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 /* 50bc2ae0f5STejun Heo * global_cwq flags 51bc2ae0f5STejun Heo * 52bc2ae0f5STejun Heo * A bound gcwq 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 59bc2ae0f5STejun Heo * be executing on any CPU. The gcwq behaves as an unbound one. 60bc2ae0f5STejun Heo * 61bc2ae0f5STejun Heo * Note that DISASSOCIATED can be flipped only while holding 62b2eb83d1SLai Jiangshan * assoc_mutex of all pools on the gcwq to avoid changing binding 63bc2ae0f5STejun Heo * state while create_worker() is in progress. 64bc2ae0f5STejun Heo */ 6511ebea50STejun Heo GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */ 6611ebea50STejun Heo GCWQ_FREEZING = 1 << 1, /* freeze in progress */ 6711ebea50STejun Heo 6811ebea50STejun Heo /* pool flags */ 6911ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 70552a37e9SLai Jiangshan POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */ 71db7bccf4STejun Heo 72c8e55f36STejun Heo /* worker flags */ 73c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 74c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 75c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 76e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79e22bee78STejun Heo 805f7dabfdSLai Jiangshan WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | 81403c821dSTejun Heo WORKER_CPU_INTENSIVE, 82db7bccf4STejun Heo 833270476aSTejun Heo NR_WORKER_POOLS = 2, /* # worker pools per gcwq */ 844ce62e9eSTejun Heo 85c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 86db7bccf4STejun Heo 87e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 88e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 89e22bee78STejun Heo 903233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 913233cdbdSTejun Heo /* call for help after 10ms 923233cdbdSTejun Heo (min two ticks) */ 93e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 94e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 951da177e4SLinus Torvalds 961da177e4SLinus Torvalds /* 97e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 98e22bee78STejun Heo * all cpus. Give -20. 99e22bee78STejun Heo */ 100e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1013270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 102c8e55f36STejun Heo }; 103c8e55f36STejun Heo 1041da177e4SLinus Torvalds /* 1054690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1064690c4abSTejun Heo * 107e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 108e41e704bSTejun Heo * everyone else. 1094690c4abSTejun Heo * 110e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 111e22bee78STejun Heo * only be modified and accessed from the local cpu. 112e22bee78STejun Heo * 1138b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 1144690c4abSTejun Heo * 115e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 116e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 117e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 118f3421797STejun Heo * If GCWQ_DISASSOCIATED is set, it's identical to L. 119e22bee78STejun Heo * 12073f53c4aSTejun Heo * F: wq->flush_mutex protected. 12173f53c4aSTejun Heo * 1224690c4abSTejun Heo * W: workqueue_lock protected. 1234690c4abSTejun Heo */ 1244690c4abSTejun Heo 1252eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 126c34056a3STejun Heo 127bd7bdd43STejun Heo struct worker_pool { 128bd7bdd43STejun Heo struct global_cwq *gcwq; /* I: the owning gcwq */ 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 141b2eb83d1SLai Jiangshan struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */ 142bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 143bd7bdd43STejun Heo }; 144bd7bdd43STejun Heo 1454690c4abSTejun Heo /* 146e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 147e22bee78STejun Heo * and all works are queued and processed here regardless of their 148e22bee78STejun Heo * target workqueues. 1498b03ae3cSTejun Heo */ 1508b03ae3cSTejun Heo struct global_cwq { 1518b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1528b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 153db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 154c8e55f36STejun Heo 155bd7bdd43STejun Heo /* workers are chained either in busy_hash or pool idle_list */ 15642f8570fSSasha Levin DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 157c8e55f36STejun Heo /* L: hash of busy workers */ 158c8e55f36STejun Heo 159330dad5bSJoonsoo Kim struct worker_pool pools[NR_WORKER_POOLS]; 160330dad5bSJoonsoo Kim /* normal and highpri pools */ 1618b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1628b03ae3cSTejun Heo 1638b03ae3cSTejun Heo /* 164502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1650f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1660f900049STejun Heo * aligned at two's power of the number of flag bits. 1671da177e4SLinus Torvalds */ 1681da177e4SLinus Torvalds struct cpu_workqueue_struct { 169bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1704690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 17173f53c4aSTejun Heo int work_color; /* L: current color */ 17273f53c4aSTejun Heo int flush_color; /* L: flushing color */ 17373f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 17473f53c4aSTejun Heo /* L: nr of in_flight works */ 1751e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 176a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1771e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1780f900049STejun Heo }; 1791da177e4SLinus Torvalds 1801da177e4SLinus Torvalds /* 18173f53c4aSTejun Heo * Structure used to wait for workqueue flush. 18273f53c4aSTejun Heo */ 18373f53c4aSTejun Heo struct wq_flusher { 18473f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 18573f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 18673f53c4aSTejun Heo struct completion done; /* flush completion */ 18773f53c4aSTejun Heo }; 1881da177e4SLinus Torvalds 18973f53c4aSTejun Heo /* 190f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 191f2e005aaSTejun Heo * used to determine whether there's something to be done. 192f2e005aaSTejun Heo */ 193f2e005aaSTejun Heo #ifdef CONFIG_SMP 194f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 195f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 196f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 197f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 198f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 1999c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 200f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 201f2e005aaSTejun Heo #else 202f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 203f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 204f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 205f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 206f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 207f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 208f2e005aaSTejun Heo #endif 2091da177e4SLinus Torvalds 2101da177e4SLinus Torvalds /* 2111da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2121da177e4SLinus Torvalds * per-CPU workqueues: 2131da177e4SLinus Torvalds */ 2141da177e4SLinus Torvalds struct workqueue_struct { 2159c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 216bdbc5dd7STejun Heo union { 217bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 218bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 219bdbc5dd7STejun Heo unsigned long v; 220bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2214690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 22273f53c4aSTejun Heo 22373f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 22473f53c4aSTejun Heo int work_color; /* F: current work color */ 22573f53c4aSTejun Heo int flush_color; /* F: current flush color */ 22673f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 22773f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 22873f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 22973f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 23073f53c4aSTejun Heo 231f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 232e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 233e22bee78STejun Heo 2349c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 235dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2364e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2374e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2384e6045f1SJohannes Berg #endif 239b196be89STejun Heo char name[]; /* I: workqueue name */ 2401da177e4SLinus Torvalds }; 2411da177e4SLinus Torvalds 242d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 243d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 244044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2451aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 246044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 247d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 248044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 249f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 250044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 25124d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 252d320c038STejun Heo 25397bd2347STejun Heo #define CREATE_TRACE_POINTS 25497bd2347STejun Heo #include <trace/events/workqueue.h> 25597bd2347STejun Heo 2564ce62e9eSTejun Heo #define for_each_worker_pool(pool, gcwq) \ 2573270476aSTejun Heo for ((pool) = &(gcwq)->pools[0]; \ 2583270476aSTejun Heo (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++) 2594ce62e9eSTejun Heo 260db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 26142f8570fSSasha Levin hash_for_each(gcwq->busy_hash, i, pos, worker, hentry) 262db7bccf4STejun Heo 263f3421797STejun Heo static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, 264f3421797STejun Heo unsigned int sw) 265f3421797STejun Heo { 266f3421797STejun Heo if (cpu < nr_cpu_ids) { 267f3421797STejun Heo if (sw & 1) { 268f3421797STejun Heo cpu = cpumask_next(cpu, mask); 269f3421797STejun Heo if (cpu < nr_cpu_ids) 270f3421797STejun Heo return cpu; 271f3421797STejun Heo } 272f3421797STejun Heo if (sw & 2) 273f3421797STejun Heo return WORK_CPU_UNBOUND; 274f3421797STejun Heo } 275f3421797STejun Heo return WORK_CPU_NONE; 276f3421797STejun Heo } 277f3421797STejun Heo 278f3421797STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 279f3421797STejun Heo struct workqueue_struct *wq) 280f3421797STejun Heo { 281f3421797STejun Heo return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 282f3421797STejun Heo } 283f3421797STejun Heo 28409884951STejun Heo /* 28509884951STejun Heo * CPU iterators 28609884951STejun Heo * 28709884951STejun Heo * An extra gcwq is defined for an invalid cpu number 28809884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 28909884951STejun Heo * specific CPU. The following iterators are similar to 29009884951STejun Heo * for_each_*_cpu() iterators but also considers the unbound gcwq. 29109884951STejun Heo * 29209884951STejun Heo * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND 29309884951STejun Heo * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND 29409884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 29509884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 29609884951STejun Heo */ 297f3421797STejun Heo #define for_each_gcwq_cpu(cpu) \ 298f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ 299f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 300f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) 301f3421797STejun Heo 302f3421797STejun Heo #define for_each_online_gcwq_cpu(cpu) \ 303f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ 304f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 305f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) 306f3421797STejun Heo 307f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 308f3421797STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ 309f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 310f3421797STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) 311f3421797STejun Heo 312dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 313dc186ad7SThomas Gleixner 314dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 315dc186ad7SThomas Gleixner 31699777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 31799777288SStanislaw Gruszka { 31899777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 31999777288SStanislaw Gruszka } 32099777288SStanislaw Gruszka 321dc186ad7SThomas Gleixner /* 322dc186ad7SThomas Gleixner * fixup_init is called when: 323dc186ad7SThomas Gleixner * - an active object is initialized 324dc186ad7SThomas Gleixner */ 325dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 326dc186ad7SThomas Gleixner { 327dc186ad7SThomas Gleixner struct work_struct *work = addr; 328dc186ad7SThomas Gleixner 329dc186ad7SThomas Gleixner switch (state) { 330dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 331dc186ad7SThomas Gleixner cancel_work_sync(work); 332dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 333dc186ad7SThomas Gleixner return 1; 334dc186ad7SThomas Gleixner default: 335dc186ad7SThomas Gleixner return 0; 336dc186ad7SThomas Gleixner } 337dc186ad7SThomas Gleixner } 338dc186ad7SThomas Gleixner 339dc186ad7SThomas Gleixner /* 340dc186ad7SThomas Gleixner * fixup_activate is called when: 341dc186ad7SThomas Gleixner * - an active object is activated 342dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 343dc186ad7SThomas Gleixner */ 344dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 345dc186ad7SThomas Gleixner { 346dc186ad7SThomas Gleixner struct work_struct *work = addr; 347dc186ad7SThomas Gleixner 348dc186ad7SThomas Gleixner switch (state) { 349dc186ad7SThomas Gleixner 350dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 351dc186ad7SThomas Gleixner /* 352dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 353dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 354dc186ad7SThomas Gleixner * is tracked in the object tracker. 355dc186ad7SThomas Gleixner */ 35622df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 357dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 358dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 359dc186ad7SThomas Gleixner return 0; 360dc186ad7SThomas Gleixner } 361dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 362dc186ad7SThomas Gleixner return 0; 363dc186ad7SThomas Gleixner 364dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 365dc186ad7SThomas Gleixner WARN_ON(1); 366dc186ad7SThomas Gleixner 367dc186ad7SThomas Gleixner default: 368dc186ad7SThomas Gleixner return 0; 369dc186ad7SThomas Gleixner } 370dc186ad7SThomas Gleixner } 371dc186ad7SThomas Gleixner 372dc186ad7SThomas Gleixner /* 373dc186ad7SThomas Gleixner * fixup_free is called when: 374dc186ad7SThomas Gleixner * - an active object is freed 375dc186ad7SThomas Gleixner */ 376dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 377dc186ad7SThomas Gleixner { 378dc186ad7SThomas Gleixner struct work_struct *work = addr; 379dc186ad7SThomas Gleixner 380dc186ad7SThomas Gleixner switch (state) { 381dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 382dc186ad7SThomas Gleixner cancel_work_sync(work); 383dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 384dc186ad7SThomas Gleixner return 1; 385dc186ad7SThomas Gleixner default: 386dc186ad7SThomas Gleixner return 0; 387dc186ad7SThomas Gleixner } 388dc186ad7SThomas Gleixner } 389dc186ad7SThomas Gleixner 390dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 391dc186ad7SThomas Gleixner .name = "work_struct", 39299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 393dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 394dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 395dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 396dc186ad7SThomas Gleixner }; 397dc186ad7SThomas Gleixner 398dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 399dc186ad7SThomas Gleixner { 400dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 401dc186ad7SThomas Gleixner } 402dc186ad7SThomas Gleixner 403dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 404dc186ad7SThomas Gleixner { 405dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 406dc186ad7SThomas Gleixner } 407dc186ad7SThomas Gleixner 408dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 409dc186ad7SThomas Gleixner { 410dc186ad7SThomas Gleixner if (onstack) 411dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 412dc186ad7SThomas Gleixner else 413dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 414dc186ad7SThomas Gleixner } 415dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 416dc186ad7SThomas Gleixner 417dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 418dc186ad7SThomas Gleixner { 419dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 420dc186ad7SThomas Gleixner } 421dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 422dc186ad7SThomas Gleixner 423dc186ad7SThomas Gleixner #else 424dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 425dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 426dc186ad7SThomas Gleixner #endif 427dc186ad7SThomas Gleixner 42895402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 42995402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4301da177e4SLinus Torvalds static LIST_HEAD(workqueues); 431a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4321da177e4SLinus Torvalds 43314441960SOleg Nesterov /* 434e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 435e22bee78STejun Heo * which is expected to be used frequently by other cpus via 436e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 43714441960SOleg Nesterov */ 4388b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 4394ce62e9eSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]); 440f756d5e2SNathan Lynch 441f3421797STejun Heo /* 442f3421797STejun Heo * Global cpu workqueue and nr_running counter for unbound gcwq. The 443f3421797STejun Heo * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its 444f3421797STejun Heo * workers have WORKER_UNBOUND set. 445f3421797STejun Heo */ 446f3421797STejun Heo static struct global_cwq unbound_global_cwq; 4474ce62e9eSTejun Heo static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = { 4484ce62e9eSTejun Heo [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */ 4494ce62e9eSTejun Heo }; 450f3421797STejun Heo 451c34056a3STejun Heo static int worker_thread(void *__worker); 4521da177e4SLinus Torvalds 4533270476aSTejun Heo static int worker_pool_pri(struct worker_pool *pool) 4543270476aSTejun Heo { 4553270476aSTejun Heo return pool - pool->gcwq->pools; 4563270476aSTejun Heo } 4573270476aSTejun Heo 4588b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 4591da177e4SLinus Torvalds { 460f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 4618b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 462f3421797STejun Heo else 463f3421797STejun Heo return &unbound_global_cwq; 4641da177e4SLinus Torvalds } 4651da177e4SLinus Torvalds 46663d95a91STejun Heo static atomic_t *get_pool_nr_running(struct worker_pool *pool) 467b1f4ec17SOleg Nesterov { 46863d95a91STejun Heo int cpu = pool->gcwq->cpu; 4693270476aSTejun Heo int idx = worker_pool_pri(pool); 47063d95a91STejun Heo 471f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 4724ce62e9eSTejun Heo return &per_cpu(pool_nr_running, cpu)[idx]; 473f3421797STejun Heo else 4744ce62e9eSTejun Heo return &unbound_pool_nr_running[idx]; 475b1f4ec17SOleg Nesterov } 476b1f4ec17SOleg Nesterov 4774690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 4784690c4abSTejun Heo struct workqueue_struct *wq) 479a848e3b6SOleg Nesterov { 480f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 481e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 482bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 483f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 484f3421797STejun Heo return wq->cpu_wq.single; 485f3421797STejun Heo return NULL; 486f3421797STejun Heo } 487a848e3b6SOleg Nesterov 48873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 48973f53c4aSTejun Heo { 49073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 49173f53c4aSTejun Heo } 49273f53c4aSTejun Heo 49373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 49473f53c4aSTejun Heo { 49573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 49673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 49773f53c4aSTejun Heo } 49873f53c4aSTejun Heo 49973f53c4aSTejun Heo static int work_next_color(int color) 50073f53c4aSTejun Heo { 50173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 5021da177e4SLinus Torvalds } 5031da177e4SLinus Torvalds 5044594bf15SDavid Howells /* 505b5490077STejun Heo * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data 506b5490077STejun Heo * contain the pointer to the queued cwq. Once execution starts, the flag 507b5490077STejun Heo * is cleared and the high bits contain OFFQ flags and CPU number. 5087a22ad75STejun Heo * 509bbb68dfaSTejun Heo * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling() 510bbb68dfaSTejun Heo * and clear_work_data() can be used to set the cwq, cpu or clear 511bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 512bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5137a22ad75STejun Heo * 514bbb68dfaSTejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to 515bbb68dfaSTejun Heo * a work. gcwq is available once the work has been queued anywhere after 516bbb68dfaSTejun Heo * initialization until it is sync canceled. cwq is available only while 517bbb68dfaSTejun Heo * the work item is queued. 