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 61b2eb83d1SLai Jiangshan * assoc_mutex 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 */ 76fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 77f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 78e22bee78STejun Heo 795f7dabfdSLai Jiangshan WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | 80403c821dSTejun Heo WORKER_CPU_INTENSIVE, 81db7bccf4STejun Heo 823270476aSTejun Heo NR_WORKER_POOLS = 2, /* # worker pools per gcwq */ 834ce62e9eSTejun Heo 84c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 85c8e55f36STejun Heo BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, 86c8e55f36STejun Heo BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, 87db7bccf4STejun Heo 88e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 89e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 90e22bee78STejun Heo 913233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 923233cdbdSTejun Heo /* call for help after 10ms 933233cdbdSTejun Heo (min two ticks) */ 94e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 95e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 961da177e4SLinus Torvalds 971da177e4SLinus Torvalds /* 98e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 99e22bee78STejun Heo * all cpus. Give -20. 100e22bee78STejun Heo */ 101e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1023270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 103c8e55f36STejun Heo }; 104c8e55f36STejun Heo 1051da177e4SLinus Torvalds /* 1064690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1074690c4abSTejun Heo * 108e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 109e41e704bSTejun Heo * everyone else. 1104690c4abSTejun Heo * 111e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 112e22bee78STejun Heo * only be modified and accessed from the local cpu. 113e22bee78STejun Heo * 1148b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 1154690c4abSTejun Heo * 116e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 117e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 118e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 119f3421797STejun Heo * If GCWQ_DISASSOCIATED is set, it's identical to L. 120e22bee78STejun Heo * 12173f53c4aSTejun Heo * F: wq->flush_mutex protected. 12273f53c4aSTejun Heo * 1234690c4abSTejun Heo * W: workqueue_lock protected. 1244690c4abSTejun Heo */ 1254690c4abSTejun Heo 1268b03ae3cSTejun Heo struct global_cwq; 127bd7bdd43STejun Heo struct worker_pool; 128c34056a3STejun Heo 129e22bee78STejun Heo /* 130e22bee78STejun Heo * The poor guys doing the actual heavy lifting. All on-duty workers 131e22bee78STejun Heo * are either serving the manager role, on idle list or on busy hash. 132e22bee78STejun Heo */ 133c34056a3STejun Heo struct worker { 134c8e55f36STejun Heo /* on idle list while idle, on busy hash table while busy */ 135c8e55f36STejun Heo union { 136c8e55f36STejun Heo struct list_head entry; /* L: while idle */ 137c8e55f36STejun Heo struct hlist_node hentry; /* L: while busy */ 138c8e55f36STejun Heo }; 139c8e55f36STejun Heo 140c34056a3STejun Heo struct work_struct *current_work; /* L: work being processed */ 1418cca0eeaSTejun Heo struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ 142affee4b2STejun Heo struct list_head scheduled; /* L: scheduled works */ 143c34056a3STejun Heo struct task_struct *task; /* I: worker task */ 144bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 145e22bee78STejun Heo /* 64 bytes boundary on 64bit, 32 on 32bit */ 146e22bee78STejun Heo unsigned long last_active; /* L: last active timestamp */ 147e22bee78STejun Heo unsigned int flags; /* X: flags */ 148c34056a3STejun Heo int id; /* I: worker id */ 14925511a47STejun Heo 15025511a47STejun Heo /* for rebinding worker to CPU */ 15125511a47STejun Heo struct work_struct rebind_work; /* L: for busy worker */ 152c34056a3STejun Heo }; 153c34056a3STejun Heo 154bd7bdd43STejun Heo struct worker_pool { 155bd7bdd43STejun Heo struct global_cwq *gcwq; /* I: the owning gcwq */ 15611ebea50STejun Heo unsigned int flags; /* X: flags */ 157bd7bdd43STejun Heo 158bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 159bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 160ea1abd61SLai Jiangshan 161ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 162bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 163bd7bdd43STejun Heo 164bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 165bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 166bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 167bd7bdd43STejun Heo 168b2eb83d1SLai Jiangshan struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */ 169bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 170bd7bdd43STejun Heo }; 171bd7bdd43STejun Heo 1724690c4abSTejun Heo /* 173e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 174e22bee78STejun Heo * and all works are queued and processed here regardless of their 175e22bee78STejun Heo * target workqueues. 1768b03ae3cSTejun Heo */ 1778b03ae3cSTejun Heo struct global_cwq { 1788b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1798b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 180db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 181c8e55f36STejun Heo 182bd7bdd43STejun Heo /* workers are chained either in busy_hash or pool idle_list */ 183c8e55f36STejun Heo struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; 184c8e55f36STejun Heo /* L: hash of busy workers */ 185c8e55f36STejun Heo 186330dad5bSJoonsoo Kim struct worker_pool pools[NR_WORKER_POOLS]; 187330dad5bSJoonsoo Kim /* normal and highpri pools */ 1888b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1898b03ae3cSTejun Heo 1908b03ae3cSTejun Heo /* 191502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1920f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1930f900049STejun Heo * aligned at two's power of the number of flag bits. 1941da177e4SLinus Torvalds */ 1951da177e4SLinus Torvalds struct cpu_workqueue_struct { 196bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1974690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19873f53c4aSTejun Heo int work_color; /* L: current color */ 19973f53c4aSTejun Heo int flush_color; /* L: flushing color */ 20073f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20173f53c4aSTejun Heo /* L: nr of in_flight works */ 2021e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 203a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2041e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2050f900049STejun Heo }; 2061da177e4SLinus Torvalds 2071da177e4SLinus Torvalds /* 20873f53c4aSTejun Heo * Structure used to wait for workqueue flush. 20973f53c4aSTejun Heo */ 21073f53c4aSTejun Heo struct wq_flusher { 21173f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 21273f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 21373f53c4aSTejun Heo struct completion done; /* flush completion */ 21473f53c4aSTejun Heo }; 2151da177e4SLinus Torvalds 21673f53c4aSTejun Heo /* 217f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 218f2e005aaSTejun Heo * used to determine whether there's something to be done. 219f2e005aaSTejun Heo */ 220f2e005aaSTejun Heo #ifdef CONFIG_SMP 221f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 222f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 223f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 224f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 225f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 2269c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 227f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 228f2e005aaSTejun Heo #else 229f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 230f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 231f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 232f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 233f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 234f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 235f2e005aaSTejun Heo #endif 2361da177e4SLinus Torvalds 2371da177e4SLinus Torvalds /* 2381da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2391da177e4SLinus Torvalds * per-CPU workqueues: 2401da177e4SLinus Torvalds */ 2411da177e4SLinus Torvalds struct workqueue_struct { 2429c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 243bdbc5dd7STejun Heo union { 244bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 245bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 246bdbc5dd7STejun Heo unsigned long v; 247bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2484690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 24973f53c4aSTejun Heo 25073f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 25173f53c4aSTejun Heo int work_color; /* F: current work color */ 25273f53c4aSTejun Heo int flush_color; /* F: current flush color */ 25373f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 25473f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 25573f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 25673f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 25773f53c4aSTejun Heo 258f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 259e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 260e22bee78STejun Heo 2619c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 262dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2634e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2644e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2654e6045f1SJohannes Berg #endif 266b196be89STejun Heo char name[]; /* I: workqueue name */ 2671da177e4SLinus Torvalds }; 2681da177e4SLinus Torvalds 269d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 270d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 271044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2721aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 273044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 274d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 275044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 276f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 277044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 27824d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 279d320c038STejun Heo 28097bd2347STejun Heo #define CREATE_TRACE_POINTS 28197bd2347STejun Heo #include <trace/events/workqueue.h> 28297bd2347STejun Heo 2834ce62e9eSTejun Heo #define for_each_worker_pool(pool, gcwq) \ 2843270476aSTejun Heo for ((pool) = &(gcwq)->pools[0]; \ 2853270476aSTejun Heo (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++) 2864ce62e9eSTejun Heo 287db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 288db7bccf4STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ 289db7bccf4STejun Heo hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) 290db7bccf4STejun Heo 291f3421797STejun Heo static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, 292f3421797STejun Heo unsigned int sw) 293f3421797STejun Heo { 294f3421797STejun Heo if (cpu < nr_cpu_ids) { 295f3421797STejun Heo if (sw & 1) { 296f3421797STejun Heo cpu = cpumask_next(cpu, mask); 297f3421797STejun Heo if (cpu < nr_cpu_ids) 298f3421797STejun Heo return cpu; 299f3421797STejun Heo } 300f3421797STejun Heo if (sw & 2) 301f3421797STejun Heo return WORK_CPU_UNBOUND; 302f3421797STejun Heo } 303f3421797STejun Heo return WORK_CPU_NONE; 304f3421797STejun Heo } 305f3421797STejun Heo 306f3421797STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 307f3421797STejun Heo struct workqueue_struct *wq) 308f3421797STejun Heo { 309f3421797STejun Heo return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 310f3421797STejun Heo } 311f3421797STejun Heo 31209884951STejun Heo /* 31309884951STejun Heo * CPU iterators 31409884951STejun Heo * 31509884951STejun Heo * An extra gcwq is defined for an invalid cpu number 31609884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 31709884951STejun Heo * specific CPU. The following iterators are similar to 31809884951STejun Heo * for_each_*_cpu() iterators but also considers the unbound gcwq. 31909884951STejun Heo * 32009884951STejun Heo * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND 32109884951STejun Heo * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND 32209884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 32309884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 32409884951STejun Heo */ 325f3421797STejun Heo #define for_each_gcwq_cpu(cpu) \ 326f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ 327f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 328f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) 329f3421797STejun Heo 330f3421797STejun Heo #define for_each_online_gcwq_cpu(cpu) \ 331f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ 332f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 333f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) 334f3421797STejun Heo 335f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 336f3421797STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ 337f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 338f3421797STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) 339f3421797STejun Heo 340dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 341dc186ad7SThomas Gleixner 342dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 343dc186ad7SThomas Gleixner 34499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 34599777288SStanislaw Gruszka { 34699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 34799777288SStanislaw Gruszka } 34899777288SStanislaw Gruszka 349dc186ad7SThomas Gleixner /* 350dc186ad7SThomas Gleixner * fixup_init is called when: 351dc186ad7SThomas Gleixner * - an active object is initialized 352dc186ad7SThomas Gleixner */ 353dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 354dc186ad7SThomas Gleixner { 355dc186ad7SThomas Gleixner struct work_struct *work = addr; 356dc186ad7SThomas Gleixner 357dc186ad7SThomas Gleixner switch (state) { 358dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 359dc186ad7SThomas Gleixner cancel_work_sync(work); 360dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 361dc186ad7SThomas Gleixner return 1; 362dc186ad7SThomas Gleixner default: 363dc186ad7SThomas Gleixner return 0; 364dc186ad7SThomas Gleixner } 365dc186ad7SThomas Gleixner } 366dc186ad7SThomas Gleixner 367dc186ad7SThomas Gleixner /* 368dc186ad7SThomas Gleixner * fixup_activate is called when: 369dc186ad7SThomas Gleixner * - an active object is activated 370dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 371dc186ad7SThomas Gleixner */ 372dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 373dc186ad7SThomas Gleixner { 374dc186ad7SThomas Gleixner struct work_struct *work = addr; 375dc186ad7SThomas Gleixner 376dc186ad7SThomas Gleixner switch (state) { 377dc186ad7SThomas Gleixner 378dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 379dc186ad7SThomas Gleixner /* 380dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 381dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 382dc186ad7SThomas Gleixner * is tracked in the object tracker. 383dc186ad7SThomas Gleixner */ 38422df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 385dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 386dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 387dc186ad7SThomas Gleixner return 0; 388dc186ad7SThomas Gleixner } 389dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 390dc186ad7SThomas Gleixner return 0; 391dc186ad7SThomas Gleixner 392dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 393dc186ad7SThomas Gleixner WARN_ON(1); 394dc186ad7SThomas Gleixner 395dc186ad7SThomas Gleixner default: 396dc186ad7SThomas Gleixner return 0; 397dc186ad7SThomas Gleixner } 398dc186ad7SThomas Gleixner } 399dc186ad7SThomas Gleixner 400dc186ad7SThomas Gleixner /* 401dc186ad7SThomas Gleixner * fixup_free is called when: 402dc186ad7SThomas Gleixner * - an active object is freed 403dc186ad7SThomas Gleixner */ 404dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 405dc186ad7SThomas Gleixner { 406dc186ad7SThomas Gleixner struct work_struct *work = addr; 407dc186ad7SThomas Gleixner 408dc186ad7SThomas Gleixner switch (state) { 409dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 410dc186ad7SThomas Gleixner cancel_work_sync(work); 411dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 412dc186ad7SThomas Gleixner return 1; 413dc186ad7SThomas Gleixner default: 414dc186ad7SThomas Gleixner return 0; 415dc186ad7SThomas Gleixner } 416dc186ad7SThomas Gleixner } 417dc186ad7SThomas Gleixner 418dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 419dc186ad7SThomas Gleixner .name = "work_struct", 42099777288SStanislaw Gruszka .debug_hint = work_debug_hint, 421dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 422dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 423dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 424dc186ad7SThomas Gleixner }; 425dc186ad7SThomas Gleixner 426dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 427dc186ad7SThomas Gleixner { 428dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 429dc186ad7SThomas Gleixner } 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 432dc186ad7SThomas Gleixner { 433dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 434dc186ad7SThomas Gleixner } 435dc186ad7SThomas Gleixner 436dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 437dc186ad7SThomas Gleixner { 438dc186ad7SThomas Gleixner if (onstack) 439dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 440dc186ad7SThomas Gleixner else 441dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 442dc186ad7SThomas Gleixner } 443dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 444dc186ad7SThomas Gleixner 445dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 446dc186ad7SThomas Gleixner { 447dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 448dc186ad7SThomas Gleixner } 449dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 450dc186ad7SThomas Gleixner 451dc186ad7SThomas Gleixner #else 452dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 453dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 454dc186ad7SThomas Gleixner #endif 455dc186ad7SThomas Gleixner 45695402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 45795402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4581da177e4SLinus Torvalds static LIST_HEAD(workqueues); 459a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4601da177e4SLinus Torvalds 46114441960SOleg Nesterov /* 462e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 463e22bee78STejun Heo * which is expected to be used frequently by other cpus via 464e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 46514441960SOleg Nesterov */ 4668b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 4674ce62e9eSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]); 468f756d5e2SNathan Lynch 469f3421797STejun Heo /* 470f3421797STejun Heo * Global cpu workqueue and nr_running counter for unbound gcwq. The 471f3421797STejun Heo * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its 472f3421797STejun Heo * workers have WORKER_UNBOUND set. 473f3421797STejun Heo */ 474f3421797STejun Heo static struct global_cwq unbound_global_cwq; 4754ce62e9eSTejun Heo static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = { 4764ce62e9eSTejun Heo [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */ 4774ce62e9eSTejun Heo }; 478f3421797STejun Heo 479c34056a3STejun Heo static int worker_thread(void *__worker); 4801da177e4SLinus Torvalds 4813270476aSTejun Heo static int worker_pool_pri(struct worker_pool *pool) 4823270476aSTejun Heo { 4833270476aSTejun Heo return pool - pool->gcwq->pools; 4843270476aSTejun Heo } 4853270476aSTejun Heo 4868b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 4871da177e4SLinus Torvalds { 488f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 4898b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 490f3421797STejun Heo else 491f3421797STejun Heo return &unbound_global_cwq; 4921da177e4SLinus Torvalds } 4931da177e4SLinus Torvalds 49463d95a91STejun Heo static atomic_t *get_pool_nr_running(struct worker_pool *pool) 495b1f4ec17SOleg Nesterov { 49663d95a91STejun Heo int cpu = pool->gcwq->cpu; 4973270476aSTejun Heo int idx = worker_pool_pri(pool); 49863d95a91STejun Heo 499f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 5004ce62e9eSTejun Heo return &per_cpu(pool_nr_running, cpu)[idx]; 501f3421797STejun Heo else 5024ce62e9eSTejun Heo return &unbound_pool_nr_running[idx]; 503b1f4ec17SOleg Nesterov } 504b1f4ec17SOleg Nesterov 5054690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 5064690c4abSTejun Heo struct workqueue_struct *wq) 507a848e3b6SOleg Nesterov { 508f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 509e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 510bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 511f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 512f3421797STejun Heo return wq->cpu_wq.single; 513f3421797STejun Heo return NULL; 514f3421797STejun Heo } 515a848e3b6SOleg Nesterov 51673f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 51773f53c4aSTejun Heo { 51873f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 51973f53c4aSTejun Heo } 52073f53c4aSTejun Heo 52173f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 52273f53c4aSTejun Heo { 52373f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 52473f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 52573f53c4aSTejun Heo } 52673f53c4aSTejun Heo 52773f53c4aSTejun Heo static int work_next_color(int color) 52873f53c4aSTejun Heo { 52973f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 5301da177e4SLinus Torvalds } 5311da177e4SLinus Torvalds 5324594bf15SDavid Howells /* 533b5490077STejun Heo * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data 534b5490077STejun Heo * contain the pointer to the queued cwq. Once execution starts, the flag 535b5490077STejun Heo * is cleared and the high bits contain OFFQ flags and CPU number. 