11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4442f8570fSSasha Levin #include <linux/hashtable.h> 45e22bee78STejun Heo 46ea138446STejun Heo #include "workqueue_internal.h" 471da177e4SLinus Torvalds 48c8e55f36STejun Heo enum { 49bc2ae0f5STejun Heo /* 5024647570STejun Heo * worker_pool flags 51bc2ae0f5STejun Heo * 5224647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 53bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 54bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 55bc2ae0f5STejun Heo * is in effect. 56bc2ae0f5STejun Heo * 57bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 58bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 5924647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 60bc2ae0f5STejun Heo * 61bc2ae0f5STejun Heo * Note that DISASSOCIATED can be flipped only while holding 6224647570STejun Heo * assoc_mutex to avoid changing binding state while 6324647570STejun Heo * create_worker() is in progress. 64bc2ae0f5STejun Heo */ 6511ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 66552a37e9SLai Jiangshan POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */ 6724647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 6835b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 69db7bccf4STejun Heo 70c8e55f36STejun Heo /* worker flags */ 71c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 72c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 73c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 74e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 75fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 76f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 77e22bee78STejun Heo 785f7dabfdSLai Jiangshan WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | 79403c821dSTejun Heo WORKER_CPU_INTENSIVE, 80db7bccf4STejun Heo 81e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 824ce62e9eSTejun Heo 83c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 84db7bccf4STejun Heo 85e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 86e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 87e22bee78STejun Heo 883233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 893233cdbdSTejun Heo /* call for help after 10ms 903233cdbdSTejun Heo (min two ticks) */ 91e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 92e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds /* 95e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 96e22bee78STejun Heo * all cpus. Give -20. 97e22bee78STejun Heo */ 98e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 993270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 100c8e55f36STejun Heo }; 101c8e55f36STejun Heo 1021da177e4SLinus Torvalds /* 1034690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1044690c4abSTejun Heo * 105e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 106e41e704bSTejun Heo * everyone else. 1074690c4abSTejun Heo * 108e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 109e22bee78STejun Heo * only be modified and accessed from the local cpu. 110e22bee78STejun Heo * 111d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1124690c4abSTejun Heo * 113d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 114d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 115d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 116d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 117e22bee78STejun Heo * 11873f53c4aSTejun Heo * F: wq->flush_mutex protected. 11973f53c4aSTejun Heo * 1204690c4abSTejun Heo * W: workqueue_lock protected. 1214690c4abSTejun Heo */ 1224690c4abSTejun Heo 1232eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 124c34056a3STejun Heo 125bd7bdd43STejun Heo struct worker_pool { 126d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 127ec22ca5eSTejun Heo unsigned int cpu; /* I: the associated cpu */ 1289daf9e67STejun Heo int id; /* I: pool ID */ 12911ebea50STejun Heo unsigned int flags; /* X: flags */ 130bd7bdd43STejun Heo 131bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 132bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 133ea1abd61SLai Jiangshan 134ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 135bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 136bd7bdd43STejun Heo 137bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 138bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 139bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 140bd7bdd43STejun Heo 141c9e7cf27STejun Heo /* workers are chained either in busy_hash or idle_list */ 142c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 143c9e7cf27STejun Heo /* L: hash of busy workers */ 144c9e7cf27STejun Heo 14524647570STejun Heo struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */ 146bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 147e19e397aSTejun Heo 148e19e397aSTejun Heo /* 149e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 150e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 151e19e397aSTejun Heo * cacheline. 152e19e397aSTejun Heo */ 153e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 1548b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1558b03ae3cSTejun Heo 1568b03ae3cSTejun Heo /* 157112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 158112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 159112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 160112202d9STejun Heo * number of flag bits. 1611da177e4SLinus Torvalds */ 162112202d9STejun Heo struct pool_workqueue { 163bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1644690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 16573f53c4aSTejun Heo int work_color; /* L: current color */ 16673f53c4aSTejun Heo int flush_color; /* L: flushing color */ 16773f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 16873f53c4aSTejun Heo /* L: nr of in_flight works */ 1691e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 170a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1711e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 17230cdf249STejun Heo struct list_head pwqs_node; /* I: node on wq->pwqs */ 173e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 1741da177e4SLinus Torvalds 1751da177e4SLinus Torvalds /* 17673f53c4aSTejun Heo * Structure used to wait for workqueue flush. 17773f53c4aSTejun Heo */ 17873f53c4aSTejun Heo struct wq_flusher { 17973f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 18073f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 18173f53c4aSTejun Heo struct completion done; /* flush completion */ 18273f53c4aSTejun Heo }; 1831da177e4SLinus Torvalds 18473f53c4aSTejun Heo /* 185f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 186f2e005aaSTejun Heo * used to determine whether there's something to be done. 187f2e005aaSTejun Heo */ 188f2e005aaSTejun Heo #ifdef CONFIG_SMP 189f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 190f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 191f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 192f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 193f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 1949c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 195f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 196f2e005aaSTejun Heo #else 197f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 198f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 199f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 200f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 201f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 202f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 203f2e005aaSTejun Heo #endif 2041da177e4SLinus Torvalds 2051da177e4SLinus Torvalds /* 2061da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2071da177e4SLinus Torvalds * per-CPU workqueues: 2081da177e4SLinus Torvalds */ 2091da177e4SLinus Torvalds struct workqueue_struct { 2109c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 211bdbc5dd7STejun Heo union { 212112202d9STejun Heo struct pool_workqueue __percpu *pcpu; 213112202d9STejun Heo struct pool_workqueue *single; 214bdbc5dd7STejun Heo unsigned long v; 215112202d9STejun Heo } pool_wq; /* I: pwq's */ 21630cdf249STejun Heo struct list_head pwqs; /* I: all pwqs of this wq */ 2174690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 21873f53c4aSTejun Heo 21973f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 22073f53c4aSTejun Heo int work_color; /* F: current work color */ 22173f53c4aSTejun Heo int flush_color; /* F: current flush color */ 222112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 22373f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 22473f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 22573f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 22673f53c4aSTejun Heo 227f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 228e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 229e22bee78STejun Heo 2309c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 231112202d9STejun Heo int saved_max_active; /* W: saved pwq max_active */ 2324e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2334e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2344e6045f1SJohannes Berg #endif 235b196be89STejun Heo char name[]; /* I: workqueue name */ 2361da177e4SLinus Torvalds }; 2371da177e4SLinus Torvalds 238e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 239e904e6c2STejun Heo 240d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 241d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 242044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2431aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 244044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 245d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 246044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 247f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 248044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 24924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 250d320c038STejun Heo 25197bd2347STejun Heo #define CREATE_TRACE_POINTS 25297bd2347STejun Heo #include <trace/events/workqueue.h> 25397bd2347STejun Heo 25438db41d9STejun Heo #define for_each_std_worker_pool(pool, cpu) \ 255a60dc39cSTejun Heo for ((pool) = &std_worker_pools(cpu)[0]; \ 256a60dc39cSTejun Heo (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++) 2574ce62e9eSTejun Heo 258b67bfe0dSSasha Levin #define for_each_busy_worker(worker, i, pool) \ 259b67bfe0dSSasha Levin hash_for_each(pool->busy_hash, i, worker, hentry) 260db7bccf4STejun Heo 261706026c2STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 262f3421797STejun Heo unsigned int sw) 263f3421797STejun Heo { 264f3421797STejun Heo if (cpu < nr_cpu_ids) { 265f3421797STejun Heo if (sw & 1) { 266f3421797STejun Heo cpu = cpumask_next(cpu, mask); 267f3421797STejun Heo if (cpu < nr_cpu_ids) 268f3421797STejun Heo return cpu; 269f3421797STejun Heo } 270f3421797STejun Heo if (sw & 2) 271f3421797STejun Heo return WORK_CPU_UNBOUND; 272f3421797STejun Heo } 2736be19588SLai Jiangshan return WORK_CPU_END; 274f3421797STejun Heo } 275f3421797STejun Heo 276112202d9STejun Heo static inline int __next_pwq_cpu(int cpu, const struct cpumask *mask, 277f3421797STejun Heo struct workqueue_struct *wq) 278f3421797STejun Heo { 279706026c2STejun Heo return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 280f3421797STejun Heo } 281f3421797STejun Heo 28209884951STejun Heo /* 28309884951STejun Heo * CPU iterators 28409884951STejun Heo * 285706026c2STejun Heo * An extra cpu number is defined using an invalid cpu number 28609884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 287706026c2STejun Heo * specific CPU. The following iterators are similar to for_each_*_cpu() 288706026c2STejun Heo * iterators but also considers the unbound CPU. 28909884951STejun Heo * 290706026c2STejun Heo * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND 291706026c2STejun Heo * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND 292112202d9STejun Heo * for_each_pwq_cpu() : possible CPUs for bound workqueues, 29309884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 29409884951STejun Heo */ 295706026c2STejun Heo #define for_each_wq_cpu(cpu) \ 296706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \ 2976be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 298706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3)) 299f3421797STejun Heo 300706026c2STejun Heo #define for_each_online_wq_cpu(cpu) \ 301706026c2STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \ 3026be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 303706026c2STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3)) 304f3421797STejun Heo 305112202d9STejun Heo #define for_each_pwq_cpu(cpu, wq) \ 306112202d9STejun Heo for ((cpu) = __next_pwq_cpu(-1, cpu_possible_mask, (wq)); \ 3076be19588SLai Jiangshan (cpu) < WORK_CPU_END; \ 308112202d9STejun Heo (cpu) = __next_pwq_cpu((cpu), cpu_possible_mask, (wq))) 309f3421797STejun Heo 310dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 311dc186ad7SThomas Gleixner 312dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 313dc186ad7SThomas Gleixner 31499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 31599777288SStanislaw Gruszka { 31699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 31799777288SStanislaw Gruszka } 31899777288SStanislaw Gruszka 319dc186ad7SThomas Gleixner /* 320dc186ad7SThomas Gleixner * fixup_init is called when: 321dc186ad7SThomas Gleixner * - an active object is initialized 322dc186ad7SThomas Gleixner */ 323dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 324dc186ad7SThomas Gleixner { 325dc186ad7SThomas Gleixner struct work_struct *work = addr; 326dc186ad7SThomas Gleixner 327dc186ad7SThomas Gleixner switch (state) { 328dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 329dc186ad7SThomas Gleixner cancel_work_sync(work); 330dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 331dc186ad7SThomas Gleixner return 1; 332dc186ad7SThomas Gleixner default: 333dc186ad7SThomas Gleixner return 0; 334dc186ad7SThomas Gleixner } 335dc186ad7SThomas Gleixner } 336dc186ad7SThomas Gleixner 337dc186ad7SThomas Gleixner /* 338dc186ad7SThomas Gleixner * fixup_activate is called when: 339dc186ad7SThomas Gleixner * - an active object is activated 340dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 341dc186ad7SThomas Gleixner */ 342dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 343dc186ad7SThomas Gleixner { 344dc186ad7SThomas Gleixner struct work_struct *work = addr; 345dc186ad7SThomas Gleixner 346dc186ad7SThomas Gleixner switch (state) { 347dc186ad7SThomas Gleixner 348dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 349dc186ad7SThomas Gleixner /* 350dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 351dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 352dc186ad7SThomas Gleixner * is tracked in the object tracker. 353dc186ad7SThomas Gleixner */ 35422df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 355dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 356dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 357dc186ad7SThomas Gleixner return 0; 358dc186ad7SThomas Gleixner } 359dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 360dc186ad7SThomas Gleixner return 0; 361dc186ad7SThomas Gleixner 362dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 363dc186ad7SThomas Gleixner WARN_ON(1); 364dc186ad7SThomas Gleixner 365dc186ad7SThomas Gleixner default: 366dc186ad7SThomas Gleixner return 0; 367dc186ad7SThomas Gleixner } 368dc186ad7SThomas Gleixner } 369dc186ad7SThomas Gleixner 370dc186ad7SThomas Gleixner /* 371dc186ad7SThomas Gleixner * fixup_free is called when: 372dc186ad7SThomas Gleixner * - an active object is freed 373dc186ad7SThomas Gleixner */ 374dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 375dc186ad7SThomas Gleixner { 376dc186ad7SThomas Gleixner struct work_struct *work = addr; 377dc186ad7SThomas Gleixner 378dc186ad7SThomas Gleixner switch (state) { 379dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 380dc186ad7SThomas Gleixner cancel_work_sync(work); 381dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 382dc186ad7SThomas Gleixner return 1; 383dc186ad7SThomas Gleixner default: 384dc186ad7SThomas Gleixner return 0; 385dc186ad7SThomas Gleixner } 386dc186ad7SThomas Gleixner } 387dc186ad7SThomas Gleixner 388dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 389dc186ad7SThomas Gleixner .name = "work_struct", 39099777288SStanislaw Gruszka .debug_hint = work_debug_hint, 391dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 392dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 393dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 394dc186ad7SThomas Gleixner }; 395dc186ad7SThomas Gleixner 396dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 397dc186ad7SThomas Gleixner { 398dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 399dc186ad7SThomas Gleixner } 400dc186ad7SThomas Gleixner 401dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 402dc186ad7SThomas Gleixner { 403dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 404dc186ad7SThomas Gleixner } 405dc186ad7SThomas Gleixner 406dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 407dc186ad7SThomas Gleixner { 408dc186ad7SThomas Gleixner if (onstack) 409dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 410dc186ad7SThomas Gleixner else 411dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 412dc186ad7SThomas Gleixner } 413dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 414dc186ad7SThomas Gleixner 415dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 416dc186ad7SThomas Gleixner { 417dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 418dc186ad7SThomas Gleixner } 419dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 420dc186ad7SThomas Gleixner 421dc186ad7SThomas Gleixner #else 422dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 423dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 424dc186ad7SThomas Gleixner #endif 425dc186ad7SThomas Gleixner 42695402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 42795402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4281da177e4SLinus Torvalds static LIST_HEAD(workqueues); 429a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4301da177e4SLinus Torvalds 43114441960SOleg Nesterov /* 432e19e397aSTejun Heo * The CPU and unbound standard worker pools. The unbound ones have 433e19e397aSTejun Heo * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set. 43414441960SOleg Nesterov */ 435e19e397aSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 436a60dc39cSTejun Heo cpu_std_worker_pools); 437a60dc39cSTejun Heo static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS]; 438f3421797STejun Heo 4399daf9e67STejun Heo /* idr of all pools */ 4409daf9e67STejun Heo static DEFINE_MUTEX(worker_pool_idr_mutex); 4419daf9e67STejun Heo static DEFINE_IDR(worker_pool_idr); 4429daf9e67STejun Heo 443c34056a3STejun Heo static int worker_thread(void *__worker); 4441da177e4SLinus Torvalds 445a60dc39cSTejun Heo static struct worker_pool *std_worker_pools(int cpu) 4461da177e4SLinus Torvalds { 447f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 448a60dc39cSTejun Heo return per_cpu(cpu_std_worker_pools, cpu); 449f3421797STejun Heo else 450a60dc39cSTejun Heo return unbound_std_worker_pools; 4511da177e4SLinus Torvalds } 4521da177e4SLinus Torvalds 4534e8f0a60STejun Heo static int std_worker_pool_pri(struct worker_pool *pool) 4544e8f0a60STejun Heo { 455a60dc39cSTejun Heo return pool - std_worker_pools(pool->cpu); 4564e8f0a60STejun Heo } 4574e8f0a60STejun Heo 4589daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4599daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4609daf9e67STejun Heo { 4619daf9e67STejun Heo int ret; 4629daf9e67STejun Heo 4639daf9e67STejun Heo mutex_lock(&worker_pool_idr_mutex); 4649daf9e67STejun Heo idr_pre_get(&worker_pool_idr, GFP_KERNEL); 4659daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 4669daf9e67STejun Heo mutex_unlock(&worker_pool_idr_mutex); 4679daf9e67STejun Heo 4689daf9e67STejun Heo return ret; 4699daf9e67STejun Heo } 4709daf9e67STejun Heo 4717c3eed5cSTejun Heo /* 4727c3eed5cSTejun Heo * Lookup worker_pool by id. The idr currently is built during boot and 4737c3eed5cSTejun Heo * never modified. Don't worry about locking for now. 4747c3eed5cSTejun Heo */ 4757c3eed5cSTejun Heo static struct worker_pool *worker_pool_by_id(int pool_id) 4767c3eed5cSTejun Heo { 4777c3eed5cSTejun Heo return idr_find(&worker_pool_idr, pool_id); 4787c3eed5cSTejun Heo } 4797c3eed5cSTejun Heo 480d565ed63STejun Heo static struct worker_pool *get_std_worker_pool(int cpu, bool highpri) 481d565ed63STejun Heo { 482a60dc39cSTejun Heo struct worker_pool *pools = std_worker_pools(cpu); 483d565ed63STejun Heo 484a60dc39cSTejun Heo return &pools[highpri]; 485d565ed63STejun Heo } 486d565ed63STejun Heo 487112202d9STejun Heo static struct pool_workqueue *get_pwq(unsigned int cpu, 4884690c4abSTejun Heo struct workqueue_struct *wq) 489a848e3b6SOleg Nesterov { 490f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 491e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 492112202d9STejun Heo return per_cpu_ptr(wq->pool_wq.pcpu, cpu); 493f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 494112202d9STejun Heo return wq->pool_wq.single; 495f3421797STejun Heo return NULL; 496f3421797STejun Heo } 497a848e3b6SOleg Nesterov 49873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 49973f53c4aSTejun Heo { 50073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 50173f53c4aSTejun Heo } 50273f53c4aSTejun Heo 50373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 50473f53c4aSTejun Heo { 50573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 50673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 50773f53c4aSTejun Heo } 50873f53c4aSTejun Heo 50973f53c4aSTejun Heo static int work_next_color(int color) 51073f53c4aSTejun Heo { 51173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 512b1f4ec17SOleg Nesterov } 513b1f4ec17SOleg Nesterov 5144594bf15SDavid Howells /* 515112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 516112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5177c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5187a22ad75STejun Heo * 519112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 520112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 521bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 522bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5237a22ad75STejun Heo * 524112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5257c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 526112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5277c3eed5cSTejun Heo * available only while the work item is queued. 