { 13668852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13678852aac2STejun Heo return; 13688852aac2STejun Heo } 13698852aac2STejun Heo 13707beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13717beb2edfSTejun Heo 137260c057bcSLai Jiangshan dwork->wq = wq; 13731265057fSTejun Heo dwork->cpu = cpu; 13747beb2edfSTejun Heo timer->expires = jiffies + delay; 13757beb2edfSTejun Heo 13767beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13777beb2edfSTejun Heo add_timer_on(timer, cpu); 13787beb2edfSTejun Heo else 13797beb2edfSTejun Heo add_timer(timer); 13807beb2edfSTejun Heo } 13811da177e4SLinus Torvalds 13820fcb78c2SRolf Eike Beer /** 13830fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13840fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13850fcb78c2SRolf Eike Beer * @wq: workqueue to use 1386af9997e4SRandy Dunlap * @dwork: work to queue 13870fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13880fcb78c2SRolf Eike Beer * 1389715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1390715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1391715f1300STejun Heo * execution. 13920fcb78c2SRolf Eike Beer */ 1393d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 139452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13957a6bc1cdSVenkatesh Pallipadi { 139652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1397d4283e93STejun Heo bool ret = false; 13988930cabaSTejun Heo unsigned long flags; 13998930cabaSTejun Heo 14008930cabaSTejun Heo /* read the comment in __queue_work() */ 14018930cabaSTejun Heo local_irq_save(flags); 14027a6bc1cdSVenkatesh Pallipadi 140322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14047beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1405d4283e93STejun Heo ret = true; 14067a6bc1cdSVenkatesh Pallipadi } 14078930cabaSTejun Heo 14088930cabaSTejun Heo local_irq_restore(flags); 14097a6bc1cdSVenkatesh Pallipadi return ret; 14107a6bc1cdSVenkatesh Pallipadi } 1411ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14121da177e4SLinus Torvalds 1413c8e55f36STejun Heo /** 14140a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 14150a13c00eSTejun Heo * @wq: workqueue to use 14160a13c00eSTejun Heo * @dwork: delayable work to queue 14170a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 14180a13c00eSTejun Heo * 1419715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 14200a13c00eSTejun Heo */ 1421d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 14220a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 14230a13c00eSTejun Heo { 142457469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14250a13c00eSTejun Heo } 14260a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 14270a13c00eSTejun Heo 14280a13c00eSTejun Heo /** 14298376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14308376fe22STejun Heo * @cpu: CPU number to execute work on 14318376fe22STejun Heo * @wq: workqueue to use 14328376fe22STejun Heo * @dwork: work to queue 14338376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14348376fe22STejun Heo * 14358376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14368376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14378376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14388376fe22STejun Heo * current state. 14398376fe22STejun Heo * 14408376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14418376fe22STejun Heo * pending and its timer was modified. 14428376fe22STejun Heo * 1443e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14448376fe22STejun Heo * See try_to_grab_pending() for details. 14458376fe22STejun Heo */ 14468376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14478376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14488376fe22STejun Heo { 14498376fe22STejun Heo unsigned long flags; 14508376fe22STejun Heo int ret; 14518376fe22STejun Heo 14528376fe22STejun Heo do { 14538376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14548376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14558376fe22STejun Heo 14568376fe22STejun Heo if (likely(ret >= 0)) { 14578376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14588376fe22STejun Heo local_irq_restore(flags); 14598376fe22STejun Heo } 14608376fe22STejun Heo 14618376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14628376fe22STejun Heo return ret; 14638376fe22STejun Heo } 14648376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14658376fe22STejun Heo 14668376fe22STejun Heo /** 14678376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14688376fe22STejun Heo * @wq: workqueue to use 14698376fe22STejun Heo * @dwork: work to queue 14708376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14718376fe22STejun Heo * 14728376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14738376fe22STejun Heo */ 14748376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14758376fe22STejun Heo unsigned long delay) 14768376fe22STejun Heo { 14778376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14788376fe22STejun Heo } 14798376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14808376fe22STejun Heo 14818376fe22STejun Heo /** 1482c8e55f36STejun Heo * worker_enter_idle - enter idle state 1483c8e55f36STejun Heo * @worker: worker which is entering idle state 1484c8e55f36STejun Heo * 1485c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1486c8e55f36STejun Heo * necessary. 1487c8e55f36STejun Heo * 1488c8e55f36STejun Heo * LOCKING: 1489d565ed63STejun Heo * spin_lock_irq(pool->lock). 1490c8e55f36STejun Heo */ 1491c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14921da177e4SLinus Torvalds { 1493bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1494c8e55f36STejun Heo 1495c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1496c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1497c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1498c8e55f36STejun Heo 1499cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1500cb444766STejun Heo worker->flags |= WORKER_IDLE; 1501bd7bdd43STejun Heo pool->nr_idle++; 1502e22bee78STejun Heo worker->last_active = jiffies; 1503c8e55f36STejun Heo 1504c8e55f36STejun Heo /* idle_list is LIFO */ 1505bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1506db7bccf4STejun Heo 150763d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1508628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1509cb444766STejun Heo 1510544ecf31STejun Heo /* 1511706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1512d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1513628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1514628c78e7STejun Heo * unbind is not in progress. 1515544ecf31STejun Heo */ 151624647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1517bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 151863d95a91STejun Heo atomic_read(get_pool_nr_running(pool))); 1519c8e55f36STejun Heo } 1520c8e55f36STejun Heo 1521c8e55f36STejun Heo /** 1522c8e55f36STejun Heo * worker_leave_idle - leave idle state 1523c8e55f36STejun Heo * @worker: worker which is leaving idle state 1524c8e55f36STejun Heo * 1525c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1526c8e55f36STejun Heo * 1527c8e55f36STejun Heo * LOCKING: 1528d565ed63STejun Heo * spin_lock_irq(pool->lock). 1529c8e55f36STejun Heo */ 1530c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1531c8e55f36STejun Heo { 1532bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1533c8e55f36STejun Heo 1534c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1535d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1536bd7bdd43STejun Heo pool->nr_idle--; 1537c8e55f36STejun Heo list_del_init(&worker->entry); 1538c8e55f36STejun Heo } 1539c8e55f36STejun Heo 1540e22bee78STejun Heo /** 1541706026c2STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool 1542e22bee78STejun Heo * @worker: self 1543e22bee78STejun Heo * 1544e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1545e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1546e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1547e22bee78STejun Heo * guaranteed to execute on the cpu. 1548e22bee78STejun Heo * 1549e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1550e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1551e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1552e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1553706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1554e22bee78STejun Heo * [dis]associated in the meantime. 1555e22bee78STejun Heo * 1556706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 155724647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1558f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1559f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1560f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1561e22bee78STejun Heo * 1562e22bee78STejun Heo * CONTEXT: 1563d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1564e22bee78STejun Heo * held. 1565e22bee78STejun Heo * 1566e22bee78STejun Heo * RETURNS: 1567706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1568e22bee78STejun Heo * bound), %false if offline. 1569e22bee78STejun Heo */ 1570e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1571d565ed63STejun Heo __acquires(&pool->lock) 1572e22bee78STejun Heo { 157324647570STejun Heo struct worker_pool *pool = worker->pool; 1574e22bee78STejun Heo struct task_struct *task = worker->task; 1575e22bee78STejun Heo 1576e22bee78STejun Heo while (true) { 1577e22bee78STejun Heo /* 1578e22bee78STejun Heo * The following call may fail, succeed or succeed 1579e22bee78STejun Heo * without actually migrating the task to the cpu if 1580e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 158124647570STejun Heo * against POOL_DISASSOCIATED. 1582e22bee78STejun Heo */ 158324647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 1584ec22ca5eSTejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu)); 1585e22bee78STejun Heo 1586d565ed63STejun Heo spin_lock_irq(&pool->lock); 158724647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1588e22bee78STejun Heo return false; 1589ec22ca5eSTejun Heo if (task_cpu(task) == pool->cpu && 1590e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1591ec22ca5eSTejun Heo get_cpu_mask(pool->cpu))) 1592e22bee78STejun Heo return true; 1593d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1594e22bee78STejun Heo 15955035b20fSTejun Heo /* 15965035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 15975035b20fSTejun Heo * and retry migration. cond_resched() is required here; 15985035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 15995035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16005035b20fSTejun Heo */ 1601e22bee78STejun Heo cpu_relax(); 16025035b20fSTejun Heo cond_resched(); 1603e22bee78STejun Heo } 1604e22bee78STejun Heo } 1605e22bee78STejun Heo 1606e22bee78STejun Heo /* 1607ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 16085f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 160925511a47STejun Heo */ 161025511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 161125511a47STejun Heo { 16125f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 16135f7dabfdSLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 16145f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 161525511a47STejun Heo 1616ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1617ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1618d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 161925511a47STejun Heo } 162025511a47STejun Heo 162125511a47STejun Heo /* 162225511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1623403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1624403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1625403c821dSTejun Heo * executed twice without intervening cpu down. 1626e22bee78STejun Heo */ 162725511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1628e22bee78STejun Heo { 1629e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1630e22bee78STejun Heo 1631eab6d828SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 1632eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1633e22bee78STejun Heo 1634d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1635e22bee78STejun Heo } 1636e22bee78STejun Heo 163725511a47STejun Heo /** 163894cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 163994cf58bbSTejun Heo * @pool: pool of interest 164025511a47STejun Heo * 164194cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 164225511a47STejun Heo * is different for idle and busy ones. 164325511a47STejun Heo * 1644ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1645ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1646ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1647ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 164825511a47STejun Heo * 1649ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1650ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1651ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1652ea1abd61SLai Jiangshan * rebind. 165325511a47STejun Heo * 1654ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1655ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1656ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1657ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 165825511a47STejun Heo */ 165994cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 166025511a47STejun Heo { 1661ea1abd61SLai Jiangshan struct worker *worker, *n; 166225511a47STejun Heo struct hlist_node *pos; 166325511a47STejun Heo int i; 166425511a47STejun Heo 1665b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 1666d565ed63STejun Heo lockdep_assert_held(&pool->lock); 166725511a47STejun Heo 16685f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1669ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1670ea1abd61SLai Jiangshan /* 167194cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 167294cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1673ea1abd61SLai Jiangshan */ 1674ea1abd61SLai Jiangshan list_del_init(&worker->entry); 167525511a47STejun Heo 167625511a47STejun Heo /* 167794cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 167894cf58bbSTejun Heo * idle_worker_rebind(). 