5367a22ad75STejun Heo * 537bbb68dfaSTejun Heo * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling() 538bbb68dfaSTejun Heo * and clear_work_data() can be used to set the cwq, cpu or clear 539bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 540bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5417a22ad75STejun Heo * 542bbb68dfaSTejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to 543bbb68dfaSTejun Heo * a work. gcwq is available once the work has been queued anywhere after 544bbb68dfaSTejun Heo * initialization until it is sync canceled. cwq is available only while 545bbb68dfaSTejun Heo * the work item is queued. 546bbb68dfaSTejun Heo * 547bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 548bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 549bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 550bbb68dfaSTejun Heo * try to steal the PENDING bit. 5514594bf15SDavid Howells */ 5527a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5537a22ad75STejun Heo unsigned long flags) 5547a22ad75STejun Heo { 5557a22ad75STejun Heo BUG_ON(!work_pending(work)); 5567a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5577a22ad75STejun Heo } 5587a22ad75STejun Heo 5597a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5604690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5614690c4abSTejun Heo unsigned long extra_flags) 562365970a1SDavid Howells { 5637a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 564e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 565365970a1SDavid Howells } 566365970a1SDavid Howells 5678930cabaSTejun Heo static void set_work_cpu_and_clear_pending(struct work_struct *work, 5688930cabaSTejun Heo unsigned int cpu) 5694d707b9fSOleg Nesterov { 57023657bb1STejun Heo /* 57123657bb1STejun Heo * The following wmb is paired with the implied mb in 57223657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 57323657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 57423657bb1STejun Heo * owner. 57523657bb1STejun Heo */ 57623657bb1STejun Heo smp_wmb(); 577b5490077STejun Heo set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0); 5784d707b9fSOleg Nesterov } 5794d707b9fSOleg Nesterov 5807a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 581365970a1SDavid Howells { 58223657bb1STejun Heo smp_wmb(); /* see set_work_cpu_and_clear_pending() */ 5837a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 5847a22ad75STejun Heo } 5857a22ad75STejun Heo 5867a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5877a22ad75STejun Heo { 588e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5897a22ad75STejun Heo 590e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 591e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 592e120153dSTejun Heo else 593e120153dSTejun Heo return NULL; 5947a22ad75STejun Heo } 5957a22ad75STejun Heo 5967a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 5977a22ad75STejun Heo { 598e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5997a22ad75STejun Heo unsigned int cpu; 6007a22ad75STejun Heo 601e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 602e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 603bd7bdd43STejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq; 6047a22ad75STejun Heo 605b5490077STejun Heo cpu = data >> WORK_OFFQ_CPU_SHIFT; 606bdbc5dd7STejun Heo if (cpu == WORK_CPU_NONE) 6077a22ad75STejun Heo return NULL; 6087a22ad75STejun Heo 609f3421797STejun Heo BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); 6107a22ad75STejun Heo return get_gcwq(cpu); 611365970a1SDavid Howells } 612365970a1SDavid Howells 613bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 614bbb68dfaSTejun Heo { 615bbb68dfaSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 616bbb68dfaSTejun Heo unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE; 617bbb68dfaSTejun Heo 618bbb68dfaSTejun Heo set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING, 619bbb68dfaSTejun Heo WORK_STRUCT_PENDING); 620bbb68dfaSTejun Heo } 621bbb68dfaSTejun Heo 622bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 623bbb68dfaSTejun Heo { 624bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 625bbb68dfaSTejun Heo 626bbb68dfaSTejun Heo return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING); 627bbb68dfaSTejun Heo } 628bbb68dfaSTejun Heo 629e22bee78STejun Heo /* 6303270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6313270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 6323270476aSTejun Heo * they're being called with gcwq->lock held. 633e22bee78STejun Heo */ 634e22bee78STejun Heo 63563d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 636649027d7STejun Heo { 6373270476aSTejun Heo return !atomic_read(get_pool_nr_running(pool)); 638649027d7STejun Heo } 639649027d7STejun Heo 640e22bee78STejun Heo /* 641e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 642e22bee78STejun Heo * running workers. 643974271c4STejun Heo * 644974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 645974271c4STejun Heo * function will always return %true for unbound gcwq as long as the 646974271c4STejun Heo * worklist isn't empty. 647e22bee78STejun Heo */ 64863d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 649e22bee78STejun Heo { 65063d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 651e22bee78STejun Heo } 652e22bee78STejun Heo 653e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 65463d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 655e22bee78STejun Heo { 65663d95a91STejun Heo return pool->nr_idle; 657e22bee78STejun Heo } 658e22bee78STejun Heo 659e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 66063d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 661e22bee78STejun Heo { 66263d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 663e22bee78STejun Heo 6643270476aSTejun Heo return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1; 665e22bee78STejun Heo } 666e22bee78STejun Heo 667e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 66863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 669e22bee78STejun Heo { 67063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 671e22bee78STejun Heo } 672e22bee78STejun Heo 673e22bee78STejun Heo /* Do I need to be the manager? */ 67463d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 675e22bee78STejun Heo { 67663d95a91STejun Heo return need_to_create_worker(pool) || 67711ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 678e22bee78STejun Heo } 679e22bee78STejun Heo 680e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 68163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 682e22bee78STejun Heo { 683552a37e9SLai Jiangshan bool managing = pool->flags & POOL_MANAGING_WORKERS; 68463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 68563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 686e22bee78STejun Heo 687ea1abd61SLai Jiangshan /* 688ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 689ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 690ea1abd61SLai Jiangshan */ 691ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 692ea1abd61SLai Jiangshan return false; 693ea1abd61SLai Jiangshan 694e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 695e22bee78STejun Heo } 696e22bee78STejun Heo 697e22bee78STejun Heo /* 698e22bee78STejun Heo * Wake up functions. 699e22bee78STejun Heo */ 700e22bee78STejun Heo 7017e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 70263d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7037e11629dSTejun Heo { 70463d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7057e11629dSTejun Heo return NULL; 7067e11629dSTejun Heo 70763d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7087e11629dSTejun Heo } 7097e11629dSTejun Heo 7107e11629dSTejun Heo /** 7117e11629dSTejun Heo * wake_up_worker - wake up an idle worker 71263d95a91STejun Heo * @pool: worker pool to wake worker from 7137e11629dSTejun Heo * 71463d95a91STejun Heo * Wake up the first idle worker of @pool. 7157e11629dSTejun Heo * 7167e11629dSTejun Heo * CONTEXT: 7177e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 7187e11629dSTejun Heo */ 71963d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7207e11629dSTejun Heo { 72163d95a91STejun Heo struct worker *worker = first_worker(pool); 7227e11629dSTejun Heo 7237e11629dSTejun Heo if (likely(worker)) 7247e11629dSTejun Heo wake_up_process(worker->task); 7257e11629dSTejun Heo } 7267e11629dSTejun Heo 7274690c4abSTejun Heo /** 728e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 729e22bee78STejun Heo * @task: task waking up 730e22bee78STejun Heo * @cpu: CPU @task is waking up to 731e22bee78STejun Heo * 732e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 733e22bee78STejun Heo * being awoken. 734e22bee78STejun Heo * 735e22bee78STejun Heo * CONTEXT: 736e22bee78STejun Heo * spin_lock_irq(rq->lock) 737e22bee78STejun Heo */ 738e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 739e22bee78STejun Heo { 740e22bee78STejun Heo struct worker *worker = kthread_data(task); 741e22bee78STejun Heo 7422d64672eSSteven Rostedt if (!(worker->flags & WORKER_NOT_RUNNING)) 74363d95a91STejun Heo atomic_inc(get_pool_nr_running(worker->pool)); 744e22bee78STejun Heo } 745e22bee78STejun Heo 746e22bee78STejun Heo /** 747e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 748e22bee78STejun Heo * @task: task going to sleep 749e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 750e22bee78STejun Heo * 751e22bee78STejun Heo * This function is called during schedule() when a busy worker is 752e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 753e22bee78STejun Heo * returning pointer to its task. 754e22bee78STejun Heo * 755e22bee78STejun Heo * CONTEXT: 756e22bee78STejun Heo * spin_lock_irq(rq->lock) 757e22bee78STejun Heo * 758e22bee78STejun Heo * RETURNS: 759e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 760e22bee78STejun Heo */ 761e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 762e22bee78STejun Heo unsigned int cpu) 763e22bee78STejun Heo { 764e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 765bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 76663d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 767e22bee78STejun Heo 7682d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 769e22bee78STejun Heo return NULL; 770e22bee78STejun Heo 771e22bee78STejun Heo /* this can only happen on the local cpu */ 772e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 773e22bee78STejun Heo 774e22bee78STejun Heo /* 775e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 776e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 777e22bee78STejun Heo * Please read comment there. 778e22bee78STejun Heo * 779628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 780628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 781628c78e7STejun Heo * disabled, which in turn means that none else could be 782628c78e7STejun Heo * manipulating idle_list, so dereferencing idle_list without gcwq 783628c78e7STejun Heo * lock is safe. 784e22bee78STejun Heo */ 785bd7bdd43STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist)) 78663d95a91STejun Heo to_wakeup = first_worker(pool); 787e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 788e22bee78STejun Heo } 789e22bee78STejun Heo 790e22bee78STejun Heo /** 791e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 792cb444766STejun Heo * @worker: self 793d302f017STejun Heo * @flags: flags to set 794d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 795d302f017STejun Heo * 796e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 797e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 798e22bee78STejun Heo * woken up. 799d302f017STejun Heo * 800cb444766STejun Heo * CONTEXT: 801cb444766STejun Heo * spin_lock_irq(gcwq->lock) 802d302f017STejun Heo */ 803d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 804d302f017STejun Heo bool wakeup) 805d302f017STejun Heo { 806bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 807e22bee78STejun Heo 808cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 809cb444766STejun Heo 810e22bee78STejun Heo /* 811e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 812e22bee78STejun Heo * wake up an idle worker as necessary if requested by 813e22bee78STejun Heo * @wakeup. 814e22bee78STejun Heo */ 815e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 816e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 81763d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 818e22bee78STejun Heo 819e22bee78STejun Heo if (wakeup) { 820e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 821bd7bdd43STejun Heo !list_empty(&pool->worklist)) 82263d95a91STejun Heo wake_up_worker(pool); 823e22bee78STejun Heo } else 824e22bee78STejun Heo atomic_dec(nr_running); 825e22bee78STejun Heo } 826e22bee78STejun Heo 827d302f017STejun Heo worker->flags |= flags; 828d302f017STejun Heo } 829d302f017STejun Heo 830d302f017STejun Heo /** 831e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 832cb444766STejun Heo * @worker: self 833d302f017STejun Heo * @flags: flags to clear 834d302f017STejun Heo * 835e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 836d302f017STejun Heo * 837cb444766STejun Heo * CONTEXT: 838cb444766STejun Heo * spin_lock_irq(gcwq->lock) 839d302f017STejun Heo */ 840d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 841d302f017STejun Heo { 84263d95a91STejun Heo struct worker_pool *pool = worker->pool; 843e22bee78STejun Heo unsigned int oflags = worker->flags; 844e22bee78STejun Heo 845cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 846cb444766STejun Heo 847d302f017STejun Heo worker->flags &= ~flags; 848e22bee78STejun Heo 84942c025f3STejun Heo /* 85042c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 85142c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 85242c025f3STejun Heo * of multiple flags, not a single flag. 85342c025f3STejun Heo */ 854e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 855e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 85663d95a91STejun Heo atomic_inc(get_pool_nr_running(pool)); 857d302f017STejun Heo } 858d302f017STejun Heo 859d302f017STejun Heo /** 860c8e55f36STejun Heo * busy_worker_head - return the busy hash head for a work 861c8e55f36STejun Heo * @gcwq: gcwq of interest 862c8e55f36STejun Heo * @work: work to be hashed 863c8e55f36STejun Heo * 864c8e55f36STejun Heo * Return hash head of @gcwq for @work. 865c8e55f36STejun Heo * 866c8e55f36STejun Heo * CONTEXT: 867c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 868c8e55f36STejun Heo * 869c8e55f36STejun Heo * RETURNS: 870c8e55f36STejun Heo * Pointer to the hash head. 871c8e55f36STejun Heo */ 872c8e55f36STejun Heo static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, 873c8e55f36STejun Heo struct work_struct *work) 874c8e55f36STejun Heo { 875c8e55f36STejun Heo const int base_shift = ilog2(sizeof(struct work_struct)); 876c8e55f36STejun Heo unsigned long v = (unsigned long)work; 877c8e55f36STejun Heo 878c8e55f36STejun Heo /* simple shift and fold hash, do we need something better? */ 879c8e55f36STejun Heo v >>= base_shift; 880c8e55f36STejun Heo v += v >> BUSY_WORKER_HASH_ORDER; 881c8e55f36STejun Heo v &= BUSY_WORKER_HASH_MASK; 882c8e55f36STejun Heo 883c8e55f36STejun Heo return &gcwq->busy_hash[v]; 884c8e55f36STejun Heo } 885c8e55f36STejun Heo 886c8e55f36STejun Heo /** 8878cca0eeaSTejun Heo * __find_worker_executing_work - find worker which is executing a work 8888cca0eeaSTejun Heo * @gcwq: gcwq of interest 8898cca0eeaSTejun Heo * @bwh: hash head as returned by busy_worker_head() 8908cca0eeaSTejun Heo * @work: work to find worker for 8918cca0eeaSTejun Heo * 8928cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. @bwh should be 8938cca0eeaSTejun Heo * the hash head obtained by calling busy_worker_head() with the same 8948cca0eeaSTejun Heo * work. 8958cca0eeaSTejun Heo * 8968cca0eeaSTejun Heo * CONTEXT: 8978cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8988cca0eeaSTejun Heo * 8998cca0eeaSTejun Heo * RETURNS: 9008cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9018cca0eeaSTejun Heo * otherwise. 9028cca0eeaSTejun Heo */ 9038cca0eeaSTejun Heo static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, 9048cca0eeaSTejun Heo struct hlist_head *bwh, 9058cca0eeaSTejun Heo struct work_struct *work) 9068cca0eeaSTejun Heo { 9078cca0eeaSTejun Heo struct worker *worker; 9088cca0eeaSTejun Heo struct hlist_node *tmp; 9098cca0eeaSTejun Heo 9108cca0eeaSTejun Heo hlist_for_each_entry(worker, tmp, bwh, hentry) 9118cca0eeaSTejun Heo if (worker->current_work == work) 9128cca0eeaSTejun Heo return worker; 9138cca0eeaSTejun Heo return NULL; 9148cca0eeaSTejun Heo } 9158cca0eeaSTejun Heo 9168cca0eeaSTejun Heo /** 9178cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 9188cca0eeaSTejun Heo * @gcwq: gcwq of interest 9198cca0eeaSTejun Heo * @work: work to find worker for 9208cca0eeaSTejun Heo * 9218cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. This function is 9228cca0eeaSTejun Heo * identical to __find_worker_executing_work() except that this 9238cca0eeaSTejun Heo * function calculates @bwh itself. 9248cca0eeaSTejun Heo * 9258cca0eeaSTejun Heo * CONTEXT: 9268cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 9278cca0eeaSTejun Heo * 9288cca0eeaSTejun Heo * RETURNS: 9298cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9308cca0eeaSTejun Heo * otherwise. 9318cca0eeaSTejun Heo */ 9328cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 9338cca0eeaSTejun Heo struct work_struct *work) 9348cca0eeaSTejun Heo { 9358cca0eeaSTejun Heo return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), 9368cca0eeaSTejun Heo work); 9378cca0eeaSTejun Heo } 9388cca0eeaSTejun Heo 9398cca0eeaSTejun Heo /** 940bf4ede01STejun Heo * move_linked_works - move linked works to a list 941bf4ede01STejun Heo * @work: start of series of works to be scheduled 942bf4ede01STejun Heo * @head: target list to append @work to 943bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 944bf4ede01STejun Heo * 945bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 946bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 947bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 948bf4ede01STejun Heo * 949bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 950bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 951bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 952bf4ede01STejun Heo * 953bf4ede01STejun Heo * CONTEXT: 954bf4ede01STejun Heo * spin_lock_irq(gcwq->lock). 955bf4ede01STejun Heo */ 956bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 957bf4ede01STejun Heo struct work_struct **nextp) 958bf4ede01STejun Heo { 959bf4ede01STejun Heo struct work_struct *n; 960bf4ede01STejun Heo 961bf4ede01STejun Heo /* 962bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 963bf4ede01STejun Heo * use NULL for list head. 964bf4ede01STejun Heo */ 965bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 966bf4ede01STejun Heo list_move_tail(&work->entry, head); 967bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 968bf4ede01STejun Heo break; 969bf4ede01STejun Heo } 970bf4ede01STejun Heo 971bf4ede01STejun Heo /* 972bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 973bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 974bf4ede01STejun Heo * needs to be updated. 975bf4ede01STejun Heo */ 976bf4ede01STejun Heo if (nextp) 977bf4ede01STejun Heo *nextp = n; 978bf4ede01STejun Heo } 979bf4ede01STejun Heo 9803aa62497SLai Jiangshan static void cwq_activate_delayed_work(struct work_struct *work) 981bf4ede01STejun Heo { 9823aa62497SLai Jiangshan struct cpu_workqueue_struct *cwq = get_work_cwq(work); 983bf4ede01STejun Heo 984bf4ede01STejun Heo trace_workqueue_activate_work(work); 985bf4ede01STejun Heo move_linked_works(work, &cwq->pool->worklist, NULL); 986bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 987bf4ede01STejun Heo cwq->nr_active++; 988bf4ede01STejun Heo } 989bf4ede01STejun Heo 9903aa62497SLai Jiangshan static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 9913aa62497SLai Jiangshan { 9923aa62497SLai Jiangshan struct work_struct *work = list_first_entry(&cwq->delayed_works, 9933aa62497SLai Jiangshan struct work_struct, entry); 9943aa62497SLai Jiangshan 9953aa62497SLai Jiangshan cwq_activate_delayed_work(work); 9963aa62497SLai Jiangshan } 9973aa62497SLai Jiangshan 998bf4ede01STejun Heo /** 999bf4ede01STejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 1000bf4ede01STejun Heo * @cwq: cwq of interest 1001bf4ede01STejun Heo * @color: color of work which left the queue 1002bf4ede01STejun Heo * 1003bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1004bf4ede01STejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 1005bf4ede01STejun Heo * 1006bf4ede01STejun Heo * CONTEXT: 1007bf4ede01STejun Heo * spin_lock_irq(gcwq->lock). 