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> 44e22bee78STejun Heo 45e22bee78STejun Heo #include "workqueue_sched.h" 461da177e4SLinus Torvalds 47c8e55f36STejun Heo enum { 48bc2ae0f5STejun Heo /* 49bc2ae0f5STejun Heo * global_cwq flags 50bc2ae0f5STejun Heo * 51bc2ae0f5STejun Heo * A bound gcwq is either associated or disassociated with its CPU. 52bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 53bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 54bc2ae0f5STejun Heo * is in effect. 55bc2ae0f5STejun Heo * 56bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 57bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 58bc2ae0f5STejun Heo * be executing on any CPU. The gcwq behaves as an unbound one. 59bc2ae0f5STejun Heo * 60bc2ae0f5STejun Heo * Note that DISASSOCIATED can be flipped only while holding 61bc2ae0f5STejun Heo * managership of all pools on the gcwq to avoid changing binding 62bc2ae0f5STejun Heo * state while create_worker() is in progress. 63bc2ae0f5STejun Heo */ 6411ebea50STejun Heo GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */ 6511ebea50STejun Heo GCWQ_FREEZING = 1 << 1, /* freeze in progress */ 6611ebea50STejun Heo 6711ebea50STejun Heo /* pool flags */ 6811ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 69552a37e9SLai Jiangshan POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */ 70db7bccf4STejun Heo 71c8e55f36STejun Heo /* worker flags */ 72c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 73c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 74c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 75e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 76e22bee78STejun Heo WORKER_REBIND = 1 << 5, /* mom is home, come back */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79e22bee78STejun Heo 80403c821dSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_REBIND | 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 */ 86c8e55f36STejun Heo BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, 87c8e55f36STejun Heo BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, 88db7bccf4STejun Heo 89e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 90e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 91e22bee78STejun Heo 923233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 933233cdbdSTejun Heo /* call for help after 10ms 943233cdbdSTejun Heo (min two ticks) */ 95e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 96e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* 99e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 100e22bee78STejun Heo * all cpus. Give -20. 101e22bee78STejun Heo */ 102e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1033270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 104c8e55f36STejun Heo }; 105c8e55f36STejun Heo 1061da177e4SLinus Torvalds /* 1074690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1084690c4abSTejun Heo * 109e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 110e41e704bSTejun Heo * everyone else. 1114690c4abSTejun Heo * 112e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 113e22bee78STejun Heo * only be modified and accessed from the local cpu. 114e22bee78STejun Heo * 1158b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 1164690c4abSTejun Heo * 117e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 118e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 119e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 120f3421797STejun Heo * If GCWQ_DISASSOCIATED is set, it's identical to L. 121e22bee78STejun Heo * 12273f53c4aSTejun Heo * F: wq->flush_mutex protected. 12373f53c4aSTejun Heo * 1244690c4abSTejun Heo * W: workqueue_lock protected. 1254690c4abSTejun Heo */ 1264690c4abSTejun Heo 1278b03ae3cSTejun Heo struct global_cwq; 128bd7bdd43STejun Heo struct worker_pool; 129c34056a3STejun Heo 130e22bee78STejun Heo /* 131e22bee78STejun Heo * The poor guys doing the actual heavy lifting. All on-duty workers 132e22bee78STejun Heo * are either serving the manager role, on idle list or on busy hash. 133e22bee78STejun Heo */ 134c34056a3STejun Heo struct worker { 135c8e55f36STejun Heo /* on idle list while idle, on busy hash table while busy */ 136c8e55f36STejun Heo union { 137c8e55f36STejun Heo struct list_head entry; /* L: while idle */ 138c8e55f36STejun Heo struct hlist_node hentry; /* L: while busy */ 139c8e55f36STejun Heo }; 140c8e55f36STejun Heo 141c34056a3STejun Heo struct work_struct *current_work; /* L: work being processed */ 1428cca0eeaSTejun Heo struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ 143affee4b2STejun Heo struct list_head scheduled; /* L: scheduled works */ 144c34056a3STejun Heo struct task_struct *task; /* I: worker task */ 145bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 146e22bee78STejun Heo /* 64 bytes boundary on 64bit, 32 on 32bit */ 147e22bee78STejun Heo unsigned long last_active; /* L: last active timestamp */ 148e22bee78STejun Heo unsigned int flags; /* X: flags */ 149c34056a3STejun Heo int id; /* I: worker id */ 15025511a47STejun Heo 15125511a47STejun Heo /* for rebinding worker to CPU */ 15225511a47STejun Heo struct work_struct rebind_work; /* L: for busy worker */ 153c34056a3STejun Heo }; 154c34056a3STejun Heo 155bd7bdd43STejun Heo struct worker_pool { 156bd7bdd43STejun Heo struct global_cwq *gcwq; /* I: the owning gcwq */ 15711ebea50STejun Heo unsigned int flags; /* X: flags */ 158bd7bdd43STejun Heo 159bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 160bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 161ea1abd61SLai Jiangshan 162ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 163bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 164bd7bdd43STejun Heo 165bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 166bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 167bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 168bd7bdd43STejun Heo 16960373152STejun Heo struct mutex manager_mutex; /* mutex manager should hold */ 170bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 171bd7bdd43STejun Heo }; 172bd7bdd43STejun Heo 1734690c4abSTejun Heo /* 174e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 175e22bee78STejun Heo * and all works are queued and processed here regardless of their 176e22bee78STejun Heo * target workqueues. 1778b03ae3cSTejun Heo */ 1788b03ae3cSTejun Heo struct global_cwq { 1798b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1808b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 181db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 182c8e55f36STejun Heo 183bd7bdd43STejun Heo /* workers are chained either in busy_hash or pool idle_list */ 184c8e55f36STejun Heo struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; 185c8e55f36STejun Heo /* L: hash of busy workers */ 186c8e55f36STejun Heo 187330dad5bSJoonsoo Kim struct worker_pool pools[NR_WORKER_POOLS]; 188330dad5bSJoonsoo Kim /* normal and highpri pools */ 1898b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1908b03ae3cSTejun Heo 1918b03ae3cSTejun Heo /* 192502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1930f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1940f900049STejun Heo * aligned at two's power of the number of flag bits. 1951da177e4SLinus Torvalds */ 1961da177e4SLinus Torvalds struct cpu_workqueue_struct { 197bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1984690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19973f53c4aSTejun Heo int work_color; /* L: current color */ 20073f53c4aSTejun Heo int flush_color; /* L: flushing color */ 20173f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20273f53c4aSTejun Heo /* L: nr of in_flight works */ 2031e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 204a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2051e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2060f900049STejun Heo }; 2071da177e4SLinus Torvalds 2081da177e4SLinus Torvalds /* 20973f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21073f53c4aSTejun Heo */ 21173f53c4aSTejun Heo struct wq_flusher { 21273f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 21373f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 21473f53c4aSTejun Heo struct completion done; /* flush completion */ 21573f53c4aSTejun Heo }; 2161da177e4SLinus Torvalds 21773f53c4aSTejun Heo /* 218f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 219f2e005aaSTejun Heo * used to determine whether there's something to be done. 220f2e005aaSTejun Heo */ 221f2e005aaSTejun Heo #ifdef CONFIG_SMP 222f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 223f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 224f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 225f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 226f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 2279c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 228f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 229f2e005aaSTejun Heo #else 230f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 231f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 232f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 233f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 234f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 235f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 236f2e005aaSTejun Heo #endif 2371da177e4SLinus Torvalds 2381da177e4SLinus Torvalds /* 2391da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2401da177e4SLinus Torvalds * per-CPU workqueues: 2411da177e4SLinus Torvalds */ 2421da177e4SLinus Torvalds struct workqueue_struct { 2439c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 244bdbc5dd7STejun Heo union { 245bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 246bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 247bdbc5dd7STejun Heo unsigned long v; 248bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2494690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 25073f53c4aSTejun Heo 25173f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 25273f53c4aSTejun Heo int work_color; /* F: current work color */ 25373f53c4aSTejun Heo int flush_color; /* F: current flush color */ 25473f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 25573f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 25673f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 25773f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 25873f53c4aSTejun Heo 259f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 260e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 261e22bee78STejun Heo 2629c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 263dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2644e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2654e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2664e6045f1SJohannes Berg #endif 267b196be89STejun Heo char name[]; /* I: workqueue name */ 2681da177e4SLinus Torvalds }; 2691da177e4SLinus Torvalds 270d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 271d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 272044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2731aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 274044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 275d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 276044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 277f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 278044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 27924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 280d320c038STejun Heo 28197bd2347STejun Heo #define CREATE_TRACE_POINTS 28297bd2347STejun Heo #include <trace/events/workqueue.h> 28397bd2347STejun Heo 2844ce62e9eSTejun Heo #define for_each_worker_pool(pool, gcwq) \ 2853270476aSTejun Heo for ((pool) = &(gcwq)->pools[0]; \ 2863270476aSTejun Heo (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++) 2874ce62e9eSTejun Heo 288db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 289db7bccf4STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ 290db7bccf4STejun Heo hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) 291db7bccf4STejun Heo 292f3421797STejun Heo static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, 293f3421797STejun Heo unsigned int sw) 294f3421797STejun Heo { 295f3421797STejun Heo if (cpu < nr_cpu_ids) { 296f3421797STejun Heo if (sw & 1) { 297f3421797STejun Heo cpu = cpumask_next(cpu, mask); 298f3421797STejun Heo if (cpu < nr_cpu_ids) 299f3421797STejun Heo return cpu; 300f3421797STejun Heo } 301f3421797STejun Heo if (sw & 2) 302f3421797STejun Heo return WORK_CPU_UNBOUND; 303f3421797STejun Heo } 304f3421797STejun Heo return WORK_CPU_NONE; 305f3421797STejun Heo } 306f3421797STejun Heo 307f3421797STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 308f3421797STejun Heo struct workqueue_struct *wq) 309f3421797STejun Heo { 310f3421797STejun Heo return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 311f3421797STejun Heo } 312f3421797STejun Heo 31309884951STejun Heo /* 31409884951STejun Heo * CPU iterators 31509884951STejun Heo * 31609884951STejun Heo * An extra gcwq is defined for an invalid cpu number 31709884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 31809884951STejun Heo * specific CPU. The following iterators are similar to 31909884951STejun Heo * for_each_*_cpu() iterators but also considers the unbound gcwq. 32009884951STejun Heo * 32109884951STejun Heo * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND 32209884951STejun Heo * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND 32309884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 32409884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 32509884951STejun Heo */ 326f3421797STejun Heo #define for_each_gcwq_cpu(cpu) \ 327f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ 328f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 329f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) 330f3421797STejun Heo 331f3421797STejun Heo #define for_each_online_gcwq_cpu(cpu) \ 332f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ 333f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 334f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) 335f3421797STejun Heo 336f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 337f3421797STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ 338f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 339f3421797STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) 340f3421797STejun Heo 341dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 342dc186ad7SThomas Gleixner 343dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 344dc186ad7SThomas Gleixner 34599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 34699777288SStanislaw Gruszka { 34799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 34899777288SStanislaw Gruszka } 34999777288SStanislaw Gruszka 350dc186ad7SThomas Gleixner /* 351dc186ad7SThomas Gleixner * fixup_init is called when: 352dc186ad7SThomas Gleixner * - an active object is initialized 353dc186ad7SThomas Gleixner */ 354dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 355dc186ad7SThomas Gleixner { 356dc186ad7SThomas Gleixner struct work_struct *work = addr; 357dc186ad7SThomas Gleixner 358dc186ad7SThomas Gleixner switch (state) { 359dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 360dc186ad7SThomas Gleixner cancel_work_sync(work); 361dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 362dc186ad7SThomas Gleixner return 1; 363dc186ad7SThomas Gleixner default: 364dc186ad7SThomas Gleixner return 0; 365dc186ad7SThomas Gleixner } 366dc186ad7SThomas Gleixner } 367dc186ad7SThomas Gleixner 368dc186ad7SThomas Gleixner /* 369dc186ad7SThomas Gleixner * fixup_activate is called when: 370dc186ad7SThomas Gleixner * - an active object is activated 371dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 372dc186ad7SThomas Gleixner */ 373dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 374dc186ad7SThomas Gleixner { 375dc186ad7SThomas Gleixner struct work_struct *work = addr; 376dc186ad7SThomas Gleixner 377dc186ad7SThomas Gleixner switch (state) { 378dc186ad7SThomas Gleixner 379dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 380dc186ad7SThomas Gleixner /* 381dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 382dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 383dc186ad7SThomas Gleixner * is tracked in the object tracker. 384dc186ad7SThomas Gleixner */ 38522df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 386dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 387dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 388dc186ad7SThomas Gleixner return 0; 389dc186ad7SThomas Gleixner } 390dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 391dc186ad7SThomas Gleixner return 0; 392dc186ad7SThomas Gleixner 393dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 394dc186ad7SThomas Gleixner WARN_ON(1); 395dc186ad7SThomas Gleixner 396dc186ad7SThomas Gleixner default: 397dc186ad7SThomas Gleixner return 0; 398dc186ad7SThomas Gleixner } 399dc186ad7SThomas Gleixner } 400dc186ad7SThomas Gleixner 401dc186ad7SThomas Gleixner /* 402dc186ad7SThomas Gleixner * fixup_free is called when: 403dc186ad7SThomas Gleixner * - an active object is freed 404dc186ad7SThomas Gleixner */ 405dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 406dc186ad7SThomas Gleixner { 407dc186ad7SThomas Gleixner struct work_struct *work = addr; 408dc186ad7SThomas Gleixner 409dc186ad7SThomas Gleixner switch (state) { 410dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 411dc186ad7SThomas Gleixner cancel_work_sync(work); 412dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 413dc186ad7SThomas Gleixner return 1; 414dc186ad7SThomas Gleixner default: 415dc186ad7SThomas Gleixner return 0; 416dc186ad7SThomas Gleixner } 417dc186ad7SThomas Gleixner } 418dc186ad7SThomas Gleixner 419dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 420dc186ad7SThomas Gleixner .name = "work_struct", 42199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 422dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 423dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 424dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 425dc186ad7SThomas Gleixner }; 426dc186ad7SThomas Gleixner 427dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 428dc186ad7SThomas Gleixner { 429dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 430dc186ad7SThomas Gleixner } 431dc186ad7SThomas Gleixner 432dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 433dc186ad7SThomas Gleixner { 434dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 435dc186ad7SThomas Gleixner } 436dc186ad7SThomas Gleixner 437dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 438dc186ad7SThomas Gleixner { 439dc186ad7SThomas Gleixner if (onstack) 440dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 441dc186ad7SThomas Gleixner else 442dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 443dc186ad7SThomas Gleixner } 444dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 445dc186ad7SThomas Gleixner 446dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 447dc186ad7SThomas Gleixner { 448dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 449dc186ad7SThomas Gleixner } 450dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 451dc186ad7SThomas Gleixner 452dc186ad7SThomas Gleixner #else 453dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 454dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 455dc186ad7SThomas Gleixner #endif 456dc186ad7SThomas Gleixner 45795402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 45895402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4591da177e4SLinus Torvalds static LIST_HEAD(workqueues); 460a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4611da177e4SLinus Torvalds 46214441960SOleg Nesterov /* 463e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 464e22bee78STejun Heo * which is expected to be used frequently by other cpus via 465e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 46614441960SOleg Nesterov */ 4678b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 4684ce62e9eSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]); 469f756d5e2SNathan Lynch 470f3421797STejun Heo /* 471f3421797STejun Heo * Global cpu workqueue and nr_running counter for unbound gcwq. The 472f3421797STejun Heo * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its 473f3421797STejun Heo * workers have WORKER_UNBOUND set. 474f3421797STejun Heo */ 475f3421797STejun Heo static struct global_cwq unbound_global_cwq; 4764ce62e9eSTejun Heo static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = { 4774ce62e9eSTejun Heo [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */ 4784ce62e9eSTejun Heo }; 479f3421797STejun Heo 480c34056a3STejun Heo static int worker_thread(void *__worker); 4811da177e4SLinus Torvalds 4823270476aSTejun Heo static int worker_pool_pri(struct worker_pool *pool) 4833270476aSTejun Heo { 4843270476aSTejun Heo return pool - pool->gcwq->pools; 4853270476aSTejun Heo } 4863270476aSTejun Heo 4878b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 4881da177e4SLinus Torvalds { 489f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 4908b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 491f3421797STejun Heo else 492f3421797STejun Heo return &unbound_global_cwq; 4931da177e4SLinus Torvalds } 4941da177e4SLinus Torvalds 49563d95a91STejun Heo static atomic_t *get_pool_nr_running(struct worker_pool *pool) 496b1f4ec17SOleg Nesterov { 49763d95a91STejun Heo int cpu = pool->gcwq->cpu; 4983270476aSTejun Heo int idx = worker_pool_pri(pool); 49963d95a91STejun Heo 500f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 5014ce62e9eSTejun Heo return &per_cpu(pool_nr_running, cpu)[idx]; 502f3421797STejun Heo else 5034ce62e9eSTejun Heo return &unbound_pool_nr_running[idx]; 504b1f4ec17SOleg Nesterov } 505b1f4ec17SOleg Nesterov 5064690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 5074690c4abSTejun Heo struct workqueue_struct *wq) 508a848e3b6SOleg Nesterov { 509f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 510e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 511bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 512f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 513f3421797STejun Heo return wq->cpu_wq.single; 514f3421797STejun Heo return NULL; 515f3421797STejun Heo } 516a848e3b6SOleg Nesterov 51773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 51873f53c4aSTejun Heo { 51973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 52073f53c4aSTejun Heo } 52173f53c4aSTejun Heo 52273f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 52373f53c4aSTejun Heo { 52473f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 52573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 52673f53c4aSTejun Heo } 52773f53c4aSTejun Heo 52873f53c4aSTejun Heo static int work_next_color(int color) 52973f53c4aSTejun Heo { 53073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 5311da177e4SLinus Torvalds } 5321da177e4SLinus Torvalds 5334594bf15SDavid Howells /* 534b5490077STejun Heo * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data 535b5490077STejun Heo * contain the pointer to the queued cwq. Once execution starts, the flag 536b5490077STejun Heo * is cleared and the high bits contain OFFQ flags and CPU number. 