167925511a47STejun Heo */ 168025511a47STejun Heo wake_up_process(worker->task); 168125511a47STejun Heo } 168225511a47STejun Heo 1683ea1abd61SLai Jiangshan /* rebind busy workers */ 1684c9e7cf27STejun Heo for_each_busy_worker(worker, i, pos, pool) { 168525511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1686e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 168725511a47STejun Heo 168825511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 168925511a47STejun Heo work_data_bits(rebind_work))) 169025511a47STejun Heo continue; 169125511a47STejun Heo 169225511a47STejun Heo debug_work_activate(rebind_work); 169390beca5dSTejun Heo 169490beca5dSTejun Heo /* 169594cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 169694cf58bbSTejun Heo * and @cwq->pool consistent for sanity. 169790beca5dSTejun Heo */ 1698e34cdddbSTejun Heo if (std_worker_pool_pri(worker->pool)) 1699e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1700e2b6a6d5SJoonsoo Kim else 1701e2b6a6d5SJoonsoo Kim wq = system_wq; 1702ec58815aSTejun Heo 1703ec22ca5eSTejun Heo insert_work(get_cwq(pool->cpu, wq), rebind_work, 170425511a47STejun Heo worker->scheduled.next, 170525511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1706ec58815aSTejun Heo } 170725511a47STejun Heo } 170825511a47STejun Heo 1709c34056a3STejun Heo static struct worker *alloc_worker(void) 1710c34056a3STejun Heo { 1711c34056a3STejun Heo struct worker *worker; 1712c34056a3STejun Heo 1713c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1714c8e55f36STejun Heo if (worker) { 1715c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1716affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 171725511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1718e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1719e22bee78STejun Heo worker->flags = WORKER_PREP; 1720c8e55f36STejun Heo } 1721c34056a3STejun Heo return worker; 1722c34056a3STejun Heo } 1723c34056a3STejun Heo 1724c34056a3STejun Heo /** 1725c34056a3STejun Heo * create_worker - create a new workqueue worker 172663d95a91STejun Heo * @pool: pool the new worker will belong to 1727c34056a3STejun Heo * 172863d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1729c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1730c34056a3STejun Heo * destroy_worker(). 1731c34056a3STejun Heo * 1732c34056a3STejun Heo * CONTEXT: 1733c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1734c34056a3STejun Heo * 1735c34056a3STejun Heo * RETURNS: 1736c34056a3STejun Heo * Pointer to the newly created worker. 1737c34056a3STejun Heo */ 1738bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1739c34056a3STejun Heo { 1740e34cdddbSTejun Heo const char *pri = std_worker_pool_pri(pool) ? "H" : ""; 1741c34056a3STejun Heo struct worker *worker = NULL; 1742f3421797STejun Heo int id = -1; 1743c34056a3STejun Heo 1744d565ed63STejun Heo spin_lock_irq(&pool->lock); 1745bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1746d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1747bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1748c34056a3STejun Heo goto fail; 1749d565ed63STejun Heo spin_lock_irq(&pool->lock); 1750c34056a3STejun Heo } 1751d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1752c34056a3STejun Heo 1753c34056a3STejun Heo worker = alloc_worker(); 1754c34056a3STejun Heo if (!worker) 1755c34056a3STejun Heo goto fail; 1756c34056a3STejun Heo 1757bd7bdd43STejun Heo worker->pool = pool; 1758c34056a3STejun Heo worker->id = id; 1759c34056a3STejun Heo 1760ec22ca5eSTejun Heo if (pool->cpu != WORK_CPU_UNBOUND) 176194dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1762ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1763ec22ca5eSTejun Heo "kworker/%u:%d%s", pool->cpu, id, pri); 1764f3421797STejun Heo else 1765f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 17663270476aSTejun Heo "kworker/u:%d%s", id, pri); 1767c34056a3STejun Heo if (IS_ERR(worker->task)) 1768c34056a3STejun Heo goto fail; 1769c34056a3STejun Heo 1770e34cdddbSTejun Heo if (std_worker_pool_pri(pool)) 17713270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 17723270476aSTejun Heo 1773db7bccf4STejun Heo /* 1774bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 177524647570STejun Heo * %POOL_DISASSOCIATED. The caller is responsible for ensuring the 1776bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1777bc2ae0f5STejun Heo * above the flag definition for details. 1778bc2ae0f5STejun Heo * 1779bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1780bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1781db7bccf4STejun Heo */ 178224647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) { 1783ec22ca5eSTejun Heo kthread_bind(worker->task, pool->cpu); 1784bc2ae0f5STejun Heo } else { 1785db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1786f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1787f3421797STejun Heo } 1788c34056a3STejun Heo 1789c34056a3STejun Heo return worker; 1790c34056a3STejun Heo fail: 1791c34056a3STejun Heo if (id >= 0) { 1792d565ed63STejun Heo spin_lock_irq(&pool->lock); 1793bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1794d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1795c34056a3STejun Heo } 1796c34056a3STejun Heo kfree(worker); 1797c34056a3STejun Heo return NULL; 1798c34056a3STejun Heo } 1799c34056a3STejun Heo 1800c34056a3STejun Heo /** 1801c34056a3STejun Heo * start_worker - start a newly created worker 1802c34056a3STejun Heo * @worker: worker to start 1803c34056a3STejun Heo * 1804706026c2STejun Heo * Make the pool aware of @worker and start it. 1805c34056a3STejun Heo * 1806c34056a3STejun Heo * CONTEXT: 1807d565ed63STejun Heo * spin_lock_irq(pool->lock). 1808c34056a3STejun Heo */ 1809c34056a3STejun Heo static void start_worker(struct worker *worker) 1810c34056a3STejun Heo { 1811cb444766STejun Heo worker->flags |= WORKER_STARTED; 1812bd7bdd43STejun Heo worker->pool->nr_workers++; 1813c8e55f36STejun Heo worker_enter_idle(worker); 1814c34056a3STejun Heo wake_up_process(worker->task); 1815c34056a3STejun Heo } 1816c34056a3STejun Heo 1817c34056a3STejun Heo /** 1818c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1819c34056a3STejun Heo * @worker: worker to be destroyed 1820c34056a3STejun Heo * 1821706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1822c8e55f36STejun Heo * 1823c8e55f36STejun Heo * CONTEXT: 1824d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1825c34056a3STejun Heo */ 1826c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1827c34056a3STejun Heo { 1828bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1829c34056a3STejun Heo int id = worker->id; 1830c34056a3STejun Heo 1831c34056a3STejun Heo /* sanity check frenzy */ 1832c34056a3STejun Heo BUG_ON(worker->current_work); 1833affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1834c34056a3STejun Heo 1835c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1836bd7bdd43STejun Heo pool->nr_workers--; 1837c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1838bd7bdd43STejun Heo pool->nr_idle--; 1839c8e55f36STejun Heo 1840c8e55f36STejun Heo list_del_init(&worker->entry); 1841cb444766STejun Heo worker->flags |= WORKER_DIE; 1842c8e55f36STejun Heo 1843d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1844c8e55f36STejun Heo 1845c34056a3STejun Heo kthread_stop(worker->task); 1846c34056a3STejun Heo kfree(worker); 1847c34056a3STejun Heo 1848d565ed63STejun Heo spin_lock_irq(&pool->lock); 1849bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1850c34056a3STejun Heo } 1851c34056a3STejun Heo 185263d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1853e22bee78STejun Heo { 185463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1855e22bee78STejun Heo 1856d565ed63STejun Heo spin_lock_irq(&pool->lock); 1857e22bee78STejun Heo 185863d95a91STejun Heo if (too_many_workers(pool)) { 1859e22bee78STejun Heo struct worker *worker; 1860e22bee78STejun Heo unsigned long expires; 1861e22bee78STejun Heo 1862e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 186363d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1864e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1865e22bee78STejun Heo 1866e22bee78STejun Heo if (time_before(jiffies, expires)) 186763d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1868e22bee78STejun Heo else { 1869e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 187011ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 187163d95a91STejun Heo wake_up_worker(pool); 1872e22bee78STejun Heo } 1873e22bee78STejun Heo } 1874e22bee78STejun Heo 1875d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1876e22bee78STejun Heo } 1877e22bee78STejun Heo 1878e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1879e22bee78STejun Heo { 1880e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1881e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1882f3421797STejun Heo unsigned int cpu; 1883e22bee78STejun Heo 1884e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1885e22bee78STejun Heo return false; 1886e22bee78STejun Heo 1887e22bee78STejun Heo /* mayday mayday mayday */ 1888ec22ca5eSTejun Heo cpu = cwq->pool->cpu; 1889f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1890f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1891f3421797STejun Heo cpu = 0; 1892f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1893e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1894e22bee78STejun Heo return true; 1895e22bee78STejun Heo } 1896e22bee78STejun Heo 1897706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1898e22bee78STejun Heo { 189963d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1900e22bee78STejun Heo struct work_struct *work; 1901e22bee78STejun Heo 1902d565ed63STejun Heo spin_lock_irq(&pool->lock); 1903e22bee78STejun Heo 190463d95a91STejun Heo if (need_to_create_worker(pool)) { 1905e22bee78STejun Heo /* 1906e22bee78STejun Heo * We've been trying to create a new worker but 1907e22bee78STejun Heo * haven't been successful. We might be hitting an 1908e22bee78STejun Heo * allocation deadlock. Send distress signals to 1909e22bee78STejun Heo * rescuers. 1910e22bee78STejun Heo */ 191163d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1912e22bee78STejun Heo send_mayday(work); 1913e22bee78STejun Heo } 1914e22bee78STejun Heo 1915d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1916e22bee78STejun Heo 191763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1918e22bee78STejun Heo } 1919e22bee78STejun Heo 1920e22bee78STejun Heo /** 1921e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 192263d95a91STejun Heo * @pool: pool to create a new worker for 1923e22bee78STejun Heo * 192463d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1925e22bee78STejun Heo * have at least one idle worker on return from this function. If 1926e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 192763d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1928e22bee78STejun Heo * possible allocation deadlock. 1929e22bee78STejun Heo * 1930e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1931e22bee78STejun Heo * may_start_working() true. 1932e22bee78STejun Heo * 1933e22bee78STejun Heo * LOCKING: 1934d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1935e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1936e22bee78STejun Heo * manager. 1937e22bee78STejun Heo * 1938e22bee78STejun Heo * RETURNS: 1939d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1940e22bee78STejun Heo * otherwise. 1941e22bee78STejun Heo */ 194263d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1943d565ed63STejun Heo __releases(&pool->lock) 1944d565ed63STejun Heo __acquires(&pool->lock) 1945e22bee78STejun Heo { 194663d95a91STejun Heo if (!need_to_create_worker(pool)) 1947e22bee78STejun Heo return false; 1948e22bee78STejun Heo restart: 1949d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19509f9c2364STejun Heo 1951e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 195263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1953e22bee78STejun Heo 1954e22bee78STejun Heo while (true) { 1955e22bee78STejun Heo struct worker *worker; 1956e22bee78STejun Heo 1957bc2ae0f5STejun Heo worker = create_worker(pool); 1958e22bee78STejun Heo if (worker) { 195963d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1960d565ed63STejun Heo spin_lock_irq(&pool->lock); 1961e22bee78STejun Heo start_worker(worker); 196263d95a91STejun Heo BUG_ON(need_to_create_worker(pool)); 1963e22bee78STejun Heo return true; 1964e22bee78STejun Heo } 1965e22bee78STejun Heo 196663d95a91STejun Heo if (!need_to_create_worker(pool)) 1967e22bee78STejun Heo break; 1968e22bee78STejun Heo 1969e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1970e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19719f9c2364STejun Heo 197263d95a91STejun Heo if (!need_to_create_worker(pool)) 1973e22bee78STejun Heo break; 1974e22bee78STejun Heo } 1975e22bee78STejun Heo 197663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1977d565ed63STejun Heo spin_lock_irq(&pool->lock); 197863d95a91STejun Heo if (need_to_create_worker(pool)) 1979e22bee78STejun Heo goto restart; 1980e22bee78STejun Heo return true; 1981e22bee78STejun Heo } 1982e22bee78STejun Heo 1983e22bee78STejun Heo /** 1984e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 198563d95a91STejun Heo * @pool: pool to destroy workers for 1986e22bee78STejun Heo * 198763d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1988e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1989e22bee78STejun Heo * 1990e22bee78STejun Heo * LOCKING: 1991d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1992e22bee78STejun Heo * multiple times. Called only from manager. 