1008bf4ede01STejun Heo */ 1009b3f9f405SLai Jiangshan static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color) 1010bf4ede01STejun Heo { 1011bf4ede01STejun Heo /* ignore uncolored works */ 1012bf4ede01STejun Heo if (color == WORK_NO_COLOR) 1013bf4ede01STejun Heo return; 1014bf4ede01STejun Heo 1015bf4ede01STejun Heo cwq->nr_in_flight[color]--; 1016bf4ede01STejun Heo 1017bf4ede01STejun Heo cwq->nr_active--; 1018bf4ede01STejun Heo if (!list_empty(&cwq->delayed_works)) { 1019bf4ede01STejun Heo /* one down, submit a delayed one */ 1020bf4ede01STejun Heo if (cwq->nr_active < cwq->max_active) 1021bf4ede01STejun Heo cwq_activate_first_delayed(cwq); 1022bf4ede01STejun Heo } 1023bf4ede01STejun Heo 1024bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1025bf4ede01STejun Heo if (likely(cwq->flush_color != color)) 1026bf4ede01STejun Heo return; 1027bf4ede01STejun Heo 1028bf4ede01STejun Heo /* are there still in-flight works? */ 1029bf4ede01STejun Heo if (cwq->nr_in_flight[color]) 1030bf4ede01STejun Heo return; 1031bf4ede01STejun Heo 1032bf4ede01STejun Heo /* this cwq is done, clear flush_color */ 1033bf4ede01STejun Heo cwq->flush_color = -1; 1034bf4ede01STejun Heo 1035bf4ede01STejun Heo /* 1036bf4ede01STejun Heo * If this was the last cwq, wake up the first flusher. It 1037bf4ede01STejun Heo * will handle the rest. 1038bf4ede01STejun Heo */ 1039bf4ede01STejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 1040bf4ede01STejun Heo complete(&cwq->wq->first_flusher->done); 1041bf4ede01STejun Heo } 1042bf4ede01STejun Heo 104336e227d2STejun Heo /** 1044bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 104536e227d2STejun Heo * @work: work item to steal 104636e227d2STejun Heo * @is_dwork: @work is a delayed_work 1047bbb68dfaSTejun Heo * @flags: place to store irq state 104836e227d2STejun Heo * 104936e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 105036e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 105136e227d2STejun Heo * 105236e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 105336e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 105436e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1055bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1056bbb68dfaSTejun Heo * for arbitrarily long 105736e227d2STejun Heo * 1058bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1059e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1060e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1061e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1062bbb68dfaSTejun Heo * 1063bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1064bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1065bbb68dfaSTejun Heo * 1066e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1067bf4ede01STejun Heo */ 1068bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1069bbb68dfaSTejun Heo unsigned long *flags) 1070bf4ede01STejun Heo { 1071bf4ede01STejun Heo struct global_cwq *gcwq; 1072bf4ede01STejun 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); 11103aa62497SLai Jiangshan 11113aa62497SLai Jiangshan /* 11123aa62497SLai Jiangshan * A delayed work item cannot be grabbed directly 11133aa62497SLai Jiangshan * because it might have linked NO_COLOR work items 11143aa62497SLai Jiangshan * which, if left on the delayed_list, will confuse 11153aa62497SLai Jiangshan * cwq->nr_active management later on and cause 11163aa62497SLai Jiangshan * stall. Make sure the work item is activated 11173aa62497SLai Jiangshan * before grabbing. 11183aa62497SLai Jiangshan */ 11193aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 11203aa62497SLai Jiangshan cwq_activate_delayed_work(work); 11213aa62497SLai Jiangshan 1122bf4ede01STejun Heo list_del_init(&work->entry); 1123bf4ede01STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 1124b3f9f405SLai Jiangshan get_work_color(work)); 112536e227d2STejun Heo 1126bbb68dfaSTejun Heo spin_unlock(&gcwq->lock); 112736e227d2STejun Heo return 1; 1128bf4ede01STejun Heo } 1129bf4ede01STejun Heo } 1130bbb68dfaSTejun Heo spin_unlock(&gcwq->lock); 1131bbb68dfaSTejun Heo fail: 1132bbb68dfaSTejun Heo local_irq_restore(*flags); 1133bbb68dfaSTejun Heo if (work_is_canceling(work)) 1134bbb68dfaSTejun Heo return -ENOENT; 1135bbb68dfaSTejun Heo cpu_relax(); 113636e227d2STejun Heo return -EAGAIN; 1137bf4ede01STejun Heo } 1138bf4ede01STejun Heo 1139bf4ede01STejun Heo /** 11407e11629dSTejun Heo * insert_work - insert a work into gcwq 11414690c4abSTejun Heo * @cwq: cwq @work belongs to 11424690c4abSTejun Heo * @work: work to insert 11434690c4abSTejun Heo * @head: insertion point 11444690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11454690c4abSTejun Heo * 11467e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 11477e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 11484690c4abSTejun Heo * 11494690c4abSTejun Heo * CONTEXT: 11508b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 11511da177e4SLinus Torvalds */ 1152b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 11534690c4abSTejun Heo struct work_struct *work, struct list_head *head, 11544690c4abSTejun Heo unsigned int extra_flags) 1155b89deed3SOleg Nesterov { 115663d95a91STejun Heo struct worker_pool *pool = cwq->pool; 1157e1d8aa9fSFrederic Weisbecker 11584690c4abSTejun Heo /* we own @work, set data and link */ 11597a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 11604690c4abSTejun Heo 11616e84d644SOleg Nesterov /* 11626e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 11636e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 11646e84d644SOleg Nesterov */ 11656e84d644SOleg Nesterov smp_wmb(); 11664690c4abSTejun Heo 11671a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 1168e22bee78STejun Heo 1169e22bee78STejun Heo /* 1170e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 1171e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 1172e22bee78STejun Heo * lying around lazily while there are works to be processed. 1173e22bee78STejun Heo */ 1174e22bee78STejun Heo smp_mb(); 1175e22bee78STejun Heo 117663d95a91STejun Heo if (__need_more_worker(pool)) 117763d95a91STejun Heo wake_up_worker(pool); 1178b89deed3SOleg Nesterov } 1179b89deed3SOleg Nesterov 1180c8efcc25STejun Heo /* 1181c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 1182c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 1183c8efcc25STejun Heo * cold paths. 1184c8efcc25STejun Heo */ 1185c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1186c8efcc25STejun Heo { 1187c8efcc25STejun Heo unsigned long flags; 1188c8efcc25STejun Heo unsigned int cpu; 1189c8efcc25STejun Heo 1190c8efcc25STejun Heo for_each_gcwq_cpu(cpu) { 1191c8efcc25STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 1192c8efcc25STejun Heo struct worker *worker; 1193c8efcc25STejun Heo struct hlist_node *pos; 1194c8efcc25STejun Heo int i; 1195c8efcc25STejun Heo 1196c8efcc25STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1197c8efcc25STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 1198c8efcc25STejun Heo if (worker->task != current) 1199c8efcc25STejun Heo continue; 1200c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 1201c8efcc25STejun Heo /* 1202c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 1203c8efcc25STejun Heo * is headed to the same workqueue. 1204c8efcc25STejun Heo */ 1205c8efcc25STejun Heo return worker->current_cwq->wq == wq; 1206c8efcc25STejun Heo } 1207c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 1208c8efcc25STejun Heo } 1209c8efcc25STejun Heo return false; 1210c8efcc25STejun Heo } 1211c8efcc25STejun Heo 12124690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 12131da177e4SLinus Torvalds struct work_struct *work) 12141da177e4SLinus Torvalds { 1215502ca9d8STejun Heo struct global_cwq *gcwq; 1216502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 12171e19ffc6STejun Heo struct list_head *worklist; 12188a2e8e5dSTejun Heo unsigned int work_flags; 1219b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12208930cabaSTejun Heo 12218930cabaSTejun Heo /* 12228930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12238930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12248930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12258930cabaSTejun Heo * happen with IRQ disabled. 12268930cabaSTejun Heo */ 12278930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12281da177e4SLinus Torvalds 1229dc186ad7SThomas Gleixner debug_work_activate(work); 12301e19ffc6STejun Heo 1231c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 12329c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1233c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1234e41e704bSTejun Heo return; 1235e41e704bSTejun Heo 1236c7fc77f7STejun Heo /* determine gcwq to use */ 1237c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1238c7fc77f7STejun Heo struct global_cwq *last_gcwq; 1239c7fc77f7STejun Heo 124057469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1241f3421797STejun Heo cpu = raw_smp_processor_id(); 1242f3421797STejun Heo 124318aa9effSTejun Heo /* 1244dbf2576eSTejun Heo * It's multi cpu. If @work was previously on a different 1245dbf2576eSTejun Heo * cpu, it might still be running there, in which case the 1246dbf2576eSTejun Heo * work needs to be queued on that cpu to guarantee 1247dbf2576eSTejun Heo * non-reentrancy. 124818aa9effSTejun Heo */ 1249502ca9d8STejun Heo gcwq = get_gcwq(cpu); 1250dbf2576eSTejun Heo last_gcwq = get_work_gcwq(work); 1251dbf2576eSTejun Heo 1252dbf2576eSTejun Heo if (last_gcwq && last_gcwq != gcwq) { 125318aa9effSTejun Heo struct worker *worker; 125418aa9effSTejun Heo 12558930cabaSTejun Heo spin_lock(&last_gcwq->lock); 125618aa9effSTejun Heo 125718aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 125818aa9effSTejun Heo 125918aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 126018aa9effSTejun Heo gcwq = last_gcwq; 126118aa9effSTejun Heo else { 126218aa9effSTejun Heo /* meh... not running there, queue here */ 12638930cabaSTejun Heo spin_unlock(&last_gcwq->lock); 12648930cabaSTejun Heo spin_lock(&gcwq->lock); 126518aa9effSTejun Heo } 12668930cabaSTejun Heo } else { 12678930cabaSTejun Heo spin_lock(&gcwq->lock); 12688930cabaSTejun Heo } 1269f3421797STejun Heo } else { 1270f3421797STejun Heo gcwq = get_gcwq(WORK_CPU_UNBOUND); 12718930cabaSTejun Heo spin_lock(&gcwq->lock); 1272502ca9d8STejun Heo } 1273502ca9d8STejun Heo 1274502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 1275502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 1276b75cac93SJoonsoo Kim trace_workqueue_queue_work(req_cpu, cwq, work); 1277502ca9d8STejun Heo 1278f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 12798930cabaSTejun Heo spin_unlock(&gcwq->lock); 1280f5b2552bSDan Carpenter return; 1281f5b2552bSDan Carpenter } 12821e19ffc6STejun Heo 128373f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 12848a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 12851e19ffc6STejun Heo 12861e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1287cdadf009STejun Heo trace_workqueue_activate_work(work); 12881e19ffc6STejun Heo cwq->nr_active++; 12893270476aSTejun Heo worklist = &cwq->pool->worklist; 12908a2e8e5dSTejun Heo } else { 12918a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 12921e19ffc6STejun Heo worklist = &cwq->delayed_works; 12938a2e8e5dSTejun Heo } 12941e19ffc6STejun Heo 12958a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 12961e19ffc6STejun Heo 12978930cabaSTejun Heo spin_unlock(&gcwq->lock); 12981da177e4SLinus Torvalds } 12991da177e4SLinus Torvalds 13000fcb78c2SRolf Eike Beer /** 1301c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1302c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1303c1a220e7SZhang Rui * @wq: workqueue to use 1304c1a220e7SZhang Rui * @work: work to queue 1305c1a220e7SZhang Rui * 1306d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1307c1a220e7SZhang Rui * 1308c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1309c1a220e7SZhang Rui * can't go away. 1310c1a220e7SZhang Rui */ 1311d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1312d4283e93STejun Heo struct work_struct *work) 1313c1a220e7SZhang Rui { 1314d4283e93STejun Heo bool ret = false; 13158930cabaSTejun Heo unsigned long flags; 13168930cabaSTejun Heo 13178930cabaSTejun Heo local_irq_save(flags); 1318c1a220e7SZhang Rui 131922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13204690c4abSTejun Heo __queue_work(cpu, wq, work); 1321d4283e93STejun Heo ret = true; 1322c1a220e7SZhang Rui } 13238930cabaSTejun Heo 13248930cabaSTejun Heo local_irq_restore(flags); 1325c1a220e7SZhang Rui return ret; 1326c1a220e7SZhang Rui } 1327c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1328c1a220e7SZhang Rui 13290a13c00eSTejun Heo /** 13300a13c00eSTejun Heo * queue_work - queue work on a workqueue 13310a13c00eSTejun Heo * @wq: workqueue to use 13320a13c00eSTejun Heo * @work: work to queue 13330a13c00eSTejun Heo * 1334d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 13350a13c00eSTejun Heo * 13360a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 13370a13c00eSTejun Heo * it can be processed by another CPU. 13380a13c00eSTejun Heo */ 1339d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13400a13c00eSTejun Heo { 134157469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13420a13c00eSTejun Heo } 13430a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13440a13c00eSTejun Heo 1345d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13461da177e4SLinus Torvalds { 134752bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13487a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 13491da177e4SLinus Torvalds 1350e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 13511265057fSTejun Heo __queue_work(dwork->cpu, cwq->wq, &dwork->work); 13521da177e4SLinus Torvalds } 1353d8e794dfSTejun Heo EXPORT_SYMBOL_GPL(delayed_work_timer_fn); 13541da177e4SLinus Torvalds 13557beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 135652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13571da177e4SLinus Torvalds { 13587beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13597beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13607beb2edfSTejun Heo unsigned int lcpu; 13611da177e4SLinus Torvalds 13627beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13637beb2edfSTejun Heo timer->data != (unsigned long)dwork); 13647beb2edfSTejun Heo BUG_ON(timer_pending(timer)); 13657beb2edfSTejun Heo BUG_ON(!list_empty(&work->entry)); 13667beb2edfSTejun Heo 13678852aac2STejun Heo /* 13688852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13698852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13708852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13718852aac2STejun Heo * on that there's no such delay when @delay is 0. 13728852aac2STejun Heo */ 13738852aac2STejun Heo if (!delay) { 13748852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13758852aac2STejun Heo return; 13768852aac2STejun Heo } 13778852aac2STejun Heo 13787beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13797beb2edfSTejun Heo 13807beb2edfSTejun Heo /* 13817beb2edfSTejun Heo * This stores cwq for the moment, for the timer_fn. Note that the 13827beb2edfSTejun Heo * work's gcwq is preserved to allow reentrance detection for 13837beb2edfSTejun Heo * delayed works. 13847beb2edfSTejun Heo */ 13857beb2edfSTejun Heo if (!(wq->flags & WQ_UNBOUND)) { 13867beb2edfSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 13877beb2edfSTejun Heo 1388e42986deSJoonsoo Kim /* 1389e42986deSJoonsoo Kim * If we cannot get the last gcwq from @work directly, 1390e42986deSJoonsoo Kim * select the last CPU such that it avoids unnecessarily 1391e42986deSJoonsoo Kim * triggering non-reentrancy check in __queue_work(). 1392e42986deSJoonsoo Kim */ 1393e42986deSJoonsoo Kim lcpu = cpu; 1394e42986deSJoonsoo Kim if (gcwq) 13957beb2edfSTejun Heo lcpu = gcwq->cpu; 1396e42986deSJoonsoo Kim if (lcpu == WORK_CPU_UNBOUND) 13977beb2edfSTejun Heo lcpu = raw_smp_processor_id(); 13987beb2edfSTejun Heo } else { 13997beb2edfSTejun Heo lcpu = WORK_CPU_UNBOUND; 14001da177e4SLinus Torvalds } 14017beb2edfSTejun Heo 14027beb2edfSTejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 14037beb2edfSTejun Heo 14041265057fSTejun Heo dwork->cpu = cpu; 14057beb2edfSTejun Heo timer->expires = jiffies + delay; 14067beb2edfSTejun Heo 14077beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14087beb2edfSTejun Heo add_timer_on(timer, cpu); 14097beb2edfSTejun Heo else 14107beb2edfSTejun Heo add_timer(timer); 14117beb2edfSTejun Heo } 14121da177e4SLinus Torvalds 14130fcb78c2SRolf Eike Beer /** 14140fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14150fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14160fcb78c2SRolf Eike Beer * @wq: workqueue to use 1417af9997e4SRandy Dunlap * @dwork: work to queue 14180fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14190fcb78c2SRolf Eike Beer * 1420715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1421715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1422715f1300STejun Heo * execution. 14230fcb78c2SRolf Eike Beer */ 1424d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 142552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14267a6bc1cdSVenkatesh Pallipadi { 142752bad64dSDavid Howells struct work_struct *work = &dwork->work; 1428d4283e93STejun Heo bool ret = false; 14298930cabaSTejun Heo unsigned long flags; 14308930cabaSTejun Heo 14318930cabaSTejun Heo /* read the comment in __queue_work() */ 14328930cabaSTejun Heo local_irq_save(flags); 14337a6bc1cdSVenkatesh Pallipadi 143422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14357beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1436d4283e93STejun Heo ret = true; 14377a6bc1cdSVenkatesh Pallipadi } 14388930cabaSTejun Heo 14398930cabaSTejun Heo local_irq_restore(flags); 14407a6bc1cdSVenkatesh Pallipadi return ret; 14417a6bc1cdSVenkatesh Pallipadi } 1442ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14431da177e4SLinus Torvalds 1444c8e55f36STejun Heo /** 14450a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 14460a13c00eSTejun Heo * @wq: workqueue to use 14470a13c00eSTejun Heo * @dwork: delayable work to queue 14480a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 14490a13c00eSTejun Heo * 1450715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 14510a13c00eSTejun Heo */ 1452d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 14530a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 14540a13c00eSTejun Heo { 145557469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14560a13c00eSTejun Heo } 14570a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 14580a13c00eSTejun Heo 14590a13c00eSTejun Heo /** 14608376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14618376fe22STejun Heo * @cpu: CPU number to execute work on 14628376fe22STejun Heo * @wq: workqueue to use 14638376fe22STejun Heo * @dwork: work to queue 14648376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14658376fe22STejun Heo * 14668376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14678376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14688376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14698376fe22STejun Heo * current state. 14708376fe22STejun Heo * 14718376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14728376fe22STejun Heo * pending and its timer was modified. 14738376fe22STejun Heo * 1474e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14758376fe22STejun Heo * See try_to_grab_pending() for details. 14768376fe22STejun Heo */ 14778376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14788376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14798376fe22STejun Heo { 14808376fe22STejun Heo unsigned long flags; 14818376fe22STejun Heo int ret; 14828376fe22STejun Heo 14838376fe22STejun Heo do { 14848376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14858376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14868376fe22STejun Heo 14878376fe22STejun Heo if (likely(ret >= 0)) { 14888376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14898376fe22STejun Heo local_irq_restore(flags); 14908376fe22STejun Heo } 14918376fe22STejun Heo 14928376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14938376fe22STejun Heo return ret; 14948376fe22STejun Heo } 14958376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14968376fe22STejun Heo 14978376fe22STejun Heo /** 14988376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14998376fe22STejun Heo * @wq: workqueue to use 15008376fe22STejun Heo * @dwork: work to queue 15018376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15028376fe22STejun Heo * 15038376fe22STejun Heo * mod_delayed_work_on() on local CPU. 15048376fe22STejun Heo */ 15058376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 15068376fe22STejun Heo unsigned long delay) 15078376fe22STejun Heo { 15088376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 15098376fe22STejun Heo } 15108376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 15118376fe22STejun Heo 15128376fe22STejun Heo /** 1513c8e55f36STejun Heo * worker_enter_idle - enter idle state 1514c8e55f36STejun Heo * @worker: worker which is entering idle state 1515c8e55f36STejun Heo * 1516c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1517c8e55f36STejun Heo * necessary. 