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 261da177e4SLinus Torvalds #include <linux/module.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 { 48db7bccf4STejun Heo /* global_cwq flags */ 49e22bee78STejun Heo GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 50e22bee78STejun Heo GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */ 51e22bee78STejun Heo GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 52db7bccf4STejun Heo GCWQ_FREEZING = 1 << 3, /* freeze in progress */ 53649027d7STejun Heo GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */ 54db7bccf4STejun Heo 55c8e55f36STejun Heo /* worker flags */ 56c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 57c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 58c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 59e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 60db7bccf4STejun Heo WORKER_ROGUE = 1 << 4, /* not bound to any cpu */ 61e22bee78STejun Heo WORKER_REBIND = 1 << 5, /* mom is home, come back */ 62fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 63f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 64e22bee78STejun Heo 65fb0e7bebSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND | 66f3421797STejun Heo WORKER_CPU_INTENSIVE | WORKER_UNBOUND, 67db7bccf4STejun Heo 68db7bccf4STejun Heo /* gcwq->trustee_state */ 69db7bccf4STejun Heo TRUSTEE_START = 0, /* start */ 70db7bccf4STejun Heo TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */ 71db7bccf4STejun Heo TRUSTEE_BUTCHER = 2, /* butcher workers */ 72db7bccf4STejun Heo TRUSTEE_RELEASE = 3, /* release workers */ 73db7bccf4STejun Heo TRUSTEE_DONE = 4, /* trustee is done */ 74c8e55f36STejun Heo 75c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 76c8e55f36STejun Heo BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, 77c8e55f36STejun Heo BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, 78db7bccf4STejun Heo 79e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 80e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 81e22bee78STejun Heo 823233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 833233cdbdSTejun Heo /* call for help after 10ms 843233cdbdSTejun Heo (min two ticks) */ 85e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 86e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 87db7bccf4STejun Heo TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */ 881da177e4SLinus Torvalds 891da177e4SLinus Torvalds /* 90e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 91e22bee78STejun Heo * all cpus. Give -20. 92e22bee78STejun Heo */ 93e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 94c8e55f36STejun Heo }; 95c8e55f36STejun Heo 961da177e4SLinus Torvalds /* 974690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 984690c4abSTejun Heo * 99e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 100e41e704bSTejun Heo * everyone else. 1014690c4abSTejun Heo * 102e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 103e22bee78STejun Heo * only be modified and accessed from the local cpu. 104e22bee78STejun Heo * 1058b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 1064690c4abSTejun Heo * 107e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 108e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 109e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 110f3421797STejun Heo * If GCWQ_DISASSOCIATED is set, it's identical to L. 111e22bee78STejun Heo * 11273f53c4aSTejun Heo * F: wq->flush_mutex protected. 11373f53c4aSTejun Heo * 1144690c4abSTejun Heo * W: workqueue_lock protected. 1154690c4abSTejun Heo */ 1164690c4abSTejun Heo 1178b03ae3cSTejun Heo struct global_cwq; 118c34056a3STejun Heo 119e22bee78STejun Heo /* 120e22bee78STejun Heo * The poor guys doing the actual heavy lifting. All on-duty workers 121e22bee78STejun Heo * are either serving the manager role, on idle list or on busy hash. 122e22bee78STejun Heo */ 123c34056a3STejun Heo struct worker { 124c8e55f36STejun Heo /* on idle list while idle, on busy hash table while busy */ 125c8e55f36STejun Heo union { 126c8e55f36STejun Heo struct list_head entry; /* L: while idle */ 127c8e55f36STejun Heo struct hlist_node hentry; /* L: while busy */ 128c8e55f36STejun Heo }; 129c8e55f36STejun Heo 130c34056a3STejun Heo struct work_struct *current_work; /* L: work being processed */ 1318cca0eeaSTejun Heo struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ 132affee4b2STejun Heo struct list_head scheduled; /* L: scheduled works */ 133c34056a3STejun Heo struct task_struct *task; /* I: worker task */ 1348b03ae3cSTejun Heo struct global_cwq *gcwq; /* I: the associated gcwq */ 135e22bee78STejun Heo /* 64 bytes boundary on 64bit, 32 on 32bit */ 136e22bee78STejun Heo unsigned long last_active; /* L: last active timestamp */ 137e22bee78STejun Heo unsigned int flags; /* X: flags */ 138c34056a3STejun Heo int id; /* I: worker id */ 139e22bee78STejun Heo struct work_struct rebind_work; /* L: rebind worker to cpu */ 140c34056a3STejun Heo }; 141c34056a3STejun Heo 1424690c4abSTejun Heo /* 143e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 144e22bee78STejun Heo * and all works are queued and processed here regardless of their 145e22bee78STejun Heo * target workqueues. 1468b03ae3cSTejun Heo */ 1478b03ae3cSTejun Heo struct global_cwq { 1488b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1497e11629dSTejun Heo struct list_head worklist; /* L: list of pending works */ 1508b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 151db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 152c8e55f36STejun Heo 153c8e55f36STejun Heo int nr_workers; /* L: total number of workers */ 154c8e55f36STejun Heo int nr_idle; /* L: currently idle ones */ 155c8e55f36STejun Heo 156c8e55f36STejun Heo /* workers are chained either in the idle_list or busy_hash */ 157e22bee78STejun Heo struct list_head idle_list; /* X: list of idle workers */ 158c8e55f36STejun Heo struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; 159c8e55f36STejun Heo /* L: hash of busy workers */ 160c8e55f36STejun Heo 161e22bee78STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 162e22bee78STejun Heo struct timer_list mayday_timer; /* L: SOS timer for dworkers */ 163e22bee78STejun Heo 1648b03ae3cSTejun Heo struct ida worker_ida; /* L: for worker IDs */ 165db7bccf4STejun Heo 166db7bccf4STejun Heo struct task_struct *trustee; /* L: for gcwq shutdown */ 167db7bccf4STejun Heo unsigned int trustee_state; /* L: trustee state */ 168db7bccf4STejun Heo wait_queue_head_t trustee_wait; /* trustee wait */ 169e22bee78STejun Heo struct worker *first_idle; /* L: first idle worker */ 1708b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1718b03ae3cSTejun Heo 1728b03ae3cSTejun Heo /* 173502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1740f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1750f900049STejun Heo * aligned at two's power of the number of flag bits. 1761da177e4SLinus Torvalds */ 1771da177e4SLinus Torvalds struct cpu_workqueue_struct { 1788b03ae3cSTejun Heo struct global_cwq *gcwq; /* I: the associated gcwq */ 1794690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 18073f53c4aSTejun Heo int work_color; /* L: current color */ 18173f53c4aSTejun Heo int flush_color; /* L: flushing color */ 18273f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 18373f53c4aSTejun Heo /* L: nr of in_flight works */ 1841e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 185a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1861e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1870f900049STejun Heo }; 1881da177e4SLinus Torvalds 1891da177e4SLinus Torvalds /* 19073f53c4aSTejun Heo * Structure used to wait for workqueue flush. 19173f53c4aSTejun Heo */ 19273f53c4aSTejun Heo struct wq_flusher { 19373f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 19473f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 19573f53c4aSTejun Heo struct completion done; /* flush completion */ 19673f53c4aSTejun Heo }; 1971da177e4SLinus Torvalds 19873f53c4aSTejun Heo /* 199f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 200f2e005aaSTejun Heo * used to determine whether there's something to be done. 201f2e005aaSTejun Heo */ 202f2e005aaSTejun Heo #ifdef CONFIG_SMP 203f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 204f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 205f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 206f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 207f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 2089c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 209f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 210f2e005aaSTejun Heo #else 211f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 212f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 213f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 214f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 215f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 216f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 217f2e005aaSTejun Heo #endif 2181da177e4SLinus Torvalds 2191da177e4SLinus Torvalds /* 2201da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2211da177e4SLinus Torvalds * per-CPU workqueues: 2221da177e4SLinus Torvalds */ 2231da177e4SLinus Torvalds struct workqueue_struct { 2249c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 225bdbc5dd7STejun Heo union { 226bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 227bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 228bdbc5dd7STejun Heo unsigned long v; 229bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2304690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 23173f53c4aSTejun Heo 23273f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 23373f53c4aSTejun Heo int work_color; /* F: current work color */ 23473f53c4aSTejun Heo int flush_color; /* F: current flush color */ 23573f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 23673f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 23773f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 23873f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 23973f53c4aSTejun Heo 240f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 241e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 242e22bee78STejun Heo 2439c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 244dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2454690c4abSTejun Heo const char *name; /* I: workqueue name */ 2464e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2474e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2484e6045f1SJohannes Berg #endif 2491da177e4SLinus Torvalds }; 2501da177e4SLinus Torvalds 251d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 252d320c038STejun Heo struct workqueue_struct *system_long_wq __read_mostly; 253d320c038STejun Heo struct workqueue_struct *system_nrt_wq __read_mostly; 254f3421797STejun Heo struct workqueue_struct *system_unbound_wq __read_mostly; 25524d51addSTejun Heo struct workqueue_struct *system_freezable_wq __read_mostly; 256d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 257d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 258d320c038STejun Heo EXPORT_SYMBOL_GPL(system_nrt_wq); 259f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 26024d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 261d320c038STejun Heo 26297bd2347STejun Heo #define CREATE_TRACE_POINTS 26397bd2347STejun Heo #include <trace/events/workqueue.h> 26497bd2347STejun Heo 265db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 266db7bccf4STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ 267db7bccf4STejun Heo hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) 268db7bccf4STejun Heo 269f3421797STejun Heo static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, 270f3421797STejun Heo unsigned int sw) 271f3421797STejun Heo { 272f3421797STejun Heo if (cpu < nr_cpu_ids) { 273f3421797STejun Heo if (sw & 1) { 274f3421797STejun Heo cpu = cpumask_next(cpu, mask); 275f3421797STejun Heo if (cpu < nr_cpu_ids) 276f3421797STejun Heo return cpu; 277f3421797STejun Heo } 278f3421797STejun Heo if (sw & 2) 279f3421797STejun Heo return WORK_CPU_UNBOUND; 280f3421797STejun Heo } 281f3421797STejun Heo return WORK_CPU_NONE; 282f3421797STejun Heo } 283f3421797STejun Heo 284f3421797STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 285f3421797STejun Heo struct workqueue_struct *wq) 286f3421797STejun Heo { 287f3421797STejun Heo return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 288f3421797STejun Heo } 289f3421797STejun Heo 29009884951STejun Heo /* 29109884951STejun Heo * CPU iterators 29209884951STejun Heo * 29309884951STejun Heo * An extra gcwq is defined for an invalid cpu number 29409884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 29509884951STejun Heo * specific CPU. The following iterators are similar to 29609884951STejun Heo * for_each_*_cpu() iterators but also considers the unbound gcwq. 29709884951STejun Heo * 29809884951STejun Heo * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND 29909884951STejun Heo * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND 30009884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 30109884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 