1993e22bee78STejun Heo * 1994e22bee78STejun Heo * RETURNS: 1995d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1996e22bee78STejun Heo * otherwise. 1997e22bee78STejun Heo */ 199863d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1999e22bee78STejun Heo { 2000e22bee78STejun Heo bool ret = false; 2001e22bee78STejun Heo 200263d95a91STejun Heo while (too_many_workers(pool)) { 2003e22bee78STejun Heo struct worker *worker; 2004e22bee78STejun Heo unsigned long expires; 2005e22bee78STejun Heo 200663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2007e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2008e22bee78STejun Heo 2009e22bee78STejun Heo if (time_before(jiffies, expires)) { 201063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2011e22bee78STejun Heo break; 2012e22bee78STejun Heo } 2013e22bee78STejun Heo 2014e22bee78STejun Heo destroy_worker(worker); 2015e22bee78STejun Heo ret = true; 2016e22bee78STejun Heo } 2017e22bee78STejun Heo 2018e22bee78STejun Heo return ret; 2019e22bee78STejun Heo } 2020e22bee78STejun Heo 2021e22bee78STejun Heo /** 2022e22bee78STejun Heo * manage_workers - manage worker pool 2023e22bee78STejun Heo * @worker: self 2024e22bee78STejun Heo * 2025706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2026e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2027706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2028e22bee78STejun Heo * 2029e22bee78STejun Heo * The caller can safely start processing works on false return. On 2030e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2031e22bee78STejun Heo * and may_start_working() is true. 2032e22bee78STejun Heo * 2033e22bee78STejun Heo * CONTEXT: 2034d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2035e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2036e22bee78STejun Heo * 2037e22bee78STejun Heo * RETURNS: 2038d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2039d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2040e22bee78STejun Heo */ 2041e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2042e22bee78STejun Heo { 204363d95a91STejun Heo struct worker_pool *pool = worker->pool; 2044e22bee78STejun Heo bool ret = false; 2045e22bee78STejun Heo 2046ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2047e22bee78STejun Heo return ret; 2048e22bee78STejun Heo 2049552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2050ee378aa4SLai Jiangshan 2051ee378aa4SLai Jiangshan /* 2052ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2053ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2054ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2055ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2056ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2057b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2058ee378aa4SLai Jiangshan * manager against CPU hotplug. 2059ee378aa4SLai Jiangshan * 2060b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2061d565ed63STejun Heo * progress. trylock first without dropping @pool->lock. 2062ee378aa4SLai Jiangshan */ 2063b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2064d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2065b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2066ee378aa4SLai Jiangshan /* 2067ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2068b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2069ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2070706026c2STejun Heo * @pool's state and ours could have deviated. 2071ee378aa4SLai Jiangshan * 2072b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2073ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2074706026c2STejun Heo * on @pool's current state. Try it and adjust 2075ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2076ee378aa4SLai Jiangshan */ 2077ee378aa4SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 2078ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2079ee378aa4SLai Jiangshan else 2080ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2081ee378aa4SLai Jiangshan 2082ee378aa4SLai Jiangshan ret = true; 2083ee378aa4SLai Jiangshan } 2084ee378aa4SLai Jiangshan 208511ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2086e22bee78STejun Heo 2087e22bee78STejun Heo /* 2088e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2089e22bee78STejun Heo * on return. 2090e22bee78STejun Heo */ 209163d95a91STejun Heo ret |= maybe_destroy_workers(pool); 209263d95a91STejun Heo ret |= maybe_create_worker(pool); 2093e22bee78STejun Heo 2094552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2095b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2096e22bee78STejun Heo return ret; 2097e22bee78STejun Heo } 2098e22bee78STejun Heo 2099a62428c0STejun Heo /** 2100a62428c0STejun Heo * process_one_work - process single work 2101c34056a3STejun Heo * @worker: self 2102a62428c0STejun Heo * @work: work to process 2103a62428c0STejun Heo * 2104a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2105a62428c0STejun Heo * process a single work including synchronization against and 2106a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2107a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2108a62428c0STejun Heo * call this function to process a work. 2109a62428c0STejun Heo * 2110a62428c0STejun Heo * CONTEXT: 2111d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2112a62428c0STejun Heo */ 2113c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2114d565ed63STejun Heo __releases(&pool->lock) 2115d565ed63STejun Heo __acquires(&pool->lock) 21161da177e4SLinus Torvalds { 21177e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 2118bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2119fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 212073f53c4aSTejun Heo int work_color; 21217e11629dSTejun Heo struct worker *collision; 21224e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21234e6045f1SJohannes Berg /* 2124a62428c0STejun Heo * It is permissible to free the struct work_struct from 2125a62428c0STejun Heo * inside the function that is called from it, this we need to 2126a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2127a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2128a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21294e6045f1SJohannes Berg */ 21304d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21314d82a1deSPeter Zijlstra 21324d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21334e6045f1SJohannes Berg #endif 21346fec10a1STejun Heo /* 21356fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21366fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 213724647570STejun Heo * unbound or a disassociated pool. 21386fec10a1STejun Heo */ 21395f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 214024647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2141ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 214225511a47STejun Heo 21437e11629dSTejun Heo /* 21447e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21457e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21467e11629dSTejun Heo * already processing the work. If so, defer the work to the 21477e11629dSTejun Heo * currently executing one. 21487e11629dSTejun Heo */ 2149c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21507e11629dSTejun Heo if (unlikely(collision)) { 21517e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21527e11629dSTejun Heo return; 21537e11629dSTejun Heo } 21541da177e4SLinus Torvalds 21558930cabaSTejun Heo /* claim and dequeue */ 21561da177e4SLinus Torvalds debug_work_deactivate(work); 2157c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2158c34056a3STejun Heo worker->current_work = work; 2159a2c1c57bSTejun Heo worker->current_func = work->func; 21608cca0eeaSTejun Heo worker->current_cwq = cwq; 216173f53c4aSTejun Heo work_color = get_work_color(work); 21627a22ad75STejun Heo 2163a62428c0STejun Heo list_del_init(&work->entry); 2164a62428c0STejun Heo 2165649027d7STejun Heo /* 2166fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2167fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2168fb0e7bebSTejun Heo */ 2169fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2170fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2171fb0e7bebSTejun Heo 2172974271c4STejun Heo /* 2173d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2174974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2175974271c4STejun Heo */ 217663d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 217763d95a91STejun Heo wake_up_worker(pool); 2178974271c4STejun Heo 21798930cabaSTejun Heo /* 21807c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2181d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 218223657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 218323657bb1STejun Heo * disabled. 21848930cabaSTejun Heo */ 21857c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21861da177e4SLinus Torvalds 2187d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2188365970a1SDavid Howells 2189e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 21903295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2191e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2192a2c1c57bSTejun Heo worker->current_func(work); 2193e36c886aSArjan van de Ven /* 2194e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2195e36c886aSArjan van de Ven * point will only record its address. 2196e36c886aSArjan van de Ven */ 2197e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21983295f0efSIngo Molnar lock_map_release(&lockdep_map); 21993295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 22001da177e4SLinus Torvalds 2201d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2202044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2203044c782cSValentin Ilie " last function: %pf\n", 2204a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2205a2c1c57bSTejun Heo worker->current_func); 2206d5abe669SPeter Zijlstra debug_show_held_locks(current); 2207d5abe669SPeter Zijlstra dump_stack(); 2208d5abe669SPeter Zijlstra } 2209d5abe669SPeter Zijlstra 2210d565ed63STejun Heo spin_lock_irq(&pool->lock); 2211a62428c0STejun Heo 2212fb0e7bebSTejun Heo /* clear cpu intensive status */ 2213fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2214fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2215fb0e7bebSTejun Heo 2216a62428c0STejun Heo /* we're done with it, release */ 221742f8570fSSasha Levin hash_del(&worker->hentry); 2218c34056a3STejun Heo worker->current_work = NULL; 2219a2c1c57bSTejun Heo worker->current_func = NULL; 22208cca0eeaSTejun Heo worker->current_cwq = NULL; 2221b3f9f405SLai Jiangshan cwq_dec_nr_in_flight(cwq, work_color); 22221da177e4SLinus Torvalds } 22231da177e4SLinus Torvalds 2224affee4b2STejun Heo /** 2225affee4b2STejun Heo * process_scheduled_works - process scheduled works 2226affee4b2STejun Heo * @worker: self 2227affee4b2STejun Heo * 2228affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2229affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2230affee4b2STejun Heo * fetches a work from the top and executes it. 2231affee4b2STejun Heo * 2232affee4b2STejun Heo * CONTEXT: 2233d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2234affee4b2STejun Heo * multiple times. 2235affee4b2STejun Heo */ 2236affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22371da177e4SLinus Torvalds { 2238affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2239affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2240a62428c0STejun Heo struct work_struct, entry); 2241c34056a3STejun Heo process_one_work(worker, work); 2242a62428c0STejun Heo } 22431da177e4SLinus Torvalds } 22441da177e4SLinus Torvalds 22454690c4abSTejun Heo /** 22464690c4abSTejun Heo * worker_thread - the worker thread function 2247c34056a3STejun Heo * @__worker: self 22484690c4abSTejun Heo * 2249706026c2STejun Heo * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools 2250706026c2STejun Heo * of these per each cpu. These workers process all works regardless of 2251e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2252e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2253e22bee78STejun Heo * rescuer_thread(). 