1518c8e55f36STejun Heo * 1519c8e55f36STejun Heo * LOCKING: 1520c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1521c8e55f36STejun Heo */ 1522c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15231da177e4SLinus Torvalds { 1524bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1525bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1526c8e55f36STejun Heo 1527c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1528c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1529c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1530c8e55f36STejun Heo 1531cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1532cb444766STejun Heo worker->flags |= WORKER_IDLE; 1533bd7bdd43STejun Heo pool->nr_idle++; 1534e22bee78STejun Heo worker->last_active = jiffies; 1535c8e55f36STejun Heo 1536c8e55f36STejun Heo /* idle_list is LIFO */ 1537bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1538db7bccf4STejun Heo 153963d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1540628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1541cb444766STejun Heo 1542544ecf31STejun Heo /* 1543628c78e7STejun Heo * Sanity check nr_running. Because gcwq_unbind_fn() releases 1544628c78e7STejun Heo * gcwq->lock between setting %WORKER_UNBOUND and zapping 1545628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1546628c78e7STejun Heo * unbind is not in progress. 1547544ecf31STejun Heo */ 1548628c78e7STejun Heo WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) && 1549bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 155063d95a91STejun Heo atomic_read(get_pool_nr_running(pool))); 1551c8e55f36STejun Heo } 1552c8e55f36STejun Heo 1553c8e55f36STejun Heo /** 1554c8e55f36STejun Heo * worker_leave_idle - leave idle state 1555c8e55f36STejun Heo * @worker: worker which is leaving idle state 1556c8e55f36STejun Heo * 1557c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1558c8e55f36STejun Heo * 1559c8e55f36STejun Heo * LOCKING: 1560c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1561c8e55f36STejun Heo */ 1562c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1563c8e55f36STejun Heo { 1564bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1565c8e55f36STejun Heo 1566c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1567d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1568bd7bdd43STejun Heo pool->nr_idle--; 1569c8e55f36STejun Heo list_del_init(&worker->entry); 1570c8e55f36STejun Heo } 1571c8e55f36STejun Heo 1572e22bee78STejun Heo /** 1573e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1574e22bee78STejun Heo * @worker: self 1575e22bee78STejun Heo * 1576e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1577e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1578e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1579e22bee78STejun Heo * guaranteed to execute on the cpu. 1580e22bee78STejun Heo * 1581e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1582e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1583e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1584e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1585e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1586e22bee78STejun Heo * [dis]associated in the meantime. 1587e22bee78STejun Heo * 1588f2d5a0eeSTejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies the 1589f2d5a0eeSTejun Heo * binding against %GCWQ_DISASSOCIATED which is set during 1590f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1591f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1592f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1593e22bee78STejun Heo * 1594e22bee78STejun Heo * CONTEXT: 1595e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1596e22bee78STejun Heo * held. 1597e22bee78STejun Heo * 1598e22bee78STejun Heo * RETURNS: 1599e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1600e22bee78STejun Heo * bound), %false if offline. 1601e22bee78STejun Heo */ 1602e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1603972fa1c5SNamhyung Kim __acquires(&gcwq->lock) 1604e22bee78STejun Heo { 1605bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1606e22bee78STejun Heo struct task_struct *task = worker->task; 1607e22bee78STejun Heo 1608e22bee78STejun Heo while (true) { 1609e22bee78STejun Heo /* 1610e22bee78STejun Heo * The following call may fail, succeed or succeed 1611e22bee78STejun Heo * without actually migrating the task to the cpu if 1612e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1613e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1614e22bee78STejun Heo */ 1615f3421797STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) 1616e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1617e22bee78STejun Heo 1618e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1619e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1620e22bee78STejun Heo return false; 1621e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1622e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1623e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1624e22bee78STejun Heo return true; 1625e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1626e22bee78STejun Heo 16275035b20fSTejun Heo /* 16285035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16295035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16305035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16315035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16325035b20fSTejun Heo */ 1633e22bee78STejun Heo cpu_relax(); 16345035b20fSTejun Heo cond_resched(); 1635e22bee78STejun Heo } 1636e22bee78STejun Heo } 1637e22bee78STejun Heo 1638e22bee78STejun Heo /* 1639ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 16405f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 164125511a47STejun Heo */ 164225511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 164325511a47STejun Heo { 164425511a47STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 164525511a47STejun Heo 16465f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 16475f7dabfdSLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 16485f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 164925511a47STejun Heo 1650ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1651ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1652ea1abd61SLai Jiangshan spin_unlock_irq(&gcwq->lock); 165325511a47STejun Heo } 165425511a47STejun Heo 165525511a47STejun Heo /* 165625511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1657403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1658403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1659403c821dSTejun Heo * executed twice without intervening cpu down. 1660e22bee78STejun Heo */ 166125511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1662e22bee78STejun Heo { 1663e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1664bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1665e22bee78STejun Heo 1666eab6d828SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 1667eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1668e22bee78STejun Heo 1669e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1670e22bee78STejun Heo } 1671e22bee78STejun Heo 167225511a47STejun Heo /** 167325511a47STejun Heo * rebind_workers - rebind all workers of a gcwq to the associated CPU 167425511a47STejun Heo * @gcwq: gcwq of interest 167525511a47STejun Heo * 167625511a47STejun Heo * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding 167725511a47STejun Heo * is different for idle and busy ones. 167825511a47STejun Heo * 1679ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1680ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1681ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1682ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 168325511a47STejun Heo * 1684ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1685ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1686ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1687ea1abd61SLai Jiangshan * rebind. 168825511a47STejun Heo * 1689ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1690ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1691ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1692ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 169325511a47STejun Heo */ 169425511a47STejun Heo static void rebind_workers(struct global_cwq *gcwq) 169525511a47STejun Heo { 169625511a47STejun Heo struct worker_pool *pool; 1697ea1abd61SLai Jiangshan struct worker *worker, *n; 169825511a47STejun Heo struct hlist_node *pos; 169925511a47STejun Heo int i; 170025511a47STejun Heo 170125511a47STejun Heo lockdep_assert_held(&gcwq->lock); 170225511a47STejun Heo 170325511a47STejun Heo for_each_worker_pool(pool, gcwq) 1704b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 170525511a47STejun Heo 17065f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 170725511a47STejun Heo for_each_worker_pool(pool, gcwq) { 1708ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1709ea1abd61SLai Jiangshan /* 1710ea1abd61SLai Jiangshan * idle workers should be off @pool->idle_list 1711ea1abd61SLai Jiangshan * until rebind is complete to avoid receiving 1712ea1abd61SLai Jiangshan * premature local wake-ups. 1713ea1abd61SLai Jiangshan */ 1714ea1abd61SLai Jiangshan list_del_init(&worker->entry); 171525511a47STejun Heo 171625511a47STejun Heo /* 17175f7dabfdSLai Jiangshan * worker_thread() will see the above dequeuing 17185f7dabfdSLai Jiangshan * and call idle_worker_rebind(). 171925511a47STejun Heo */ 172025511a47STejun Heo wake_up_process(worker->task); 172125511a47STejun Heo } 172225511a47STejun Heo } 172325511a47STejun Heo 1724ea1abd61SLai Jiangshan /* rebind busy workers */ 172525511a47STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 172625511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1727e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 172825511a47STejun Heo 172925511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 173025511a47STejun Heo work_data_bits(rebind_work))) 173125511a47STejun Heo continue; 173225511a47STejun Heo 173325511a47STejun Heo debug_work_activate(rebind_work); 173490beca5dSTejun Heo 173590beca5dSTejun Heo /* 1736e2b6a6d5SJoonsoo Kim * wq doesn't really matter but let's keep @worker->pool 1737e2b6a6d5SJoonsoo Kim * and @cwq->pool consistent for sanity. 173890beca5dSTejun Heo */ 1739e2b6a6d5SJoonsoo Kim if (worker_pool_pri(worker->pool)) 1740e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1741e2b6a6d5SJoonsoo Kim else 1742e2b6a6d5SJoonsoo Kim wq = system_wq; 1743ec58815aSTejun Heo 1744e2b6a6d5SJoonsoo Kim insert_work(get_cwq(gcwq->cpu, wq), rebind_work, 174525511a47STejun Heo worker->scheduled.next, 174625511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1747ec58815aSTejun Heo } 174825511a47STejun Heo } 174925511a47STejun Heo 1750c34056a3STejun Heo static struct worker *alloc_worker(void) 1751c34056a3STejun Heo { 1752c34056a3STejun Heo struct worker *worker; 1753c34056a3STejun Heo 1754c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1755c8e55f36STejun Heo if (worker) { 1756c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1757affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 175825511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1759e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1760e22bee78STejun Heo worker->flags = WORKER_PREP; 1761c8e55f36STejun Heo } 1762c34056a3STejun Heo return worker; 1763c34056a3STejun Heo } 1764c34056a3STejun Heo 1765c34056a3STejun Heo /** 1766c34056a3STejun Heo * create_worker - create a new workqueue worker 176763d95a91STejun Heo * @pool: pool the new worker will belong to 1768c34056a3STejun Heo * 176963d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1770c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1771c34056a3STejun Heo * destroy_worker(). 1772c34056a3STejun Heo * 1773c34056a3STejun Heo * CONTEXT: 1774c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1775c34056a3STejun Heo * 1776c34056a3STejun Heo * RETURNS: 1777c34056a3STejun Heo * Pointer to the newly created worker. 1778c34056a3STejun Heo */ 1779bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1780c34056a3STejun Heo { 178163d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 17823270476aSTejun Heo const char *pri = worker_pool_pri(pool) ? "H" : ""; 1783c34056a3STejun Heo struct worker *worker = NULL; 1784f3421797STejun Heo int id = -1; 1785c34056a3STejun Heo 17868b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1787bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 17888b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1789bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1790c34056a3STejun Heo goto fail; 17918b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1792c34056a3STejun Heo } 17938b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1794c34056a3STejun Heo 1795c34056a3STejun Heo worker = alloc_worker(); 1796c34056a3STejun Heo if (!worker) 1797c34056a3STejun Heo goto fail; 1798c34056a3STejun Heo 1799bd7bdd43STejun Heo worker->pool = pool; 1800c34056a3STejun Heo worker->id = id; 1801c34056a3STejun Heo 1802bc2ae0f5STejun Heo if (gcwq->cpu != WORK_CPU_UNBOUND) 180394dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 18043270476aSTejun Heo worker, cpu_to_node(gcwq->cpu), 18053270476aSTejun Heo "kworker/%u:%d%s", gcwq->cpu, id, pri); 1806f3421797STejun Heo else 1807f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 18083270476aSTejun Heo "kworker/u:%d%s", id, pri); 1809c34056a3STejun Heo if (IS_ERR(worker->task)) 1810c34056a3STejun Heo goto fail; 1811c34056a3STejun Heo 18123270476aSTejun Heo if (worker_pool_pri(pool)) 18133270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 18143270476aSTejun Heo 1815db7bccf4STejun Heo /* 1816bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 1817bc2ae0f5STejun Heo * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the 1818bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1819bc2ae0f5STejun Heo * above the flag definition for details. 1820bc2ae0f5STejun Heo * 1821bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1822bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1823db7bccf4STejun Heo */ 1824bc2ae0f5STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) { 18258b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1826bc2ae0f5STejun Heo } else { 1827db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1828f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1829f3421797STejun Heo } 1830c34056a3STejun Heo 1831c34056a3STejun Heo return worker; 1832c34056a3STejun Heo fail: 1833c34056a3STejun Heo if (id >= 0) { 18348b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1835bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 18368b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1837c34056a3STejun Heo } 1838c34056a3STejun Heo kfree(worker); 1839c34056a3STejun Heo return NULL; 1840c34056a3STejun Heo } 1841c34056a3STejun Heo 1842c34056a3STejun Heo /** 1843c34056a3STejun Heo * start_worker - start a newly created worker 1844c34056a3STejun Heo * @worker: worker to start 1845c34056a3STejun Heo * 1846c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1847c34056a3STejun Heo * 1848c34056a3STejun Heo * CONTEXT: 18498b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1850c34056a3STejun Heo */ 1851c34056a3STejun Heo static void start_worker(struct worker *worker) 1852c34056a3STejun Heo { 1853cb444766STejun Heo worker->flags |= WORKER_STARTED; 1854bd7bdd43STejun Heo worker->pool->nr_workers++; 1855c8e55f36STejun Heo worker_enter_idle(worker); 1856c34056a3STejun Heo wake_up_process(worker->task); 1857c34056a3STejun Heo } 1858c34056a3STejun Heo 1859c34056a3STejun Heo /** 1860c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1861c34056a3STejun Heo * @worker: worker to be destroyed 1862c34056a3STejun Heo * 1863c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1864c8e55f36STejun Heo * 1865c8e55f36STejun Heo * CONTEXT: 1866c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1867c34056a3STejun Heo */ 1868c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1869c34056a3STejun Heo { 1870bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1871bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1872c34056a3STejun Heo int id = worker->id; 1873c34056a3STejun Heo 1874c34056a3STejun Heo /* sanity check frenzy */ 1875c34056a3STejun Heo BUG_ON(worker->current_work); 1876affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1877c34056a3STejun Heo 1878c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1879bd7bdd43STejun Heo pool->nr_workers--; 1880c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1881bd7bdd43STejun Heo pool->nr_idle--; 1882c8e55f36STejun Heo 1883c8e55f36STejun Heo list_del_init(&worker->entry); 1884cb444766STejun Heo worker->flags |= WORKER_DIE; 1885c8e55f36STejun Heo 1886c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1887c8e55f36STejun Heo 1888c34056a3STejun Heo kthread_stop(worker->task); 1889c34056a3STejun Heo kfree(worker); 1890c34056a3STejun Heo 18918b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1892bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1893c34056a3STejun Heo } 1894c34056a3STejun Heo 189563d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1896e22bee78STejun Heo { 189763d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 189863d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1899e22bee78STejun Heo 1900e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1901e22bee78STejun Heo 190263d95a91STejun Heo if (too_many_workers(pool)) { 1903e22bee78STejun Heo struct worker *worker; 1904e22bee78STejun Heo unsigned long expires; 1905e22bee78STejun Heo 1906e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 190763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1908e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1909e22bee78STejun Heo 1910e22bee78STejun Heo if (time_before(jiffies, expires)) 191163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1912e22bee78STejun Heo else { 1913e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 191411ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 191563d95a91STejun Heo wake_up_worker(pool); 1916e22bee78STejun Heo } 1917e22bee78STejun Heo } 1918e22bee78STejun Heo 1919e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1920e22bee78STejun Heo } 1921e22bee78STejun Heo 1922e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1923e22bee78STejun Heo { 1924e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1925e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1926f3421797STejun Heo unsigned int cpu; 1927e22bee78STejun Heo 1928e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1929e22bee78STejun Heo return false; 1930e22bee78STejun Heo 1931e22bee78STejun Heo /* mayday mayday mayday */ 1932bd7bdd43STejun Heo cpu = cwq->pool->gcwq->cpu; 1933f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1934f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1935f3421797STejun Heo cpu = 0; 1936f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1937e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1938e22bee78STejun Heo return true; 1939e22bee78STejun Heo } 1940e22bee78STejun Heo 194163d95a91STejun Heo static void gcwq_mayday_timeout(unsigned long __pool) 1942e22bee78STejun Heo { 194363d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 194463d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1945e22bee78STejun Heo struct work_struct *work; 1946e22bee78STejun Heo 1947e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1948e22bee78STejun Heo 194963d95a91STejun Heo if (need_to_create_worker(pool)) { 1950e22bee78STejun Heo /* 1951e22bee78STejun Heo * We've been trying to create a new worker but 1952e22bee78STejun Heo * haven't been successful. We might be hitting an 1953e22bee78STejun Heo * allocation deadlock. Send distress signals to 1954e22bee78STejun Heo * rescuers. 1955e22bee78STejun Heo */ 195663d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1957e22bee78STejun Heo send_mayday(work); 1958e22bee78STejun Heo } 1959e22bee78STejun Heo 1960e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1961e22bee78STejun Heo 196263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1963e22bee78STejun Heo } 1964e22bee78STejun Heo 1965e22bee78STejun Heo /** 1966e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 196763d95a91STejun Heo * @pool: pool to create a new worker for 1968e22bee78STejun Heo * 196963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1970e22bee78STejun Heo * have at least one idle worker on return from this function. If 1971e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 197263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1973e22bee78STejun Heo * possible allocation deadlock. 1974e22bee78STejun Heo * 1975e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1976e22bee78STejun Heo * may_start_working() true. 1977e22bee78STejun Heo * 1978e22bee78STejun Heo * LOCKING: 1979e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1980e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1981e22bee78STejun Heo * manager. 1982e22bee78STejun Heo * 1983e22bee78STejun Heo * RETURNS: 1984e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1985e22bee78STejun Heo * otherwise. 