528bbb68dfaSTejun Heo * 529bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 530bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 531bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 532bbb68dfaSTejun Heo * try to steal the PENDING bit. 5334594bf15SDavid Howells */ 5347a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5357a22ad75STejun Heo unsigned long flags) 5367a22ad75STejun Heo { 5376183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5387a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5397a22ad75STejun Heo } 5407a22ad75STejun Heo 541112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5424690c4abSTejun Heo unsigned long extra_flags) 543365970a1SDavid Howells { 544112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 545112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 546365970a1SDavid Howells } 547365970a1SDavid Howells 5484468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5494468a00fSLai Jiangshan int pool_id) 5504468a00fSLai Jiangshan { 5514468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5524468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5534468a00fSLai Jiangshan } 5544468a00fSLai Jiangshan 5557c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5567c3eed5cSTejun Heo int pool_id) 5574d707b9fSOleg Nesterov { 55823657bb1STejun Heo /* 55923657bb1STejun Heo * The following wmb is paired with the implied mb in 56023657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 56123657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 56223657bb1STejun Heo * owner. 56323657bb1STejun Heo */ 56423657bb1STejun Heo smp_wmb(); 5657c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5664d707b9fSOleg Nesterov } 5674d707b9fSOleg Nesterov 5687a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 569365970a1SDavid Howells { 5707c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5717c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5727a22ad75STejun Heo } 5737a22ad75STejun Heo 574112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5757a22ad75STejun Heo { 576e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5777a22ad75STejun Heo 578112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 579e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 580e120153dSTejun Heo else 581e120153dSTejun Heo return NULL; 5827a22ad75STejun Heo } 5837a22ad75STejun Heo 5847c3eed5cSTejun Heo /** 5857c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5867c3eed5cSTejun Heo * @work: the work item of interest 5877c3eed5cSTejun Heo * 5887c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 5897c3eed5cSTejun Heo */ 5907c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 5917a22ad75STejun Heo { 592e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5937c3eed5cSTejun Heo struct worker_pool *pool; 5947c3eed5cSTejun Heo int pool_id; 5957a22ad75STejun Heo 596112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 597112202d9STejun Heo return ((struct pool_workqueue *) 5987c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 5997a22ad75STejun Heo 6007c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6017c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6027a22ad75STejun Heo return NULL; 6037a22ad75STejun Heo 6047c3eed5cSTejun Heo pool = worker_pool_by_id(pool_id); 6057c3eed5cSTejun Heo WARN_ON_ONCE(!pool); 6067c3eed5cSTejun Heo return pool; 6077c3eed5cSTejun Heo } 6087c3eed5cSTejun Heo 6097c3eed5cSTejun Heo /** 6107c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6117c3eed5cSTejun Heo * @work: the work item of interest 6127c3eed5cSTejun Heo * 6137c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6147c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6157c3eed5cSTejun Heo */ 6167c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6177c3eed5cSTejun Heo { 61854d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6197c3eed5cSTejun Heo 620112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 621112202d9STejun Heo return ((struct pool_workqueue *) 62254d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 62354d5b7d0SLai Jiangshan 62454d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 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 639112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (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 { 650e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 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 { 675e19e397aSTejun Heo return !list_empty(&pool->worklist) && 676e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 677e22bee78STejun Heo } 678e22bee78STejun Heo 679e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 68063d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 681e22bee78STejun Heo { 68263d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 683e22bee78STejun Heo } 684e22bee78STejun Heo 685e22bee78STejun Heo /* Do I need to be the manager? */ 68663d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 687e22bee78STejun Heo { 68863d95a91STejun Heo return need_to_create_worker(pool) || 68911ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 690e22bee78STejun Heo } 691e22bee78STejun Heo 692e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 69363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 694e22bee78STejun Heo { 695552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 69663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 69763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 698e22bee78STejun Heo 699ea1abd61SLai Jiangshan /* 700ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 701ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 702ea1abd61SLai Jiangshan */ 703ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 704ea1abd61SLai Jiangshan return false; 705ea1abd61SLai Jiangshan 706e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 707e22bee78STejun Heo } 708e22bee78STejun Heo 709e22bee78STejun Heo /* 710e22bee78STejun Heo * Wake up functions. 711e22bee78STejun Heo */ 712e22bee78STejun Heo 7137e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 71463d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7157e11629dSTejun Heo { 71663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7177e11629dSTejun Heo return NULL; 7187e11629dSTejun Heo 71963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7207e11629dSTejun Heo } 7217e11629dSTejun Heo 7227e11629dSTejun Heo /** 7237e11629dSTejun Heo * wake_up_worker - wake up an idle worker 72463d95a91STejun Heo * @pool: worker pool to wake worker from 7257e11629dSTejun Heo * 72663d95a91STejun Heo * Wake up the first idle worker of @pool. 7277e11629dSTejun Heo * 7287e11629dSTejun Heo * CONTEXT: 729d565ed63STejun Heo * spin_lock_irq(pool->lock). 7307e11629dSTejun Heo */ 73163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7327e11629dSTejun Heo { 73363d95a91STejun Heo struct worker *worker = first_worker(pool); 7347e11629dSTejun Heo 7357e11629dSTejun Heo if (likely(worker)) 7367e11629dSTejun Heo wake_up_process(worker->task); 7377e11629dSTejun Heo } 7387e11629dSTejun Heo 7394690c4abSTejun Heo /** 740e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 741e22bee78STejun Heo * @task: task waking up 742e22bee78STejun Heo * @cpu: CPU @task is waking up to 743e22bee78STejun Heo * 744e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 745e22bee78STejun Heo * being awoken. 746e22bee78STejun Heo * 747e22bee78STejun Heo * CONTEXT: 748e22bee78STejun Heo * spin_lock_irq(rq->lock) 749e22bee78STejun Heo */ 750e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 751e22bee78STejun Heo { 752e22bee78STejun Heo struct worker *worker = kthread_data(task); 753e22bee78STejun Heo 75436576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 755ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 756e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 757e22bee78STejun Heo } 75836576000SJoonsoo Kim } 759e22bee78STejun Heo 760e22bee78STejun Heo /** 761e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 762e22bee78STejun Heo * @task: task going to sleep 763e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 764e22bee78STejun Heo * 765e22bee78STejun Heo * This function is called during schedule() when a busy worker is 766e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 767e22bee78STejun Heo * returning pointer to its task. 768e22bee78STejun Heo * 769e22bee78STejun Heo * CONTEXT: 770e22bee78STejun Heo * spin_lock_irq(rq->lock) 771e22bee78STejun Heo * 772e22bee78STejun Heo * RETURNS: 773e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 774e22bee78STejun Heo */ 775e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 776e22bee78STejun Heo unsigned int cpu) 777e22bee78STejun Heo { 778e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 779111c225aSTejun Heo struct worker_pool *pool; 780e22bee78STejun Heo 781111c225aSTejun Heo /* 782111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 783111c225aSTejun Heo * workers, also reach here, let's not access anything before 784111c225aSTejun Heo * checking NOT_RUNNING. 785111c225aSTejun Heo */ 7862d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 787e22bee78STejun Heo return NULL; 788e22bee78STejun Heo 789111c225aSTejun Heo pool = worker->pool; 790111c225aSTejun Heo 791e22bee78STejun Heo /* this can only happen on the local cpu */ 7926183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 7936183c009STejun Heo return NULL; 794e22bee78STejun Heo 795e22bee78STejun Heo /* 796e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 797e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 798e22bee78STejun Heo * Please read comment there. 799e22bee78STejun Heo * 800628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 801628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 802628c78e7STejun Heo * disabled, which in turn means that none else could be 803d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 804628c78e7STejun Heo * lock is safe. 805e22bee78STejun Heo */ 806e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 807e19e397aSTejun Heo !list_empty(&pool->worklist)) 80863d95a91STejun Heo to_wakeup = first_worker(pool); 809e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 810e22bee78STejun Heo } 811e22bee78STejun Heo 812e22bee78STejun Heo /** 813e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 814cb444766STejun Heo * @worker: self 815d302f017STejun Heo * @flags: flags to set 816d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 817d302f017STejun Heo * 818e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 819e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 820e22bee78STejun Heo * woken up. 821d302f017STejun Heo * 822cb444766STejun Heo * CONTEXT: 823d565ed63STejun Heo * spin_lock_irq(pool->lock) 824d302f017STejun Heo */ 825d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 826d302f017STejun Heo bool wakeup) 827d302f017STejun Heo { 828bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 829e22bee78STejun Heo 830cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 831cb444766STejun Heo 832e22bee78STejun Heo /* 833e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 834e22bee78STejun Heo * wake up an idle worker as necessary if requested by 835e22bee78STejun Heo * @wakeup. 836e22bee78STejun Heo */ 837e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 838e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 839e22bee78STejun Heo if (wakeup) { 840e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 841bd7bdd43STejun Heo !list_empty(&pool->worklist)) 84263d95a91STejun Heo wake_up_worker(pool); 843e22bee78STejun Heo } else 844e19e397aSTejun Heo atomic_dec(&pool->nr_running); 845e22bee78STejun Heo } 846e22bee78STejun Heo 847d302f017STejun Heo worker->flags |= flags; 848d302f017STejun Heo } 849d302f017STejun Heo 850d302f017STejun Heo /** 851e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 852cb444766STejun Heo * @worker: self 853d302f017STejun Heo * @flags: flags to clear 854d302f017STejun Heo * 855e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 856d302f017STejun Heo * 857cb444766STejun Heo * CONTEXT: 858d565ed63STejun Heo * spin_lock_irq(pool->lock) 859d302f017STejun Heo */ 860d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 861d302f017STejun Heo { 86263d95a91STejun Heo struct worker_pool *pool = worker->pool; 863e22bee78STejun Heo unsigned int oflags = worker->flags; 864e22bee78STejun Heo 865cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 866cb444766STejun Heo 867d302f017STejun Heo worker->flags &= ~flags; 868e22bee78STejun Heo 86942c025f3STejun Heo /* 87042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 87142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 87242c025f3STejun Heo * of multiple flags, not a single flag. 87342c025f3STejun Heo */ 874e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 875e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 876e19e397aSTejun Heo atomic_inc(&pool->nr_running); 877d302f017STejun Heo } 878d302f017STejun Heo 879d302f017STejun Heo /** 8808cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 881c9e7cf27STejun Heo * @pool: pool of interest 8828cca0eeaSTejun Heo * @work: work to find worker for 8838cca0eeaSTejun Heo * 884c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 885c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 886a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 887a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 888a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 889a2c1c57bSTejun Heo * being executed. 890a2c1c57bSTejun Heo * 891a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 892a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 893a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 894a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 895a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 896a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 897a2c1c57bSTejun Heo * 898a2c1c57bSTejun Heo * This function checks the work item address, work function and workqueue 899a2c1c57bSTejun Heo * to avoid false positives. Note that this isn't complete as one may 900a2c1c57bSTejun Heo * construct a work function which can introduce dependency onto itself 901a2c1c57bSTejun Heo * through a recycled work item. Well, if somebody wants to shoot oneself 902a2c1c57bSTejun Heo * in the foot that badly, there's only so much we can do, and if such 903a2c1c57bSTejun Heo * deadlock actually occurs, it should be easy to locate the culprit work 904a2c1c57bSTejun Heo * function. 9058cca0eeaSTejun Heo * 9068cca0eeaSTejun Heo * CONTEXT: 907d565ed63STejun Heo * spin_lock_irq(pool->lock). 9088cca0eeaSTejun Heo * 9098cca0eeaSTejun Heo * RETURNS: 9108cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9118cca0eeaSTejun Heo * otherwise. 9128cca0eeaSTejun Heo */ 913c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9148cca0eeaSTejun Heo struct work_struct *work) 9158cca0eeaSTejun Heo { 91642f8570fSSasha Levin struct worker *worker; 91742f8570fSSasha Levin 918b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 919a2c1c57bSTejun Heo (unsigned long)work) 920a2c1c57bSTejun Heo if (worker->current_work == work && 921a2c1c57bSTejun Heo worker->current_func == work->func) 92242f8570fSSasha Levin return worker; 92342f8570fSSasha Levin 92442f8570fSSasha Levin return NULL; 9258cca0eeaSTejun Heo } 9268cca0eeaSTejun Heo 9278cca0eeaSTejun Heo /** 928bf4ede01STejun Heo * move_linked_works - move linked works to a list 929bf4ede01STejun Heo * @work: start of series of works to be scheduled 930bf4ede01STejun Heo * @head: target list to append @work to 931bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 932bf4ede01STejun Heo * 933bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 934bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 935bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 936bf4ede01STejun Heo * 937bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 938bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 939bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 940bf4ede01STejun Heo * 941bf4ede01STejun Heo * CONTEXT: 942d565ed63STejun Heo * spin_lock_irq(pool->lock). 943bf4ede01STejun Heo */ 944bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 945bf4ede01STejun Heo struct work_struct **nextp) 946bf4ede01STejun Heo { 947bf4ede01STejun Heo struct work_struct *n; 948bf4ede01STejun Heo 949bf4ede01STejun Heo /* 950bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 951bf4ede01STejun Heo * use NULL for list head. 952bf4ede01STejun Heo */ 953bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 954bf4ede01STejun Heo list_move_tail(&work->entry, head); 955bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 956bf4ede01STejun Heo break; 957bf4ede01STejun Heo } 958bf4ede01STejun Heo 959bf4ede01STejun Heo /* 960bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 961bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 962bf4ede01STejun Heo * needs to be updated. 963bf4ede01STejun Heo */ 964bf4ede01STejun Heo if (nextp) 965bf4ede01STejun Heo *nextp = n; 966bf4ede01STejun Heo } 967bf4ede01STejun Heo 968112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 969bf4ede01STejun Heo { 970112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 971bf4ede01STejun Heo 972bf4ede01STejun Heo trace_workqueue_activate_work(work); 973112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 974bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 975112202d9STejun Heo pwq->nr_active++; 976bf4ede01STejun Heo } 977bf4ede01STejun Heo 978112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 9793aa62497SLai Jiangshan { 980112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 9813aa62497SLai Jiangshan struct work_struct, entry); 9823aa62497SLai Jiangshan 983112202d9STejun Heo pwq_activate_delayed_work(work); 9843aa62497SLai Jiangshan } 9853aa62497SLai Jiangshan 986bf4ede01STejun Heo /** 987112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 988112202d9STejun Heo * @pwq: pwq of interest 989bf4ede01STejun Heo * @color: color of work which left the queue 990bf4ede01STejun Heo * 991bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 992112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 993bf4ede01STejun Heo * 994bf4ede01STejun Heo * CONTEXT: 995d565ed63STejun Heo * spin_lock_irq(pool->lock). 