22544690c4abSTejun Heo */ 2255c34056a3STejun Heo static int worker_thread(void *__worker) 22561da177e4SLinus Torvalds { 2257c34056a3STejun Heo struct worker *worker = __worker; 2258bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22591da177e4SLinus Torvalds 2260e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2261e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2262c8e55f36STejun Heo woke_up: 2263d565ed63STejun Heo spin_lock_irq(&pool->lock); 2264affee4b2STejun Heo 22655f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22665f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2267d565ed63STejun Heo spin_unlock_irq(&pool->lock); 226825511a47STejun Heo 22695f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 227025511a47STejun Heo if (worker->flags & WORKER_DIE) { 2271e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2272c8e55f36STejun Heo return 0; 2273c8e55f36STejun Heo } 2274c8e55f36STejun Heo 22755f7dabfdSLai Jiangshan /* otherwise, rebind */ 227625511a47STejun Heo idle_worker_rebind(worker); 227725511a47STejun Heo goto woke_up; 227825511a47STejun Heo } 227925511a47STejun Heo 2280c8e55f36STejun Heo worker_leave_idle(worker); 2281db7bccf4STejun Heo recheck: 2282e22bee78STejun Heo /* no more worker necessary? */ 228363d95a91STejun Heo if (!need_more_worker(pool)) 2284e22bee78STejun Heo goto sleep; 2285e22bee78STejun Heo 2286e22bee78STejun Heo /* do we need to manage? */ 228763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2288e22bee78STejun Heo goto recheck; 2289e22bee78STejun Heo 2290c8e55f36STejun Heo /* 2291c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2292c8e55f36STejun Heo * preparing to process a work or actually processing it. 2293c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2294c8e55f36STejun Heo */ 2295c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 2296c8e55f36STejun Heo 2297e22bee78STejun Heo /* 2298e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2299e22bee78STejun Heo * at least one idle worker or that someone else has already 2300e22bee78STejun Heo * assumed the manager role. 2301e22bee78STejun Heo */ 2302e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2303e22bee78STejun Heo 2304e22bee78STejun Heo do { 2305affee4b2STejun Heo struct work_struct *work = 2306bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2307affee4b2STejun Heo struct work_struct, entry); 2308affee4b2STejun Heo 2309c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2310affee4b2STejun Heo /* optimization path, not strictly necessary */ 2311affee4b2STejun Heo process_one_work(worker, work); 2312affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2313affee4b2STejun Heo process_scheduled_works(worker); 2314affee4b2STejun Heo } else { 2315c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2316affee4b2STejun Heo process_scheduled_works(worker); 2317affee4b2STejun Heo } 231863d95a91STejun Heo } while (keep_working(pool)); 2319affee4b2STejun Heo 2320e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2321d313dd85STejun Heo sleep: 232263d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2323e22bee78STejun Heo goto recheck; 2324d313dd85STejun Heo 2325c8e55f36STejun Heo /* 2326d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2327d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2328d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2329d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2330d565ed63STejun Heo * event. 2331c8e55f36STejun Heo */ 2332c8e55f36STejun Heo worker_enter_idle(worker); 2333c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2334d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23351da177e4SLinus Torvalds schedule(); 2336c8e55f36STejun Heo goto woke_up; 23371da177e4SLinus Torvalds } 23381da177e4SLinus Torvalds 2339e22bee78STejun Heo /** 2340e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2341111c225aSTejun Heo * @__rescuer: self 2342e22bee78STejun Heo * 2343e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2344e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2345e22bee78STejun Heo * 2346706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2347e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2348e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2349e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2350e22bee78STejun Heo * the problem rescuer solves. 2351e22bee78STejun Heo * 2352706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2353706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2354e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2355e22bee78STejun Heo * 2356e22bee78STejun Heo * This should happen rarely. 2357e22bee78STejun Heo */ 2358111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2359e22bee78STejun Heo { 2360111c225aSTejun Heo struct worker *rescuer = __rescuer; 2361111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2362e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2363f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2364e22bee78STejun Heo unsigned int cpu; 2365e22bee78STejun Heo 2366e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2367111c225aSTejun Heo 2368111c225aSTejun Heo /* 2369111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2370111c225aSTejun Heo * doesn't participate in concurrency management. 2371111c225aSTejun Heo */ 2372111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2373e22bee78STejun Heo repeat: 2374e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23751da177e4SLinus Torvalds 2376412d32e6SMike Galbraith if (kthread_should_stop()) { 2377412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2378111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2379e22bee78STejun Heo return 0; 2380412d32e6SMike Galbraith } 23811da177e4SLinus Torvalds 2382f3421797STejun Heo /* 2383f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2384f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2385f3421797STejun Heo */ 2386f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2387f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2388f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2389bd7bdd43STejun Heo struct worker_pool *pool = cwq->pool; 2390e22bee78STejun Heo struct work_struct *work, *n; 2391e22bee78STejun Heo 2392e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2393f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2394e22bee78STejun Heo 2395e22bee78STejun Heo /* migrate to the target cpu if possible */ 2396bd7bdd43STejun Heo rescuer->pool = pool; 2397e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2398e22bee78STejun Heo 2399e22bee78STejun Heo /* 2400e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2401e22bee78STejun Heo * process'em. 2402e22bee78STejun Heo */ 2403e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2404bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2405e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2406e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2407e22bee78STejun Heo 2408e22bee78STejun Heo process_scheduled_works(rescuer); 24097576958aSTejun Heo 24107576958aSTejun Heo /* 2411d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24127576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24137576958aSTejun Heo * and stalling the execution. 24147576958aSTejun Heo */ 241563d95a91STejun Heo if (keep_working(pool)) 241663d95a91STejun Heo wake_up_worker(pool); 24177576958aSTejun Heo 2418d565ed63STejun Heo spin_unlock_irq(&pool->lock); 24191da177e4SLinus Torvalds } 24201da177e4SLinus Torvalds 2421111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2422111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2423e22bee78STejun Heo schedule(); 2424e22bee78STejun Heo goto repeat; 24251da177e4SLinus Torvalds } 24261da177e4SLinus Torvalds 2427fc2e4d70SOleg Nesterov struct wq_barrier { 2428fc2e4d70SOleg Nesterov struct work_struct work; 2429fc2e4d70SOleg Nesterov struct completion done; 2430fc2e4d70SOleg Nesterov }; 2431fc2e4d70SOleg Nesterov 2432fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2433fc2e4d70SOleg Nesterov { 2434fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2435fc2e4d70SOleg Nesterov complete(&barr->done); 2436fc2e4d70SOleg Nesterov } 2437fc2e4d70SOleg Nesterov 24384690c4abSTejun Heo /** 24394690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 24404690c4abSTejun Heo * @cwq: cwq to insert barrier into 24414690c4abSTejun Heo * @barr: wq_barrier to insert 2442affee4b2STejun Heo * @target: target work to attach @barr to 2443affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24444690c4abSTejun Heo * 2445affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2446affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2447affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2448affee4b2STejun Heo * cpu. 2449affee4b2STejun Heo * 2450affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2451affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2452affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2453affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2454affee4b2STejun Heo * after a work with LINKED flag set. 2455affee4b2STejun Heo * 2456affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2457affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 24584690c4abSTejun Heo * 24594690c4abSTejun Heo * CONTEXT: 2460d565ed63STejun Heo * spin_lock_irq(pool->lock). 24614690c4abSTejun Heo */ 246283c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2463affee4b2STejun Heo struct wq_barrier *barr, 2464affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2465fc2e4d70SOleg Nesterov { 2466affee4b2STejun Heo struct list_head *head; 2467affee4b2STejun Heo unsigned int linked = 0; 2468affee4b2STejun Heo 2469dc186ad7SThomas Gleixner /* 2470d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2471dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2472dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2473dc186ad7SThomas Gleixner * might deadlock. 2474dc186ad7SThomas Gleixner */ 2475ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 247622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2477fc2e4d70SOleg Nesterov init_completion(&barr->done); 247883c22520SOleg Nesterov 2479affee4b2STejun Heo /* 2480affee4b2STejun Heo * If @target is currently being executed, schedule the 2481affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2482affee4b2STejun Heo */ 2483affee4b2STejun Heo if (worker) 2484affee4b2STejun Heo head = worker->scheduled.next; 2485affee4b2STejun Heo else { 2486affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2487affee4b2STejun Heo 2488affee4b2STejun Heo head = target->entry.next; 2489affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2490affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2491affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2492affee4b2STejun Heo } 2493affee4b2STejun Heo 2494dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2495affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2496affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2497fc2e4d70SOleg Nesterov } 2498fc2e4d70SOleg Nesterov 249973f53c4aSTejun Heo /** 250073f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 250173f53c4aSTejun Heo * @wq: workqueue being flushed 250273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 250373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 250473f53c4aSTejun Heo * 250573f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 250673f53c4aSTejun Heo * 250773f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 250873f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 250973f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 251073f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 251173f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 251273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 251373f53c4aSTejun Heo * 251473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 251573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 251673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 251773f53c4aSTejun Heo * is returned. 251873f53c4aSTejun Heo * 251973f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 252073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 252173f53c4aSTejun Heo * advanced to @work_color. 252273f53c4aSTejun Heo * 252373f53c4aSTejun Heo * CONTEXT: 252473f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 252573f53c4aSTejun Heo * 252673f53c4aSTejun Heo * RETURNS: 252773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 252873f53c4aSTejun Heo * otherwise. 252973f53c4aSTejun Heo */ 253073f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 253173f53c4aSTejun Heo int flush_color, int work_color) 25321da177e4SLinus Torvalds { 253373f53c4aSTejun Heo bool wait = false; 253473f53c4aSTejun Heo unsigned int cpu; 25351da177e4SLinus Torvalds 253673f53c4aSTejun Heo if (flush_color >= 0) { 253773f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 253873f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2539dc186ad7SThomas Gleixner } 254014441960SOleg Nesterov 2541f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 254273f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2543d565ed63STejun Heo struct worker_pool *pool = cwq->pool; 25441da177e4SLinus Torvalds 2545d565ed63STejun Heo spin_lock_irq(&pool->lock); 254673f53c4aSTejun Heo 254773f53c4aSTejun Heo if (flush_color >= 0) { 254873f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 254973f53c4aSTejun Heo 255073f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 255173f53c4aSTejun Heo cwq->flush_color = flush_color; 255273f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 255373f53c4aSTejun Heo wait = true; 25541da177e4SLinus Torvalds } 255573f53c4aSTejun Heo } 255673f53c4aSTejun Heo 255773f53c4aSTejun Heo if (work_color >= 0) { 255873f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 255973f53c4aSTejun Heo cwq->work_color = work_color; 256073f53c4aSTejun Heo } 256173f53c4aSTejun Heo 2562d565ed63STejun Heo spin_unlock_irq(&pool->lock); 25631da177e4SLinus Torvalds } 25641da177e4SLinus Torvalds 256573f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 256673f53c4aSTejun Heo complete(&wq->first_flusher->done); 256773f53c4aSTejun Heo 256873f53c4aSTejun Heo return wait; 256983c22520SOleg Nesterov } 25701da177e4SLinus Torvalds 25710fcb78c2SRolf Eike Beer /** 25721da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25730fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25741da177e4SLinus Torvalds * 25751da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 25761da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 25771da177e4SLinus Torvalds * 2578fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2579fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 25801da177e4SLinus Torvalds */ 25817ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25821da177e4SLinus Torvalds { 258373f53c4aSTejun Heo struct wq_flusher this_flusher = { 258473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 