1986e22bee78STejun Heo */ 198763d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 198806bd6ebfSNamhyung Kim __releases(&gcwq->lock) 198906bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 1990e22bee78STejun Heo { 199163d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 199263d95a91STejun Heo 199363d95a91STejun Heo if (!need_to_create_worker(pool)) 1994e22bee78STejun Heo return false; 1995e22bee78STejun Heo restart: 19969f9c2364STejun Heo spin_unlock_irq(&gcwq->lock); 19979f9c2364STejun Heo 1998e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 199963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2000e22bee78STejun Heo 2001e22bee78STejun Heo while (true) { 2002e22bee78STejun Heo struct worker *worker; 2003e22bee78STejun Heo 2004bc2ae0f5STejun Heo worker = create_worker(pool); 2005e22bee78STejun Heo if (worker) { 200663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2007e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 2008e22bee78STejun Heo start_worker(worker); 200963d95a91STejun Heo BUG_ON(need_to_create_worker(pool)); 2010e22bee78STejun Heo return true; 2011e22bee78STejun Heo } 2012e22bee78STejun Heo 201363d95a91STejun Heo if (!need_to_create_worker(pool)) 2014e22bee78STejun Heo break; 2015e22bee78STejun Heo 2016e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2017e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 20189f9c2364STejun Heo 201963d95a91STejun Heo if (!need_to_create_worker(pool)) 2020e22bee78STejun Heo break; 2021e22bee78STejun Heo } 2022e22bee78STejun Heo 202363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2024e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 202563d95a91STejun Heo if (need_to_create_worker(pool)) 2026e22bee78STejun Heo goto restart; 2027e22bee78STejun Heo return true; 2028e22bee78STejun Heo } 2029e22bee78STejun Heo 2030e22bee78STejun Heo /** 2031e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 203263d95a91STejun Heo * @pool: pool to destroy workers for 2033e22bee78STejun Heo * 203463d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 2035e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 2036e22bee78STejun Heo * 2037e22bee78STejun Heo * LOCKING: 2038e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2039e22bee78STejun Heo * multiple times. Called only from manager. 2040e22bee78STejun Heo * 2041e22bee78STejun Heo * RETURNS: 2042e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 2043e22bee78STejun Heo * otherwise. 2044e22bee78STejun Heo */ 204563d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2046e22bee78STejun Heo { 2047e22bee78STejun Heo bool ret = false; 2048e22bee78STejun Heo 204963d95a91STejun Heo while (too_many_workers(pool)) { 2050e22bee78STejun Heo struct worker *worker; 2051e22bee78STejun Heo unsigned long expires; 2052e22bee78STejun Heo 205363d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2054e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2055e22bee78STejun Heo 2056e22bee78STejun Heo if (time_before(jiffies, expires)) { 205763d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2058e22bee78STejun Heo break; 2059e22bee78STejun Heo } 2060e22bee78STejun Heo 2061e22bee78STejun Heo destroy_worker(worker); 2062e22bee78STejun Heo ret = true; 2063e22bee78STejun Heo } 2064e22bee78STejun Heo 2065e22bee78STejun Heo return ret; 2066e22bee78STejun Heo } 2067e22bee78STejun Heo 2068e22bee78STejun Heo /** 2069e22bee78STejun Heo * manage_workers - manage worker pool 2070e22bee78STejun Heo * @worker: self 2071e22bee78STejun Heo * 2072e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 2073e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2074e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 2075e22bee78STejun Heo * 2076e22bee78STejun Heo * The caller can safely start processing works on false return. On 2077e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2078e22bee78STejun Heo * and may_start_working() is true. 2079e22bee78STejun Heo * 2080e22bee78STejun Heo * CONTEXT: 2081e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2082e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2083e22bee78STejun Heo * 2084e22bee78STejun Heo * RETURNS: 2085e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 2086e22bee78STejun Heo * some action was taken. 2087e22bee78STejun Heo */ 2088e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2089e22bee78STejun Heo { 209063d95a91STejun Heo struct worker_pool *pool = worker->pool; 2091e22bee78STejun Heo bool ret = false; 2092e22bee78STejun Heo 2093ee378aa4SLai Jiangshan if (pool->flags & POOL_MANAGING_WORKERS) 2094e22bee78STejun Heo return ret; 2095e22bee78STejun Heo 2096552a37e9SLai Jiangshan pool->flags |= POOL_MANAGING_WORKERS; 2097ee378aa4SLai Jiangshan 2098ee378aa4SLai Jiangshan /* 2099ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2100ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 2101ee378aa4SLai Jiangshan * grab %POOL_MANAGING_WORKERS to achieve this because that can 2102ee378aa4SLai Jiangshan * lead to idle worker depletion (all become busy thinking someone 2103ee378aa4SLai Jiangshan * else is managing) which in turn can result in deadlock under 2104b2eb83d1SLai Jiangshan * extreme circumstances. Use @pool->assoc_mutex to synchronize 2105ee378aa4SLai Jiangshan * manager against CPU hotplug. 2106ee378aa4SLai Jiangshan * 2107b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2108ee378aa4SLai Jiangshan * progress. trylock first without dropping @gcwq->lock. 2109ee378aa4SLai Jiangshan */ 2110b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2111ee378aa4SLai Jiangshan spin_unlock_irq(&pool->gcwq->lock); 2112b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2113ee378aa4SLai Jiangshan /* 2114ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2115b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2116ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2117ee378aa4SLai Jiangshan * @gcwq's state and ours could have deviated. 2118ee378aa4SLai Jiangshan * 2119b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2120ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2121ee378aa4SLai Jiangshan * on @gcwq's current state. Try it and adjust 2122ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2123ee378aa4SLai Jiangshan */ 2124ee378aa4SLai Jiangshan if (worker_maybe_bind_and_lock(worker)) 2125ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2126ee378aa4SLai Jiangshan else 2127ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2128ee378aa4SLai Jiangshan 2129ee378aa4SLai Jiangshan ret = true; 2130ee378aa4SLai Jiangshan } 2131ee378aa4SLai Jiangshan 213211ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2133e22bee78STejun Heo 2134e22bee78STejun Heo /* 2135e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2136e22bee78STejun Heo * on return. 2137e22bee78STejun Heo */ 213863d95a91STejun Heo ret |= maybe_destroy_workers(pool); 213963d95a91STejun Heo ret |= maybe_create_worker(pool); 2140e22bee78STejun Heo 2141552a37e9SLai Jiangshan pool->flags &= ~POOL_MANAGING_WORKERS; 2142b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 2143e22bee78STejun Heo return ret; 2144e22bee78STejun Heo } 2145e22bee78STejun Heo 2146a62428c0STejun Heo /** 2147a62428c0STejun Heo * process_one_work - process single work 2148c34056a3STejun Heo * @worker: self 2149a62428c0STejun Heo * @work: work to process 2150a62428c0STejun Heo * 2151a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2152a62428c0STejun Heo * process a single work including synchronization against and 2153a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2154a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2155a62428c0STejun Heo * call this function to process a work. 2156a62428c0STejun Heo * 2157a62428c0STejun Heo * CONTEXT: 21588b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 2159a62428c0STejun Heo */ 2160c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 216106bd6ebfSNamhyung Kim __releases(&gcwq->lock) 216206bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 21631da177e4SLinus Torvalds { 21647e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 2165bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2166bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2167c8e55f36STejun Heo struct hlist_head *bwh = busy_worker_head(gcwq, work); 2168fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 21696bb49e59SDavid Howells work_func_t f = work->func; 217073f53c4aSTejun Heo int work_color; 21717e11629dSTejun Heo struct worker *collision; 21724e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21734e6045f1SJohannes Berg /* 2174a62428c0STejun Heo * It is permissible to free the struct work_struct from 2175a62428c0STejun Heo * inside the function that is called from it, this we need to 2176a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2177a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2178a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21794e6045f1SJohannes Berg */ 21804d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21814d82a1deSPeter Zijlstra 21824d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21834e6045f1SJohannes Berg #endif 21846fec10a1STejun Heo /* 21856fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21866fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 21876fec10a1STejun Heo * unbound or a disassociated gcwq. 21886fec10a1STejun Heo */ 21895f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 21906fec10a1STejun Heo !(gcwq->flags & GCWQ_DISASSOCIATED) && 219125511a47STejun Heo raw_smp_processor_id() != gcwq->cpu); 219225511a47STejun Heo 21937e11629dSTejun Heo /* 21947e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21957e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21967e11629dSTejun Heo * already processing the work. If so, defer the work to the 21977e11629dSTejun Heo * currently executing one. 21987e11629dSTejun Heo */ 21997e11629dSTejun Heo collision = __find_worker_executing_work(gcwq, bwh, work); 22007e11629dSTejun Heo if (unlikely(collision)) { 22017e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 22027e11629dSTejun Heo return; 22037e11629dSTejun Heo } 22041da177e4SLinus Torvalds 22058930cabaSTejun Heo /* claim and dequeue */ 22061da177e4SLinus Torvalds debug_work_deactivate(work); 2207c8e55f36STejun Heo hlist_add_head(&worker->hentry, bwh); 2208c34056a3STejun Heo worker->current_work = work; 22098cca0eeaSTejun Heo worker->current_cwq = cwq; 221073f53c4aSTejun Heo work_color = get_work_color(work); 22117a22ad75STejun Heo 2212a62428c0STejun Heo list_del_init(&work->entry); 2213a62428c0STejun Heo 2214649027d7STejun Heo /* 2215fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2216fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2217fb0e7bebSTejun Heo */ 2218fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2219fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2220fb0e7bebSTejun Heo 2221974271c4STejun Heo /* 2222974271c4STejun Heo * Unbound gcwq isn't concurrency managed and work items should be 2223974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2224974271c4STejun Heo */ 222563d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 222663d95a91STejun Heo wake_up_worker(pool); 2227974271c4STejun Heo 22288930cabaSTejun Heo /* 222923657bb1STejun Heo * Record the last CPU and clear PENDING which should be the last 223023657bb1STejun Heo * update to @work. Also, do this inside @gcwq->lock so that 223123657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 223223657bb1STejun Heo * disabled. 22338930cabaSTejun Heo */ 22348930cabaSTejun Heo set_work_cpu_and_clear_pending(work, gcwq->cpu); 22351da177e4SLinus Torvalds 22361da177e4SLinus Torvalds spin_unlock_irq(&gcwq->lock); 2237365970a1SDavid Howells 2238e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 22393295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2240e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 224165f27f38SDavid Howells f(work); 2242e36c886aSArjan van de Ven /* 2243e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2244e36c886aSArjan van de Ven * point will only record its address. 2245e36c886aSArjan van de Ven */ 2246e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22473295f0efSIngo Molnar lock_map_release(&lockdep_map); 22483295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 22491da177e4SLinus Torvalds 2250d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2251044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2252044c782cSValentin Ilie " last function: %pf\n", 2253044c782cSValentin Ilie current->comm, preempt_count(), task_pid_nr(current), f); 2254d5abe669SPeter Zijlstra debug_show_held_locks(current); 2255d5abe669SPeter Zijlstra dump_stack(); 2256d5abe669SPeter Zijlstra } 2257d5abe669SPeter Zijlstra 22588b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2259a62428c0STejun Heo 2260fb0e7bebSTejun Heo /* clear cpu intensive status */ 2261fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2262fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2263fb0e7bebSTejun Heo 2264a62428c0STejun Heo /* we're done with it, release */ 2265c8e55f36STejun Heo hlist_del_init(&worker->hentry); 2266c34056a3STejun Heo worker->current_work = NULL; 22678cca0eeaSTejun Heo worker->current_cwq = NULL; 2268b3f9f405SLai Jiangshan cwq_dec_nr_in_flight(cwq, work_color); 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds 2271affee4b2STejun Heo /** 2272affee4b2STejun Heo * process_scheduled_works - process scheduled works 2273affee4b2STejun Heo * @worker: self 2274affee4b2STejun Heo * 2275affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2276affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2277affee4b2STejun Heo * fetches a work from the top and executes it. 2278affee4b2STejun Heo * 2279affee4b2STejun Heo * CONTEXT: 22808b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2281affee4b2STejun Heo * multiple times. 2282affee4b2STejun Heo */ 2283affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22841da177e4SLinus Torvalds { 2285affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2286affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2287a62428c0STejun Heo struct work_struct, entry); 2288c34056a3STejun Heo process_one_work(worker, work); 2289a62428c0STejun Heo } 22901da177e4SLinus Torvalds } 22911da177e4SLinus Torvalds 22924690c4abSTejun Heo /** 22934690c4abSTejun Heo * worker_thread - the worker thread function 2294c34056a3STejun Heo * @__worker: self 22954690c4abSTejun Heo * 2296e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 2297e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 2298e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2299e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2300e22bee78STejun Heo * rescuer_thread(). 23014690c4abSTejun Heo */ 2302c34056a3STejun Heo static int worker_thread(void *__worker) 23031da177e4SLinus Torvalds { 2304c34056a3STejun Heo struct worker *worker = __worker; 2305bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2306bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 23071da177e4SLinus Torvalds 2308e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2309e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2310c8e55f36STejun Heo woke_up: 23118b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2312affee4b2STejun Heo 23135f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 23145f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2315c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 231625511a47STejun Heo 23175f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 231825511a47STejun Heo if (worker->flags & WORKER_DIE) { 2319e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2320c8e55f36STejun Heo return 0; 2321c8e55f36STejun Heo } 2322c8e55f36STejun Heo 23235f7dabfdSLai Jiangshan /* otherwise, rebind */ 232425511a47STejun Heo idle_worker_rebind(worker); 232525511a47STejun Heo goto woke_up; 232625511a47STejun Heo } 232725511a47STejun Heo 2328c8e55f36STejun Heo worker_leave_idle(worker); 2329db7bccf4STejun Heo recheck: 2330e22bee78STejun Heo /* no more worker necessary? */ 233163d95a91STejun Heo if (!need_more_worker(pool)) 2332e22bee78STejun Heo goto sleep; 2333e22bee78STejun Heo 2334e22bee78STejun Heo /* do we need to manage? */ 233563d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2336e22bee78STejun Heo goto recheck; 2337e22bee78STejun Heo 2338c8e55f36STejun Heo /* 2339c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2340c8e55f36STejun Heo * preparing to process a work or actually processing it. 2341c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2342c8e55f36STejun Heo */ 2343c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 2344c8e55f36STejun Heo 2345e22bee78STejun Heo /* 2346e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2347e22bee78STejun Heo * at least one idle worker or that someone else has already 2348e22bee78STejun Heo * assumed the manager role. 2349e22bee78STejun Heo */ 2350e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2351e22bee78STejun Heo 2352e22bee78STejun Heo do { 2353affee4b2STejun Heo struct work_struct *work = 2354bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2355affee4b2STejun Heo struct work_struct, entry); 2356affee4b2STejun Heo 2357c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2358affee4b2STejun Heo /* optimization path, not strictly necessary */ 2359affee4b2STejun Heo process_one_work(worker, work); 2360affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2361affee4b2STejun Heo process_scheduled_works(worker); 2362affee4b2STejun Heo } else { 2363c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2364affee4b2STejun Heo process_scheduled_works(worker); 2365affee4b2STejun Heo } 236663d95a91STejun Heo } while (keep_working(pool)); 2367affee4b2STejun Heo 2368e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2369d313dd85STejun Heo sleep: 237063d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2371e22bee78STejun Heo goto recheck; 2372d313dd85STejun Heo 2373c8e55f36STejun Heo /* 2374e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 2375e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 2376e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 2377e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 2378e22bee78STejun Heo * prevent losing any event. 2379c8e55f36STejun Heo */ 2380c8e55f36STejun Heo worker_enter_idle(worker); 2381c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 23828b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 23831da177e4SLinus Torvalds schedule(); 2384c8e55f36STejun Heo goto woke_up; 23851da177e4SLinus Torvalds } 23861da177e4SLinus Torvalds 2387e22bee78STejun Heo /** 2388e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2389e22bee78STejun Heo * @__wq: the associated workqueue 2390e22bee78STejun Heo * 2391e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2392e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2393e22bee78STejun Heo * 2394e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 2395e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2396e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2397e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2398e22bee78STejun Heo * the problem rescuer solves. 2399e22bee78STejun Heo * 2400e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 2401e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 2402e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2403e22bee78STejun Heo * 2404e22bee78STejun Heo * This should happen rarely. 2405e22bee78STejun Heo */ 2406e22bee78STejun Heo static int rescuer_thread(void *__wq) 2407e22bee78STejun Heo { 2408e22bee78STejun Heo struct workqueue_struct *wq = __wq; 2409e22bee78STejun Heo struct worker *rescuer = wq->rescuer; 2410e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2411f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2412e22bee78STejun Heo unsigned int cpu; 2413e22bee78STejun Heo 2414e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2415e22bee78STejun Heo repeat: 2416e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 24171da177e4SLinus Torvalds 2418412d32e6SMike Galbraith if (kthread_should_stop()) { 2419412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2420e22bee78STejun Heo return 0; 2421412d32e6SMike Galbraith } 24221da177e4SLinus Torvalds 2423f3421797STejun Heo /* 2424f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2425f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2426f3421797STejun Heo */ 2427f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2428f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2429f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2430bd7bdd43STejun Heo struct worker_pool *pool = cwq->pool; 2431bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2432e22bee78STejun Heo struct work_struct *work, *n; 2433e22bee78STejun Heo 2434e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2435f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2436e22bee78STejun Heo 2437e22bee78STejun Heo /* migrate to the target cpu if possible */ 2438bd7bdd43STejun Heo rescuer->pool = pool; 2439e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2440e22bee78STejun Heo 2441e22bee78STejun Heo /* 2442e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2443e22bee78STejun Heo * process'em. 2444e22bee78STejun Heo */ 2445e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2446bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2447e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2448e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2449e22bee78STejun Heo 2450e22bee78STejun Heo process_scheduled_works(rescuer); 24517576958aSTejun Heo 24527576958aSTejun Heo /* 24537576958aSTejun Heo * Leave this gcwq. If keep_working() is %true, notify a 24547576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24557576958aSTejun Heo * and stalling the execution. 