518bbb68dfaSTejun Heo * 519bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 520bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 521bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 522bbb68dfaSTejun Heo * try to steal the PENDING bit. 5234594bf15SDavid Howells */ 5247a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5257a22ad75STejun Heo unsigned long flags) 5267a22ad75STejun Heo { 5277a22ad75STejun Heo BUG_ON(!work_pending(work)); 5287a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5297a22ad75STejun Heo } 5307a22ad75STejun Heo 5317a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5324690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5334690c4abSTejun Heo unsigned long extra_flags) 534365970a1SDavid Howells { 5357a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 536e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 537365970a1SDavid Howells } 538365970a1SDavid Howells 5398930cabaSTejun Heo static void set_work_cpu_and_clear_pending(struct work_struct *work, 5408930cabaSTejun Heo unsigned int cpu) 5414d707b9fSOleg Nesterov { 54223657bb1STejun Heo /* 54323657bb1STejun Heo * The following wmb is paired with the implied mb in 54423657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 54523657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 54623657bb1STejun Heo * owner. 54723657bb1STejun Heo */ 54823657bb1STejun Heo smp_wmb(); 549b5490077STejun Heo set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0); 5504d707b9fSOleg Nesterov } 5514d707b9fSOleg Nesterov 5527a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 553365970a1SDavid Howells { 55423657bb1STejun Heo smp_wmb(); /* see set_work_cpu_and_clear_pending() */ 5557a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 5567a22ad75STejun Heo } 5577a22ad75STejun Heo 5587a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5597a22ad75STejun Heo { 560e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5617a22ad75STejun Heo 562e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 563e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 564e120153dSTejun Heo else 565e120153dSTejun Heo return NULL; 5667a22ad75STejun Heo } 5677a22ad75STejun Heo 5687a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 5697a22ad75STejun Heo { 570e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5717a22ad75STejun Heo unsigned int cpu; 5727a22ad75STejun Heo 573e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 574e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 575bd7bdd43STejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq; 5767a22ad75STejun Heo 577b5490077STejun Heo cpu = data >> WORK_OFFQ_CPU_SHIFT; 578bdbc5dd7STejun Heo if (cpu == WORK_CPU_NONE) 5797a22ad75STejun Heo return NULL; 5807a22ad75STejun Heo 581f3421797STejun Heo BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); 5827a22ad75STejun Heo return get_gcwq(cpu); 583365970a1SDavid Howells } 584365970a1SDavid Howells 585bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 586bbb68dfaSTejun Heo { 587bbb68dfaSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 588bbb68dfaSTejun Heo unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE; 589bbb68dfaSTejun Heo 590bbb68dfaSTejun Heo set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING, 591bbb68dfaSTejun Heo WORK_STRUCT_PENDING); 592bbb68dfaSTejun Heo } 593bbb68dfaSTejun Heo 594bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 595bbb68dfaSTejun Heo { 596bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 597bbb68dfaSTejun Heo 598bbb68dfaSTejun Heo return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING); 599bbb68dfaSTejun Heo } 600bbb68dfaSTejun Heo 601e22bee78STejun Heo /* 6023270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6033270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 6043270476aSTejun Heo * they're being called with gcwq->lock held. 605e22bee78STejun Heo */ 606e22bee78STejun Heo 60763d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 608649027d7STejun Heo { 6093270476aSTejun Heo return !atomic_read(get_pool_nr_running(pool)); 610649027d7STejun Heo } 611649027d7STejun Heo 612e22bee78STejun Heo /* 613e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 614e22bee78STejun Heo * running workers. 615974271c4STejun Heo * 616974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 617974271c4STejun Heo * function will always return %true for unbound gcwq as long as the 618974271c4STejun Heo * worklist isn't empty. 619e22bee78STejun Heo */ 62063d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 621e22bee78STejun Heo { 62263d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 623e22bee78STejun Heo } 624e22bee78STejun Heo 625e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 62663d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 627e22bee78STejun Heo { 62863d95a91STejun Heo return pool->nr_idle; 629e22bee78STejun Heo } 630e22bee78STejun Heo 631e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 63263d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 633e22bee78STejun Heo { 63463d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 635e22bee78STejun Heo 6363270476aSTejun Heo return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1; 637e22bee78STejun Heo } 638e22bee78STejun Heo 639e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 64063d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 641e22bee78STejun Heo { 64263d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 643e22bee78STejun Heo } 644e22bee78STejun Heo 645e22bee78STejun Heo /* Do I need to be the manager? */ 64663d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 647e22bee78STejun Heo { 64863d95a91STejun Heo return need_to_create_worker(pool) || 64911ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 650e22bee78STejun Heo } 651e22bee78STejun Heo 652e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 65363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 654e22bee78STejun Heo { 655552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 65663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 65763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 658e22bee78STejun Heo 659ea1abd61SLai Jiangshan /* 660ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 661ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 662ea1abd61SLai Jiangshan */ 663ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 664ea1abd61SLai Jiangshan return false; 665ea1abd61SLai Jiangshan 666e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 667e22bee78STejun Heo } 668e22bee78STejun Heo 669e22bee78STejun Heo /* 670e22bee78STejun Heo * Wake up functions. 671e22bee78STejun Heo */ 672e22bee78STejun Heo 6737e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 67463d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 6757e11629dSTejun Heo { 67663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 6777e11629dSTejun Heo return NULL; 6787e11629dSTejun Heo 67963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 6807e11629dSTejun Heo } 6817e11629dSTejun Heo 6827e11629dSTejun Heo /** 6837e11629dSTejun Heo * wake_up_worker - wake up an idle worker 68463d95a91STejun Heo * @pool: worker pool to wake worker from 6857e11629dSTejun Heo * 68663d95a91STejun Heo * Wake up the first idle worker of @pool. 6877e11629dSTejun Heo * 6887e11629dSTejun Heo * CONTEXT: 6897e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 6907e11629dSTejun Heo */ 69163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 6927e11629dSTejun Heo { 69363d95a91STejun Heo struct worker *worker = first_worker(pool); 6947e11629dSTejun Heo 6957e11629dSTejun Heo if (likely(worker)) 6967e11629dSTejun Heo wake_up_process(worker->task); 6977e11629dSTejun Heo } 6987e11629dSTejun Heo 6994690c4abSTejun Heo /** 700e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 701e22bee78STejun Heo * @task: task waking up 702e22bee78STejun Heo * @cpu: CPU @task is waking up to 703e22bee78STejun Heo * 704e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 705e22bee78STejun Heo * being awoken. 706e22bee78STejun Heo * 707e22bee78STejun Heo * CONTEXT: 708e22bee78STejun Heo * spin_lock_irq(rq->lock) 709e22bee78STejun Heo */ 710e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 711e22bee78STejun Heo { 712e22bee78STejun Heo struct worker *worker = kthread_data(task); 713e22bee78STejun Heo 71436576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 71536576000SJoonsoo Kim WARN_ON_ONCE(worker->pool->gcwq->cpu != cpu); 71663d95a91STejun Heo atomic_inc(get_pool_nr_running(worker->pool)); 717e22bee78STejun Heo } 71836576000SJoonsoo Kim } 719e22bee78STejun Heo 720e22bee78STejun Heo /** 721e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 722e22bee78STejun Heo * @task: task going to sleep 723e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 724e22bee78STejun Heo * 725e22bee78STejun Heo * This function is called during schedule() when a busy worker is 726e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 727e22bee78STejun Heo * returning pointer to its task. 728e22bee78STejun Heo * 729e22bee78STejun Heo * CONTEXT: 730e22bee78STejun Heo * spin_lock_irq(rq->lock) 731e22bee78STejun Heo * 732e22bee78STejun Heo * RETURNS: 733e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 734e22bee78STejun Heo */ 735e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 736e22bee78STejun Heo unsigned int cpu) 737e22bee78STejun Heo { 738e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 739111c225aSTejun Heo struct worker_pool *pool; 740111c225aSTejun Heo atomic_t *nr_running; 741e22bee78STejun Heo 742111c225aSTejun Heo /* 743111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 744111c225aSTejun Heo * workers, also reach here, let's not access anything before 745111c225aSTejun Heo * checking NOT_RUNNING. 746111c225aSTejun Heo */ 7472d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 748e22bee78STejun Heo return NULL; 749e22bee78STejun Heo 750111c225aSTejun Heo pool = worker->pool; 751111c225aSTejun Heo nr_running = get_pool_nr_running(pool); 752111c225aSTejun Heo 753e22bee78STejun Heo /* this can only happen on the local cpu */ 754e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 755e22bee78STejun Heo 756e22bee78STejun Heo /* 757e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 758e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 759e22bee78STejun Heo * Please read comment there. 760e22bee78STejun Heo * 761628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 762628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 763628c78e7STejun Heo * disabled, which in turn means that none else could be 764628c78e7STejun Heo * manipulating idle_list, so dereferencing idle_list without gcwq 765628c78e7STejun Heo * lock is safe. 766e22bee78STejun Heo */ 767bd7bdd43STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist)) 76863d95a91STejun Heo to_wakeup = first_worker(pool); 769e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 770e22bee78STejun Heo } 771e22bee78STejun Heo 772e22bee78STejun Heo /** 773e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 774cb444766STejun Heo * @worker: self 775d302f017STejun Heo * @flags: flags to set 776d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 777d302f017STejun Heo * 778e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 779e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 780e22bee78STejun Heo * woken up. 781d302f017STejun Heo * 782cb444766STejun Heo * CONTEXT: 783cb444766STejun Heo * spin_lock_irq(gcwq->lock) 784d302f017STejun Heo */ 785d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 786d302f017STejun Heo bool wakeup) 787d302f017STejun Heo { 788bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 789e22bee78STejun Heo 790cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 791cb444766STejun Heo 792e22bee78STejun Heo /* 793e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 794e22bee78STejun Heo * wake up an idle worker as necessary if requested by 795e22bee78STejun Heo * @wakeup. 796e22bee78STejun Heo */ 797e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 798e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 79963d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 800e22bee78STejun Heo 801e22bee78STejun Heo if (wakeup) { 802e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 803bd7bdd43STejun Heo !list_empty(&pool->worklist)) 80463d95a91STejun Heo wake_up_worker(pool); 805e22bee78STejun Heo } else 806e22bee78STejun Heo atomic_dec(nr_running); 807e22bee78STejun Heo } 808e22bee78STejun Heo 809d302f017STejun Heo worker->flags |= flags; 810d302f017STejun Heo } 811d302f017STejun Heo 812d302f017STejun Heo /** 813e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 814cb444766STejun Heo * @worker: self 815d302f017STejun Heo * @flags: flags to clear 816d302f017STejun Heo * 817e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 818d302f017STejun Heo * 819cb444766STejun Heo * CONTEXT: 820cb444766STejun Heo * spin_lock_irq(gcwq->lock) 821d302f017STejun Heo */ 822d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 823d302f017STejun Heo { 82463d95a91STejun Heo struct worker_pool *pool = worker->pool; 825e22bee78STejun Heo unsigned int oflags = worker->flags; 826e22bee78STejun Heo 827cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 828cb444766STejun Heo 829d302f017STejun Heo worker->flags &= ~flags; 830e22bee78STejun Heo 83142c025f3STejun Heo /* 83242c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 83342c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 83442c025f3STejun Heo * of multiple flags, not a single flag. 83542c025f3STejun Heo */ 836e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 837e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 83863d95a91STejun Heo atomic_inc(get_pool_nr_running(pool)); 839d302f017STejun Heo } 840d302f017STejun Heo 841d302f017STejun Heo /** 8428cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 8438cca0eeaSTejun Heo * @gcwq: gcwq of interest 8448cca0eeaSTejun Heo * @work: work to find worker for 8458cca0eeaSTejun Heo * 846a2c1c57bSTejun Heo * Find a worker which is executing @work on @gcwq by searching 847a2c1c57bSTejun Heo * @gcwq->busy_hash which is keyed by the address of @work. For a worker 848a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 849a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 850a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 851a2c1c57bSTejun Heo * being executed. 852a2c1c57bSTejun Heo * 853a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 854a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 855a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 856a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 857a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 858a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 859a2c1c57bSTejun Heo * 860a2c1c57bSTejun Heo * This function checks the work item address, work function and workqueue 861a2c1c57bSTejun Heo * to avoid false positives. Note that this isn't complete as one may 862a2c1c57bSTejun Heo * construct a work function which can introduce dependency onto itself 863a2c1c57bSTejun Heo * through a recycled work item. Well, if somebody wants to shoot oneself 864a2c1c57bSTejun Heo * in the foot that badly, there's only so much we can do, and if such 865a2c1c57bSTejun Heo * deadlock actually occurs, it should be easy to locate the culprit work 866a2c1c57bSTejun Heo * function. 8678cca0eeaSTejun Heo * 8688cca0eeaSTejun Heo * CONTEXT: 8698cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8708cca0eeaSTejun Heo * 8718cca0eeaSTejun Heo * RETURNS: 8728cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8738cca0eeaSTejun Heo * otherwise. 8748cca0eeaSTejun Heo */ 8758cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 8768cca0eeaSTejun Heo struct work_struct *work) 8778cca0eeaSTejun Heo { 87842f8570fSSasha Levin struct worker *worker; 87942f8570fSSasha Levin struct hlist_node *tmp; 88042f8570fSSasha Levin 881a2c1c57bSTejun Heo hash_for_each_possible(gcwq->busy_hash, worker, tmp, hentry, 882a2c1c57bSTejun Heo (unsigned long)work) 883a2c1c57bSTejun Heo if (worker->current_work == work && 884a2c1c57bSTejun Heo worker->current_func == work->func) 88542f8570fSSasha Levin return worker; 88642f8570fSSasha Levin 88742f8570fSSasha Levin return NULL; 8888cca0eeaSTejun Heo } 8898cca0eeaSTejun Heo 8908cca0eeaSTejun Heo /** 891bf4ede01STejun Heo * move_linked_works - move linked works to a list 892bf4ede01STejun Heo * @work: start of series of works to be scheduled 893bf4ede01STejun Heo * @head: target list to append @work to 894bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 895bf4ede01STejun Heo * 896bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 897bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 898bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 899bf4ede01STejun Heo * 900bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 901bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 902bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 903bf4ede01STejun Heo * 904bf4ede01STejun Heo * CONTEXT: 905bf4ede01STejun Heo * spin_lock_irq(gcwq->lock). 906bf4ede01STejun Heo */ 907bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 908bf4ede01STejun Heo struct work_struct **nextp) 909bf4ede01STejun Heo { 910bf4ede01STejun Heo struct work_struct *n; 911bf4ede01STejun Heo 912bf4ede01STejun Heo /* 913bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 914bf4ede01STejun Heo * use NULL for list head. 915bf4ede01STejun Heo */ 916bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 917bf4ede01STejun Heo list_move_tail(&work->entry, head); 918bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 919bf4ede01STejun Heo break; 920bf4ede01STejun Heo } 921bf4ede01STejun Heo 922bf4ede01STejun Heo /* 923bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 924bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 925bf4ede01STejun Heo * needs to be updated. 926bf4ede01STejun Heo */ 927bf4ede01STejun Heo if (nextp) 928bf4ede01STejun Heo *nextp = n; 929bf4ede01STejun Heo } 930bf4ede01STejun Heo 9313aa62497SLai Jiangshan static void cwq_activate_delayed_work(struct work_struct *work) 932bf4ede01STejun Heo { 9333aa62497SLai Jiangshan struct cpu_workqueue_struct *cwq = get_work_cwq(work); 934bf4ede01STejun Heo 935bf4ede01STejun Heo trace_workqueue_activate_work(work); 936bf4ede01STejun Heo move_linked_works(work, &cwq->pool->worklist, NULL); 937bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 938bf4ede01STejun Heo cwq->nr_active++; 939bf4ede01STejun Heo } 940bf4ede01STejun Heo 9413aa62497SLai Jiangshan static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 9423aa62497SLai Jiangshan { 9433aa62497SLai Jiangshan struct work_struct *work = list_first_entry(&cwq->delayed_works, 9443aa62497SLai Jiangshan struct work_struct, entry); 9453aa62497SLai Jiangshan 9463aa62497SLai Jiangshan cwq_activate_delayed_work(work); 9473aa62497SLai Jiangshan } 9483aa62497SLai Jiangshan 949bf4ede01STejun Heo /** 950bf4ede01STejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 951bf4ede01STejun Heo * @cwq: cwq of interest 952bf4ede01STejun Heo * @color: color of work which left the queue 953bf4ede01STejun Heo * 954bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 955bf4ede01STejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 956bf4ede01STejun Heo * 957bf4ede01STejun Heo * CONTEXT: 958bf4ede01STejun Heo * spin_lock_irq(gcwq->lock). 