258573f53c4aSTejun Heo .flush_color = -1, 258673f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 258773f53c4aSTejun Heo }; 258873f53c4aSTejun Heo int next_color; 2589b1f4ec17SOleg Nesterov 25903295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25913295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 259273f53c4aSTejun Heo 259373f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 259473f53c4aSTejun Heo 259573f53c4aSTejun Heo /* 259673f53c4aSTejun Heo * Start-to-wait phase 259773f53c4aSTejun Heo */ 259873f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 259973f53c4aSTejun Heo 260073f53c4aSTejun Heo if (next_color != wq->flush_color) { 260173f53c4aSTejun Heo /* 260273f53c4aSTejun Heo * Color space is not full. The current work_color 260373f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 260473f53c4aSTejun Heo * by one. 260573f53c4aSTejun Heo */ 260673f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 260773f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 260873f53c4aSTejun Heo wq->work_color = next_color; 260973f53c4aSTejun Heo 261073f53c4aSTejun Heo if (!wq->first_flusher) { 261173f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 261273f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 261373f53c4aSTejun Heo 261473f53c4aSTejun Heo wq->first_flusher = &this_flusher; 261573f53c4aSTejun Heo 261673f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 261773f53c4aSTejun Heo wq->work_color)) { 261873f53c4aSTejun Heo /* nothing to flush, done */ 261973f53c4aSTejun Heo wq->flush_color = next_color; 262073f53c4aSTejun Heo wq->first_flusher = NULL; 262173f53c4aSTejun Heo goto out_unlock; 262273f53c4aSTejun Heo } 262373f53c4aSTejun Heo } else { 262473f53c4aSTejun Heo /* wait in queue */ 262573f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 262673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 262773f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 262873f53c4aSTejun Heo } 262973f53c4aSTejun Heo } else { 263073f53c4aSTejun Heo /* 263173f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 263273f53c4aSTejun Heo * The next flush completion will assign us 263373f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 263473f53c4aSTejun Heo */ 263573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 263673f53c4aSTejun Heo } 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 264173f53c4aSTejun Heo 264273f53c4aSTejun Heo /* 264373f53c4aSTejun Heo * Wake-up-and-cascade phase 264473f53c4aSTejun Heo * 264573f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 264673f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 264773f53c4aSTejun Heo */ 264873f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 264973f53c4aSTejun Heo return; 265073f53c4aSTejun Heo 265173f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 265273f53c4aSTejun Heo 26534ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26544ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26554ce48b37STejun Heo goto out_unlock; 26564ce48b37STejun Heo 265773f53c4aSTejun Heo wq->first_flusher = NULL; 265873f53c4aSTejun Heo 265973f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 266073f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 266173f53c4aSTejun Heo 266273f53c4aSTejun Heo while (true) { 266373f53c4aSTejun Heo struct wq_flusher *next, *tmp; 266473f53c4aSTejun Heo 266573f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 266673f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 266773f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 266873f53c4aSTejun Heo break; 266973f53c4aSTejun Heo list_del_init(&next->list); 267073f53c4aSTejun Heo complete(&next->done); 267173f53c4aSTejun Heo } 267273f53c4aSTejun Heo 267373f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 267473f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 267573f53c4aSTejun Heo 267673f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 267773f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 267873f53c4aSTejun Heo 267973f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 268073f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 268173f53c4aSTejun Heo /* 268273f53c4aSTejun Heo * Assign the same color to all overflowed 268373f53c4aSTejun Heo * flushers, advance work_color and append to 268473f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 268573f53c4aSTejun Heo * phase for these overflowed flushers. 268673f53c4aSTejun Heo */ 268773f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 268873f53c4aSTejun Heo tmp->flush_color = wq->work_color; 268973f53c4aSTejun Heo 269073f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 269173f53c4aSTejun Heo 269273f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 269373f53c4aSTejun Heo &wq->flusher_queue); 269473f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 269573f53c4aSTejun Heo } 269673f53c4aSTejun Heo 269773f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 269873f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 269973f53c4aSTejun Heo break; 270073f53c4aSTejun Heo } 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo /* 270373f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 270473f53c4aSTejun Heo * the new first flusher and arm cwqs. 270573f53c4aSTejun Heo */ 270673f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 270773f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 270873f53c4aSTejun Heo 270973f53c4aSTejun Heo list_del_init(&next->list); 271073f53c4aSTejun Heo wq->first_flusher = next; 271173f53c4aSTejun Heo 271273f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 271373f53c4aSTejun Heo break; 271473f53c4aSTejun Heo 271573f53c4aSTejun Heo /* 271673f53c4aSTejun Heo * Meh... this color is already done, clear first 271773f53c4aSTejun Heo * flusher and repeat cascading. 271873f53c4aSTejun Heo */ 271973f53c4aSTejun Heo wq->first_flusher = NULL; 272073f53c4aSTejun Heo } 272173f53c4aSTejun Heo 272273f53c4aSTejun Heo out_unlock: 272373f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27241da177e4SLinus Torvalds } 2725ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27261da177e4SLinus Torvalds 27279c5a2ba7STejun Heo /** 27289c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27299c5a2ba7STejun Heo * @wq: workqueue to drain 27309c5a2ba7STejun Heo * 27319c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27329c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27339c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27349c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27359c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27369c5a2ba7STejun Heo * takes too long. 27379c5a2ba7STejun Heo */ 27389c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27399c5a2ba7STejun Heo { 27409c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27419c5a2ba7STejun Heo unsigned int cpu; 27429c5a2ba7STejun Heo 27439c5a2ba7STejun Heo /* 27449c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27459c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27469c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27479c5a2ba7STejun Heo */ 27489c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27499c5a2ba7STejun Heo if (!wq->nr_drainers++) 27509c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 27519c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27529c5a2ba7STejun Heo reflush: 27539c5a2ba7STejun Heo flush_workqueue(wq); 27549c5a2ba7STejun Heo 27559c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 27569c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2757fa2563e4SThomas Tuttle bool drained; 27589c5a2ba7STejun Heo 2759d565ed63STejun Heo spin_lock_irq(&cwq->pool->lock); 2760fa2563e4SThomas Tuttle drained = !cwq->nr_active && list_empty(&cwq->delayed_works); 2761d565ed63STejun Heo spin_unlock_irq(&cwq->pool->lock); 2762fa2563e4SThomas Tuttle 2763fa2563e4SThomas Tuttle if (drained) 27649c5a2ba7STejun Heo continue; 27659c5a2ba7STejun Heo 27669c5a2ba7STejun Heo if (++flush_cnt == 10 || 27679c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2768044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 27699c5a2ba7STejun Heo wq->name, flush_cnt); 27709c5a2ba7STejun Heo goto reflush; 27719c5a2ba7STejun Heo } 27729c5a2ba7STejun Heo 27739c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27749c5a2ba7STejun Heo if (!--wq->nr_drainers) 27759c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 27769c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27779c5a2ba7STejun Heo } 27789c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27799c5a2ba7STejun Heo 2780606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2781baf59022STejun Heo { 2782baf59022STejun Heo struct worker *worker = NULL; 2783c9e7cf27STejun Heo struct worker_pool *pool; 2784baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2785baf59022STejun Heo 2786baf59022STejun Heo might_sleep(); 2787c9e7cf27STejun Heo pool = get_work_pool(work); 2788c9e7cf27STejun Heo if (!pool) 2789baf59022STejun Heo return false; 2790baf59022STejun Heo 2791d565ed63STejun Heo spin_lock_irq(&pool->lock); 2792baf59022STejun Heo if (!list_empty(&work->entry)) { 2793baf59022STejun Heo /* 2794baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2795d565ed63STejun Heo * If it was re-queued to a different pool under us, we 2796baf59022STejun Heo * are not going to wait. 2797baf59022STejun Heo */ 2798baf59022STejun Heo smp_rmb(); 2799baf59022STejun Heo cwq = get_work_cwq(work); 2800d565ed63STejun Heo if (unlikely(!cwq || pool != cwq->pool)) 2801baf59022STejun Heo goto already_gone; 2802606a5020STejun Heo } else { 2803c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2804baf59022STejun Heo if (!worker) 2805baf59022STejun Heo goto already_gone; 2806baf59022STejun Heo cwq = worker->current_cwq; 2807606a5020STejun Heo } 2808baf59022STejun Heo 2809baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2810d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2811baf59022STejun Heo 2812e159489bSTejun Heo /* 2813e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2814e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2815e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2816e159489bSTejun Heo * access. 2817e159489bSTejun Heo */ 2818e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2819baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2820e159489bSTejun Heo else 2821e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2822baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2823e159489bSTejun Heo 2824baf59022STejun Heo return true; 2825baf59022STejun Heo already_gone: 2826d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2827baf59022STejun Heo return false; 2828baf59022STejun Heo } 2829baf59022STejun Heo 2830db700897SOleg Nesterov /** 2831401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2832401a8d04STejun Heo * @work: the work to flush 2833db700897SOleg Nesterov * 2834606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2835606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2836401a8d04STejun Heo * 2837401a8d04STejun Heo * RETURNS: 2838401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2839401a8d04STejun Heo * %false if it was already idle. 2840db700897SOleg Nesterov */ 2841401a8d04STejun Heo bool flush_work(struct work_struct *work) 2842db700897SOleg Nesterov { 2843db700897SOleg Nesterov struct wq_barrier barr; 2844db700897SOleg Nesterov 28450976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28460976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28470976dfc1SStephen Boyd 2848606a5020STejun Heo if (start_flush_work(work, &barr)) { 2849db700897SOleg Nesterov wait_for_completion(&barr.done); 2850dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2851401a8d04STejun Heo return true; 2852606a5020STejun Heo } else { 2853401a8d04STejun Heo return false; 2854db700897SOleg Nesterov } 2855606a5020STejun Heo } 2856db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2857db700897SOleg Nesterov 285836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2859401a8d04STejun Heo { 2860bbb68dfaSTejun Heo unsigned long flags; 28611f1f642eSOleg Nesterov int ret; 28621f1f642eSOleg Nesterov 28631f1f642eSOleg Nesterov do { 2864bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2865bbb68dfaSTejun Heo /* 2866bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2867bbb68dfaSTejun Heo * would be waiting for before retrying. 2868bbb68dfaSTejun Heo */ 2869bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2870606a5020STejun Heo flush_work(work); 28711f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28721f1f642eSOleg Nesterov 2873bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2874bbb68dfaSTejun Heo mark_work_canceling(work); 2875bbb68dfaSTejun Heo local_irq_restore(flags); 2876bbb68dfaSTejun Heo 2877606a5020STejun Heo flush_work(work); 28787a22ad75STejun Heo clear_work_data(work); 28791f1f642eSOleg Nesterov return ret; 28801f1f642eSOleg Nesterov } 28811f1f642eSOleg Nesterov 28826e84d644SOleg Nesterov /** 2883401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2884401a8d04STejun Heo * @work: the work to cancel 28856e84d644SOleg Nesterov * 2886401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2887401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2888401a8d04STejun Heo * another workqueue. On return from this function, @work is 2889401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28901f1f642eSOleg Nesterov * 2891401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2892401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28936e84d644SOleg Nesterov * 2894401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28956e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2896401a8d04STejun Heo * 2897401a8d04STejun Heo * RETURNS: 2898401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28996e84d644SOleg Nesterov */ 2900401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29016e84d644SOleg Nesterov { 290236e227d2STejun Heo return __cancel_work_timer(work, false); 2903b89deed3SOleg Nesterov } 290428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2905b89deed3SOleg Nesterov 29066e84d644SOleg Nesterov /** 2907401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2908401a8d04STejun Heo * @dwork: the delayed work to flush 29096e84d644SOleg Nesterov * 2910401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2911401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2912401a8d04STejun Heo * considers the last queueing instance of @dwork. 