30209884951STejun Heo */ 303f3421797STejun Heo #define for_each_gcwq_cpu(cpu) \ 304f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ 305f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 306f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) 307f3421797STejun Heo 308f3421797STejun Heo #define for_each_online_gcwq_cpu(cpu) \ 309f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ 310f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 311f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) 312f3421797STejun Heo 313f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 314f3421797STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ 315f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 316f3421797STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) 317f3421797STejun Heo 318dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 319dc186ad7SThomas Gleixner 320dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 321dc186ad7SThomas Gleixner 32299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 32399777288SStanislaw Gruszka { 32499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 32599777288SStanislaw Gruszka } 32699777288SStanislaw Gruszka 327dc186ad7SThomas Gleixner /* 328dc186ad7SThomas Gleixner * fixup_init is called when: 329dc186ad7SThomas Gleixner * - an active object is initialized 330dc186ad7SThomas Gleixner */ 331dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 332dc186ad7SThomas Gleixner { 333dc186ad7SThomas Gleixner struct work_struct *work = addr; 334dc186ad7SThomas Gleixner 335dc186ad7SThomas Gleixner switch (state) { 336dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 337dc186ad7SThomas Gleixner cancel_work_sync(work); 338dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 339dc186ad7SThomas Gleixner return 1; 340dc186ad7SThomas Gleixner default: 341dc186ad7SThomas Gleixner return 0; 342dc186ad7SThomas Gleixner } 343dc186ad7SThomas Gleixner } 344dc186ad7SThomas Gleixner 345dc186ad7SThomas Gleixner /* 346dc186ad7SThomas Gleixner * fixup_activate is called when: 347dc186ad7SThomas Gleixner * - an active object is activated 348dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 349dc186ad7SThomas Gleixner */ 350dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 351dc186ad7SThomas Gleixner { 352dc186ad7SThomas Gleixner struct work_struct *work = addr; 353dc186ad7SThomas Gleixner 354dc186ad7SThomas Gleixner switch (state) { 355dc186ad7SThomas Gleixner 356dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 357dc186ad7SThomas Gleixner /* 358dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 359dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 360dc186ad7SThomas Gleixner * is tracked in the object tracker. 361dc186ad7SThomas Gleixner */ 36222df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 363dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 364dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 365dc186ad7SThomas Gleixner return 0; 366dc186ad7SThomas Gleixner } 367dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 368dc186ad7SThomas Gleixner return 0; 369dc186ad7SThomas Gleixner 370dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 371dc186ad7SThomas Gleixner WARN_ON(1); 372dc186ad7SThomas Gleixner 373dc186ad7SThomas Gleixner default: 374dc186ad7SThomas Gleixner return 0; 375dc186ad7SThomas Gleixner } 376dc186ad7SThomas Gleixner } 377dc186ad7SThomas Gleixner 378dc186ad7SThomas Gleixner /* 379dc186ad7SThomas Gleixner * fixup_free is called when: 380dc186ad7SThomas Gleixner * - an active object is freed 381dc186ad7SThomas Gleixner */ 382dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 383dc186ad7SThomas Gleixner { 384dc186ad7SThomas Gleixner struct work_struct *work = addr; 385dc186ad7SThomas Gleixner 386dc186ad7SThomas Gleixner switch (state) { 387dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 388dc186ad7SThomas Gleixner cancel_work_sync(work); 389dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 390dc186ad7SThomas Gleixner return 1; 391dc186ad7SThomas Gleixner default: 392dc186ad7SThomas Gleixner return 0; 393dc186ad7SThomas Gleixner } 394dc186ad7SThomas Gleixner } 395dc186ad7SThomas Gleixner 396dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 397dc186ad7SThomas Gleixner .name = "work_struct", 39899777288SStanislaw Gruszka .debug_hint = work_debug_hint, 399dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 400dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 401dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 402dc186ad7SThomas Gleixner }; 403dc186ad7SThomas Gleixner 404dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 405dc186ad7SThomas Gleixner { 406dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 407dc186ad7SThomas Gleixner } 408dc186ad7SThomas Gleixner 409dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 410dc186ad7SThomas Gleixner { 411dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 412dc186ad7SThomas Gleixner } 413dc186ad7SThomas Gleixner 414dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 415dc186ad7SThomas Gleixner { 416dc186ad7SThomas Gleixner if (onstack) 417dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 418dc186ad7SThomas Gleixner else 419dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 420dc186ad7SThomas Gleixner } 421dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 422dc186ad7SThomas Gleixner 423dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 424dc186ad7SThomas Gleixner { 425dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 426dc186ad7SThomas Gleixner } 427dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 428dc186ad7SThomas Gleixner 429dc186ad7SThomas Gleixner #else 430dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 431dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 432dc186ad7SThomas Gleixner #endif 433dc186ad7SThomas Gleixner 43495402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 43595402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4361da177e4SLinus Torvalds static LIST_HEAD(workqueues); 437a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4381da177e4SLinus Torvalds 43914441960SOleg Nesterov /* 440e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 441e22bee78STejun Heo * which is expected to be used frequently by other cpus via 442e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 44314441960SOleg Nesterov */ 4448b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 445e22bee78STejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running); 446f756d5e2SNathan Lynch 447f3421797STejun Heo /* 448f3421797STejun Heo * Global cpu workqueue and nr_running counter for unbound gcwq. The 449f3421797STejun Heo * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its 450f3421797STejun Heo * workers have WORKER_UNBOUND set. 451f3421797STejun Heo */ 452f3421797STejun Heo static struct global_cwq unbound_global_cwq; 453f3421797STejun Heo static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */ 454f3421797STejun Heo 455c34056a3STejun Heo static int worker_thread(void *__worker); 4561da177e4SLinus Torvalds 4578b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 4581da177e4SLinus Torvalds { 459f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 4608b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 461f3421797STejun Heo else 462f3421797STejun Heo return &unbound_global_cwq; 4631da177e4SLinus Torvalds } 4641da177e4SLinus Torvalds 465e22bee78STejun Heo static atomic_t *get_gcwq_nr_running(unsigned int cpu) 466b1f4ec17SOleg Nesterov { 467f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 468e22bee78STejun Heo return &per_cpu(gcwq_nr_running, cpu); 469f3421797STejun Heo else 470f3421797STejun Heo return &unbound_gcwq_nr_running; 471b1f4ec17SOleg Nesterov } 472b1f4ec17SOleg Nesterov 4734690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 4744690c4abSTejun Heo struct workqueue_struct *wq) 475a848e3b6SOleg Nesterov { 476f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 477f3421797STejun Heo if (likely(cpu < nr_cpu_ids)) { 478f3421797STejun Heo #ifdef CONFIG_SMP 479bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 480f3421797STejun Heo #else 481f3421797STejun Heo return wq->cpu_wq.single; 482bdbc5dd7STejun Heo #endif 483a848e3b6SOleg Nesterov } 484f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 485f3421797STejun Heo return wq->cpu_wq.single; 486f3421797STejun Heo return NULL; 487f3421797STejun Heo } 488a848e3b6SOleg Nesterov 48973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 49073f53c4aSTejun Heo { 49173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 49273f53c4aSTejun Heo } 49373f53c4aSTejun Heo 49473f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 49573f53c4aSTejun Heo { 49673f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 49773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 49873f53c4aSTejun Heo } 49973f53c4aSTejun Heo 50073f53c4aSTejun Heo static int work_next_color(int color) 50173f53c4aSTejun Heo { 50273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 5031da177e4SLinus Torvalds } 5041da177e4SLinus Torvalds 5054594bf15SDavid Howells /* 506e120153dSTejun Heo * A work's data points to the cwq with WORK_STRUCT_CWQ set while the 507e120153dSTejun Heo * work is on queue. Once execution starts, WORK_STRUCT_CWQ is 508e120153dSTejun Heo * cleared and the work data contains the cpu number it was last on. 5097a22ad75STejun Heo * 5107a22ad75STejun Heo * set_work_{cwq|cpu}() and clear_work_data() can be used to set the 5117a22ad75STejun Heo * cwq, cpu or clear work->data. These functions should only be 5127a22ad75STejun Heo * called while the work is owned - ie. while the PENDING bit is set. 5137a22ad75STejun Heo * 5147a22ad75STejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq 5157a22ad75STejun Heo * corresponding to a work. gcwq is available once the work has been 5167a22ad75STejun Heo * queued anywhere after initialization. cwq is available only from 5177a22ad75STejun Heo * queueing until execution starts. 5184594bf15SDavid Howells */ 5197a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5207a22ad75STejun Heo unsigned long flags) 5217a22ad75STejun Heo { 5227a22ad75STejun Heo BUG_ON(!work_pending(work)); 5237a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5247a22ad75STejun Heo } 5257a22ad75STejun Heo 5267a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5274690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5284690c4abSTejun Heo unsigned long extra_flags) 529365970a1SDavid Howells { 5307a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 531e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 532365970a1SDavid Howells } 533365970a1SDavid Howells 5347a22ad75STejun Heo static void set_work_cpu(struct work_struct *work, unsigned int cpu) 5354d707b9fSOleg Nesterov { 5367a22ad75STejun Heo set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING); 5374d707b9fSOleg Nesterov } 5384d707b9fSOleg Nesterov 5397a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 540365970a1SDavid Howells { 5417a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 5427a22ad75STejun Heo } 5437a22ad75STejun Heo 5447a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5457a22ad75STejun Heo { 546e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5477a22ad75STejun Heo 548e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 549e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 550e120153dSTejun Heo else 551e120153dSTejun Heo return NULL; 5527a22ad75STejun Heo } 5537a22ad75STejun Heo 5547a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 5557a22ad75STejun Heo { 556e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5577a22ad75STejun Heo unsigned int cpu; 5587a22ad75STejun Heo 559e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 560e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 561e120153dSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq; 5627a22ad75STejun Heo 5637a22ad75STejun Heo cpu = data >> WORK_STRUCT_FLAG_BITS; 564bdbc5dd7STejun Heo if (cpu == WORK_CPU_NONE) 5657a22ad75STejun Heo return NULL; 5667a22ad75STejun Heo 567f3421797STejun Heo BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); 5687a22ad75STejun Heo return get_gcwq(cpu); 569365970a1SDavid Howells } 570365970a1SDavid Howells 571e22bee78STejun Heo /* 572e22bee78STejun Heo * Policy functions. These define the policies on how the global 573e22bee78STejun Heo * worker pool is managed. Unless noted otherwise, these functions 574e22bee78STejun Heo * assume that they're being called with gcwq->lock held. 575e22bee78STejun Heo */ 576e22bee78STejun Heo 577649027d7STejun Heo static bool __need_more_worker(struct global_cwq *gcwq) 578649027d7STejun Heo { 579649027d7STejun Heo return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) || 580649027d7STejun Heo gcwq->flags & GCWQ_HIGHPRI_PENDING; 581649027d7STejun Heo } 582649027d7STejun Heo 583e22bee78STejun Heo /* 584e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 585e22bee78STejun Heo * running workers. 