959bf4ede01STejun Heo */ 960b3f9f405SLai Jiangshan static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color) 961bf4ede01STejun Heo { 962bf4ede01STejun Heo /* ignore uncolored works */ 963bf4ede01STejun Heo if (color == WORK_NO_COLOR) 964bf4ede01STejun Heo return; 965bf4ede01STejun Heo 966bf4ede01STejun Heo cwq->nr_in_flight[color]--; 967bf4ede01STejun Heo 968bf4ede01STejun Heo cwq->nr_active--; 969bf4ede01STejun Heo if (!list_empty(&cwq->delayed_works)) { 970bf4ede01STejun Heo /* one down, submit a delayed one */ 971bf4ede01STejun Heo if (cwq->nr_active < cwq->max_active) 972bf4ede01STejun Heo cwq_activate_first_delayed(cwq); 973bf4ede01STejun Heo } 974bf4ede01STejun Heo 975bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 976bf4ede01STejun Heo if (likely(cwq->flush_color != color)) 977bf4ede01STejun Heo return; 978bf4ede01STejun Heo 979bf4ede01STejun Heo /* are there still in-flight works? */ 980bf4ede01STejun Heo if (cwq->nr_in_flight[color]) 981bf4ede01STejun Heo return; 982bf4ede01STejun Heo 983bf4ede01STejun Heo /* this cwq is done, clear flush_color */ 984bf4ede01STejun Heo cwq->flush_color = -1; 985bf4ede01STejun Heo 986bf4ede01STejun Heo /* 987bf4ede01STejun Heo * If this was the last cwq, wake up the first flusher. It 988bf4ede01STejun Heo * will handle the rest. 989bf4ede01STejun Heo */ 990bf4ede01STejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 991bf4ede01STejun Heo complete(&cwq->wq->first_flusher->done); 992bf4ede01STejun Heo } 993bf4ede01STejun Heo 99436e227d2STejun Heo /** 995bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 99636e227d2STejun Heo * @work: work item to steal 99736e227d2STejun Heo * @is_dwork: @work is a delayed_work 998bbb68dfaSTejun Heo * @flags: place to store irq state 99936e227d2STejun Heo * 100036e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 100136e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 100236e227d2STejun Heo * 100336e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 100436e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 100536e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1006bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1007bbb68dfaSTejun Heo * for arbitrarily long 100836e227d2STejun Heo * 1009bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1010e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1011e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1012e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1013bbb68dfaSTejun Heo * 1014bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1015bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1016bbb68dfaSTejun Heo * 1017e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1018bf4ede01STejun Heo */ 1019bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1020bbb68dfaSTejun Heo unsigned long *flags) 1021bf4ede01STejun Heo { 1022bf4ede01STejun Heo struct global_cwq *gcwq; 1023bf4ede01STejun Heo 1024bbb68dfaSTejun Heo local_irq_save(*flags); 1025bbb68dfaSTejun Heo 102636e227d2STejun Heo /* try to steal the timer if it exists */ 102736e227d2STejun Heo if (is_dwork) { 102836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 102936e227d2STejun Heo 1030e0aecdd8STejun Heo /* 1031e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1032e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1033e0aecdd8STejun Heo * running on the local CPU. 1034e0aecdd8STejun Heo */ 103536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 103636e227d2STejun Heo return 1; 103736e227d2STejun Heo } 103836e227d2STejun Heo 103936e227d2STejun Heo /* try to claim PENDING the normal way */ 1040bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1041bf4ede01STejun Heo return 0; 1042bf4ede01STejun Heo 1043bf4ede01STejun Heo /* 1044bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1045bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1046bf4ede01STejun Heo */ 1047bf4ede01STejun Heo gcwq = get_work_gcwq(work); 1048bf4ede01STejun Heo if (!gcwq) 1049bbb68dfaSTejun Heo goto fail; 1050bf4ede01STejun Heo 1051bbb68dfaSTejun Heo spin_lock(&gcwq->lock); 1052bf4ede01STejun Heo if (!list_empty(&work->entry)) { 1053bf4ede01STejun Heo /* 1054bf4ede01STejun Heo * This work is queued, but perhaps we locked the wrong gcwq. 1055bf4ede01STejun Heo * In that case we must see the new value after rmb(), see 1056bf4ede01STejun Heo * insert_work()->wmb(). 1057bf4ede01STejun Heo */ 1058bf4ede01STejun Heo smp_rmb(); 1059bf4ede01STejun Heo if (gcwq == get_work_gcwq(work)) { 1060bf4ede01STejun Heo debug_work_deactivate(work); 10613aa62497SLai Jiangshan 10623aa62497SLai Jiangshan /* 10633aa62497SLai Jiangshan * A delayed work item cannot be grabbed directly 10643aa62497SLai Jiangshan * because it might have linked NO_COLOR work items 10653aa62497SLai Jiangshan * which, if left on the delayed_list, will confuse 10663aa62497SLai Jiangshan * cwq->nr_active management later on and cause 10673aa62497SLai Jiangshan * stall. Make sure the work item is activated 10683aa62497SLai Jiangshan * before grabbing. 10693aa62497SLai Jiangshan */ 10703aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 10713aa62497SLai Jiangshan cwq_activate_delayed_work(work); 10723aa62497SLai Jiangshan 1073bf4ede01STejun Heo list_del_init(&work->entry); 1074bf4ede01STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 1075b3f9f405SLai Jiangshan get_work_color(work)); 107636e227d2STejun Heo 1077bbb68dfaSTejun Heo spin_unlock(&gcwq->lock); 107836e227d2STejun Heo return 1; 1079bf4ede01STejun Heo } 1080bf4ede01STejun Heo } 1081bbb68dfaSTejun Heo spin_unlock(&gcwq->lock); 1082bbb68dfaSTejun Heo fail: 1083bbb68dfaSTejun Heo local_irq_restore(*flags); 1084bbb68dfaSTejun Heo if (work_is_canceling(work)) 1085bbb68dfaSTejun Heo return -ENOENT; 1086bbb68dfaSTejun Heo cpu_relax(); 108736e227d2STejun Heo return -EAGAIN; 1088bf4ede01STejun Heo } 1089bf4ede01STejun Heo 1090bf4ede01STejun Heo /** 10917e11629dSTejun Heo * insert_work - insert a work into gcwq 10924690c4abSTejun Heo * @cwq: cwq @work belongs to 10934690c4abSTejun Heo * @work: work to insert 10944690c4abSTejun Heo * @head: insertion point 10954690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 10964690c4abSTejun Heo * 10977e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 10987e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 10994690c4abSTejun Heo * 11004690c4abSTejun Heo * CONTEXT: 11018b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 11021da177e4SLinus Torvalds */ 1103b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 11044690c4abSTejun Heo struct work_struct *work, struct list_head *head, 11054690c4abSTejun Heo unsigned int extra_flags) 1106b89deed3SOleg Nesterov { 110763d95a91STejun Heo struct worker_pool *pool = cwq->pool; 1108e1d8aa9fSFrederic Weisbecker 11094690c4abSTejun Heo /* we own @work, set data and link */ 11107a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 11114690c4abSTejun Heo 11126e84d644SOleg Nesterov /* 11136e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 11146e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 11156e84d644SOleg Nesterov */ 11166e84d644SOleg Nesterov smp_wmb(); 11174690c4abSTejun Heo 11181a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1119e22bee78STejun Heo 1120e22bee78STejun Heo /* 1121e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1122e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1123e22bee78STejun Heo * lying around lazily while there are works to be processed. 1124e22bee78STejun Heo */ 1125e22bee78STejun Heo smp_mb(); 1126e22bee78STejun Heo 112763d95a91STejun Heo if (__need_more_worker(pool)) 112863d95a91STejun Heo wake_up_worker(pool); 1129b89deed3SOleg Nesterov } 1130b89deed3SOleg Nesterov 1131c8efcc25STejun Heo /* 1132c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 1133c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 1134c8efcc25STejun Heo * cold paths. 1135c8efcc25STejun Heo */ 1136c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1137c8efcc25STejun Heo { 1138c8efcc25STejun Heo unsigned long flags; 1139c8efcc25STejun Heo unsigned int cpu; 1140c8efcc25STejun Heo 1141c8efcc25STejun Heo for_each_gcwq_cpu(cpu) { 1142c8efcc25STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 1143c8efcc25STejun Heo struct worker *worker; 1144c8efcc25STejun Heo struct hlist_node *pos; 1145c8efcc25STejun Heo int i; 1146c8efcc25STejun Heo 1147c8efcc25STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1148c8efcc25STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 1149c8efcc25STejun Heo if (worker->task != current) 1150c8efcc25STejun Heo continue; 1151c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 1152c8efcc25STejun Heo /* 1153c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 1154c8efcc25STejun Heo * is headed to the same workqueue. 1155c8efcc25STejun Heo */ 1156c8efcc25STejun Heo return worker->current_cwq->wq == wq; 1157c8efcc25STejun Heo } 1158c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 1159c8efcc25STejun Heo } 1160c8efcc25STejun Heo return false; 1161c8efcc25STejun Heo } 1162c8efcc25STejun Heo 11634690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 11641da177e4SLinus Torvalds struct work_struct *work) 11651da177e4SLinus Torvalds { 1166502ca9d8STejun Heo struct global_cwq *gcwq; 1167502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 11681e19ffc6STejun Heo struct list_head *worklist; 11698a2e8e5dSTejun Heo unsigned int work_flags; 1170b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 11718930cabaSTejun Heo 11728930cabaSTejun Heo /* 11738930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 11748930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 11758930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 11768930cabaSTejun Heo * happen with IRQ disabled. 11778930cabaSTejun Heo */ 11788930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 11791da177e4SLinus Torvalds 1180dc186ad7SThomas Gleixner debug_work_activate(work); 11811e19ffc6STejun Heo 1182c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 11839c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1184c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1185e41e704bSTejun Heo return; 1186e41e704bSTejun Heo 1187c7fc77f7STejun Heo /* determine gcwq to use */ 1188c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1189c7fc77f7STejun Heo struct global_cwq *last_gcwq; 1190c7fc77f7STejun Heo 119157469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1192f3421797STejun Heo cpu = raw_smp_processor_id(); 1193f3421797STejun Heo 119418aa9effSTejun Heo /* 1195dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1196dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1197dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1198dbf2576eSTejun Heo * non-reentrancy. 119918aa9effSTejun Heo */ 1200502ca9d8STejun Heo gcwq = get_gcwq(cpu); 1201dbf2576eSTejun Heo last_gcwq = get_work_gcwq(work); 1202dbf2576eSTejun Heo 1203dbf2576eSTejun Heo if (last_gcwq && last_gcwq != gcwq) { 120418aa9effSTejun Heo struct worker *worker; 120518aa9effSTejun Heo 12068930cabaSTejun Heo spin_lock(&last_gcwq->lock); 120718aa9effSTejun Heo 120818aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 120918aa9effSTejun Heo 121018aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 121118aa9effSTejun Heo gcwq = last_gcwq; 121218aa9effSTejun Heo else { 121318aa9effSTejun Heo /* meh... not running there, queue here */ 12148930cabaSTejun Heo spin_unlock(&last_gcwq->lock); 12158930cabaSTejun Heo spin_lock(&gcwq->lock); 121618aa9effSTejun Heo } 12178930cabaSTejun Heo } else { 12188930cabaSTejun Heo spin_lock(&gcwq->lock); 12198930cabaSTejun Heo } 1220f3421797STejun Heo } else { 1221f3421797STejun Heo gcwq = get_gcwq(WORK_CPU_UNBOUND); 12228930cabaSTejun Heo spin_lock(&gcwq->lock); 1223502ca9d8STejun Heo } 1224502ca9d8STejun Heo 1225502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 1226502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 1227b75cac93SJoonsoo Kim trace_workqueue_queue_work(req_cpu, cwq, work); 1228502ca9d8STejun Heo 1229f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 12308930cabaSTejun Heo spin_unlock(&gcwq->lock); 1231f5b2552bSDan Carpenter return; 1232f5b2552bSDan Carpenter } 12331e19ffc6STejun Heo 123473f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 12358a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 12361e19ffc6STejun Heo 12371e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1238cdadf009STejun Heo trace_workqueue_activate_work(work); 12391e19ffc6STejun Heo cwq->nr_active++; 12403270476aSTejun Heo worklist = &cwq->pool->worklist; 12418a2e8e5dSTejun Heo } else { 12428a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 12431e19ffc6STejun Heo worklist = &cwq->delayed_works; 12448a2e8e5dSTejun Heo } 12451e19ffc6STejun Heo 12468a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 12471e19ffc6STejun Heo 12488930cabaSTejun Heo spin_unlock(&gcwq->lock); 12491da177e4SLinus Torvalds } 12501da177e4SLinus Torvalds 12510fcb78c2SRolf Eike Beer /** 1252c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1253c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1254c1a220e7SZhang Rui * @wq: workqueue to use 1255c1a220e7SZhang Rui * @work: work to queue 1256c1a220e7SZhang Rui * 1257d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1258c1a220e7SZhang Rui * 1259c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1260c1a220e7SZhang Rui * can't go away. 1261c1a220e7SZhang Rui */ 1262d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1263d4283e93STejun Heo struct work_struct *work) 1264c1a220e7SZhang Rui { 1265d4283e93STejun Heo bool ret = false; 12668930cabaSTejun Heo unsigned long flags; 12678930cabaSTejun Heo 12688930cabaSTejun Heo local_irq_save(flags); 1269c1a220e7SZhang Rui 127022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 12714690c4abSTejun Heo __queue_work(cpu, wq, work); 1272d4283e93STejun Heo ret = true; 1273c1a220e7SZhang Rui } 12748930cabaSTejun Heo 12758930cabaSTejun Heo local_irq_restore(flags); 1276c1a220e7SZhang Rui return ret; 1277c1a220e7SZhang Rui } 1278c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1279c1a220e7SZhang Rui 12800a13c00eSTejun Heo /** 12810a13c00eSTejun Heo * queue_work - queue work on a workqueue 12820a13c00eSTejun Heo * @wq: workqueue to use 12830a13c00eSTejun Heo * @work: work to queue 12840a13c00eSTejun Heo * 1285d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 12860a13c00eSTejun Heo * 12870a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 12880a13c00eSTejun Heo * it can be processed by another CPU. 12890a13c00eSTejun Heo */ 1290d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 12910a13c00eSTejun Heo { 129257469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 12930a13c00eSTejun Heo } 12940a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 12950a13c00eSTejun Heo 1296d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 12971da177e4SLinus Torvalds { 129852bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 12997a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 13001da177e4SLinus Torvalds 1301e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 13021265057fSTejun Heo __queue_work(dwork->cpu, cwq->wq, &dwork->work); 13031da177e4SLinus Torvalds } 1304d8e794dfSTejun Heo EXPORT_SYMBOL_GPL(delayed_work_timer_fn); 13051da177e4SLinus Torvalds 13067beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 130752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13081da177e4SLinus Torvalds { 13097beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13107beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13117beb2edfSTejun Heo unsigned int lcpu; 13121da177e4SLinus Torvalds 13137beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13147beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1315fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1316fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13177beb2edfSTejun Heo 13188852aac2STejun Heo /* 13198852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13208852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13218852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13228852aac2STejun Heo * on that there's no such delay when @delay is 0. 13238852aac2STejun Heo */ 13248852aac2STejun Heo if (!delay) { 13258852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13268852aac2STejun Heo return; 13278852aac2STejun Heo } 13288852aac2STejun Heo 13297beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13307beb2edfSTejun Heo 13317beb2edfSTejun Heo /* 13327beb2edfSTejun Heo * This stores cwq for the moment, for the timer_fn. Note that the 13337beb2edfSTejun Heo * work's gcwq is preserved to allow reentrance detection for 13347beb2edfSTejun Heo * delayed works. 13357beb2edfSTejun Heo */ 13367beb2edfSTejun Heo if (!(wq->flags & WQ_UNBOUND)) { 13377beb2edfSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 13387beb2edfSTejun Heo 1339e42986deSJoonsoo Kim /* 1340e42986deSJoonsoo Kim * If we cannot get the last gcwq from @work directly, 1341e42986deSJoonsoo Kim * select the last CPU such that it avoids unnecessarily 1342e42986deSJoonsoo Kim * triggering non-reentrancy check in __queue_work(). 1343e42986deSJoonsoo Kim */ 1344e42986deSJoonsoo Kim lcpu = cpu; 1345e42986deSJoonsoo Kim if (gcwq) 13467beb2edfSTejun Heo lcpu = gcwq->cpu; 1347e42986deSJoonsoo Kim if (lcpu == WORK_CPU_UNBOUND) 13487beb2edfSTejun Heo lcpu = raw_smp_processor_id(); 13497beb2edfSTejun Heo } else { 13507beb2edfSTejun Heo lcpu = WORK_CPU_UNBOUND; 13511da177e4SLinus Torvalds } 13527beb2edfSTejun Heo 13537beb2edfSTejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 13547beb2edfSTejun Heo 13551265057fSTejun Heo dwork->cpu = cpu; 13567beb2edfSTejun Heo timer->expires = jiffies + delay; 13577beb2edfSTejun Heo 13587beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13597beb2edfSTejun Heo add_timer_on(timer, cpu); 13607beb2edfSTejun Heo else 13617beb2edfSTejun Heo add_timer(timer); 13627beb2edfSTejun Heo } 13631da177e4SLinus Torvalds 13640fcb78c2SRolf Eike Beer /** 13650fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13660fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13670fcb78c2SRolf Eike Beer * @wq: workqueue to use 1368af9997e4SRandy Dunlap * @dwork: work to queue 13690fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13700fcb78c2SRolf Eike Beer * 1371715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1372715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1373715f1300STejun Heo * execution. 13740fcb78c2SRolf Eike Beer */ 1375d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 137652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13777a6bc1cdSVenkatesh Pallipadi { 137852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1379d4283e93STejun Heo bool ret = false; 13808930cabaSTejun Heo unsigned long flags; 13818930cabaSTejun Heo 13828930cabaSTejun Heo /* read the comment in __queue_work() */ 13838930cabaSTejun Heo local_irq_save(flags); 13847a6bc1cdSVenkatesh Pallipadi 138522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13867beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1387d4283e93STejun Heo ret = true; 13887a6bc1cdSVenkatesh Pallipadi } 13898930cabaSTejun Heo 13908930cabaSTejun Heo local_irq_restore(flags); 13917a6bc1cdSVenkatesh Pallipadi return ret; 13927a6bc1cdSVenkatesh Pallipadi } 1393ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 13941da177e4SLinus Torvalds 1395c8e55f36STejun Heo /** 13960a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 13970a13c00eSTejun Heo * @wq: workqueue to use 13980a13c00eSTejun Heo * @dwork: delayable work to queue 13990a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 14000a13c00eSTejun Heo * 1401715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 14020a13c00eSTejun Heo */ 1403d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 14040a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 14050a13c00eSTejun Heo { 140657469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14070a13c00eSTejun Heo } 14080a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 14090a13c00eSTejun Heo 14100a13c00eSTejun Heo /** 14118376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14128376fe22STejun Heo * @cpu: CPU number to execute work on 14138376fe22STejun Heo * @wq: workqueue to use 14148376fe22STejun Heo * @dwork: work to queue 14158376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14168376fe22STejun Heo * 14178376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14188376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14198376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14208376fe22STejun Heo * current state. 14218376fe22STejun Heo * 14228376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14238376fe22STejun Heo * pending and its timer was modified. 14248376fe22STejun Heo * 1425e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14268376fe22STejun Heo * See try_to_grab_pending() for details. 14278376fe22STejun Heo */ 14288376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14298376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14308376fe22STejun Heo { 14318376fe22STejun Heo unsigned long flags; 14328376fe22STejun Heo int ret; 14338376fe22STejun Heo 14348376fe22STejun Heo do { 14358376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14368376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14378376fe22STejun Heo 14388376fe22STejun Heo if (likely(ret >= 0)) { 14398376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14408376fe22STejun Heo local_irq_restore(flags); 14418376fe22STejun Heo } 14428376fe22STejun Heo 14438376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14448376fe22STejun Heo return ret; 14458376fe22STejun Heo } 14468376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14478376fe22STejun Heo 14488376fe22STejun Heo /** 14498376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14508376fe22STejun Heo * @wq: workqueue to use 14518376fe22STejun Heo * @dwork: work to queue 14528376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14538376fe22STejun Heo * 14548376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14558376fe22STejun Heo */ 14568376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14578376fe22STejun Heo unsigned long delay) 14588376fe22STejun Heo { 14598376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14608376fe22STejun Heo } 14618376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14628376fe22STejun Heo 14638376fe22STejun Heo /** 1464c8e55f36STejun Heo * worker_enter_idle - enter idle state 1465c8e55f36STejun Heo * @worker: worker which is entering idle state 1466c8e55f36STejun Heo * 1467c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1468c8e55f36STejun Heo * necessary. 