5377a22ad75STejun Heo * 538bbb68dfaSTejun Heo * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling() 539bbb68dfaSTejun Heo * and clear_work_data() can be used to set the cwq, cpu or clear 540bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 541bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5427a22ad75STejun Heo * 543bbb68dfaSTejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to 544bbb68dfaSTejun Heo * a work. gcwq is available once the work has been queued anywhere after 545bbb68dfaSTejun Heo * initialization until it is sync canceled. cwq is available only while 546bbb68dfaSTejun Heo * the work item is queued. 547bbb68dfaSTejun Heo * 548bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 549bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 550bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 551bbb68dfaSTejun Heo * try to steal the PENDING bit. 5524594bf15SDavid Howells */ 5537a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5547a22ad75STejun Heo unsigned long flags) 5557a22ad75STejun Heo { 5567a22ad75STejun Heo BUG_ON(!work_pending(work)); 5577a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5587a22ad75STejun Heo } 5597a22ad75STejun Heo 5607a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5614690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5624690c4abSTejun Heo unsigned long extra_flags) 563365970a1SDavid Howells { 5647a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 565e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 566365970a1SDavid Howells } 567365970a1SDavid Howells 5688930cabaSTejun Heo static void set_work_cpu_and_clear_pending(struct work_struct *work, 5698930cabaSTejun Heo unsigned int cpu) 5704d707b9fSOleg Nesterov { 57123657bb1STejun Heo /* 57223657bb1STejun Heo * The following wmb is paired with the implied mb in 57323657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 57423657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 57523657bb1STejun Heo * owner. 57623657bb1STejun Heo */ 57723657bb1STejun Heo smp_wmb(); 578b5490077STejun Heo set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0); 5794d707b9fSOleg Nesterov } 5804d707b9fSOleg Nesterov 5817a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 582365970a1SDavid Howells { 58323657bb1STejun Heo smp_wmb(); /* see set_work_cpu_and_clear_pending() */ 5847a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 5857a22ad75STejun Heo } 5867a22ad75STejun Heo 5877a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5887a22ad75STejun Heo { 589e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5907a22ad75STejun Heo 591e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 592e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 593e120153dSTejun Heo else 594e120153dSTejun Heo return NULL; 5957a22ad75STejun Heo } 5967a22ad75STejun Heo 5977a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 5987a22ad75STejun Heo { 599e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6007a22ad75STejun Heo unsigned int cpu; 6017a22ad75STejun Heo 602e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 603e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 604bd7bdd43STejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq; 6057a22ad75STejun Heo 606b5490077STejun Heo cpu = data >> WORK_OFFQ_CPU_SHIFT; 607bdbc5dd7STejun Heo if (cpu == WORK_CPU_NONE) 6087a22ad75STejun Heo return NULL; 6097a22ad75STejun Heo 610f3421797STejun Heo BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); 6117a22ad75STejun Heo return get_gcwq(cpu); 612365970a1SDavid Howells } 613365970a1SDavid Howells 614bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 615bbb68dfaSTejun Heo { 616bbb68dfaSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 617bbb68dfaSTejun Heo unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE; 618bbb68dfaSTejun Heo 619bbb68dfaSTejun Heo set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING, 620bbb68dfaSTejun Heo WORK_STRUCT_PENDING); 621bbb68dfaSTejun Heo } 622bbb68dfaSTejun Heo 623bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 624bbb68dfaSTejun Heo { 625bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 626bbb68dfaSTejun Heo 627bbb68dfaSTejun Heo return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING); 628bbb68dfaSTejun Heo } 629bbb68dfaSTejun Heo 630e22bee78STejun Heo /* 6313270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6323270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 6333270476aSTejun Heo * they're being called with gcwq->lock held. 634e22bee78STejun Heo */ 635e22bee78STejun Heo 63663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 637649027d7STejun Heo { 6383270476aSTejun Heo return !atomic_read(get_pool_nr_running(pool)); 639649027d7STejun Heo } 640649027d7STejun Heo 641e22bee78STejun Heo /* 642e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 643e22bee78STejun Heo * running workers. 644974271c4STejun Heo * 645974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 646974271c4STejun Heo * function will always return %true for unbound gcwq as long as the 647974271c4STejun Heo * worklist isn't empty. 648e22bee78STejun Heo */ 64963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 650e22bee78STejun Heo { 65163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 652e22bee78STejun Heo } 653e22bee78STejun Heo 654e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 65563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 656e22bee78STejun Heo { 65763d95a91STejun Heo return pool->nr_idle; 658e22bee78STejun Heo } 659e22bee78STejun Heo 660e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 66163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 662e22bee78STejun Heo { 66363d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 664e22bee78STejun Heo 6653270476aSTejun Heo return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1; 666e22bee78STejun Heo } 667e22bee78STejun Heo 668e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 66963d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 670e22bee78STejun Heo { 67163d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 672e22bee78STejun Heo } 673e22bee78STejun Heo 674e22bee78STejun Heo /* Do I need to be the manager? */ 67563d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 676e22bee78STejun Heo { 67763d95a91STejun Heo return need_to_create_worker(pool) || 67811ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 679e22bee78STejun Heo } 680e22bee78STejun Heo 681e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 68263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 683e22bee78STejun Heo { 684552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 68563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 68663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 687e22bee78STejun Heo 688ea1abd61SLai Jiangshan /* 689ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 690ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 691ea1abd61SLai Jiangshan */ 692ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 693ea1abd61SLai Jiangshan return false; 694ea1abd61SLai Jiangshan 695e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 696e22bee78STejun Heo } 697e22bee78STejun Heo 698e22bee78STejun Heo /* 699e22bee78STejun Heo * Wake up functions. 700e22bee78STejun Heo */ 701e22bee78STejun Heo 7027e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 70363d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7047e11629dSTejun Heo { 70563d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7067e11629dSTejun Heo return NULL; 7077e11629dSTejun Heo 70863d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7097e11629dSTejun Heo } 7107e11629dSTejun Heo 7117e11629dSTejun Heo /** 7127e11629dSTejun Heo * wake_up_worker - wake up an idle worker 71363d95a91STejun Heo * @pool: worker pool to wake worker from 7147e11629dSTejun Heo * 71563d95a91STejun Heo * Wake up the first idle worker of @pool. 7167e11629dSTejun Heo * 7177e11629dSTejun Heo * CONTEXT: 7187e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 7197e11629dSTejun Heo */ 72063d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7217e11629dSTejun Heo { 72263d95a91STejun Heo struct worker *worker = first_worker(pool); 7237e11629dSTejun Heo 7247e11629dSTejun Heo if (likely(worker)) 7257e11629dSTejun Heo wake_up_process(worker->task); 7267e11629dSTejun Heo } 7277e11629dSTejun Heo 7284690c4abSTejun Heo /** 729e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 730e22bee78STejun Heo * @task: task waking up 731e22bee78STejun Heo * @cpu: CPU @task is waking up to 732e22bee78STejun Heo * 733e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 734e22bee78STejun Heo * being awoken. 735e22bee78STejun Heo * 736e22bee78STejun Heo * CONTEXT: 737e22bee78STejun Heo * spin_lock_irq(rq->lock) 738e22bee78STejun Heo */ 739e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 740e22bee78STejun Heo { 741e22bee78STejun Heo struct worker *worker = kthread_data(task); 742e22bee78STejun Heo 7432d64672eSSteven Rostedt if (!(worker->flags & WORKER_NOT_RUNNING)) 74463d95a91STejun Heo atomic_inc(get_pool_nr_running(worker->pool)); 745e22bee78STejun Heo } 746e22bee78STejun Heo 747e22bee78STejun Heo /** 748e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 749e22bee78STejun Heo * @task: task going to sleep 750e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 751e22bee78STejun Heo * 752e22bee78STejun Heo * This function is called during schedule() when a busy worker is 753e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 754e22bee78STejun Heo * returning pointer to its task. 755e22bee78STejun Heo * 756e22bee78STejun Heo * CONTEXT: 757e22bee78STejun Heo * spin_lock_irq(rq->lock) 758e22bee78STejun Heo * 759e22bee78STejun Heo * RETURNS: 760e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 761e22bee78STejun Heo */ 762e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 763e22bee78STejun Heo unsigned int cpu) 764e22bee78STejun Heo { 765e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 766bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 76763d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 768e22bee78STejun Heo 7692d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 770e22bee78STejun Heo return NULL; 771e22bee78STejun Heo 772e22bee78STejun Heo /* this can only happen on the local cpu */ 773e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 774e22bee78STejun Heo 775e22bee78STejun Heo /* 776e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 777e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 778e22bee78STejun Heo * Please read comment there. 779e22bee78STejun Heo * 780628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 781628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 782628c78e7STejun Heo * disabled, which in turn means that none else could be 783628c78e7STejun Heo * manipulating idle_list, so dereferencing idle_list without gcwq 784628c78e7STejun Heo * lock is safe. 785e22bee78STejun Heo */ 786bd7bdd43STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist)) 78763d95a91STejun Heo to_wakeup = first_worker(pool); 788e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 789e22bee78STejun Heo } 790e22bee78STejun Heo 791e22bee78STejun Heo /** 792e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 793cb444766STejun Heo * @worker: self 794d302f017STejun Heo * @flags: flags to set 795d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 796d302f017STejun Heo * 797e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 798e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 799e22bee78STejun Heo * woken up. 800d302f017STejun Heo * 801cb444766STejun Heo * CONTEXT: 802cb444766STejun Heo * spin_lock_irq(gcwq->lock) 803d302f017STejun Heo */ 804d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 805d302f017STejun Heo bool wakeup) 806d302f017STejun Heo { 807bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 808e22bee78STejun Heo 809cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 810cb444766STejun Heo 811e22bee78STejun Heo /* 812e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 813e22bee78STejun Heo * wake up an idle worker as necessary if requested by 814e22bee78STejun Heo * @wakeup. 815e22bee78STejun Heo */ 816e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 817e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 81863d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 819e22bee78STejun Heo 820e22bee78STejun Heo if (wakeup) { 821e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 822bd7bdd43STejun Heo !list_empty(&pool->worklist)) 82363d95a91STejun Heo wake_up_worker(pool); 824e22bee78STejun Heo } else 825e22bee78STejun Heo atomic_dec(nr_running); 826e22bee78STejun Heo } 827e22bee78STejun Heo 828d302f017STejun Heo worker->flags |= flags; 829d302f017STejun Heo } 830d302f017STejun Heo 831d302f017STejun Heo /** 832e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 833cb444766STejun Heo * @worker: self 834d302f017STejun Heo * @flags: flags to clear 835d302f017STejun Heo * 836e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 837d302f017STejun Heo * 838cb444766STejun Heo * CONTEXT: 839cb444766STejun Heo * spin_lock_irq(gcwq->lock) 840d302f017STejun Heo */ 841d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 842d302f017STejun Heo { 84363d95a91STejun Heo struct worker_pool *pool = worker->pool; 844e22bee78STejun Heo unsigned int oflags = worker->flags; 845e22bee78STejun Heo 846cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 847cb444766STejun Heo 848d302f017STejun Heo worker->flags &= ~flags; 849e22bee78STejun Heo 85042c025f3STejun Heo /* 85142c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 85242c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 85342c025f3STejun Heo * of multiple flags, not a single flag. 85442c025f3STejun Heo */ 855e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 856e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 85763d95a91STejun Heo atomic_inc(get_pool_nr_running(pool)); 858d302f017STejun Heo } 859d302f017STejun Heo 860d302f017STejun Heo /** 861c8e55f36STejun Heo * busy_worker_head - return the busy hash head for a work 862c8e55f36STejun Heo * @gcwq: gcwq of interest 863c8e55f36STejun Heo * @work: work to be hashed 864c8e55f36STejun Heo * 865c8e55f36STejun Heo * Return hash head of @gcwq for @work. 866c8e55f36STejun Heo * 867c8e55f36STejun Heo * CONTEXT: 868c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 869c8e55f36STejun Heo * 870c8e55f36STejun Heo * RETURNS: 871c8e55f36STejun Heo * Pointer to the hash head. 872c8e55f36STejun Heo */ 873c8e55f36STejun Heo static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, 874c8e55f36STejun Heo struct work_struct *work) 875c8e55f36STejun Heo { 876c8e55f36STejun Heo const int base_shift = ilog2(sizeof(struct work_struct)); 877c8e55f36STejun Heo unsigned long v = (unsigned long)work; 878c8e55f36STejun Heo 879c8e55f36STejun Heo /* simple shift and fold hash, do we need something better? */ 880c8e55f36STejun Heo v >>= base_shift; 881c8e55f36STejun Heo v += v >> BUSY_WORKER_HASH_ORDER; 882c8e55f36STejun Heo v &= BUSY_WORKER_HASH_MASK; 883c8e55f36STejun Heo 884c8e55f36STejun Heo return &gcwq->busy_hash[v]; 885c8e55f36STejun Heo } 886c8e55f36STejun Heo 887c8e55f36STejun Heo /** 8888cca0eeaSTejun Heo * __find_worker_executing_work - find worker which is executing a work 8898cca0eeaSTejun Heo * @gcwq: gcwq of interest 8908cca0eeaSTejun Heo * @bwh: hash head as returned by busy_worker_head() 8918cca0eeaSTejun Heo * @work: work to find worker for 8928cca0eeaSTejun Heo * 8938cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. @bwh should be 8948cca0eeaSTejun Heo * the hash head obtained by calling busy_worker_head() with the same 8958cca0eeaSTejun Heo * work. 8968cca0eeaSTejun Heo * 8978cca0eeaSTejun Heo * CONTEXT: 8988cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8998cca0eeaSTejun Heo * 9008cca0eeaSTejun Heo * RETURNS: 9018cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9028cca0eeaSTejun Heo * otherwise. 9038cca0eeaSTejun Heo */ 9048cca0eeaSTejun Heo static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, 9058cca0eeaSTejun Heo struct hlist_head *bwh, 9068cca0eeaSTejun Heo struct work_struct *work) 9078cca0eeaSTejun Heo { 9088cca0eeaSTejun Heo struct worker *worker; 9098cca0eeaSTejun Heo struct hlist_node *tmp; 9108cca0eeaSTejun Heo 9118cca0eeaSTejun Heo hlist_for_each_entry(worker, tmp, bwh, hentry) 9128cca0eeaSTejun Heo if (worker->current_work == work) 9138cca0eeaSTejun Heo return worker; 9148cca0eeaSTejun Heo return NULL; 9158cca0eeaSTejun Heo } 9168cca0eeaSTejun Heo 9178cca0eeaSTejun Heo /** 9188cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 9198cca0eeaSTejun Heo * @gcwq: gcwq of interest 9208cca0eeaSTejun Heo * @work: work to find worker for 9218cca0eeaSTejun Heo * 9228cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. This function is 9238cca0eeaSTejun Heo * identical to __find_worker_executing_work() except that this 9248cca0eeaSTejun Heo * function calculates @bwh itself. 9258cca0eeaSTejun Heo * 9268cca0eeaSTejun Heo * CONTEXT: 9278cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 9288cca0eeaSTejun Heo * 9298cca0eeaSTejun Heo * RETURNS: 9308cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9318cca0eeaSTejun Heo * otherwise. 9328cca0eeaSTejun Heo */ 9338cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 9348cca0eeaSTejun Heo struct work_struct *work) 9358cca0eeaSTejun Heo { 9368cca0eeaSTejun Heo return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), 9378cca0eeaSTejun Heo work); 9388cca0eeaSTejun Heo } 9398cca0eeaSTejun Heo 9408cca0eeaSTejun Heo /** 941bf4ede01STejun Heo * move_linked_works - move linked works to a list 942bf4ede01STejun Heo * @work: start of series of works to be scheduled 943bf4ede01STejun Heo * @head: target list to append @work to 944bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 945bf4ede01STejun Heo * 946bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 947bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 948bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 949bf4ede01STejun Heo * 950bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 951bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 952bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 953bf4ede01STejun Heo * 954bf4ede01STejun Heo * CONTEXT: 955bf4ede01STejun Heo * spin_lock_irq(gcwq->lock). 956bf4ede01STejun Heo */ 957bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 958bf4ede01STejun Heo struct work_struct **nextp) 959bf4ede01STejun Heo { 960bf4ede01STejun Heo struct work_struct *n; 961bf4ede01STejun Heo 962bf4ede01STejun Heo /* 963bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 964bf4ede01STejun Heo * use NULL for list head. 965bf4ede01STejun Heo */ 966bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 967bf4ede01STejun Heo list_move_tail(&work->entry, head); 968bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 969bf4ede01STejun Heo break; 970bf4ede01STejun Heo } 971bf4ede01STejun Heo 972bf4ede01STejun Heo /* 973bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 974bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 975bf4ede01STejun Heo * needs to be updated. 976bf4ede01STejun Heo */ 977bf4ede01STejun Heo if (nextp) 978bf4ede01STejun Heo *nextp = n; 979bf4ede01STejun Heo } 980bf4ede01STejun Heo 981bf4ede01STejun Heo static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 982bf4ede01STejun Heo { 983bf4ede01STejun Heo struct work_struct *work = list_first_entry(&cwq->delayed_works, 984bf4ede01STejun Heo struct work_struct, entry); 985bf4ede01STejun Heo 986bf4ede01STejun Heo trace_workqueue_activate_work(work); 987bf4ede01STejun Heo move_linked_works(work, &cwq->pool->worklist, NULL); 988bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 989bf4ede01STejun Heo cwq->nr_active++; 990bf4ede01STejun Heo } 991bf4ede01STejun Heo 992bf4ede01STejun Heo /** 993bf4ede01STejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 994bf4ede01STejun Heo * @cwq: cwq of interest 995bf4ede01STejun Heo * @color: color of work which left the queue 996bf4ede01STejun Heo * @delayed: for a delayed work 997bf4ede01STejun Heo * 998bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 999bf4ede01STejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 1000bf4ede01STejun Heo * 1001bf4ede01STejun Heo * CONTEXT: 1002bf4ede01STejun Heo * spin_lock_irq(gcwq->lock). 