586e22bee78STejun Heo */ 587e22bee78STejun Heo static bool need_more_worker(struct global_cwq *gcwq) 588e22bee78STejun Heo { 589649027d7STejun Heo return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq); 590e22bee78STejun Heo } 591e22bee78STejun Heo 592e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 593e22bee78STejun Heo static bool may_start_working(struct global_cwq *gcwq) 594e22bee78STejun Heo { 595e22bee78STejun Heo return gcwq->nr_idle; 596e22bee78STejun Heo } 597e22bee78STejun Heo 598e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 599e22bee78STejun Heo static bool keep_working(struct global_cwq *gcwq) 600e22bee78STejun Heo { 601e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 602e22bee78STejun Heo 60330310045STejun Heo return !list_empty(&gcwq->worklist) && 60430310045STejun Heo (atomic_read(nr_running) <= 1 || 60530310045STejun Heo gcwq->flags & GCWQ_HIGHPRI_PENDING); 606e22bee78STejun Heo } 607e22bee78STejun Heo 608e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 609e22bee78STejun Heo static bool need_to_create_worker(struct global_cwq *gcwq) 610e22bee78STejun Heo { 611e22bee78STejun Heo return need_more_worker(gcwq) && !may_start_working(gcwq); 612e22bee78STejun Heo } 613e22bee78STejun Heo 614e22bee78STejun Heo /* Do I need to be the manager? */ 615e22bee78STejun Heo static bool need_to_manage_workers(struct global_cwq *gcwq) 616e22bee78STejun Heo { 617e22bee78STejun Heo return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS; 618e22bee78STejun Heo } 619e22bee78STejun Heo 620e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 621e22bee78STejun Heo static bool too_many_workers(struct global_cwq *gcwq) 622e22bee78STejun Heo { 623e22bee78STejun Heo bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS; 624e22bee78STejun Heo int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */ 625e22bee78STejun Heo int nr_busy = gcwq->nr_workers - nr_idle; 626e22bee78STejun Heo 627e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 628e22bee78STejun Heo } 629e22bee78STejun Heo 630e22bee78STejun Heo /* 631e22bee78STejun Heo * Wake up functions. 632e22bee78STejun Heo */ 633e22bee78STejun Heo 6347e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 6357e11629dSTejun Heo static struct worker *first_worker(struct global_cwq *gcwq) 6367e11629dSTejun Heo { 6377e11629dSTejun Heo if (unlikely(list_empty(&gcwq->idle_list))) 6387e11629dSTejun Heo return NULL; 6397e11629dSTejun Heo 6407e11629dSTejun Heo return list_first_entry(&gcwq->idle_list, struct worker, entry); 6417e11629dSTejun Heo } 6427e11629dSTejun Heo 6437e11629dSTejun Heo /** 6447e11629dSTejun Heo * wake_up_worker - wake up an idle worker 6457e11629dSTejun Heo * @gcwq: gcwq to wake worker for 6467e11629dSTejun Heo * 6477e11629dSTejun Heo * Wake up the first idle worker of @gcwq. 6487e11629dSTejun Heo * 6497e11629dSTejun Heo * CONTEXT: 6507e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 6517e11629dSTejun Heo */ 6527e11629dSTejun Heo static void wake_up_worker(struct global_cwq *gcwq) 6537e11629dSTejun Heo { 6547e11629dSTejun Heo struct worker *worker = first_worker(gcwq); 6557e11629dSTejun Heo 6567e11629dSTejun Heo if (likely(worker)) 6577e11629dSTejun Heo wake_up_process(worker->task); 6587e11629dSTejun Heo } 6597e11629dSTejun Heo 6604690c4abSTejun Heo /** 661e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 662e22bee78STejun Heo * @task: task waking up 663e22bee78STejun Heo * @cpu: CPU @task is waking up to 664e22bee78STejun Heo * 665e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 666e22bee78STejun Heo * being awoken. 667e22bee78STejun Heo * 668e22bee78STejun Heo * CONTEXT: 669e22bee78STejun Heo * spin_lock_irq(rq->lock) 670e22bee78STejun Heo */ 671e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 672e22bee78STejun Heo { 673e22bee78STejun Heo struct worker *worker = kthread_data(task); 674e22bee78STejun Heo 6752d64672eSSteven Rostedt if (!(worker->flags & WORKER_NOT_RUNNING)) 676e22bee78STejun Heo atomic_inc(get_gcwq_nr_running(cpu)); 677e22bee78STejun Heo } 678e22bee78STejun Heo 679e22bee78STejun Heo /** 680e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 681e22bee78STejun Heo * @task: task going to sleep 682e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 683e22bee78STejun Heo * 684e22bee78STejun Heo * This function is called during schedule() when a busy worker is 685e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 686e22bee78STejun Heo * returning pointer to its task. 687e22bee78STejun Heo * 688e22bee78STejun Heo * CONTEXT: 689e22bee78STejun Heo * spin_lock_irq(rq->lock) 690e22bee78STejun Heo * 691e22bee78STejun Heo * RETURNS: 692e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 693e22bee78STejun Heo */ 694e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 695e22bee78STejun Heo unsigned int cpu) 696e22bee78STejun Heo { 697e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 698e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 699e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(cpu); 700e22bee78STejun Heo 7012d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 702e22bee78STejun Heo return NULL; 703e22bee78STejun Heo 704e22bee78STejun Heo /* this can only happen on the local cpu */ 705e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 706e22bee78STejun Heo 707e22bee78STejun Heo /* 708e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 709e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 710e22bee78STejun Heo * Please read comment there. 711e22bee78STejun Heo * 712e22bee78STejun Heo * NOT_RUNNING is clear. This means that trustee is not in 713e22bee78STejun Heo * charge and we're running on the local cpu w/ rq lock held 714e22bee78STejun Heo * and preemption disabled, which in turn means that none else 715e22bee78STejun Heo * could be manipulating idle_list, so dereferencing idle_list 716e22bee78STejun Heo * without gcwq lock is safe. 717e22bee78STejun Heo */ 718e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist)) 719e22bee78STejun Heo to_wakeup = first_worker(gcwq); 720e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 721e22bee78STejun Heo } 722e22bee78STejun Heo 723e22bee78STejun Heo /** 724e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 725cb444766STejun Heo * @worker: self 726d302f017STejun Heo * @flags: flags to set 727d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 728d302f017STejun Heo * 729e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 730e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 731e22bee78STejun Heo * woken up. 732d302f017STejun Heo * 733cb444766STejun Heo * CONTEXT: 734cb444766STejun Heo * spin_lock_irq(gcwq->lock) 735d302f017STejun Heo */ 736d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 737d302f017STejun Heo bool wakeup) 738d302f017STejun Heo { 739e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 740e22bee78STejun Heo 741cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 742cb444766STejun Heo 743e22bee78STejun Heo /* 744e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 745e22bee78STejun Heo * wake up an idle worker as necessary if requested by 746e22bee78STejun Heo * @wakeup. 747e22bee78STejun Heo */ 748e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 749e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 750e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 751e22bee78STejun Heo 752e22bee78STejun Heo if (wakeup) { 753e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 754e22bee78STejun Heo !list_empty(&gcwq->worklist)) 755e22bee78STejun Heo wake_up_worker(gcwq); 756e22bee78STejun Heo } else 757e22bee78STejun Heo atomic_dec(nr_running); 758e22bee78STejun Heo } 759e22bee78STejun Heo 760d302f017STejun Heo worker->flags |= flags; 761d302f017STejun Heo } 762d302f017STejun Heo 763d302f017STejun Heo /** 764e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 765cb444766STejun Heo * @worker: self 766d302f017STejun Heo * @flags: flags to clear 767d302f017STejun Heo * 768e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 769d302f017STejun Heo * 770cb444766STejun Heo * CONTEXT: 771cb444766STejun Heo * spin_lock_irq(gcwq->lock) 772d302f017STejun Heo */ 773d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 774d302f017STejun Heo { 775e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 776e22bee78STejun Heo unsigned int oflags = worker->flags; 777e22bee78STejun Heo 778cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 779cb444766STejun Heo 780d302f017STejun Heo worker->flags &= ~flags; 781e22bee78STejun Heo 78242c025f3STejun Heo /* 78342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 78442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 78542c025f3STejun Heo * of multiple flags, not a single flag. 78642c025f3STejun Heo */ 787e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 788e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 789e22bee78STejun Heo atomic_inc(get_gcwq_nr_running(gcwq->cpu)); 790d302f017STejun Heo } 791d302f017STejun Heo 792d302f017STejun Heo /** 793c8e55f36STejun Heo * busy_worker_head - return the busy hash head for a work 794c8e55f36STejun Heo * @gcwq: gcwq of interest 795c8e55f36STejun Heo * @work: work to be hashed 796c8e55f36STejun Heo * 797c8e55f36STejun Heo * Return hash head of @gcwq for @work. 798c8e55f36STejun Heo * 799c8e55f36STejun Heo * CONTEXT: 800c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 801c8e55f36STejun Heo * 802c8e55f36STejun Heo * RETURNS: 803c8e55f36STejun Heo * Pointer to the hash head. 804c8e55f36STejun Heo */ 805c8e55f36STejun Heo static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, 806c8e55f36STejun Heo struct work_struct *work) 807c8e55f36STejun Heo { 808c8e55f36STejun Heo const int base_shift = ilog2(sizeof(struct work_struct)); 809c8e55f36STejun Heo unsigned long v = (unsigned long)work; 810c8e55f36STejun Heo 811c8e55f36STejun Heo /* simple shift and fold hash, do we need something better? */ 812c8e55f36STejun Heo v >>= base_shift; 813c8e55f36STejun Heo v += v >> BUSY_WORKER_HASH_ORDER; 814c8e55f36STejun Heo v &= BUSY_WORKER_HASH_MASK; 815c8e55f36STejun Heo 816c8e55f36STejun Heo return &gcwq->busy_hash[v]; 817c8e55f36STejun Heo } 818c8e55f36STejun Heo 819c8e55f36STejun Heo /** 8208cca0eeaSTejun Heo * __find_worker_executing_work - find worker which is executing a work 8218cca0eeaSTejun Heo * @gcwq: gcwq of interest 8228cca0eeaSTejun Heo * @bwh: hash head as returned by busy_worker_head() 8238cca0eeaSTejun Heo * @work: work to find worker for 8248cca0eeaSTejun Heo * 8258cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. @bwh should be 8268cca0eeaSTejun Heo * the hash head obtained by calling busy_worker_head() with the same 8278cca0eeaSTejun Heo * work. 8288cca0eeaSTejun Heo * 8298cca0eeaSTejun Heo * CONTEXT: 8308cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8318cca0eeaSTejun Heo * 8328cca0eeaSTejun Heo * RETURNS: 8338cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8348cca0eeaSTejun Heo * otherwise. 8358cca0eeaSTejun Heo */ 8368cca0eeaSTejun Heo static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, 8378cca0eeaSTejun Heo struct hlist_head *bwh, 8388cca0eeaSTejun Heo struct work_struct *work) 8398cca0eeaSTejun Heo { 8408cca0eeaSTejun Heo struct worker *worker; 8418cca0eeaSTejun Heo struct hlist_node *tmp; 8428cca0eeaSTejun Heo 8438cca0eeaSTejun Heo hlist_for_each_entry(worker, tmp, bwh, hentry) 8448cca0eeaSTejun Heo if (worker->current_work == work) 8458cca0eeaSTejun Heo return worker; 8468cca0eeaSTejun Heo return NULL; 8478cca0eeaSTejun Heo } 8488cca0eeaSTejun Heo 8498cca0eeaSTejun Heo /** 8508cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 8518cca0eeaSTejun Heo * @gcwq: gcwq of interest 8528cca0eeaSTejun Heo * @work: work to find worker for 8538cca0eeaSTejun Heo * 8548cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. This function is 8558cca0eeaSTejun Heo * identical to __find_worker_executing_work() except that this 8568cca0eeaSTejun Heo * function calculates @bwh itself. 8578cca0eeaSTejun Heo * 8588cca0eeaSTejun Heo * CONTEXT: 8598cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8608cca0eeaSTejun Heo * 8618cca0eeaSTejun Heo * RETURNS: 8628cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8638cca0eeaSTejun Heo * otherwise. 8648cca0eeaSTejun Heo */ 8658cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 8668cca0eeaSTejun Heo struct work_struct *work) 8678cca0eeaSTejun Heo { 8688cca0eeaSTejun Heo return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), 8698cca0eeaSTejun Heo work); 8708cca0eeaSTejun Heo } 8718cca0eeaSTejun Heo 8728cca0eeaSTejun Heo /** 873649027d7STejun Heo * gcwq_determine_ins_pos - find insertion position 874649027d7STejun Heo * @gcwq: gcwq of interest 875649027d7STejun Heo * @cwq: cwq a work is being queued for 876649027d7STejun Heo * 877649027d7STejun Heo * A work for @cwq is about to be queued on @gcwq, determine insertion 878649027d7STejun Heo * position for the work. If @cwq is for HIGHPRI wq, the work is 879649027d7STejun Heo * queued at the head of the queue but in FIFO order with respect to 880649027d7STejun Heo * other HIGHPRI works; otherwise, at the end of the queue. This 881649027d7STejun Heo * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that 882649027d7STejun Heo * there are HIGHPRI works pending. 883649027d7STejun Heo * 884649027d7STejun Heo * CONTEXT: 885649027d7STejun Heo * spin_lock_irq(gcwq->lock). 