1469c8e55f36STejun Heo * 1470c8e55f36STejun Heo * LOCKING: 1471c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1472c8e55f36STejun Heo */ 1473c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14741da177e4SLinus Torvalds { 1475bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1476bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1477c8e55f36STejun Heo 1478c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1479c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1480c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1481c8e55f36STejun Heo 1482cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1483cb444766STejun Heo worker->flags |= WORKER_IDLE; 1484bd7bdd43STejun Heo pool->nr_idle++; 1485e22bee78STejun Heo worker->last_active = jiffies; 1486c8e55f36STejun Heo 1487c8e55f36STejun Heo /* idle_list is LIFO */ 1488bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1489db7bccf4STejun Heo 149063d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1491628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1492cb444766STejun Heo 1493544ecf31STejun Heo /* 1494628c78e7STejun Heo * Sanity check nr_running. Because gcwq_unbind_fn() releases 1495628c78e7STejun Heo * gcwq->lock between setting %WORKER_UNBOUND and zapping 1496628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1497628c78e7STejun Heo * unbind is not in progress. 1498544ecf31STejun Heo */ 1499628c78e7STejun Heo WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) && 1500bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 150163d95a91STejun Heo atomic_read(get_pool_nr_running(pool))); 1502c8e55f36STejun Heo } 1503c8e55f36STejun Heo 1504c8e55f36STejun Heo /** 1505c8e55f36STejun Heo * worker_leave_idle - leave idle state 1506c8e55f36STejun Heo * @worker: worker which is leaving idle state 1507c8e55f36STejun Heo * 1508c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1509c8e55f36STejun Heo * 1510c8e55f36STejun Heo * LOCKING: 1511c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1512c8e55f36STejun Heo */ 1513c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1514c8e55f36STejun Heo { 1515bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1516c8e55f36STejun Heo 1517c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1518d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1519bd7bdd43STejun Heo pool->nr_idle--; 1520c8e55f36STejun Heo list_del_init(&worker->entry); 1521c8e55f36STejun Heo } 1522c8e55f36STejun Heo 1523e22bee78STejun Heo /** 1524e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1525e22bee78STejun Heo * @worker: self 1526e22bee78STejun Heo * 1527e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1528e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1529e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1530e22bee78STejun Heo * guaranteed to execute on the cpu. 1531e22bee78STejun Heo * 1532e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1533e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1534e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1535e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1536e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1537e22bee78STejun Heo * [dis]associated in the meantime. 1538e22bee78STejun Heo * 1539f2d5a0eeSTejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies the 1540f2d5a0eeSTejun Heo * binding against %GCWQ_DISASSOCIATED which is set during 1541f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1542f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1543f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1544e22bee78STejun Heo * 1545e22bee78STejun Heo * CONTEXT: 1546e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1547e22bee78STejun Heo * held. 1548e22bee78STejun Heo * 1549e22bee78STejun Heo * RETURNS: 1550e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1551e22bee78STejun Heo * bound), %false if offline. 1552e22bee78STejun Heo */ 1553e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1554972fa1c5SNamhyung Kim __acquires(&gcwq->lock) 1555e22bee78STejun Heo { 1556bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1557e22bee78STejun Heo struct task_struct *task = worker->task; 1558e22bee78STejun Heo 1559e22bee78STejun Heo while (true) { 1560e22bee78STejun Heo /* 1561e22bee78STejun Heo * The following call may fail, succeed or succeed 1562e22bee78STejun Heo * without actually migrating the task to the cpu if 1563e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1564e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1565e22bee78STejun Heo */ 1566f3421797STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) 1567e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1568e22bee78STejun Heo 1569e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1570e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1571e22bee78STejun Heo return false; 1572e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1573e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1574e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1575e22bee78STejun Heo return true; 1576e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1577e22bee78STejun Heo 15785035b20fSTejun Heo /* 15795035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 15805035b20fSTejun Heo * and retry migration. cond_resched() is required here; 15815035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 15825035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 15835035b20fSTejun Heo */ 1584e22bee78STejun Heo cpu_relax(); 15855035b20fSTejun Heo cond_resched(); 1586e22bee78STejun Heo } 1587e22bee78STejun Heo } 1588e22bee78STejun Heo 1589e22bee78STejun Heo /* 1590ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 15915f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 159225511a47STejun Heo */ 159325511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 159425511a47STejun Heo { 159525511a47STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 159625511a47STejun Heo 15975f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 15985f7dabfdSLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 15995f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 160025511a47STejun Heo 1601ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1602ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1603ea1abd61SLai Jiangshan spin_unlock_irq(&gcwq->lock); 160425511a47STejun Heo } 160525511a47STejun Heo 160625511a47STejun Heo /* 160725511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1608403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1609403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1610403c821dSTejun Heo * executed twice without intervening cpu down. 1611e22bee78STejun Heo */ 161225511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1613e22bee78STejun Heo { 1614e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1615bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1616e22bee78STejun Heo 1617eab6d828SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 1618eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1619e22bee78STejun Heo 1620e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1621e22bee78STejun Heo } 1622e22bee78STejun Heo 162325511a47STejun Heo /** 162425511a47STejun Heo * rebind_workers - rebind all workers of a gcwq to the associated CPU 162525511a47STejun Heo * @gcwq: gcwq of interest 162625511a47STejun Heo * 162725511a47STejun Heo * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding 162825511a47STejun Heo * is different for idle and busy ones. 162925511a47STejun Heo * 1630ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1631ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1632ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1633ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 163425511a47STejun Heo * 1635ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1636ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1637ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1638ea1abd61SLai Jiangshan * rebind. 163925511a47STejun Heo * 1640ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1641ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1642ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1643ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 164425511a47STejun Heo */ 164525511a47STejun Heo static void rebind_workers(struct global_cwq *gcwq) 164625511a47STejun Heo { 164725511a47STejun Heo struct worker_pool *pool; 1648ea1abd61SLai Jiangshan struct worker *worker, *n; 164925511a47STejun Heo struct hlist_node *pos; 165025511a47STejun Heo int i; 165125511a47STejun Heo 165225511a47STejun Heo lockdep_assert_held(&gcwq->lock); 165325511a47STejun Heo 165425511a47STejun Heo for_each_worker_pool(pool, gcwq) 1655b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 165625511a47STejun Heo 16575f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 165825511a47STejun Heo for_each_worker_pool(pool, gcwq) { 1659ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1660ea1abd61SLai Jiangshan /* 1661ea1abd61SLai Jiangshan * idle workers should be off @pool->idle_list 1662ea1abd61SLai Jiangshan * until rebind is complete to avoid receiving 1663ea1abd61SLai Jiangshan * premature local wake-ups. 1664ea1abd61SLai Jiangshan */ 1665ea1abd61SLai Jiangshan list_del_init(&worker->entry); 166625511a47STejun Heo 166725511a47STejun Heo /* 16685f7dabfdSLai Jiangshan * worker_thread() will see the above dequeuing 16695f7dabfdSLai Jiangshan * and call idle_worker_rebind(). 167025511a47STejun Heo */ 167125511a47STejun Heo wake_up_process(worker->task); 167225511a47STejun Heo } 167325511a47STejun Heo } 167425511a47STejun Heo 1675ea1abd61SLai Jiangshan /* rebind busy workers */ 167625511a47STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 167725511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1678e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 167925511a47STejun Heo 168025511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 168125511a47STejun Heo work_data_bits(rebind_work))) 168225511a47STejun Heo continue; 168325511a47STejun Heo 168425511a47STejun Heo debug_work_activate(rebind_work); 168590beca5dSTejun Heo 168690beca5dSTejun Heo /* 1687e2b6a6d5SJoonsoo Kim * wq doesn't really matter but let's keep @worker->pool 1688e2b6a6d5SJoonsoo Kim * and @cwq->pool consistent for sanity. 168990beca5dSTejun Heo */ 1690e2b6a6d5SJoonsoo Kim if (worker_pool_pri(worker->pool)) 1691e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1692e2b6a6d5SJoonsoo Kim else 1693e2b6a6d5SJoonsoo Kim wq = system_wq; 1694ec58815aSTejun Heo 1695e2b6a6d5SJoonsoo Kim insert_work(get_cwq(gcwq->cpu, wq), rebind_work, 169625511a47STejun Heo worker->scheduled.next, 169725511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1698ec58815aSTejun Heo } 169925511a47STejun Heo } 170025511a47STejun Heo 1701c34056a3STejun Heo static struct worker *alloc_worker(void) 1702c34056a3STejun Heo { 1703c34056a3STejun Heo struct worker *worker; 1704c34056a3STejun Heo 1705c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1706c8e55f36STejun Heo if (worker) { 1707c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1708affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 170925511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1710e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1711e22bee78STejun Heo worker->flags = WORKER_PREP; 1712c8e55f36STejun Heo } 1713c34056a3STejun Heo return worker; 1714c34056a3STejun Heo } 1715c34056a3STejun Heo 1716c34056a3STejun Heo /** 1717c34056a3STejun Heo * create_worker - create a new workqueue worker 171863d95a91STejun Heo * @pool: pool the new worker will belong to 1719c34056a3STejun Heo * 172063d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1721c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1722c34056a3STejun Heo * destroy_worker(). 1723c34056a3STejun Heo * 1724c34056a3STejun Heo * CONTEXT: 1725c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1726c34056a3STejun Heo * 1727c34056a3STejun Heo * RETURNS: 1728c34056a3STejun Heo * Pointer to the newly created worker. 1729c34056a3STejun Heo */ 1730bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1731c34056a3STejun Heo { 173263d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 17333270476aSTejun Heo const char *pri = worker_pool_pri(pool) ? "H" : ""; 1734c34056a3STejun Heo struct worker *worker = NULL; 1735f3421797STejun Heo int id = -1; 1736c34056a3STejun Heo 17378b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1738bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 17398b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1740bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1741c34056a3STejun Heo goto fail; 17428b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1743c34056a3STejun Heo } 17448b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1745c34056a3STejun Heo 1746c34056a3STejun Heo worker = alloc_worker(); 1747c34056a3STejun Heo if (!worker) 1748c34056a3STejun Heo goto fail; 1749c34056a3STejun Heo 1750bd7bdd43STejun Heo worker->pool = pool; 1751c34056a3STejun Heo worker->id = id; 1752c34056a3STejun Heo 1753bc2ae0f5STejun Heo if (gcwq->cpu != WORK_CPU_UNBOUND) 175494dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 17553270476aSTejun Heo worker, cpu_to_node(gcwq->cpu), 17563270476aSTejun Heo "kworker/%u:%d%s", gcwq->cpu, id, pri); 1757f3421797STejun Heo else 1758f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 17593270476aSTejun Heo "kworker/u:%d%s", id, pri); 1760c34056a3STejun Heo if (IS_ERR(worker->task)) 1761c34056a3STejun Heo goto fail; 1762c34056a3STejun Heo 17633270476aSTejun Heo if (worker_pool_pri(pool)) 17643270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 17653270476aSTejun Heo 1766db7bccf4STejun Heo /* 1767bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 1768bc2ae0f5STejun Heo * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the 1769bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1770bc2ae0f5STejun Heo * above the flag definition for details. 1771bc2ae0f5STejun Heo * 1772bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1773bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1774db7bccf4STejun Heo */ 1775bc2ae0f5STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) { 17768b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1777bc2ae0f5STejun Heo } else { 1778db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1779f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1780f3421797STejun Heo } 1781c34056a3STejun Heo 1782c34056a3STejun Heo return worker; 1783c34056a3STejun Heo fail: 1784c34056a3STejun Heo if (id >= 0) { 17858b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1786bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 17878b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1788c34056a3STejun Heo } 1789c34056a3STejun Heo kfree(worker); 1790c34056a3STejun Heo return NULL; 1791c34056a3STejun Heo } 1792c34056a3STejun Heo 1793c34056a3STejun Heo /** 1794c34056a3STejun Heo * start_worker - start a newly created worker 1795c34056a3STejun Heo * @worker: worker to start 1796c34056a3STejun Heo * 1797c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1798c34056a3STejun Heo * 1799c34056a3STejun Heo * CONTEXT: 18008b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1801c34056a3STejun Heo */ 1802c34056a3STejun Heo static void start_worker(struct worker *worker) 1803c34056a3STejun Heo { 1804cb444766STejun Heo worker->flags |= WORKER_STARTED; 1805bd7bdd43STejun Heo worker->pool->nr_workers++; 1806c8e55f36STejun Heo worker_enter_idle(worker); 1807c34056a3STejun Heo wake_up_process(worker->task); 1808c34056a3STejun Heo } 1809c34056a3STejun Heo 1810c34056a3STejun Heo /** 1811c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1812c34056a3STejun Heo * @worker: worker to be destroyed 1813c34056a3STejun Heo * 1814c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1815c8e55f36STejun Heo * 1816c8e55f36STejun Heo * CONTEXT: 1817c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1818c34056a3STejun Heo */ 1819c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1820c34056a3STejun Heo { 1821bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1822bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1823c34056a3STejun Heo int id = worker->id; 1824c34056a3STejun Heo 1825c34056a3STejun Heo /* sanity check frenzy */ 1826c34056a3STejun Heo BUG_ON(worker->current_work); 1827affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1828c34056a3STejun Heo 1829c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1830bd7bdd43STejun Heo pool->nr_workers--; 1831c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1832bd7bdd43STejun Heo pool->nr_idle--; 1833c8e55f36STejun Heo 1834c8e55f36STejun Heo list_del_init(&worker->entry); 1835cb444766STejun Heo worker->flags |= WORKER_DIE; 1836c8e55f36STejun Heo 1837c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1838c8e55f36STejun Heo 1839c34056a3STejun Heo kthread_stop(worker->task); 1840c34056a3STejun Heo kfree(worker); 1841c34056a3STejun Heo 18428b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1843bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1844c34056a3STejun Heo } 1845c34056a3STejun Heo 184663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1847e22bee78STejun Heo { 184863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 184963d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1850e22bee78STejun Heo 1851e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1852e22bee78STejun Heo 185363d95a91STejun Heo if (too_many_workers(pool)) { 1854e22bee78STejun Heo struct worker *worker; 1855e22bee78STejun Heo unsigned long expires; 1856e22bee78STejun Heo 1857e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 185863d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1859e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1860e22bee78STejun Heo 1861e22bee78STejun Heo if (time_before(jiffies, expires)) 186263d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1863e22bee78STejun Heo else { 1864e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 186511ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 186663d95a91STejun Heo wake_up_worker(pool); 1867e22bee78STejun Heo } 1868e22bee78STejun Heo } 1869e22bee78STejun Heo 1870e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1871e22bee78STejun Heo } 1872e22bee78STejun Heo 1873e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1874e22bee78STejun Heo { 1875e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1876e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1877f3421797STejun Heo unsigned int cpu; 1878e22bee78STejun Heo 1879e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1880e22bee78STejun Heo return false; 1881e22bee78STejun Heo 1882e22bee78STejun Heo /* mayday mayday mayday */ 1883bd7bdd43STejun Heo cpu = cwq->pool->gcwq->cpu; 1884f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1885f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1886f3421797STejun Heo cpu = 0; 1887f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1888e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1889e22bee78STejun Heo return true; 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 189263d95a91STejun Heo static void gcwq_mayday_timeout(unsigned long __pool) 1893e22bee78STejun Heo { 189463d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 189563d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1896e22bee78STejun Heo struct work_struct *work; 1897e22bee78STejun Heo 1898e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1899e22bee78STejun Heo 190063d95a91STejun Heo if (need_to_create_worker(pool)) { 1901e22bee78STejun Heo /* 1902e22bee78STejun Heo * We've been trying to create a new worker but 1903e22bee78STejun Heo * haven't been successful. We might be hitting an 1904e22bee78STejun Heo * allocation deadlock. Send distress signals to 1905e22bee78STejun Heo * rescuers. 1906e22bee78STejun Heo */ 190763d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1908e22bee78STejun Heo send_mayday(work); 1909e22bee78STejun Heo } 1910e22bee78STejun Heo 1911e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1912e22bee78STejun Heo 191363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1914e22bee78STejun Heo } 1915e22bee78STejun Heo 1916e22bee78STejun Heo /** 1917e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 191863d95a91STejun Heo * @pool: pool to create a new worker for 1919e22bee78STejun Heo * 192063d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1921e22bee78STejun Heo * have at least one idle worker on return from this function. If 1922e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 192363d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1924e22bee78STejun Heo * possible allocation deadlock. 1925e22bee78STejun Heo * 1926e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1927e22bee78STejun Heo * may_start_working() true. 1928e22bee78STejun Heo * 1929e22bee78STejun Heo * LOCKING: 1930e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1931e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1932e22bee78STejun Heo * manager. 1933e22bee78STejun Heo * 1934e22bee78STejun Heo * RETURNS: 1935e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1936e22bee78STejun Heo * otherwise. 