1003bf4ede01STejun Heo */ 1004bf4ede01STejun Heo static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color, 1005bf4ede01STejun Heo bool delayed) 1006bf4ede01STejun Heo { 1007bf4ede01STejun Heo /* ignore uncolored works */ 1008bf4ede01STejun Heo if (color == WORK_NO_COLOR) 1009bf4ede01STejun Heo return; 1010bf4ede01STejun Heo 1011bf4ede01STejun Heo cwq->nr_in_flight[color]--; 1012bf4ede01STejun Heo 1013bf4ede01STejun Heo if (!delayed) { 1014bf4ede01STejun Heo cwq->nr_active--; 1015bf4ede01STejun Heo if (!list_empty(&cwq->delayed_works)) { 1016bf4ede01STejun Heo /* one down, submit a delayed one */ 1017bf4ede01STejun Heo if (cwq->nr_active < cwq->max_active) 1018bf4ede01STejun Heo cwq_activate_first_delayed(cwq); 1019bf4ede01STejun Heo } 1020bf4ede01STejun Heo } 1021bf4ede01STejun Heo 1022bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1023bf4ede01STejun Heo if (likely(cwq->flush_color != color)) 1024bf4ede01STejun Heo return; 1025bf4ede01STejun Heo 1026bf4ede01STejun Heo /* are there still in-flight works? */ 1027bf4ede01STejun Heo if (cwq->nr_in_flight[color]) 1028bf4ede01STejun Heo return; 1029bf4ede01STejun Heo 1030bf4ede01STejun Heo /* this cwq is done, clear flush_color */ 1031bf4ede01STejun Heo cwq->flush_color = -1; 1032bf4ede01STejun Heo 1033bf4ede01STejun Heo /* 1034bf4ede01STejun Heo * If this was the last cwq, wake up the first flusher. It 1035bf4ede01STejun Heo * will handle the rest. 1036bf4ede01STejun Heo */ 1037bf4ede01STejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 1038bf4ede01STejun Heo complete(&cwq->wq->first_flusher->done); 1039bf4ede01STejun Heo } 1040bf4ede01STejun Heo 104136e227d2STejun Heo /** 1042bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 104336e227d2STejun Heo * @work: work item to steal 104436e227d2STejun Heo * @is_dwork: @work is a delayed_work 1045bbb68dfaSTejun Heo * @flags: place to store irq state 104636e227d2STejun Heo * 104736e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 104836e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 104936e227d2STejun Heo * 105036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 105136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 105236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1053bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1054bbb68dfaSTejun Heo * for arbitrarily long 105536e227d2STejun Heo * 1056bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1057e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1058e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1059e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1060bbb68dfaSTejun Heo * 1061bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1062bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1063bbb68dfaSTejun Heo * 1064e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1065bf4ede01STejun Heo */ 1066bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1067bbb68dfaSTejun Heo unsigned long *flags) 1068bf4ede01STejun Heo { 1069bf4ede01STejun Heo struct global_cwq *gcwq; 1070bf4ede01STejun Heo 1071bbb68dfaSTejun Heo WARN_ON_ONCE(in_irq()); 1072bbb68dfaSTejun Heo 1073bbb68dfaSTejun Heo local_irq_save(*flags); 1074bbb68dfaSTejun Heo 107536e227d2STejun Heo /* try to steal the timer if it exists */ 107636e227d2STejun Heo if (is_dwork) { 107736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 107836e227d2STejun Heo 1079e0aecdd8STejun Heo /* 1080e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1081e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1082e0aecdd8STejun Heo * running on the local CPU. 1083e0aecdd8STejun Heo */ 108436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 108536e227d2STejun Heo return 1; 108636e227d2STejun Heo } 108736e227d2STejun Heo 108836e227d2STejun Heo /* try to claim PENDING the normal way */ 1089bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1090bf4ede01STejun Heo return 0; 1091bf4ede01STejun Heo 1092bf4ede01STejun Heo /* 1093bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1094bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1095bf4ede01STejun Heo */ 1096bf4ede01STejun Heo gcwq = get_work_gcwq(work); 1097bf4ede01STejun Heo if (!gcwq) 1098bbb68dfaSTejun Heo goto fail; 1099bf4ede01STejun Heo 1100bbb68dfaSTejun Heo spin_lock(&gcwq->lock); 1101bf4ede01STejun Heo if (!list_empty(&work->entry)) { 1102bf4ede01STejun Heo /* 1103bf4ede01STejun Heo * This work is queued, but perhaps we locked the wrong gcwq. 1104bf4ede01STejun Heo * In that case we must see the new value after rmb(), see 1105bf4ede01STejun Heo * insert_work()->wmb(). 1106bf4ede01STejun Heo */ 1107bf4ede01STejun Heo smp_rmb(); 1108bf4ede01STejun Heo if (gcwq == get_work_gcwq(work)) { 1109bf4ede01STejun Heo debug_work_deactivate(work); 1110bf4ede01STejun Heo list_del_init(&work->entry); 1111bf4ede01STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 1112bf4ede01STejun Heo get_work_color(work), 1113bf4ede01STejun Heo *work_data_bits(work) & WORK_STRUCT_DELAYED); 111436e227d2STejun Heo 1115bbb68dfaSTejun Heo spin_unlock(&gcwq->lock); 111636e227d2STejun Heo return 1; 1117bf4ede01STejun Heo } 1118bf4ede01STejun Heo } 1119bbb68dfaSTejun Heo spin_unlock(&gcwq->lock); 1120bbb68dfaSTejun Heo fail: 1121bbb68dfaSTejun Heo local_irq_restore(*flags); 1122bbb68dfaSTejun Heo if (work_is_canceling(work)) 1123bbb68dfaSTejun Heo return -ENOENT; 1124bbb68dfaSTejun Heo cpu_relax(); 112536e227d2STejun Heo return -EAGAIN; 1126bf4ede01STejun Heo } 1127bf4ede01STejun Heo 1128bf4ede01STejun Heo /** 11297e11629dSTejun Heo * insert_work - insert a work into gcwq 11304690c4abSTejun Heo * @cwq: cwq @work belongs to 11314690c4abSTejun Heo * @work: work to insert 11324690c4abSTejun Heo * @head: insertion point 11334690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11344690c4abSTejun Heo * 11357e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 11367e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 11374690c4abSTejun Heo * 11384690c4abSTejun Heo * CONTEXT: 11398b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 11401da177e4SLinus Torvalds */ 1141b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 11424690c4abSTejun Heo struct work_struct *work, struct list_head *head, 11434690c4abSTejun Heo unsigned int extra_flags) 1144b89deed3SOleg Nesterov { 114563d95a91STejun Heo struct worker_pool *pool = cwq->pool; 1146e1d8aa9fSFrederic Weisbecker 11474690c4abSTejun Heo /* we own @work, set data and link */ 11487a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 11494690c4abSTejun Heo 11506e84d644SOleg Nesterov /* 11516e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 11526e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 11536e84d644SOleg Nesterov */ 11546e84d644SOleg Nesterov smp_wmb(); 11554690c4abSTejun Heo 11561a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1157e22bee78STejun Heo 1158e22bee78STejun Heo /* 1159e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1160e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1161e22bee78STejun Heo * lying around lazily while there are works to be processed. 1162e22bee78STejun Heo */ 1163e22bee78STejun Heo smp_mb(); 1164e22bee78STejun Heo 116563d95a91STejun Heo if (__need_more_worker(pool)) 116663d95a91STejun Heo wake_up_worker(pool); 1167b89deed3SOleg Nesterov } 1168b89deed3SOleg Nesterov 1169c8efcc25STejun Heo /* 1170c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 1171c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 1172c8efcc25STejun Heo * cold paths. 1173c8efcc25STejun Heo */ 1174c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1175c8efcc25STejun Heo { 1176c8efcc25STejun Heo unsigned long flags; 1177c8efcc25STejun Heo unsigned int cpu; 1178c8efcc25STejun Heo 1179c8efcc25STejun Heo for_each_gcwq_cpu(cpu) { 1180c8efcc25STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 1181c8efcc25STejun Heo struct worker *worker; 1182c8efcc25STejun Heo struct hlist_node *pos; 1183c8efcc25STejun Heo int i; 1184c8efcc25STejun Heo 1185c8efcc25STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1186c8efcc25STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 1187c8efcc25STejun Heo if (worker->task != current) 1188c8efcc25STejun Heo continue; 1189c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 1190c8efcc25STejun Heo /* 1191c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 1192c8efcc25STejun Heo * is headed to the same workqueue. 1193c8efcc25STejun Heo */ 1194c8efcc25STejun Heo return worker->current_cwq->wq == wq; 1195c8efcc25STejun Heo } 1196c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 1197c8efcc25STejun Heo } 1198c8efcc25STejun Heo return false; 1199c8efcc25STejun Heo } 1200c8efcc25STejun Heo 12014690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 12021da177e4SLinus Torvalds struct work_struct *work) 12031da177e4SLinus Torvalds { 1204502ca9d8STejun Heo struct global_cwq *gcwq; 1205502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 12061e19ffc6STejun Heo struct list_head *worklist; 12078a2e8e5dSTejun Heo unsigned int work_flags; 1208b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12098930cabaSTejun Heo 12108930cabaSTejun Heo /* 12118930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12128930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12138930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12148930cabaSTejun Heo * happen with IRQ disabled. 12158930cabaSTejun Heo */ 12168930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12171da177e4SLinus Torvalds 1218dc186ad7SThomas Gleixner debug_work_activate(work); 12191e19ffc6STejun Heo 1220c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 12219c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1222c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1223e41e704bSTejun Heo return; 1224e41e704bSTejun Heo 1225c7fc77f7STejun Heo /* determine gcwq to use */ 1226c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1227c7fc77f7STejun Heo struct global_cwq *last_gcwq; 1228c7fc77f7STejun Heo 122957469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1230f3421797STejun Heo cpu = raw_smp_processor_id(); 1231f3421797STejun Heo 123218aa9effSTejun Heo /* 1233dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1234dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1235dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1236dbf2576eSTejun Heo * non-reentrancy. 123718aa9effSTejun Heo */ 1238502ca9d8STejun Heo gcwq = get_gcwq(cpu); 1239dbf2576eSTejun Heo last_gcwq = get_work_gcwq(work); 1240dbf2576eSTejun Heo 1241dbf2576eSTejun Heo if (last_gcwq && last_gcwq != gcwq) { 124218aa9effSTejun Heo struct worker *worker; 124318aa9effSTejun Heo 12448930cabaSTejun Heo spin_lock(&last_gcwq->lock); 124518aa9effSTejun Heo 124618aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 124718aa9effSTejun Heo 124818aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 124918aa9effSTejun Heo gcwq = last_gcwq; 125018aa9effSTejun Heo else { 125118aa9effSTejun Heo /* meh... not running there, queue here */ 12528930cabaSTejun Heo spin_unlock(&last_gcwq->lock); 12538930cabaSTejun Heo spin_lock(&gcwq->lock); 125418aa9effSTejun Heo } 12558930cabaSTejun Heo } else { 12568930cabaSTejun Heo spin_lock(&gcwq->lock); 12578930cabaSTejun Heo } 1258f3421797STejun Heo } else { 1259f3421797STejun Heo gcwq = get_gcwq(WORK_CPU_UNBOUND); 12608930cabaSTejun Heo spin_lock(&gcwq->lock); 1261502ca9d8STejun Heo } 1262502ca9d8STejun Heo 1263502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 1264502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 1265b75cac93SJoonsoo Kim trace_workqueue_queue_work(req_cpu, cwq, work); 1266502ca9d8STejun Heo 1267f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 12688930cabaSTejun Heo spin_unlock(&gcwq->lock); 1269f5b2552bSDan Carpenter return; 1270f5b2552bSDan Carpenter } 12711e19ffc6STejun Heo 127273f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 12738a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 12741e19ffc6STejun Heo 12751e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1276cdadf009STejun Heo trace_workqueue_activate_work(work); 12771e19ffc6STejun Heo cwq->nr_active++; 12783270476aSTejun Heo worklist = &cwq->pool->worklist; 12798a2e8e5dSTejun Heo } else { 12808a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 12811e19ffc6STejun Heo worklist = &cwq->delayed_works; 12828a2e8e5dSTejun Heo } 12831e19ffc6STejun Heo 12848a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 12851e19ffc6STejun Heo 12868930cabaSTejun Heo spin_unlock(&gcwq->lock); 12871da177e4SLinus Torvalds } 12881da177e4SLinus Torvalds 12890fcb78c2SRolf Eike Beer /** 1290c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1291c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1292c1a220e7SZhang Rui * @wq: workqueue to use 1293c1a220e7SZhang Rui * @work: work to queue 1294c1a220e7SZhang Rui * 1295d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1296c1a220e7SZhang Rui * 1297c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1298c1a220e7SZhang Rui * can't go away. 1299c1a220e7SZhang Rui */ 1300d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1301d4283e93STejun Heo struct work_struct *work) 1302c1a220e7SZhang Rui { 1303d4283e93STejun Heo bool ret = false; 13048930cabaSTejun Heo unsigned long flags; 13058930cabaSTejun Heo 13068930cabaSTejun Heo local_irq_save(flags); 1307c1a220e7SZhang Rui 130822df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13094690c4abSTejun Heo __queue_work(cpu, wq, work); 1310d4283e93STejun Heo ret = true; 1311c1a220e7SZhang Rui } 13128930cabaSTejun Heo 13138930cabaSTejun Heo local_irq_restore(flags); 1314c1a220e7SZhang Rui return ret; 1315c1a220e7SZhang Rui } 1316c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1317c1a220e7SZhang Rui 13180a13c00eSTejun Heo /** 13190a13c00eSTejun Heo * queue_work - queue work on a workqueue 13200a13c00eSTejun Heo * @wq: workqueue to use 13210a13c00eSTejun Heo * @work: work to queue 13220a13c00eSTejun Heo * 1323d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 13240a13c00eSTejun Heo * 13250a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 13260a13c00eSTejun Heo * it can be processed by another CPU. 13270a13c00eSTejun Heo */ 1328d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13290a13c00eSTejun Heo { 133057469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13310a13c00eSTejun Heo } 13320a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13330a13c00eSTejun Heo 1334d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13351da177e4SLinus Torvalds { 133652bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13377a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 13381da177e4SLinus Torvalds 1339e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 13401265057fSTejun Heo __queue_work(dwork->cpu, cwq->wq, &dwork->work); 13411da177e4SLinus Torvalds } 1342d8e794dfSTejun Heo EXPORT_SYMBOL_GPL(delayed_work_timer_fn); 13431da177e4SLinus Torvalds 13447beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 13457beb2edfSTejun Heo struct delayed_work *dwork, unsigned long delay) 13467beb2edfSTejun Heo { 13477beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13487beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13497beb2edfSTejun Heo unsigned int lcpu; 13507beb2edfSTejun Heo 13517beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13527beb2edfSTejun Heo timer->data != (unsigned long)dwork); 13537beb2edfSTejun Heo BUG_ON(timer_pending(timer)); 13547beb2edfSTejun Heo BUG_ON(!list_empty(&work->entry)); 13557beb2edfSTejun Heo 13567beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13577beb2edfSTejun Heo 13587beb2edfSTejun Heo /* 13597beb2edfSTejun Heo * This stores cwq for the moment, for the timer_fn. Note that the 13607beb2edfSTejun Heo * work's gcwq is preserved to allow reentrance detection for 13617beb2edfSTejun Heo * delayed works. 13627beb2edfSTejun Heo */ 13637beb2edfSTejun Heo if (!(wq->flags & WQ_UNBOUND)) { 13647beb2edfSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 13657beb2edfSTejun Heo 1366e42986deSJoonsoo Kim /* 1367e42986deSJoonsoo Kim * If we cannot get the last gcwq from @work directly, 1368e42986deSJoonsoo Kim * select the last CPU such that it avoids unnecessarily 1369e42986deSJoonsoo Kim * triggering non-reentrancy check in __queue_work(). 1370e42986deSJoonsoo Kim */ 1371e42986deSJoonsoo Kim lcpu = cpu; 1372e42986deSJoonsoo Kim if (gcwq) 13737beb2edfSTejun Heo lcpu = gcwq->cpu; 1374e42986deSJoonsoo Kim if (lcpu == WORK_CPU_UNBOUND) 13757beb2edfSTejun Heo lcpu = raw_smp_processor_id(); 13767beb2edfSTejun Heo } else { 13777beb2edfSTejun Heo lcpu = WORK_CPU_UNBOUND; 13787beb2edfSTejun Heo } 13797beb2edfSTejun Heo 13807beb2edfSTejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 13817beb2edfSTejun Heo 13821265057fSTejun Heo dwork->cpu = cpu; 13837beb2edfSTejun Heo timer->expires = jiffies + delay; 13847beb2edfSTejun Heo 13857beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13867beb2edfSTejun Heo add_timer_on(timer, cpu); 13877beb2edfSTejun Heo else 13887beb2edfSTejun Heo add_timer(timer); 13897beb2edfSTejun Heo } 13907beb2edfSTejun Heo 13910fcb78c2SRolf Eike Beer /** 13920fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 13930fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 13940fcb78c2SRolf Eike Beer * @wq: workqueue to use 1395af9997e4SRandy Dunlap * @dwork: work to queue 13960fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 13970fcb78c2SRolf Eike Beer * 1398715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1399715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1400715f1300STejun Heo * execution. 14010fcb78c2SRolf Eike Beer */ 1402d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 140352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14047a6bc1cdSVenkatesh Pallipadi { 140552bad64dSDavid Howells struct work_struct *work = &dwork->work; 1406d4283e93STejun Heo bool ret = false; 14078930cabaSTejun Heo unsigned long flags; 14088930cabaSTejun Heo 1409715f1300STejun Heo if (!delay) 1410715f1300STejun Heo return queue_work_on(cpu, wq, &dwork->work); 1411715f1300STejun Heo 14128930cabaSTejun Heo /* read the comment in __queue_work() */ 14138930cabaSTejun Heo local_irq_save(flags); 14147a6bc1cdSVenkatesh Pallipadi 141522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14167beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1417d4283e93STejun Heo ret = true; 14187a6bc1cdSVenkatesh Pallipadi } 14198930cabaSTejun Heo 14208930cabaSTejun Heo local_irq_restore(flags); 14217a6bc1cdSVenkatesh Pallipadi return ret; 14227a6bc1cdSVenkatesh Pallipadi } 1423ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14241da177e4SLinus Torvalds 1425c8e55f36STejun Heo /** 14260a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 14270a13c00eSTejun Heo * @wq: workqueue to use 14280a13c00eSTejun Heo * @dwork: delayable work to queue 14290a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 14300a13c00eSTejun Heo * 1431715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 14320a13c00eSTejun Heo */ 1433d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 14340a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 14350a13c00eSTejun Heo { 143657469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14370a13c00eSTejun Heo } 14380a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 14390a13c00eSTejun Heo 14400a13c00eSTejun Heo /** 14418376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14428376fe22STejun Heo * @cpu: CPU number to execute work on 14438376fe22STejun Heo * @wq: workqueue to use 14448376fe22STejun Heo * @dwork: work to queue 14458376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14468376fe22STejun Heo * 14478376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14488376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14498376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14508376fe22STejun Heo * current state. 14518376fe22STejun Heo * 14528376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14538376fe22STejun Heo * pending and its timer was modified. 14548376fe22STejun Heo * 1455e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14568376fe22STejun Heo * See try_to_grab_pending() for details. 14578376fe22STejun Heo */ 14588376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14598376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14608376fe22STejun Heo { 14618376fe22STejun Heo unsigned long flags; 14628376fe22STejun Heo int ret; 14638376fe22STejun Heo 14648376fe22STejun Heo do { 14658376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14668376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14678376fe22STejun Heo 14688376fe22STejun Heo if (likely(ret >= 0)) { 14698376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14708376fe22STejun Heo local_irq_restore(flags); 14718376fe22STejun Heo } 14728376fe22STejun Heo 14738376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14748376fe22STejun Heo return ret; 14758376fe22STejun Heo } 14768376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14778376fe22STejun Heo 14788376fe22STejun Heo /** 14798376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14808376fe22STejun Heo * @wq: workqueue to use 14818376fe22STejun Heo * @dwork: work to queue 14828376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14838376fe22STejun Heo * 14848376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14858376fe22STejun Heo */ 14868376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14878376fe22STejun Heo unsigned long delay) 14888376fe22STejun Heo { 14898376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14908376fe22STejun Heo } 14918376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 14928376fe22STejun Heo 14938376fe22STejun Heo /** 1494c8e55f36STejun Heo * worker_enter_idle - enter idle state 1495c8e55f36STejun Heo * @worker: worker which is entering idle state 1496c8e55f36STejun Heo * 1497c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1498c8e55f36STejun Heo * necessary. 