24567576958aSTejun Heo */ 245763d95a91STejun Heo if (keep_working(pool)) 245863d95a91STejun Heo wake_up_worker(pool); 24597576958aSTejun Heo 2460e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 24611da177e4SLinus Torvalds } 24621da177e4SLinus Torvalds 2463e22bee78STejun Heo schedule(); 2464e22bee78STejun Heo goto repeat; 24651da177e4SLinus Torvalds } 24661da177e4SLinus Torvalds 2467fc2e4d70SOleg Nesterov struct wq_barrier { 2468fc2e4d70SOleg Nesterov struct work_struct work; 2469fc2e4d70SOleg Nesterov struct completion done; 2470fc2e4d70SOleg Nesterov }; 2471fc2e4d70SOleg Nesterov 2472fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2473fc2e4d70SOleg Nesterov { 2474fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2475fc2e4d70SOleg Nesterov complete(&barr->done); 2476fc2e4d70SOleg Nesterov } 2477fc2e4d70SOleg Nesterov 24784690c4abSTejun Heo /** 24794690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 24804690c4abSTejun Heo * @cwq: cwq to insert barrier into 24814690c4abSTejun Heo * @barr: wq_barrier to insert 2482affee4b2STejun Heo * @target: target work to attach @barr to 2483affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24844690c4abSTejun Heo * 2485affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2486affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2487affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2488affee4b2STejun Heo * cpu. 2489affee4b2STejun Heo * 2490affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2491affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2492affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2493affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2494affee4b2STejun Heo * after a work with LINKED flag set. 2495affee4b2STejun Heo * 2496affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2497affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 24984690c4abSTejun Heo * 24994690c4abSTejun Heo * CONTEXT: 25008b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 25014690c4abSTejun Heo */ 250283c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2503affee4b2STejun Heo struct wq_barrier *barr, 2504affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2505fc2e4d70SOleg Nesterov { 2506affee4b2STejun Heo struct list_head *head; 2507affee4b2STejun Heo unsigned int linked = 0; 2508affee4b2STejun Heo 2509dc186ad7SThomas Gleixner /* 25108b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 2511dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2512dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2513dc186ad7SThomas Gleixner * might deadlock. 2514dc186ad7SThomas Gleixner */ 2515ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 251622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2517fc2e4d70SOleg Nesterov init_completion(&barr->done); 251883c22520SOleg Nesterov 2519affee4b2STejun Heo /* 2520affee4b2STejun Heo * If @target is currently being executed, schedule the 2521affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2522affee4b2STejun Heo */ 2523affee4b2STejun Heo if (worker) 2524affee4b2STejun Heo head = worker->scheduled.next; 2525affee4b2STejun Heo else { 2526affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2527affee4b2STejun Heo 2528affee4b2STejun Heo head = target->entry.next; 2529affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2530affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2531affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2532affee4b2STejun Heo } 2533affee4b2STejun Heo 2534dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2535affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2536affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2537fc2e4d70SOleg Nesterov } 2538fc2e4d70SOleg Nesterov 253973f53c4aSTejun Heo /** 254073f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 254173f53c4aSTejun Heo * @wq: workqueue being flushed 254273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 254373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 254473f53c4aSTejun Heo * 254573f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 254673f53c4aSTejun Heo * 254773f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 254873f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 254973f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 255073f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 255173f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 255273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 255373f53c4aSTejun Heo * 255473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 255573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 255673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 255773f53c4aSTejun Heo * is returned. 255873f53c4aSTejun Heo * 255973f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 256073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 256173f53c4aSTejun Heo * advanced to @work_color. 256273f53c4aSTejun Heo * 256373f53c4aSTejun Heo * CONTEXT: 256473f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 256573f53c4aSTejun Heo * 256673f53c4aSTejun Heo * RETURNS: 256773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 256873f53c4aSTejun Heo * otherwise. 256973f53c4aSTejun Heo */ 257073f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 257173f53c4aSTejun Heo int flush_color, int work_color) 25721da177e4SLinus Torvalds { 257373f53c4aSTejun Heo bool wait = false; 257473f53c4aSTejun Heo unsigned int cpu; 25751da177e4SLinus Torvalds 257673f53c4aSTejun Heo if (flush_color >= 0) { 257773f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 257873f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2579dc186ad7SThomas Gleixner } 258014441960SOleg Nesterov 2581f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 258273f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2583bd7bdd43STejun Heo struct global_cwq *gcwq = cwq->pool->gcwq; 25841da177e4SLinus Torvalds 25858b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo if (flush_color >= 0) { 258873f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 258973f53c4aSTejun Heo 259073f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 259173f53c4aSTejun Heo cwq->flush_color = flush_color; 259273f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 259373f53c4aSTejun Heo wait = true; 25941da177e4SLinus Torvalds } 259573f53c4aSTejun Heo } 259673f53c4aSTejun Heo 259773f53c4aSTejun Heo if (work_color >= 0) { 259873f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 259973f53c4aSTejun Heo cwq->work_color = work_color; 260073f53c4aSTejun Heo } 260173f53c4aSTejun Heo 26028b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 26031da177e4SLinus Torvalds } 26041da177e4SLinus Torvalds 260573f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 260673f53c4aSTejun Heo complete(&wq->first_flusher->done); 260773f53c4aSTejun Heo 260873f53c4aSTejun Heo return wait; 260983c22520SOleg Nesterov } 26101da177e4SLinus Torvalds 26110fcb78c2SRolf Eike Beer /** 26121da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26130fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26141da177e4SLinus Torvalds * 26151da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 26161da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 26171da177e4SLinus Torvalds * 2618fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2619fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 26201da177e4SLinus Torvalds */ 26217ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26221da177e4SLinus Torvalds { 262373f53c4aSTejun Heo struct wq_flusher this_flusher = { 262473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 262573f53c4aSTejun Heo .flush_color = -1, 262673f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 262773f53c4aSTejun Heo }; 262873f53c4aSTejun Heo int next_color; 2629b1f4ec17SOleg Nesterov 26303295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26313295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 263273f53c4aSTejun Heo 263373f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 263473f53c4aSTejun Heo 263573f53c4aSTejun Heo /* 263673f53c4aSTejun Heo * Start-to-wait phase 263773f53c4aSTejun Heo */ 263873f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo if (next_color != wq->flush_color) { 264173f53c4aSTejun Heo /* 264273f53c4aSTejun Heo * Color space is not full. The current work_color 264373f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 264473f53c4aSTejun Heo * by one. 264573f53c4aSTejun Heo */ 264673f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 264773f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 264873f53c4aSTejun Heo wq->work_color = next_color; 264973f53c4aSTejun Heo 265073f53c4aSTejun Heo if (!wq->first_flusher) { 265173f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 265273f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 265373f53c4aSTejun Heo 265473f53c4aSTejun Heo wq->first_flusher = &this_flusher; 265573f53c4aSTejun Heo 265673f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 265773f53c4aSTejun Heo wq->work_color)) { 265873f53c4aSTejun Heo /* nothing to flush, done */ 265973f53c4aSTejun Heo wq->flush_color = next_color; 266073f53c4aSTejun Heo wq->first_flusher = NULL; 266173f53c4aSTejun Heo goto out_unlock; 266273f53c4aSTejun Heo } 266373f53c4aSTejun Heo } else { 266473f53c4aSTejun Heo /* wait in queue */ 266573f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 266673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 266773f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 266873f53c4aSTejun Heo } 266973f53c4aSTejun Heo } else { 267073f53c4aSTejun Heo /* 267173f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 267273f53c4aSTejun Heo * The next flush completion will assign us 267373f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 267473f53c4aSTejun Heo */ 267573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 267673f53c4aSTejun Heo } 267773f53c4aSTejun Heo 267873f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 267973f53c4aSTejun Heo 268073f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 268173f53c4aSTejun Heo 268273f53c4aSTejun Heo /* 268373f53c4aSTejun Heo * Wake-up-and-cascade phase 268473f53c4aSTejun Heo * 268573f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 268673f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 268773f53c4aSTejun Heo */ 268873f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 268973f53c4aSTejun Heo return; 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 269273f53c4aSTejun Heo 26934ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26944ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26954ce48b37STejun Heo goto out_unlock; 26964ce48b37STejun Heo 269773f53c4aSTejun Heo wq->first_flusher = NULL; 269873f53c4aSTejun Heo 269973f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 270073f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo while (true) { 270373f53c4aSTejun Heo struct wq_flusher *next, *tmp; 270473f53c4aSTejun Heo 270573f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 270673f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 270773f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 270873f53c4aSTejun Heo break; 270973f53c4aSTejun Heo list_del_init(&next->list); 271073f53c4aSTejun Heo complete(&next->done); 271173f53c4aSTejun Heo } 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 271473f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 271573f53c4aSTejun Heo 271673f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 271773f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 271873f53c4aSTejun Heo 271973f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 272073f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 272173f53c4aSTejun Heo /* 272273f53c4aSTejun Heo * Assign the same color to all overflowed 272373f53c4aSTejun Heo * flushers, advance work_color and append to 272473f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 272573f53c4aSTejun Heo * phase for these overflowed flushers. 272673f53c4aSTejun Heo */ 272773f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 272873f53c4aSTejun Heo tmp->flush_color = wq->work_color; 272973f53c4aSTejun Heo 273073f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 273173f53c4aSTejun Heo 273273f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 273373f53c4aSTejun Heo &wq->flusher_queue); 273473f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 273573f53c4aSTejun Heo } 273673f53c4aSTejun Heo 273773f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 273873f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 273973f53c4aSTejun Heo break; 274073f53c4aSTejun Heo } 274173f53c4aSTejun Heo 274273f53c4aSTejun Heo /* 274373f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 274473f53c4aSTejun Heo * the new first flusher and arm cwqs. 274573f53c4aSTejun Heo */ 274673f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 274773f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 274873f53c4aSTejun Heo 274973f53c4aSTejun Heo list_del_init(&next->list); 275073f53c4aSTejun Heo wq->first_flusher = next; 275173f53c4aSTejun Heo 275273f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 275373f53c4aSTejun Heo break; 275473f53c4aSTejun Heo 275573f53c4aSTejun Heo /* 275673f53c4aSTejun Heo * Meh... this color is already done, clear first 275773f53c4aSTejun Heo * flusher and repeat cascading. 275873f53c4aSTejun Heo */ 275973f53c4aSTejun Heo wq->first_flusher = NULL; 276073f53c4aSTejun Heo } 276173f53c4aSTejun Heo 276273f53c4aSTejun Heo out_unlock: 276373f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27641da177e4SLinus Torvalds } 2765ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27661da177e4SLinus Torvalds 27679c5a2ba7STejun Heo /** 27689c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27699c5a2ba7STejun Heo * @wq: workqueue to drain 27709c5a2ba7STejun Heo * 27719c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27729c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27739c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27749c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27759c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27769c5a2ba7STejun Heo * takes too long. 27779c5a2ba7STejun Heo */ 27789c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27799c5a2ba7STejun Heo { 27809c5a2ba7STejun Heo unsigned int flush_cnt = 0; 27819c5a2ba7STejun Heo unsigned int cpu; 27829c5a2ba7STejun Heo 27839c5a2ba7STejun Heo /* 27849c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27859c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 27869c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 27879c5a2ba7STejun Heo */ 27889c5a2ba7STejun Heo spin_lock(&workqueue_lock); 27899c5a2ba7STejun Heo if (!wq->nr_drainers++) 27909c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 27919c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 27929c5a2ba7STejun Heo reflush: 27939c5a2ba7STejun Heo flush_workqueue(wq); 27949c5a2ba7STejun Heo 27959c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 27969c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2797fa2563e4SThomas Tuttle bool drained; 27989c5a2ba7STejun Heo 2799bd7bdd43STejun Heo spin_lock_irq(&cwq->pool->gcwq->lock); 2800fa2563e4SThomas Tuttle drained = !cwq->nr_active && list_empty(&cwq->delayed_works); 2801bd7bdd43STejun Heo spin_unlock_irq(&cwq->pool->gcwq->lock); 2802fa2563e4SThomas Tuttle 2803fa2563e4SThomas Tuttle if (drained) 28049c5a2ba7STejun Heo continue; 28059c5a2ba7STejun Heo 28069c5a2ba7STejun Heo if (++flush_cnt == 10 || 28079c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2808044c782cSValentin Ilie pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n", 28099c5a2ba7STejun Heo wq->name, flush_cnt); 28109c5a2ba7STejun Heo goto reflush; 28119c5a2ba7STejun Heo } 28129c5a2ba7STejun Heo 28139c5a2ba7STejun Heo spin_lock(&workqueue_lock); 28149c5a2ba7STejun Heo if (!--wq->nr_drainers) 28159c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 28169c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 28179c5a2ba7STejun Heo } 28189c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28199c5a2ba7STejun Heo 2820606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2821baf59022STejun Heo { 2822baf59022STejun Heo struct worker *worker = NULL; 2823baf59022STejun Heo struct global_cwq *gcwq; 2824baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2825baf59022STejun Heo 2826baf59022STejun Heo might_sleep(); 2827baf59022STejun Heo gcwq = get_work_gcwq(work); 2828baf59022STejun Heo if (!gcwq) 2829baf59022STejun Heo return false; 2830baf59022STejun Heo 2831baf59022STejun Heo spin_lock_irq(&gcwq->lock); 2832baf59022STejun Heo if (!list_empty(&work->entry)) { 2833baf59022STejun Heo /* 2834baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2835baf59022STejun Heo * If it was re-queued to a different gcwq under us, we 2836baf59022STejun Heo * are not going to wait. 2837baf59022STejun Heo */ 2838baf59022STejun Heo smp_rmb(); 2839baf59022STejun Heo cwq = get_work_cwq(work); 2840bd7bdd43STejun Heo if (unlikely(!cwq || gcwq != cwq->pool->gcwq)) 2841baf59022STejun Heo goto already_gone; 2842606a5020STejun Heo } else { 2843baf59022STejun Heo worker = find_worker_executing_work(gcwq, work); 2844baf59022STejun Heo if (!worker) 2845baf59022STejun Heo goto already_gone; 2846baf59022STejun Heo cwq = worker->current_cwq; 2847606a5020STejun Heo } 2848baf59022STejun Heo 2849baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2850baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2851baf59022STejun Heo 2852e159489bSTejun Heo /* 2853e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2854e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2855e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2856e159489bSTejun Heo * access. 2857e159489bSTejun Heo */ 2858e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2859baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2860e159489bSTejun Heo else 2861e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2862baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2863e159489bSTejun Heo 2864baf59022STejun Heo return true; 2865baf59022STejun Heo already_gone: 2866baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2867baf59022STejun Heo return false; 2868baf59022STejun Heo } 2869baf59022STejun Heo 2870db700897SOleg Nesterov /** 2871401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2872401a8d04STejun Heo * @work: the work to flush 2873db700897SOleg Nesterov * 2874606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2875606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2876401a8d04STejun Heo * 2877401a8d04STejun Heo * RETURNS: 2878401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2879401a8d04STejun Heo * %false if it was already idle. 2880db700897SOleg Nesterov */ 2881401a8d04STejun Heo bool flush_work(struct work_struct *work) 2882db700897SOleg Nesterov { 2883db700897SOleg Nesterov struct wq_barrier barr; 2884db700897SOleg Nesterov 28850976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28860976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28870976dfc1SStephen Boyd 2888606a5020STejun Heo if (start_flush_work(work, &barr)) { 2889db700897SOleg Nesterov wait_for_completion(&barr.done); 2890dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2891401a8d04STejun Heo return true; 2892606a5020STejun Heo } else { 2893401a8d04STejun Heo return false; 2894db700897SOleg Nesterov } 2895606a5020STejun Heo } 2896db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2897db700897SOleg Nesterov 289836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2899401a8d04STejun Heo { 2900bbb68dfaSTejun Heo unsigned long flags; 29011f1f642eSOleg Nesterov int ret; 29021f1f642eSOleg Nesterov 29031f1f642eSOleg Nesterov do { 2904bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2905bbb68dfaSTejun Heo /* 2906bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2907bbb68dfaSTejun Heo * would be waiting for before retrying. 2908bbb68dfaSTejun Heo */ 2909bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2910606a5020STejun Heo flush_work(work); 29111f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29121f1f642eSOleg Nesterov 2913bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2914bbb68dfaSTejun Heo mark_work_canceling(work); 2915bbb68dfaSTejun Heo local_irq_restore(flags); 2916bbb68dfaSTejun Heo 2917606a5020STejun Heo flush_work(work); 29187a22ad75STejun Heo clear_work_data(work); 29191f1f642eSOleg Nesterov return ret; 29201f1f642eSOleg Nesterov } 29211f1f642eSOleg Nesterov 29226e84d644SOleg Nesterov /** 2923401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2924401a8d04STejun Heo * @work: the work to cancel 29256e84d644SOleg Nesterov * 2926401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2927401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2928401a8d04STejun Heo * another workqueue. On return from this function, @work is 2929401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29301f1f642eSOleg Nesterov * 2931401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2932401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29336e84d644SOleg Nesterov * 2934401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29356e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2936401a8d04STejun Heo * 2937401a8d04STejun Heo * RETURNS: 2938401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29396e84d644SOleg Nesterov */ 2940401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29416e84d644SOleg Nesterov { 294236e227d2STejun Heo return __cancel_work_timer(work, false); 2943b89deed3SOleg Nesterov } 294428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2945b89deed3SOleg Nesterov 29466e84d644SOleg Nesterov /** 2947401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2948401a8d04STejun Heo * @dwork: the delayed work to flush 29496e84d644SOleg Nesterov * 2950401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2951401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2952401a8d04STejun Heo * considers the last queueing instance of @dwork. 