886649027d7STejun Heo * 887649027d7STejun Heo * RETURNS: 888649027d7STejun Heo * Pointer to inserstion position. 889649027d7STejun Heo */ 890649027d7STejun Heo static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq, 8911da177e4SLinus Torvalds struct cpu_workqueue_struct *cwq) 8921da177e4SLinus Torvalds { 893649027d7STejun Heo struct work_struct *twork; 8941da177e4SLinus Torvalds 895649027d7STejun Heo if (likely(!(cwq->wq->flags & WQ_HIGHPRI))) 896649027d7STejun Heo return &gcwq->worklist; 8971da177e4SLinus Torvalds 898649027d7STejun Heo list_for_each_entry(twork, &gcwq->worklist, entry) { 899649027d7STejun Heo struct cpu_workqueue_struct *tcwq = get_work_cwq(twork); 900649027d7STejun Heo 901649027d7STejun Heo if (!(tcwq->wq->flags & WQ_HIGHPRI)) 902649027d7STejun Heo break; 9031da177e4SLinus Torvalds } 9041da177e4SLinus Torvalds 905649027d7STejun Heo gcwq->flags |= GCWQ_HIGHPRI_PENDING; 906649027d7STejun Heo return &twork->entry; 907649027d7STejun Heo } 908649027d7STejun Heo 909649027d7STejun Heo /** 9107e11629dSTejun Heo * insert_work - insert a work into gcwq 9114690c4abSTejun Heo * @cwq: cwq @work belongs to 9124690c4abSTejun Heo * @work: work to insert 9134690c4abSTejun Heo * @head: insertion point 9144690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 9154690c4abSTejun Heo * 9167e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 9177e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 9184690c4abSTejun Heo * 9194690c4abSTejun Heo * CONTEXT: 9208b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 9211da177e4SLinus Torvalds */ 922b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 9234690c4abSTejun Heo struct work_struct *work, struct list_head *head, 9244690c4abSTejun Heo unsigned int extra_flags) 925b89deed3SOleg Nesterov { 926e22bee78STejun Heo struct global_cwq *gcwq = cwq->gcwq; 927e1d8aa9fSFrederic Weisbecker 9284690c4abSTejun Heo /* we own @work, set data and link */ 9297a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 9304690c4abSTejun Heo 9316e84d644SOleg Nesterov /* 9326e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 9336e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 9346e84d644SOleg Nesterov */ 9356e84d644SOleg Nesterov smp_wmb(); 9364690c4abSTejun Heo 9371a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 938e22bee78STejun Heo 939e22bee78STejun Heo /* 940e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 941e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 942e22bee78STejun Heo * lying around lazily while there are works to be processed. 943e22bee78STejun Heo */ 944e22bee78STejun Heo smp_mb(); 945e22bee78STejun Heo 946649027d7STejun Heo if (__need_more_worker(gcwq)) 947e22bee78STejun Heo wake_up_worker(gcwq); 948b89deed3SOleg Nesterov } 949b89deed3SOleg Nesterov 950c8efcc25STejun Heo /* 951c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 952c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 953c8efcc25STejun Heo * cold paths. 954c8efcc25STejun Heo */ 955c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 956c8efcc25STejun Heo { 957c8efcc25STejun Heo unsigned long flags; 958c8efcc25STejun Heo unsigned int cpu; 959c8efcc25STejun Heo 960c8efcc25STejun Heo for_each_gcwq_cpu(cpu) { 961c8efcc25STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 962c8efcc25STejun Heo struct worker *worker; 963c8efcc25STejun Heo struct hlist_node *pos; 964c8efcc25STejun Heo int i; 965c8efcc25STejun Heo 966c8efcc25STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 967c8efcc25STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 968c8efcc25STejun Heo if (worker->task != current) 969c8efcc25STejun Heo continue; 970c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 971c8efcc25STejun Heo /* 972c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 973c8efcc25STejun Heo * is headed to the same workqueue. 974c8efcc25STejun Heo */ 975c8efcc25STejun Heo return worker->current_cwq->wq == wq; 976c8efcc25STejun Heo } 977c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 978c8efcc25STejun Heo } 979c8efcc25STejun Heo return false; 980c8efcc25STejun Heo } 981c8efcc25STejun Heo 9824690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 9831da177e4SLinus Torvalds struct work_struct *work) 9841da177e4SLinus Torvalds { 985502ca9d8STejun Heo struct global_cwq *gcwq; 986502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 9871e19ffc6STejun Heo struct list_head *worklist; 9888a2e8e5dSTejun Heo unsigned int work_flags; 9891da177e4SLinus Torvalds unsigned long flags; 9901da177e4SLinus Torvalds 991dc186ad7SThomas Gleixner debug_work_activate(work); 9921e19ffc6STejun Heo 993c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 9949c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 995c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 996e41e704bSTejun Heo return; 997e41e704bSTejun Heo 998c7fc77f7STejun Heo /* determine gcwq to use */ 999c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1000c7fc77f7STejun Heo struct global_cwq *last_gcwq; 1001c7fc77f7STejun Heo 1002f3421797STejun Heo if (unlikely(cpu == WORK_CPU_UNBOUND)) 1003f3421797STejun Heo cpu = raw_smp_processor_id(); 1004f3421797STejun Heo 100518aa9effSTejun Heo /* 100618aa9effSTejun Heo * It's multi cpu. If @wq is non-reentrant and @work 100718aa9effSTejun Heo * was previously on a different cpu, it might still 100818aa9effSTejun Heo * be running there, in which case the work needs to 100918aa9effSTejun Heo * be queued on that cpu to guarantee non-reentrance. 101018aa9effSTejun Heo */ 1011502ca9d8STejun Heo gcwq = get_gcwq(cpu); 101218aa9effSTejun Heo if (wq->flags & WQ_NON_REENTRANT && 101318aa9effSTejun Heo (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) { 101418aa9effSTejun Heo struct worker *worker; 101518aa9effSTejun Heo 101618aa9effSTejun Heo spin_lock_irqsave(&last_gcwq->lock, flags); 101718aa9effSTejun Heo 101818aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 101918aa9effSTejun Heo 102018aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 102118aa9effSTejun Heo gcwq = last_gcwq; 102218aa9effSTejun Heo else { 102318aa9effSTejun Heo /* meh... not running there, queue here */ 102418aa9effSTejun Heo spin_unlock_irqrestore(&last_gcwq->lock, flags); 102518aa9effSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 102618aa9effSTejun Heo } 102718aa9effSTejun Heo } else 10288b03ae3cSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1029f3421797STejun Heo } else { 1030f3421797STejun Heo gcwq = get_gcwq(WORK_CPU_UNBOUND); 1031f3421797STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1032502ca9d8STejun Heo } 1033502ca9d8STejun Heo 1034502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 1035502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 1036cdadf009STejun Heo trace_workqueue_queue_work(cpu, cwq, work); 1037502ca9d8STejun Heo 10384690c4abSTejun Heo BUG_ON(!list_empty(&work->entry)); 10391e19ffc6STejun Heo 104073f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 10418a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 10421e19ffc6STejun Heo 10431e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1044cdadf009STejun Heo trace_workqueue_activate_work(work); 10451e19ffc6STejun Heo cwq->nr_active++; 1046649027d7STejun Heo worklist = gcwq_determine_ins_pos(gcwq, cwq); 10478a2e8e5dSTejun Heo } else { 10488a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 10491e19ffc6STejun Heo worklist = &cwq->delayed_works; 10508a2e8e5dSTejun Heo } 10511e19ffc6STejun Heo 10528a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 10531e19ffc6STejun Heo 10548b03ae3cSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 10551da177e4SLinus Torvalds } 10561da177e4SLinus Torvalds 10570fcb78c2SRolf Eike Beer /** 10580fcb78c2SRolf Eike Beer * queue_work - queue work on a workqueue 10590fcb78c2SRolf Eike Beer * @wq: workqueue to use 10600fcb78c2SRolf Eike Beer * @work: work to queue 10610fcb78c2SRolf Eike Beer * 1062057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 10631da177e4SLinus Torvalds * 106400dfcaf7SOleg Nesterov * We queue the work to the CPU on which it was submitted, but if the CPU dies 106500dfcaf7SOleg Nesterov * it can be processed by another CPU. 10661da177e4SLinus Torvalds */ 10677ad5b3a5SHarvey Harrison int queue_work(struct workqueue_struct *wq, struct work_struct *work) 10681da177e4SLinus Torvalds { 1069ef1ca236SOleg Nesterov int ret; 10701da177e4SLinus Torvalds 1071ef1ca236SOleg Nesterov ret = queue_work_on(get_cpu(), wq, work); 1072a848e3b6SOleg Nesterov put_cpu(); 1073ef1ca236SOleg Nesterov 10741da177e4SLinus Torvalds return ret; 10751da177e4SLinus Torvalds } 1076ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_work); 10771da177e4SLinus Torvalds 1078c1a220e7SZhang Rui /** 1079c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1080c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1081c1a220e7SZhang Rui * @wq: workqueue to use 1082c1a220e7SZhang Rui * @work: work to queue 1083c1a220e7SZhang Rui * 1084c1a220e7SZhang Rui * Returns 0 if @work was already on a queue, non-zero otherwise. 1085c1a220e7SZhang Rui * 1086c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1087c1a220e7SZhang Rui * can't go away. 1088c1a220e7SZhang Rui */ 1089c1a220e7SZhang Rui int 1090c1a220e7SZhang Rui queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work) 1091c1a220e7SZhang Rui { 1092c1a220e7SZhang Rui int ret = 0; 1093c1a220e7SZhang Rui 109422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 10954690c4abSTejun Heo __queue_work(cpu, wq, work); 1096c1a220e7SZhang Rui ret = 1; 1097c1a220e7SZhang Rui } 1098c1a220e7SZhang Rui return ret; 1099c1a220e7SZhang Rui } 1100c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1101c1a220e7SZhang Rui 11026d141c3fSLi Zefan static void delayed_work_timer_fn(unsigned long __data) 11031da177e4SLinus Torvalds { 110452bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 11057a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 11061da177e4SLinus Torvalds 11074690c4abSTejun Heo __queue_work(smp_processor_id(), cwq->wq, &dwork->work); 11081da177e4SLinus Torvalds } 11091da177e4SLinus Torvalds 11100fcb78c2SRolf Eike Beer /** 11110fcb78c2SRolf Eike Beer * queue_delayed_work - queue work on a workqueue after delay 11120fcb78c2SRolf Eike Beer * @wq: workqueue to use 1113af9997e4SRandy Dunlap * @dwork: delayable work to queue 11140fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 11150fcb78c2SRolf Eike Beer * 1116057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 11170fcb78c2SRolf Eike Beer */ 11187ad5b3a5SHarvey Harrison int queue_delayed_work(struct workqueue_struct *wq, 111952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 11201da177e4SLinus Torvalds { 112152bad64dSDavid Howells if (delay == 0) 112263bc0362SOleg Nesterov return queue_work(wq, &dwork->work); 11231da177e4SLinus Torvalds 112463bc0362SOleg Nesterov return queue_delayed_work_on(-1, wq, dwork, delay); 11251da177e4SLinus Torvalds } 1126ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work); 11271da177e4SLinus Torvalds 11280fcb78c2SRolf Eike Beer /** 11290fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 11300fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 11310fcb78c2SRolf Eike Beer * @wq: workqueue to use 1132af9997e4SRandy Dunlap * @dwork: work to queue 11330fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 11340fcb78c2SRolf Eike Beer * 1135057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 11360fcb78c2SRolf Eike Beer */ 11377a6bc1cdSVenkatesh Pallipadi int queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 113852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 11397a6bc1cdSVenkatesh Pallipadi { 11407a6bc1cdSVenkatesh Pallipadi int ret = 0; 114152bad64dSDavid Howells struct timer_list *timer = &dwork->timer; 114252bad64dSDavid Howells struct work_struct *work = &dwork->work; 11437a6bc1cdSVenkatesh Pallipadi 114422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1145c7fc77f7STejun Heo unsigned int lcpu; 11467a22ad75STejun Heo 11477a6bc1cdSVenkatesh Pallipadi BUG_ON(timer_pending(timer)); 11487a6bc1cdSVenkatesh Pallipadi BUG_ON(!list_empty(&work->entry)); 11497a6bc1cdSVenkatesh Pallipadi 11508a3e77ccSAndrew Liu timer_stats_timer_set_start_info(&dwork->timer); 11518a3e77ccSAndrew Liu 11527a22ad75STejun Heo /* 11537a22ad75STejun Heo * This stores cwq for the moment, for the timer_fn. 11547a22ad75STejun Heo * Note that the work's gcwq is preserved to allow 11557a22ad75STejun Heo * reentrance detection for delayed works. 11567a22ad75STejun Heo */ 1157c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1158c7fc77f7STejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 1159c7fc77f7STejun Heo 1160c7fc77f7STejun Heo if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND) 1161c7fc77f7STejun Heo lcpu = gcwq->cpu; 1162c7fc77f7STejun Heo else 1163c7fc77f7STejun Heo lcpu = raw_smp_processor_id(); 1164c7fc77f7STejun Heo } else 1165c7fc77f7STejun Heo lcpu = WORK_CPU_UNBOUND; 1166c7fc77f7STejun Heo 11677a22ad75STejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 1168c7fc77f7STejun Heo 11697a6bc1cdSVenkatesh Pallipadi timer->expires = jiffies + delay; 117052bad64dSDavid Howells timer->data = (unsigned long)dwork; 11717a6bc1cdSVenkatesh Pallipadi timer->function = delayed_work_timer_fn; 117263bc0362SOleg Nesterov 117363bc0362SOleg Nesterov if (unlikely(cpu >= 0)) 11747a6bc1cdSVenkatesh Pallipadi add_timer_on(timer, cpu); 117563bc0362SOleg Nesterov else 117663bc0362SOleg Nesterov add_timer(timer); 11777a6bc1cdSVenkatesh Pallipadi ret = 1; 11787a6bc1cdSVenkatesh Pallipadi } 11797a6bc1cdSVenkatesh Pallipadi return ret; 11807a6bc1cdSVenkatesh Pallipadi } 1181ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 11821da177e4SLinus Torvalds 1183c8e55f36STejun Heo /** 1184c8e55f36STejun Heo * worker_enter_idle - enter idle state 1185c8e55f36STejun Heo * @worker: worker which is entering idle state 1186c8e55f36STejun Heo * 1187c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1188c8e55f36STejun Heo * necessary. 