1499c8e55f36STejun Heo * 1500c8e55f36STejun Heo * LOCKING: 1501c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1502c8e55f36STejun Heo */ 1503c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15041da177e4SLinus Torvalds { 1505bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1506bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1507c8e55f36STejun Heo 1508c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1509c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1510c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1511c8e55f36STejun Heo 1512cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1513cb444766STejun Heo worker->flags |= WORKER_IDLE; 1514bd7bdd43STejun Heo pool->nr_idle++; 1515e22bee78STejun Heo worker->last_active = jiffies; 1516c8e55f36STejun Heo 1517c8e55f36STejun Heo /* idle_list is LIFO */ 1518bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1519db7bccf4STejun Heo 152063d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1521628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1522cb444766STejun Heo 1523544ecf31STejun Heo /* 1524628c78e7STejun Heo * Sanity check nr_running. Because gcwq_unbind_fn() releases 1525628c78e7STejun Heo * gcwq->lock between setting %WORKER_UNBOUND and zapping 1526628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1527628c78e7STejun Heo * unbind is not in progress. 1528544ecf31STejun Heo */ 1529628c78e7STejun Heo WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) && 1530bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 153163d95a91STejun Heo atomic_read(get_pool_nr_running(pool))); 1532c8e55f36STejun Heo } 1533c8e55f36STejun Heo 1534c8e55f36STejun Heo /** 1535c8e55f36STejun Heo * worker_leave_idle - leave idle state 1536c8e55f36STejun Heo * @worker: worker which is leaving idle state 1537c8e55f36STejun Heo * 1538c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1539c8e55f36STejun Heo * 1540c8e55f36STejun Heo * LOCKING: 1541c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1542c8e55f36STejun Heo */ 1543c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1544c8e55f36STejun Heo { 1545bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1546c8e55f36STejun Heo 1547c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1548d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1549bd7bdd43STejun Heo pool->nr_idle--; 1550c8e55f36STejun Heo list_del_init(&worker->entry); 1551c8e55f36STejun Heo } 1552c8e55f36STejun Heo 1553e22bee78STejun Heo /** 1554e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1555e22bee78STejun Heo * @worker: self 1556e22bee78STejun Heo * 1557e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1558e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1559e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1560e22bee78STejun Heo * guaranteed to execute on the cpu. 1561e22bee78STejun Heo * 1562e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1563e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1564e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1565e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1566e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1567e22bee78STejun Heo * [dis]associated in the meantime. 1568e22bee78STejun Heo * 1569f2d5a0eeSTejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies the 1570f2d5a0eeSTejun Heo * binding against %GCWQ_DISASSOCIATED which is set during 1571f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1572f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1573f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1574e22bee78STejun Heo * 1575e22bee78STejun Heo * CONTEXT: 1576e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1577e22bee78STejun Heo * held. 1578e22bee78STejun Heo * 1579e22bee78STejun Heo * RETURNS: 1580e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1581e22bee78STejun Heo * bound), %false if offline. 1582e22bee78STejun Heo */ 1583e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1584972fa1c5SNamhyung Kim __acquires(&gcwq->lock) 1585e22bee78STejun Heo { 1586bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1587e22bee78STejun Heo struct task_struct *task = worker->task; 1588e22bee78STejun Heo 1589e22bee78STejun Heo while (true) { 1590e22bee78STejun Heo /* 1591e22bee78STejun Heo * The following call may fail, succeed or succeed 1592e22bee78STejun Heo * without actually migrating the task to the cpu if 1593e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1594e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1595e22bee78STejun Heo */ 1596f3421797STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) 1597e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1598e22bee78STejun Heo 1599e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1600e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1601e22bee78STejun Heo return false; 1602e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1603e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1604e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1605e22bee78STejun Heo return true; 1606e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1607e22bee78STejun Heo 16085035b20fSTejun Heo /* 16095035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16105035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16115035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16125035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16135035b20fSTejun Heo */ 1614e22bee78STejun Heo cpu_relax(); 16155035b20fSTejun Heo cond_resched(); 1616e22bee78STejun Heo } 1617e22bee78STejun Heo } 1618e22bee78STejun Heo 1619e22bee78STejun Heo /* 1620ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 162125511a47STejun Heo * %WORKER_REBIND before leaving idle and call this function. 162225511a47STejun Heo */ 162325511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 162425511a47STejun Heo { 162525511a47STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 162625511a47STejun Heo 1627ec58815aSTejun Heo /* 1628ea1abd61SLai Jiangshan * CPU may go down again inbetween. If rebinding fails, reinstate 1629ea1abd61SLai Jiangshan * UNBOUND. We're off idle_list and nobody else can do it for us. 1630ec58815aSTejun Heo */ 1631ea1abd61SLai Jiangshan if (!worker_maybe_bind_and_lock(worker)) 1632ea1abd61SLai Jiangshan worker->flags |= WORKER_UNBOUND; 1633ea1abd61SLai Jiangshan 1634ea1abd61SLai Jiangshan worker_clr_flags(worker, WORKER_REBIND); 1635ea1abd61SLai Jiangshan 1636ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1637ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1638ea1abd61SLai Jiangshan spin_unlock_irq(&gcwq->lock); 163925511a47STejun Heo } 164025511a47STejun Heo 164125511a47STejun Heo /* 164225511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1643403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1644403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1645403c821dSTejun Heo * executed twice without intervening cpu down. 1646e22bee78STejun Heo */ 164725511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1648e22bee78STejun Heo { 1649e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1650bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1651e22bee78STejun Heo 1652960bd11bSLai Jiangshan worker_maybe_bind_and_lock(worker); 1653960bd11bSLai Jiangshan 1654960bd11bSLai Jiangshan /* 1655960bd11bSLai Jiangshan * %WORKER_REBIND must be cleared even if the above binding failed; 1656960bd11bSLai Jiangshan * otherwise, we may confuse the next CPU_UP cycle or oops / get 1657960bd11bSLai Jiangshan * stuck by calling idle_worker_rebind() prematurely. If CPU went 1658960bd11bSLai Jiangshan * down again inbetween, %WORKER_UNBOUND would be set, so clearing 1659960bd11bSLai Jiangshan * %WORKER_REBIND is always safe. 1660960bd11bSLai Jiangshan */ 1661e22bee78STejun Heo worker_clr_flags(worker, WORKER_REBIND); 1662e22bee78STejun Heo 1663e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1664e22bee78STejun Heo } 1665e22bee78STejun Heo 166625511a47STejun Heo /** 166725511a47STejun Heo * rebind_workers - rebind all workers of a gcwq to the associated CPU 166825511a47STejun Heo * @gcwq: gcwq of interest 166925511a47STejun Heo * 167025511a47STejun Heo * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding 167125511a47STejun Heo * is different for idle and busy ones. 167225511a47STejun Heo * 1673ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1674ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1675ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1676ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 167725511a47STejun Heo * 1678ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1679ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1680ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1681ea1abd61SLai Jiangshan * rebind. 168225511a47STejun Heo * 1683ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1684ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1685ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1686ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 168725511a47STejun Heo */ 168825511a47STejun Heo static void rebind_workers(struct global_cwq *gcwq) 168925511a47STejun Heo { 169025511a47STejun Heo struct worker_pool *pool; 1691ea1abd61SLai Jiangshan struct worker *worker, *n; 169225511a47STejun Heo struct hlist_node *pos; 169325511a47STejun Heo int i; 169425511a47STejun Heo 169525511a47STejun Heo lockdep_assert_held(&gcwq->lock); 169625511a47STejun Heo 169725511a47STejun Heo for_each_worker_pool(pool, gcwq) 169825511a47STejun Heo lockdep_assert_held(&pool->manager_mutex); 169925511a47STejun Heo 1700ea1abd61SLai Jiangshan /* set REBIND and kick idle ones */ 170125511a47STejun Heo for_each_worker_pool(pool, gcwq) { 1702ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 170396e65306SLai Jiangshan unsigned long worker_flags = worker->flags; 170496e65306SLai Jiangshan 170596e65306SLai Jiangshan /* morph UNBOUND to REBIND atomically */ 170696e65306SLai Jiangshan worker_flags &= ~WORKER_UNBOUND; 170796e65306SLai Jiangshan worker_flags |= WORKER_REBIND; 170896e65306SLai Jiangshan ACCESS_ONCE(worker->flags) = worker_flags; 170925511a47STejun Heo 1710ea1abd61SLai Jiangshan /* 1711ea1abd61SLai Jiangshan * idle workers should be off @pool->idle_list 1712ea1abd61SLai Jiangshan * until rebind is complete to avoid receiving 1713ea1abd61SLai Jiangshan * premature local wake-ups. 1714ea1abd61SLai Jiangshan */ 1715ea1abd61SLai Jiangshan list_del_init(&worker->entry); 171625511a47STejun Heo 171725511a47STejun Heo /* worker_thread() will call idle_worker_rebind() */ 171825511a47STejun Heo wake_up_process(worker->task); 171925511a47STejun Heo } 172025511a47STejun Heo } 172125511a47STejun Heo 1722ea1abd61SLai Jiangshan /* rebind busy workers */ 172325511a47STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 172496e65306SLai Jiangshan unsigned long worker_flags = worker->flags; 172525511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1726e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 172725511a47STejun Heo 172896e65306SLai Jiangshan /* morph UNBOUND to REBIND atomically */ 172996e65306SLai Jiangshan worker_flags &= ~WORKER_UNBOUND; 173096e65306SLai Jiangshan worker_flags |= WORKER_REBIND; 173196e65306SLai Jiangshan ACCESS_ONCE(worker->flags) = worker_flags; 173225511a47STejun Heo 173325511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 173425511a47STejun Heo work_data_bits(rebind_work))) 173525511a47STejun Heo continue; 173625511a47STejun Heo 173725511a47STejun Heo debug_work_activate(rebind_work); 1738e2b6a6d5SJoonsoo Kim 1739e2b6a6d5SJoonsoo Kim /* 1740e2b6a6d5SJoonsoo Kim * wq doesn't really matter but let's keep @worker->pool 1741e2b6a6d5SJoonsoo Kim * and @cwq->pool consistent for sanity. 1742e2b6a6d5SJoonsoo Kim */ 1743e2b6a6d5SJoonsoo Kim if (worker_pool_pri(worker->pool)) 1744e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1745e2b6a6d5SJoonsoo Kim else 1746e2b6a6d5SJoonsoo Kim wq = system_wq; 1747e2b6a6d5SJoonsoo Kim 1748e2b6a6d5SJoonsoo Kim insert_work(get_cwq(gcwq->cpu, wq), rebind_work, 174925511a47STejun Heo worker->scheduled.next, 175025511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 175125511a47STejun Heo } 175225511a47STejun Heo } 175325511a47STejun Heo 1754c34056a3STejun Heo static struct worker *alloc_worker(void) 1755c34056a3STejun Heo { 1756c34056a3STejun Heo struct worker *worker; 1757c34056a3STejun Heo 1758c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1759c8e55f36STejun Heo if (worker) { 1760c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1761affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 176225511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1763e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1764e22bee78STejun Heo worker->flags = WORKER_PREP; 1765c8e55f36STejun Heo } 1766c34056a3STejun Heo return worker; 1767c34056a3STejun Heo } 1768c34056a3STejun Heo 1769c34056a3STejun Heo /** 1770c34056a3STejun Heo * create_worker - create a new workqueue worker 177163d95a91STejun Heo * @pool: pool the new worker will belong to 1772c34056a3STejun Heo * 177363d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1774c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1775c34056a3STejun Heo * destroy_worker(). 1776c34056a3STejun Heo * 1777c34056a3STejun Heo * CONTEXT: 1778c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1779c34056a3STejun Heo * 1780c34056a3STejun Heo * RETURNS: 1781c34056a3STejun Heo * Pointer to the newly created worker. 1782c34056a3STejun Heo */ 1783bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1784c34056a3STejun Heo { 178563d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 17863270476aSTejun Heo const char *pri = worker_pool_pri(pool) ? "H" : ""; 1787c34056a3STejun Heo struct worker *worker = NULL; 1788f3421797STejun Heo int id = -1; 1789c34056a3STejun Heo 17908b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1791bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 17928b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1793bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1794c34056a3STejun Heo goto fail; 17958b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1796c34056a3STejun Heo } 17978b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1798c34056a3STejun Heo 1799c34056a3STejun Heo worker = alloc_worker(); 1800c34056a3STejun Heo if (!worker) 1801c34056a3STejun Heo goto fail; 1802c34056a3STejun Heo 1803bd7bdd43STejun Heo worker->pool = pool; 1804c34056a3STejun Heo worker->id = id; 1805c34056a3STejun Heo 1806bc2ae0f5STejun Heo if (gcwq->cpu != WORK_CPU_UNBOUND) 180794dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 18083270476aSTejun Heo worker, cpu_to_node(gcwq->cpu), 18093270476aSTejun Heo "kworker/%u:%d%s", gcwq->cpu, id, pri); 1810f3421797STejun Heo else 1811f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 18123270476aSTejun Heo "kworker/u:%d%s", id, pri); 1813c34056a3STejun Heo if (IS_ERR(worker->task)) 1814c34056a3STejun Heo goto fail; 1815c34056a3STejun Heo 18163270476aSTejun Heo if (worker_pool_pri(pool)) 18173270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 18183270476aSTejun Heo 1819db7bccf4STejun Heo /* 1820bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 1821bc2ae0f5STejun Heo * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the 1822bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1823bc2ae0f5STejun Heo * above the flag definition for details. 1824bc2ae0f5STejun Heo * 1825bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1826bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1827db7bccf4STejun Heo */ 1828bc2ae0f5STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) { 18298b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1830bc2ae0f5STejun Heo } else { 1831db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1832f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1833f3421797STejun Heo } 1834c34056a3STejun Heo 1835c34056a3STejun Heo return worker; 1836c34056a3STejun Heo fail: 1837c34056a3STejun Heo if (id >= 0) { 18388b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1839bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 18408b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1841c34056a3STejun Heo } 1842c34056a3STejun Heo kfree(worker); 1843c34056a3STejun Heo return NULL; 1844c34056a3STejun Heo } 1845c34056a3STejun Heo 1846c34056a3STejun Heo /** 1847c34056a3STejun Heo * start_worker - start a newly created worker 1848c34056a3STejun Heo * @worker: worker to start 1849c34056a3STejun Heo * 1850c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1851c34056a3STejun Heo * 1852c34056a3STejun Heo * CONTEXT: 18538b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1854c34056a3STejun Heo */ 1855c34056a3STejun Heo static void start_worker(struct worker *worker) 1856c34056a3STejun Heo { 1857cb444766STejun Heo worker->flags |= WORKER_STARTED; 1858bd7bdd43STejun Heo worker->pool->nr_workers++; 1859c8e55f36STejun Heo worker_enter_idle(worker); 1860c34056a3STejun Heo wake_up_process(worker->task); 1861c34056a3STejun Heo } 1862c34056a3STejun Heo 1863c34056a3STejun Heo /** 1864c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1865c34056a3STejun Heo * @worker: worker to be destroyed 1866c34056a3STejun Heo * 1867c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1868c8e55f36STejun Heo * 1869c8e55f36STejun Heo * CONTEXT: 1870c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1871c34056a3STejun Heo */ 1872c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1873c34056a3STejun Heo { 1874bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1875bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1876c34056a3STejun Heo int id = worker->id; 1877c34056a3STejun Heo 1878c34056a3STejun Heo /* sanity check frenzy */ 1879c34056a3STejun Heo BUG_ON(worker->current_work); 1880affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1881c34056a3STejun Heo 1882c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1883bd7bdd43STejun Heo pool->nr_workers--; 1884c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1885bd7bdd43STejun Heo pool->nr_idle--; 1886c8e55f36STejun Heo 1887c8e55f36STejun Heo list_del_init(&worker->entry); 1888cb444766STejun Heo worker->flags |= WORKER_DIE; 1889c8e55f36STejun Heo 1890c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1891c8e55f36STejun Heo 1892c34056a3STejun Heo kthread_stop(worker->task); 1893c34056a3STejun Heo kfree(worker); 1894c34056a3STejun Heo 18958b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1896bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1897c34056a3STejun Heo } 1898c34056a3STejun Heo 189963d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1900e22bee78STejun Heo { 190163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 190263d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1903e22bee78STejun Heo 1904e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1905e22bee78STejun Heo 190663d95a91STejun Heo if (too_many_workers(pool)) { 1907e22bee78STejun Heo struct worker *worker; 1908e22bee78STejun Heo unsigned long expires; 1909e22bee78STejun Heo 1910e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 191163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1912e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1913e22bee78STejun Heo 1914e22bee78STejun Heo if (time_before(jiffies, expires)) 191563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1916e22bee78STejun Heo else { 1917e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 191811ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 191963d95a91STejun Heo wake_up_worker(pool); 1920e22bee78STejun Heo } 1921e22bee78STejun Heo } 1922e22bee78STejun Heo 1923e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1924e22bee78STejun Heo } 1925e22bee78STejun Heo 1926e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1927e22bee78STejun Heo { 1928e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1929e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1930f3421797STejun Heo unsigned int cpu; 1931e22bee78STejun Heo 1932e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1933e22bee78STejun Heo return false; 1934e22bee78STejun Heo 1935e22bee78STejun Heo /* mayday mayday mayday */ 1936bd7bdd43STejun Heo cpu = cwq->pool->gcwq->cpu; 1937f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1938f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1939f3421797STejun Heo cpu = 0; 1940f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1941e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1942e22bee78STejun Heo return true; 1943e22bee78STejun Heo } 1944e22bee78STejun Heo 194563d95a91STejun Heo static void gcwq_mayday_timeout(unsigned long __pool) 1946e22bee78STejun Heo { 194763d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 194863d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1949e22bee78STejun Heo struct work_struct *work; 1950e22bee78STejun Heo 1951e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1952e22bee78STejun Heo 195363d95a91STejun Heo if (need_to_create_worker(pool)) { 1954e22bee78STejun Heo /* 1955e22bee78STejun Heo * We've been trying to create a new worker but 1956e22bee78STejun Heo * haven't been successful. We might be hitting an 1957e22bee78STejun Heo * allocation deadlock. Send distress signals to 1958e22bee78STejun Heo * rescuers. 