29131f1f642eSOleg Nesterov * 2914401a8d04STejun Heo * RETURNS: 2915401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2916401a8d04STejun Heo * %false if it was already idle. 29176e84d644SOleg Nesterov */ 2918401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2919401a8d04STejun Heo { 29208930cabaSTejun Heo local_irq_disable(); 2921401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 292260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29238930cabaSTejun Heo local_irq_enable(); 2924401a8d04STejun Heo return flush_work(&dwork->work); 2925401a8d04STejun Heo } 2926401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2927401a8d04STejun Heo 2928401a8d04STejun Heo /** 292957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 293057b30ae7STejun Heo * @dwork: delayed_work to cancel 293109383498STejun Heo * 293257b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 293357b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 293457b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 293557b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 293657b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 293709383498STejun Heo * 293857b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 293909383498STejun Heo */ 294057b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 294109383498STejun Heo { 294257b30ae7STejun Heo unsigned long flags; 294357b30ae7STejun Heo int ret; 294457b30ae7STejun Heo 294557b30ae7STejun Heo do { 294657b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 294757b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 294857b30ae7STejun Heo 294957b30ae7STejun Heo if (unlikely(ret < 0)) 295057b30ae7STejun Heo return false; 295157b30ae7STejun Heo 29527c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29537c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 295457b30ae7STejun Heo local_irq_restore(flags); 2955c0158ca6SDan Magenheimer return ret; 295609383498STejun Heo } 295757b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 295809383498STejun Heo 295909383498STejun Heo /** 2960401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2961401a8d04STejun Heo * @dwork: the delayed work cancel 2962401a8d04STejun Heo * 2963401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2964401a8d04STejun Heo * 2965401a8d04STejun Heo * RETURNS: 2966401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2967401a8d04STejun Heo */ 2968401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29696e84d644SOleg Nesterov { 297036e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29716e84d644SOleg Nesterov } 2972f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29731da177e4SLinus Torvalds 29740fcb78c2SRolf Eike Beer /** 2975c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2976c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2977c1a220e7SZhang Rui * @work: job to be done 2978c1a220e7SZhang Rui * 2979c1a220e7SZhang Rui * This puts a job on a specific cpu 2980c1a220e7SZhang Rui */ 2981d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 2982c1a220e7SZhang Rui { 2983d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2984c1a220e7SZhang Rui } 2985c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2986c1a220e7SZhang Rui 29870fcb78c2SRolf Eike Beer /** 2988ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 29891da177e4SLinus Torvalds * @work: job to be done 29901da177e4SLinus Torvalds * 2991d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 2992d4283e93STejun Heo * %true otherwise. 299352bad64dSDavid Howells * 29940fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 29950fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 29960fcb78c2SRolf Eike Beer * workqueue otherwise. 29971da177e4SLinus Torvalds */ 2998d4283e93STejun Heo bool schedule_work(struct work_struct *work) 29991da177e4SLinus Torvalds { 30000fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 30011da177e4SLinus Torvalds } 30020fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 30031da177e4SLinus Torvalds 30040fcb78c2SRolf Eike Beer /** 30050fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 30060fcb78c2SRolf Eike Beer * @cpu: cpu to use 30070fcb78c2SRolf Eike Beer * @dwork: job to be done 30080fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 30090fcb78c2SRolf Eike Beer * 30100fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 30110fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 30120fcb78c2SRolf Eike Beer */ 3013d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 3014d4283e93STejun Heo unsigned long delay) 30151da177e4SLinus Torvalds { 3016d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 30171da177e4SLinus Torvalds } 3018ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 30191da177e4SLinus Torvalds 3020b6136773SAndrew Morton /** 30210a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 30220a13c00eSTejun Heo * @dwork: job to be done 30230a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 30240a13c00eSTejun Heo * 30250a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 30260a13c00eSTejun Heo * workqueue. 30270a13c00eSTejun Heo */ 3028d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 30290a13c00eSTejun Heo { 30300a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30310a13c00eSTejun Heo } 30320a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30330a13c00eSTejun Heo 30340a13c00eSTejun Heo /** 303531ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3036b6136773SAndrew Morton * @func: the function to call 3037b6136773SAndrew Morton * 303831ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 303931ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3040b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 304131ddd871STejun Heo * 304231ddd871STejun Heo * RETURNS: 304331ddd871STejun Heo * 0 on success, -errno on failure. 3044b6136773SAndrew Morton */ 304565f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 304615316ba8SChristoph Lameter { 304715316ba8SChristoph Lameter int cpu; 304838f51568SNamhyung Kim struct work_struct __percpu *works; 304915316ba8SChristoph Lameter 3050b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3051b6136773SAndrew Morton if (!works) 305215316ba8SChristoph Lameter return -ENOMEM; 3053b6136773SAndrew Morton 305495402b38SGautham R Shenoy get_online_cpus(); 305593981800STejun Heo 305615316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30579bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30589bfb1839SIngo Molnar 30599bfb1839SIngo Molnar INIT_WORK(work, func); 30608de6d308SOleg Nesterov schedule_work_on(cpu, work); 306115316ba8SChristoph Lameter } 306293981800STejun Heo 306393981800STejun Heo for_each_online_cpu(cpu) 30648616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 306593981800STejun Heo 306695402b38SGautham R Shenoy put_online_cpus(); 3067b6136773SAndrew Morton free_percpu(works); 306815316ba8SChristoph Lameter return 0; 306915316ba8SChristoph Lameter } 307015316ba8SChristoph Lameter 3071eef6a7d5SAlan Stern /** 3072eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3073eef6a7d5SAlan Stern * 3074eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3075eef6a7d5SAlan Stern * completion. 3076eef6a7d5SAlan Stern * 3077eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3078eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3079eef6a7d5SAlan Stern * will lead to deadlock: 3080eef6a7d5SAlan Stern * 3081eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3082eef6a7d5SAlan Stern * a lock held by your code or its caller. 3083eef6a7d5SAlan Stern * 3084eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3085eef6a7d5SAlan Stern * 3086eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3087eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3088eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3089eef6a7d5SAlan Stern * 3090eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3091eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3092eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3093eef6a7d5SAlan Stern * cancel_work_sync() instead. 3094eef6a7d5SAlan Stern */ 30951da177e4SLinus Torvalds void flush_scheduled_work(void) 30961da177e4SLinus Torvalds { 3097d320c038STejun Heo flush_workqueue(system_wq); 30981da177e4SLinus Torvalds } 3099ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 31001da177e4SLinus Torvalds 31011da177e4SLinus Torvalds /** 31021fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31031fa44ecaSJames Bottomley * @fn: the function to execute 31041fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31051fa44ecaSJames Bottomley * be available when the work executes) 31061fa44ecaSJames Bottomley * 31071fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31081fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31091fa44ecaSJames Bottomley * 31101fa44ecaSJames Bottomley * Returns: 0 - function was executed 31111fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31121fa44ecaSJames Bottomley */ 311365f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31141fa44ecaSJames Bottomley { 31151fa44ecaSJames Bottomley if (!in_interrupt()) { 311665f27f38SDavid Howells fn(&ew->work); 31171fa44ecaSJames Bottomley return 0; 31181fa44ecaSJames Bottomley } 31191fa44ecaSJames Bottomley 312065f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31211fa44ecaSJames Bottomley schedule_work(&ew->work); 31221fa44ecaSJames Bottomley 31231fa44ecaSJames Bottomley return 1; 31241fa44ecaSJames Bottomley } 31251fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31261fa44ecaSJames Bottomley 31271da177e4SLinus Torvalds int keventd_up(void) 31281da177e4SLinus Torvalds { 3129d320c038STejun Heo return system_wq != NULL; 31301da177e4SLinus Torvalds } 31311da177e4SLinus Torvalds 3132bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 31331da177e4SLinus Torvalds { 31343af24433SOleg Nesterov /* 31350f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 31360f900049STejun Heo * Make sure that the alignment isn't lower than that of 31370f900049STejun Heo * unsigned long long. 31383af24433SOleg Nesterov */ 31390f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 31400f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 31410f900049STejun Heo __alignof__(unsigned long long)); 31423af24433SOleg Nesterov 3143e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3144f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 3145931ac77eSTejun Heo else { 31460f900049STejun Heo void *ptr; 3147e1d8aa9fSFrederic Weisbecker 31480f900049STejun Heo /* 3149f3421797STejun Heo * Allocate enough room to align cwq and put an extra 3150f3421797STejun Heo * pointer at the end pointing back to the originally 3151f3421797STejun Heo * allocated pointer which will be used for free. 31520f900049STejun Heo */ 3153bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3154bdbc5dd7STejun Heo if (ptr) { 3155bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 3156bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 3157bdbc5dd7STejun Heo } 31583af24433SOleg Nesterov } 31593af24433SOleg Nesterov 31600415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3161bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 3162bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 31630f900049STejun Heo } 31640f900049STejun Heo 3165bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 316606ba38a9SOleg Nesterov { 3167e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3168bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 3169f3421797STejun Heo else if (wq->cpu_wq.single) { 3170f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 3171f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 317206ba38a9SOleg Nesterov } 317306ba38a9SOleg Nesterov } 317406ba38a9SOleg Nesterov 3175f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3176f3421797STejun Heo const char *name) 3177b71ab8c2STejun Heo { 3178f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3179f3421797STejun Heo 3180f3421797STejun Heo if (max_active < 1 || max_active > lim) 3181044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3182f3421797STejun Heo max_active, name, 1, lim); 3183b71ab8c2STejun Heo 3184f3421797STejun Heo return clamp_val(max_active, 1, lim); 3185b71ab8c2STejun Heo } 3186b71ab8c2STejun Heo 3187b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 318897e37d7bSTejun Heo unsigned int flags, 31891e19ffc6STejun Heo int max_active, 3190eb13ba87SJohannes Berg struct lock_class_key *key, 3191b196be89STejun Heo const char *lock_name, ...) 31923af24433SOleg Nesterov { 3193b196be89STejun Heo va_list args, args1; 31943af24433SOleg Nesterov struct workqueue_struct *wq; 3195c34056a3STejun Heo unsigned int cpu; 3196b196be89STejun Heo size_t namelen; 3197b196be89STejun Heo 3198b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3199b196be89STejun Heo va_start(args, lock_name); 3200b196be89STejun Heo va_copy(args1, args); 3201b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3202b196be89STejun Heo 3203b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3204b196be89STejun Heo if (!wq) 3205b196be89STejun Heo goto err; 3206b196be89STejun Heo 3207b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3208b196be89STejun Heo va_end(args); 3209b196be89STejun Heo va_end(args1); 32103af24433SOleg Nesterov 3211f3421797STejun Heo /* 32126370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 32136370a6adSTejun Heo * have a rescuer to guarantee forward progress. 