1189c8e55f36STejun Heo * 1190c8e55f36STejun Heo * LOCKING: 1191c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1192c8e55f36STejun Heo */ 1193c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 11941da177e4SLinus Torvalds { 1195c8e55f36STejun Heo struct global_cwq *gcwq = worker->gcwq; 1196c8e55f36STejun Heo 1197c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1198c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1199c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1200c8e55f36STejun Heo 1201cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1202cb444766STejun Heo worker->flags |= WORKER_IDLE; 1203c8e55f36STejun Heo gcwq->nr_idle++; 1204e22bee78STejun Heo worker->last_active = jiffies; 1205c8e55f36STejun Heo 1206c8e55f36STejun Heo /* idle_list is LIFO */ 1207c8e55f36STejun Heo list_add(&worker->entry, &gcwq->idle_list); 1208db7bccf4STejun Heo 1209e22bee78STejun Heo if (likely(!(worker->flags & WORKER_ROGUE))) { 1210e22bee78STejun Heo if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer)) 1211e22bee78STejun Heo mod_timer(&gcwq->idle_timer, 1212e22bee78STejun Heo jiffies + IDLE_WORKER_TIMEOUT); 1213e22bee78STejun Heo } else 1214db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 1215cb444766STejun Heo 1216cb444766STejun Heo /* sanity check nr_running */ 1217cb444766STejun Heo WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle && 1218cb444766STejun Heo atomic_read(get_gcwq_nr_running(gcwq->cpu))); 1219c8e55f36STejun Heo } 1220c8e55f36STejun Heo 1221c8e55f36STejun Heo /** 1222c8e55f36STejun Heo * worker_leave_idle - leave idle state 1223c8e55f36STejun Heo * @worker: worker which is leaving idle state 1224c8e55f36STejun Heo * 1225c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1226c8e55f36STejun Heo * 1227c8e55f36STejun Heo * LOCKING: 1228c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1229c8e55f36STejun Heo */ 1230c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1231c8e55f36STejun Heo { 1232c8e55f36STejun Heo struct global_cwq *gcwq = worker->gcwq; 1233c8e55f36STejun Heo 1234c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1235d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1236c8e55f36STejun Heo gcwq->nr_idle--; 1237c8e55f36STejun Heo list_del_init(&worker->entry); 1238c8e55f36STejun Heo } 1239c8e55f36STejun Heo 1240e22bee78STejun Heo /** 1241e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1242e22bee78STejun Heo * @worker: self 1243e22bee78STejun Heo * 1244e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1245e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1246e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1247e22bee78STejun Heo * guaranteed to execute on the cpu. 1248e22bee78STejun Heo * 1249e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1250e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1251e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1252e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1253e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1254e22bee78STejun Heo * [dis]associated in the meantime. 1255e22bee78STejun Heo * 1256e22bee78STejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies 1257e22bee78STejun Heo * the binding against GCWQ_DISASSOCIATED which is set during 1258e22bee78STejun Heo * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters 1259e22bee78STejun Heo * idle state or fetches works without dropping lock, it can guarantee 1260e22bee78STejun Heo * the scheduling requirement described in the first paragraph. 1261e22bee78STejun Heo * 1262e22bee78STejun Heo * CONTEXT: 1263e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1264e22bee78STejun Heo * held. 1265e22bee78STejun Heo * 1266e22bee78STejun Heo * RETURNS: 1267e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1268e22bee78STejun Heo * bound), %false if offline. 1269e22bee78STejun Heo */ 1270e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1271972fa1c5SNamhyung Kim __acquires(&gcwq->lock) 1272e22bee78STejun Heo { 1273e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1274e22bee78STejun Heo struct task_struct *task = worker->task; 1275e22bee78STejun Heo 1276e22bee78STejun Heo while (true) { 1277e22bee78STejun Heo /* 1278e22bee78STejun Heo * The following call may fail, succeed or succeed 1279e22bee78STejun Heo * without actually migrating the task to the cpu if 1280e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1281e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1282e22bee78STejun Heo */ 1283f3421797STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) 1284e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1285e22bee78STejun Heo 1286e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1287e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1288e22bee78STejun Heo return false; 1289e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1290e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1291e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1292e22bee78STejun Heo return true; 1293e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1294e22bee78STejun Heo 12955035b20fSTejun Heo /* 12965035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 12975035b20fSTejun Heo * and retry migration. cond_resched() is required here; 12985035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 12995035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 13005035b20fSTejun Heo */ 1301e22bee78STejun Heo cpu_relax(); 13025035b20fSTejun Heo cond_resched(); 1303e22bee78STejun Heo } 1304e22bee78STejun Heo } 1305e22bee78STejun Heo 1306e22bee78STejun Heo /* 1307e22bee78STejun Heo * Function for worker->rebind_work used to rebind rogue busy workers 1308e22bee78STejun Heo * to the associated cpu which is coming back online. This is 1309e22bee78STejun Heo * scheduled by cpu up but can race with other cpu hotplug operations 1310e22bee78STejun Heo * and may be executed twice without intervening cpu down. 1311e22bee78STejun Heo */ 1312e22bee78STejun Heo static void worker_rebind_fn(struct work_struct *work) 1313e22bee78STejun Heo { 1314e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1315e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1316e22bee78STejun Heo 1317e22bee78STejun Heo if (worker_maybe_bind_and_lock(worker)) 1318e22bee78STejun Heo worker_clr_flags(worker, WORKER_REBIND); 1319e22bee78STejun Heo 1320e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1321e22bee78STejun Heo } 1322e22bee78STejun Heo 1323c34056a3STejun Heo static struct worker *alloc_worker(void) 1324c34056a3STejun Heo { 1325c34056a3STejun Heo struct worker *worker; 1326c34056a3STejun Heo 1327c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1328c8e55f36STejun Heo if (worker) { 1329c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1330affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1331e22bee78STejun Heo INIT_WORK(&worker->rebind_work, worker_rebind_fn); 1332e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1333e22bee78STejun Heo worker->flags = WORKER_PREP; 1334c8e55f36STejun Heo } 1335c34056a3STejun Heo return worker; 1336c34056a3STejun Heo } 1337c34056a3STejun Heo 1338c34056a3STejun Heo /** 1339c34056a3STejun Heo * create_worker - create a new workqueue worker 13407e11629dSTejun Heo * @gcwq: gcwq the new worker will belong to 1341c34056a3STejun Heo * @bind: whether to set affinity to @cpu or not 1342c34056a3STejun Heo * 13437e11629dSTejun Heo * Create a new worker which is bound to @gcwq. The returned worker 1344c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1345c34056a3STejun Heo * destroy_worker(). 1346c34056a3STejun Heo * 1347c34056a3STejun Heo * CONTEXT: 1348c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1349c34056a3STejun Heo * 1350c34056a3STejun Heo * RETURNS: 1351c34056a3STejun Heo * Pointer to the newly created worker. 1352c34056a3STejun Heo */ 13537e11629dSTejun Heo static struct worker *create_worker(struct global_cwq *gcwq, bool bind) 1354c34056a3STejun Heo { 1355f3421797STejun Heo bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND; 1356c34056a3STejun Heo struct worker *worker = NULL; 1357f3421797STejun Heo int id = -1; 1358c34056a3STejun Heo 13598b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 13608b03ae3cSTejun Heo while (ida_get_new(&gcwq->worker_ida, &id)) { 13618b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 13628b03ae3cSTejun Heo if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL)) 1363c34056a3STejun Heo goto fail; 13648b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1365c34056a3STejun Heo } 13668b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1367c34056a3STejun Heo 1368c34056a3STejun Heo worker = alloc_worker(); 1369c34056a3STejun Heo if (!worker) 1370c34056a3STejun Heo goto fail; 1371c34056a3STejun Heo 13728b03ae3cSTejun Heo worker->gcwq = gcwq; 1373c34056a3STejun Heo worker->id = id; 1374c34056a3STejun Heo 1375f3421797STejun Heo if (!on_unbound_cpu) 137694dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 137794dcf29aSEric Dumazet worker, 137894dcf29aSEric Dumazet cpu_to_node(gcwq->cpu), 1379f3421797STejun Heo "kworker/%u:%d", gcwq->cpu, id); 1380f3421797STejun Heo else 1381f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1382f3421797STejun Heo "kworker/u:%d", id); 1383c34056a3STejun Heo if (IS_ERR(worker->task)) 1384c34056a3STejun Heo goto fail; 1385c34056a3STejun Heo 1386db7bccf4STejun Heo /* 1387db7bccf4STejun Heo * A rogue worker will become a regular one if CPU comes 1388db7bccf4STejun Heo * online later on. Make sure every worker has 1389db7bccf4STejun Heo * PF_THREAD_BOUND set. 1390db7bccf4STejun Heo */ 1391f3421797STejun Heo if (bind && !on_unbound_cpu) 13928b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1393f3421797STejun Heo else { 1394db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1395f3421797STejun Heo if (on_unbound_cpu) 1396f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1397f3421797STejun Heo } 1398c34056a3STejun Heo 1399c34056a3STejun Heo return worker; 1400c34056a3STejun Heo fail: 1401c34056a3STejun Heo if (id >= 0) { 14028b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 14038b03ae3cSTejun Heo ida_remove(&gcwq->worker_ida, id); 14048b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1405c34056a3STejun Heo } 1406c34056a3STejun Heo kfree(worker); 1407c34056a3STejun Heo return NULL; 1408c34056a3STejun Heo } 1409c34056a3STejun Heo 1410c34056a3STejun Heo /** 1411c34056a3STejun Heo * start_worker - start a newly created worker 1412c34056a3STejun Heo * @worker: worker to start 1413c34056a3STejun Heo * 1414c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1415c34056a3STejun Heo * 1416c34056a3STejun Heo * CONTEXT: 14178b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1418c34056a3STejun Heo */ 1419c34056a3STejun Heo static void start_worker(struct worker *worker) 1420c34056a3STejun Heo { 1421cb444766STejun Heo worker->flags |= WORKER_STARTED; 1422c8e55f36STejun Heo worker->gcwq->nr_workers++; 1423c8e55f36STejun Heo worker_enter_idle(worker); 1424c34056a3STejun Heo wake_up_process(worker->task); 1425c34056a3STejun Heo } 1426c34056a3STejun Heo 1427c34056a3STejun Heo /** 1428c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1429c34056a3STejun Heo * @worker: worker to be destroyed 1430c34056a3STejun Heo * 1431c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1432c8e55f36STejun Heo * 1433c8e55f36STejun Heo * CONTEXT: 1434c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1435c34056a3STejun Heo */ 1436c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1437c34056a3STejun Heo { 14388b03ae3cSTejun Heo struct global_cwq *gcwq = worker->gcwq; 1439c34056a3STejun Heo int id = worker->id; 1440c34056a3STejun Heo 1441c34056a3STejun Heo /* sanity check frenzy */ 1442c34056a3STejun Heo BUG_ON(worker->current_work); 1443affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1444c34056a3STejun Heo 1445c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1446c8e55f36STejun Heo gcwq->nr_workers--; 1447c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1448c8e55f36STejun Heo gcwq->nr_idle--; 1449c8e55f36STejun Heo 1450c8e55f36STejun Heo list_del_init(&worker->entry); 1451cb444766STejun Heo worker->flags |= WORKER_DIE; 1452c8e55f36STejun Heo 1453c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1454c8e55f36STejun Heo 1455c34056a3STejun Heo kthread_stop(worker->task); 1456c34056a3STejun Heo kfree(worker); 1457c34056a3STejun Heo 14588b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 14598b03ae3cSTejun Heo ida_remove(&gcwq->worker_ida, id); 