996bf4ede01STejun Heo */ 997112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 998bf4ede01STejun Heo { 999bf4ede01STejun Heo /* ignore uncolored works */ 1000bf4ede01STejun Heo if (color == WORK_NO_COLOR) 1001bf4ede01STejun Heo return; 1002bf4ede01STejun Heo 1003112202d9STejun Heo pwq->nr_in_flight[color]--; 1004bf4ede01STejun Heo 1005112202d9STejun Heo pwq->nr_active--; 1006112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1007bf4ede01STejun Heo /* one down, submit a delayed one */ 1008112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1009112202d9STejun Heo pwq_activate_first_delayed(pwq); 1010bf4ede01STejun Heo } 1011bf4ede01STejun Heo 1012bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1013112202d9STejun Heo if (likely(pwq->flush_color != color)) 1014bf4ede01STejun Heo return; 1015bf4ede01STejun Heo 1016bf4ede01STejun Heo /* are there still in-flight works? */ 1017112202d9STejun Heo if (pwq->nr_in_flight[color]) 1018bf4ede01STejun Heo return; 1019bf4ede01STejun Heo 1020112202d9STejun Heo /* this pwq is done, clear flush_color */ 1021112202d9STejun Heo pwq->flush_color = -1; 1022bf4ede01STejun Heo 1023bf4ede01STejun Heo /* 1024112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1025bf4ede01STejun Heo * will handle the rest. 1026bf4ede01STejun Heo */ 1027112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1028112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 1029bf4ede01STejun Heo } 1030bf4ede01STejun Heo 103136e227d2STejun Heo /** 1032bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 103336e227d2STejun Heo * @work: work item to steal 103436e227d2STejun Heo * @is_dwork: @work is a delayed_work 1035bbb68dfaSTejun Heo * @flags: place to store irq state 103636e227d2STejun Heo * 103736e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 103836e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 103936e227d2STejun Heo * 104036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 104136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 104236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1043bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1044bbb68dfaSTejun Heo * for arbitrarily long 104536e227d2STejun Heo * 1046bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1047e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1048e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1049e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1050bbb68dfaSTejun Heo * 1051bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1052bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1053bbb68dfaSTejun Heo * 1054e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1055bf4ede01STejun Heo */ 1056bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1057bbb68dfaSTejun Heo unsigned long *flags) 1058bf4ede01STejun Heo { 1059d565ed63STejun Heo struct worker_pool *pool; 1060112202d9STejun Heo struct pool_workqueue *pwq; 1061bf4ede01STejun Heo 1062bbb68dfaSTejun Heo local_irq_save(*flags); 1063bbb68dfaSTejun Heo 106436e227d2STejun Heo /* try to steal the timer if it exists */ 106536e227d2STejun Heo if (is_dwork) { 106636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 106736e227d2STejun Heo 1068e0aecdd8STejun Heo /* 1069e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1070e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1071e0aecdd8STejun Heo * running on the local CPU. 1072e0aecdd8STejun Heo */ 107336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 107436e227d2STejun Heo return 1; 107536e227d2STejun Heo } 107636e227d2STejun Heo 107736e227d2STejun Heo /* try to claim PENDING the normal way */ 1078bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1079bf4ede01STejun Heo return 0; 1080bf4ede01STejun Heo 1081bf4ede01STejun Heo /* 1082bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1083bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1084bf4ede01STejun Heo */ 1085d565ed63STejun Heo pool = get_work_pool(work); 1086d565ed63STejun Heo if (!pool) 1087bbb68dfaSTejun Heo goto fail; 1088bf4ede01STejun Heo 1089d565ed63STejun Heo spin_lock(&pool->lock); 1090bf4ede01STejun Heo /* 1091112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1092112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1093112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1094112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1095112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 10960b3dae68SLai Jiangshan * item is currently queued on that pool. 1097bf4ede01STejun Heo */ 1098112202d9STejun Heo pwq = get_work_pwq(work); 1099112202d9STejun Heo if (pwq && pwq->pool == pool) { 1100bf4ede01STejun Heo debug_work_deactivate(work); 11013aa62497SLai Jiangshan 11023aa62497SLai Jiangshan /* 110316062836STejun Heo * A delayed work item cannot be grabbed directly because 110416062836STejun Heo * it might have linked NO_COLOR work items which, if left 1105112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 110616062836STejun Heo * management later on and cause stall. Make sure the work 110716062836STejun Heo * item is activated before grabbing. 11083aa62497SLai Jiangshan */ 11093aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1110112202d9STejun Heo pwq_activate_delayed_work(work); 11113aa62497SLai Jiangshan 1112bf4ede01STejun Heo list_del_init(&work->entry); 1113112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 111436e227d2STejun Heo 1115112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11164468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11174468a00fSLai Jiangshan 1118d565ed63STejun Heo spin_unlock(&pool->lock); 111936e227d2STejun Heo return 1; 1120bf4ede01STejun Heo } 1121d565ed63STejun Heo spin_unlock(&pool->lock); 1122bbb68dfaSTejun Heo fail: 1123bbb68dfaSTejun Heo local_irq_restore(*flags); 1124bbb68dfaSTejun Heo if (work_is_canceling(work)) 1125bbb68dfaSTejun Heo return -ENOENT; 1126bbb68dfaSTejun Heo cpu_relax(); 112736e227d2STejun Heo return -EAGAIN; 1128bf4ede01STejun Heo } 1129bf4ede01STejun Heo 1130bf4ede01STejun Heo /** 1131706026c2STejun Heo * insert_work - insert a work into a pool 1132112202d9STejun Heo * @pwq: pwq @work belongs to 11334690c4abSTejun Heo * @work: work to insert 11344690c4abSTejun Heo * @head: insertion point 11354690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11364690c4abSTejun Heo * 1137112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1138706026c2STejun Heo * work_struct flags. 11394690c4abSTejun Heo * 11404690c4abSTejun Heo * CONTEXT: 1141d565ed63STejun Heo * spin_lock_irq(pool->lock). 1142365970a1SDavid Howells */ 1143112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1144112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1145b89deed3SOleg Nesterov { 1146112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1147e1d8aa9fSFrederic Weisbecker 11484690c4abSTejun Heo /* we own @work, set data and link */ 1149112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 11501a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1151e22bee78STejun Heo 1152e22bee78STejun Heo /* 1153e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1154e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1155e22bee78STejun Heo * lying around lazily while there are works to be processed. 1156e22bee78STejun Heo */ 1157e22bee78STejun Heo smp_mb(); 1158e22bee78STejun Heo 115963d95a91STejun Heo if (__need_more_worker(pool)) 116063d95a91STejun Heo wake_up_worker(pool); 1161b89deed3SOleg Nesterov } 1162b89deed3SOleg Nesterov 1163c8efcc25STejun Heo /* 1164c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 11658d03ecfeSTejun Heo * same workqueue. 1166c8efcc25STejun Heo */ 1167c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1168c8efcc25STejun Heo { 1169c8efcc25STejun Heo struct worker *worker; 1170c8efcc25STejun Heo 11718d03ecfeSTejun Heo worker = current_wq_worker(); 1172c8efcc25STejun Heo /* 11738d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 11748d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1175c8efcc25STejun Heo */ 1176112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1177c8efcc25STejun Heo } 1178c8efcc25STejun Heo 11794690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 11801da177e4SLinus Torvalds struct work_struct *work) 11811da177e4SLinus Torvalds { 1182112202d9STejun Heo struct pool_workqueue *pwq; 11831e19ffc6STejun Heo struct list_head *worklist; 11848a2e8e5dSTejun Heo unsigned int work_flags; 1185b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 11868930cabaSTejun Heo 11878930cabaSTejun Heo /* 11888930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 11898930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 11908930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 11918930cabaSTejun Heo * happen with IRQ disabled. 11928930cabaSTejun Heo */ 11938930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 11941da177e4SLinus Torvalds 1195dc186ad7SThomas Gleixner debug_work_activate(work); 11961e19ffc6STejun Heo 1197c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 11989c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1199c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1200e41e704bSTejun Heo return; 1201e41e704bSTejun Heo 1202112202d9STejun Heo /* determine the pwq to use */ 1203c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1204c9e7cf27STejun Heo struct worker_pool *last_pool; 1205c7fc77f7STejun Heo 120657469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1207f3421797STejun Heo cpu = raw_smp_processor_id(); 1208f3421797STejun Heo 120918aa9effSTejun Heo /* 1210dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1211dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1212dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1213dbf2576eSTejun Heo * non-reentrancy. 121418aa9effSTejun Heo */ 1215112202d9STejun Heo pwq = get_pwq(cpu, wq); 1216c9e7cf27STejun Heo last_pool = get_work_pool(work); 1217dbf2576eSTejun Heo 1218112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 121918aa9effSTejun Heo struct worker *worker; 122018aa9effSTejun Heo 1221d565ed63STejun Heo spin_lock(&last_pool->lock); 122218aa9effSTejun Heo 1223c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 122418aa9effSTejun Heo 1225112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1226112202d9STejun Heo pwq = get_pwq(last_pool->cpu, wq); 12278594fadeSLai Jiangshan } else { 122818aa9effSTejun Heo /* meh... not running there, queue here */ 1229d565ed63STejun Heo spin_unlock(&last_pool->lock); 1230112202d9STejun Heo spin_lock(&pwq->pool->lock); 123118aa9effSTejun Heo } 12328930cabaSTejun Heo } else { 1233112202d9STejun Heo spin_lock(&pwq->pool->lock); 12348930cabaSTejun Heo } 1235f3421797STejun Heo } else { 1236112202d9STejun Heo pwq = get_pwq(WORK_CPU_UNBOUND, wq); 1237112202d9STejun Heo spin_lock(&pwq->pool->lock); 1238502ca9d8STejun Heo } 1239502ca9d8STejun Heo 1240112202d9STejun Heo /* pwq determined, queue */ 1241112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1242502ca9d8STejun Heo 1243f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1244112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1245f5b2552bSDan Carpenter return; 1246f5b2552bSDan Carpenter } 12471e19ffc6STejun Heo 1248112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1249112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 12501e19ffc6STejun Heo 1251112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1252cdadf009STejun Heo trace_workqueue_activate_work(work); 1253112202d9STejun Heo pwq->nr_active++; 1254112202d9STejun Heo worklist = &pwq->pool->worklist; 12558a2e8e5dSTejun Heo } else { 12568a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1257112202d9STejun Heo worklist = &pwq->delayed_works; 12588a2e8e5dSTejun Heo } 12591e19ffc6STejun Heo 1260112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 12611e19ffc6STejun Heo 1262112202d9STejun Heo spin_unlock(&pwq->pool->lock); 12631da177e4SLinus Torvalds } 12641da177e4SLinus Torvalds 12650fcb78c2SRolf Eike Beer /** 1266c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1267c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1268c1a220e7SZhang Rui * @wq: workqueue to use 1269c1a220e7SZhang Rui * @work: work to queue 1270c1a220e7SZhang Rui * 1271d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1272c1a220e7SZhang Rui * 1273c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1274c1a220e7SZhang Rui * can't go away. 1275c1a220e7SZhang Rui */ 1276d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1277d4283e93STejun Heo struct work_struct *work) 1278c1a220e7SZhang Rui { 1279d4283e93STejun Heo bool ret = false; 12808930cabaSTejun Heo unsigned long flags; 12818930cabaSTejun Heo 12828930cabaSTejun Heo local_irq_save(flags); 1283c1a220e7SZhang Rui 128422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 12854690c4abSTejun Heo __queue_work(cpu, wq, work); 1286d4283e93STejun Heo ret = true; 1287c1a220e7SZhang Rui } 12888930cabaSTejun Heo 12898930cabaSTejun Heo local_irq_restore(flags); 1290c1a220e7SZhang Rui return ret; 1291c1a220e7SZhang Rui } 1292c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1293c1a220e7SZhang Rui 12940a13c00eSTejun Heo /** 12950a13c00eSTejun Heo * queue_work - queue work on a workqueue 12960a13c00eSTejun Heo * @wq: workqueue to use 12970a13c00eSTejun Heo * @work: work to queue 12980a13c00eSTejun Heo * 1299d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 13000a13c00eSTejun Heo * 13010a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 13020a13c00eSTejun Heo * it can be processed by another CPU. 13030a13c00eSTejun Heo */ 1304d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13050a13c00eSTejun Heo { 130657469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13070a13c00eSTejun Heo } 13080a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13090a13c00eSTejun Heo 1310d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13111da177e4SLinus Torvalds { 131252bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13131da177e4SLinus Torvalds 1314e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 131560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13161da177e4SLinus Torvalds } 13171438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13181da177e4SLinus Torvalds 13197beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 132052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13211da177e4SLinus Torvalds { 13227beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13237beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13241da177e4SLinus Torvalds 13257beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13267beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1327fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1328fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13297beb2edfSTejun Heo 13308852aac2STejun Heo /* 13318852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13328852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13338852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13348852aac2STejun Heo * on that there's no such delay when @delay is 0. 13358852aac2STejun Heo */ 13368852aac2STejun Heo if (!delay) { 13378852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13388852aac2STejun Heo return; 13398852aac2STejun Heo } 13408852aac2STejun Heo 13417beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13427beb2edfSTejun Heo 134360c057bcSLai Jiangshan dwork->wq = wq; 13441265057fSTejun Heo dwork->cpu = cpu; 13457beb2edfSTejun Heo timer->expires = jiffies + delay; 13467beb2edfSTejun Heo 13477beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13487beb2edfSTejun Heo add_timer_on(timer, cpu); 13497beb2edfSTejun Heo else 13507beb2edfSTejun Heo add_timer(timer); 13517beb2edfSTejun Heo } 13521da177e4SLinus Torvalds 13530fcb78c2SRolf Eike Beer /** 13540fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13550fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13560fcb78c2SRolf Eike Beer * @wq: workqueue to use 1357af9997e4SRandy Dunlap * @dwork: work to queue 13580fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13590fcb78c2SRolf Eike Beer * 1360715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1361715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1362715f1300STejun Heo * execution. 13630fcb78c2SRolf Eike Beer */ 1364d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 136552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13667a6bc1cdSVenkatesh Pallipadi { 136752bad64dSDavid Howells struct work_struct *work = &dwork->work; 1368d4283e93STejun Heo bool ret = false; 13698930cabaSTejun Heo unsigned long flags; 13708930cabaSTejun Heo 13718930cabaSTejun Heo /* read the comment in __queue_work() */ 13728930cabaSTejun Heo local_irq_save(flags); 13737a6bc1cdSVenkatesh Pallipadi 137422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13757beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1376d4283e93STejun Heo ret = true; 13777a6bc1cdSVenkatesh Pallipadi } 13788930cabaSTejun Heo 13798930cabaSTejun Heo local_irq_restore(flags); 13807a6bc1cdSVenkatesh Pallipadi return ret; 13817a6bc1cdSVenkatesh Pallipadi } 1382ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 13831da177e4SLinus Torvalds 1384c8e55f36STejun Heo /** 13850a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 13860a13c00eSTejun Heo * @wq: workqueue to use 13870a13c00eSTejun Heo * @dwork: delayable work to queue 13880a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 13890a13c00eSTejun Heo * 1390715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 13910a13c00eSTejun Heo */ 1392d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 13930a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 13940a13c00eSTejun Heo { 139557469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 13960a13c00eSTejun Heo } 13970a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 13980a13c00eSTejun Heo 13990a13c00eSTejun Heo /** 14008376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14018376fe22STejun Heo * @cpu: CPU number to execute work on 14028376fe22STejun Heo * @wq: workqueue to use 14038376fe22STejun Heo * @dwork: work to queue 14048376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14058376fe22STejun Heo * 14068376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14078376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14088376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14098376fe22STejun Heo * current state. 14108376fe22STejun Heo * 14118376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14128376fe22STejun Heo * pending and its timer was modified. 14138376fe22STejun Heo * 1414e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14158376fe22STejun Heo * See try_to_grab_pending() for details. 14168376fe22STejun Heo */ 14178376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14188376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14198376fe22STejun Heo { 14208376fe22STejun Heo unsigned long flags; 14218376fe22STejun Heo int ret; 14228376fe22STejun Heo 14238376fe22STejun Heo do { 14248376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14258376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14268376fe22STejun Heo 14278376fe22STejun Heo if (likely(ret >= 0)) { 14288376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14298376fe22STejun Heo local_irq_restore(flags); 14308376fe22STejun Heo } 14318376fe22STejun Heo 14328376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14338376fe22STejun Heo return ret; 14348376fe22STejun Heo } 14358376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14368376fe22STejun Heo 14378376fe22STejun Heo /** 14388376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14398376fe22STejun Heo * @wq: workqueue to use 14408376fe22STejun Heo * @dwork: work to queue 14418376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14428376fe22STejun Heo * 14438376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14448376fe22STejun Heo */ 14458376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14468376fe22STejun Heo unsigned long delay) 14478376fe22STejun Heo { 14488376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14498376fe22STejun Heo } 14508376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14518376fe22STejun Heo 14528376fe22STejun Heo /** 1453c8e55f36STejun Heo * worker_enter_idle - enter idle state 1454c8e55f36STejun Heo * @worker: worker which is entering idle state 1455c8e55f36STejun Heo * 1456c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1457c8e55f36STejun Heo * necessary. 1458c8e55f36STejun Heo * 1459c8e55f36STejun Heo * LOCKING: 1460d565ed63STejun Heo * spin_lock_irq(pool->lock). 