1959e22bee78STejun Heo */ 196063d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1961e22bee78STejun Heo send_mayday(work); 1962e22bee78STejun Heo } 1963e22bee78STejun Heo 1964e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1965e22bee78STejun Heo 196663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1967e22bee78STejun Heo } 1968e22bee78STejun Heo 1969e22bee78STejun Heo /** 1970e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 197163d95a91STejun Heo * @pool: pool to create a new worker for 1972e22bee78STejun Heo * 197363d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1974e22bee78STejun Heo * have at least one idle worker on return from this function. If 1975e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 197663d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1977e22bee78STejun Heo * possible allocation deadlock. 1978e22bee78STejun Heo * 1979e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1980e22bee78STejun Heo * may_start_working() true. 1981e22bee78STejun Heo * 1982e22bee78STejun Heo * LOCKING: 1983e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1984e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1985e22bee78STejun Heo * manager. 1986e22bee78STejun Heo * 1987e22bee78STejun Heo * RETURNS: 1988e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1989e22bee78STejun Heo * otherwise. 1990e22bee78STejun Heo */ 199163d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 199206bd6ebfSNamhyung Kim __releases(&gcwq->lock) 199306bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 1994e22bee78STejun Heo { 199563d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 199663d95a91STejun Heo 199763d95a91STejun Heo if (!need_to_create_worker(pool)) 1998e22bee78STejun Heo return false; 1999e22bee78STejun Heo restart: 20009f9c2364STejun Heo spin_unlock_irq(&gcwq->lock); 20019f9c2364STejun Heo 2002e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 200363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2004e22bee78STejun Heo 2005e22bee78STejun Heo while (true) { 2006e22bee78STejun Heo struct worker *worker; 2007e22bee78STejun Heo 2008bc2ae0f5STejun Heo worker = create_worker(pool); 2009e22bee78STejun Heo if (worker) { 201063d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2011e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 2012e22bee78STejun Heo start_worker(worker); 201363d95a91STejun Heo BUG_ON(need_to_create_worker(pool)); 2014e22bee78STejun Heo return true; 2015e22bee78STejun Heo } 2016e22bee78STejun Heo 201763d95a91STejun Heo if (!need_to_create_worker(pool)) 2018e22bee78STejun Heo break; 2019e22bee78STejun Heo 2020e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2021e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 20229f9c2364STejun Heo 202363d95a91STejun Heo if (!need_to_create_worker(pool)) 2024e22bee78STejun Heo break; 2025e22bee78STejun Heo } 2026e22bee78STejun Heo 202763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2028e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 202963d95a91STejun Heo if (need_to_create_worker(pool)) 2030e22bee78STejun Heo goto restart; 2031e22bee78STejun Heo return true; 2032e22bee78STejun Heo } 2033e22bee78STejun Heo 2034e22bee78STejun Heo /** 2035e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 203663d95a91STejun Heo * @pool: pool to destroy workers for 2037e22bee78STejun Heo * 203863d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 2039e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 2040e22bee78STejun Heo * 2041e22bee78STejun Heo * LOCKING: 2042e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2043e22bee78STejun Heo * multiple times. Called only from manager. 2044e22bee78STejun Heo * 2045e22bee78STejun Heo * RETURNS: 2046e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 2047e22bee78STejun Heo * otherwise. 2048e22bee78STejun Heo */ 204963d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2050e22bee78STejun Heo { 2051e22bee78STejun Heo bool ret = false; 2052e22bee78STejun Heo 205363d95a91STejun Heo while (too_many_workers(pool)) { 2054e22bee78STejun Heo struct worker *worker; 2055e22bee78STejun Heo unsigned long expires; 2056e22bee78STejun Heo 205763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2058e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2059e22bee78STejun Heo 2060e22bee78STejun Heo if (time_before(jiffies, expires)) { 206163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2062e22bee78STejun Heo break; 2063e22bee78STejun Heo } 2064e22bee78STejun Heo 2065e22bee78STejun Heo destroy_worker(worker); 2066e22bee78STejun Heo ret = true; 2067e22bee78STejun Heo } 2068e22bee78STejun Heo 2069e22bee78STejun Heo return ret; 2070e22bee78STejun Heo } 2071e22bee78STejun Heo 2072e22bee78STejun Heo /** 2073e22bee78STejun Heo * manage_workers - manage worker pool 2074e22bee78STejun Heo * @worker: self 2075e22bee78STejun Heo * 2076e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 2077e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2078e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 2079e22bee78STejun Heo * 2080e22bee78STejun Heo * The caller can safely start processing works on false return. On 2081e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2082e22bee78STejun Heo * and may_start_working() is true. 2083e22bee78STejun Heo * 2084e22bee78STejun Heo * CONTEXT: 2085e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2086e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2087e22bee78STejun Heo * 2088e22bee78STejun Heo * RETURNS: 2089e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 2090e22bee78STejun Heo * some action was taken. 2091e22bee78STejun Heo */ 2092e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2093e22bee78STejun Heo { 209463d95a91STejun Heo struct worker_pool *pool = worker->pool; 2095e22bee78STejun Heo bool ret = false; 2096e22bee78STejun Heo 2097ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2098e22bee78STejun Heo return ret; 2099e22bee78STejun Heo 2100552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2101ee378aa4SLai Jiangshan 2102ee378aa4SLai Jiangshan /* 2103ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2104ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2105ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2106ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2107ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2108ee378aa4SLai Jiangshan * extreme circumstances. Use @pool->manager_mutex to synchronize 2109ee378aa4SLai Jiangshan * manager against CPU hotplug. 2110ee378aa4SLai Jiangshan * 2111ee378aa4SLai Jiangshan * manager_mutex would always be free unless CPU hotplug is in 2112ee378aa4SLai Jiangshan * progress. trylock first without dropping @gcwq->lock. 2113ee378aa4SLai Jiangshan */ 2114ee378aa4SLai Jiangshan if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2115ee378aa4SLai Jiangshan spin_unlock_irq(&pool->gcwq->lock); 2116ee378aa4SLai Jiangshan mutex_lock(&pool->manager_mutex); 2117ee378aa4SLai Jiangshan /* 2118ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2119ee378aa4SLai Jiangshan * for manager_mutex. Hotplug itself can't handle us 2120ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2121ee378aa4SLai Jiangshan * @gcwq's state and ours could have deviated. 2122ee378aa4SLai Jiangshan * 2123ee378aa4SLai Jiangshan * As hotplug is now excluded via manager_mutex, we can 2124ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2125ee378aa4SLai Jiangshan * on @gcwq's current state. Try it and adjust 2126ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2127ee378aa4SLai Jiangshan */ 2128ee378aa4SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 2129ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2130ee378aa4SLai Jiangshan else 2131ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2132ee378aa4SLai Jiangshan 2133ee378aa4SLai Jiangshan ret = true; 2134ee378aa4SLai Jiangshan } 2135ee378aa4SLai Jiangshan 213611ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2137e22bee78STejun Heo 2138e22bee78STejun Heo /* 2139e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2140e22bee78STejun Heo * on return. 2141e22bee78STejun Heo */ 214263d95a91STejun Heo ret |= maybe_destroy_workers(pool); 214363d95a91STejun Heo ret |= maybe_create_worker(pool); 2144e22bee78STejun Heo 2145552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 214660373152STejun Heo mutex_unlock(&pool->manager_mutex); 2147e22bee78STejun Heo return ret; 2148e22bee78STejun Heo } 2149e22bee78STejun Heo 2150a62428c0STejun Heo /** 2151a62428c0STejun Heo * process_one_work - process single work 2152c34056a3STejun Heo * @worker: self 2153a62428c0STejun Heo * @work: work to process 2154a62428c0STejun Heo * 2155a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2156a62428c0STejun Heo * process a single work including synchronization against and 2157a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2158a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2159a62428c0STejun Heo * call this function to process a work. 2160a62428c0STejun Heo * 2161a62428c0STejun Heo * CONTEXT: 21628b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 2163a62428c0STejun Heo */ 2164c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 216506bd6ebfSNamhyung Kim __releases(&gcwq->lock) 216606bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 21671da177e4SLinus Torvalds { 21687e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 2169bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2170bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2171c8e55f36STejun Heo struct hlist_head *bwh = busy_worker_head(gcwq, work); 2172fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 21736bb49e59SDavid Howells work_func_t f = work->func; 217473f53c4aSTejun Heo int work_color; 21757e11629dSTejun Heo struct worker *collision; 21764e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21774e6045f1SJohannes Berg /* 2178a62428c0STejun Heo * It is permissible to free the struct work_struct from 2179a62428c0STejun Heo * inside the function that is called from it, this we need to 2180a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2181a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2182a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21834e6045f1SJohannes Berg */ 21844d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21854d82a1deSPeter Zijlstra 21864d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21874e6045f1SJohannes Berg #endif 21886fec10a1STejun Heo /* 21896fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21906fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 21916fec10a1STejun Heo * unbound or a disassociated gcwq. 21926fec10a1STejun Heo */ 219325511a47STejun Heo WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) && 21946fec10a1STejun Heo !(gcwq->flags & GCWQ_DISASSOCIATED) && 219525511a47STejun Heo raw_smp_processor_id() != gcwq->cpu); 219625511a47STejun Heo 21977e11629dSTejun Heo /* 21987e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21997e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 22007e11629dSTejun Heo * already processing the work. If so, defer the work to the 22017e11629dSTejun Heo * currently executing one. 22027e11629dSTejun Heo */ 22037e11629dSTejun Heo collision = __find_worker_executing_work(gcwq, bwh, work); 22047e11629dSTejun Heo if (unlikely(collision)) { 22057e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 22067e11629dSTejun Heo return; 22077e11629dSTejun Heo } 22081da177e4SLinus Torvalds 22098930cabaSTejun Heo /* claim and dequeue */ 22101da177e4SLinus Torvalds debug_work_deactivate(work); 2211c8e55f36STejun Heo hlist_add_head(&worker->hentry, bwh); 2212c34056a3STejun Heo worker->current_work = work; 22138cca0eeaSTejun Heo worker->current_cwq = cwq; 221473f53c4aSTejun Heo work_color = get_work_color(work); 22157a22ad75STejun Heo 2216a62428c0STejun Heo list_del_init(&work->entry); 2217a62428c0STejun Heo 2218649027d7STejun Heo /* 2219fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2220fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2221fb0e7bebSTejun Heo */ 2222fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2223fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2224fb0e7bebSTejun Heo 2225974271c4STejun Heo /* 2226974271c4STejun Heo * Unbound gcwq isn't concurrency managed and work items should be 2227974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2228974271c4STejun Heo */ 222963d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 223063d95a91STejun Heo wake_up_worker(pool); 2231974271c4STejun Heo 22328930cabaSTejun Heo /* 223323657bb1STejun Heo * Record the last CPU and clear PENDING which should be the last 223423657bb1STejun Heo * update to @work. Also, do this inside @gcwq->lock so that 223523657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 223623657bb1STejun Heo * disabled. 22378930cabaSTejun Heo */ 22388930cabaSTejun Heo set_work_cpu_and_clear_pending(work, gcwq->cpu); 22391da177e4SLinus Torvalds 22408930cabaSTejun Heo spin_unlock_irq(&gcwq->lock); 2241959d1af8STejun Heo 2242e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 22433295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2244e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 224565f27f38SDavid Howells f(work); 2246e36c886aSArjan van de Ven /* 2247e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2248e36c886aSArjan van de Ven * point will only record its address. 2249e36c886aSArjan van de Ven */ 2250e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22513295f0efSIngo Molnar lock_map_release(&lockdep_map); 22523295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 22531da177e4SLinus Torvalds 2254d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2255044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2256044c782cSValentin Ilie " last function: %pf\n", 2257044c782cSValentin Ilie current->comm, preempt_count(), task_pid_nr(current), f); 2258d5abe669SPeter Zijlstra debug_show_held_locks(current); 2259d5abe669SPeter Zijlstra dump_stack(); 2260d5abe669SPeter Zijlstra } 2261d5abe669SPeter Zijlstra 22628b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2263a62428c0STejun Heo 2264fb0e7bebSTejun Heo /* clear cpu intensive status */ 2265fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2266fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2267fb0e7bebSTejun Heo 2268a62428c0STejun Heo /* we're done with it, release */ 2269c8e55f36STejun Heo hlist_del_init(&worker->hentry); 2270c34056a3STejun Heo worker->current_work = NULL; 22718cca0eeaSTejun Heo worker->current_cwq = NULL; 22728a2e8e5dSTejun Heo cwq_dec_nr_in_flight(cwq, work_color, false); 22731da177e4SLinus Torvalds } 22741da177e4SLinus Torvalds 2275affee4b2STejun Heo /** 2276affee4b2STejun Heo * process_scheduled_works - process scheduled works 2277affee4b2STejun Heo * @worker: self 2278affee4b2STejun Heo * 2279affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2280affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2281affee4b2STejun Heo * fetches a work from the top and executes it. 2282affee4b2STejun Heo * 2283affee4b2STejun Heo * CONTEXT: 22848b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2285affee4b2STejun Heo * multiple times. 2286affee4b2STejun Heo */ 2287affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22881da177e4SLinus Torvalds { 2289affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2290affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2291a62428c0STejun Heo struct work_struct, entry); 2292c34056a3STejun Heo process_one_work(worker, work); 2293a62428c0STejun Heo } 22941da177e4SLinus Torvalds } 22951da177e4SLinus Torvalds 22964690c4abSTejun Heo /** 22974690c4abSTejun Heo * worker_thread - the worker thread function 2298c34056a3STejun Heo * @__worker: self 22994690c4abSTejun Heo * 2300e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 2301e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 2302e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2303e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2304e22bee78STejun Heo * rescuer_thread(). 23054690c4abSTejun Heo */ 2306c34056a3STejun Heo static int worker_thread(void *__worker) 23071da177e4SLinus Torvalds { 2308c34056a3STejun Heo struct worker *worker = __worker; 2309bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2310bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 23111da177e4SLinus Torvalds 2312e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2313e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2314c8e55f36STejun Heo woke_up: 23158b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2316affee4b2STejun Heo 231725511a47STejun Heo /* 231825511a47STejun Heo * DIE can be set only while idle and REBIND set while busy has 231925511a47STejun Heo * @worker->rebind_work scheduled. Checking here is enough. 232025511a47STejun Heo */ 232125511a47STejun Heo if (unlikely(worker->flags & (WORKER_REBIND | WORKER_DIE))) { 2322c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 232325511a47STejun Heo 232425511a47STejun Heo if (worker->flags & WORKER_DIE) { 2325e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2326c8e55f36STejun Heo return 0; 2327c8e55f36STejun Heo } 2328c8e55f36STejun Heo 232925511a47STejun Heo idle_worker_rebind(worker); 233025511a47STejun Heo goto woke_up; 233125511a47STejun Heo } 233225511a47STejun Heo 2333c8e55f36STejun Heo worker_leave_idle(worker); 2334db7bccf4STejun Heo recheck: 2335e22bee78STejun Heo /* no more worker necessary? */ 233663d95a91STejun Heo if (!need_more_worker(pool)) 2337e22bee78STejun Heo goto sleep; 2338e22bee78STejun Heo 2339e22bee78STejun Heo /* do we need to manage? */ 234063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2341e22bee78STejun Heo goto recheck; 2342e22bee78STejun Heo 2343c8e55f36STejun Heo /* 2344c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2345c8e55f36STejun Heo * preparing to process a work or actually processing it. 2346c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2347c8e55f36STejun Heo */ 2348c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 2349c8e55f36STejun Heo 2350e22bee78STejun Heo /* 2351e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2352e22bee78STejun Heo * at least one idle worker or that someone else has already 2353e22bee78STejun Heo * assumed the manager role. 2354e22bee78STejun Heo */ 2355e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2356e22bee78STejun Heo 2357e22bee78STejun Heo do { 2358affee4b2STejun Heo struct work_struct *work = 2359bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2360affee4b2STejun Heo struct work_struct, entry); 2361affee4b2STejun Heo 2362c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2363affee4b2STejun Heo /* optimization path, not strictly necessary */ 2364affee4b2STejun Heo process_one_work(worker, work); 2365affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2366affee4b2STejun Heo process_scheduled_works(worker); 2367affee4b2STejun Heo } else { 2368c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2369affee4b2STejun Heo process_scheduled_works(worker); 2370affee4b2STejun Heo } 237163d95a91STejun Heo } while (keep_working(pool)); 2372affee4b2STejun Heo 2373e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2374d313dd85STejun Heo sleep: 237563d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2376e22bee78STejun Heo goto recheck; 2377d313dd85STejun Heo 2378c8e55f36STejun Heo /* 2379e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 2380e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 2381e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 2382e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 2383e22bee78STejun Heo * prevent losing any event. 2384c8e55f36STejun Heo */ 2385c8e55f36STejun Heo worker_enter_idle(worker); 2386c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 23878b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 23881da177e4SLinus Torvalds schedule(); 2389c8e55f36STejun Heo goto woke_up; 23901da177e4SLinus Torvalds } 23911da177e4SLinus Torvalds 2392e22bee78STejun Heo /** 2393e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2394e22bee78STejun Heo * @__wq: the associated workqueue 2395e22bee78STejun Heo * 2396e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2397e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2398e22bee78STejun Heo * 2399e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 2400e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2401e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2402e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2403e22bee78STejun Heo * the problem rescuer solves. 2404e22bee78STejun Heo * 2405e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 2406e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 2407e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2408e22bee78STejun Heo * 2409e22bee78STejun Heo * This should happen rarely. 