32146370a6adSTejun Heo */ 32156370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 32166370a6adSTejun Heo flags |= WQ_RESCUER; 32176370a6adSTejun Heo 3218d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3219b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 32203af24433SOleg Nesterov 3221b196be89STejun Heo /* init wq */ 322297e37d7bSTejun Heo wq->flags = flags; 3223a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 322473f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 322573f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 322673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 322773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 32283af24433SOleg Nesterov 3229eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3230cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 32313af24433SOleg Nesterov 3232bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 3233bdbc5dd7STejun Heo goto err; 3234bdbc5dd7STejun Heo 3235f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 32361537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 32371537663fSTejun Heo 32380f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 3239a60dc39cSTejun Heo cwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI); 3240c34056a3STejun Heo cwq->wq = wq; 324173f53c4aSTejun Heo cwq->flush_color = -1; 32421e19ffc6STejun Heo cwq->max_active = max_active; 32431e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3244e22bee78STejun Heo } 32451537663fSTejun Heo 3246e22bee78STejun Heo if (flags & WQ_RESCUER) { 3247e22bee78STejun Heo struct worker *rescuer; 3248e22bee78STejun Heo 3249f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3250e22bee78STejun Heo goto err; 3251e22bee78STejun Heo 3252e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3253e22bee78STejun Heo if (!rescuer) 3254e22bee78STejun Heo goto err; 3255e22bee78STejun Heo 3256111c225aSTejun Heo rescuer->rescue_wq = wq; 3257111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3258b196be89STejun Heo wq->name); 3259e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3260e22bee78STejun Heo goto err; 3261e22bee78STejun Heo 3262e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3263e22bee78STejun Heo wake_up_process(rescuer->task); 32643af24433SOleg Nesterov } 32651537663fSTejun Heo 32663af24433SOleg Nesterov /* 3267a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3268a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3269a0a1a5fdSTejun Heo * workqueue to workqueues list. 32703af24433SOleg Nesterov */ 32713af24433SOleg Nesterov spin_lock(&workqueue_lock); 3272a0a1a5fdSTejun Heo 327358a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3274f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3275a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3276a0a1a5fdSTejun Heo 32773af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3278a0a1a5fdSTejun Heo 32793af24433SOleg Nesterov spin_unlock(&workqueue_lock); 32803af24433SOleg Nesterov 32813af24433SOleg Nesterov return wq; 32824690c4abSTejun Heo err: 32834690c4abSTejun Heo if (wq) { 3284bdbc5dd7STejun Heo free_cwqs(wq); 3285f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3286e22bee78STejun Heo kfree(wq->rescuer); 32874690c4abSTejun Heo kfree(wq); 32883af24433SOleg Nesterov } 32894690c4abSTejun Heo return NULL; 32901da177e4SLinus Torvalds } 3291d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 32921da177e4SLinus Torvalds 32933af24433SOleg Nesterov /** 32943af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 32953af24433SOleg Nesterov * @wq: target workqueue 32963af24433SOleg Nesterov * 32973af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 32983af24433SOleg Nesterov */ 32993af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 33003af24433SOleg Nesterov { 3301c8e55f36STejun Heo unsigned int cpu; 33023af24433SOleg Nesterov 33039c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 33049c5a2ba7STejun Heo drain_workqueue(wq); 3305c8efcc25STejun Heo 3306a0a1a5fdSTejun Heo /* 3307a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3308a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3309a0a1a5fdSTejun Heo */ 331095402b38SGautham R Shenoy spin_lock(&workqueue_lock); 33113af24433SOleg Nesterov list_del(&wq->list); 331295402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 33133af24433SOleg Nesterov 3314e22bee78STejun Heo /* sanity check */ 3315f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 331673f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 331773f53c4aSTejun Heo int i; 33183af24433SOleg Nesterov 331973f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 332073f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 33211e19ffc6STejun Heo BUG_ON(cwq->nr_active); 33221e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 332373f53c4aSTejun Heo } 33241537663fSTejun Heo 3325e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3326e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3327f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 33288d9df9f0SXiaotian Feng kfree(wq->rescuer); 3329e22bee78STejun Heo } 3330e22bee78STejun Heo 3331bdbc5dd7STejun Heo free_cwqs(wq); 33323af24433SOleg Nesterov kfree(wq); 33333af24433SOleg Nesterov } 33343af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33353af24433SOleg Nesterov 3336dcd989cbSTejun Heo /** 33379f4bd4cdSLai Jiangshan * cwq_set_max_active - adjust max_active of a cwq 33389f4bd4cdSLai Jiangshan * @cwq: target cpu_workqueue_struct 33399f4bd4cdSLai Jiangshan * @max_active: new max_active value. 33409f4bd4cdSLai Jiangshan * 33419f4bd4cdSLai Jiangshan * Set @cwq->max_active to @max_active and activate delayed works if 33429f4bd4cdSLai Jiangshan * increased. 33439f4bd4cdSLai Jiangshan * 33449f4bd4cdSLai Jiangshan * CONTEXT: 3345d565ed63STejun Heo * spin_lock_irq(pool->lock). 33469f4bd4cdSLai Jiangshan */ 33479f4bd4cdSLai Jiangshan static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active) 33489f4bd4cdSLai Jiangshan { 33499f4bd4cdSLai Jiangshan cwq->max_active = max_active; 33509f4bd4cdSLai Jiangshan 33519f4bd4cdSLai Jiangshan while (!list_empty(&cwq->delayed_works) && 33529f4bd4cdSLai Jiangshan cwq->nr_active < cwq->max_active) 33539f4bd4cdSLai Jiangshan cwq_activate_first_delayed(cwq); 33549f4bd4cdSLai Jiangshan } 33559f4bd4cdSLai Jiangshan 33569f4bd4cdSLai Jiangshan /** 3357dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3358dcd989cbSTejun Heo * @wq: target workqueue 3359dcd989cbSTejun Heo * @max_active: new max_active value. 3360dcd989cbSTejun Heo * 3361dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3362dcd989cbSTejun Heo * 3363dcd989cbSTejun Heo * CONTEXT: 3364dcd989cbSTejun Heo * Don't call from IRQ context. 3365dcd989cbSTejun Heo */ 3366dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3367dcd989cbSTejun Heo { 3368dcd989cbSTejun Heo unsigned int cpu; 3369dcd989cbSTejun Heo 3370f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3371dcd989cbSTejun Heo 3372dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3373dcd989cbSTejun Heo 3374dcd989cbSTejun Heo wq->saved_max_active = max_active; 3375dcd989cbSTejun Heo 3376f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 337735b6bb63STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 337835b6bb63STejun Heo struct worker_pool *pool = cwq->pool; 3379dcd989cbSTejun Heo 3380d565ed63STejun Heo spin_lock_irq(&pool->lock); 3381dcd989cbSTejun Heo 338258a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 338335b6bb63STejun Heo !(pool->flags & POOL_FREEZING)) 338435b6bb63STejun Heo cwq_set_max_active(cwq, max_active); 3385dcd989cbSTejun Heo 3386d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3387dcd989cbSTejun Heo } 3388dcd989cbSTejun Heo 3389dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3390dcd989cbSTejun Heo } 3391dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3392dcd989cbSTejun Heo 3393dcd989cbSTejun Heo /** 3394dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3395dcd989cbSTejun Heo * @cpu: CPU in question 3396dcd989cbSTejun Heo * @wq: target workqueue 3397dcd989cbSTejun Heo * 3398dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3399dcd989cbSTejun Heo * no synchronization around this function and the test result is 3400dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3401dcd989cbSTejun Heo * 3402dcd989cbSTejun Heo * RETURNS: 3403dcd989cbSTejun Heo * %true if congested, %false otherwise. 3404dcd989cbSTejun Heo */ 3405dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3406dcd989cbSTejun Heo { 3407dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3408dcd989cbSTejun Heo 3409dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3410dcd989cbSTejun Heo } 3411dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3412dcd989cbSTejun Heo 3413dcd989cbSTejun Heo /** 3414dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3415dcd989cbSTejun Heo * @work: the work to be tested 3416dcd989cbSTejun Heo * 3417dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3418dcd989cbSTejun Heo * synchronization around this function and the test result is 3419dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3420dcd989cbSTejun Heo * 3421dcd989cbSTejun Heo * RETURNS: 3422dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3423dcd989cbSTejun Heo */ 3424dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3425dcd989cbSTejun Heo { 3426c9e7cf27STejun Heo struct worker_pool *pool = get_work_pool(work); 3427dcd989cbSTejun Heo unsigned long flags; 3428dcd989cbSTejun Heo unsigned int ret = 0; 3429dcd989cbSTejun Heo 3430dcd989cbSTejun Heo if (work_pending(work)) 3431dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3432038366c5SLai Jiangshan 3433038366c5SLai Jiangshan if (pool) { 3434038366c5SLai Jiangshan spin_lock_irqsave(&pool->lock, flags); 3435c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3436dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3437d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 3438038366c5SLai Jiangshan } 3439dcd989cbSTejun Heo 3440dcd989cbSTejun Heo return ret; 3441dcd989cbSTejun Heo } 3442dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3443dcd989cbSTejun Heo 3444db7bccf4STejun Heo /* 3445db7bccf4STejun Heo * CPU hotplug. 3446db7bccf4STejun Heo * 3447e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3448e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3449706026c2STejun Heo * pool which make migrating pending and scheduled works very 3450e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 345194cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 3452e22bee78STejun Heo * blocked draining impractical. 3453e22bee78STejun Heo * 345424647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 3455628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3456628c78e7STejun Heo * cpu comes back online. 3457db7bccf4STejun Heo */ 3458db7bccf4STejun Heo 3459706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 3460db7bccf4STejun Heo { 346138db41d9STejun Heo int cpu = smp_processor_id(); 34624ce62e9eSTejun Heo struct worker_pool *pool; 3463db7bccf4STejun Heo struct worker *worker; 3464db7bccf4STejun Heo struct hlist_node *pos; 3465db7bccf4STejun Heo int i; 3466db7bccf4STejun Heo 346738db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 346838db41d9STejun Heo BUG_ON(cpu != smp_processor_id()); 3469db7bccf4STejun Heo 347094cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 347194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 3472e22bee78STejun Heo 3473f2d5a0eeSTejun Heo /* 347494cf58bbSTejun Heo * We've claimed all manager positions. Make all workers 347594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 347694cf58bbSTejun Heo * except for the ones which are still executing works from 347794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 347894cf58bbSTejun Heo * this, they may become diasporas. 3479f2d5a0eeSTejun Heo */ 34804ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3481403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3482db7bccf4STejun Heo 3483c9e7cf27STejun Heo for_each_busy_worker(worker, i, pos, pool) 3484403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3485db7bccf4STejun Heo 348624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 3487f2d5a0eeSTejun Heo 348894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 348994cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 349094cf58bbSTejun Heo } 3491e22bee78STejun Heo 3492e22bee78STejun Heo /* 3493628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3494628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3495628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3496628c78e7STejun Heo */ 3497628c78e7STejun Heo schedule(); 3498628c78e7STejun Heo 3499628c78e7STejun Heo /* 3500628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3501628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 350238db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 350338db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 350438db41d9STejun Heo * pools which are served by workers tied to the CPU. 3505628c78e7STejun Heo * 3506628c78e7STejun Heo * On return from this function, the current worker would trigger 3507628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3508628c78e7STejun Heo * didn't already. 