1461c8e55f36STejun Heo */ 1462c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14631da177e4SLinus Torvalds { 1464bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1465c8e55f36STejun Heo 14666183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 14676183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 14686183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 14696183c009STejun Heo return; 1470c8e55f36STejun Heo 1471cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1472cb444766STejun Heo worker->flags |= WORKER_IDLE; 1473bd7bdd43STejun Heo pool->nr_idle++; 1474e22bee78STejun Heo worker->last_active = jiffies; 1475c8e55f36STejun Heo 1476c8e55f36STejun Heo /* idle_list is LIFO */ 1477bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1478db7bccf4STejun Heo 147963d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1480628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1481cb444766STejun Heo 1482544ecf31STejun Heo /* 1483706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1484d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1485628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1486628c78e7STejun Heo * unbind is not in progress. 1487544ecf31STejun Heo */ 148824647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1489bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1490e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1491c8e55f36STejun Heo } 1492c8e55f36STejun Heo 1493c8e55f36STejun Heo /** 1494c8e55f36STejun Heo * worker_leave_idle - leave idle state 1495c8e55f36STejun Heo * @worker: worker which is leaving idle state 1496c8e55f36STejun Heo * 1497c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1498c8e55f36STejun Heo * 1499c8e55f36STejun Heo * LOCKING: 1500d565ed63STejun Heo * spin_lock_irq(pool->lock). 1501c8e55f36STejun Heo */ 1502c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1503c8e55f36STejun Heo { 1504bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1505c8e55f36STejun Heo 15066183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15076183c009STejun Heo return; 1508d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1509bd7bdd43STejun Heo pool->nr_idle--; 1510c8e55f36STejun Heo list_del_init(&worker->entry); 1511c8e55f36STejun Heo } 1512c8e55f36STejun Heo 1513e22bee78STejun Heo /** 1514f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1515f36dc67bSLai Jiangshan * @pool: target worker_pool 1516f36dc67bSLai Jiangshan * 1517f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1518e22bee78STejun Heo * 1519e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1520e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1521e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1522e22bee78STejun Heo * guaranteed to execute on the cpu. 1523e22bee78STejun Heo * 1524f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1525e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1526e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1527e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1528706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1529e22bee78STejun Heo * [dis]associated in the meantime. 1530e22bee78STejun Heo * 1531706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 153224647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1533f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1534f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1535f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1536e22bee78STejun Heo * 1537e22bee78STejun Heo * CONTEXT: 1538d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1539e22bee78STejun Heo * held. 1540e22bee78STejun Heo * 1541e22bee78STejun Heo * RETURNS: 1542706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1543e22bee78STejun Heo * bound), %false if offline. 1544e22bee78STejun Heo */ 1545f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1546d565ed63STejun Heo __acquires(&pool->lock) 1547e22bee78STejun Heo { 1548e22bee78STejun Heo while (true) { 1549e22bee78STejun Heo /* 1550e22bee78STejun Heo * The following call may fail, succeed or succeed 1551e22bee78STejun Heo * without actually migrating the task to the cpu if 1552e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 155324647570STejun Heo * against POOL_DISASSOCIATED. 1554e22bee78STejun Heo */ 155524647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 1556f5faa077SLai Jiangshan set_cpus_allowed_ptr(current, get_cpu_mask(pool->cpu)); 1557e22bee78STejun Heo 1558d565ed63STejun Heo spin_lock_irq(&pool->lock); 155924647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1560e22bee78STejun Heo return false; 1561f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 1562e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1563ec22ca5eSTejun Heo get_cpu_mask(pool->cpu))) 1564e22bee78STejun Heo return true; 1565d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1566e22bee78STejun Heo 15675035b20fSTejun Heo /* 15685035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 15695035b20fSTejun Heo * and retry migration. cond_resched() is required here; 15705035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 15715035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 15725035b20fSTejun Heo */ 1573e22bee78STejun Heo cpu_relax(); 15745035b20fSTejun Heo cond_resched(); 1575e22bee78STejun Heo } 1576e22bee78STejun Heo } 1577e22bee78STejun Heo 1578e22bee78STejun Heo /* 1579ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 15805f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 158125511a47STejun Heo */ 158225511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 158325511a47STejun Heo { 15845f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 1585f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 15865f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 158725511a47STejun Heo 1588ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1589ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1590d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 159125511a47STejun Heo } 159225511a47STejun Heo 159325511a47STejun Heo /* 159425511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1595403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1596403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1597403c821dSTejun Heo * executed twice without intervening cpu down. 1598e22bee78STejun Heo */ 159925511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1600e22bee78STejun Heo { 1601e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1602e22bee78STejun Heo 1603f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 1604eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1605e22bee78STejun Heo 1606d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1607e22bee78STejun Heo } 1608e22bee78STejun Heo 160925511a47STejun Heo /** 161094cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 161194cf58bbSTejun Heo * @pool: pool of interest 161225511a47STejun Heo * 161394cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 161425511a47STejun Heo * is different for idle and busy ones. 161525511a47STejun Heo * 1616ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1617ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1618ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1619ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 162025511a47STejun Heo * 1621ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1622ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1623ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1624ea1abd61SLai Jiangshan * rebind. 162525511a47STejun Heo * 1626ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1627ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1628ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1629ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 163025511a47STejun Heo */ 163194cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 163225511a47STejun Heo { 1633ea1abd61SLai Jiangshan struct worker *worker, *n; 163425511a47STejun Heo int i; 163525511a47STejun Heo 1636b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 1637d565ed63STejun Heo lockdep_assert_held(&pool->lock); 163825511a47STejun Heo 16395f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1640ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1641ea1abd61SLai Jiangshan /* 164294cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 164394cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1644ea1abd61SLai Jiangshan */ 1645ea1abd61SLai Jiangshan list_del_init(&worker->entry); 164625511a47STejun Heo 164725511a47STejun Heo /* 164894cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 164994cf58bbSTejun Heo * idle_worker_rebind(). 165025511a47STejun Heo */ 165125511a47STejun Heo wake_up_process(worker->task); 165225511a47STejun Heo } 165325511a47STejun Heo 1654ea1abd61SLai Jiangshan /* rebind busy workers */ 1655b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) { 165625511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1657e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 165825511a47STejun Heo 165925511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 166025511a47STejun Heo work_data_bits(rebind_work))) 166125511a47STejun Heo continue; 166225511a47STejun Heo 166325511a47STejun Heo debug_work_activate(rebind_work); 166490beca5dSTejun Heo 166590beca5dSTejun Heo /* 166694cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 1667112202d9STejun Heo * and @pwq->pool consistent for sanity. 166890beca5dSTejun Heo */ 1669e34cdddbSTejun Heo if (std_worker_pool_pri(worker->pool)) 1670e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1671e2b6a6d5SJoonsoo Kim else 1672e2b6a6d5SJoonsoo Kim wq = system_wq; 1673ec58815aSTejun Heo 1674112202d9STejun Heo insert_work(get_pwq(pool->cpu, wq), rebind_work, 167525511a47STejun Heo worker->scheduled.next, 167625511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1677ec58815aSTejun Heo } 167825511a47STejun Heo } 167925511a47STejun Heo 1680c34056a3STejun Heo static struct worker *alloc_worker(void) 1681c34056a3STejun Heo { 1682c34056a3STejun Heo struct worker *worker; 1683c34056a3STejun Heo 1684c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1685c8e55f36STejun Heo if (worker) { 1686c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1687affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 168825511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1689e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1690e22bee78STejun Heo worker->flags = WORKER_PREP; 1691c8e55f36STejun Heo } 1692c34056a3STejun Heo return worker; 1693c34056a3STejun Heo } 1694c34056a3STejun Heo 1695c34056a3STejun Heo /** 1696c34056a3STejun Heo * create_worker - create a new workqueue worker 169763d95a91STejun Heo * @pool: pool the new worker will belong to 1698c34056a3STejun Heo * 169963d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1700c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1701c34056a3STejun Heo * destroy_worker(). 1702c34056a3STejun Heo * 1703c34056a3STejun Heo * CONTEXT: 1704c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1705c34056a3STejun Heo * 1706c34056a3STejun Heo * RETURNS: 1707c34056a3STejun Heo * Pointer to the newly created worker. 1708c34056a3STejun Heo */ 1709bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1710c34056a3STejun Heo { 1711e34cdddbSTejun Heo const char *pri = std_worker_pool_pri(pool) ? "H" : ""; 1712c34056a3STejun Heo struct worker *worker = NULL; 1713f3421797STejun Heo int id = -1; 1714c34056a3STejun Heo 1715d565ed63STejun Heo spin_lock_irq(&pool->lock); 1716bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1717d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1718bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1719c34056a3STejun Heo goto fail; 1720d565ed63STejun Heo spin_lock_irq(&pool->lock); 1721c34056a3STejun Heo } 1722d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1723c34056a3STejun Heo 1724c34056a3STejun Heo worker = alloc_worker(); 1725c34056a3STejun Heo if (!worker) 1726c34056a3STejun Heo goto fail; 1727c34056a3STejun Heo 1728bd7bdd43STejun Heo worker->pool = pool; 1729c34056a3STejun Heo worker->id = id; 1730c34056a3STejun Heo 1731ec22ca5eSTejun Heo if (pool->cpu != WORK_CPU_UNBOUND) 173294dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1733ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1734ec22ca5eSTejun Heo "kworker/%u:%d%s", pool->cpu, id, pri); 1735f3421797STejun Heo else 1736f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 17373270476aSTejun Heo "kworker/u:%d%s", id, pri); 1738c34056a3STejun Heo if (IS_ERR(worker->task)) 1739c34056a3STejun Heo goto fail; 1740c34056a3STejun Heo 1741e34cdddbSTejun Heo if (std_worker_pool_pri(pool)) 17423270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 17433270476aSTejun Heo 1744db7bccf4STejun Heo /* 1745bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 174624647570STejun Heo * %POOL_DISASSOCIATED. The caller is responsible for ensuring the 1747bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1748bc2ae0f5STejun Heo * above the flag definition for details. 1749bc2ae0f5STejun Heo * 1750bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1751bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1752db7bccf4STejun Heo */ 175324647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) { 1754ec22ca5eSTejun Heo kthread_bind(worker->task, pool->cpu); 1755bc2ae0f5STejun Heo } else { 1756db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1757f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1758f3421797STejun Heo } 1759c34056a3STejun Heo 1760c34056a3STejun Heo return worker; 1761c34056a3STejun Heo fail: 1762c34056a3STejun Heo if (id >= 0) { 1763d565ed63STejun Heo spin_lock_irq(&pool->lock); 1764bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1765d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1766c34056a3STejun Heo } 1767c34056a3STejun Heo kfree(worker); 1768c34056a3STejun Heo return NULL; 1769c34056a3STejun Heo } 1770c34056a3STejun Heo 1771c34056a3STejun Heo /** 1772c34056a3STejun Heo * start_worker - start a newly created worker 1773c34056a3STejun Heo * @worker: worker to start 1774c34056a3STejun Heo * 1775706026c2STejun Heo * Make the pool aware of @worker and start it. 1776c34056a3STejun Heo * 1777c34056a3STejun Heo * CONTEXT: 1778d565ed63STejun Heo * spin_lock_irq(pool->lock). 1779c34056a3STejun Heo */ 1780c34056a3STejun Heo static void start_worker(struct worker *worker) 1781c34056a3STejun Heo { 1782cb444766STejun Heo worker->flags |= WORKER_STARTED; 1783bd7bdd43STejun Heo worker->pool->nr_workers++; 1784c8e55f36STejun Heo worker_enter_idle(worker); 1785c34056a3STejun Heo wake_up_process(worker->task); 1786c34056a3STejun Heo } 1787c34056a3STejun Heo 1788c34056a3STejun Heo /** 1789c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1790c34056a3STejun Heo * @worker: worker to be destroyed 1791c34056a3STejun Heo * 1792706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1793c8e55f36STejun Heo * 1794c8e55f36STejun Heo * CONTEXT: 1795d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1796c34056a3STejun Heo */ 1797c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1798c34056a3STejun Heo { 1799bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1800c34056a3STejun Heo int id = worker->id; 1801c34056a3STejun Heo 1802c34056a3STejun Heo /* sanity check frenzy */ 18036183c009STejun Heo if (WARN_ON(worker->current_work) || 18046183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18056183c009STejun Heo return; 1806c34056a3STejun Heo 1807c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1808bd7bdd43STejun Heo pool->nr_workers--; 1809c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1810bd7bdd43STejun Heo pool->nr_idle--; 1811c8e55f36STejun Heo 1812c8e55f36STejun Heo list_del_init(&worker->entry); 1813cb444766STejun Heo worker->flags |= WORKER_DIE; 1814c8e55f36STejun Heo 1815d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1816c8e55f36STejun Heo 1817c34056a3STejun Heo kthread_stop(worker->task); 1818c34056a3STejun Heo kfree(worker); 1819c34056a3STejun Heo 1820d565ed63STejun Heo spin_lock_irq(&pool->lock); 1821bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1822c34056a3STejun Heo } 1823c34056a3STejun Heo 182463d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1825e22bee78STejun Heo { 182663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1827e22bee78STejun Heo 1828d565ed63STejun Heo spin_lock_irq(&pool->lock); 1829e22bee78STejun Heo 183063d95a91STejun Heo if (too_many_workers(pool)) { 1831e22bee78STejun Heo struct worker *worker; 1832e22bee78STejun Heo unsigned long expires; 1833e22bee78STejun Heo 1834e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 183563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1836e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1837e22bee78STejun Heo 1838e22bee78STejun Heo if (time_before(jiffies, expires)) 183963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1840e22bee78STejun Heo else { 1841e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 184211ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 184363d95a91STejun Heo wake_up_worker(pool); 1844e22bee78STejun Heo } 1845e22bee78STejun Heo } 1846e22bee78STejun Heo 1847d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1848e22bee78STejun Heo } 1849e22bee78STejun Heo 1850e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1851e22bee78STejun Heo { 1852112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1853112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1854f3421797STejun Heo unsigned int cpu; 1855e22bee78STejun Heo 1856e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1857e22bee78STejun Heo return false; 1858e22bee78STejun Heo 1859e22bee78STejun Heo /* mayday mayday mayday */ 1860112202d9STejun Heo cpu = pwq->pool->cpu; 1861f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1862f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1863f3421797STejun Heo cpu = 0; 1864f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1865e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1866e22bee78STejun Heo return true; 1867e22bee78STejun Heo } 1868e22bee78STejun Heo 1869706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1870e22bee78STejun Heo { 187163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1872e22bee78STejun Heo struct work_struct *work; 1873e22bee78STejun Heo 1874d565ed63STejun Heo spin_lock_irq(&pool->lock); 1875e22bee78STejun Heo 187663d95a91STejun Heo if (need_to_create_worker(pool)) { 1877e22bee78STejun Heo /* 1878e22bee78STejun Heo * We've been trying to create a new worker but 1879e22bee78STejun Heo * haven't been successful. We might be hitting an 1880e22bee78STejun Heo * allocation deadlock. Send distress signals to 1881e22bee78STejun Heo * rescuers. 1882e22bee78STejun Heo */ 188363d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1884e22bee78STejun Heo send_mayday(work); 1885e22bee78STejun Heo } 1886e22bee78STejun Heo 1887d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1888e22bee78STejun Heo 188963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892e22bee78STejun Heo /** 1893e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 189463d95a91STejun Heo * @pool: pool to create a new worker for 1895e22bee78STejun Heo * 189663d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1897e22bee78STejun Heo * have at least one idle worker on return from this function. If 1898e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 189963d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1900e22bee78STejun Heo * possible allocation deadlock. 1901e22bee78STejun Heo * 1902e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1903e22bee78STejun Heo * may_start_working() true. 1904e22bee78STejun Heo * 1905e22bee78STejun Heo * LOCKING: 1906d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1907e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1908e22bee78STejun Heo * manager. 1909e22bee78STejun Heo * 1910e22bee78STejun Heo * RETURNS: 1911d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1912e22bee78STejun Heo * otherwise. 