1937e22bee78STejun Heo */ 193863d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 193906bd6ebfSNamhyung Kim __releases(&gcwq->lock) 194006bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 1941e22bee78STejun Heo { 194263d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 194363d95a91STejun Heo 194463d95a91STejun Heo if (!need_to_create_worker(pool)) 1945e22bee78STejun Heo return false; 1946e22bee78STejun Heo restart: 19479f9c2364STejun Heo spin_unlock_irq(&gcwq->lock); 19489f9c2364STejun Heo 1949e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 195063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1951e22bee78STejun Heo 1952e22bee78STejun Heo while (true) { 1953e22bee78STejun Heo struct worker *worker; 1954e22bee78STejun Heo 1955bc2ae0f5STejun Heo worker = create_worker(pool); 1956e22bee78STejun Heo if (worker) { 195763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1958e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1959e22bee78STejun Heo start_worker(worker); 196063d95a91STejun Heo BUG_ON(need_to_create_worker(pool)); 1961e22bee78STejun Heo return true; 1962e22bee78STejun Heo } 1963e22bee78STejun Heo 196463d95a91STejun Heo if (!need_to_create_worker(pool)) 1965e22bee78STejun Heo break; 1966e22bee78STejun Heo 1967e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1968e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19699f9c2364STejun Heo 197063d95a91STejun Heo if (!need_to_create_worker(pool)) 1971e22bee78STejun Heo break; 1972e22bee78STejun Heo } 1973e22bee78STejun Heo 197463d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1975e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 197663d95a91STejun Heo if (need_to_create_worker(pool)) 1977e22bee78STejun Heo goto restart; 1978e22bee78STejun Heo return true; 1979e22bee78STejun Heo } 1980e22bee78STejun Heo 1981e22bee78STejun Heo /** 1982e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 198363d95a91STejun Heo * @pool: pool to destroy workers for 1984e22bee78STejun Heo * 198563d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1986e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1987e22bee78STejun Heo * 1988e22bee78STejun Heo * LOCKING: 1989e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1990e22bee78STejun Heo * multiple times. Called only from manager. 1991e22bee78STejun Heo * 1992e22bee78STejun Heo * RETURNS: 1993e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1994e22bee78STejun Heo * otherwise. 1995e22bee78STejun Heo */ 199663d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1997e22bee78STejun Heo { 1998e22bee78STejun Heo bool ret = false; 1999e22bee78STejun Heo 200063d95a91STejun Heo while (too_many_workers(pool)) { 2001e22bee78STejun Heo struct worker *worker; 2002e22bee78STejun Heo unsigned long expires; 2003e22bee78STejun Heo 200463d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2005e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2006e22bee78STejun Heo 2007e22bee78STejun Heo if (time_before(jiffies, expires)) { 200863d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2009e22bee78STejun Heo break; 2010e22bee78STejun Heo } 2011e22bee78STejun Heo 2012e22bee78STejun Heo destroy_worker(worker); 2013e22bee78STejun Heo ret = true; 2014e22bee78STejun Heo } 2015e22bee78STejun Heo 2016e22bee78STejun Heo return ret; 2017e22bee78STejun Heo } 2018e22bee78STejun Heo 2019e22bee78STejun Heo /** 2020e22bee78STejun Heo * manage_workers - manage worker pool 2021e22bee78STejun Heo * @worker: self 2022e22bee78STejun Heo * 2023e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 2024e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2025e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 2026e22bee78STejun Heo * 2027e22bee78STejun Heo * The caller can safely start processing works on false return. On 2028e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2029e22bee78STejun Heo * and may_start_working() is true. 2030e22bee78STejun Heo * 2031e22bee78STejun Heo * CONTEXT: 2032e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2033e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2034e22bee78STejun Heo * 2035e22bee78STejun Heo * RETURNS: 2036e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 2037e22bee78STejun Heo * some action was taken. 2038e22bee78STejun Heo */ 2039e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2040e22bee78STejun Heo { 204163d95a91STejun Heo struct worker_pool *pool = worker->pool; 2042e22bee78STejun Heo bool ret = false; 2043e22bee78STejun Heo 2044ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2045e22bee78STejun Heo return ret; 2046e22bee78STejun Heo 2047552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2048ee378aa4SLai Jiangshan 2049ee378aa4SLai Jiangshan /* 2050ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2051ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2052ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2053ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2054ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2055b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2056ee378aa4SLai Jiangshan * manager against CPU hotplug. 2057ee378aa4SLai Jiangshan * 2058b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2059ee378aa4SLai Jiangshan * progress. trylock first without dropping @gcwq->lock. 2060ee378aa4SLai Jiangshan */ 2061b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2062ee378aa4SLai Jiangshan spin_unlock_irq(&pool->gcwq->lock); 2063b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2064ee378aa4SLai Jiangshan /* 2065ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2066b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2067ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2068ee378aa4SLai Jiangshan * @gcwq's state and ours could have deviated. 2069ee378aa4SLai Jiangshan * 2070b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2071ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2072ee378aa4SLai Jiangshan * on @gcwq's current state. Try it and adjust 2073ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2074ee378aa4SLai Jiangshan */ 2075ee378aa4SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 2076ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2077ee378aa4SLai Jiangshan else 2078ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2079ee378aa4SLai Jiangshan 2080ee378aa4SLai Jiangshan ret = true; 2081ee378aa4SLai Jiangshan } 2082ee378aa4SLai Jiangshan 208311ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2084e22bee78STejun Heo 2085e22bee78STejun Heo /* 2086e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2087e22bee78STejun Heo * on return. 2088e22bee78STejun Heo */ 208963d95a91STejun Heo ret |= maybe_destroy_workers(pool); 209063d95a91STejun Heo ret |= maybe_create_worker(pool); 2091e22bee78STejun Heo 2092552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2093b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2094e22bee78STejun Heo return ret; 2095e22bee78STejun Heo } 2096e22bee78STejun Heo 2097a62428c0STejun Heo /** 2098a62428c0STejun Heo * process_one_work - process single work 2099c34056a3STejun Heo * @worker: self 2100a62428c0STejun Heo * @work: work to process 2101a62428c0STejun Heo * 2102a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2103a62428c0STejun Heo * process a single work including synchronization against and 2104a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2105a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2106a62428c0STejun Heo * call this function to process a work. 2107a62428c0STejun Heo * 2108a62428c0STejun Heo * CONTEXT: 21098b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 2110a62428c0STejun Heo */ 2111c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 211206bd6ebfSNamhyung Kim __releases(&gcwq->lock) 211306bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 21141da177e4SLinus Torvalds { 21157e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 2116bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2117bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2118fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 211973f53c4aSTejun Heo int work_color; 21207e11629dSTejun Heo struct worker *collision; 21214e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21224e6045f1SJohannes Berg /* 2123a62428c0STejun Heo * It is permissible to free the struct work_struct from 2124a62428c0STejun Heo * inside the function that is called from it, this we need to 2125a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2126a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2127a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21284e6045f1SJohannes Berg */ 21294d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21304d82a1deSPeter Zijlstra 21314d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21324e6045f1SJohannes Berg #endif 21336fec10a1STejun Heo /* 21346fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21356fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 21366fec10a1STejun Heo * unbound or a disassociated gcwq. 21376fec10a1STejun Heo */ 21385f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 21396fec10a1STejun Heo !(gcwq->flags & GCWQ_DISASSOCIATED) && 214025511a47STejun Heo raw_smp_processor_id() != gcwq->cpu); 214125511a47STejun Heo 21427e11629dSTejun Heo /* 21437e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21447e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21457e11629dSTejun Heo * already processing the work. If so, defer the work to the 21467e11629dSTejun Heo * currently executing one. 21477e11629dSTejun Heo */ 214842f8570fSSasha Levin collision = find_worker_executing_work(gcwq, work); 21497e11629dSTejun Heo if (unlikely(collision)) { 21507e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21517e11629dSTejun Heo return; 21527e11629dSTejun Heo } 21531da177e4SLinus Torvalds 21548930cabaSTejun Heo /* claim and dequeue */ 21551da177e4SLinus Torvalds debug_work_deactivate(work); 2156023f27d3STejun Heo hash_add(gcwq->busy_hash, &worker->hentry, (unsigned long)work); 2157c34056a3STejun Heo worker->current_work = work; 2158a2c1c57bSTejun Heo worker->current_func = work->func; 21598cca0eeaSTejun Heo worker->current_cwq = cwq; 216073f53c4aSTejun Heo work_color = get_work_color(work); 21617a22ad75STejun Heo 2162a62428c0STejun Heo list_del_init(&work->entry); 2163a62428c0STejun Heo 2164649027d7STejun Heo /* 2165fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2166fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2167fb0e7bebSTejun Heo */ 2168fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2169fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2170fb0e7bebSTejun Heo 2171974271c4STejun Heo /* 2172974271c4STejun Heo * Unbound gcwq isn't concurrency managed and work items should be 2173974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2174974271c4STejun Heo */ 217563d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 217663d95a91STejun Heo wake_up_worker(pool); 2177974271c4STejun Heo 21788930cabaSTejun Heo /* 217923657bb1STejun Heo * Record the last CPU and clear PENDING which should be the last 218023657bb1STejun Heo * update to @work. Also, do this inside @gcwq->lock so that 218123657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 218223657bb1STejun Heo * disabled. 21838930cabaSTejun Heo */ 21848930cabaSTejun Heo set_work_cpu_and_clear_pending(work, gcwq->cpu); 21851da177e4SLinus Torvalds 21861da177e4SLinus Torvalds spin_unlock_irq(&gcwq->lock); 2187365970a1SDavid Howells 2188e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 21893295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2190e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2191a2c1c57bSTejun Heo worker->current_func(work); 2192e36c886aSArjan van de Ven /* 2193e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2194e36c886aSArjan van de Ven * point will only record its address. 2195e36c886aSArjan van de Ven */ 2196e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21973295f0efSIngo Molnar lock_map_release(&lockdep_map); 21983295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 21991da177e4SLinus Torvalds 2200d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2201044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2202044c782cSValentin Ilie " last function: %pf\n", 2203a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2204a2c1c57bSTejun Heo worker->current_func); 2205d5abe669SPeter Zijlstra debug_show_held_locks(current); 2206d5abe669SPeter Zijlstra dump_stack(); 2207d5abe669SPeter Zijlstra } 2208d5abe669SPeter Zijlstra 22098b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2210a62428c0STejun Heo 2211fb0e7bebSTejun Heo /* clear cpu intensive status */ 2212fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2213fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2214fb0e7bebSTejun Heo 2215a62428c0STejun Heo /* we're done with it, release */ 221642f8570fSSasha Levin hash_del(&worker->hentry); 2217c34056a3STejun Heo worker->current_work = NULL; 2218a2c1c57bSTejun Heo worker->current_func = NULL; 22198cca0eeaSTejun Heo worker->current_cwq = NULL; 2220b3f9f405SLai Jiangshan cwq_dec_nr_in_flight(cwq, work_color); 22211da177e4SLinus Torvalds } 22221da177e4SLinus Torvalds 2223affee4b2STejun Heo /** 2224affee4b2STejun Heo * process_scheduled_works - process scheduled works 2225affee4b2STejun Heo * @worker: self 2226affee4b2STejun Heo * 2227affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2228affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2229affee4b2STejun Heo * fetches a work from the top and executes it. 2230affee4b2STejun Heo * 2231affee4b2STejun Heo * CONTEXT: 22328b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2233affee4b2STejun Heo * multiple times. 2234affee4b2STejun Heo */ 2235affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22361da177e4SLinus Torvalds { 2237affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2238affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2239a62428c0STejun Heo struct work_struct, entry); 2240c34056a3STejun Heo process_one_work(worker, work); 2241a62428c0STejun Heo } 22421da177e4SLinus Torvalds } 22431da177e4SLinus Torvalds 22444690c4abSTejun Heo /** 22454690c4abSTejun Heo * worker_thread - the worker thread function 2246c34056a3STejun Heo * @__worker: self 22474690c4abSTejun Heo * 2248e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 2249e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 2250e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2251e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2252e22bee78STejun Heo * rescuer_thread(). 22534690c4abSTejun Heo */ 2254c34056a3STejun Heo static int worker_thread(void *__worker) 22551da177e4SLinus Torvalds { 2256c34056a3STejun Heo struct worker *worker = __worker; 2257bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2258bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 22591da177e4SLinus Torvalds 2260e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2261e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2262c8e55f36STejun Heo woke_up: 22638b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2264affee4b2STejun Heo 22655f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22665f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2267c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 226825511a47STejun Heo 22695f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 227025511a47STejun Heo if (worker->flags & WORKER_DIE) { 2271e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2272c8e55f36STejun Heo return 0; 2273c8e55f36STejun Heo } 2274c8e55f36STejun Heo 22755f7dabfdSLai Jiangshan /* otherwise, rebind */ 227625511a47STejun Heo idle_worker_rebind(worker); 227725511a47STejun Heo goto woke_up; 227825511a47STejun Heo } 227925511a47STejun Heo 2280c8e55f36STejun Heo worker_leave_idle(worker); 2281db7bccf4STejun Heo recheck: 2282e22bee78STejun Heo /* no more worker necessary? */ 228363d95a91STejun Heo if (!need_more_worker(pool)) 2284e22bee78STejun Heo goto sleep; 2285e22bee78STejun Heo 2286e22bee78STejun Heo /* do we need to manage? */ 228763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2288e22bee78STejun Heo goto recheck; 2289e22bee78STejun Heo 2290c8e55f36STejun Heo /* 2291c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2292c8e55f36STejun Heo * preparing to process a work or actually processing it. 2293c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2294c8e55f36STejun Heo */ 2295c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 2296c8e55f36STejun Heo 2297e22bee78STejun Heo /* 2298e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2299e22bee78STejun Heo * at least one idle worker or that someone else has already 2300e22bee78STejun Heo * assumed the manager role. 2301e22bee78STejun Heo */ 2302e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2303e22bee78STejun Heo 2304e22bee78STejun Heo do { 2305affee4b2STejun Heo struct work_struct *work = 2306bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2307affee4b2STejun Heo struct work_struct, entry); 2308affee4b2STejun Heo 2309c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2310affee4b2STejun Heo /* optimization path, not strictly necessary */ 2311affee4b2STejun Heo process_one_work(worker, work); 2312affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2313affee4b2STejun Heo process_scheduled_works(worker); 2314affee4b2STejun Heo } else { 2315c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2316affee4b2STejun Heo process_scheduled_works(worker); 2317affee4b2STejun Heo } 231863d95a91STejun Heo } while (keep_working(pool)); 2319affee4b2STejun Heo 2320e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2321d313dd85STejun Heo sleep: 232263d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2323e22bee78STejun Heo goto recheck; 2324d313dd85STejun Heo 2325c8e55f36STejun Heo /* 2326e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 2327e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 2328e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 2329e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 2330e22bee78STejun Heo * prevent losing any event. 2331c8e55f36STejun Heo */ 2332c8e55f36STejun Heo worker_enter_idle(worker); 2333c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 23348b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 23351da177e4SLinus Torvalds schedule(); 2336c8e55f36STejun Heo goto woke_up; 23371da177e4SLinus Torvalds } 23381da177e4SLinus Torvalds 2339e22bee78STejun Heo /** 2340e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2341111c225aSTejun Heo * @__rescuer: self 2342e22bee78STejun Heo * 2343e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2344e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2345e22bee78STejun Heo * 2346e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 2347e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2348e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2349e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2350e22bee78STejun Heo * the problem rescuer solves. 2351e22bee78STejun Heo * 2352e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 2353e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 2354e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2355e22bee78STejun Heo * 2356e22bee78STejun Heo * This should happen rarely. 2357e22bee78STejun Heo */ 2358111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2359e22bee78STejun Heo { 2360111c225aSTejun Heo struct worker *rescuer = __rescuer; 2361111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2362e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2363f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2364e22bee78STejun Heo unsigned int cpu; 2365e22bee78STejun Heo 2366e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2367111c225aSTejun Heo 2368111c225aSTejun Heo /* 2369111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2370111c225aSTejun Heo * doesn't participate in concurrency management. 2371111c225aSTejun Heo */ 2372111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2373e22bee78STejun Heo repeat: 2374e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23751da177e4SLinus Torvalds 2376412d32e6SMike Galbraith if (kthread_should_stop()) { 2377412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2378111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2379e22bee78STejun Heo return 0; 2380412d32e6SMike Galbraith } 23811da177e4SLinus Torvalds 2382f3421797STejun Heo /* 2383f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2384f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2385f3421797STejun Heo */ 2386f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2387f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2388f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2389bd7bdd43STejun Heo struct worker_pool *pool = cwq->pool; 2390bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2391e22bee78STejun Heo struct work_struct *work, *n; 2392e22bee78STejun Heo 2393e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2394f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2395e22bee78STejun Heo 2396e22bee78STejun Heo /* migrate to the target cpu if possible */ 2397bd7bdd43STejun Heo rescuer->pool = pool; 2398e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2399e22bee78STejun Heo 2400e22bee78STejun Heo /* 2401e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2402e22bee78STejun Heo * process'em. 2403e22bee78STejun Heo */ 2404e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2405bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2406e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2407e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2408e22bee78STejun Heo 2409e22bee78STejun Heo process_scheduled_works(rescuer); 24107576958aSTejun Heo 24117576958aSTejun Heo /* 24127576958aSTejun Heo * Leave this gcwq. If keep_working() is %true, notify a 24137576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24147576958aSTejun Heo * and stalling the execution. 