29531f1f642eSOleg Nesterov * 2954401a8d04STejun Heo * RETURNS: 2955401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2956401a8d04STejun Heo * %false if it was already idle. 29576e84d644SOleg Nesterov */ 2958401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2959401a8d04STejun Heo { 29608930cabaSTejun Heo local_irq_disable(); 2961401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 29621265057fSTejun Heo __queue_work(dwork->cpu, 2963401a8d04STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 29648930cabaSTejun Heo local_irq_enable(); 2965401a8d04STejun Heo return flush_work(&dwork->work); 2966401a8d04STejun Heo } 2967401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2968401a8d04STejun Heo 2969401a8d04STejun Heo /** 297057b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 297157b30ae7STejun Heo * @dwork: delayed_work to cancel 297209383498STejun Heo * 297357b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 297457b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 297557b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 297657b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 297757b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 297809383498STejun Heo * 297957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 298009383498STejun Heo */ 298157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 298209383498STejun Heo { 298357b30ae7STejun Heo unsigned long flags; 298457b30ae7STejun Heo int ret; 298557b30ae7STejun Heo 298657b30ae7STejun Heo do { 298757b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 298857b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 298957b30ae7STejun Heo 299057b30ae7STejun Heo if (unlikely(ret < 0)) 299157b30ae7STejun Heo return false; 299257b30ae7STejun Heo 299357b30ae7STejun Heo set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work)); 299457b30ae7STejun Heo local_irq_restore(flags); 2995c0158ca6SDan Magenheimer return ret; 299609383498STejun Heo } 299757b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 299809383498STejun Heo 299909383498STejun Heo /** 3000401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3001401a8d04STejun Heo * @dwork: the delayed work cancel 3002401a8d04STejun Heo * 3003401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3004401a8d04STejun Heo * 3005401a8d04STejun Heo * RETURNS: 3006401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3007401a8d04STejun Heo */ 3008401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 30096e84d644SOleg Nesterov { 301036e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30116e84d644SOleg Nesterov } 3012f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30131da177e4SLinus Torvalds 30140fcb78c2SRolf Eike Beer /** 3015c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 3016c1a220e7SZhang Rui * @cpu: cpu to put the work task on 3017c1a220e7SZhang Rui * @work: job to be done 3018c1a220e7SZhang Rui * 3019c1a220e7SZhang Rui * This puts a job on a specific cpu 3020c1a220e7SZhang Rui */ 3021d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 3022c1a220e7SZhang Rui { 3023d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 3024c1a220e7SZhang Rui } 3025c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 3026c1a220e7SZhang Rui 30270fcb78c2SRolf Eike Beer /** 3028ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 30291da177e4SLinus Torvalds * @work: job to be done 30301da177e4SLinus Torvalds * 3031d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 3032d4283e93STejun Heo * %true otherwise. 303352bad64dSDavid Howells * 30340fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 30350fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 30360fcb78c2SRolf Eike Beer * workqueue otherwise. 30371da177e4SLinus Torvalds */ 3038d4283e93STejun Heo bool schedule_work(struct work_struct *work) 30391da177e4SLinus Torvalds { 30400fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 30411da177e4SLinus Torvalds } 30420fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 30431da177e4SLinus Torvalds 30440fcb78c2SRolf Eike Beer /** 30450fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 30460fcb78c2SRolf Eike Beer * @cpu: cpu to use 30470fcb78c2SRolf Eike Beer * @dwork: job to be done 30480fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 30490fcb78c2SRolf Eike Beer * 30500fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 30510fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 30520fcb78c2SRolf Eike Beer */ 3053d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 3054d4283e93STejun Heo unsigned long delay) 30551da177e4SLinus Torvalds { 3056d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 30571da177e4SLinus Torvalds } 3058ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 30591da177e4SLinus Torvalds 3060b6136773SAndrew Morton /** 30610a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 30620a13c00eSTejun Heo * @dwork: job to be done 30630a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 30640a13c00eSTejun Heo * 30650a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 30660a13c00eSTejun Heo * workqueue. 30670a13c00eSTejun Heo */ 3068d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 30690a13c00eSTejun Heo { 30700a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30710a13c00eSTejun Heo } 30720a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30730a13c00eSTejun Heo 30740a13c00eSTejun Heo /** 307531ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3076b6136773SAndrew Morton * @func: the function to call 3077b6136773SAndrew Morton * 307831ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 307931ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3080b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 308131ddd871STejun Heo * 308231ddd871STejun Heo * RETURNS: 308331ddd871STejun Heo * 0 on success, -errno on failure. 3084b6136773SAndrew Morton */ 308565f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 308615316ba8SChristoph Lameter { 308715316ba8SChristoph Lameter int cpu; 308838f51568SNamhyung Kim struct work_struct __percpu *works; 308915316ba8SChristoph Lameter 3090b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3091b6136773SAndrew Morton if (!works) 309215316ba8SChristoph Lameter return -ENOMEM; 3093b6136773SAndrew Morton 309495402b38SGautham R Shenoy get_online_cpus(); 309593981800STejun Heo 309615316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30979bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30989bfb1839SIngo Molnar 30999bfb1839SIngo Molnar INIT_WORK(work, func); 31008de6d308SOleg Nesterov schedule_work_on(cpu, work); 310115316ba8SChristoph Lameter } 310293981800STejun Heo 310393981800STejun Heo for_each_online_cpu(cpu) 31048616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 310593981800STejun Heo 310695402b38SGautham R Shenoy put_online_cpus(); 3107b6136773SAndrew Morton free_percpu(works); 310815316ba8SChristoph Lameter return 0; 310915316ba8SChristoph Lameter } 311015316ba8SChristoph Lameter 3111eef6a7d5SAlan Stern /** 3112eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3113eef6a7d5SAlan Stern * 3114eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3115eef6a7d5SAlan Stern * completion. 3116eef6a7d5SAlan Stern * 3117eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3118eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3119eef6a7d5SAlan Stern * will lead to deadlock: 3120eef6a7d5SAlan Stern * 3121eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3122eef6a7d5SAlan Stern * a lock held by your code or its caller. 3123eef6a7d5SAlan Stern * 3124eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3125eef6a7d5SAlan Stern * 3126eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3127eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3128eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3129eef6a7d5SAlan Stern * 3130eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3131eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3132eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3133eef6a7d5SAlan Stern * cancel_work_sync() instead. 3134eef6a7d5SAlan Stern */ 31351da177e4SLinus Torvalds void flush_scheduled_work(void) 31361da177e4SLinus Torvalds { 3137d320c038STejun Heo flush_workqueue(system_wq); 31381da177e4SLinus Torvalds } 3139ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 31401da177e4SLinus Torvalds 31411da177e4SLinus Torvalds /** 31421fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31431fa44ecaSJames Bottomley * @fn: the function to execute 31441fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31451fa44ecaSJames Bottomley * be available when the work executes) 31461fa44ecaSJames Bottomley * 31471fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31481fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31491fa44ecaSJames Bottomley * 31501fa44ecaSJames Bottomley * Returns: 0 - function was executed 31511fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31521fa44ecaSJames Bottomley */ 315365f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31541fa44ecaSJames Bottomley { 31551fa44ecaSJames Bottomley if (!in_interrupt()) { 315665f27f38SDavid Howells fn(&ew->work); 31571fa44ecaSJames Bottomley return 0; 31581fa44ecaSJames Bottomley } 31591fa44ecaSJames Bottomley 316065f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31611fa44ecaSJames Bottomley schedule_work(&ew->work); 31621fa44ecaSJames Bottomley 31631fa44ecaSJames Bottomley return 1; 31641fa44ecaSJames Bottomley } 31651fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31661fa44ecaSJames Bottomley 31671da177e4SLinus Torvalds int keventd_up(void) 31681da177e4SLinus Torvalds { 3169d320c038STejun Heo return system_wq != NULL; 31701da177e4SLinus Torvalds } 31711da177e4SLinus Torvalds 3172bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 31731da177e4SLinus Torvalds { 31743af24433SOleg Nesterov /* 31750f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 31760f900049STejun Heo * Make sure that the alignment isn't lower than that of 31770f900049STejun Heo * unsigned long long. 31783af24433SOleg Nesterov */ 31790f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 31800f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 31810f900049STejun Heo __alignof__(unsigned long long)); 31823af24433SOleg Nesterov 3183e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3184f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 3185931ac77eSTejun Heo else { 31860f900049STejun Heo void *ptr; 3187e1d8aa9fSFrederic Weisbecker 31880f900049STejun Heo /* 3189f3421797STejun Heo * Allocate enough room to align cwq and put an extra 3190f3421797STejun Heo * pointer at the end pointing back to the originally 3191f3421797STejun Heo * allocated pointer which will be used for free. 31920f900049STejun Heo */ 3193bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3194bdbc5dd7STejun Heo if (ptr) { 3195bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 3196bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 3197bdbc5dd7STejun Heo } 31983af24433SOleg Nesterov } 31993af24433SOleg Nesterov 32000415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3201bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 3202bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 32030f900049STejun Heo } 32040f900049STejun Heo 3205bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 320606ba38a9SOleg Nesterov { 3207e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3208bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 3209f3421797STejun Heo else if (wq->cpu_wq.single) { 3210f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 3211f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 321206ba38a9SOleg Nesterov } 321306ba38a9SOleg Nesterov } 321406ba38a9SOleg Nesterov 3215f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3216f3421797STejun Heo const char *name) 3217b71ab8c2STejun Heo { 3218f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3219f3421797STejun Heo 3220f3421797STejun Heo if (max_active < 1 || max_active > lim) 3221044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3222f3421797STejun Heo max_active, name, 1, lim); 3223b71ab8c2STejun Heo 3224f3421797STejun Heo return clamp_val(max_active, 1, lim); 3225b71ab8c2STejun Heo } 3226b71ab8c2STejun Heo 3227b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 322897e37d7bSTejun Heo unsigned int flags, 32291e19ffc6STejun Heo int max_active, 3230eb13ba87SJohannes Berg struct lock_class_key *key, 3231b196be89STejun Heo const char *lock_name, ...) 32323af24433SOleg Nesterov { 3233b196be89STejun Heo va_list args, args1; 32343af24433SOleg Nesterov struct workqueue_struct *wq; 3235c34056a3STejun Heo unsigned int cpu; 3236b196be89STejun Heo size_t namelen; 3237b196be89STejun Heo 3238b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3239b196be89STejun Heo va_start(args, lock_name); 3240b196be89STejun Heo va_copy(args1, args); 3241b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3242b196be89STejun Heo 3243b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3244b196be89STejun Heo if (!wq) 3245b196be89STejun Heo goto err; 3246b196be89STejun Heo 3247b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3248b196be89STejun Heo va_end(args); 3249b196be89STejun Heo va_end(args1); 32503af24433SOleg Nesterov 3251f3421797STejun Heo /* 32526370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 32536370a6adSTejun Heo * have a rescuer to guarantee forward progress. 32546370a6adSTejun Heo */ 32556370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 32566370a6adSTejun Heo flags |= WQ_RESCUER; 32576370a6adSTejun Heo 3258d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3259b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 32603af24433SOleg Nesterov 3261b196be89STejun Heo /* init wq */ 326297e37d7bSTejun Heo wq->flags = flags; 3263a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 326473f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 326573f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 326673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 326773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 32683af24433SOleg Nesterov 3269eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3270cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 32713af24433SOleg Nesterov 3272bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 3273bdbc5dd7STejun Heo goto err; 3274bdbc5dd7STejun Heo 3275f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 32761537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 32778b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 32783270476aSTejun Heo int pool_idx = (bool)(flags & WQ_HIGHPRI); 32791537663fSTejun Heo 32800f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 32813270476aSTejun Heo cwq->pool = &gcwq->pools[pool_idx]; 3282c34056a3STejun Heo cwq->wq = wq; 328373f53c4aSTejun Heo cwq->flush_color = -1; 32841e19ffc6STejun Heo cwq->max_active = max_active; 32851e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3286e22bee78STejun Heo } 32871537663fSTejun Heo 3288e22bee78STejun Heo if (flags & WQ_RESCUER) { 3289e22bee78STejun Heo struct worker *rescuer; 3290e22bee78STejun Heo 3291f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3292e22bee78STejun Heo goto err; 3293e22bee78STejun Heo 3294e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3295e22bee78STejun Heo if (!rescuer) 3296e22bee78STejun Heo goto err; 3297e22bee78STejun Heo 3298b196be89STejun Heo rescuer->task = kthread_create(rescuer_thread, wq, "%s", 3299b196be89STejun Heo wq->name); 3300e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3301e22bee78STejun Heo goto err; 3302e22bee78STejun Heo 3303e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3304e22bee78STejun Heo wake_up_process(rescuer->task); 33053af24433SOleg Nesterov } 33061537663fSTejun Heo 33073af24433SOleg Nesterov /* 3308a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3309a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3310a0a1a5fdSTejun Heo * workqueue to workqueues list. 33113af24433SOleg Nesterov */ 33123af24433SOleg Nesterov spin_lock(&workqueue_lock); 3313a0a1a5fdSTejun Heo 331458a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3315f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3316a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3317a0a1a5fdSTejun Heo 33183af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3319a0a1a5fdSTejun Heo 33203af24433SOleg Nesterov spin_unlock(&workqueue_lock); 33213af24433SOleg Nesterov 33223af24433SOleg Nesterov return wq; 33234690c4abSTejun Heo err: 33244690c4abSTejun Heo if (wq) { 3325bdbc5dd7STejun Heo free_cwqs(wq); 3326f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3327e22bee78STejun Heo kfree(wq->rescuer); 33284690c4abSTejun Heo kfree(wq); 33293af24433SOleg Nesterov } 33304690c4abSTejun Heo return NULL; 33311da177e4SLinus Torvalds } 3332d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 33331da177e4SLinus Torvalds 33343af24433SOleg Nesterov /** 33353af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 33363af24433SOleg Nesterov * @wq: target workqueue 33373af24433SOleg Nesterov * 33383af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 33393af24433SOleg Nesterov */ 33403af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 33413af24433SOleg Nesterov { 3342c8e55f36STejun Heo unsigned int cpu; 33433af24433SOleg Nesterov 33449c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 33459c5a2ba7STejun Heo drain_workqueue(wq); 3346c8efcc25STejun Heo 3347a0a1a5fdSTejun Heo /* 3348a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3349a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3350a0a1a5fdSTejun Heo */ 335195402b38SGautham R Shenoy spin_lock(&workqueue_lock); 33523af24433SOleg Nesterov list_del(&wq->list); 335395402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 33543af24433SOleg Nesterov 3355e22bee78STejun Heo /* sanity check */ 3356f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 335773f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 335873f53c4aSTejun Heo int i; 33593af24433SOleg Nesterov 336073f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 336173f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 33621e19ffc6STejun Heo BUG_ON(cwq->nr_active); 33631e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 336473f53c4aSTejun Heo } 33651537663fSTejun Heo 3366e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3367e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3368f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 33698d9df9f0SXiaotian Feng kfree(wq->rescuer); 3370e22bee78STejun Heo } 3371e22bee78STejun Heo 3372bdbc5dd7STejun Heo free_cwqs(wq); 33733af24433SOleg Nesterov kfree(wq); 33743af24433SOleg Nesterov } 33753af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 33763af24433SOleg Nesterov 3377dcd989cbSTejun Heo /** 33789f4bd4cdSLai Jiangshan * cwq_set_max_active - adjust max_active of a cwq 33799f4bd4cdSLai Jiangshan * @cwq: target cpu_workqueue_struct 33809f4bd4cdSLai Jiangshan * @max_active: new max_active value. 33819f4bd4cdSLai Jiangshan * 33829f4bd4cdSLai Jiangshan * Set @cwq->max_active to @max_active and activate delayed works if 33839f4bd4cdSLai Jiangshan * increased. 33849f4bd4cdSLai Jiangshan * 33859f4bd4cdSLai Jiangshan * CONTEXT: 33869f4bd4cdSLai Jiangshan * spin_lock_irq(gcwq->lock). 33879f4bd4cdSLai Jiangshan */ 33889f4bd4cdSLai Jiangshan static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active) 33899f4bd4cdSLai Jiangshan { 33909f4bd4cdSLai Jiangshan cwq->max_active = max_active; 33919f4bd4cdSLai Jiangshan 33929f4bd4cdSLai Jiangshan while (!list_empty(&cwq->delayed_works) && 33939f4bd4cdSLai Jiangshan cwq->nr_active < cwq->max_active) 33949f4bd4cdSLai Jiangshan cwq_activate_first_delayed(cwq); 33959f4bd4cdSLai Jiangshan } 33969f4bd4cdSLai Jiangshan 33979f4bd4cdSLai Jiangshan /** 3398dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3399dcd989cbSTejun Heo * @wq: target workqueue 3400dcd989cbSTejun Heo * @max_active: new max_active value. 3401dcd989cbSTejun Heo * 3402dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3403dcd989cbSTejun Heo * 3404dcd989cbSTejun Heo * CONTEXT: 3405dcd989cbSTejun Heo * Don't call from IRQ context. 