1460c34056a3STejun Heo } 1461c34056a3STejun Heo 1462e22bee78STejun Heo static void idle_worker_timeout(unsigned long __gcwq) 1463e22bee78STejun Heo { 1464e22bee78STejun Heo struct global_cwq *gcwq = (void *)__gcwq; 1465e22bee78STejun Heo 1466e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1467e22bee78STejun Heo 1468e22bee78STejun Heo if (too_many_workers(gcwq)) { 1469e22bee78STejun Heo struct worker *worker; 1470e22bee78STejun Heo unsigned long expires; 1471e22bee78STejun Heo 1472e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 1473e22bee78STejun Heo worker = list_entry(gcwq->idle_list.prev, struct worker, entry); 1474e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1475e22bee78STejun Heo 1476e22bee78STejun Heo if (time_before(jiffies, expires)) 1477e22bee78STejun Heo mod_timer(&gcwq->idle_timer, expires); 1478e22bee78STejun Heo else { 1479e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 1480e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGE_WORKERS; 1481e22bee78STejun Heo wake_up_worker(gcwq); 1482e22bee78STejun Heo } 1483e22bee78STejun Heo } 1484e22bee78STejun Heo 1485e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1486e22bee78STejun Heo } 1487e22bee78STejun Heo 1488e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1489e22bee78STejun Heo { 1490e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1491e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1492f3421797STejun Heo unsigned int cpu; 1493e22bee78STejun Heo 1494e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1495e22bee78STejun Heo return false; 1496e22bee78STejun Heo 1497e22bee78STejun Heo /* mayday mayday mayday */ 1498f3421797STejun Heo cpu = cwq->gcwq->cpu; 1499f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1500f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1501f3421797STejun Heo cpu = 0; 1502f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1503e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1504e22bee78STejun Heo return true; 1505e22bee78STejun Heo } 1506e22bee78STejun Heo 1507e22bee78STejun Heo static void gcwq_mayday_timeout(unsigned long __gcwq) 1508e22bee78STejun Heo { 1509e22bee78STejun Heo struct global_cwq *gcwq = (void *)__gcwq; 1510e22bee78STejun Heo struct work_struct *work; 1511e22bee78STejun Heo 1512e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1513e22bee78STejun Heo 1514e22bee78STejun Heo if (need_to_create_worker(gcwq)) { 1515e22bee78STejun Heo /* 1516e22bee78STejun Heo * We've been trying to create a new worker but 1517e22bee78STejun Heo * haven't been successful. We might be hitting an 1518e22bee78STejun Heo * allocation deadlock. Send distress signals to 1519e22bee78STejun Heo * rescuers. 1520e22bee78STejun Heo */ 1521e22bee78STejun Heo list_for_each_entry(work, &gcwq->worklist, entry) 1522e22bee78STejun Heo send_mayday(work); 1523e22bee78STejun Heo } 1524e22bee78STejun Heo 1525e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1526e22bee78STejun Heo 1527e22bee78STejun Heo mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL); 1528e22bee78STejun Heo } 1529e22bee78STejun Heo 1530e22bee78STejun Heo /** 1531e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 1532e22bee78STejun Heo * @gcwq: gcwq to create a new worker for 1533e22bee78STejun Heo * 1534e22bee78STejun Heo * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to 1535e22bee78STejun Heo * have at least one idle worker on return from this function. If 1536e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 1537e22bee78STejun Heo * sent to all rescuers with works scheduled on @gcwq to resolve 1538e22bee78STejun Heo * possible allocation deadlock. 1539e22bee78STejun Heo * 1540e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1541e22bee78STejun Heo * may_start_working() true. 1542e22bee78STejun Heo * 1543e22bee78STejun Heo * LOCKING: 1544e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1545e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1546e22bee78STejun Heo * manager. 1547e22bee78STejun Heo * 1548e22bee78STejun Heo * RETURNS: 1549e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1550e22bee78STejun Heo * otherwise. 1551e22bee78STejun Heo */ 1552e22bee78STejun Heo static bool maybe_create_worker(struct global_cwq *gcwq) 155306bd6ebfSNamhyung Kim __releases(&gcwq->lock) 155406bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 1555e22bee78STejun Heo { 1556e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1557e22bee78STejun Heo return false; 1558e22bee78STejun Heo restart: 15599f9c2364STejun Heo spin_unlock_irq(&gcwq->lock); 15609f9c2364STejun Heo 1561e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 1562e22bee78STejun Heo mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1563e22bee78STejun Heo 1564e22bee78STejun Heo while (true) { 1565e22bee78STejun Heo struct worker *worker; 1566e22bee78STejun Heo 1567e22bee78STejun Heo worker = create_worker(gcwq, true); 1568e22bee78STejun Heo if (worker) { 1569e22bee78STejun Heo del_timer_sync(&gcwq->mayday_timer); 1570e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1571e22bee78STejun Heo start_worker(worker); 1572e22bee78STejun Heo BUG_ON(need_to_create_worker(gcwq)); 1573e22bee78STejun Heo return true; 1574e22bee78STejun Heo } 1575e22bee78STejun Heo 1576e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1577e22bee78STejun Heo break; 1578e22bee78STejun Heo 1579e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1580e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 15819f9c2364STejun Heo 1582e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1583e22bee78STejun Heo break; 1584e22bee78STejun Heo } 1585e22bee78STejun Heo 1586e22bee78STejun Heo del_timer_sync(&gcwq->mayday_timer); 1587e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1588e22bee78STejun Heo if (need_to_create_worker(gcwq)) 1589e22bee78STejun Heo goto restart; 1590e22bee78STejun Heo return true; 1591e22bee78STejun Heo } 1592e22bee78STejun Heo 1593e22bee78STejun Heo /** 1594e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 1595e22bee78STejun Heo * @gcwq: gcwq to destroy workers for 1596e22bee78STejun Heo * 1597e22bee78STejun Heo * Destroy @gcwq workers which have been idle for longer than 1598e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1599e22bee78STejun Heo * 1600e22bee78STejun Heo * LOCKING: 1601e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1602e22bee78STejun Heo * multiple times. Called only from manager. 1603e22bee78STejun Heo * 1604e22bee78STejun Heo * RETURNS: 1605e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1606e22bee78STejun Heo * otherwise. 1607e22bee78STejun Heo */ 1608e22bee78STejun Heo static bool maybe_destroy_workers(struct global_cwq *gcwq) 1609e22bee78STejun Heo { 1610e22bee78STejun Heo bool ret = false; 1611e22bee78STejun Heo 1612e22bee78STejun Heo while (too_many_workers(gcwq)) { 1613e22bee78STejun Heo struct worker *worker; 1614e22bee78STejun Heo unsigned long expires; 1615e22bee78STejun Heo 1616e22bee78STejun Heo worker = list_entry(gcwq->idle_list.prev, struct worker, entry); 1617e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1618e22bee78STejun Heo 1619e22bee78STejun Heo if (time_before(jiffies, expires)) { 1620e22bee78STejun Heo mod_timer(&gcwq->idle_timer, expires); 1621e22bee78STejun Heo break; 1622e22bee78STejun Heo } 1623e22bee78STejun Heo 1624e22bee78STejun Heo destroy_worker(worker); 1625e22bee78STejun Heo ret = true; 1626e22bee78STejun Heo } 1627e22bee78STejun Heo 1628e22bee78STejun Heo return ret; 1629e22bee78STejun Heo } 1630e22bee78STejun Heo 1631e22bee78STejun Heo /** 1632e22bee78STejun Heo * manage_workers - manage worker pool 1633e22bee78STejun Heo * @worker: self 1634e22bee78STejun Heo * 1635e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 1636e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1637e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 1638e22bee78STejun Heo * 1639e22bee78STejun Heo * The caller can safely start processing works on false return. On 1640e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1641e22bee78STejun Heo * and may_start_working() is true. 1642e22bee78STejun Heo * 1643e22bee78STejun Heo * CONTEXT: 1644e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1645e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1646e22bee78STejun Heo * 1647e22bee78STejun Heo * RETURNS: 1648e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 1649e22bee78STejun Heo * some action was taken. 1650e22bee78STejun Heo */ 1651e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1652e22bee78STejun Heo { 1653e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1654e22bee78STejun Heo bool ret = false; 1655e22bee78STejun Heo 1656e22bee78STejun Heo if (gcwq->flags & GCWQ_MANAGING_WORKERS) 1657e22bee78STejun Heo return ret; 1658e22bee78STejun Heo 1659e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGE_WORKERS; 1660e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGING_WORKERS; 1661e22bee78STejun Heo 1662e22bee78STejun Heo /* 1663e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 1664e22bee78STejun Heo * on return. 1665e22bee78STejun Heo */ 1666e22bee78STejun Heo ret |= maybe_destroy_workers(gcwq); 1667e22bee78STejun Heo ret |= maybe_create_worker(gcwq); 1668e22bee78STejun Heo 1669e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGING_WORKERS; 1670e22bee78STejun Heo 1671e22bee78STejun Heo /* 1672e22bee78STejun Heo * The trustee might be waiting to take over the manager 1673e22bee78STejun Heo * position, tell it we're done. 1674e22bee78STejun Heo */ 1675e22bee78STejun Heo if (unlikely(gcwq->trustee)) 1676e22bee78STejun Heo wake_up_all(&gcwq->trustee_wait); 1677e22bee78STejun Heo 1678e22bee78STejun Heo return ret; 1679e22bee78STejun Heo } 1680e22bee78STejun Heo 1681a62428c0STejun Heo /** 1682affee4b2STejun Heo * move_linked_works - move linked works to a list 1683affee4b2STejun Heo * @work: start of series of works to be scheduled 1684affee4b2STejun Heo * @head: target list to append @work to 1685affee4b2STejun Heo * @nextp: out paramter for nested worklist walking 1686affee4b2STejun Heo * 1687affee4b2STejun Heo * Schedule linked works starting from @work to @head. Work series to 1688affee4b2STejun Heo * be scheduled starts at @work and includes any consecutive work with 1689affee4b2STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1690affee4b2STejun Heo * 1691affee4b2STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1692affee4b2STejun Heo * the last scheduled work. This allows move_linked_works() to be 1693affee4b2STejun Heo * nested inside outer list_for_each_entry_safe(). 1694affee4b2STejun Heo * 1695affee4b2STejun Heo * CONTEXT: 16968b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1697affee4b2STejun Heo */ 1698affee4b2STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1699affee4b2STejun Heo struct work_struct **nextp) 1700affee4b2STejun Heo { 1701affee4b2STejun Heo struct work_struct *n; 1702affee4b2STejun Heo 1703affee4b2STejun Heo /* 1704affee4b2STejun Heo * Linked worklist will always end before the end of the list, 1705affee4b2STejun Heo * use NULL for list head. 1706affee4b2STejun Heo */ 1707affee4b2STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1708affee4b2STejun Heo list_move_tail(&work->entry, head); 1709affee4b2STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1710affee4b2STejun Heo break; 1711affee4b2STejun Heo } 1712affee4b2STejun Heo 1713affee4b2STejun Heo /* 1714affee4b2STejun Heo * If we're already inside safe list traversal and have moved 1715affee4b2STejun Heo * multiple works to the scheduled queue, the next position 1716affee4b2STejun Heo * needs to be updated. 