2410e22bee78STejun Heo */ 2411e22bee78STejun Heo static int rescuer_thread(void *__wq) 2412e22bee78STejun Heo { 2413e22bee78STejun Heo struct workqueue_struct *wq = __wq; 2414e22bee78STejun Heo struct worker *rescuer = wq->rescuer; 2415e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2416f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2417e22bee78STejun Heo unsigned int cpu; 2418e22bee78STejun Heo 2419e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2420e22bee78STejun Heo repeat: 2421e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 24221da177e4SLinus Torvalds 24231da177e4SLinus Torvalds if (kthread_should_stop()) 2424e22bee78STejun Heo return 0; 24251da177e4SLinus Torvalds 2426f3421797STejun Heo /* 2427f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2428f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2429f3421797STejun Heo */ 2430f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2431f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2432f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2433bd7bdd43STejun Heo struct worker_pool *pool = cwq->pool; 2434bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2435e22bee78STejun Heo struct work_struct *work, *n; 2436e22bee78STejun Heo 2437e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2438f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2439e22bee78STejun Heo 2440e22bee78STejun Heo /* migrate to the target cpu if possible */ 2441bd7bdd43STejun Heo rescuer->pool = pool; 2442e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2443e22bee78STejun Heo 2444e22bee78STejun Heo /* 2445e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2446e22bee78STejun Heo * process'em. 2447e22bee78STejun Heo */ 2448e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2449bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2450e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2451e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2452e22bee78STejun Heo 2453e22bee78STejun Heo process_scheduled_works(rescuer); 24547576958aSTejun Heo 24557576958aSTejun Heo /* 24567576958aSTejun Heo * Leave this gcwq. If keep_working() is %true, notify a 24577576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24587576958aSTejun Heo * and stalling the execution. 24597576958aSTejun Heo */ 246063d95a91STejun Heo if (keep_working(pool)) 246163d95a91STejun Heo wake_up_worker(pool); 24627576958aSTejun Heo 2463e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 24641da177e4SLinus Torvalds } 24651da177e4SLinus Torvalds 2466e22bee78STejun Heo schedule(); 2467e22bee78STejun Heo goto repeat; 24681da177e4SLinus Torvalds } 24691da177e4SLinus Torvalds 2470fc2e4d70SOleg Nesterov struct wq_barrier { 2471fc2e4d70SOleg Nesterov struct work_struct work; 2472fc2e4d70SOleg Nesterov struct completion done; 2473fc2e4d70SOleg Nesterov }; 2474fc2e4d70SOleg Nesterov 2475fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2476fc2e4d70SOleg Nesterov { 2477fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2478fc2e4d70SOleg Nesterov complete(&barr->done); 2479fc2e4d70SOleg Nesterov } 2480fc2e4d70SOleg Nesterov 24814690c4abSTejun Heo /** 24824690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 24834690c4abSTejun Heo * @cwq: cwq to insert barrier into 24844690c4abSTejun Heo * @barr: wq_barrier to insert 2485affee4b2STejun Heo * @target: target work to attach @barr to 2486affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24874690c4abSTejun Heo * 2488affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2489affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2490affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2491affee4b2STejun Heo * cpu. 2492affee4b2STejun Heo * 2493affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2494affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2495affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2496affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2497affee4b2STejun Heo * after a work with LINKED flag set. 2498affee4b2STejun Heo * 2499affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2500affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 25014690c4abSTejun Heo * 25024690c4abSTejun Heo * CONTEXT: 25038b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 25044690c4abSTejun Heo */ 250583c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2506affee4b2STejun Heo struct wq_barrier *barr, 2507affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2508fc2e4d70SOleg Nesterov { 2509affee4b2STejun Heo struct list_head *head; 2510affee4b2STejun Heo unsigned int linked = 0; 2511affee4b2STejun Heo 2512dc186ad7SThomas Gleixner /* 25138b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 2514dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2515dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2516dc186ad7SThomas Gleixner * might deadlock. 2517dc186ad7SThomas Gleixner */ 2518ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 251922df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2520fc2e4d70SOleg Nesterov init_completion(&barr->done); 252183c22520SOleg Nesterov 2522affee4b2STejun Heo /* 2523affee4b2STejun Heo * If @target is currently being executed, schedule the 2524affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2525affee4b2STejun Heo */ 2526affee4b2STejun Heo if (worker) 2527affee4b2STejun Heo head = worker->scheduled.next; 2528affee4b2STejun Heo else { 2529affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2530affee4b2STejun Heo 2531affee4b2STejun Heo head = target->entry.next; 2532affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2533affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2534affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2535affee4b2STejun Heo } 2536affee4b2STejun Heo 2537dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2538affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2539affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2540fc2e4d70SOleg Nesterov } 2541fc2e4d70SOleg Nesterov 254273f53c4aSTejun Heo /** 254373f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 254473f53c4aSTejun Heo * @wq: workqueue being flushed 254573f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 254673f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 254773f53c4aSTejun Heo * 254873f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 254973f53c4aSTejun Heo * 255073f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 255173f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 255273f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 255373f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 255473f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 255573f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 255673f53c4aSTejun Heo * 255773f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 255873f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 255973f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 256073f53c4aSTejun Heo * is returned. 256173f53c4aSTejun Heo * 256273f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 256373f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 256473f53c4aSTejun Heo * advanced to @work_color. 256573f53c4aSTejun Heo * 256673f53c4aSTejun Heo * CONTEXT: 256773f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 256873f53c4aSTejun Heo * 256973f53c4aSTejun Heo * RETURNS: 257073f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 257173f53c4aSTejun Heo * otherwise. 257273f53c4aSTejun Heo */ 257373f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 257473f53c4aSTejun Heo int flush_color, int work_color) 25751da177e4SLinus Torvalds { 257673f53c4aSTejun Heo bool wait = false; 257773f53c4aSTejun Heo unsigned int cpu; 25781da177e4SLinus Torvalds 257973f53c4aSTejun Heo if (flush_color >= 0) { 258073f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 258173f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2582dc186ad7SThomas Gleixner } 258314441960SOleg Nesterov 2584f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 258573f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2586bd7bdd43STejun Heo struct global_cwq *gcwq = cwq->pool->gcwq; 25871da177e4SLinus Torvalds 25888b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 258973f53c4aSTejun Heo 259073f53c4aSTejun Heo if (flush_color >= 0) { 259173f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 259273f53c4aSTejun Heo 259373f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 259473f53c4aSTejun Heo cwq->flush_color = flush_color; 259573f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 259673f53c4aSTejun Heo wait = true; 25971da177e4SLinus Torvalds } 259873f53c4aSTejun Heo } 259973f53c4aSTejun Heo 260073f53c4aSTejun Heo if (work_color >= 0) { 260173f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 260273f53c4aSTejun Heo cwq->work_color = work_color; 260373f53c4aSTejun Heo } 260473f53c4aSTejun Heo 26058b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 26061da177e4SLinus Torvalds } 26071da177e4SLinus Torvalds 260873f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 260973f53c4aSTejun Heo complete(&wq->first_flusher->done); 261073f53c4aSTejun Heo 261173f53c4aSTejun Heo return wait; 261283c22520SOleg Nesterov } 26131da177e4SLinus Torvalds 26140fcb78c2SRolf Eike Beer /** 26151da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26160fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26171da177e4SLinus Torvalds * 26181da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 26191da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 26201da177e4SLinus Torvalds * 2621fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2622fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 26231da177e4SLinus Torvalds */ 26247ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26251da177e4SLinus Torvalds { 262673f53c4aSTejun Heo struct wq_flusher this_flusher = { 262773f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 262873f53c4aSTejun Heo .flush_color = -1, 262973f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 263073f53c4aSTejun Heo }; 263173f53c4aSTejun Heo int next_color; 2632b1f4ec17SOleg Nesterov 26333295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26343295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 263573f53c4aSTejun Heo 263673f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo /* 263973f53c4aSTejun Heo * Start-to-wait phase 264073f53c4aSTejun Heo */ 264173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 264273f53c4aSTejun Heo 264373f53c4aSTejun Heo if (next_color != wq->flush_color) { 264473f53c4aSTejun Heo /* 264573f53c4aSTejun Heo * Color space is not full. The current work_color 264673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 264773f53c4aSTejun Heo * by one. 264873f53c4aSTejun Heo */ 264973f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 265073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 265173f53c4aSTejun Heo wq->work_color = next_color; 265273f53c4aSTejun Heo 265373f53c4aSTejun Heo if (!wq->first_flusher) { 265473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 265573f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 265673f53c4aSTejun Heo 265773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 265873f53c4aSTejun Heo 265973f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 266073f53c4aSTejun Heo wq->work_color)) { 266173f53c4aSTejun Heo /* nothing to flush, done */ 266273f53c4aSTejun Heo wq->flush_color = next_color; 266373f53c4aSTejun Heo wq->first_flusher = NULL; 266473f53c4aSTejun Heo goto out_unlock; 266573f53c4aSTejun Heo } 266673f53c4aSTejun Heo } else { 266773f53c4aSTejun Heo /* wait in queue */ 266873f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 266973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 267073f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 267173f53c4aSTejun Heo } 267273f53c4aSTejun Heo } else { 267373f53c4aSTejun Heo /* 267473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 267573f53c4aSTejun Heo * The next flush completion will assign us 267673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 267773f53c4aSTejun Heo */ 267873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 267973f53c4aSTejun Heo } 268073f53c4aSTejun Heo 268173f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 268273f53c4aSTejun Heo 268373f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 268473f53c4aSTejun Heo 268573f53c4aSTejun Heo /* 268673f53c4aSTejun Heo * Wake-up-and-cascade phase 268773f53c4aSTejun Heo * 268873f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 268973f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 269073f53c4aSTejun Heo */ 269173f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 269273f53c4aSTejun Heo return; 269373f53c4aSTejun Heo 269473f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 269573f53c4aSTejun Heo 26964ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26974ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26984ce48b37STejun Heo goto out_unlock; 26994ce48b37STejun Heo 270073f53c4aSTejun Heo wq->first_flusher = NULL; 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 270373f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 270473f53c4aSTejun Heo 270573f53c4aSTejun Heo while (true) { 270673f53c4aSTejun Heo struct wq_flusher *next, *tmp; 270773f53c4aSTejun Heo 270873f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 270973f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 271073f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 271173f53c4aSTejun Heo break; 271273f53c4aSTejun Heo list_del_init(&next->list); 271373f53c4aSTejun Heo complete(&next->done); 271473f53c4aSTejun Heo } 271573f53c4aSTejun Heo 271673f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 271773f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 271873f53c4aSTejun Heo 271973f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 272073f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 272173f53c4aSTejun Heo 272273f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 272373f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 272473f53c4aSTejun Heo /* 272573f53c4aSTejun Heo * Assign the same color to all overflowed 272673f53c4aSTejun Heo * flushers, advance work_color and append to 272773f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 272873f53c4aSTejun Heo * phase for these overflowed flushers. 272973f53c4aSTejun Heo */ 273073f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 273173f53c4aSTejun Heo tmp->flush_color = wq->work_color; 273273f53c4aSTejun Heo 273373f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 273473f53c4aSTejun Heo 273573f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 273673f53c4aSTejun Heo &wq->flusher_queue); 273773f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 273873f53c4aSTejun Heo } 273973f53c4aSTejun Heo 274073f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 274173f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 274273f53c4aSTejun Heo break; 274373f53c4aSTejun Heo } 274473f53c4aSTejun Heo 274573f53c4aSTejun Heo /* 274673f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 274773f53c4aSTejun Heo * the new first flusher and arm cwqs. 274873f53c4aSTejun Heo */ 274973f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 275073f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 275173f53c4aSTejun Heo 275273f53c4aSTejun Heo list_del_init(&next->list); 275373f53c4aSTejun Heo wq->first_flusher = next; 275473f53c4aSTejun Heo 275573f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 275673f53c4aSTejun Heo break; 275773f53c4aSTejun Heo 275873f53c4aSTejun Heo /* 275973f53c4aSTejun Heo * Meh... this color is already done, clear first 276073f53c4aSTejun Heo * flusher and repeat cascading. 276173f53c4aSTejun Heo */ 276273f53c4aSTejun Heo wq->first_flusher = NULL; 276373f53c4aSTejun Heo } 276473f53c4aSTejun Heo 276573f53c4aSTejun Heo out_unlock: 276673f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27671da177e4SLinus Torvalds } 2768ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27691da177e4SLinus Torvalds 27709c5a2ba7STejun Heo /** 27719c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27729c5a2ba7STejun Heo * @wq: workqueue to drain 27739c5a2ba7STejun Heo * 27749c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27759c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27769c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27779c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27789c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27799c5a2ba7STejun Heo * takes too long. 27809c5a2ba7STejun Heo */ 27819c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27829c5a2ba7STejun Heo { 27839c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27849c5a2ba7STejun Heo unsigned int cpu; 27859c5a2ba7STejun Heo 27869c5a2ba7STejun Heo /* 27879c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27889c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27899c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27909c5a2ba7STejun Heo */ 27919c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27929c5a2ba7STejun Heo if (!wq->nr_drainers++) 27939c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 27949c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27959c5a2ba7STejun Heo reflush: 27969c5a2ba7STejun Heo flush_workqueue(wq); 27979c5a2ba7STejun Heo 27989c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 27999c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2800fa2563e4SThomas Tuttle bool drained; 28019c5a2ba7STejun Heo 2802bd7bdd43STejun Heo spin_lock_irq(&cwq->pool->gcwq->lock); 2803fa2563e4SThomas Tuttle drained = !cwq->nr_active && list_empty(&cwq->delayed_works); 2804bd7bdd43STejun Heo spin_unlock_irq(&cwq->pool->gcwq->lock); 2805fa2563e4SThomas Tuttle 2806fa2563e4SThomas Tuttle if (drained) 28079c5a2ba7STejun Heo continue; 28089c5a2ba7STejun Heo 28099c5a2ba7STejun Heo if (++flush_cnt == 10 || 28109c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2811044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 28129c5a2ba7STejun Heo wq->name, flush_cnt); 28139c5a2ba7STejun Heo goto reflush; 28149c5a2ba7STejun Heo } 28159c5a2ba7STejun Heo 28169c5a2ba7STejun Heo spin_lock(&workqueue_lock); 28179c5a2ba7STejun Heo if (!--wq->nr_drainers) 28189c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 28199c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 28209c5a2ba7STejun Heo } 28219c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28229c5a2ba7STejun Heo 2823606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2824baf59022STejun Heo { 2825baf59022STejun Heo struct worker *worker = NULL; 2826baf59022STejun Heo struct global_cwq *gcwq; 2827baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2828baf59022STejun Heo 2829baf59022STejun Heo might_sleep(); 2830baf59022STejun Heo gcwq = get_work_gcwq(work); 2831baf59022STejun Heo if (!gcwq) 2832baf59022STejun Heo return false; 2833baf59022STejun Heo 2834baf59022STejun Heo spin_lock_irq(&gcwq->lock); 2835baf59022STejun Heo if (!list_empty(&work->entry)) { 2836baf59022STejun Heo /* 2837baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2838baf59022STejun Heo * If it was re-queued to a different gcwq under us, we 2839baf59022STejun Heo * are not going to wait. 2840baf59022STejun Heo */ 2841baf59022STejun Heo smp_rmb(); 2842baf59022STejun Heo cwq = get_work_cwq(work); 2843bd7bdd43STejun Heo if (unlikely(!cwq || gcwq != cwq->pool->gcwq)) 2844baf59022STejun Heo goto already_gone; 2845606a5020STejun Heo } else { 2846baf59022STejun Heo worker = find_worker_executing_work(gcwq, work); 2847baf59022STejun Heo if (!worker) 2848baf59022STejun Heo goto already_gone; 2849baf59022STejun Heo cwq = worker->current_cwq; 2850606a5020STejun Heo } 2851baf59022STejun Heo 2852baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2853baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2854baf59022STejun Heo 2855e159489bSTejun Heo /* 2856e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2857e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2858e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2859e159489bSTejun Heo * access. 2860e159489bSTejun Heo */ 2861e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2862baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2863e159489bSTejun Heo else 2864e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2865baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2866e159489bSTejun Heo 2867baf59022STejun Heo return true; 2868baf59022STejun Heo already_gone: 2869baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2870baf59022STejun Heo return false; 2871baf59022STejun Heo } 2872baf59022STejun Heo 2873db700897SOleg Nesterov /** 2874401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2875401a8d04STejun Heo * @work: the work to flush 2876db700897SOleg Nesterov * 2877606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2878606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2879401a8d04STejun Heo * 2880401a8d04STejun Heo * RETURNS: 2881401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2882401a8d04STejun Heo * %false if it was already idle. 2883db700897SOleg Nesterov */ 2884401a8d04STejun Heo bool flush_work(struct work_struct *work) 2885db700897SOleg Nesterov { 2886db700897SOleg Nesterov struct wq_barrier barr; 2887db700897SOleg Nesterov 28880976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28890976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28900976dfc1SStephen Boyd 2891606a5020STejun Heo if (start_flush_work(work, &barr)) { 2892db700897SOleg Nesterov wait_for_completion(&barr.done); 2893dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2894401a8d04STejun Heo return true; 2895606a5020STejun Heo } else { 2896401a8d04STejun Heo return false; 2897db700897SOleg Nesterov } 2898606a5020STejun Heo } 2899db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2900db700897SOleg Nesterov 290136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 29021f1f642eSOleg Nesterov { 2903bbb68dfaSTejun Heo unsigned long flags; 29041f1f642eSOleg Nesterov int ret; 29051f1f642eSOleg Nesterov 29061f1f642eSOleg Nesterov do { 2907bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2908bbb68dfaSTejun Heo /* 2909bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2910bbb68dfaSTejun Heo * would be waiting for before retrying. 