1913e22bee78STejun Heo */ 191463d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1915d565ed63STejun Heo __releases(&pool->lock) 1916d565ed63STejun Heo __acquires(&pool->lock) 1917e22bee78STejun Heo { 191863d95a91STejun Heo if (!need_to_create_worker(pool)) 1919e22bee78STejun Heo return false; 1920e22bee78STejun Heo restart: 1921d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19229f9c2364STejun Heo 1923e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 192463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1925e22bee78STejun Heo 1926e22bee78STejun Heo while (true) { 1927e22bee78STejun Heo struct worker *worker; 1928e22bee78STejun Heo 1929bc2ae0f5STejun Heo worker = create_worker(pool); 1930e22bee78STejun Heo if (worker) { 193163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1932d565ed63STejun Heo spin_lock_irq(&pool->lock); 1933e22bee78STejun Heo start_worker(worker); 19346183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19356183c009STejun Heo goto restart; 1936e22bee78STejun Heo return true; 1937e22bee78STejun Heo } 1938e22bee78STejun Heo 193963d95a91STejun Heo if (!need_to_create_worker(pool)) 1940e22bee78STejun Heo break; 1941e22bee78STejun Heo 1942e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1943e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19449f9c2364STejun Heo 194563d95a91STejun Heo if (!need_to_create_worker(pool)) 1946e22bee78STejun Heo break; 1947e22bee78STejun Heo } 1948e22bee78STejun Heo 194963d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1950d565ed63STejun Heo spin_lock_irq(&pool->lock); 195163d95a91STejun Heo if (need_to_create_worker(pool)) 1952e22bee78STejun Heo goto restart; 1953e22bee78STejun Heo return true; 1954e22bee78STejun Heo } 1955e22bee78STejun Heo 1956e22bee78STejun Heo /** 1957e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 195863d95a91STejun Heo * @pool: pool to destroy workers for 1959e22bee78STejun Heo * 196063d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1961e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1962e22bee78STejun Heo * 1963e22bee78STejun Heo * LOCKING: 1964d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1965e22bee78STejun Heo * multiple times. Called only from manager. 1966e22bee78STejun Heo * 1967e22bee78STejun Heo * RETURNS: 1968d565ed63STejun Heo * false if no action was taken and pool->lock stayed locked, true 1969e22bee78STejun Heo * otherwise. 1970e22bee78STejun Heo */ 197163d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1972e22bee78STejun Heo { 1973e22bee78STejun Heo bool ret = false; 1974e22bee78STejun Heo 197563d95a91STejun Heo while (too_many_workers(pool)) { 1976e22bee78STejun Heo struct worker *worker; 1977e22bee78STejun Heo unsigned long expires; 1978e22bee78STejun Heo 197963d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1980e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1981e22bee78STejun Heo 1982e22bee78STejun Heo if (time_before(jiffies, expires)) { 198363d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1984e22bee78STejun Heo break; 1985e22bee78STejun Heo } 1986e22bee78STejun Heo 1987e22bee78STejun Heo destroy_worker(worker); 1988e22bee78STejun Heo ret = true; 1989e22bee78STejun Heo } 1990e22bee78STejun Heo 1991e22bee78STejun Heo return ret; 1992e22bee78STejun Heo } 1993e22bee78STejun Heo 1994e22bee78STejun Heo /** 1995e22bee78STejun Heo * manage_workers - manage worker pool 1996e22bee78STejun Heo * @worker: self 1997e22bee78STejun Heo * 1998706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1999e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2000706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2001e22bee78STejun Heo * 2002e22bee78STejun Heo * The caller can safely start processing works on false return. On 2003e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2004e22bee78STejun Heo * and may_start_working() is true. 2005e22bee78STejun Heo * 2006e22bee78STejun Heo * CONTEXT: 2007d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2008e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2009e22bee78STejun Heo * 2010e22bee78STejun Heo * RETURNS: 2011d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2012d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2013e22bee78STejun Heo */ 2014e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2015e22bee78STejun Heo { 201663d95a91STejun Heo struct worker_pool *pool = worker->pool; 2017e22bee78STejun Heo bool ret = false; 2018e22bee78STejun Heo 2019ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2020e22bee78STejun Heo return ret; 2021e22bee78STejun Heo 2022552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2023ee378aa4SLai Jiangshan 2024ee378aa4SLai Jiangshan /* 2025ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2026ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2027ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2028ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2029ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2030b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2031ee378aa4SLai Jiangshan * manager against CPU hotplug. 2032ee378aa4SLai Jiangshan * 2033b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2034d565ed63STejun Heo * progress. trylock first without dropping @pool->lock. 2035ee378aa4SLai Jiangshan */ 2036b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2037d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2038b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2039ee378aa4SLai Jiangshan /* 2040ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2041b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2042ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2043706026c2STejun Heo * @pool's state and ours could have deviated. 2044ee378aa4SLai Jiangshan * 2045b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2046ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2047706026c2STejun Heo * on @pool's current state. Try it and adjust 2048ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2049ee378aa4SLai Jiangshan */ 2050f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(pool)) 2051ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2052ee378aa4SLai Jiangshan else 2053ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2054ee378aa4SLai Jiangshan 2055ee378aa4SLai Jiangshan ret = true; 2056ee378aa4SLai Jiangshan } 2057ee378aa4SLai Jiangshan 205811ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2059e22bee78STejun Heo 2060e22bee78STejun Heo /* 2061e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2062e22bee78STejun Heo * on return. 2063e22bee78STejun Heo */ 206463d95a91STejun Heo ret |= maybe_destroy_workers(pool); 206563d95a91STejun Heo ret |= maybe_create_worker(pool); 2066e22bee78STejun Heo 2067552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2068b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2069e22bee78STejun Heo return ret; 2070e22bee78STejun Heo } 2071e22bee78STejun Heo 2072a62428c0STejun Heo /** 2073a62428c0STejun Heo * process_one_work - process single work 2074c34056a3STejun Heo * @worker: self 2075a62428c0STejun Heo * @work: work to process 2076a62428c0STejun Heo * 2077a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2078a62428c0STejun Heo * process a single work including synchronization against and 2079a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2080a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2081a62428c0STejun Heo * call this function to process a work. 2082a62428c0STejun Heo * 2083a62428c0STejun Heo * CONTEXT: 2084d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2085a62428c0STejun Heo */ 2086c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2087d565ed63STejun Heo __releases(&pool->lock) 2088d565ed63STejun Heo __acquires(&pool->lock) 20891da177e4SLinus Torvalds { 2090112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2091bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2092112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 209373f53c4aSTejun Heo int work_color; 20947e11629dSTejun Heo struct worker *collision; 20954e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20964e6045f1SJohannes Berg /* 2097a62428c0STejun Heo * It is permissible to free the struct work_struct from 2098a62428c0STejun Heo * inside the function that is called from it, this we need to 2099a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2100a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2101a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21024e6045f1SJohannes Berg */ 21034d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21044d82a1deSPeter Zijlstra 21054d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21064e6045f1SJohannes Berg #endif 21076fec10a1STejun Heo /* 21086fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21096fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 211024647570STejun Heo * unbound or a disassociated pool. 21116fec10a1STejun Heo */ 21125f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 211324647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2114ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 211525511a47STejun Heo 21167e11629dSTejun Heo /* 21177e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21187e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21197e11629dSTejun Heo * already processing the work. If so, defer the work to the 21207e11629dSTejun Heo * currently executing one. 21217e11629dSTejun Heo */ 2122c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21237e11629dSTejun Heo if (unlikely(collision)) { 21247e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21257e11629dSTejun Heo return; 21267e11629dSTejun Heo } 21271da177e4SLinus Torvalds 21288930cabaSTejun Heo /* claim and dequeue */ 21291da177e4SLinus Torvalds debug_work_deactivate(work); 2130c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2131c34056a3STejun Heo worker->current_work = work; 2132a2c1c57bSTejun Heo worker->current_func = work->func; 2133112202d9STejun Heo worker->current_pwq = pwq; 213473f53c4aSTejun Heo work_color = get_work_color(work); 21357a22ad75STejun Heo 2136a62428c0STejun Heo list_del_init(&work->entry); 2137a62428c0STejun Heo 2138649027d7STejun Heo /* 2139fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2140fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2141fb0e7bebSTejun Heo */ 2142fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2143fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2144fb0e7bebSTejun Heo 2145974271c4STejun Heo /* 2146d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2147974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2148974271c4STejun Heo */ 214963d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 215063d95a91STejun Heo wake_up_worker(pool); 2151974271c4STejun Heo 21528930cabaSTejun Heo /* 21537c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2154d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 215523657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 215623657bb1STejun Heo * disabled. 21578930cabaSTejun Heo */ 21587c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21591da177e4SLinus Torvalds 2160d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2161365970a1SDavid Howells 2162112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21633295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2164e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2165a2c1c57bSTejun Heo worker->current_func(work); 2166e36c886aSArjan van de Ven /* 2167e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2168e36c886aSArjan van de Ven * point will only record its address. 2169e36c886aSArjan van de Ven */ 2170e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21713295f0efSIngo Molnar lock_map_release(&lockdep_map); 2172112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21731da177e4SLinus Torvalds 2174d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2175044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2176044c782cSValentin Ilie " last function: %pf\n", 2177a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2178a2c1c57bSTejun Heo worker->current_func); 2179d5abe669SPeter Zijlstra debug_show_held_locks(current); 2180d5abe669SPeter Zijlstra dump_stack(); 2181d5abe669SPeter Zijlstra } 2182d5abe669SPeter Zijlstra 2183d565ed63STejun Heo spin_lock_irq(&pool->lock); 2184a62428c0STejun Heo 2185fb0e7bebSTejun Heo /* clear cpu intensive status */ 2186fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2187fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2188fb0e7bebSTejun Heo 2189a62428c0STejun Heo /* we're done with it, release */ 219042f8570fSSasha Levin hash_del(&worker->hentry); 2191c34056a3STejun Heo worker->current_work = NULL; 2192a2c1c57bSTejun Heo worker->current_func = NULL; 2193112202d9STejun Heo worker->current_pwq = NULL; 2194112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21951da177e4SLinus Torvalds } 21961da177e4SLinus Torvalds 2197affee4b2STejun Heo /** 2198affee4b2STejun Heo * process_scheduled_works - process scheduled works 2199affee4b2STejun Heo * @worker: self 2200affee4b2STejun Heo * 2201affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2202affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2203affee4b2STejun Heo * fetches a work from the top and executes it. 2204affee4b2STejun Heo * 2205affee4b2STejun Heo * CONTEXT: 2206d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2207affee4b2STejun Heo * multiple times. 2208affee4b2STejun Heo */ 2209affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22101da177e4SLinus Torvalds { 2211affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2212affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2213a62428c0STejun Heo struct work_struct, entry); 2214c34056a3STejun Heo process_one_work(worker, work); 2215a62428c0STejun Heo } 22161da177e4SLinus Torvalds } 22171da177e4SLinus Torvalds 22184690c4abSTejun Heo /** 22194690c4abSTejun Heo * worker_thread - the worker thread function 2220c34056a3STejun Heo * @__worker: self 22214690c4abSTejun Heo * 2222706026c2STejun Heo * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools 2223706026c2STejun Heo * of these per each cpu. These workers process all works regardless of 2224e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2225e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2226e22bee78STejun Heo * rescuer_thread(). 22274690c4abSTejun Heo */ 2228c34056a3STejun Heo static int worker_thread(void *__worker) 22291da177e4SLinus Torvalds { 2230c34056a3STejun Heo struct worker *worker = __worker; 2231bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22321da177e4SLinus Torvalds 2233e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2234e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2235c8e55f36STejun Heo woke_up: 2236d565ed63STejun Heo spin_lock_irq(&pool->lock); 2237affee4b2STejun Heo 22385f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22395f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2240d565ed63STejun Heo spin_unlock_irq(&pool->lock); 224125511a47STejun Heo 22425f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 224325511a47STejun Heo if (worker->flags & WORKER_DIE) { 2244e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2245c8e55f36STejun Heo return 0; 2246c8e55f36STejun Heo } 2247c8e55f36STejun Heo 22485f7dabfdSLai Jiangshan /* otherwise, rebind */ 224925511a47STejun Heo idle_worker_rebind(worker); 225025511a47STejun Heo goto woke_up; 225125511a47STejun Heo } 225225511a47STejun Heo 2253c8e55f36STejun Heo worker_leave_idle(worker); 2254db7bccf4STejun Heo recheck: 2255e22bee78STejun Heo /* no more worker necessary? */ 225663d95a91STejun Heo if (!need_more_worker(pool)) 2257e22bee78STejun Heo goto sleep; 2258e22bee78STejun Heo 2259e22bee78STejun Heo /* do we need to manage? */ 226063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2261e22bee78STejun Heo goto recheck; 2262e22bee78STejun Heo 2263c8e55f36STejun Heo /* 2264c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2265c8e55f36STejun Heo * preparing to process a work or actually processing it. 2266c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2267c8e55f36STejun Heo */ 22686183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2269c8e55f36STejun Heo 2270e22bee78STejun Heo /* 2271e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2272e22bee78STejun Heo * at least one idle worker or that someone else has already 2273e22bee78STejun Heo * assumed the manager role. 2274e22bee78STejun Heo */ 2275e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2276e22bee78STejun Heo 2277e22bee78STejun Heo do { 2278affee4b2STejun Heo struct work_struct *work = 2279bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2280affee4b2STejun Heo struct work_struct, entry); 2281affee4b2STejun Heo 2282c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2283affee4b2STejun Heo /* optimization path, not strictly necessary */ 2284affee4b2STejun Heo process_one_work(worker, work); 2285affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2286affee4b2STejun Heo process_scheduled_works(worker); 2287affee4b2STejun Heo } else { 2288c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2289affee4b2STejun Heo process_scheduled_works(worker); 2290affee4b2STejun Heo } 229163d95a91STejun Heo } while (keep_working(pool)); 2292affee4b2STejun Heo 2293e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2294d313dd85STejun Heo sleep: 229563d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2296e22bee78STejun Heo goto recheck; 2297d313dd85STejun Heo 2298c8e55f36STejun Heo /* 2299d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2300d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2301d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2302d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2303d565ed63STejun Heo * event. 2304c8e55f36STejun Heo */ 2305c8e55f36STejun Heo worker_enter_idle(worker); 2306c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2307d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23081da177e4SLinus Torvalds schedule(); 2309c8e55f36STejun Heo goto woke_up; 23101da177e4SLinus Torvalds } 23111da177e4SLinus Torvalds 2312e22bee78STejun Heo /** 2313e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2314111c225aSTejun Heo * @__rescuer: self 2315e22bee78STejun Heo * 2316e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2317e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2318e22bee78STejun Heo * 2319706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2320e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2321e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2322e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2323e22bee78STejun Heo * the problem rescuer solves. 2324e22bee78STejun Heo * 2325706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2326706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2327e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2328e22bee78STejun Heo * 2329e22bee78STejun Heo * This should happen rarely. 2330e22bee78STejun Heo */ 2331111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2332e22bee78STejun Heo { 2333111c225aSTejun Heo struct worker *rescuer = __rescuer; 2334111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2335e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2336f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2337e22bee78STejun Heo unsigned int cpu; 2338e22bee78STejun Heo 2339e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2340111c225aSTejun Heo 2341111c225aSTejun Heo /* 2342111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2343111c225aSTejun Heo * doesn't participate in concurrency management. 2344111c225aSTejun Heo */ 2345111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2346e22bee78STejun Heo repeat: 2347e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23481da177e4SLinus Torvalds 2349412d32e6SMike Galbraith if (kthread_should_stop()) { 2350412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2351111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2352e22bee78STejun Heo return 0; 2353412d32e6SMike Galbraith } 23541da177e4SLinus Torvalds 2355f3421797STejun Heo /* 2356f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2357f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2358f3421797STejun Heo */ 2359f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2360f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2361112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(tcpu, wq); 2362112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2363e22bee78STejun Heo struct work_struct *work, *n; 2364e22bee78STejun Heo 2365e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2366f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2367e22bee78STejun Heo 2368e22bee78STejun Heo /* migrate to the target cpu if possible */ 2369f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2370b3104104SLai Jiangshan rescuer->pool = pool; 2371e22bee78STejun Heo 2372e22bee78STejun Heo /* 2373e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2374e22bee78STejun Heo * process'em. 2375e22bee78STejun Heo */ 23766183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2377bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2378112202d9STejun Heo if (get_work_pwq(work) == pwq) 2379e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2380e22bee78STejun Heo 2381e22bee78STejun Heo process_scheduled_works(rescuer); 23827576958aSTejun Heo 23837576958aSTejun Heo /* 2384d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23857576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23867576958aSTejun Heo * and stalling the execution. 