3406dcd989cbSTejun Heo */ 3407dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3408dcd989cbSTejun Heo { 3409dcd989cbSTejun Heo unsigned int cpu; 3410dcd989cbSTejun Heo 3411f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3412dcd989cbSTejun Heo 3413dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3414dcd989cbSTejun Heo 3415dcd989cbSTejun Heo wq->saved_max_active = max_active; 3416dcd989cbSTejun Heo 3417f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 3418dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3419dcd989cbSTejun Heo 3420dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 3421dcd989cbSTejun Heo 342258a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 3423dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 342470369b11SLai Jiangshan cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active); 3425dcd989cbSTejun Heo 3426dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 3427dcd989cbSTejun Heo } 3428dcd989cbSTejun Heo 3429dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3430dcd989cbSTejun Heo } 3431dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3432dcd989cbSTejun Heo 3433dcd989cbSTejun Heo /** 3434dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3435dcd989cbSTejun Heo * @cpu: CPU in question 3436dcd989cbSTejun Heo * @wq: target workqueue 3437dcd989cbSTejun Heo * 3438dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3439dcd989cbSTejun Heo * no synchronization around this function and the test result is 3440dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3441dcd989cbSTejun Heo * 3442dcd989cbSTejun Heo * RETURNS: 3443dcd989cbSTejun Heo * %true if congested, %false otherwise. 3444dcd989cbSTejun Heo */ 3445dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3446dcd989cbSTejun Heo { 3447dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3448dcd989cbSTejun Heo 3449dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3450dcd989cbSTejun Heo } 3451dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3452dcd989cbSTejun Heo 3453dcd989cbSTejun Heo /** 3454dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 3455dcd989cbSTejun Heo * @work: the work of interest 3456dcd989cbSTejun Heo * 3457dcd989cbSTejun Heo * RETURNS: 3458bdbc5dd7STejun Heo * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. 3459dcd989cbSTejun Heo */ 3460dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 3461dcd989cbSTejun Heo { 3462dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3463dcd989cbSTejun Heo 3464bdbc5dd7STejun Heo return gcwq ? gcwq->cpu : WORK_CPU_NONE; 3465dcd989cbSTejun Heo } 3466dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 3467dcd989cbSTejun Heo 3468dcd989cbSTejun Heo /** 3469dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3470dcd989cbSTejun Heo * @work: the work to be tested 3471dcd989cbSTejun Heo * 3472dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3473dcd989cbSTejun Heo * synchronization around this function and the test result is 3474dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3475dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 3476dcd989cbSTejun Heo * running state. 3477dcd989cbSTejun Heo * 3478dcd989cbSTejun Heo * RETURNS: 3479dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3480dcd989cbSTejun Heo */ 3481dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3482dcd989cbSTejun Heo { 3483dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3484dcd989cbSTejun Heo unsigned long flags; 3485dcd989cbSTejun Heo unsigned int ret = 0; 3486dcd989cbSTejun Heo 3487dcd989cbSTejun Heo if (!gcwq) 3488dcd989cbSTejun Heo return false; 3489dcd989cbSTejun Heo 3490dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 3491dcd989cbSTejun Heo 3492dcd989cbSTejun Heo if (work_pending(work)) 3493dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3494dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 3495dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3496dcd989cbSTejun Heo 3497dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 3498dcd989cbSTejun Heo 3499dcd989cbSTejun Heo return ret; 3500dcd989cbSTejun Heo } 3501dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3502dcd989cbSTejun Heo 3503db7bccf4STejun Heo /* 3504db7bccf4STejun Heo * CPU hotplug. 3505db7bccf4STejun Heo * 3506e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3507e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3508e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 3509e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 3510e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 3511e22bee78STejun Heo * blocked draining impractical. 3512e22bee78STejun Heo * 3513628c78e7STejun Heo * This is solved by allowing a gcwq to be disassociated from the CPU 3514628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 3515628c78e7STejun Heo * cpu comes back online. 3516db7bccf4STejun Heo */ 3517db7bccf4STejun Heo 351860373152STejun Heo /* claim manager positions of all pools */ 3519b2eb83d1SLai Jiangshan static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq) 352060373152STejun Heo { 352160373152STejun Heo struct worker_pool *pool; 352260373152STejun Heo 352360373152STejun Heo for_each_worker_pool(pool, gcwq) 3524b2eb83d1SLai Jiangshan mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools); 35258db25e78STejun Heo spin_lock_irq(&gcwq->lock); 352660373152STejun Heo } 352760373152STejun Heo 352860373152STejun Heo /* release manager positions */ 3529b2eb83d1SLai Jiangshan static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq) 353060373152STejun Heo { 353160373152STejun Heo struct worker_pool *pool; 353260373152STejun Heo 35338db25e78STejun Heo spin_unlock_irq(&gcwq->lock); 353460373152STejun Heo for_each_worker_pool(pool, gcwq) 3535b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 353660373152STejun Heo } 353760373152STejun Heo 3538628c78e7STejun Heo static void gcwq_unbind_fn(struct work_struct *work) 3539db7bccf4STejun Heo { 3540628c78e7STejun Heo struct global_cwq *gcwq = get_gcwq(smp_processor_id()); 35414ce62e9eSTejun Heo struct worker_pool *pool; 3542db7bccf4STejun Heo struct worker *worker; 3543db7bccf4STejun Heo struct hlist_node *pos; 3544db7bccf4STejun Heo int i; 3545db7bccf4STejun Heo 3546db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3547db7bccf4STejun Heo 3548b2eb83d1SLai Jiangshan gcwq_claim_assoc_and_lock(gcwq); 3549e22bee78STejun Heo 3550f2d5a0eeSTejun Heo /* 3551f2d5a0eeSTejun Heo * We've claimed all manager positions. Make all workers unbound 3552f2d5a0eeSTejun Heo * and set DISASSOCIATED. Before this, all workers except for the 3553f2d5a0eeSTejun Heo * ones which are still executing works from before the last CPU 3554f2d5a0eeSTejun Heo * down must be on the cpu. After this, they may become diasporas. 3555f2d5a0eeSTejun Heo */ 355660373152STejun Heo for_each_worker_pool(pool, gcwq) 35574ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3558403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3559db7bccf4STejun Heo 3560db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 3561403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3562db7bccf4STejun Heo 3563f2d5a0eeSTejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3564f2d5a0eeSTejun Heo 3565b2eb83d1SLai Jiangshan gcwq_release_assoc_and_unlock(gcwq); 3566e22bee78STejun Heo 3567e22bee78STejun Heo /* 3568628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3569628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 3570628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 3571628c78e7STejun Heo */ 3572628c78e7STejun Heo schedule(); 3573628c78e7STejun Heo 3574628c78e7STejun Heo /* 3575628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 3576628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 3577628c78e7STejun Heo * are always true as long as the worklist is not empty. @gcwq now 3578628c78e7STejun Heo * behaves as unbound (in terms of concurrency management) gcwq 3579628c78e7STejun Heo * which is served by workers tied to the CPU. 3580628c78e7STejun Heo * 3581628c78e7STejun Heo * On return from this function, the current worker would trigger 3582628c78e7STejun Heo * unbound chain execution of pending work items if other workers 3583628c78e7STejun Heo * didn't already. 3584e22bee78STejun Heo */ 35854ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 35864ce62e9eSTejun Heo atomic_set(get_pool_nr_running(pool), 0); 3587db7bccf4STejun Heo } 3588db7bccf4STejun Heo 35898db25e78STejun Heo /* 35908db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 35918db25e78STejun Heo * This will be registered high priority CPU notifier. 35928db25e78STejun Heo */ 35939fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 35941da177e4SLinus Torvalds unsigned long action, 35951da177e4SLinus Torvalds void *hcpu) 35961da177e4SLinus Torvalds { 35973af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3598db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 35994ce62e9eSTejun Heo struct worker_pool *pool; 36001da177e4SLinus Torvalds 36018db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 36023af24433SOleg Nesterov case CPU_UP_PREPARE: 36034ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 36043ce63377STejun Heo struct worker *worker; 36053ce63377STejun Heo 36063ce63377STejun Heo if (pool->nr_workers) 36073ce63377STejun Heo continue; 36083ce63377STejun Heo 36093ce63377STejun Heo worker = create_worker(pool); 36103ce63377STejun Heo if (!worker) 36113ce63377STejun Heo return NOTIFY_BAD; 36123ce63377STejun Heo 36133ce63377STejun Heo spin_lock_irq(&gcwq->lock); 36143ce63377STejun Heo start_worker(worker); 36153ce63377STejun Heo spin_unlock_irq(&gcwq->lock); 36163af24433SOleg Nesterov } 36171da177e4SLinus Torvalds break; 36181da177e4SLinus Torvalds 361965758202STejun Heo case CPU_DOWN_FAILED: 362065758202STejun Heo case CPU_ONLINE: 3621b2eb83d1SLai Jiangshan gcwq_claim_assoc_and_lock(gcwq); 36228db25e78STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 36238db25e78STejun Heo rebind_workers(gcwq); 3624b2eb83d1SLai Jiangshan gcwq_release_assoc_and_unlock(gcwq); 36258db25e78STejun Heo break; 362665758202STejun Heo } 362765758202STejun Heo return NOTIFY_OK; 362865758202STejun Heo } 362965758202STejun Heo 363065758202STejun Heo /* 363165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 363265758202STejun Heo * This will be registered as low priority CPU notifier. 363365758202STejun Heo */ 36349fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 363565758202STejun Heo unsigned long action, 363665758202STejun Heo void *hcpu) 363765758202STejun Heo { 36388db25e78STejun Heo unsigned int cpu = (unsigned long)hcpu; 36398db25e78STejun Heo struct work_struct unbind_work; 36408db25e78STejun Heo 364165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 364265758202STejun Heo case CPU_DOWN_PREPARE: 36438db25e78STejun Heo /* unbinding should happen on the local CPU */ 36448db25e78STejun Heo INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn); 36457635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 36468db25e78STejun Heo flush_work(&unbind_work); 36478db25e78STejun Heo break; 364865758202STejun Heo } 364965758202STejun Heo return NOTIFY_OK; 365065758202STejun Heo } 365165758202STejun Heo 36522d3854a3SRusty Russell #ifdef CONFIG_SMP 36538ccad40dSRusty Russell 36542d3854a3SRusty Russell struct work_for_cpu { 3655ed48ece2STejun Heo struct work_struct work; 36562d3854a3SRusty Russell long (*fn)(void *); 36572d3854a3SRusty Russell void *arg; 36582d3854a3SRusty Russell long ret; 36592d3854a3SRusty Russell }; 36602d3854a3SRusty Russell 3661ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 36622d3854a3SRusty Russell { 3663ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 3664ed48ece2STejun Heo 36652d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 36662d3854a3SRusty Russell } 36672d3854a3SRusty Russell 36682d3854a3SRusty Russell /** 36692d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 36702d3854a3SRusty Russell * @cpu: the cpu to run on 36712d3854a3SRusty Russell * @fn: the function to run 36722d3854a3SRusty Russell * @arg: the function arg 36732d3854a3SRusty Russell * 367431ad9081SRusty Russell * This will return the value @fn returns. 367531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 36766b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 36772d3854a3SRusty Russell */ 36782d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 36792d3854a3SRusty Russell { 3680ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 36812d3854a3SRusty Russell 3682ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 3683ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 3684ed48ece2STejun Heo flush_work(&wfc.work); 36852d3854a3SRusty Russell return wfc.ret; 36862d3854a3SRusty Russell } 36872d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 36882d3854a3SRusty Russell #endif /* CONFIG_SMP */ 36892d3854a3SRusty Russell 3690a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3691e7577c50SRusty Russell 3692a0a1a5fdSTejun Heo /** 3693a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3694a0a1a5fdSTejun Heo * 369558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 369658a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 369758a69cb4STejun Heo * gcwq->worklist. 3698a0a1a5fdSTejun Heo * 3699a0a1a5fdSTejun Heo * CONTEXT: 37008b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3701a0a1a5fdSTejun Heo */ 3702a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3703a0a1a5fdSTejun Heo { 3704a0a1a5fdSTejun Heo unsigned int cpu; 3705a0a1a5fdSTejun Heo 3706a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3707a0a1a5fdSTejun Heo 3708a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3709a0a1a5fdSTejun Heo workqueue_freezing = true; 3710a0a1a5fdSTejun Heo 3711f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37128b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3713bdbc5dd7STejun Heo struct workqueue_struct *wq; 37148b03ae3cSTejun Heo 37158b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 37168b03ae3cSTejun Heo 3717db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3718db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3719db7bccf4STejun Heo 3720a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3721a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3722a0a1a5fdSTejun Heo 372358a69cb4STejun Heo if (cwq && wq->flags & WQ_FREEZABLE) 3724a0a1a5fdSTejun Heo cwq->max_active = 0; 37251da177e4SLinus Torvalds } 37268b03ae3cSTejun Heo 37278b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3728a0a1a5fdSTejun Heo } 3729a0a1a5fdSTejun Heo 3730a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3731a0a1a5fdSTejun Heo } 3732a0a1a5fdSTejun Heo 3733a0a1a5fdSTejun Heo /** 373458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3735a0a1a5fdSTejun Heo * 3736a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3737a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3738a0a1a5fdSTejun Heo * 3739a0a1a5fdSTejun Heo * CONTEXT: 3740a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3741a0a1a5fdSTejun Heo * 3742a0a1a5fdSTejun Heo * RETURNS: 374358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 374458a69cb4STejun Heo * is complete. 3745a0a1a5fdSTejun Heo */ 3746a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3747a0a1a5fdSTejun Heo { 3748a0a1a5fdSTejun Heo unsigned int cpu; 3749a0a1a5fdSTejun Heo bool busy = false; 3750a0a1a5fdSTejun Heo 3751a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3752a0a1a5fdSTejun Heo 3753a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3754a0a1a5fdSTejun Heo 3755f3421797STejun Heo for_each_gcwq_cpu(cpu) { 3756bdbc5dd7STejun Heo struct workqueue_struct *wq; 3757a0a1a5fdSTejun Heo /* 3758a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3759a0a1a5fdSTejun Heo * to peek without lock. 3760a0a1a5fdSTejun Heo */ 3761a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3762a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3763a0a1a5fdSTejun Heo 376458a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3765a0a1a5fdSTejun Heo continue; 3766a0a1a5fdSTejun Heo 3767a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3768a0a1a5fdSTejun Heo if (cwq->nr_active) { 3769a0a1a5fdSTejun Heo busy = true; 3770a0a1a5fdSTejun Heo goto out_unlock; 3771a0a1a5fdSTejun Heo } 3772a0a1a5fdSTejun Heo } 3773a0a1a5fdSTejun Heo } 3774a0a1a5fdSTejun Heo out_unlock: 3775a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3776a0a1a5fdSTejun Heo return busy; 3777a0a1a5fdSTejun Heo } 3778a0a1a5fdSTejun Heo 3779a0a1a5fdSTejun Heo /** 3780a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3781a0a1a5fdSTejun Heo * 3782a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 37837e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3784a0a1a5fdSTejun Heo * 3785a0a1a5fdSTejun Heo * CONTEXT: 37868b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3787a0a1a5fdSTejun Heo */ 3788a0a1a5fdSTejun Heo void thaw_workqueues(void) 3789a0a1a5fdSTejun Heo { 3790a0a1a5fdSTejun Heo unsigned int cpu; 3791a0a1a5fdSTejun Heo 3792a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3793a0a1a5fdSTejun Heo 3794a0a1a5fdSTejun Heo if (!workqueue_freezing) 3795a0a1a5fdSTejun Heo goto out_unlock; 3796a0a1a5fdSTejun Heo 3797f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37988b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 37994ce62e9eSTejun Heo struct worker_pool *pool; 3800bdbc5dd7STejun Heo struct workqueue_struct *wq; 38018b03ae3cSTejun Heo 38028b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 38038b03ae3cSTejun Heo 3804db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3805db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3806db7bccf4STejun Heo 3807a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3808a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3809a0a1a5fdSTejun Heo 381058a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3811a0a1a5fdSTejun Heo continue; 3812a0a1a5fdSTejun Heo 3813a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 38149f4bd4cdSLai Jiangshan cwq_set_max_active(cwq, wq->saved_max_active); 3815a0a1a5fdSTejun Heo } 38168b03ae3cSTejun Heo 38174ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 38184ce62e9eSTejun Heo wake_up_worker(pool); 3819e22bee78STejun Heo 38208b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3821a0a1a5fdSTejun Heo } 3822a0a1a5fdSTejun Heo 3823a0a1a5fdSTejun Heo workqueue_freezing = false; 3824a0a1a5fdSTejun Heo out_unlock: 3825a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3826a0a1a5fdSTejun Heo } 3827a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3828a0a1a5fdSTejun Heo 38296ee0578bSSuresh Siddha static int __init init_workqueues(void) 38301da177e4SLinus Torvalds { 3831c34056a3STejun Heo unsigned int cpu; 3832c8e55f36STejun Heo int i; 3833c34056a3STejun Heo 3834b5490077STejun Heo /* make sure we have enough bits for OFFQ CPU number */ 3835b5490077STejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) < 3836b5490077STejun Heo WORK_CPU_LAST); 3837b5490077STejun Heo 383865758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 3839a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 38408b03ae3cSTejun Heo 38418b03ae3cSTejun Heo /* initialize gcwqs */ 3842f3421797STejun Heo for_each_gcwq_cpu(cpu) { 38438b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38444ce62e9eSTejun Heo struct worker_pool *pool; 38458b03ae3cSTejun Heo 38468b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 38478b03ae3cSTejun Heo gcwq->cpu = cpu; 3848f3421797STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 38498b03ae3cSTejun Heo 3850c8e55f36STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) 3851c8e55f36STejun Heo INIT_HLIST_HEAD(&gcwq->busy_hash[i]); 3852c8e55f36STejun Heo 38534ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 38544ce62e9eSTejun Heo pool->gcwq = gcwq; 38554ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 38564ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3857e22bee78STejun Heo 38584ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 38594ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 38604ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3861e22bee78STejun Heo 38624ce62e9eSTejun Heo setup_timer(&pool->mayday_timer, gcwq_mayday_timeout, 38634ce62e9eSTejun Heo (unsigned long)pool); 38644ce62e9eSTejun Heo 3865b2eb83d1SLai Jiangshan mutex_init(&pool->assoc_mutex); 38664ce62e9eSTejun Heo ida_init(&pool->worker_ida); 38674ce62e9eSTejun Heo } 38688b03ae3cSTejun Heo } 38698b03ae3cSTejun Heo 3870e22bee78STejun Heo /* create the initial worker */ 3871f3421797STejun Heo for_each_online_gcwq_cpu(cpu) { 3872e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38734ce62e9eSTejun Heo struct worker_pool *pool; 3874e22bee78STejun Heo 3875477a3c33STejun Heo if (cpu != WORK_CPU_UNBOUND) 3876477a3c33STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 38774ce62e9eSTejun Heo 38784ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 38794ce62e9eSTejun Heo struct worker *worker; 38804ce62e9eSTejun Heo 3881bc2ae0f5STejun Heo worker = create_worker(pool); 3882e22bee78STejun Heo BUG_ON(!worker); 3883e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3884e22bee78STejun Heo start_worker(worker); 3885e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3886e22bee78STejun Heo } 38874ce62e9eSTejun Heo } 3888e22bee78STejun Heo 3889d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 38901aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 3891d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3892f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3893f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 389424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 389524d51addSTejun Heo WQ_FREEZABLE, 0); 38961aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 3897ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 38986ee0578bSSuresh Siddha return 0; 38991da177e4SLinus Torvalds } 39006ee0578bSSuresh Siddha early_initcall(init_workqueues); 3901