3509e22bee78STejun Heo */ 351038db41d9STejun Heo for_each_std_worker_pool(pool, cpu) 35114ce62e9eSTejun Heo atomic_set(get_pool_nr_running(pool), 0); 3512db7bccf4STejun Heo } 3513db7bccf4STejun Heo 35148db25e78STejun Heo /* 35158db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 35168db25e78STejun Heo * This will be registered high priority CPU notifier. 35178db25e78STejun Heo */ 35189fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 35191da177e4SLinus Torvalds unsigned long action, 35201da177e4SLinus Torvalds void *hcpu) 35211da177e4SLinus Torvalds { 35223af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 35234ce62e9eSTejun Heo struct worker_pool *pool; 35241da177e4SLinus Torvalds 35258db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 35263af24433SOleg Nesterov case CPU_UP_PREPARE: 352738db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 35283ce63377STejun Heo struct worker *worker; 35293ce63377STejun Heo 35303ce63377STejun Heo if (pool->nr_workers) 35313ce63377STejun Heo continue; 35323ce63377STejun Heo 35333ce63377STejun Heo worker = create_worker(pool); 35343ce63377STejun Heo if (!worker) 35353ce63377STejun Heo return NOTIFY_BAD; 35363ce63377STejun Heo 3537d565ed63STejun Heo spin_lock_irq(&pool->lock); 35383ce63377STejun Heo start_worker(worker); 3539d565ed63STejun Heo spin_unlock_irq(&pool->lock); 35403af24433SOleg Nesterov } 35411da177e4SLinus Torvalds break; 35421da177e4SLinus Torvalds 354365758202STejun Heo case CPU_DOWN_FAILED: 354465758202STejun Heo case CPU_ONLINE: 354538db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 354694cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 354794cf58bbSTejun Heo spin_lock_irq(&pool->lock); 354894cf58bbSTejun Heo 354924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 355094cf58bbSTejun Heo rebind_workers(pool); 355194cf58bbSTejun Heo 355294cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 355394cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 355494cf58bbSTejun Heo } 35558db25e78STejun Heo break; 355665758202STejun Heo } 355765758202STejun Heo return NOTIFY_OK; 355865758202STejun Heo } 355965758202STejun Heo 356065758202STejun Heo /* 356165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 356265758202STejun Heo * This will be registered as low priority CPU notifier. 356365758202STejun Heo */ 35649fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 356565758202STejun Heo unsigned long action, 356665758202STejun Heo void *hcpu) 356765758202STejun Heo { 35688db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 35698db25e78STejun Heo struct work_struct unbind_work; 35708db25e78STejun Heo 357165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 357265758202STejun Heo case CPU_DOWN_PREPARE: 35738db25e78STejun Heo /* unbinding should happen on the local CPU */ 3574706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 35757635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 35768db25e78STejun Heo flush_work(&unbind_work); 35778db25e78STejun Heo break; 357865758202STejun Heo } 357965758202STejun Heo return NOTIFY_OK; 358065758202STejun Heo } 358165758202STejun Heo 35822d3854a3SRusty Russell #ifdef CONFIG_SMP 35838ccad40dSRusty Russell 35842d3854a3SRusty Russell struct work_for_cpu { 3585ed48ece2STejun Heo struct work_struct work; 35862d3854a3SRusty Russell long (*fn)(void *); 35872d3854a3SRusty Russell void *arg; 35882d3854a3SRusty Russell long ret; 35892d3854a3SRusty Russell }; 35902d3854a3SRusty Russell 3591ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 35922d3854a3SRusty Russell { 3593ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3594ed48ece2STejun Heo 35952d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 35962d3854a3SRusty Russell } 35972d3854a3SRusty Russell 35982d3854a3SRusty Russell /** 35992d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 36002d3854a3SRusty Russell * @cpu: the cpu to run on 36012d3854a3SRusty Russell * @fn: the function to run 36022d3854a3SRusty Russell * @arg: the function arg 36032d3854a3SRusty Russell * 360431ad9081SRusty Russell * This will return the value @fn returns. 360531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 36066b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 36072d3854a3SRusty Russell */ 36082d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 36092d3854a3SRusty Russell { 3610ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 36112d3854a3SRusty Russell 3612ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3613ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3614ed48ece2STejun Heo flush_work(&wfc.work); 36152d3854a3SRusty Russell return wfc.ret; 36162d3854a3SRusty Russell } 36172d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 36182d3854a3SRusty Russell #endif /* CONFIG_SMP */ 36192d3854a3SRusty Russell 3620a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3621e7577c50SRusty Russell 3622a0a1a5fdSTejun Heo /** 3623a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3624a0a1a5fdSTejun Heo * 362558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 362658a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 3627706026c2STejun Heo * pool->worklist. 3628a0a1a5fdSTejun Heo * 3629a0a1a5fdSTejun Heo * CONTEXT: 3630d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3631a0a1a5fdSTejun Heo */ 3632a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3633a0a1a5fdSTejun Heo { 3634a0a1a5fdSTejun Heo unsigned int cpu; 3635a0a1a5fdSTejun Heo 3636a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3637a0a1a5fdSTejun Heo 3638a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3639a0a1a5fdSTejun Heo workqueue_freezing = true; 3640a0a1a5fdSTejun Heo 3641706026c2STejun Heo for_each_wq_cpu(cpu) { 364235b6bb63STejun Heo struct worker_pool *pool; 3643bdbc5dd7STejun Heo struct workqueue_struct *wq; 36448b03ae3cSTejun Heo 364538db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3646a1056305STejun Heo spin_lock_irq(&pool->lock); 3647d565ed63STejun Heo 364835b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 364935b6bb63STejun Heo pool->flags |= POOL_FREEZING; 3650db7bccf4STejun Heo 3651a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3652a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3653a0a1a5fdSTejun Heo 3654a1056305STejun Heo if (cwq && cwq->pool == pool && 3655a1056305STejun Heo (wq->flags & WQ_FREEZABLE)) 3656a0a1a5fdSTejun Heo cwq->max_active = 0; 36571da177e4SLinus Torvalds } 36588b03ae3cSTejun Heo 3659a1056305STejun Heo spin_unlock_irq(&pool->lock); 3660a1056305STejun Heo } 3661a0a1a5fdSTejun Heo } 3662a0a1a5fdSTejun Heo 3663a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3664a0a1a5fdSTejun Heo } 3665a0a1a5fdSTejun Heo 3666a0a1a5fdSTejun Heo /** 366758a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3668a0a1a5fdSTejun Heo * 3669a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3670a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3671a0a1a5fdSTejun Heo * 3672a0a1a5fdSTejun Heo * CONTEXT: 3673a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3674a0a1a5fdSTejun Heo * 3675a0a1a5fdSTejun Heo * RETURNS: 367658a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 367758a69cb4STejun Heo * is complete. 3678a0a1a5fdSTejun Heo */ 3679a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3680a0a1a5fdSTejun Heo { 3681a0a1a5fdSTejun Heo unsigned int cpu; 3682a0a1a5fdSTejun Heo bool busy = false; 3683a0a1a5fdSTejun Heo 3684a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3685a0a1a5fdSTejun Heo 3686a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3687a0a1a5fdSTejun Heo 3688706026c2STejun Heo for_each_wq_cpu(cpu) { 3689bdbc5dd7STejun Heo struct workqueue_struct *wq; 3690a0a1a5fdSTejun Heo /* 3691a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3692a0a1a5fdSTejun Heo * to peek without lock. 3693a0a1a5fdSTejun Heo */ 3694a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3695a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3696a0a1a5fdSTejun Heo 369758a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3698a0a1a5fdSTejun Heo continue; 3699a0a1a5fdSTejun Heo 3700a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3701a0a1a5fdSTejun Heo if (cwq->nr_active) { 3702a0a1a5fdSTejun Heo busy = true; 3703a0a1a5fdSTejun Heo goto out_unlock; 3704a0a1a5fdSTejun Heo } 3705a0a1a5fdSTejun Heo } 3706a0a1a5fdSTejun Heo } 3707a0a1a5fdSTejun Heo out_unlock: 3708a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3709a0a1a5fdSTejun Heo return busy; 3710a0a1a5fdSTejun Heo } 3711a0a1a5fdSTejun Heo 3712a0a1a5fdSTejun Heo /** 3713a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3714a0a1a5fdSTejun Heo * 3715a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 3716706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 3717a0a1a5fdSTejun Heo * 3718a0a1a5fdSTejun Heo * CONTEXT: 3719d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3720a0a1a5fdSTejun Heo */ 3721a0a1a5fdSTejun Heo void thaw_workqueues(void) 3722a0a1a5fdSTejun Heo { 3723a0a1a5fdSTejun Heo unsigned int cpu; 3724a0a1a5fdSTejun Heo 3725a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3726a0a1a5fdSTejun Heo 3727a0a1a5fdSTejun Heo if (!workqueue_freezing) 3728a0a1a5fdSTejun Heo goto out_unlock; 3729a0a1a5fdSTejun Heo 3730706026c2STejun Heo for_each_wq_cpu(cpu) { 37314ce62e9eSTejun Heo struct worker_pool *pool; 3732bdbc5dd7STejun Heo struct workqueue_struct *wq; 37338b03ae3cSTejun Heo 373438db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3735a1056305STejun Heo spin_lock_irq(&pool->lock); 3736d565ed63STejun Heo 373735b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 373835b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 3739db7bccf4STejun Heo 3740a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3741a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3742a0a1a5fdSTejun Heo 3743a1056305STejun Heo if (!cwq || cwq->pool != pool || 3744a1056305STejun Heo !(wq->flags & WQ_FREEZABLE)) 3745a0a1a5fdSTejun Heo continue; 3746a0a1a5fdSTejun Heo 3747a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 37489f4bd4cdSLai Jiangshan cwq_set_max_active(cwq, wq->saved_max_active); 3749a0a1a5fdSTejun Heo } 37508b03ae3cSTejun Heo 37514ce62e9eSTejun Heo wake_up_worker(pool); 3752a1056305STejun Heo 3753a1056305STejun Heo spin_unlock_irq(&pool->lock); 3754d565ed63STejun Heo } 3755a0a1a5fdSTejun Heo } 3756a0a1a5fdSTejun Heo 3757a0a1a5fdSTejun Heo workqueue_freezing = false; 3758a0a1a5fdSTejun Heo out_unlock: 3759a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3760a0a1a5fdSTejun Heo } 3761a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3762a0a1a5fdSTejun Heo 37636ee0578bSSuresh Siddha static int __init init_workqueues(void) 37641da177e4SLinus Torvalds { 3765c34056a3STejun Heo unsigned int cpu; 3766c34056a3STejun Heo 37677c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 37687c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 37696be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 3770b5490077STejun Heo 377165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3772a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 37738b03ae3cSTejun Heo 3774706026c2STejun Heo /* initialize CPU pools */ 3775706026c2STejun Heo for_each_wq_cpu(cpu) { 37764ce62e9eSTejun Heo struct worker_pool *pool; 37778b03ae3cSTejun Heo 377838db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3779d565ed63STejun Heo spin_lock_init(&pool->lock); 3780ec22ca5eSTejun Heo pool->cpu = cpu; 378124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 37824ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37834ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3784c9e7cf27STejun Heo hash_init(pool->busy_hash); 3785e22bee78STejun Heo 37864ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 37874ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 37884ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3789e22bee78STejun Heo 3790706026c2STejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 37914ce62e9eSTejun Heo (unsigned long)pool); 37924ce62e9eSTejun Heo 3793b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 37944ce62e9eSTejun Heo ida_init(&pool->worker_ida); 37959daf9e67STejun Heo 37969daf9e67STejun Heo /* alloc pool ID */ 37979daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 37984ce62e9eSTejun Heo } 37998b03ae3cSTejun Heo } 38008b03ae3cSTejun Heo 3801e22bee78STejun Heo /* create the initial worker */ 3802706026c2STejun Heo for_each_online_wq_cpu(cpu) { 38034ce62e9eSTejun Heo struct worker_pool *pool; 3804e22bee78STejun Heo 380538db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 38064ce62e9eSTejun Heo struct worker *worker; 38074ce62e9eSTejun Heo 380824647570STejun Heo if (cpu != WORK_CPU_UNBOUND) 380924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 381024647570STejun Heo 3811bc2ae0f5STejun Heo worker = create_worker(pool); 3812e22bee78STejun Heo BUG_ON(!worker); 3813d565ed63STejun Heo spin_lock_irq(&pool->lock); 3814e22bee78STejun Heo start_worker(worker); 3815d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3816e22bee78STejun Heo } 38174ce62e9eSTejun Heo } 3818e22bee78STejun Heo 3819d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 38201aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3821d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3822f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3823f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 382424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 382524d51addSTejun Heo WQ_FREEZABLE, 0); 38261aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3827ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 38286ee0578bSSuresh Siddha return 0; 38291da177e4SLinus Torvalds } 38306ee0578bSSuresh Siddha early_initcall(init_workqueues); 3831