2911bbb68dfaSTejun Heo */ 2912bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2913606a5020STejun Heo flush_work(work); 29141f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29151f1f642eSOleg Nesterov 2916bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2917bbb68dfaSTejun Heo mark_work_canceling(work); 2918bbb68dfaSTejun Heo local_irq_restore(flags); 2919bbb68dfaSTejun Heo 2920606a5020STejun Heo flush_work(work); 29217a22ad75STejun Heo clear_work_data(work); 29221f1f642eSOleg Nesterov return ret; 29231f1f642eSOleg Nesterov } 29241f1f642eSOleg Nesterov 29256e84d644SOleg Nesterov /** 2926401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2927401a8d04STejun Heo * @work: the work to cancel 29286e84d644SOleg Nesterov * 2929401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2930401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2931401a8d04STejun Heo * another workqueue. On return from this function, @work is 2932401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29331f1f642eSOleg Nesterov * 2934401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2935401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29366e84d644SOleg Nesterov * 2937401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29386e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2939401a8d04STejun Heo * 2940401a8d04STejun Heo * RETURNS: 2941401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29426e84d644SOleg Nesterov */ 2943401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29446e84d644SOleg Nesterov { 294536e227d2STejun Heo return __cancel_work_timer(work, false); 2946b89deed3SOleg Nesterov } 294728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2948b89deed3SOleg Nesterov 29496e84d644SOleg Nesterov /** 2950401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2951401a8d04STejun Heo * @dwork: the delayed work to flush 29526e84d644SOleg Nesterov * 2953401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2954401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2955401a8d04STejun Heo * considers the last queueing instance of @dwork. 29561f1f642eSOleg Nesterov * 2957401a8d04STejun Heo * RETURNS: 2958401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2959401a8d04STejun Heo * %false if it was already idle. 29606e84d644SOleg Nesterov */ 2961401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2962401a8d04STejun Heo { 29638930cabaSTejun Heo local_irq_disable(); 2964401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 29651265057fSTejun Heo __queue_work(dwork->cpu, 2966401a8d04STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 29678930cabaSTejun Heo local_irq_enable(); 2968401a8d04STejun Heo return flush_work(&dwork->work); 2969401a8d04STejun Heo } 2970401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2971401a8d04STejun Heo 2972401a8d04STejun Heo /** 297357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 297457b30ae7STejun Heo * @dwork: delayed_work to cancel 297557b30ae7STejun Heo * 297657b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 297757b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 297857b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 297957b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 298057b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 298157b30ae7STejun Heo * 298257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 298357b30ae7STejun Heo */ 298457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 298557b30ae7STejun Heo { 298657b30ae7STejun Heo unsigned long flags; 298757b30ae7STejun Heo int ret; 298857b30ae7STejun Heo 298957b30ae7STejun Heo do { 299057b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 299157b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 299257b30ae7STejun Heo 299357b30ae7STejun Heo if (unlikely(ret < 0)) 299457b30ae7STejun Heo return false; 299557b30ae7STejun Heo 299657b30ae7STejun Heo set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work)); 299757b30ae7STejun Heo local_irq_restore(flags); 299857b30ae7STejun Heo return true; 299957b30ae7STejun Heo } 300057b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 300157b30ae7STejun Heo 300257b30ae7STejun Heo /** 3003401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3004401a8d04STejun Heo * @dwork: the delayed work cancel 3005401a8d04STejun Heo * 3006401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3007401a8d04STejun Heo * 3008401a8d04STejun Heo * RETURNS: 3009401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3010401a8d04STejun Heo */ 3011401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30126e84d644SOleg Nesterov { 301336e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30146e84d644SOleg Nesterov } 3015f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30161da177e4SLinus Torvalds 3017d4283e93STejun Heo /** 30180a13c00eSTejun Heo * schedule_work_on - put work task on a specific cpu 30190a13c00eSTejun Heo * @cpu: cpu to put the work task on 30200a13c00eSTejun Heo * @work: job to be done 30210a13c00eSTejun Heo * 30220a13c00eSTejun Heo * This puts a job on a specific cpu 30230a13c00eSTejun Heo */ 3024d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 30250a13c00eSTejun Heo { 30260a13c00eSTejun Heo return queue_work_on(cpu, system_wq, work); 30270a13c00eSTejun Heo } 30280a13c00eSTejun Heo EXPORT_SYMBOL(schedule_work_on); 30290a13c00eSTejun Heo 30300fcb78c2SRolf Eike Beer /** 30310fcb78c2SRolf Eike Beer * schedule_work - put work task in global workqueue 30320fcb78c2SRolf Eike Beer * @work: job to be done 30330fcb78c2SRolf Eike Beer * 3034d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 3035d4283e93STejun Heo * %true otherwise. 30365b0f437dSBart Van Assche * 30375b0f437dSBart Van Assche * This puts a job in the kernel-global workqueue if it was not already 30385b0f437dSBart Van Assche * queued and leaves it in the same position on the kernel-global 30395b0f437dSBart Van Assche * workqueue otherwise. 30400fcb78c2SRolf Eike Beer */ 3041d4283e93STejun Heo bool schedule_work(struct work_struct *work) 30421da177e4SLinus Torvalds { 3043d320c038STejun Heo return queue_work(system_wq, work); 30441da177e4SLinus Torvalds } 3045ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_work); 30461da177e4SLinus Torvalds 30470fcb78c2SRolf Eike Beer /** 30480fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 30490fcb78c2SRolf Eike Beer * @cpu: cpu to use 305052bad64dSDavid Howells * @dwork: job to be done 30510fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 30520fcb78c2SRolf Eike Beer * 30530fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 30540fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 30550fcb78c2SRolf Eike Beer */ 3056d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 3057d4283e93STejun Heo unsigned long delay) 30581da177e4SLinus Torvalds { 3059d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 30601da177e4SLinus Torvalds } 3061ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 30621da177e4SLinus Torvalds 3063b6136773SAndrew Morton /** 30640a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 30650a13c00eSTejun Heo * @dwork: job to be done 30660a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 30670a13c00eSTejun Heo * 30680a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 30690a13c00eSTejun Heo * workqueue. 30700a13c00eSTejun Heo */ 3071d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 30720a13c00eSTejun Heo { 30730a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30740a13c00eSTejun Heo } 30750a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30760a13c00eSTejun Heo 30770a13c00eSTejun Heo /** 307831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3079b6136773SAndrew Morton * @func: the function to call 3080b6136773SAndrew Morton * 308131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 308231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3083b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 308431ddd871STejun Heo * 308531ddd871STejun Heo * RETURNS: 308631ddd871STejun Heo * 0 on success, -errno on failure. 3087b6136773SAndrew Morton */ 308865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 308915316ba8SChristoph Lameter { 309015316ba8SChristoph Lameter int cpu; 309138f51568SNamhyung Kim struct work_struct __percpu *works; 309215316ba8SChristoph Lameter 3093b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3094b6136773SAndrew Morton if (!works) 309515316ba8SChristoph Lameter return -ENOMEM; 3096b6136773SAndrew Morton 309795402b38SGautham R Shenoy get_online_cpus(); 309893981800STejun Heo 309915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 31009bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 31019bfb1839SIngo Molnar 31029bfb1839SIngo Molnar INIT_WORK(work, func); 31038de6d308SOleg Nesterov schedule_work_on(cpu, work); 310415316ba8SChristoph Lameter } 310593981800STejun Heo 310693981800STejun Heo for_each_online_cpu(cpu) 31078616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 310893981800STejun Heo 310995402b38SGautham R Shenoy put_online_cpus(); 3110b6136773SAndrew Morton free_percpu(works); 311115316ba8SChristoph Lameter return 0; 311215316ba8SChristoph Lameter } 311315316ba8SChristoph Lameter 3114eef6a7d5SAlan Stern /** 3115eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3116eef6a7d5SAlan Stern * 3117eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3118eef6a7d5SAlan Stern * completion. 3119eef6a7d5SAlan Stern * 3120eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3121eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3122eef6a7d5SAlan Stern * will lead to deadlock: 3123eef6a7d5SAlan Stern * 3124eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3125eef6a7d5SAlan Stern * a lock held by your code or its caller. 3126eef6a7d5SAlan Stern * 3127eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3128eef6a7d5SAlan Stern * 3129eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3130eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3131eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3132eef6a7d5SAlan Stern * 3133eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3134eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3135eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3136eef6a7d5SAlan Stern * cancel_work_sync() instead. 3137eef6a7d5SAlan Stern */ 31381da177e4SLinus Torvalds void flush_scheduled_work(void) 31391da177e4SLinus Torvalds { 3140d320c038STejun Heo flush_workqueue(system_wq); 31411da177e4SLinus Torvalds } 3142ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 31431da177e4SLinus Torvalds 31441da177e4SLinus Torvalds /** 31451fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31461fa44ecaSJames Bottomley * @fn: the function to execute 31471fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31481fa44ecaSJames Bottomley * be available when the work executes) 31491fa44ecaSJames Bottomley * 31501fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31511fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31521fa44ecaSJames Bottomley * 31531fa44ecaSJames Bottomley * Returns: 0 - function was executed 31541fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31551fa44ecaSJames Bottomley */ 315665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31571fa44ecaSJames Bottomley { 31581fa44ecaSJames Bottomley if (!in_interrupt()) { 315965f27f38SDavid Howells fn(&ew->work); 31601fa44ecaSJames Bottomley return 0; 31611fa44ecaSJames Bottomley } 31621fa44ecaSJames Bottomley 316365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31641fa44ecaSJames Bottomley schedule_work(&ew->work); 31651fa44ecaSJames Bottomley 31661fa44ecaSJames Bottomley return 1; 31671fa44ecaSJames Bottomley } 31681fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31691fa44ecaSJames Bottomley 31701da177e4SLinus Torvalds int keventd_up(void) 31711da177e4SLinus Torvalds { 3172d320c038STejun Heo return system_wq != NULL; 31731da177e4SLinus Torvalds } 31741da177e4SLinus Torvalds 3175bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 31761da177e4SLinus Torvalds { 31773af24433SOleg Nesterov /* 31780f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 31790f900049STejun Heo * Make sure that the alignment isn't lower than that of 31800f900049STejun Heo * unsigned long long. 31813af24433SOleg Nesterov */ 31820f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 31830f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 31840f900049STejun Heo __alignof__(unsigned long long)); 31853af24433SOleg Nesterov 3186e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3187f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 3188931ac77eSTejun Heo else { 31890f900049STejun Heo void *ptr; 3190e1d8aa9fSFrederic Weisbecker 31910f900049STejun Heo /* 3192f3421797STejun Heo * Allocate enough room to align cwq and put an extra 3193f3421797STejun Heo * pointer at the end pointing back to the originally 3194f3421797STejun Heo * allocated pointer which will be used for free. 31950f900049STejun Heo */ 3196bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3197bdbc5dd7STejun Heo if (ptr) { 3198bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 3199bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 3200bdbc5dd7STejun Heo } 32013af24433SOleg Nesterov } 32023af24433SOleg Nesterov 32030415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3204bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 3205bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 32060f900049STejun Heo } 32070f900049STejun Heo 3208bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 320906ba38a9SOleg Nesterov { 3210e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3211bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 3212f3421797STejun Heo else if (wq->cpu_wq.single) { 3213f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 3214f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 321506ba38a9SOleg Nesterov } 321606ba38a9SOleg Nesterov } 321706ba38a9SOleg Nesterov 3218f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3219f3421797STejun Heo const char *name) 3220b71ab8c2STejun Heo { 3221f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3222f3421797STejun Heo 3223f3421797STejun Heo if (max_active < 1 || max_active > lim) 3224044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3225f3421797STejun Heo max_active, name, 1, lim); 3226b71ab8c2STejun Heo 3227f3421797STejun Heo return clamp_val(max_active, 1, lim); 3228b71ab8c2STejun Heo } 3229b71ab8c2STejun Heo 3230b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 323197e37d7bSTejun Heo unsigned int flags, 32321e19ffc6STejun Heo int max_active, 3233eb13ba87SJohannes Berg struct lock_class_key *key, 3234b196be89STejun Heo const char *lock_name, ...) 32353af24433SOleg Nesterov { 3236b196be89STejun Heo va_list args, args1; 32373af24433SOleg Nesterov struct workqueue_struct *wq; 3238c34056a3STejun Heo unsigned int cpu; 3239b196be89STejun Heo size_t namelen; 3240b196be89STejun Heo 3241b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3242b196be89STejun Heo va_start(args, lock_name); 3243b196be89STejun Heo va_copy(args1, args); 3244b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3245b196be89STejun Heo 3246b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3247b196be89STejun Heo if (!wq) 3248b196be89STejun Heo goto err; 3249b196be89STejun Heo 3250b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3251b196be89STejun Heo va_end(args); 3252b196be89STejun Heo va_end(args1); 32533af24433SOleg Nesterov 3254f3421797STejun Heo /* 32556370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 32566370a6adSTejun Heo * have a rescuer to guarantee forward progress. 32576370a6adSTejun Heo */ 32586370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 32596370a6adSTejun Heo flags |= WQ_RESCUER; 32606370a6adSTejun Heo 3261d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3262b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 32633af24433SOleg Nesterov 3264b196be89STejun Heo /* init wq */ 326597e37d7bSTejun Heo wq->flags = flags; 3266a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 326773f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 326873f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 326973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 327073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 32713af24433SOleg Nesterov 3272eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3273cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 32743af24433SOleg Nesterov 3275bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 3276bdbc5dd7STejun Heo goto err; 3277bdbc5dd7STejun Heo 3278f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 32791537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 32808b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 32813270476aSTejun Heo int pool_idx = (bool)(flags & WQ_HIGHPRI); 32821537663fSTejun Heo 32830f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 32843270476aSTejun Heo cwq->pool = &gcwq->pools[pool_idx]; 3285c34056a3STejun Heo cwq->wq = wq; 328673f53c4aSTejun Heo cwq->flush_color = -1; 32871e19ffc6STejun Heo cwq->max_active = max_active; 32881e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3289e22bee78STejun Heo } 32901537663fSTejun Heo 3291e22bee78STejun Heo if (flags & WQ_RESCUER) { 3292e22bee78STejun Heo struct worker *rescuer; 3293e22bee78STejun Heo 3294f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3295e22bee78STejun Heo goto err; 3296e22bee78STejun Heo 3297e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3298e22bee78STejun Heo if (!rescuer) 3299e22bee78STejun Heo goto err; 3300e22bee78STejun Heo 3301b196be89STejun Heo rescuer->task = kthread_create(rescuer_thread, wq, "%s", 3302b196be89STejun Heo wq->name); 3303e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3304e22bee78STejun Heo goto err; 3305e22bee78STejun Heo 3306e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3307e22bee78STejun Heo wake_up_process(rescuer->task); 33083af24433SOleg Nesterov } 33091537663fSTejun Heo 33103af24433SOleg Nesterov /* 3311a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3312a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3313a0a1a5fdSTejun Heo * workqueue to workqueues list. 33143af24433SOleg Nesterov */ 33153af24433SOleg Nesterov spin_lock(&workqueue_lock); 3316a0a1a5fdSTejun Heo 331758a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3318f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3319a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3320a0a1a5fdSTejun Heo 33213af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3322a0a1a5fdSTejun Heo 33233af24433SOleg Nesterov spin_unlock(&workqueue_lock); 33243af24433SOleg Nesterov 33253af24433SOleg Nesterov return wq; 33264690c4abSTejun Heo err: 33274690c4abSTejun Heo if (wq) { 3328bdbc5dd7STejun Heo free_cwqs(wq); 3329f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3330e22bee78STejun Heo kfree(wq->rescuer); 33314690c4abSTejun Heo kfree(wq); 33323af24433SOleg Nesterov } 33334690c4abSTejun Heo return NULL; 33341da177e4SLinus Torvalds } 3335d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 33361da177e4SLinus Torvalds 33373af24433SOleg Nesterov /** 33383af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 33393af24433SOleg Nesterov * @wq: target workqueue 33403af24433SOleg Nesterov * 33413af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 33423af24433SOleg Nesterov */ 33433af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 33443af24433SOleg Nesterov { 3345c8e55f36STejun Heo unsigned int cpu; 33463af24433SOleg Nesterov 33479c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 33489c5a2ba7STejun Heo drain_workqueue(wq); 3349c8efcc25STejun Heo 3350a0a1a5fdSTejun Heo /* 3351a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3352a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3353a0a1a5fdSTejun Heo */ 335495402b38SGautham R Shenoy spin_lock(&workqueue_lock); 33553af24433SOleg Nesterov list_del(&wq->list); 335695402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 33573af24433SOleg Nesterov 3358e22bee78STejun Heo /* sanity check */ 3359f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 336073f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 336173f53c4aSTejun Heo int i; 33623af24433SOleg Nesterov 336373f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 336473f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 33651e19ffc6STejun Heo BUG_ON(cwq->nr_active); 33661e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 336773f53c4aSTejun Heo } 33681537663fSTejun Heo 3369e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3370e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3371f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 33728d9df9f0SXiaotian Feng kfree(wq->rescuer); 3373e22bee78STejun Heo } 3374e22bee78STejun Heo 3375bdbc5dd7STejun Heo free_cwqs(wq); 33763af24433SOleg Nesterov kfree(wq); 33773af24433SOleg Nesterov } 33783af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33793af24433SOleg Nesterov 3380dcd989cbSTejun Heo /** 3381dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3382dcd989cbSTejun Heo * @wq: target workqueue 3383dcd989cbSTejun Heo * @max_active: new max_active value. 3384dcd989cbSTejun Heo * 3385dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3386dcd989cbSTejun Heo * 3387dcd989cbSTejun Heo * CONTEXT: 3388dcd989cbSTejun Heo * Don't call from IRQ context. 3389dcd989cbSTejun Heo */ 3390dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3391dcd989cbSTejun Heo { 3392dcd989cbSTejun Heo unsigned int cpu; 3393dcd989cbSTejun Heo 3394f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3395dcd989cbSTejun Heo 3396dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3397dcd989cbSTejun Heo 3398dcd989cbSTejun Heo wq->saved_max_active = max_active; 3399dcd989cbSTejun Heo 3400f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 3401dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3402dcd989cbSTejun Heo 3403dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 3404dcd989cbSTejun Heo 340558a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 3406dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 3407dcd989cbSTejun Heo get_cwq(gcwq->cpu, wq)->max_active = max_active; 3408dcd989cbSTejun Heo 3409dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 3410dcd989cbSTejun Heo } 3411dcd989cbSTejun Heo 3412dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3413dcd989cbSTejun Heo } 3414dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3415dcd989cbSTejun Heo 3416dcd989cbSTejun Heo /** 3417dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3418dcd989cbSTejun Heo * @cpu: CPU in question 3419dcd989cbSTejun Heo * @wq: target workqueue 3420dcd989cbSTejun Heo * 3421dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3422dcd989cbSTejun Heo * no synchronization around this function and the test result is 3423dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3424dcd989cbSTejun Heo * 3425dcd989cbSTejun Heo * RETURNS: 3426dcd989cbSTejun Heo * %true if congested, %false otherwise. 