23877576958aSTejun Heo */ 238863d95a91STejun Heo if (keep_working(pool)) 238963d95a91STejun Heo wake_up_worker(pool); 23907576958aSTejun Heo 2391b3104104SLai Jiangshan rescuer->pool = NULL; 2392d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23931da177e4SLinus Torvalds } 23941da177e4SLinus Torvalds 2395111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2396111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2397e22bee78STejun Heo schedule(); 2398e22bee78STejun Heo goto repeat; 23991da177e4SLinus Torvalds } 24001da177e4SLinus Torvalds 2401fc2e4d70SOleg Nesterov struct wq_barrier { 2402fc2e4d70SOleg Nesterov struct work_struct work; 2403fc2e4d70SOleg Nesterov struct completion done; 2404fc2e4d70SOleg Nesterov }; 2405fc2e4d70SOleg Nesterov 2406fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2407fc2e4d70SOleg Nesterov { 2408fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2409fc2e4d70SOleg Nesterov complete(&barr->done); 2410fc2e4d70SOleg Nesterov } 2411fc2e4d70SOleg Nesterov 24124690c4abSTejun Heo /** 24134690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2414112202d9STejun Heo * @pwq: pwq to insert barrier into 24154690c4abSTejun Heo * @barr: wq_barrier to insert 2416affee4b2STejun Heo * @target: target work to attach @barr to 2417affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24184690c4abSTejun Heo * 2419affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2420affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2421affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2422affee4b2STejun Heo * cpu. 2423affee4b2STejun Heo * 2424affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2425affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2426affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2427affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2428affee4b2STejun Heo * after a work with LINKED flag set. 2429affee4b2STejun Heo * 2430affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2431112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24324690c4abSTejun Heo * 24334690c4abSTejun Heo * CONTEXT: 2434d565ed63STejun Heo * spin_lock_irq(pool->lock). 24354690c4abSTejun Heo */ 2436112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2437affee4b2STejun Heo struct wq_barrier *barr, 2438affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2439fc2e4d70SOleg Nesterov { 2440affee4b2STejun Heo struct list_head *head; 2441affee4b2STejun Heo unsigned int linked = 0; 2442affee4b2STejun Heo 2443dc186ad7SThomas Gleixner /* 2444d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2445dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2446dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2447dc186ad7SThomas Gleixner * might deadlock. 2448dc186ad7SThomas Gleixner */ 2449ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 245022df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2451fc2e4d70SOleg Nesterov init_completion(&barr->done); 245283c22520SOleg Nesterov 2453affee4b2STejun Heo /* 2454affee4b2STejun Heo * If @target is currently being executed, schedule the 2455affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2456affee4b2STejun Heo */ 2457affee4b2STejun Heo if (worker) 2458affee4b2STejun Heo head = worker->scheduled.next; 2459affee4b2STejun Heo else { 2460affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2461affee4b2STejun Heo 2462affee4b2STejun Heo head = target->entry.next; 2463affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2464affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2465affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2466affee4b2STejun Heo } 2467affee4b2STejun Heo 2468dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2469112202d9STejun Heo insert_work(pwq, &barr->work, head, 2470affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2471fc2e4d70SOleg Nesterov } 2472fc2e4d70SOleg Nesterov 247373f53c4aSTejun Heo /** 2474112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 247573f53c4aSTejun Heo * @wq: workqueue being flushed 247673f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 247773f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 247873f53c4aSTejun Heo * 2479112202d9STejun Heo * Prepare pwqs for workqueue flushing. 248073f53c4aSTejun Heo * 2481112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2482112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2483112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2484112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2485112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 248673f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 248773f53c4aSTejun Heo * 248873f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 248973f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 249073f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 249173f53c4aSTejun Heo * is returned. 249273f53c4aSTejun Heo * 2493112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 249473f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 249573f53c4aSTejun Heo * advanced to @work_color. 249673f53c4aSTejun Heo * 249773f53c4aSTejun Heo * CONTEXT: 249873f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 249973f53c4aSTejun Heo * 250073f53c4aSTejun Heo * RETURNS: 250173f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 250273f53c4aSTejun Heo * otherwise. 250373f53c4aSTejun Heo */ 2504112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 250573f53c4aSTejun Heo int flush_color, int work_color) 25061da177e4SLinus Torvalds { 250773f53c4aSTejun Heo bool wait = false; 250873f53c4aSTejun Heo unsigned int cpu; 25091da177e4SLinus Torvalds 251073f53c4aSTejun Heo if (flush_color >= 0) { 25116183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2512112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2513dc186ad7SThomas Gleixner } 251414441960SOleg Nesterov 2515112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 2516112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 2517112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25181da177e4SLinus Torvalds 2519d565ed63STejun Heo spin_lock_irq(&pool->lock); 252073f53c4aSTejun Heo 252173f53c4aSTejun Heo if (flush_color >= 0) { 25226183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 252373f53c4aSTejun Heo 2524112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2525112202d9STejun Heo pwq->flush_color = flush_color; 2526112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 252773f53c4aSTejun Heo wait = true; 25281da177e4SLinus Torvalds } 252973f53c4aSTejun Heo } 253073f53c4aSTejun Heo 253173f53c4aSTejun Heo if (work_color >= 0) { 25326183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2533112202d9STejun Heo pwq->work_color = work_color; 253473f53c4aSTejun Heo } 253573f53c4aSTejun Heo 2536d565ed63STejun Heo spin_unlock_irq(&pool->lock); 25371da177e4SLinus Torvalds } 25381da177e4SLinus Torvalds 2539112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 254073f53c4aSTejun Heo complete(&wq->first_flusher->done); 254173f53c4aSTejun Heo 254273f53c4aSTejun Heo return wait; 254383c22520SOleg Nesterov } 25441da177e4SLinus Torvalds 25450fcb78c2SRolf Eike Beer /** 25461da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25470fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25481da177e4SLinus Torvalds * 25491da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 25501da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 25511da177e4SLinus Torvalds * 2552fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2553fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 25541da177e4SLinus Torvalds */ 25557ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25561da177e4SLinus Torvalds { 255773f53c4aSTejun Heo struct wq_flusher this_flusher = { 255873f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 255973f53c4aSTejun Heo .flush_color = -1, 256073f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 256173f53c4aSTejun Heo }; 256273f53c4aSTejun Heo int next_color; 2563b1f4ec17SOleg Nesterov 25643295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25653295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 256673f53c4aSTejun Heo 256773f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 256873f53c4aSTejun Heo 256973f53c4aSTejun Heo /* 257073f53c4aSTejun Heo * Start-to-wait phase 257173f53c4aSTejun Heo */ 257273f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 257373f53c4aSTejun Heo 257473f53c4aSTejun Heo if (next_color != wq->flush_color) { 257573f53c4aSTejun Heo /* 257673f53c4aSTejun Heo * Color space is not full. The current work_color 257773f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 257873f53c4aSTejun Heo * by one. 257973f53c4aSTejun Heo */ 25806183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 258173f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 258273f53c4aSTejun Heo wq->work_color = next_color; 258373f53c4aSTejun Heo 258473f53c4aSTejun Heo if (!wq->first_flusher) { 258573f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25866183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 258773f53c4aSTejun Heo 258873f53c4aSTejun Heo wq->first_flusher = &this_flusher; 258973f53c4aSTejun Heo 2590112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 259173f53c4aSTejun Heo wq->work_color)) { 259273f53c4aSTejun Heo /* nothing to flush, done */ 259373f53c4aSTejun Heo wq->flush_color = next_color; 259473f53c4aSTejun Heo wq->first_flusher = NULL; 259573f53c4aSTejun Heo goto out_unlock; 259673f53c4aSTejun Heo } 259773f53c4aSTejun Heo } else { 259873f53c4aSTejun Heo /* wait in queue */ 25996183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 260073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2601112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 260273f53c4aSTejun Heo } 260373f53c4aSTejun Heo } else { 260473f53c4aSTejun Heo /* 260573f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 260673f53c4aSTejun Heo * The next flush completion will assign us 260773f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 260873f53c4aSTejun Heo */ 260973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 261073f53c4aSTejun Heo } 261173f53c4aSTejun Heo 261273f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 261373f53c4aSTejun Heo 261473f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 261573f53c4aSTejun Heo 261673f53c4aSTejun Heo /* 261773f53c4aSTejun Heo * Wake-up-and-cascade phase 261873f53c4aSTejun Heo * 261973f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 262073f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 262173f53c4aSTejun Heo */ 262273f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 262373f53c4aSTejun Heo return; 262473f53c4aSTejun Heo 262573f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 262673f53c4aSTejun Heo 26274ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26284ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26294ce48b37STejun Heo goto out_unlock; 26304ce48b37STejun Heo 263173f53c4aSTejun Heo wq->first_flusher = NULL; 263273f53c4aSTejun Heo 26336183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26346183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 263573f53c4aSTejun Heo 263673f53c4aSTejun Heo while (true) { 263773f53c4aSTejun Heo struct wq_flusher *next, *tmp; 263873f53c4aSTejun Heo 263973f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 264073f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 264173f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 264273f53c4aSTejun Heo break; 264373f53c4aSTejun Heo list_del_init(&next->list); 264473f53c4aSTejun Heo complete(&next->done); 264573f53c4aSTejun Heo } 264673f53c4aSTejun Heo 26476183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 264873f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 264973f53c4aSTejun Heo 265073f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 265173f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 265273f53c4aSTejun Heo 265373f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 265473f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 265573f53c4aSTejun Heo /* 265673f53c4aSTejun Heo * Assign the same color to all overflowed 265773f53c4aSTejun Heo * flushers, advance work_color and append to 265873f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 265973f53c4aSTejun Heo * phase for these overflowed flushers. 266073f53c4aSTejun Heo */ 266173f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 266273f53c4aSTejun Heo tmp->flush_color = wq->work_color; 266373f53c4aSTejun Heo 266473f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 266573f53c4aSTejun Heo 266673f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 266773f53c4aSTejun Heo &wq->flusher_queue); 2668112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 266973f53c4aSTejun Heo } 267073f53c4aSTejun Heo 267173f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26726183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 267373f53c4aSTejun Heo break; 267473f53c4aSTejun Heo } 267573f53c4aSTejun Heo 267673f53c4aSTejun Heo /* 267773f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2678112202d9STejun Heo * the new first flusher and arm pwqs. 267973f53c4aSTejun Heo */ 26806183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26816183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 268273f53c4aSTejun Heo 268373f53c4aSTejun Heo list_del_init(&next->list); 268473f53c4aSTejun Heo wq->first_flusher = next; 268573f53c4aSTejun Heo 2686112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 268773f53c4aSTejun Heo break; 268873f53c4aSTejun Heo 268973f53c4aSTejun Heo /* 269073f53c4aSTejun Heo * Meh... this color is already done, clear first 269173f53c4aSTejun Heo * flusher and repeat cascading. 269273f53c4aSTejun Heo */ 269373f53c4aSTejun Heo wq->first_flusher = NULL; 269473f53c4aSTejun Heo } 269573f53c4aSTejun Heo 269673f53c4aSTejun Heo out_unlock: 269773f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 26981da177e4SLinus Torvalds } 2699ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27001da177e4SLinus Torvalds 27019c5a2ba7STejun Heo /** 27029c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27039c5a2ba7STejun Heo * @wq: workqueue to drain 27049c5a2ba7STejun Heo * 27059c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27069c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27079c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27089c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27099c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27109c5a2ba7STejun Heo * takes too long. 27119c5a2ba7STejun Heo */ 27129c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27139c5a2ba7STejun Heo { 27149c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27159c5a2ba7STejun Heo unsigned int cpu; 27169c5a2ba7STejun Heo 27179c5a2ba7STejun Heo /* 27189c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27199c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27209c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27219c5a2ba7STejun Heo */ 2722e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 27239c5a2ba7STejun Heo if (!wq->nr_drainers++) 27249c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 2725e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 27269c5a2ba7STejun Heo reflush: 27279c5a2ba7STejun Heo flush_workqueue(wq); 27289c5a2ba7STejun Heo 2729112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 2730112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 2731fa2563e4SThomas Tuttle bool drained; 27329c5a2ba7STejun Heo 2733112202d9STejun Heo spin_lock_irq(&pwq->pool->lock); 2734112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2735112202d9STejun Heo spin_unlock_irq(&pwq->pool->lock); 2736fa2563e4SThomas Tuttle 2737fa2563e4SThomas Tuttle if (drained) 27389c5a2ba7STejun Heo continue; 27399c5a2ba7STejun Heo 27409c5a2ba7STejun Heo if (++flush_cnt == 10 || 27419c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2742044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 27439c5a2ba7STejun Heo wq->name, flush_cnt); 27449c5a2ba7STejun Heo goto reflush; 27459c5a2ba7STejun Heo } 27469c5a2ba7STejun Heo 2747e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 27489c5a2ba7STejun Heo if (!--wq->nr_drainers) 27499c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 2750e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 27519c5a2ba7STejun Heo } 27529c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27539c5a2ba7STejun Heo 2754606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2755baf59022STejun Heo { 2756baf59022STejun Heo struct worker *worker = NULL; 2757c9e7cf27STejun Heo struct worker_pool *pool; 2758112202d9STejun Heo struct pool_workqueue *pwq; 2759baf59022STejun Heo 2760baf59022STejun Heo might_sleep(); 2761c9e7cf27STejun Heo pool = get_work_pool(work); 2762c9e7cf27STejun Heo if (!pool) 2763baf59022STejun Heo return false; 2764baf59022STejun Heo 2765d565ed63STejun Heo spin_lock_irq(&pool->lock); 27660b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2767112202d9STejun Heo pwq = get_work_pwq(work); 2768112202d9STejun Heo if (pwq) { 2769112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2770baf59022STejun Heo goto already_gone; 2771606a5020STejun Heo } else { 2772c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2773baf59022STejun Heo if (!worker) 2774baf59022STejun Heo goto already_gone; 2775112202d9STejun Heo pwq = worker->current_pwq; 2776606a5020STejun Heo } 2777baf59022STejun Heo 2778112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2779d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2780baf59022STejun Heo 2781e159489bSTejun Heo /* 2782e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2783e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2784e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2785e159489bSTejun Heo * access. 2786e159489bSTejun Heo */ 2787112202d9STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER) 2788112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2789e159489bSTejun Heo else 2790112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2791112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2792e159489bSTejun Heo 2793baf59022STejun Heo return true; 2794baf59022STejun Heo already_gone: 2795d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2796baf59022STejun Heo return false; 2797baf59022STejun Heo } 2798baf59022STejun Heo 2799db700897SOleg Nesterov /** 2800401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2801401a8d04STejun Heo * @work: the work to flush 2802db700897SOleg Nesterov * 2803606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2804606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2805401a8d04STejun Heo * 2806401a8d04STejun Heo * RETURNS: 2807401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2808401a8d04STejun Heo * %false if it was already idle. 2809db700897SOleg Nesterov */ 2810401a8d04STejun Heo bool flush_work(struct work_struct *work) 2811db700897SOleg Nesterov { 2812db700897SOleg Nesterov struct wq_barrier barr; 2813db700897SOleg Nesterov 28140976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28150976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28160976dfc1SStephen Boyd 2817606a5020STejun Heo if (start_flush_work(work, &barr)) { 2818db700897SOleg Nesterov wait_for_completion(&barr.done); 2819dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2820401a8d04STejun Heo return true; 2821606a5020STejun Heo } else { 2822401a8d04STejun Heo return false; 2823db700897SOleg Nesterov } 2824606a5020STejun Heo } 2825db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2826db700897SOleg Nesterov 282736e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2828401a8d04STejun Heo { 2829bbb68dfaSTejun Heo unsigned long flags; 28301f1f642eSOleg Nesterov int ret; 28311f1f642eSOleg Nesterov 28321f1f642eSOleg Nesterov do { 2833bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2834bbb68dfaSTejun Heo /* 2835bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2836bbb68dfaSTejun Heo * would be waiting for before retrying. 