24157576958aSTejun Heo */ 241663d95a91STejun Heo if (keep_working(pool)) 241763d95a91STejun Heo wake_up_worker(pool); 24187576958aSTejun Heo 2419e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 24201da177e4SLinus Torvalds } 24211da177e4SLinus Torvalds 2422111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2423111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2424e22bee78STejun Heo schedule(); 2425e22bee78STejun Heo goto repeat; 24261da177e4SLinus Torvalds } 24271da177e4SLinus Torvalds 2428fc2e4d70SOleg Nesterov struct wq_barrier { 2429fc2e4d70SOleg Nesterov struct work_struct work; 2430fc2e4d70SOleg Nesterov struct completion done; 2431fc2e4d70SOleg Nesterov }; 2432fc2e4d70SOleg Nesterov 2433fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2434fc2e4d70SOleg Nesterov { 2435fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2436fc2e4d70SOleg Nesterov complete(&barr->done); 2437fc2e4d70SOleg Nesterov } 2438fc2e4d70SOleg Nesterov 24394690c4abSTejun Heo /** 24404690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 24414690c4abSTejun Heo * @cwq: cwq to insert barrier into 24424690c4abSTejun Heo * @barr: wq_barrier to insert 2443affee4b2STejun Heo * @target: target work to attach @barr to 2444affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24454690c4abSTejun Heo * 2446affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2447affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2448affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2449affee4b2STejun Heo * cpu. 2450affee4b2STejun Heo * 2451affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2452affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2453affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2454affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2455affee4b2STejun Heo * after a work with LINKED flag set. 2456affee4b2STejun Heo * 2457affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2458affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 24594690c4abSTejun Heo * 24604690c4abSTejun Heo * CONTEXT: 24618b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 24624690c4abSTejun Heo */ 246383c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2464affee4b2STejun Heo struct wq_barrier *barr, 2465affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2466fc2e4d70SOleg Nesterov { 2467affee4b2STejun Heo struct list_head *head; 2468affee4b2STejun Heo unsigned int linked = 0; 2469affee4b2STejun Heo 2470dc186ad7SThomas Gleixner /* 24718b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 2472dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2473dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2474dc186ad7SThomas Gleixner * might deadlock. 2475dc186ad7SThomas Gleixner */ 2476ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 247722df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2478fc2e4d70SOleg Nesterov init_completion(&barr->done); 247983c22520SOleg Nesterov 2480affee4b2STejun Heo /* 2481affee4b2STejun Heo * If @target is currently being executed, schedule the 2482affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2483affee4b2STejun Heo */ 2484affee4b2STejun Heo if (worker) 2485affee4b2STejun Heo head = worker->scheduled.next; 2486affee4b2STejun Heo else { 2487affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2488affee4b2STejun Heo 2489affee4b2STejun Heo head = target->entry.next; 2490affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2491affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2492affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2493affee4b2STejun Heo } 2494affee4b2STejun Heo 2495dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2496affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2497affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2498fc2e4d70SOleg Nesterov } 2499fc2e4d70SOleg Nesterov 250073f53c4aSTejun Heo /** 250173f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 250273f53c4aSTejun Heo * @wq: workqueue being flushed 250373f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 250473f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 250573f53c4aSTejun Heo * 250673f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 250773f53c4aSTejun Heo * 250873f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 250973f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 251073f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 251173f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 251273f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 251373f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 251473f53c4aSTejun Heo * 251573f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 251673f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 251773f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 251873f53c4aSTejun Heo * is returned. 251973f53c4aSTejun Heo * 252073f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 252173f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 252273f53c4aSTejun Heo * advanced to @work_color. 252373f53c4aSTejun Heo * 252473f53c4aSTejun Heo * CONTEXT: 252573f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 252673f53c4aSTejun Heo * 252773f53c4aSTejun Heo * RETURNS: 252873f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 252973f53c4aSTejun Heo * otherwise. 253073f53c4aSTejun Heo */ 253173f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 253273f53c4aSTejun Heo int flush_color, int work_color) 25331da177e4SLinus Torvalds { 253473f53c4aSTejun Heo bool wait = false; 253573f53c4aSTejun Heo unsigned int cpu; 25361da177e4SLinus Torvalds 253773f53c4aSTejun Heo if (flush_color >= 0) { 253873f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 253973f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2540dc186ad7SThomas Gleixner } 254114441960SOleg Nesterov 2542f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 254373f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2544bd7bdd43STejun Heo struct global_cwq *gcwq = cwq->pool->gcwq; 25451da177e4SLinus Torvalds 25468b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 254773f53c4aSTejun Heo 254873f53c4aSTejun Heo if (flush_color >= 0) { 254973f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 255073f53c4aSTejun Heo 255173f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 255273f53c4aSTejun Heo cwq->flush_color = flush_color; 255373f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 255473f53c4aSTejun Heo wait = true; 25551da177e4SLinus Torvalds } 255673f53c4aSTejun Heo } 255773f53c4aSTejun Heo 255873f53c4aSTejun Heo if (work_color >= 0) { 255973f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 256073f53c4aSTejun Heo cwq->work_color = work_color; 256173f53c4aSTejun Heo } 256273f53c4aSTejun Heo 25638b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 25641da177e4SLinus Torvalds } 25651da177e4SLinus Torvalds 256673f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 256773f53c4aSTejun Heo complete(&wq->first_flusher->done); 256873f53c4aSTejun Heo 256973f53c4aSTejun Heo return wait; 257083c22520SOleg Nesterov } 25711da177e4SLinus Torvalds 25720fcb78c2SRolf Eike Beer /** 25731da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25740fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25751da177e4SLinus Torvalds * 25761da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 25771da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 25781da177e4SLinus Torvalds * 2579fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2580fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 25811da177e4SLinus Torvalds */ 25827ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25831da177e4SLinus Torvalds { 258473f53c4aSTejun Heo struct wq_flusher this_flusher = { 258573f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 258673f53c4aSTejun Heo .flush_color = -1, 258773f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 258873f53c4aSTejun Heo }; 258973f53c4aSTejun Heo int next_color; 2590b1f4ec17SOleg Nesterov 25913295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25923295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 259373f53c4aSTejun Heo 259473f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 259573f53c4aSTejun Heo 259673f53c4aSTejun Heo /* 259773f53c4aSTejun Heo * Start-to-wait phase 259873f53c4aSTejun Heo */ 259973f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 260073f53c4aSTejun Heo 260173f53c4aSTejun Heo if (next_color != wq->flush_color) { 260273f53c4aSTejun Heo /* 260373f53c4aSTejun Heo * Color space is not full. The current work_color 260473f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 260573f53c4aSTejun Heo * by one. 260673f53c4aSTejun Heo */ 260773f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 260873f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 260973f53c4aSTejun Heo wq->work_color = next_color; 261073f53c4aSTejun Heo 261173f53c4aSTejun Heo if (!wq->first_flusher) { 261273f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 261373f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 261473f53c4aSTejun Heo 261573f53c4aSTejun Heo wq->first_flusher = &this_flusher; 261673f53c4aSTejun Heo 261773f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 261873f53c4aSTejun Heo wq->work_color)) { 261973f53c4aSTejun Heo /* nothing to flush, done */ 262073f53c4aSTejun Heo wq->flush_color = next_color; 262173f53c4aSTejun Heo wq->first_flusher = NULL; 262273f53c4aSTejun Heo goto out_unlock; 262373f53c4aSTejun Heo } 262473f53c4aSTejun Heo } else { 262573f53c4aSTejun Heo /* wait in queue */ 262673f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 262773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 262873f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 262973f53c4aSTejun Heo } 263073f53c4aSTejun Heo } else { 263173f53c4aSTejun Heo /* 263273f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 263373f53c4aSTejun Heo * The next flush completion will assign us 263473f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 263573f53c4aSTejun Heo */ 263673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 263773f53c4aSTejun Heo } 263873f53c4aSTejun Heo 263973f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 264073f53c4aSTejun Heo 264173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 264273f53c4aSTejun Heo 264373f53c4aSTejun Heo /* 264473f53c4aSTejun Heo * Wake-up-and-cascade phase 264573f53c4aSTejun Heo * 264673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 264773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 264873f53c4aSTejun Heo */ 264973f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 265073f53c4aSTejun Heo return; 265173f53c4aSTejun Heo 265273f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 265373f53c4aSTejun Heo 26544ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26554ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26564ce48b37STejun Heo goto out_unlock; 26574ce48b37STejun Heo 265873f53c4aSTejun Heo wq->first_flusher = NULL; 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 266173f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 266273f53c4aSTejun Heo 266373f53c4aSTejun Heo while (true) { 266473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 266573f53c4aSTejun Heo 266673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 266773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 266873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 266973f53c4aSTejun Heo break; 267073f53c4aSTejun Heo list_del_init(&next->list); 267173f53c4aSTejun Heo complete(&next->done); 267273f53c4aSTejun Heo } 267373f53c4aSTejun Heo 267473f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 267573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 267673f53c4aSTejun Heo 267773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 267873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 267973f53c4aSTejun Heo 268073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 268173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 268273f53c4aSTejun Heo /* 268373f53c4aSTejun Heo * Assign the same color to all overflowed 268473f53c4aSTejun Heo * flushers, advance work_color and append to 268573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 268673f53c4aSTejun Heo * phase for these overflowed flushers. 268773f53c4aSTejun Heo */ 268873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 268973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 269273f53c4aSTejun Heo 269373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 269473f53c4aSTejun Heo &wq->flusher_queue); 269573f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 269673f53c4aSTejun Heo } 269773f53c4aSTejun Heo 269873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 269973f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 270073f53c4aSTejun Heo break; 270173f53c4aSTejun Heo } 270273f53c4aSTejun Heo 270373f53c4aSTejun Heo /* 270473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 270573f53c4aSTejun Heo * the new first flusher and arm cwqs. 270673f53c4aSTejun Heo */ 270773f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 270873f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 270973f53c4aSTejun Heo 271073f53c4aSTejun Heo list_del_init(&next->list); 271173f53c4aSTejun Heo wq->first_flusher = next; 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 271473f53c4aSTejun Heo break; 271573f53c4aSTejun Heo 271673f53c4aSTejun Heo /* 271773f53c4aSTejun Heo * Meh... this color is already done, clear first 271873f53c4aSTejun Heo * flusher and repeat cascading. 271973f53c4aSTejun Heo */ 272073f53c4aSTejun Heo wq->first_flusher = NULL; 272173f53c4aSTejun Heo } 272273f53c4aSTejun Heo 272373f53c4aSTejun Heo out_unlock: 272473f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27251da177e4SLinus Torvalds } 2726ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27271da177e4SLinus Torvalds 27289c5a2ba7STejun Heo /** 27299c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27309c5a2ba7STejun Heo * @wq: workqueue to drain 27319c5a2ba7STejun Heo * 27329c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27339c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27349c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27359c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27369c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27379c5a2ba7STejun Heo * takes too long. 27389c5a2ba7STejun Heo */ 27399c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27409c5a2ba7STejun Heo { 27419c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27429c5a2ba7STejun Heo unsigned int cpu; 27439c5a2ba7STejun Heo 27449c5a2ba7STejun Heo /* 27459c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27469c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27479c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27489c5a2ba7STejun Heo */ 27499c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27509c5a2ba7STejun Heo if (!wq->nr_drainers++) 27519c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 27529c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27539c5a2ba7STejun Heo reflush: 27549c5a2ba7STejun Heo flush_workqueue(wq); 27559c5a2ba7STejun Heo 27569c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 27579c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2758fa2563e4SThomas Tuttle bool drained; 27599c5a2ba7STejun Heo 2760bd7bdd43STejun Heo spin_lock_irq(&cwq->pool->gcwq->lock); 2761fa2563e4SThomas Tuttle drained = !cwq->nr_active && list_empty(&cwq->delayed_works); 2762bd7bdd43STejun Heo spin_unlock_irq(&cwq->pool->gcwq->lock); 2763fa2563e4SThomas Tuttle 2764fa2563e4SThomas Tuttle if (drained) 27659c5a2ba7STejun Heo continue; 27669c5a2ba7STejun Heo 27679c5a2ba7STejun Heo if (++flush_cnt == 10 || 27689c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2769044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 27709c5a2ba7STejun Heo wq->name, flush_cnt); 27719c5a2ba7STejun Heo goto reflush; 27729c5a2ba7STejun Heo } 27739c5a2ba7STejun Heo 27749c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27759c5a2ba7STejun Heo if (!--wq->nr_drainers) 27769c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 27779c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27789c5a2ba7STejun Heo } 27799c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27809c5a2ba7STejun Heo 2781606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2782baf59022STejun Heo { 2783baf59022STejun Heo struct worker *worker = NULL; 2784baf59022STejun Heo struct global_cwq *gcwq; 2785baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2786baf59022STejun Heo 2787baf59022STejun Heo might_sleep(); 2788baf59022STejun Heo gcwq = get_work_gcwq(work); 2789baf59022STejun Heo if (!gcwq) 2790baf59022STejun Heo return false; 2791baf59022STejun Heo 2792baf59022STejun Heo spin_lock_irq(&gcwq->lock); 2793baf59022STejun Heo if (!list_empty(&work->entry)) { 2794baf59022STejun Heo /* 2795baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2796baf59022STejun Heo * If it was re-queued to a different gcwq under us, we 2797baf59022STejun Heo * are not going to wait. 2798baf59022STejun Heo */ 2799baf59022STejun Heo smp_rmb(); 2800baf59022STejun Heo cwq = get_work_cwq(work); 2801bd7bdd43STejun Heo if (unlikely(!cwq || gcwq != cwq->pool->gcwq)) 2802baf59022STejun Heo goto already_gone; 2803606a5020STejun Heo } else { 2804baf59022STejun Heo worker = find_worker_executing_work(gcwq, work); 2805baf59022STejun Heo if (!worker) 2806baf59022STejun Heo goto already_gone; 2807baf59022STejun Heo cwq = worker->current_cwq; 2808606a5020STejun Heo } 2809baf59022STejun Heo 2810baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2811baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2812baf59022STejun Heo 2813e159489bSTejun Heo /* 2814e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2815e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2816e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2817e159489bSTejun Heo * access. 2818e159489bSTejun Heo */ 2819e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2820baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2821e159489bSTejun Heo else 2822e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2823baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2824e159489bSTejun Heo 2825baf59022STejun Heo return true; 2826baf59022STejun Heo already_gone: 2827baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2828baf59022STejun Heo return false; 2829baf59022STejun Heo } 2830baf59022STejun Heo 2831db700897SOleg Nesterov /** 2832401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2833401a8d04STejun Heo * @work: the work to flush 2834db700897SOleg Nesterov * 2835606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2836606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2837401a8d04STejun Heo * 2838401a8d04STejun Heo * RETURNS: 2839401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2840401a8d04STejun Heo * %false if it was already idle. 2841db700897SOleg Nesterov */ 2842401a8d04STejun Heo bool flush_work(struct work_struct *work) 2843db700897SOleg Nesterov { 2844db700897SOleg Nesterov struct wq_barrier barr; 2845db700897SOleg Nesterov 28460976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28470976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28480976dfc1SStephen Boyd 2849606a5020STejun Heo if (start_flush_work(work, &barr)) { 2850db700897SOleg Nesterov wait_for_completion(&barr.done); 2851dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2852401a8d04STejun Heo return true; 2853606a5020STejun Heo } else { 2854401a8d04STejun Heo return false; 2855db700897SOleg Nesterov } 2856606a5020STejun Heo } 2857db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2858db700897SOleg Nesterov 285936e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2860401a8d04STejun Heo { 2861bbb68dfaSTejun Heo unsigned long flags; 28621f1f642eSOleg Nesterov int ret; 28631f1f642eSOleg Nesterov 28641f1f642eSOleg Nesterov do { 2865bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2866bbb68dfaSTejun Heo /* 2867bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2868bbb68dfaSTejun Heo * would be waiting for before retrying. 2869bbb68dfaSTejun Heo */ 2870bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2871606a5020STejun Heo flush_work(work); 28721f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28731f1f642eSOleg Nesterov 2874bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2875bbb68dfaSTejun Heo mark_work_canceling(work); 2876bbb68dfaSTejun Heo local_irq_restore(flags); 2877bbb68dfaSTejun Heo 2878606a5020STejun Heo flush_work(work); 28797a22ad75STejun Heo clear_work_data(work); 28801f1f642eSOleg Nesterov return ret; 28811f1f642eSOleg Nesterov } 28821f1f642eSOleg Nesterov 28836e84d644SOleg Nesterov /** 2884401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2885401a8d04STejun Heo * @work: the work to cancel 28866e84d644SOleg Nesterov * 2887401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2888401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2889401a8d04STejun Heo * another workqueue. On return from this function, @work is 2890401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28911f1f642eSOleg Nesterov * 2892401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2893401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28946e84d644SOleg Nesterov * 2895401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28966e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2897401a8d04STejun Heo * 2898401a8d04STejun Heo * RETURNS: 2899401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29006e84d644SOleg Nesterov */ 2901401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29026e84d644SOleg Nesterov { 290336e227d2STejun Heo return __cancel_work_timer(work, false); 2904b89deed3SOleg Nesterov } 290528e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2906b89deed3SOleg Nesterov 29076e84d644SOleg Nesterov /** 2908401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2909401a8d04STejun Heo * @dwork: the delayed work to flush 29106e84d644SOleg Nesterov * 2911401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2912401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2913401a8d04STejun Heo * considers the last queueing instance of @dwork. 