1717affee4b2STejun Heo */ 1718affee4b2STejun Heo if (nextp) 1719affee4b2STejun Heo *nextp = n; 1720affee4b2STejun Heo } 1721affee4b2STejun Heo 17221e19ffc6STejun Heo static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 17231e19ffc6STejun Heo { 17241e19ffc6STejun Heo struct work_struct *work = list_first_entry(&cwq->delayed_works, 17251da177e4SLinus Torvalds struct work_struct, entry); 1726649027d7STejun Heo struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq); 17271e19ffc6STejun Heo 1728cdadf009STejun Heo trace_workqueue_activate_work(work); 1729649027d7STejun Heo move_linked_works(work, pos, NULL); 17308a2e8e5dSTejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 17311e19ffc6STejun Heo cwq->nr_active++; 17321e19ffc6STejun Heo } 17331e19ffc6STejun Heo 1734affee4b2STejun Heo /** 173573f53c4aSTejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 173673f53c4aSTejun Heo * @cwq: cwq of interest 173773f53c4aSTejun Heo * @color: color of work which left the queue 17388a2e8e5dSTejun Heo * @delayed: for a delayed work 173973f53c4aSTejun Heo * 174073f53c4aSTejun Heo * A work either has completed or is removed from pending queue, 174173f53c4aSTejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 174273f53c4aSTejun Heo * 174373f53c4aSTejun Heo * CONTEXT: 17448b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 174573f53c4aSTejun Heo */ 17468a2e8e5dSTejun Heo static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color, 17478a2e8e5dSTejun Heo bool delayed) 174873f53c4aSTejun Heo { 174973f53c4aSTejun Heo /* ignore uncolored works */ 175073f53c4aSTejun Heo if (color == WORK_NO_COLOR) 175173f53c4aSTejun Heo return; 175273f53c4aSTejun Heo 175373f53c4aSTejun Heo cwq->nr_in_flight[color]--; 17541e19ffc6STejun Heo 17558a2e8e5dSTejun Heo if (!delayed) { 17568a2e8e5dSTejun Heo cwq->nr_active--; 1757502ca9d8STejun Heo if (!list_empty(&cwq->delayed_works)) { 17581e19ffc6STejun Heo /* one down, submit a delayed one */ 1759502ca9d8STejun Heo if (cwq->nr_active < cwq->max_active) 17601e19ffc6STejun Heo cwq_activate_first_delayed(cwq); 1761502ca9d8STejun Heo } 17628a2e8e5dSTejun Heo } 176373f53c4aSTejun Heo 176473f53c4aSTejun Heo /* is flush in progress and are we at the flushing tip? */ 176573f53c4aSTejun Heo if (likely(cwq->flush_color != color)) 176673f53c4aSTejun Heo return; 176773f53c4aSTejun Heo 176873f53c4aSTejun Heo /* are there still in-flight works? */ 176973f53c4aSTejun Heo if (cwq->nr_in_flight[color]) 177073f53c4aSTejun Heo return; 177173f53c4aSTejun Heo 177273f53c4aSTejun Heo /* this cwq is done, clear flush_color */ 177373f53c4aSTejun Heo cwq->flush_color = -1; 177473f53c4aSTejun Heo 177573f53c4aSTejun Heo /* 177673f53c4aSTejun Heo * If this was the last cwq, wake up the first flusher. It 177773f53c4aSTejun Heo * will handle the rest. 177873f53c4aSTejun Heo */ 177973f53c4aSTejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 178073f53c4aSTejun Heo complete(&cwq->wq->first_flusher->done); 178173f53c4aSTejun Heo } 178273f53c4aSTejun Heo 178373f53c4aSTejun Heo /** 1784a62428c0STejun Heo * process_one_work - process single work 1785c34056a3STejun Heo * @worker: self 1786a62428c0STejun Heo * @work: work to process 1787a62428c0STejun Heo * 1788a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1789a62428c0STejun Heo * process a single work including synchronization against and 1790a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1791a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1792a62428c0STejun Heo * call this function to process a work. 1793a62428c0STejun Heo * 1794a62428c0STejun Heo * CONTEXT: 17958b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1796a62428c0STejun Heo */ 1797c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 179806bd6ebfSNamhyung Kim __releases(&gcwq->lock) 179906bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 18001da177e4SLinus Torvalds { 18017e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 18028b03ae3cSTejun Heo struct global_cwq *gcwq = cwq->gcwq; 1803c8e55f36STejun Heo struct hlist_head *bwh = busy_worker_head(gcwq, work); 1804fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 18056bb49e59SDavid Howells work_func_t f = work->func; 180673f53c4aSTejun Heo int work_color; 18077e11629dSTejun Heo struct worker *collision; 18084e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 18094e6045f1SJohannes Berg /* 1810a62428c0STejun Heo * It is permissible to free the struct work_struct from 1811a62428c0STejun Heo * inside the function that is called from it, this we need to 1812a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1813a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1814a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 18154e6045f1SJohannes Berg */ 18164e6045f1SJohannes Berg struct lockdep_map lockdep_map = work->lockdep_map; 18174e6045f1SJohannes Berg #endif 18187e11629dSTejun Heo /* 18197e11629dSTejun Heo * A single work shouldn't be executed concurrently by 18207e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 18217e11629dSTejun Heo * already processing the work. If so, defer the work to the 18227e11629dSTejun Heo * currently executing one. 18237e11629dSTejun Heo */ 18247e11629dSTejun Heo collision = __find_worker_executing_work(gcwq, bwh, work); 18257e11629dSTejun Heo if (unlikely(collision)) { 18267e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 18277e11629dSTejun Heo return; 18287e11629dSTejun Heo } 18291da177e4SLinus Torvalds 1830a62428c0STejun Heo /* claim and process */ 18311da177e4SLinus Torvalds debug_work_deactivate(work); 1832c8e55f36STejun Heo hlist_add_head(&worker->hentry, bwh); 1833c34056a3STejun Heo worker->current_work = work; 18348cca0eeaSTejun Heo worker->current_cwq = cwq; 183573f53c4aSTejun Heo work_color = get_work_color(work); 18367a22ad75STejun Heo 18377a22ad75STejun Heo /* record the current cpu number in the work data and dequeue */ 18387a22ad75STejun Heo set_work_cpu(work, gcwq->cpu); 1839a62428c0STejun Heo list_del_init(&work->entry); 1840a62428c0STejun Heo 1841649027d7STejun Heo /* 1842649027d7STejun Heo * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI, 1843649027d7STejun Heo * wake up another worker; otherwise, clear HIGHPRI_PENDING. 1844649027d7STejun Heo */ 1845649027d7STejun Heo if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) { 1846649027d7STejun Heo struct work_struct *nwork = list_first_entry(&gcwq->worklist, 1847649027d7STejun Heo struct work_struct, entry); 1848649027d7STejun Heo 1849649027d7STejun Heo if (!list_empty(&gcwq->worklist) && 1850649027d7STejun Heo get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI) 1851649027d7STejun Heo wake_up_worker(gcwq); 1852649027d7STejun Heo else 1853649027d7STejun Heo gcwq->flags &= ~GCWQ_HIGHPRI_PENDING; 1854649027d7STejun Heo } 1855649027d7STejun Heo 1856fb0e7bebSTejun Heo /* 1857fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 1858fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 1859fb0e7bebSTejun Heo */ 1860fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1861fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 1862fb0e7bebSTejun Heo 18638b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 18641da177e4SLinus Torvalds 186523b2e599SOleg Nesterov work_clear_pending(work); 1866e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 18673295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 1868e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 186965f27f38SDavid Howells f(work); 1870e36c886aSArjan van de Ven /* 1871e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 1872e36c886aSArjan van de Ven * point will only record its address. 1873e36c886aSArjan van de Ven */ 1874e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 18753295f0efSIngo Molnar lock_map_release(&lockdep_map); 18763295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 18771da177e4SLinus Torvalds 1878d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 1879d5abe669SPeter Zijlstra printk(KERN_ERR "BUG: workqueue leaked lock or atomic: " 1880d5abe669SPeter Zijlstra "%s/0x%08x/%d\n", 1881a62428c0STejun Heo current->comm, preempt_count(), task_pid_nr(current)); 1882d5abe669SPeter Zijlstra printk(KERN_ERR " last function: "); 1883d5abe669SPeter Zijlstra print_symbol("%s\n", (unsigned long)f); 1884d5abe669SPeter Zijlstra debug_show_held_locks(current); 1885d5abe669SPeter Zijlstra dump_stack(); 1886d5abe669SPeter Zijlstra } 1887d5abe669SPeter Zijlstra 18888b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1889a62428c0STejun Heo 1890fb0e7bebSTejun Heo /* clear cpu intensive status */ 1891fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1892fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 1893fb0e7bebSTejun Heo 1894a62428c0STejun Heo /* we're done with it, release */ 1895c8e55f36STejun Heo hlist_del_init(&worker->hentry); 1896c34056a3STejun Heo worker->current_work = NULL; 18978cca0eeaSTejun Heo worker->current_cwq = NULL; 18988a2e8e5dSTejun Heo cwq_dec_nr_in_flight(cwq, work_color, false); 18991da177e4SLinus Torvalds } 19001da177e4SLinus Torvalds 1901affee4b2STejun Heo /** 1902affee4b2STejun Heo * process_scheduled_works - process scheduled works 1903affee4b2STejun Heo * @worker: self 1904affee4b2STejun Heo * 1905affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 1906affee4b2STejun Heo * may change while processing a work, so this function repeatedly 1907affee4b2STejun Heo * fetches a work from the top and executes it. 1908affee4b2STejun Heo * 1909affee4b2STejun Heo * CONTEXT: 19108b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1911affee4b2STejun Heo * multiple times. 1912affee4b2STejun Heo */ 1913affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 19141da177e4SLinus Torvalds { 1915affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 1916affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 1917a62428c0STejun Heo struct work_struct, entry); 1918c34056a3STejun Heo process_one_work(worker, work); 1919a62428c0STejun Heo } 19201da177e4SLinus Torvalds } 19211da177e4SLinus Torvalds 19224690c4abSTejun Heo /** 19234690c4abSTejun Heo * worker_thread - the worker thread function 1924c34056a3STejun Heo * @__worker: self 19254690c4abSTejun Heo * 1926e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 1927e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 1928e22bee78STejun Heo * their specific target workqueue. The only exception is works which 1929e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 1930e22bee78STejun Heo * rescuer_thread(). 19314690c4abSTejun Heo */ 1932c34056a3STejun Heo static int worker_thread(void *__worker) 19331da177e4SLinus Torvalds { 1934c34056a3STejun Heo struct worker *worker = __worker; 19358b03ae3cSTejun Heo struct global_cwq *gcwq = worker->gcwq; 19361da177e4SLinus Torvalds 1937e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 1938e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 1939c8e55f36STejun Heo woke_up: 19408b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1941affee4b2STejun Heo 1942c8e55f36STejun Heo /* DIE can be set only while we're idle, checking here is enough */ 1943c8e55f36STejun Heo if (worker->flags & WORKER_DIE) { 1944c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1945e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 1946c8e55f36STejun Heo return 0; 1947c8e55f36STejun Heo } 1948c8e55f36STejun Heo 1949c8e55f36STejun Heo worker_leave_idle(worker); 1950db7bccf4STejun Heo recheck: 1951e22bee78STejun Heo /* no more worker necessary? */ 1952e22bee78STejun Heo if (!need_more_worker(gcwq)) 1953e22bee78STejun Heo goto sleep; 1954e22bee78STejun Heo 1955e22bee78STejun Heo /* do we need to manage? */ 1956e22bee78STejun Heo if (unlikely(!may_start_working(gcwq)) && manage_workers(worker)) 1957e22bee78STejun Heo goto recheck; 1958e22bee78STejun Heo 1959c8e55f36STejun Heo /* 1960c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 1961c8e55f36STejun Heo * preparing to process a work or actually processing it. 1962c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 1963c8e55f36STejun Heo */ 1964c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1965c8e55f36STejun Heo 1966e22bee78STejun Heo /* 1967e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 1968e22bee78STejun Heo * at least one idle worker or that someone else has already 1969e22bee78STejun Heo * assumed the manager role. 1970e22bee78STejun Heo */ 1971e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 1972e22bee78STejun Heo 1973e22bee78STejun Heo do { 1974affee4b2STejun Heo struct work_struct *work = 19757e11629dSTejun Heo list_first_entry(&gcwq->worklist, 1976affee4b2STejun Heo struct work_struct, entry); 1977affee4b2STejun Heo 1978c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 1979affee4b2STejun Heo /* optimization path, not strictly necessary */ 1980affee4b2STejun Heo process_one_work(worker, work); 1981affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 1982affee4b2STejun Heo process_scheduled_works(worker); 1983affee4b2STejun Heo } else { 1984c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 1985affee4b2STejun Heo process_scheduled_works(worker); 1986affee4b2STejun Heo } 1987e22bee78STejun Heo } while (keep_working(gcwq)); 1988affee4b2STejun Heo 1989e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 1990d313dd85STejun Heo sleep: 1991e22bee78STejun Heo if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker)) 1992e22bee78STejun Heo goto recheck; 1993d313dd85STejun Heo 1994c8e55f36STejun Heo /* 1995e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 1996e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 1997e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 1998e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 1999e22bee78STejun Heo * prevent losing any event. 2000c8e55f36STejun Heo */ 2001c8e55f36STejun Heo worker_enter_idle(worker); 2002c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 20038b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 20041da177e4SLinus Torvalds schedule(); 2005c8e55f36STejun Heo goto woke_up; 20061da177e4SLinus Torvalds } 20071da177e4SLinus Torvalds 2008e22bee78STejun Heo /** 2009e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2010e22bee78STejun Heo * @__wq: the associated workqueue 2011e22bee78STejun Heo * 2012e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2013e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2014e22bee78STejun Heo * 2015e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 2016e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2017e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2018e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2019e22bee78STejun Heo * the problem rescuer solves. 2020e22bee78STejun Heo * 2021e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 2022e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 2023e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2024e22bee78STejun Heo * 2025e22bee78STejun Heo * This should happen rarely. 