2837bbb68dfaSTejun Heo */ 2838bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2839606a5020STejun Heo flush_work(work); 28401f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28411f1f642eSOleg Nesterov 2842bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2843bbb68dfaSTejun Heo mark_work_canceling(work); 2844bbb68dfaSTejun Heo local_irq_restore(flags); 2845bbb68dfaSTejun Heo 2846606a5020STejun Heo flush_work(work); 28477a22ad75STejun Heo clear_work_data(work); 28481f1f642eSOleg Nesterov return ret; 28491f1f642eSOleg Nesterov } 28501f1f642eSOleg Nesterov 28516e84d644SOleg Nesterov /** 2852401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2853401a8d04STejun Heo * @work: the work to cancel 28546e84d644SOleg Nesterov * 2855401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2856401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2857401a8d04STejun Heo * another workqueue. On return from this function, @work is 2858401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28591f1f642eSOleg Nesterov * 2860401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2861401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28626e84d644SOleg Nesterov * 2863401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28646e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2865401a8d04STejun Heo * 2866401a8d04STejun Heo * RETURNS: 2867401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28686e84d644SOleg Nesterov */ 2869401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28706e84d644SOleg Nesterov { 287136e227d2STejun Heo return __cancel_work_timer(work, false); 2872b89deed3SOleg Nesterov } 287328e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2874b89deed3SOleg Nesterov 28756e84d644SOleg Nesterov /** 2876401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2877401a8d04STejun Heo * @dwork: the delayed work to flush 28786e84d644SOleg Nesterov * 2879401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2880401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2881401a8d04STejun Heo * considers the last queueing instance of @dwork. 28821f1f642eSOleg Nesterov * 2883401a8d04STejun Heo * RETURNS: 2884401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2885401a8d04STejun Heo * %false if it was already idle. 28866e84d644SOleg Nesterov */ 2887401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2888401a8d04STejun Heo { 28898930cabaSTejun Heo local_irq_disable(); 2890401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 289160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28928930cabaSTejun Heo local_irq_enable(); 2893401a8d04STejun Heo return flush_work(&dwork->work); 2894401a8d04STejun Heo } 2895401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2896401a8d04STejun Heo 2897401a8d04STejun Heo /** 289857b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 289957b30ae7STejun Heo * @dwork: delayed_work to cancel 290009383498STejun Heo * 290157b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 290257b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 290357b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 290457b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 290557b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 290609383498STejun Heo * 290757b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 290809383498STejun Heo */ 290957b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 291009383498STejun Heo { 291157b30ae7STejun Heo unsigned long flags; 291257b30ae7STejun Heo int ret; 291357b30ae7STejun Heo 291457b30ae7STejun Heo do { 291557b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 291657b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 291757b30ae7STejun Heo 291857b30ae7STejun Heo if (unlikely(ret < 0)) 291957b30ae7STejun Heo return false; 292057b30ae7STejun Heo 29217c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29227c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 292357b30ae7STejun Heo local_irq_restore(flags); 2924c0158ca6SDan Magenheimer return ret; 292509383498STejun Heo } 292657b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 292709383498STejun Heo 292809383498STejun Heo /** 2929401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2930401a8d04STejun Heo * @dwork: the delayed work cancel 2931401a8d04STejun Heo * 2932401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2933401a8d04STejun Heo * 2934401a8d04STejun Heo * RETURNS: 2935401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2936401a8d04STejun Heo */ 2937401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29386e84d644SOleg Nesterov { 293936e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29406e84d644SOleg Nesterov } 2941f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29421da177e4SLinus Torvalds 29430fcb78c2SRolf Eike Beer /** 2944c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2945c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2946c1a220e7SZhang Rui * @work: job to be done 2947c1a220e7SZhang Rui * 2948c1a220e7SZhang Rui * This puts a job on a specific cpu 2949c1a220e7SZhang Rui */ 2950d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 2951c1a220e7SZhang Rui { 2952d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2953c1a220e7SZhang Rui } 2954c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2955c1a220e7SZhang Rui 29560fcb78c2SRolf Eike Beer /** 2957ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 29581da177e4SLinus Torvalds * @work: job to be done 29591da177e4SLinus Torvalds * 2960d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 2961d4283e93STejun Heo * %true otherwise. 296252bad64dSDavid Howells * 29630fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 29640fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 29650fcb78c2SRolf Eike Beer * workqueue otherwise. 29661da177e4SLinus Torvalds */ 2967d4283e93STejun Heo bool schedule_work(struct work_struct *work) 29681da177e4SLinus Torvalds { 29690fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 29701da177e4SLinus Torvalds } 29710fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 29721da177e4SLinus Torvalds 29730fcb78c2SRolf Eike Beer /** 29740fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 29750fcb78c2SRolf Eike Beer * @cpu: cpu to use 29760fcb78c2SRolf Eike Beer * @dwork: job to be done 29770fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 29780fcb78c2SRolf Eike Beer * 29790fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 29800fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 29810fcb78c2SRolf Eike Beer */ 2982d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 2983d4283e93STejun Heo unsigned long delay) 29841da177e4SLinus Torvalds { 2985d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 29861da177e4SLinus Torvalds } 2987ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 29881da177e4SLinus Torvalds 2989b6136773SAndrew Morton /** 29900a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 29910a13c00eSTejun Heo * @dwork: job to be done 29920a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 29930a13c00eSTejun Heo * 29940a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 29950a13c00eSTejun Heo * workqueue. 29960a13c00eSTejun Heo */ 2997d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 29980a13c00eSTejun Heo { 29990a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30000a13c00eSTejun Heo } 30010a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30020a13c00eSTejun Heo 30030a13c00eSTejun Heo /** 300431ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3005b6136773SAndrew Morton * @func: the function to call 3006b6136773SAndrew Morton * 300731ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 300831ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3009b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 301031ddd871STejun Heo * 301131ddd871STejun Heo * RETURNS: 301231ddd871STejun Heo * 0 on success, -errno on failure. 3013b6136773SAndrew Morton */ 301465f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 301515316ba8SChristoph Lameter { 301615316ba8SChristoph Lameter int cpu; 301738f51568SNamhyung Kim struct work_struct __percpu *works; 301815316ba8SChristoph Lameter 3019b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3020b6136773SAndrew Morton if (!works) 302115316ba8SChristoph Lameter return -ENOMEM; 3022b6136773SAndrew Morton 302395402b38SGautham R Shenoy get_online_cpus(); 302493981800STejun Heo 302515316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30269bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30279bfb1839SIngo Molnar 30289bfb1839SIngo Molnar INIT_WORK(work, func); 30298de6d308SOleg Nesterov schedule_work_on(cpu, work); 303015316ba8SChristoph Lameter } 303193981800STejun Heo 303293981800STejun Heo for_each_online_cpu(cpu) 30338616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 303493981800STejun Heo 303595402b38SGautham R Shenoy put_online_cpus(); 3036b6136773SAndrew Morton free_percpu(works); 303715316ba8SChristoph Lameter return 0; 303815316ba8SChristoph Lameter } 303915316ba8SChristoph Lameter 3040eef6a7d5SAlan Stern /** 3041eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3042eef6a7d5SAlan Stern * 3043eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3044eef6a7d5SAlan Stern * completion. 3045eef6a7d5SAlan Stern * 3046eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3047eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3048eef6a7d5SAlan Stern * will lead to deadlock: 3049eef6a7d5SAlan Stern * 3050eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3051eef6a7d5SAlan Stern * a lock held by your code or its caller. 3052eef6a7d5SAlan Stern * 3053eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3054eef6a7d5SAlan Stern * 3055eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3056eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3057eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3058eef6a7d5SAlan Stern * 3059eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3060eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3061eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3062eef6a7d5SAlan Stern * cancel_work_sync() instead. 3063eef6a7d5SAlan Stern */ 30641da177e4SLinus Torvalds void flush_scheduled_work(void) 30651da177e4SLinus Torvalds { 3066d320c038STejun Heo flush_workqueue(system_wq); 30671da177e4SLinus Torvalds } 3068ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30691da177e4SLinus Torvalds 30701da177e4SLinus Torvalds /** 30711fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30721fa44ecaSJames Bottomley * @fn: the function to execute 30731fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30741fa44ecaSJames Bottomley * be available when the work executes) 30751fa44ecaSJames Bottomley * 30761fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30771fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30781fa44ecaSJames Bottomley * 30791fa44ecaSJames Bottomley * Returns: 0 - function was executed 30801fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30811fa44ecaSJames Bottomley */ 308265f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30831fa44ecaSJames Bottomley { 30841fa44ecaSJames Bottomley if (!in_interrupt()) { 308565f27f38SDavid Howells fn(&ew->work); 30861fa44ecaSJames Bottomley return 0; 30871fa44ecaSJames Bottomley } 30881fa44ecaSJames Bottomley 308965f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30901fa44ecaSJames Bottomley schedule_work(&ew->work); 30911fa44ecaSJames Bottomley 30921fa44ecaSJames Bottomley return 1; 30931fa44ecaSJames Bottomley } 30941fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30951fa44ecaSJames Bottomley 30961da177e4SLinus Torvalds int keventd_up(void) 30971da177e4SLinus Torvalds { 3098d320c038STejun Heo return system_wq != NULL; 30991da177e4SLinus Torvalds } 31001da177e4SLinus Torvalds 310130cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 31021da177e4SLinus Torvalds { 310330cdf249STejun Heo int cpu; 3104e1d8aa9fSFrederic Weisbecker 310530cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 310630cdf249STejun Heo wq->pool_wq.pcpu = alloc_percpu(struct pool_workqueue); 310730cdf249STejun Heo if (!wq->pool_wq.pcpu) 310830cdf249STejun Heo return -ENOMEM; 310930cdf249STejun Heo 311030cdf249STejun Heo for_each_possible_cpu(cpu) { 311130cdf249STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 311230cdf249STejun Heo 311330cdf249STejun Heo list_add_tail(&pwq->pwqs_node, &wq->pwqs); 311430cdf249STejun Heo } 311530cdf249STejun Heo } else { 311630cdf249STejun Heo struct pool_workqueue *pwq; 311730cdf249STejun Heo 311830cdf249STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 311930cdf249STejun Heo if (!pwq) 312030cdf249STejun Heo return -ENOMEM; 312130cdf249STejun Heo 312230cdf249STejun Heo wq->pool_wq.single = pwq; 312330cdf249STejun Heo list_add_tail(&pwq->pwqs_node, &wq->pwqs); 312430cdf249STejun Heo } 312530cdf249STejun Heo 312630cdf249STejun Heo return 0; 31270f900049STejun Heo } 31280f900049STejun Heo 3129112202d9STejun Heo static void free_pwqs(struct workqueue_struct *wq) 313006ba38a9SOleg Nesterov { 3131e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3132112202d9STejun Heo free_percpu(wq->pool_wq.pcpu); 3133e904e6c2STejun Heo else 3134e904e6c2STejun Heo kmem_cache_free(pwq_cache, wq->pool_wq.single); 313506ba38a9SOleg Nesterov } 313606ba38a9SOleg Nesterov 3137f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3138f3421797STejun Heo const char *name) 3139b71ab8c2STejun Heo { 3140f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3141f3421797STejun Heo 3142f3421797STejun Heo if (max_active < 1 || max_active > lim) 3143044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3144f3421797STejun Heo max_active, name, 1, lim); 3145b71ab8c2STejun Heo 3146f3421797STejun Heo return clamp_val(max_active, 1, lim); 3147b71ab8c2STejun Heo } 3148b71ab8c2STejun Heo 3149b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 315097e37d7bSTejun Heo unsigned int flags, 31511e19ffc6STejun Heo int max_active, 3152eb13ba87SJohannes Berg struct lock_class_key *key, 3153b196be89STejun Heo const char *lock_name, ...) 31543af24433SOleg Nesterov { 3155b196be89STejun Heo va_list args, args1; 31563af24433SOleg Nesterov struct workqueue_struct *wq; 3157c34056a3STejun Heo unsigned int cpu; 3158b196be89STejun Heo size_t namelen; 3159b196be89STejun Heo 3160b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3161b196be89STejun Heo va_start(args, lock_name); 3162b196be89STejun Heo va_copy(args1, args); 3163b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3164b196be89STejun Heo 3165b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3166b196be89STejun Heo if (!wq) 3167b196be89STejun Heo goto err; 3168b196be89STejun Heo 3169b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3170b196be89STejun Heo va_end(args); 3171b196be89STejun Heo va_end(args1); 31723af24433SOleg Nesterov 3173f3421797STejun Heo /* 31746370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 31756370a6adSTejun Heo * have a rescuer to guarantee forward progress. 31766370a6adSTejun Heo */ 31776370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 31786370a6adSTejun Heo flags |= WQ_RESCUER; 31796370a6adSTejun Heo 3180d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3181b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 31823af24433SOleg Nesterov 3183b196be89STejun Heo /* init wq */ 318497e37d7bSTejun Heo wq->flags = flags; 3185a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 318673f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3187112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 318830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 318973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 319073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 31913af24433SOleg Nesterov 3192eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3193cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 31943af24433SOleg Nesterov 319530cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3196bdbc5dd7STejun Heo goto err; 3197bdbc5dd7STejun Heo 3198112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 3199112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 32001537663fSTejun Heo 3201112202d9STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3202112202d9STejun Heo pwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI); 3203112202d9STejun Heo pwq->wq = wq; 3204112202d9STejun Heo pwq->flush_color = -1; 3205112202d9STejun Heo pwq->max_active = max_active; 3206112202d9STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3207e22bee78STejun Heo } 32081537663fSTejun Heo 3209e22bee78STejun Heo if (flags & WQ_RESCUER) { 3210e22bee78STejun Heo struct worker *rescuer; 3211e22bee78STejun Heo 3212f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3213e22bee78STejun Heo goto err; 3214e22bee78STejun Heo 3215e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3216e22bee78STejun Heo if (!rescuer) 3217e22bee78STejun Heo goto err; 3218e22bee78STejun Heo 3219111c225aSTejun Heo rescuer->rescue_wq = wq; 3220111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3221b196be89STejun Heo wq->name); 3222e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3223e22bee78STejun Heo goto err; 3224e22bee78STejun Heo 3225e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3226e22bee78STejun Heo wake_up_process(rescuer->task); 32273af24433SOleg Nesterov } 32281537663fSTejun Heo 32293af24433SOleg Nesterov /* 3230a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3231a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3232a0a1a5fdSTejun Heo * workqueue to workqueues list. 32333af24433SOleg Nesterov */ 3234e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3235a0a1a5fdSTejun Heo 323658a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3237112202d9STejun Heo for_each_pwq_cpu(cpu, wq) 3238112202d9STejun Heo get_pwq(cpu, wq)->max_active = 0; 3239a0a1a5fdSTejun Heo 32403af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3241a0a1a5fdSTejun Heo 3242e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 32433af24433SOleg Nesterov 32443af24433SOleg Nesterov return wq; 32454690c4abSTejun Heo err: 32464690c4abSTejun Heo if (wq) { 3247112202d9STejun Heo free_pwqs(wq); 3248f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3249e22bee78STejun Heo kfree(wq->rescuer); 32504690c4abSTejun Heo kfree(wq); 32513af24433SOleg Nesterov } 32524690c4abSTejun Heo return NULL; 32531da177e4SLinus Torvalds } 3254d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 32551da177e4SLinus Torvalds 32563af24433SOleg Nesterov /** 32573af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 32583af24433SOleg Nesterov * @wq: target workqueue 32593af24433SOleg Nesterov * 32603af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 32613af24433SOleg Nesterov */ 32623af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 32633af24433SOleg Nesterov { 3264c8e55f36STejun Heo unsigned int cpu; 32653af24433SOleg Nesterov 32669c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 32679c5a2ba7STejun Heo drain_workqueue(wq); 3268c8efcc25STejun Heo 32696183c009STejun Heo /* sanity checks */ 32706183c009STejun Heo for_each_pwq_cpu(cpu, wq) { 32716183c009STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 32726183c009STejun Heo int i; 32736183c009STejun Heo 32746183c009STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 32756183c009STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) 32766183c009STejun Heo return; 32776183c009STejun Heo if (WARN_ON(pwq->nr_active) || 32786183c009STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) 32796183c009STejun Heo return; 32806183c009STejun Heo } 32816183c009STejun Heo 3282a0a1a5fdSTejun Heo /* 3283a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3284a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3285a0a1a5fdSTejun Heo */ 3286e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 32873af24433SOleg Nesterov list_del(&wq->list); 3288e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 32893af24433SOleg Nesterov 3290e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3291e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3292f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 32938d9df9f0SXiaotian Feng kfree(wq->rescuer); 3294e22bee78STejun Heo } 3295e22bee78STejun Heo 3296112202d9STejun Heo free_pwqs(wq); 32973af24433SOleg Nesterov kfree(wq); 32983af24433SOleg Nesterov } 32993af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33003af24433SOleg Nesterov 3301dcd989cbSTejun Heo /** 3302112202d9STejun Heo * pwq_set_max_active - adjust max_active of a pwq 3303112202d9STejun Heo * @pwq: target pool_workqueue 33049f4bd4cdSLai Jiangshan * @max_active: new max_active value. 33059f4bd4cdSLai Jiangshan * 3306112202d9STejun Heo * Set @pwq->max_active to @max_active and activate delayed works if 33079f4bd4cdSLai Jiangshan * increased. 33089f4bd4cdSLai Jiangshan * 33099f4bd4cdSLai Jiangshan * CONTEXT: 3310d565ed63STejun Heo * spin_lock_irq(pool->lock). 33119f4bd4cdSLai Jiangshan */ 3312112202d9STejun Heo static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active) 33139f4bd4cdSLai Jiangshan { 3314112202d9STejun Heo pwq->max_active = max_active; 33159f4bd4cdSLai Jiangshan 3316112202d9STejun Heo while (!list_empty(&pwq->delayed_works) && 3317112202d9STejun Heo pwq->nr_active < pwq->max_active) 3318112202d9STejun Heo pwq_activate_first_delayed(pwq); 33199f4bd4cdSLai Jiangshan } 33209f4bd4cdSLai Jiangshan 33219f4bd4cdSLai Jiangshan /** 3322dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3323dcd989cbSTejun Heo * @wq: target workqueue 3324dcd989cbSTejun Heo * @max_active: new max_active value. 