29141f1f642eSOleg Nesterov * 2915401a8d04STejun Heo * RETURNS: 2916401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2917401a8d04STejun Heo * %false if it was already idle. 29186e84d644SOleg Nesterov */ 2919401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2920401a8d04STejun Heo { 29218930cabaSTejun Heo local_irq_disable(); 2922401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 29231265057fSTejun Heo __queue_work(dwork->cpu, 2924401a8d04STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 29258930cabaSTejun Heo local_irq_enable(); 2926401a8d04STejun Heo return flush_work(&dwork->work); 2927401a8d04STejun Heo } 2928401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2929401a8d04STejun Heo 2930401a8d04STejun Heo /** 293157b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 293257b30ae7STejun Heo * @dwork: delayed_work to cancel 293309383498STejun Heo * 293457b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 293557b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 293657b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 293757b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 293857b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 293909383498STejun Heo * 294057b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 294109383498STejun Heo */ 294257b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 294309383498STejun Heo { 294457b30ae7STejun Heo unsigned long flags; 294557b30ae7STejun Heo int ret; 294657b30ae7STejun Heo 294757b30ae7STejun Heo do { 294857b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 294957b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 295057b30ae7STejun Heo 295157b30ae7STejun Heo if (unlikely(ret < 0)) 295257b30ae7STejun Heo return false; 295357b30ae7STejun Heo 295457b30ae7STejun Heo set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work)); 295557b30ae7STejun Heo local_irq_restore(flags); 2956c0158ca6SDan Magenheimer return ret; 295709383498STejun Heo } 295857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 295909383498STejun Heo 296009383498STejun Heo /** 2961401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2962401a8d04STejun Heo * @dwork: the delayed work cancel 2963401a8d04STejun Heo * 2964401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2965401a8d04STejun Heo * 2966401a8d04STejun Heo * RETURNS: 2967401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2968401a8d04STejun Heo */ 2969401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29706e84d644SOleg Nesterov { 297136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29726e84d644SOleg Nesterov } 2973f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29741da177e4SLinus Torvalds 29750fcb78c2SRolf Eike Beer /** 2976c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2977c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2978c1a220e7SZhang Rui * @work: job to be done 2979c1a220e7SZhang Rui * 2980c1a220e7SZhang Rui * This puts a job on a specific cpu 2981c1a220e7SZhang Rui */ 2982d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 2983c1a220e7SZhang Rui { 2984d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2985c1a220e7SZhang Rui } 2986c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2987c1a220e7SZhang Rui 29880fcb78c2SRolf Eike Beer /** 2989ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 29901da177e4SLinus Torvalds * @work: job to be done 29911da177e4SLinus Torvalds * 2992d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 2993d4283e93STejun Heo * %true otherwise. 299452bad64dSDavid Howells * 29950fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 29960fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 29970fcb78c2SRolf Eike Beer * workqueue otherwise. 29981da177e4SLinus Torvalds */ 2999d4283e93STejun Heo bool schedule_work(struct work_struct *work) 30001da177e4SLinus Torvalds { 30010fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 30021da177e4SLinus Torvalds } 30030fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 30041da177e4SLinus Torvalds 30050fcb78c2SRolf Eike Beer /** 30060fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 30070fcb78c2SRolf Eike Beer * @cpu: cpu to use 30080fcb78c2SRolf Eike Beer * @dwork: job to be done 30090fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 30100fcb78c2SRolf Eike Beer * 30110fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 30120fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 30130fcb78c2SRolf Eike Beer */ 3014d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 3015d4283e93STejun Heo unsigned long delay) 30161da177e4SLinus Torvalds { 3017d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 30181da177e4SLinus Torvalds } 3019ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 30201da177e4SLinus Torvalds 3021b6136773SAndrew Morton /** 30220a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 30230a13c00eSTejun Heo * @dwork: job to be done 30240a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 30250a13c00eSTejun Heo * 30260a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 30270a13c00eSTejun Heo * workqueue. 30280a13c00eSTejun Heo */ 3029d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 30300a13c00eSTejun Heo { 30310a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30320a13c00eSTejun Heo } 30330a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30340a13c00eSTejun Heo 30350a13c00eSTejun Heo /** 303631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3037b6136773SAndrew Morton * @func: the function to call 3038b6136773SAndrew Morton * 303931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 304031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3041b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 304231ddd871STejun Heo * 304331ddd871STejun Heo * RETURNS: 304431ddd871STejun Heo * 0 on success, -errno on failure. 3045b6136773SAndrew Morton */ 304665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 304715316ba8SChristoph Lameter { 304815316ba8SChristoph Lameter int cpu; 304938f51568SNamhyung Kim struct work_struct __percpu *works; 305015316ba8SChristoph Lameter 3051b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3052b6136773SAndrew Morton if (!works) 305315316ba8SChristoph Lameter return -ENOMEM; 3054b6136773SAndrew Morton 305595402b38SGautham R Shenoy get_online_cpus(); 305693981800STejun Heo 305715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30589bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30599bfb1839SIngo Molnar 30609bfb1839SIngo Molnar INIT_WORK(work, func); 30618de6d308SOleg Nesterov schedule_work_on(cpu, work); 306215316ba8SChristoph Lameter } 306393981800STejun Heo 306493981800STejun Heo for_each_online_cpu(cpu) 30658616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 306693981800STejun Heo 306795402b38SGautham R Shenoy put_online_cpus(); 3068b6136773SAndrew Morton free_percpu(works); 306915316ba8SChristoph Lameter return 0; 307015316ba8SChristoph Lameter } 307115316ba8SChristoph Lameter 3072eef6a7d5SAlan Stern /** 3073eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3074eef6a7d5SAlan Stern * 3075eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3076eef6a7d5SAlan Stern * completion. 3077eef6a7d5SAlan Stern * 3078eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3079eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3080eef6a7d5SAlan Stern * will lead to deadlock: 3081eef6a7d5SAlan Stern * 3082eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3083eef6a7d5SAlan Stern * a lock held by your code or its caller. 3084eef6a7d5SAlan Stern * 3085eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3086eef6a7d5SAlan Stern * 3087eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3088eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3089eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3090eef6a7d5SAlan Stern * 3091eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3092eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3093eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3094eef6a7d5SAlan Stern * cancel_work_sync() instead. 3095eef6a7d5SAlan Stern */ 30961da177e4SLinus Torvalds void flush_scheduled_work(void) 30971da177e4SLinus Torvalds { 3098d320c038STejun Heo flush_workqueue(system_wq); 30991da177e4SLinus Torvalds } 3100ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 31011da177e4SLinus Torvalds 31021da177e4SLinus Torvalds /** 31031fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31041fa44ecaSJames Bottomley * @fn: the function to execute 31051fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31061fa44ecaSJames Bottomley * be available when the work executes) 31071fa44ecaSJames Bottomley * 31081fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31091fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31101fa44ecaSJames Bottomley * 31111fa44ecaSJames Bottomley * Returns: 0 - function was executed 31121fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31131fa44ecaSJames Bottomley */ 311465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31151fa44ecaSJames Bottomley { 31161fa44ecaSJames Bottomley if (!in_interrupt()) { 311765f27f38SDavid Howells fn(&ew->work); 31181fa44ecaSJames Bottomley return 0; 31191fa44ecaSJames Bottomley } 31201fa44ecaSJames Bottomley 312165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31221fa44ecaSJames Bottomley schedule_work(&ew->work); 31231fa44ecaSJames Bottomley 31241fa44ecaSJames Bottomley return 1; 31251fa44ecaSJames Bottomley } 31261fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31271fa44ecaSJames Bottomley 31281da177e4SLinus Torvalds int keventd_up(void) 31291da177e4SLinus Torvalds { 3130d320c038STejun Heo return system_wq != NULL; 31311da177e4SLinus Torvalds } 31321da177e4SLinus Torvalds 3133bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 31341da177e4SLinus Torvalds { 31353af24433SOleg Nesterov /* 31360f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 31370f900049STejun Heo * Make sure that the alignment isn't lower than that of 31380f900049STejun Heo * unsigned long long. 31393af24433SOleg Nesterov */ 31400f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 31410f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 31420f900049STejun Heo __alignof__(unsigned long long)); 31433af24433SOleg Nesterov 3144e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3145f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 3146931ac77eSTejun Heo else { 31470f900049STejun Heo void *ptr; 3148e1d8aa9fSFrederic Weisbecker 31490f900049STejun Heo /* 3150f3421797STejun Heo * Allocate enough room to align cwq and put an extra 3151f3421797STejun Heo * pointer at the end pointing back to the originally 3152f3421797STejun Heo * allocated pointer which will be used for free. 31530f900049STejun Heo */ 3154bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3155bdbc5dd7STejun Heo if (ptr) { 3156bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 3157bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 3158bdbc5dd7STejun Heo } 31593af24433SOleg Nesterov } 31603af24433SOleg Nesterov 31610415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3162bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 3163bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 31640f900049STejun Heo } 31650f900049STejun Heo 3166bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 316706ba38a9SOleg Nesterov { 3168e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3169bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 3170f3421797STejun Heo else if (wq->cpu_wq.single) { 3171f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 3172f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 317306ba38a9SOleg Nesterov } 317406ba38a9SOleg Nesterov } 317506ba38a9SOleg Nesterov 3176f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3177f3421797STejun Heo const char *name) 3178b71ab8c2STejun Heo { 3179f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3180f3421797STejun Heo 3181f3421797STejun Heo if (max_active < 1 || max_active > lim) 3182044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3183f3421797STejun Heo max_active, name, 1, lim); 3184b71ab8c2STejun Heo 3185f3421797STejun Heo return clamp_val(max_active, 1, lim); 3186b71ab8c2STejun Heo } 3187b71ab8c2STejun Heo 3188b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 318997e37d7bSTejun Heo unsigned int flags, 31901e19ffc6STejun Heo int max_active, 3191eb13ba87SJohannes Berg struct lock_class_key *key, 3192b196be89STejun Heo const char *lock_name, ...) 31933af24433SOleg Nesterov { 3194b196be89STejun Heo va_list args, args1; 31953af24433SOleg Nesterov struct workqueue_struct *wq; 3196c34056a3STejun Heo unsigned int cpu; 3197b196be89STejun Heo size_t namelen; 3198b196be89STejun Heo 3199b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3200b196be89STejun Heo va_start(args, lock_name); 3201b196be89STejun Heo va_copy(args1, args); 3202b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3203b196be89STejun Heo 3204b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3205b196be89STejun Heo if (!wq) 3206b196be89STejun Heo goto err; 3207b196be89STejun Heo 3208b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3209b196be89STejun Heo va_end(args); 3210b196be89STejun Heo va_end(args1); 32113af24433SOleg Nesterov 3212f3421797STejun Heo /* 32136370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 32146370a6adSTejun Heo * have a rescuer to guarantee forward progress. 32156370a6adSTejun Heo */ 32166370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 32176370a6adSTejun Heo flags |= WQ_RESCUER; 32186370a6adSTejun Heo 3219d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3220b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 32213af24433SOleg Nesterov 3222b196be89STejun Heo /* init wq */ 322397e37d7bSTejun Heo wq->flags = flags; 3224a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 322573f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 322673f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 322773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 322873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 32293af24433SOleg Nesterov 3230eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3231cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 32323af24433SOleg Nesterov 3233bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 3234bdbc5dd7STejun Heo goto err; 3235bdbc5dd7STejun Heo 3236f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 32371537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 32388b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 32393270476aSTejun Heo int pool_idx = (bool)(flags & WQ_HIGHPRI); 32401537663fSTejun Heo 32410f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 32423270476aSTejun Heo cwq->pool = &gcwq->pools[pool_idx]; 3243c34056a3STejun Heo cwq->wq = wq; 324473f53c4aSTejun Heo cwq->flush_color = -1; 32451e19ffc6STejun Heo cwq->max_active = max_active; 32461e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3247e22bee78STejun Heo } 32481537663fSTejun Heo 3249e22bee78STejun Heo if (flags & WQ_RESCUER) { 3250e22bee78STejun Heo struct worker *rescuer; 3251e22bee78STejun Heo 3252f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3253e22bee78STejun Heo goto err; 3254e22bee78STejun Heo 3255e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3256e22bee78STejun Heo if (!rescuer) 3257e22bee78STejun Heo goto err; 3258e22bee78STejun Heo 3259111c225aSTejun Heo rescuer->rescue_wq = wq; 3260111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3261b196be89STejun Heo wq->name); 3262e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3263e22bee78STejun Heo goto err; 3264e22bee78STejun Heo 3265e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3266e22bee78STejun Heo wake_up_process(rescuer->task); 32673af24433SOleg Nesterov } 32681537663fSTejun Heo 32693af24433SOleg Nesterov /* 3270a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3271a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3272a0a1a5fdSTejun Heo * workqueue to workqueues list. 32733af24433SOleg Nesterov */ 32743af24433SOleg Nesterov spin_lock(&workqueue_lock); 3275a0a1a5fdSTejun Heo 327658a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3277f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3278a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3279a0a1a5fdSTejun Heo 32803af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3281a0a1a5fdSTejun Heo 32823af24433SOleg Nesterov spin_unlock(&workqueue_lock); 32833af24433SOleg Nesterov 32843af24433SOleg Nesterov return wq; 32854690c4abSTejun Heo err: 32864690c4abSTejun Heo if (wq) { 3287bdbc5dd7STejun Heo free_cwqs(wq); 3288f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3289e22bee78STejun Heo kfree(wq->rescuer); 32904690c4abSTejun Heo kfree(wq); 32913af24433SOleg Nesterov } 32924690c4abSTejun Heo return NULL; 32931da177e4SLinus Torvalds } 3294d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 32951da177e4SLinus Torvalds 32963af24433SOleg Nesterov /** 32973af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 32983af24433SOleg Nesterov * @wq: target workqueue 32993af24433SOleg Nesterov * 33003af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 33013af24433SOleg Nesterov */ 33023af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 33033af24433SOleg Nesterov { 3304c8e55f36STejun Heo unsigned int cpu; 33053af24433SOleg Nesterov 33069c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 33079c5a2ba7STejun Heo drain_workqueue(wq); 3308c8efcc25STejun Heo 3309a0a1a5fdSTejun Heo /* 3310a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3311a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3312a0a1a5fdSTejun Heo */ 331395402b38SGautham R Shenoy spin_lock(&workqueue_lock); 33143af24433SOleg Nesterov list_del(&wq->list); 331595402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 33163af24433SOleg Nesterov 3317e22bee78STejun Heo /* sanity check */ 3318f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 331973f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 332073f53c4aSTejun Heo int i; 33213af24433SOleg Nesterov 332273f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 332373f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 33241e19ffc6STejun Heo BUG_ON(cwq->nr_active); 33251e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 332673f53c4aSTejun Heo } 33271537663fSTejun Heo 3328e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3329e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3330f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 33318d9df9f0SXiaotian Feng kfree(wq->rescuer); 3332e22bee78STejun Heo } 3333e22bee78STejun Heo 3334bdbc5dd7STejun Heo free_cwqs(wq); 33353af24433SOleg Nesterov kfree(wq); 33363af24433SOleg Nesterov } 33373af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33383af24433SOleg Nesterov 3339dcd989cbSTejun Heo /** 33409f4bd4cdSLai Jiangshan * cwq_set_max_active - adjust max_active of a cwq 33419f4bd4cdSLai Jiangshan * @cwq: target cpu_workqueue_struct 33429f4bd4cdSLai Jiangshan * @max_active: new max_active value. 33439f4bd4cdSLai Jiangshan * 33449f4bd4cdSLai Jiangshan * Set @cwq->max_active to @max_active and activate delayed works if 33459f4bd4cdSLai Jiangshan * increased. 33469f4bd4cdSLai Jiangshan * 33479f4bd4cdSLai Jiangshan * CONTEXT: 33489f4bd4cdSLai Jiangshan * spin_lock_irq(gcwq->lock). 33499f4bd4cdSLai Jiangshan */ 33509f4bd4cdSLai Jiangshan static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active) 33519f4bd4cdSLai Jiangshan { 33529f4bd4cdSLai Jiangshan cwq->max_active = max_active; 33539f4bd4cdSLai Jiangshan 33549f4bd4cdSLai Jiangshan while (!list_empty(&cwq->delayed_works) && 33559f4bd4cdSLai Jiangshan cwq->nr_active < cwq->max_active) 33569f4bd4cdSLai Jiangshan cwq_activate_first_delayed(cwq); 33579f4bd4cdSLai Jiangshan } 33589f4bd4cdSLai Jiangshan 33599f4bd4cdSLai Jiangshan /** 3360dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3361dcd989cbSTejun Heo * @wq: target workqueue 3362dcd989cbSTejun Heo * @max_active: new max_active value. 3363dcd989cbSTejun Heo * 3364dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3365dcd989cbSTejun Heo * 3366dcd989cbSTejun Heo * CONTEXT: 3367dcd989cbSTejun Heo * Don't call from IRQ context. 