3427dcd989cbSTejun Heo */ 3428dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3429dcd989cbSTejun Heo { 3430dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3431dcd989cbSTejun Heo 3432dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3433dcd989cbSTejun Heo } 3434dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3435dcd989cbSTejun Heo 3436dcd989cbSTejun Heo /** 3437dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 3438dcd989cbSTejun Heo * @work: the work of interest 3439dcd989cbSTejun Heo * 3440dcd989cbSTejun Heo * RETURNS: 3441bdbc5dd7STejun Heo * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. 3442dcd989cbSTejun Heo */ 3443dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 3444dcd989cbSTejun Heo { 3445dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3446dcd989cbSTejun Heo 3447bdbc5dd7STejun Heo return gcwq ? gcwq->cpu : WORK_CPU_NONE; 3448dcd989cbSTejun Heo } 3449dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 3450dcd989cbSTejun Heo 3451dcd989cbSTejun Heo /** 3452dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3453dcd989cbSTejun Heo * @work: the work to be tested 3454dcd989cbSTejun Heo * 3455dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3456dcd989cbSTejun Heo * synchronization around this function and the test result is 3457dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3458dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 3459dcd989cbSTejun Heo * running state. 3460dcd989cbSTejun Heo * 3461dcd989cbSTejun Heo * RETURNS: 3462dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3463dcd989cbSTejun Heo */ 3464dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3465dcd989cbSTejun Heo { 3466dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3467dcd989cbSTejun Heo unsigned long flags; 3468dcd989cbSTejun Heo unsigned int ret = 0; 3469dcd989cbSTejun Heo 3470dcd989cbSTejun Heo if (!gcwq) 3471dcd989cbSTejun Heo return false; 3472dcd989cbSTejun Heo 3473dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 3474dcd989cbSTejun Heo 3475dcd989cbSTejun Heo if (work_pending(work)) 3476dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3477dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 3478dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3479dcd989cbSTejun Heo 3480dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 3481dcd989cbSTejun Heo 3482dcd989cbSTejun Heo return ret; 3483dcd989cbSTejun Heo } 3484dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3485dcd989cbSTejun Heo 3486db7bccf4STejun Heo /* 3487db7bccf4STejun Heo * CPU hotplug. 3488db7bccf4STejun Heo * 3489e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3490e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3491e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 3492e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 3493e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 3494e22bee78STejun Heo * blocked draining impractical. 3495e22bee78STejun Heo * 3496628c78e7STejun Heo * This is solved by allowing a gcwq to be disassociated from the CPU 3497628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3498628c78e7STejun Heo * cpu comes back online. 3499db7bccf4STejun Heo */ 3500db7bccf4STejun Heo 350160373152STejun Heo /* claim manager positions of all pools */ 35028db25e78STejun Heo static void gcwq_claim_management_and_lock(struct global_cwq *gcwq) 350360373152STejun Heo { 350460373152STejun Heo struct worker_pool *pool; 350560373152STejun Heo 350660373152STejun Heo for_each_worker_pool(pool, gcwq) 350760373152STejun Heo mutex_lock_nested(&pool->manager_mutex, pool - gcwq->pools); 35088db25e78STejun Heo spin_lock_irq(&gcwq->lock); 350960373152STejun Heo } 351060373152STejun Heo 351160373152STejun Heo /* release manager positions */ 35128db25e78STejun Heo static void gcwq_release_management_and_unlock(struct global_cwq *gcwq) 351360373152STejun Heo { 351460373152STejun Heo struct worker_pool *pool; 351560373152STejun Heo 35168db25e78STejun Heo spin_unlock_irq(&gcwq->lock); 351760373152STejun Heo for_each_worker_pool(pool, gcwq) 351860373152STejun Heo mutex_unlock(&pool->manager_mutex); 351960373152STejun Heo } 352060373152STejun Heo 3521628c78e7STejun Heo static void gcwq_unbind_fn(struct work_struct *work) 3522db7bccf4STejun Heo { 3523628c78e7STejun Heo struct global_cwq *gcwq = get_gcwq(smp_processor_id()); 35244ce62e9eSTejun Heo struct worker_pool *pool; 3525db7bccf4STejun Heo struct worker *worker; 3526db7bccf4STejun Heo struct hlist_node *pos; 3527db7bccf4STejun Heo int i; 3528db7bccf4STejun Heo 3529db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3530db7bccf4STejun Heo 35318db25e78STejun Heo gcwq_claim_management_and_lock(gcwq); 3532e22bee78STejun Heo 3533f2d5a0eeSTejun Heo /* 3534f2d5a0eeSTejun Heo * We've claimed all manager positions. Make all workers unbound 3535f2d5a0eeSTejun Heo * and set DISASSOCIATED. Before this, all workers except for the 3536f2d5a0eeSTejun Heo * ones which are still executing works from before the last CPU 3537f2d5a0eeSTejun Heo * down must be on the cpu. After this, they may become diasporas. 3538f2d5a0eeSTejun Heo */ 353960373152STejun Heo for_each_worker_pool(pool, gcwq) 35404ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3541403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3542db7bccf4STejun Heo 3543db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 3544403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3545db7bccf4STejun Heo 3546f2d5a0eeSTejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3547f2d5a0eeSTejun Heo 35488db25e78STejun Heo gcwq_release_management_and_unlock(gcwq); 3549e22bee78STejun Heo 3550e22bee78STejun Heo /* 3551628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3552628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3553628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3554628c78e7STejun Heo */ 3555628c78e7STejun Heo schedule(); 3556628c78e7STejun Heo 3557628c78e7STejun Heo /* 3558628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3559628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 3560628c78e7STejun Heo * are always true as long as the worklist is not empty. @gcwq now 3561628c78e7STejun Heo * behaves as unbound (in terms of concurrency management) gcwq 3562628c78e7STejun Heo * which is served by workers tied to the CPU. 3563628c78e7STejun Heo * 3564628c78e7STejun Heo * On return from this function, the current worker would trigger 3565628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3566628c78e7STejun Heo * didn't already. 3567e22bee78STejun Heo */ 35684ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 35694ce62e9eSTejun Heo atomic_set(get_pool_nr_running(pool), 0); 3570db7bccf4STejun Heo } 3571db7bccf4STejun Heo 35728db25e78STejun Heo /* 35738db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 35748db25e78STejun Heo * This will be registered high priority CPU notifier. 35758db25e78STejun Heo */ 35768db25e78STejun Heo static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb, 35771da177e4SLinus Torvalds unsigned long action, 35781da177e4SLinus Torvalds void *hcpu) 35791da177e4SLinus Torvalds { 35803af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3581db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 35824ce62e9eSTejun Heo struct worker_pool *pool; 35831da177e4SLinus Torvalds 35848db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 35853af24433SOleg Nesterov case CPU_UP_PREPARE: 35864ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 35873ce63377STejun Heo struct worker *worker; 35883ce63377STejun Heo 35893ce63377STejun Heo if (pool->nr_workers) 35903ce63377STejun Heo continue; 35913ce63377STejun Heo 35923ce63377STejun Heo worker = create_worker(pool); 35933ce63377STejun Heo if (!worker) 35943ce63377STejun Heo return NOTIFY_BAD; 35953ce63377STejun Heo 35963ce63377STejun Heo spin_lock_irq(&gcwq->lock); 35973ce63377STejun Heo start_worker(worker); 35983ce63377STejun Heo spin_unlock_irq(&gcwq->lock); 35993af24433SOleg Nesterov } 36001da177e4SLinus Torvalds break; 36011da177e4SLinus Torvalds 360265758202STejun Heo case CPU_DOWN_FAILED: 360365758202STejun Heo case CPU_ONLINE: 36048db25e78STejun Heo gcwq_claim_management_and_lock(gcwq); 36058db25e78STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 36068db25e78STejun Heo rebind_workers(gcwq); 36078db25e78STejun Heo gcwq_release_management_and_unlock(gcwq); 36088db25e78STejun Heo break; 360965758202STejun Heo } 361065758202STejun Heo return NOTIFY_OK; 361165758202STejun Heo } 361265758202STejun Heo 361365758202STejun Heo /* 361465758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 361565758202STejun Heo * This will be registered as low priority CPU notifier. 361665758202STejun Heo */ 361765758202STejun Heo static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb, 361865758202STejun Heo unsigned long action, 361965758202STejun Heo void *hcpu) 362065758202STejun Heo { 36218db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 36228db25e78STejun Heo struct work_struct unbind_work; 36238db25e78STejun Heo 362465758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 362565758202STejun Heo case CPU_DOWN_PREPARE: 36268db25e78STejun Heo /* unbinding should happen on the local CPU */ 36278db25e78STejun Heo INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn); 36287635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 36298db25e78STejun Heo flush_work(&unbind_work); 36308db25e78STejun Heo break; 363165758202STejun Heo } 363265758202STejun Heo return NOTIFY_OK; 363365758202STejun Heo } 363465758202STejun Heo 36352d3854a3SRusty Russell #ifdef CONFIG_SMP 36368ccad40dSRusty Russell 36372d3854a3SRusty Russell struct work_for_cpu { 36386b44003eSAndrew Morton struct completion completion; 36392d3854a3SRusty Russell long (*fn)(void *); 36402d3854a3SRusty Russell void *arg; 36412d3854a3SRusty Russell long ret; 36422d3854a3SRusty Russell }; 36432d3854a3SRusty Russell 36446b44003eSAndrew Morton static int do_work_for_cpu(void *_wfc) 36452d3854a3SRusty Russell { 36466b44003eSAndrew Morton struct work_for_cpu *wfc = _wfc; 36472d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 36486b44003eSAndrew Morton complete(&wfc->completion); 36496b44003eSAndrew Morton return 0; 36502d3854a3SRusty Russell } 36512d3854a3SRusty Russell 36522d3854a3SRusty Russell /** 36532d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 36542d3854a3SRusty Russell * @cpu: the cpu to run on 36552d3854a3SRusty Russell * @fn: the function to run 36562d3854a3SRusty Russell * @arg: the function arg 36572d3854a3SRusty Russell * 365831ad9081SRusty Russell * This will return the value @fn returns. 365931ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 36606b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 36612d3854a3SRusty Russell */ 36622d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 36632d3854a3SRusty Russell { 36646b44003eSAndrew Morton struct task_struct *sub_thread; 36656b44003eSAndrew Morton struct work_for_cpu wfc = { 36666b44003eSAndrew Morton .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), 36676b44003eSAndrew Morton .fn = fn, 36686b44003eSAndrew Morton .arg = arg, 36696b44003eSAndrew Morton }; 36702d3854a3SRusty Russell 36716b44003eSAndrew Morton sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); 36726b44003eSAndrew Morton if (IS_ERR(sub_thread)) 36736b44003eSAndrew Morton return PTR_ERR(sub_thread); 36746b44003eSAndrew Morton kthread_bind(sub_thread, cpu); 36756b44003eSAndrew Morton wake_up_process(sub_thread); 36766b44003eSAndrew Morton wait_for_completion(&wfc.completion); 36772d3854a3SRusty Russell return wfc.ret; 36782d3854a3SRusty Russell } 36792d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 36802d3854a3SRusty Russell #endif /* CONFIG_SMP */ 36812d3854a3SRusty Russell 3682a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3683e7577c50SRusty Russell 3684a0a1a5fdSTejun Heo /** 3685a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3686a0a1a5fdSTejun Heo * 368758a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 368858a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 368958a69cb4STejun Heo * gcwq->worklist. 3690a0a1a5fdSTejun Heo * 3691a0a1a5fdSTejun Heo * CONTEXT: 36928b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3693a0a1a5fdSTejun Heo */ 3694a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3695a0a1a5fdSTejun Heo { 3696a0a1a5fdSTejun Heo unsigned int cpu; 3697a0a1a5fdSTejun Heo 3698a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3699a0a1a5fdSTejun Heo 3700a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3701a0a1a5fdSTejun Heo workqueue_freezing = true; 3702a0a1a5fdSTejun Heo 3703f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37048b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3705bdbc5dd7STejun Heo struct workqueue_struct *wq; 37068b03ae3cSTejun Heo 37078b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 37088b03ae3cSTejun Heo 3709db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3710db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3711db7bccf4STejun Heo 3712a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3713a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3714a0a1a5fdSTejun Heo 371558a69cb4STejun Heo if (cwq && wq->flags & WQ_FREEZABLE) 3716a0a1a5fdSTejun Heo cwq->max_active = 0; 37171da177e4SLinus Torvalds } 37188b03ae3cSTejun Heo 37198b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3720a0a1a5fdSTejun Heo } 3721a0a1a5fdSTejun Heo 3722a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3723a0a1a5fdSTejun Heo } 3724a0a1a5fdSTejun Heo 3725a0a1a5fdSTejun Heo /** 372658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3727a0a1a5fdSTejun Heo * 3728a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3729a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3730a0a1a5fdSTejun Heo * 3731a0a1a5fdSTejun Heo * CONTEXT: 3732a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3733a0a1a5fdSTejun Heo * 3734a0a1a5fdSTejun Heo * RETURNS: 373558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 373658a69cb4STejun Heo * is complete. 3737a0a1a5fdSTejun Heo */ 3738a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3739a0a1a5fdSTejun Heo { 3740a0a1a5fdSTejun Heo unsigned int cpu; 3741a0a1a5fdSTejun Heo bool busy = false; 3742a0a1a5fdSTejun Heo 3743a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3744a0a1a5fdSTejun Heo 3745a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3746a0a1a5fdSTejun Heo 3747f3421797STejun Heo for_each_gcwq_cpu(cpu) { 3748bdbc5dd7STejun Heo struct workqueue_struct *wq; 3749a0a1a5fdSTejun Heo /* 3750a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3751a0a1a5fdSTejun Heo * to peek without lock. 3752a0a1a5fdSTejun Heo */ 3753a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3754a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3755a0a1a5fdSTejun Heo 375658a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3757a0a1a5fdSTejun Heo continue; 3758a0a1a5fdSTejun Heo 3759a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3760a0a1a5fdSTejun Heo if (cwq->nr_active) { 3761a0a1a5fdSTejun Heo busy = true; 3762a0a1a5fdSTejun Heo goto out_unlock; 3763a0a1a5fdSTejun Heo } 3764a0a1a5fdSTejun Heo } 3765a0a1a5fdSTejun Heo } 3766a0a1a5fdSTejun Heo out_unlock: 3767a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3768a0a1a5fdSTejun Heo return busy; 3769a0a1a5fdSTejun Heo } 3770a0a1a5fdSTejun Heo 3771a0a1a5fdSTejun Heo /** 3772a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3773a0a1a5fdSTejun Heo * 3774a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 37757e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3776a0a1a5fdSTejun Heo * 3777a0a1a5fdSTejun Heo * CONTEXT: 37788b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3779a0a1a5fdSTejun Heo */ 3780a0a1a5fdSTejun Heo void thaw_workqueues(void) 3781a0a1a5fdSTejun Heo { 3782a0a1a5fdSTejun Heo unsigned int cpu; 3783a0a1a5fdSTejun Heo 3784a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3785a0a1a5fdSTejun Heo 3786a0a1a5fdSTejun Heo if (!workqueue_freezing) 3787a0a1a5fdSTejun Heo goto out_unlock; 3788a0a1a5fdSTejun Heo 3789f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37908b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 37914ce62e9eSTejun Heo struct worker_pool *pool; 3792bdbc5dd7STejun Heo struct workqueue_struct *wq; 37938b03ae3cSTejun Heo 37948b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 37958b03ae3cSTejun Heo 3796db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3797db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3798db7bccf4STejun Heo 3799a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3800a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3801a0a1a5fdSTejun Heo 380258a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3803a0a1a5fdSTejun Heo continue; 3804a0a1a5fdSTejun Heo 3805a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3806a0a1a5fdSTejun Heo cwq->max_active = wq->saved_max_active; 3807a0a1a5fdSTejun Heo 3808a0a1a5fdSTejun Heo while (!list_empty(&cwq->delayed_works) && 3809a0a1a5fdSTejun Heo cwq->nr_active < cwq->max_active) 3810a0a1a5fdSTejun Heo cwq_activate_first_delayed(cwq); 3811a0a1a5fdSTejun Heo } 38128b03ae3cSTejun Heo 38134ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 38144ce62e9eSTejun Heo wake_up_worker(pool); 3815e22bee78STejun Heo 38168b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3817a0a1a5fdSTejun Heo } 3818a0a1a5fdSTejun Heo 3819a0a1a5fdSTejun Heo workqueue_freezing = false; 3820a0a1a5fdSTejun Heo out_unlock: 3821a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3822a0a1a5fdSTejun Heo } 3823a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3824a0a1a5fdSTejun Heo 38256ee0578bSSuresh Siddha static int __init init_workqueues(void) 38261da177e4SLinus Torvalds { 3827c34056a3STejun Heo unsigned int cpu; 3828c8e55f36STejun Heo int i; 3829c34056a3STejun Heo 3830b5490077STejun Heo /* make sure we have enough bits for OFFQ CPU number */ 3831b5490077STejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) < 3832b5490077STejun Heo WORK_CPU_LAST); 3833b5490077STejun Heo 383465758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 383565758202STejun Heo cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 38368b03ae3cSTejun Heo 38378b03ae3cSTejun Heo /* initialize gcwqs */ 3838f3421797STejun Heo for_each_gcwq_cpu(cpu) { 38398b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38404ce62e9eSTejun Heo struct worker_pool *pool; 38418b03ae3cSTejun Heo 38428b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 38438b03ae3cSTejun Heo gcwq->cpu = cpu; 3844f3421797STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 38458b03ae3cSTejun Heo 3846c8e55f36STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) 3847c8e55f36STejun Heo INIT_HLIST_HEAD(&gcwq->busy_hash[i]); 3848c8e55f36STejun Heo 38494ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 38504ce62e9eSTejun Heo pool->gcwq = gcwq; 38514ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 38524ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3853e22bee78STejun Heo 38544ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 38554ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 38564ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3857e22bee78STejun Heo 38584ce62e9eSTejun Heo setup_timer(&pool->mayday_timer, gcwq_mayday_timeout, 38594ce62e9eSTejun Heo (unsigned long)pool); 38604ce62e9eSTejun Heo 386160373152STejun Heo mutex_init(&pool->manager_mutex); 38624ce62e9eSTejun Heo ida_init(&pool->worker_ida); 38634ce62e9eSTejun Heo } 38648b03ae3cSTejun Heo } 38658b03ae3cSTejun Heo 3866e22bee78STejun Heo /* create the initial worker */ 3867f3421797STejun Heo for_each_online_gcwq_cpu(cpu) { 3868e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38694ce62e9eSTejun Heo struct worker_pool *pool; 3870e22bee78STejun Heo 3871477a3c33STejun Heo if (cpu != WORK_CPU_UNBOUND) 3872477a3c33STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 38734ce62e9eSTejun Heo 38744ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 38754ce62e9eSTejun Heo struct worker *worker; 38764ce62e9eSTejun Heo 3877bc2ae0f5STejun Heo worker = create_worker(pool); 3878e22bee78STejun Heo BUG_ON(!worker); 3879e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3880e22bee78STejun Heo start_worker(worker); 3881e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3882e22bee78STejun Heo } 38834ce62e9eSTejun Heo } 3884e22bee78STejun Heo 3885d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 38861aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3887d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3888f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3889f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 389024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 389124d51addSTejun Heo WQ_FREEZABLE, 0); 38921aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3893ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 38946ee0578bSSuresh Siddha return 0; 38951da177e4SLinus Torvalds } 38966ee0578bSSuresh Siddha early_initcall(init_workqueues); 3897