3325dcd989cbSTejun Heo * 3326dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3327dcd989cbSTejun Heo * 3328dcd989cbSTejun Heo * CONTEXT: 3329dcd989cbSTejun Heo * Don't call from IRQ context. 3330dcd989cbSTejun Heo */ 3331dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3332dcd989cbSTejun Heo { 3333dcd989cbSTejun Heo unsigned int cpu; 3334dcd989cbSTejun Heo 3335f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3336dcd989cbSTejun Heo 3337e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3338dcd989cbSTejun Heo 3339dcd989cbSTejun Heo wq->saved_max_active = max_active; 3340dcd989cbSTejun Heo 3341112202d9STejun Heo for_each_pwq_cpu(cpu, wq) { 3342112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3343112202d9STejun Heo struct worker_pool *pool = pwq->pool; 3344dcd989cbSTejun Heo 3345e98d5b16STejun Heo spin_lock(&pool->lock); 3346dcd989cbSTejun Heo 334758a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 334835b6bb63STejun Heo !(pool->flags & POOL_FREEZING)) 3349112202d9STejun Heo pwq_set_max_active(pwq, max_active); 3350dcd989cbSTejun Heo 3351e98d5b16STejun Heo spin_unlock(&pool->lock); 3352dcd989cbSTejun Heo } 3353dcd989cbSTejun Heo 3354e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3355dcd989cbSTejun Heo } 3356dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3357dcd989cbSTejun Heo 3358dcd989cbSTejun Heo /** 3359dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3360dcd989cbSTejun Heo * @cpu: CPU in question 3361dcd989cbSTejun Heo * @wq: target workqueue 3362dcd989cbSTejun Heo * 3363dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3364dcd989cbSTejun Heo * no synchronization around this function and the test result is 3365dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3366dcd989cbSTejun Heo * 3367dcd989cbSTejun Heo * RETURNS: 3368dcd989cbSTejun Heo * %true if congested, %false otherwise. 3369dcd989cbSTejun Heo */ 3370dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3371dcd989cbSTejun Heo { 3372112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3373dcd989cbSTejun Heo 3374112202d9STejun Heo return !list_empty(&pwq->delayed_works); 3375dcd989cbSTejun Heo } 3376dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3377dcd989cbSTejun Heo 3378dcd989cbSTejun Heo /** 3379dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3380dcd989cbSTejun Heo * @work: the work to be tested 3381dcd989cbSTejun Heo * 3382dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3383dcd989cbSTejun Heo * synchronization around this function and the test result is 3384dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3385dcd989cbSTejun Heo * 3386dcd989cbSTejun Heo * RETURNS: 3387dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3388dcd989cbSTejun Heo */ 3389dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3390dcd989cbSTejun Heo { 3391c9e7cf27STejun Heo struct worker_pool *pool = get_work_pool(work); 3392dcd989cbSTejun Heo unsigned long flags; 3393dcd989cbSTejun Heo unsigned int ret = 0; 3394dcd989cbSTejun Heo 3395dcd989cbSTejun Heo if (work_pending(work)) 3396dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3397038366c5SLai Jiangshan 3398038366c5SLai Jiangshan if (pool) { 3399038366c5SLai Jiangshan spin_lock_irqsave(&pool->lock, flags); 3400c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3401dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3402d565ed63STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 3403038366c5SLai Jiangshan } 3404dcd989cbSTejun Heo 3405dcd989cbSTejun Heo return ret; 3406dcd989cbSTejun Heo } 3407dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3408dcd989cbSTejun Heo 3409db7bccf4STejun Heo /* 3410db7bccf4STejun Heo * CPU hotplug. 3411db7bccf4STejun Heo * 3412e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3413112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 3414706026c2STejun Heo * pool which make migrating pending and scheduled works very 3415e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 341694cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 3417e22bee78STejun Heo * blocked draining impractical. 3418e22bee78STejun Heo * 341924647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 3420628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3421628c78e7STejun Heo * cpu comes back online. 3422db7bccf4STejun Heo */ 3423db7bccf4STejun Heo 3424706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 3425db7bccf4STejun Heo { 342638db41d9STejun Heo int cpu = smp_processor_id(); 34274ce62e9eSTejun Heo struct worker_pool *pool; 3428db7bccf4STejun Heo struct worker *worker; 3429db7bccf4STejun Heo int i; 3430db7bccf4STejun Heo 343138db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 34326183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 3433db7bccf4STejun Heo 343494cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 343594cf58bbSTejun Heo spin_lock_irq(&pool->lock); 3436e22bee78STejun Heo 3437f2d5a0eeSTejun Heo /* 343894cf58bbSTejun Heo * We've claimed all manager positions. Make all workers 343994cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 344094cf58bbSTejun Heo * except for the ones which are still executing works from 344194cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 344294cf58bbSTejun Heo * this, they may become diasporas. 3443f2d5a0eeSTejun Heo */ 34444ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3445403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3446db7bccf4STejun Heo 3447b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) 3448403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3449db7bccf4STejun Heo 345024647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 3451f2d5a0eeSTejun Heo 345294cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 345394cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 345494cf58bbSTejun Heo } 3455e22bee78STejun Heo 3456e22bee78STejun Heo /* 3457628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3458628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3459628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3460628c78e7STejun Heo */ 3461628c78e7STejun Heo schedule(); 3462628c78e7STejun Heo 3463628c78e7STejun Heo /* 3464628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3465628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 346638db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 346738db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 346838db41d9STejun Heo * pools which are served by workers tied to the CPU. 3469628c78e7STejun Heo * 3470628c78e7STejun Heo * On return from this function, the current worker would trigger 3471628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3472628c78e7STejun Heo * didn't already. 3473e22bee78STejun Heo */ 347438db41d9STejun Heo for_each_std_worker_pool(pool, cpu) 3475e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 3476db7bccf4STejun Heo } 3477db7bccf4STejun Heo 34788db25e78STejun Heo /* 34798db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 34808db25e78STejun Heo * This will be registered high priority CPU notifier. 34818db25e78STejun Heo */ 34829fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 34831da177e4SLinus Torvalds unsigned long action, 34841da177e4SLinus Torvalds void *hcpu) 34851da177e4SLinus Torvalds { 34863af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 34874ce62e9eSTejun Heo struct worker_pool *pool; 34881da177e4SLinus Torvalds 34898db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 34903af24433SOleg Nesterov case CPU_UP_PREPARE: 349138db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 34923ce63377STejun Heo struct worker *worker; 34933ce63377STejun Heo 34943ce63377STejun Heo if (pool->nr_workers) 34953ce63377STejun Heo continue; 34963ce63377STejun Heo 34973ce63377STejun Heo worker = create_worker(pool); 34983ce63377STejun Heo if (!worker) 34993ce63377STejun Heo return NOTIFY_BAD; 35003ce63377STejun Heo 3501d565ed63STejun Heo spin_lock_irq(&pool->lock); 35023ce63377STejun Heo start_worker(worker); 3503d565ed63STejun Heo spin_unlock_irq(&pool->lock); 35043af24433SOleg Nesterov } 35051da177e4SLinus Torvalds break; 35061da177e4SLinus Torvalds 350765758202STejun Heo case CPU_DOWN_FAILED: 350865758202STejun Heo case CPU_ONLINE: 350938db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 351094cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 351194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 351294cf58bbSTejun Heo 351324647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 351494cf58bbSTejun Heo rebind_workers(pool); 351594cf58bbSTejun Heo 351694cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 351794cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 351894cf58bbSTejun Heo } 35198db25e78STejun Heo break; 352065758202STejun Heo } 352165758202STejun Heo return NOTIFY_OK; 352265758202STejun Heo } 352365758202STejun Heo 352465758202STejun Heo /* 352565758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 352665758202STejun Heo * This will be registered as low priority CPU notifier. 352765758202STejun Heo */ 35289fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 352965758202STejun Heo unsigned long action, 353065758202STejun Heo void *hcpu) 353165758202STejun Heo { 35328db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 35338db25e78STejun Heo struct work_struct unbind_work; 35348db25e78STejun Heo 353565758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 353665758202STejun Heo case CPU_DOWN_PREPARE: 35378db25e78STejun Heo /* unbinding should happen on the local CPU */ 3538706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 35397635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 35408db25e78STejun Heo flush_work(&unbind_work); 35418db25e78STejun Heo break; 354265758202STejun Heo } 354365758202STejun Heo return NOTIFY_OK; 354465758202STejun Heo } 354565758202STejun Heo 35462d3854a3SRusty Russell #ifdef CONFIG_SMP 35478ccad40dSRusty Russell 35482d3854a3SRusty Russell struct work_for_cpu { 3549ed48ece2STejun Heo struct work_struct work; 35502d3854a3SRusty Russell long (*fn)(void *); 35512d3854a3SRusty Russell void *arg; 35522d3854a3SRusty Russell long ret; 35532d3854a3SRusty Russell }; 35542d3854a3SRusty Russell 3555ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 35562d3854a3SRusty Russell { 3557ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3558ed48ece2STejun Heo 35592d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 35602d3854a3SRusty Russell } 35612d3854a3SRusty Russell 35622d3854a3SRusty Russell /** 35632d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 35642d3854a3SRusty Russell * @cpu: the cpu to run on 35652d3854a3SRusty Russell * @fn: the function to run 35662d3854a3SRusty Russell * @arg: the function arg 35672d3854a3SRusty Russell * 356831ad9081SRusty Russell * This will return the value @fn returns. 356931ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 35706b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 35712d3854a3SRusty Russell */ 35722d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 35732d3854a3SRusty Russell { 3574ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 35752d3854a3SRusty Russell 3576ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3577ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3578ed48ece2STejun Heo flush_work(&wfc.work); 35792d3854a3SRusty Russell return wfc.ret; 35802d3854a3SRusty Russell } 35812d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 35822d3854a3SRusty Russell #endif /* CONFIG_SMP */ 35832d3854a3SRusty Russell 3584a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3585e7577c50SRusty Russell 3586a0a1a5fdSTejun Heo /** 3587a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3588a0a1a5fdSTejun Heo * 358958a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 359058a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 3591706026c2STejun Heo * pool->worklist. 3592a0a1a5fdSTejun Heo * 3593a0a1a5fdSTejun Heo * CONTEXT: 3594d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3595a0a1a5fdSTejun Heo */ 3596a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3597a0a1a5fdSTejun Heo { 3598a0a1a5fdSTejun Heo unsigned int cpu; 3599a0a1a5fdSTejun Heo 3600e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3601a0a1a5fdSTejun Heo 36026183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 3603a0a1a5fdSTejun Heo workqueue_freezing = true; 3604a0a1a5fdSTejun Heo 3605706026c2STejun Heo for_each_wq_cpu(cpu) { 360635b6bb63STejun Heo struct worker_pool *pool; 3607bdbc5dd7STejun Heo struct workqueue_struct *wq; 36088b03ae3cSTejun Heo 360938db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3610e98d5b16STejun Heo spin_lock(&pool->lock); 3611d565ed63STejun Heo 361235b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 361335b6bb63STejun Heo pool->flags |= POOL_FREEZING; 3614db7bccf4STejun Heo 3615a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3616112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3617a0a1a5fdSTejun Heo 3618112202d9STejun Heo if (pwq && pwq->pool == pool && 3619a1056305STejun Heo (wq->flags & WQ_FREEZABLE)) 3620112202d9STejun Heo pwq->max_active = 0; 36211da177e4SLinus Torvalds } 36228b03ae3cSTejun Heo 3623e98d5b16STejun Heo spin_unlock(&pool->lock); 3624a1056305STejun Heo } 3625a0a1a5fdSTejun Heo } 3626a0a1a5fdSTejun Heo 3627e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3628a0a1a5fdSTejun Heo } 3629a0a1a5fdSTejun Heo 3630a0a1a5fdSTejun Heo /** 363158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3632a0a1a5fdSTejun Heo * 3633a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3634a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3635a0a1a5fdSTejun Heo * 3636a0a1a5fdSTejun Heo * CONTEXT: 3637a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3638a0a1a5fdSTejun Heo * 3639a0a1a5fdSTejun Heo * RETURNS: 364058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 364158a69cb4STejun Heo * is complete. 3642a0a1a5fdSTejun Heo */ 3643a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3644a0a1a5fdSTejun Heo { 3645a0a1a5fdSTejun Heo unsigned int cpu; 3646a0a1a5fdSTejun Heo bool busy = false; 3647a0a1a5fdSTejun Heo 3648e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3649a0a1a5fdSTejun Heo 36506183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 3651a0a1a5fdSTejun Heo 3652706026c2STejun Heo for_each_wq_cpu(cpu) { 3653bdbc5dd7STejun Heo struct workqueue_struct *wq; 3654a0a1a5fdSTejun Heo /* 3655a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3656a0a1a5fdSTejun Heo * to peek without lock. 3657a0a1a5fdSTejun Heo */ 3658a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3659112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3660a0a1a5fdSTejun Heo 3661112202d9STejun Heo if (!pwq || !(wq->flags & WQ_FREEZABLE)) 3662a0a1a5fdSTejun Heo continue; 3663a0a1a5fdSTejun Heo 36646183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 3665112202d9STejun Heo if (pwq->nr_active) { 3666a0a1a5fdSTejun Heo busy = true; 3667a0a1a5fdSTejun Heo goto out_unlock; 3668a0a1a5fdSTejun Heo } 3669a0a1a5fdSTejun Heo } 3670a0a1a5fdSTejun Heo } 3671a0a1a5fdSTejun Heo out_unlock: 3672e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3673a0a1a5fdSTejun Heo return busy; 3674a0a1a5fdSTejun Heo } 3675a0a1a5fdSTejun Heo 3676a0a1a5fdSTejun Heo /** 3677a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3678a0a1a5fdSTejun Heo * 3679a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 3680706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 3681a0a1a5fdSTejun Heo * 3682a0a1a5fdSTejun Heo * CONTEXT: 3683d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 3684a0a1a5fdSTejun Heo */ 3685a0a1a5fdSTejun Heo void thaw_workqueues(void) 3686a0a1a5fdSTejun Heo { 3687a0a1a5fdSTejun Heo unsigned int cpu; 3688a0a1a5fdSTejun Heo 3689e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3690a0a1a5fdSTejun Heo 3691a0a1a5fdSTejun Heo if (!workqueue_freezing) 3692a0a1a5fdSTejun Heo goto out_unlock; 3693a0a1a5fdSTejun Heo 3694706026c2STejun Heo for_each_wq_cpu(cpu) { 36954ce62e9eSTejun Heo struct worker_pool *pool; 3696bdbc5dd7STejun Heo struct workqueue_struct *wq; 36978b03ae3cSTejun Heo 369838db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3699e98d5b16STejun Heo spin_lock(&pool->lock); 3700d565ed63STejun Heo 370135b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 370235b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 3703db7bccf4STejun Heo 3704a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3705112202d9STejun Heo struct pool_workqueue *pwq = get_pwq(cpu, wq); 3706a0a1a5fdSTejun Heo 3707112202d9STejun Heo if (!pwq || pwq->pool != pool || 3708a1056305STejun Heo !(wq->flags & WQ_FREEZABLE)) 3709a0a1a5fdSTejun Heo continue; 3710a0a1a5fdSTejun Heo 3711a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3712112202d9STejun Heo pwq_set_max_active(pwq, wq->saved_max_active); 3713a0a1a5fdSTejun Heo } 37148b03ae3cSTejun Heo 37154ce62e9eSTejun Heo wake_up_worker(pool); 3716a1056305STejun Heo 3717e98d5b16STejun Heo spin_unlock(&pool->lock); 3718d565ed63STejun Heo } 3719a0a1a5fdSTejun Heo } 3720a0a1a5fdSTejun Heo 3721a0a1a5fdSTejun Heo workqueue_freezing = false; 3722a0a1a5fdSTejun Heo out_unlock: 3723e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 3724a0a1a5fdSTejun Heo } 3725a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3726a0a1a5fdSTejun Heo 37276ee0578bSSuresh Siddha static int __init init_workqueues(void) 37281da177e4SLinus Torvalds { 3729c34056a3STejun Heo unsigned int cpu; 3730c34056a3STejun Heo 37317c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 37327c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 37336be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 3734b5490077STejun Heo 3735e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 3736e904e6c2STejun Heo 3737e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 3738e904e6c2STejun Heo 373965758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3740a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 37418b03ae3cSTejun Heo 3742706026c2STejun Heo /* initialize CPU pools */ 3743706026c2STejun Heo for_each_wq_cpu(cpu) { 37444ce62e9eSTejun Heo struct worker_pool *pool; 37458b03ae3cSTejun Heo 374638db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 3747d565ed63STejun Heo spin_lock_init(&pool->lock); 3748ec22ca5eSTejun Heo pool->cpu = cpu; 374924647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 37504ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37514ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3752c9e7cf27STejun Heo hash_init(pool->busy_hash); 3753e22bee78STejun Heo 37544ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 37554ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 37564ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3757e22bee78STejun Heo 3758706026c2STejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 37594ce62e9eSTejun Heo (unsigned long)pool); 37604ce62e9eSTejun Heo 3761b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 37624ce62e9eSTejun Heo ida_init(&pool->worker_ida); 37639daf9e67STejun Heo 37649daf9e67STejun Heo /* alloc pool ID */ 37659daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 37664ce62e9eSTejun Heo } 37678b03ae3cSTejun Heo } 37688b03ae3cSTejun Heo 3769e22bee78STejun Heo /* create the initial worker */ 3770706026c2STejun Heo for_each_online_wq_cpu(cpu) { 37714ce62e9eSTejun Heo struct worker_pool *pool; 3772e22bee78STejun Heo 377338db41d9STejun Heo for_each_std_worker_pool(pool, cpu) { 37744ce62e9eSTejun Heo struct worker *worker; 37754ce62e9eSTejun Heo 377624647570STejun Heo if (cpu != WORK_CPU_UNBOUND) 377724647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 377824647570STejun Heo 3779bc2ae0f5STejun Heo worker = create_worker(pool); 3780e22bee78STejun Heo BUG_ON(!worker); 3781d565ed63STejun Heo spin_lock_irq(&pool->lock); 3782e22bee78STejun Heo start_worker(worker); 3783d565ed63STejun Heo spin_unlock_irq(&pool->lock); 3784e22bee78STejun Heo } 37854ce62e9eSTejun Heo } 3786e22bee78STejun Heo 3787d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 37881aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3789d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3790f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3791f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 379224d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 379324d51addSTejun Heo WQ_FREEZABLE, 0); 37941aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3795ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 37966ee0578bSSuresh Siddha return 0; 37971da177e4SLinus Torvalds } 37986ee0578bSSuresh Siddha early_initcall(init_workqueues); 3799