3368dcd989cbSTejun Heo */ 3369dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3370dcd989cbSTejun Heo { 3371dcd989cbSTejun Heo unsigned int cpu; 3372dcd989cbSTejun Heo 3373f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3374dcd989cbSTejun Heo 3375dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3376dcd989cbSTejun Heo 3377dcd989cbSTejun Heo wq->saved_max_active = max_active; 3378dcd989cbSTejun Heo 3379f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 3380dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3381dcd989cbSTejun Heo 3382dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 3383dcd989cbSTejun Heo 338458a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 3385dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 338670369b11SLai Jiangshan cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active); 3387dcd989cbSTejun Heo 3388dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 3389dcd989cbSTejun Heo } 3390dcd989cbSTejun Heo 3391dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3392dcd989cbSTejun Heo } 3393dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3394dcd989cbSTejun Heo 3395dcd989cbSTejun Heo /** 3396dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3397dcd989cbSTejun Heo * @cpu: CPU in question 3398dcd989cbSTejun Heo * @wq: target workqueue 3399dcd989cbSTejun Heo * 3400dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3401dcd989cbSTejun Heo * no synchronization around this function and the test result is 3402dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3403dcd989cbSTejun Heo * 3404dcd989cbSTejun Heo * RETURNS: 3405dcd989cbSTejun Heo * %true if congested, %false otherwise. 3406dcd989cbSTejun Heo */ 3407dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3408dcd989cbSTejun Heo { 3409dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3410dcd989cbSTejun Heo 3411dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3412dcd989cbSTejun Heo } 3413dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3414dcd989cbSTejun Heo 3415dcd989cbSTejun Heo /** 3416dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 3417dcd989cbSTejun Heo * @work: the work of interest 3418dcd989cbSTejun Heo * 3419dcd989cbSTejun Heo * RETURNS: 3420bdbc5dd7STejun Heo * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. 3421dcd989cbSTejun Heo */ 3422dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 3423dcd989cbSTejun Heo { 3424dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3425dcd989cbSTejun Heo 3426bdbc5dd7STejun Heo return gcwq ? gcwq->cpu : WORK_CPU_NONE; 3427dcd989cbSTejun Heo } 3428dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 3429dcd989cbSTejun Heo 3430dcd989cbSTejun Heo /** 3431dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3432dcd989cbSTejun Heo * @work: the work to be tested 3433dcd989cbSTejun Heo * 3434dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3435dcd989cbSTejun Heo * synchronization around this function and the test result is 3436dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3437dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 3438dcd989cbSTejun Heo * running state. 3439dcd989cbSTejun Heo * 3440dcd989cbSTejun Heo * RETURNS: 3441dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3442dcd989cbSTejun Heo */ 3443dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3444dcd989cbSTejun Heo { 3445dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3446dcd989cbSTejun Heo unsigned long flags; 3447dcd989cbSTejun Heo unsigned int ret = 0; 3448dcd989cbSTejun Heo 3449dcd989cbSTejun Heo if (!gcwq) 3450999767beSJoonsoo Kim return 0; 3451dcd989cbSTejun Heo 3452dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 3453dcd989cbSTejun Heo 3454dcd989cbSTejun Heo if (work_pending(work)) 3455dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3456dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 3457dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3458dcd989cbSTejun Heo 3459dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 3460dcd989cbSTejun Heo 3461dcd989cbSTejun Heo return ret; 3462dcd989cbSTejun Heo } 3463dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3464dcd989cbSTejun Heo 3465db7bccf4STejun Heo /* 3466db7bccf4STejun Heo * CPU hotplug. 3467db7bccf4STejun Heo * 3468e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3469e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3470e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 3471e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 3472e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 3473e22bee78STejun Heo * blocked draining impractical. 3474e22bee78STejun Heo * 3475628c78e7STejun Heo * This is solved by allowing a gcwq to be disassociated from the CPU 3476628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3477628c78e7STejun Heo * cpu comes back online. 3478db7bccf4STejun Heo */ 3479db7bccf4STejun Heo 348060373152STejun Heo /* claim manager positions of all pools */ 3481b2eb83d1SLai Jiangshan static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq) 348260373152STejun Heo { 348360373152STejun Heo struct worker_pool *pool; 348460373152STejun Heo 348560373152STejun Heo for_each_worker_pool(pool, gcwq) 3486b2eb83d1SLai Jiangshan mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools); 34878db25e78STejun Heo spin_lock_irq(&gcwq->lock); 348860373152STejun Heo } 348960373152STejun Heo 349060373152STejun Heo /* release manager positions */ 3491b2eb83d1SLai Jiangshan static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq) 349260373152STejun Heo { 349360373152STejun Heo struct worker_pool *pool; 349460373152STejun Heo 34958db25e78STejun Heo spin_unlock_irq(&gcwq->lock); 349660373152STejun Heo for_each_worker_pool(pool, gcwq) 3497b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 349860373152STejun Heo } 349960373152STejun Heo 3500628c78e7STejun Heo static void gcwq_unbind_fn(struct work_struct *work) 3501db7bccf4STejun Heo { 3502628c78e7STejun Heo struct global_cwq *gcwq = get_gcwq(smp_processor_id()); 35034ce62e9eSTejun Heo struct worker_pool *pool; 3504db7bccf4STejun Heo struct worker *worker; 3505db7bccf4STejun Heo struct hlist_node *pos; 3506db7bccf4STejun Heo int i; 3507db7bccf4STejun Heo 3508db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3509db7bccf4STejun Heo 3510b2eb83d1SLai Jiangshan gcwq_claim_assoc_and_lock(gcwq); 3511e22bee78STejun Heo 3512f2d5a0eeSTejun Heo /* 3513f2d5a0eeSTejun Heo * We've claimed all manager positions. Make all workers unbound 3514f2d5a0eeSTejun Heo * and set DISASSOCIATED. Before this, all workers except for the 3515f2d5a0eeSTejun Heo * ones which are still executing works from before the last CPU 3516f2d5a0eeSTejun Heo * down must be on the cpu. After this, they may become diasporas. 3517f2d5a0eeSTejun Heo */ 351860373152STejun Heo for_each_worker_pool(pool, gcwq) 35194ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3520403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3521db7bccf4STejun Heo 3522db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 3523403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3524db7bccf4STejun Heo 3525f2d5a0eeSTejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3526f2d5a0eeSTejun Heo 3527b2eb83d1SLai Jiangshan gcwq_release_assoc_and_unlock(gcwq); 3528e22bee78STejun Heo 3529e22bee78STejun Heo /* 3530628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3531628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3532628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3533628c78e7STejun Heo */ 3534628c78e7STejun Heo schedule(); 3535628c78e7STejun Heo 3536628c78e7STejun Heo /* 3537628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3538628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 3539628c78e7STejun Heo * are always true as long as the worklist is not empty. @gcwq now 3540628c78e7STejun Heo * behaves as unbound (in terms of concurrency management) gcwq 3541628c78e7STejun Heo * which is served by workers tied to the CPU. 3542628c78e7STejun Heo * 3543628c78e7STejun Heo * On return from this function, the current worker would trigger 3544628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3545628c78e7STejun Heo * didn't already. 3546e22bee78STejun Heo */ 35474ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 35484ce62e9eSTejun Heo atomic_set(get_pool_nr_running(pool), 0); 3549db7bccf4STejun Heo } 3550db7bccf4STejun Heo 35518db25e78STejun Heo /* 35528db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 35538db25e78STejun Heo * This will be registered high priority CPU notifier. 35548db25e78STejun Heo */ 35559fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 35561da177e4SLinus Torvalds unsigned long action, 35571da177e4SLinus Torvalds void *hcpu) 35581da177e4SLinus Torvalds { 35593af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3560db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 35614ce62e9eSTejun Heo struct worker_pool *pool; 35621da177e4SLinus Torvalds 35638db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 35643af24433SOleg Nesterov case CPU_UP_PREPARE: 35654ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 35663ce63377STejun Heo struct worker *worker; 35673ce63377STejun Heo 35683ce63377STejun Heo if (pool->nr_workers) 35693ce63377STejun Heo continue; 35703ce63377STejun Heo 35713ce63377STejun Heo worker = create_worker(pool); 35723ce63377STejun Heo if (!worker) 35733ce63377STejun Heo return NOTIFY_BAD; 35743ce63377STejun Heo 35753ce63377STejun Heo spin_lock_irq(&gcwq->lock); 35763ce63377STejun Heo start_worker(worker); 35773ce63377STejun Heo spin_unlock_irq(&gcwq->lock); 35783af24433SOleg Nesterov } 35791da177e4SLinus Torvalds break; 35801da177e4SLinus Torvalds 358165758202STejun Heo case CPU_DOWN_FAILED: 358265758202STejun Heo case CPU_ONLINE: 3583b2eb83d1SLai Jiangshan gcwq_claim_assoc_and_lock(gcwq); 35848db25e78STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 35858db25e78STejun Heo rebind_workers(gcwq); 3586b2eb83d1SLai Jiangshan gcwq_release_assoc_and_unlock(gcwq); 35878db25e78STejun Heo break; 358865758202STejun Heo } 358965758202STejun Heo return NOTIFY_OK; 359065758202STejun Heo } 359165758202STejun Heo 359265758202STejun Heo /* 359365758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 359465758202STejun Heo * This will be registered as low priority CPU notifier. 359565758202STejun Heo */ 35969fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 359765758202STejun Heo unsigned long action, 359865758202STejun Heo void *hcpu) 359965758202STejun Heo { 36008db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 36018db25e78STejun Heo struct work_struct unbind_work; 36028db25e78STejun Heo 360365758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 360465758202STejun Heo case CPU_DOWN_PREPARE: 36058db25e78STejun Heo /* unbinding should happen on the local CPU */ 36068db25e78STejun Heo INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn); 36077635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 36088db25e78STejun Heo flush_work(&unbind_work); 36098db25e78STejun Heo break; 361065758202STejun Heo } 361165758202STejun Heo return NOTIFY_OK; 361265758202STejun Heo } 361365758202STejun Heo 36142d3854a3SRusty Russell #ifdef CONFIG_SMP 36158ccad40dSRusty Russell 36162d3854a3SRusty Russell struct work_for_cpu { 3617ed48ece2STejun Heo struct work_struct work; 36182d3854a3SRusty Russell long (*fn)(void *); 36192d3854a3SRusty Russell void *arg; 36202d3854a3SRusty Russell long ret; 36212d3854a3SRusty Russell }; 36222d3854a3SRusty Russell 3623ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 36242d3854a3SRusty Russell { 3625ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3626ed48ece2STejun Heo 36272d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 36282d3854a3SRusty Russell } 36292d3854a3SRusty Russell 36302d3854a3SRusty Russell /** 36312d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 36322d3854a3SRusty Russell * @cpu: the cpu to run on 36332d3854a3SRusty Russell * @fn: the function to run 36342d3854a3SRusty Russell * @arg: the function arg 36352d3854a3SRusty Russell * 363631ad9081SRusty Russell * This will return the value @fn returns. 363731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 36386b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 36392d3854a3SRusty Russell */ 36402d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 36412d3854a3SRusty Russell { 3642ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 36432d3854a3SRusty Russell 3644ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3645ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3646ed48ece2STejun Heo flush_work(&wfc.work); 36472d3854a3SRusty Russell return wfc.ret; 36482d3854a3SRusty Russell } 36492d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 36502d3854a3SRusty Russell #endif /* CONFIG_SMP */ 36512d3854a3SRusty Russell 3652a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3653e7577c50SRusty Russell 3654a0a1a5fdSTejun Heo /** 3655a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3656a0a1a5fdSTejun Heo * 365758a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 365858a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 365958a69cb4STejun Heo * gcwq->worklist. 3660a0a1a5fdSTejun Heo * 3661a0a1a5fdSTejun Heo * CONTEXT: 36628b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3663a0a1a5fdSTejun Heo */ 3664a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3665a0a1a5fdSTejun Heo { 3666a0a1a5fdSTejun Heo unsigned int cpu; 3667a0a1a5fdSTejun Heo 3668a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3669a0a1a5fdSTejun Heo 3670a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3671a0a1a5fdSTejun Heo workqueue_freezing = true; 3672a0a1a5fdSTejun Heo 3673f3421797STejun Heo for_each_gcwq_cpu(cpu) { 36748b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3675bdbc5dd7STejun Heo struct workqueue_struct *wq; 36768b03ae3cSTejun Heo 36778b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 36788b03ae3cSTejun Heo 3679db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3680db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3681db7bccf4STejun Heo 3682a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3683a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3684a0a1a5fdSTejun Heo 368558a69cb4STejun Heo if (cwq && wq->flags & WQ_FREEZABLE) 3686a0a1a5fdSTejun Heo cwq->max_active = 0; 36871da177e4SLinus Torvalds } 36888b03ae3cSTejun Heo 36898b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3690a0a1a5fdSTejun Heo } 3691a0a1a5fdSTejun Heo 3692a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3693a0a1a5fdSTejun Heo } 3694a0a1a5fdSTejun Heo 3695a0a1a5fdSTejun Heo /** 369658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3697a0a1a5fdSTejun Heo * 3698a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3699a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3700a0a1a5fdSTejun Heo * 3701a0a1a5fdSTejun Heo * CONTEXT: 3702a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3703a0a1a5fdSTejun Heo * 3704a0a1a5fdSTejun Heo * RETURNS: 370558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 370658a69cb4STejun Heo * is complete. 3707a0a1a5fdSTejun Heo */ 3708a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3709a0a1a5fdSTejun Heo { 3710a0a1a5fdSTejun Heo unsigned int cpu; 3711a0a1a5fdSTejun Heo bool busy = false; 3712a0a1a5fdSTejun Heo 3713a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3714a0a1a5fdSTejun Heo 3715a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3716a0a1a5fdSTejun Heo 3717f3421797STejun Heo for_each_gcwq_cpu(cpu) { 3718bdbc5dd7STejun Heo struct workqueue_struct *wq; 3719a0a1a5fdSTejun Heo /* 3720a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3721a0a1a5fdSTejun Heo * to peek without lock. 3722a0a1a5fdSTejun Heo */ 3723a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3724a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3725a0a1a5fdSTejun Heo 372658a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3727a0a1a5fdSTejun Heo continue; 3728a0a1a5fdSTejun Heo 3729a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3730a0a1a5fdSTejun Heo if (cwq->nr_active) { 3731a0a1a5fdSTejun Heo busy = true; 3732a0a1a5fdSTejun Heo goto out_unlock; 3733a0a1a5fdSTejun Heo } 3734a0a1a5fdSTejun Heo } 3735a0a1a5fdSTejun Heo } 3736a0a1a5fdSTejun Heo out_unlock: 3737a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3738a0a1a5fdSTejun Heo return busy; 3739a0a1a5fdSTejun Heo } 3740a0a1a5fdSTejun Heo 3741a0a1a5fdSTejun Heo /** 3742a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3743a0a1a5fdSTejun Heo * 3744a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 37457e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3746a0a1a5fdSTejun Heo * 3747a0a1a5fdSTejun Heo * CONTEXT: 37488b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3749a0a1a5fdSTejun Heo */ 3750a0a1a5fdSTejun Heo void thaw_workqueues(void) 3751a0a1a5fdSTejun Heo { 3752a0a1a5fdSTejun Heo unsigned int cpu; 3753a0a1a5fdSTejun Heo 3754a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3755a0a1a5fdSTejun Heo 3756a0a1a5fdSTejun Heo if (!workqueue_freezing) 3757a0a1a5fdSTejun Heo goto out_unlock; 3758a0a1a5fdSTejun Heo 3759f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37608b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 37614ce62e9eSTejun Heo struct worker_pool *pool; 3762bdbc5dd7STejun Heo struct workqueue_struct *wq; 37638b03ae3cSTejun Heo 37648b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 37658b03ae3cSTejun Heo 3766db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3767db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3768db7bccf4STejun Heo 3769a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3770a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3771a0a1a5fdSTejun Heo 377258a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3773a0a1a5fdSTejun Heo continue; 3774a0a1a5fdSTejun Heo 3775a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 37769f4bd4cdSLai Jiangshan cwq_set_max_active(cwq, wq->saved_max_active); 3777a0a1a5fdSTejun Heo } 37788b03ae3cSTejun Heo 37794ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 37804ce62e9eSTejun Heo wake_up_worker(pool); 3781e22bee78STejun Heo 37828b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3783a0a1a5fdSTejun Heo } 3784a0a1a5fdSTejun Heo 3785a0a1a5fdSTejun Heo workqueue_freezing = false; 3786a0a1a5fdSTejun Heo out_unlock: 3787a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3788a0a1a5fdSTejun Heo } 3789a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3790a0a1a5fdSTejun Heo 37916ee0578bSSuresh Siddha static int __init init_workqueues(void) 37921da177e4SLinus Torvalds { 3793c34056a3STejun Heo unsigned int cpu; 3794c34056a3STejun Heo 3795b5490077STejun Heo /* make sure we have enough bits for OFFQ CPU number */ 3796b5490077STejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) < 3797b5490077STejun Heo WORK_CPU_LAST); 3798b5490077STejun Heo 379965758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3800a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 38018b03ae3cSTejun Heo 38028b03ae3cSTejun Heo /* initialize gcwqs */ 3803f3421797STejun Heo for_each_gcwq_cpu(cpu) { 38048b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38054ce62e9eSTejun Heo struct worker_pool *pool; 38068b03ae3cSTejun Heo 38078b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 38088b03ae3cSTejun Heo gcwq->cpu = cpu; 3809f3421797STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 38108b03ae3cSTejun Heo 381142f8570fSSasha Levin hash_init(gcwq->busy_hash); 3812c8e55f36STejun Heo 38134ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 38144ce62e9eSTejun Heo pool->gcwq = gcwq; 38154ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 38164ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3817e22bee78STejun Heo 38184ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 38194ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 38204ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3821e22bee78STejun Heo 38224ce62e9eSTejun Heo setup_timer(&pool->mayday_timer, gcwq_mayday_timeout, 38234ce62e9eSTejun Heo (unsigned long)pool); 38244ce62e9eSTejun Heo 3825b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 38264ce62e9eSTejun Heo ida_init(&pool->worker_ida); 38274ce62e9eSTejun Heo } 38288b03ae3cSTejun Heo } 38298b03ae3cSTejun Heo 3830e22bee78STejun Heo /* create the initial worker */ 3831f3421797STejun Heo for_each_online_gcwq_cpu(cpu) { 3832e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38334ce62e9eSTejun Heo struct worker_pool *pool; 3834e22bee78STejun Heo 3835477a3c33STejun Heo if (cpu != WORK_CPU_UNBOUND) 3836477a3c33STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 38374ce62e9eSTejun Heo 38384ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 38394ce62e9eSTejun Heo struct worker *worker; 38404ce62e9eSTejun Heo 3841bc2ae0f5STejun Heo worker = create_worker(pool); 3842e22bee78STejun Heo BUG_ON(!worker); 3843e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3844e22bee78STejun Heo start_worker(worker); 3845e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3846e22bee78STejun Heo } 38474ce62e9eSTejun Heo } 3848e22bee78STejun Heo 3849d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 38501aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3851d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3852f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3853f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 385424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 385524d51addSTejun Heo WQ_FREEZABLE, 0); 38561aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3857ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 38586ee0578bSSuresh Siddha return 0; 38591da177e4SLinus Torvalds } 38606ee0578bSSuresh Siddha early_initcall(init_workqueues); 3861