2026e22bee78STejun Heo */ 2027e22bee78STejun Heo static int rescuer_thread(void *__wq) 2028e22bee78STejun Heo { 2029e22bee78STejun Heo struct workqueue_struct *wq = __wq; 2030e22bee78STejun Heo struct worker *rescuer = wq->rescuer; 2031e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2032f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2033e22bee78STejun Heo unsigned int cpu; 2034e22bee78STejun Heo 2035e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2036e22bee78STejun Heo repeat: 2037e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 20381da177e4SLinus Torvalds 20391da177e4SLinus Torvalds if (kthread_should_stop()) 2040e22bee78STejun Heo return 0; 20411da177e4SLinus Torvalds 2042f3421797STejun Heo /* 2043f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2044f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2045f3421797STejun Heo */ 2046f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2047f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2048f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2049e22bee78STejun Heo struct global_cwq *gcwq = cwq->gcwq; 2050e22bee78STejun Heo struct work_struct *work, *n; 2051e22bee78STejun Heo 2052e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2053f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2054e22bee78STejun Heo 2055e22bee78STejun Heo /* migrate to the target cpu if possible */ 2056e22bee78STejun Heo rescuer->gcwq = gcwq; 2057e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2058e22bee78STejun Heo 2059e22bee78STejun Heo /* 2060e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2061e22bee78STejun Heo * process'em. 2062e22bee78STejun Heo */ 2063e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2064e22bee78STejun Heo list_for_each_entry_safe(work, n, &gcwq->worklist, entry) 2065e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2066e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2067e22bee78STejun Heo 2068e22bee78STejun Heo process_scheduled_works(rescuer); 20697576958aSTejun Heo 20707576958aSTejun Heo /* 20717576958aSTejun Heo * Leave this gcwq. If keep_working() is %true, notify a 20727576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 20737576958aSTejun Heo * and stalling the execution. 20747576958aSTejun Heo */ 20757576958aSTejun Heo if (keep_working(gcwq)) 20767576958aSTejun Heo wake_up_worker(gcwq); 20777576958aSTejun Heo 2078e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 20791da177e4SLinus Torvalds } 20801da177e4SLinus Torvalds 2081e22bee78STejun Heo schedule(); 2082e22bee78STejun Heo goto repeat; 20831da177e4SLinus Torvalds } 20841da177e4SLinus Torvalds 2085fc2e4d70SOleg Nesterov struct wq_barrier { 2086fc2e4d70SOleg Nesterov struct work_struct work; 2087fc2e4d70SOleg Nesterov struct completion done; 2088fc2e4d70SOleg Nesterov }; 2089fc2e4d70SOleg Nesterov 2090fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2091fc2e4d70SOleg Nesterov { 2092fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2093fc2e4d70SOleg Nesterov complete(&barr->done); 2094fc2e4d70SOleg Nesterov } 2095fc2e4d70SOleg Nesterov 20964690c4abSTejun Heo /** 20974690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 20984690c4abSTejun Heo * @cwq: cwq to insert barrier into 20994690c4abSTejun Heo * @barr: wq_barrier to insert 2100affee4b2STejun Heo * @target: target work to attach @barr to 2101affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 21024690c4abSTejun Heo * 2103affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2104affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2105affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2106affee4b2STejun Heo * cpu. 2107affee4b2STejun Heo * 2108affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2109affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2110affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2111affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2112affee4b2STejun Heo * after a work with LINKED flag set. 2113affee4b2STejun Heo * 2114affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2115affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 21164690c4abSTejun Heo * 21174690c4abSTejun Heo * CONTEXT: 21188b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 21194690c4abSTejun Heo */ 212083c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2121affee4b2STejun Heo struct wq_barrier *barr, 2122affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2123fc2e4d70SOleg Nesterov { 2124affee4b2STejun Heo struct list_head *head; 2125affee4b2STejun Heo unsigned int linked = 0; 2126affee4b2STejun Heo 2127dc186ad7SThomas Gleixner /* 21288b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 2129dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2130dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2131dc186ad7SThomas Gleixner * might deadlock. 2132dc186ad7SThomas Gleixner */ 2133ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 213422df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2135fc2e4d70SOleg Nesterov init_completion(&barr->done); 213683c22520SOleg Nesterov 2137affee4b2STejun Heo /* 2138affee4b2STejun Heo * If @target is currently being executed, schedule the 2139affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2140affee4b2STejun Heo */ 2141affee4b2STejun Heo if (worker) 2142affee4b2STejun Heo head = worker->scheduled.next; 2143affee4b2STejun Heo else { 2144affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2145affee4b2STejun Heo 2146affee4b2STejun Heo head = target->entry.next; 2147affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2148affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2149affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2150affee4b2STejun Heo } 2151affee4b2STejun Heo 2152dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2153affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2154affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2155fc2e4d70SOleg Nesterov } 2156fc2e4d70SOleg Nesterov 215773f53c4aSTejun Heo /** 215873f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 215973f53c4aSTejun Heo * @wq: workqueue being flushed 216073f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 216173f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 216273f53c4aSTejun Heo * 216373f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 216473f53c4aSTejun Heo * 216573f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 216673f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 216773f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 216873f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 216973f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 217073f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 217173f53c4aSTejun Heo * 217273f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 217373f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 217473f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 217573f53c4aSTejun Heo * is returned. 217673f53c4aSTejun Heo * 217773f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 217873f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 217973f53c4aSTejun Heo * advanced to @work_color. 218073f53c4aSTejun Heo * 218173f53c4aSTejun Heo * CONTEXT: 218273f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 218373f53c4aSTejun Heo * 218473f53c4aSTejun Heo * RETURNS: 218573f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 218673f53c4aSTejun Heo * otherwise. 218773f53c4aSTejun Heo */ 218873f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 218973f53c4aSTejun Heo int flush_color, int work_color) 21901da177e4SLinus Torvalds { 219173f53c4aSTejun Heo bool wait = false; 219273f53c4aSTejun Heo unsigned int cpu; 21931da177e4SLinus Torvalds 219473f53c4aSTejun Heo if (flush_color >= 0) { 219573f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 219673f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2197dc186ad7SThomas Gleixner } 219814441960SOleg Nesterov 2199f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 220073f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 22018b03ae3cSTejun Heo struct global_cwq *gcwq = cwq->gcwq; 22021da177e4SLinus Torvalds 22038b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 220473f53c4aSTejun Heo 220573f53c4aSTejun Heo if (flush_color >= 0) { 220673f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 220773f53c4aSTejun Heo 220873f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 220973f53c4aSTejun Heo cwq->flush_color = flush_color; 221073f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 221173f53c4aSTejun Heo wait = true; 22121da177e4SLinus Torvalds } 221373f53c4aSTejun Heo } 221473f53c4aSTejun Heo 221573f53c4aSTejun Heo if (work_color >= 0) { 221673f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 221773f53c4aSTejun Heo cwq->work_color = work_color; 221873f53c4aSTejun Heo } 221973f53c4aSTejun Heo 22208b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 22211da177e4SLinus Torvalds } 22221da177e4SLinus Torvalds 222373f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 222473f53c4aSTejun Heo complete(&wq->first_flusher->done); 222573f53c4aSTejun Heo 222673f53c4aSTejun Heo return wait; 222783c22520SOleg Nesterov } 22281da177e4SLinus Torvalds 22290fcb78c2SRolf Eike Beer /** 22301da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 22310fcb78c2SRolf Eike Beer * @wq: workqueue to flush 22321da177e4SLinus Torvalds * 22331da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 22341da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 22351da177e4SLinus Torvalds * 2236fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2237fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 22381da177e4SLinus Torvalds */ 22397ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 22401da177e4SLinus Torvalds { 224173f53c4aSTejun Heo struct wq_flusher this_flusher = { 224273f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 224373f53c4aSTejun Heo .flush_color = -1, 224473f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 224573f53c4aSTejun Heo }; 224673f53c4aSTejun Heo int next_color; 2247b1f4ec17SOleg Nesterov 22483295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 22493295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 225073f53c4aSTejun Heo 225173f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 225273f53c4aSTejun Heo 225373f53c4aSTejun Heo /* 225473f53c4aSTejun Heo * Start-to-wait phase 225573f53c4aSTejun Heo */ 225673f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 225773f53c4aSTejun Heo 225873f53c4aSTejun Heo if (next_color != wq->flush_color) { 225973f53c4aSTejun Heo /* 226073f53c4aSTejun Heo * Color space is not full. The current work_color 226173f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 226273f53c4aSTejun Heo * by one. 226373f53c4aSTejun Heo */ 226473f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 226573f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 226673f53c4aSTejun Heo wq->work_color = next_color; 226773f53c4aSTejun Heo 226873f53c4aSTejun Heo if (!wq->first_flusher) { 226973f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 227073f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 227173f53c4aSTejun Heo 227273f53c4aSTejun Heo wq->first_flusher = &this_flusher; 227373f53c4aSTejun Heo 227473f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 227573f53c4aSTejun Heo wq->work_color)) { 227673f53c4aSTejun Heo /* nothing to flush, done */ 227773f53c4aSTejun Heo wq->flush_color = next_color; 227873f53c4aSTejun Heo wq->first_flusher = NULL; 227973f53c4aSTejun Heo goto out_unlock; 228073f53c4aSTejun Heo } 228173f53c4aSTejun Heo } else { 228273f53c4aSTejun Heo /* wait in queue */ 228373f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 228473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 228573f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 228673f53c4aSTejun Heo } 228773f53c4aSTejun Heo } else { 228873f53c4aSTejun Heo /* 228973f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 229073f53c4aSTejun Heo * The next flush completion will assign us 229173f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 229273f53c4aSTejun Heo */ 229373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 229473f53c4aSTejun Heo } 229573f53c4aSTejun Heo 229673f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 229773f53c4aSTejun Heo 229873f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 229973f53c4aSTejun Heo 230073f53c4aSTejun Heo /* 230173f53c4aSTejun Heo * Wake-up-and-cascade phase 230273f53c4aSTejun Heo * 230373f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 230473f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 230573f53c4aSTejun Heo */ 230673f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 230773f53c4aSTejun Heo return; 230873f53c4aSTejun Heo 230973f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 231073f53c4aSTejun Heo 23114ce48b37STejun Heo /* we might have raced, check again with mutex held */ 23124ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 23134ce48b37STejun Heo goto out_unlock; 23144ce48b37STejun Heo 231573f53c4aSTejun Heo wq->first_flusher = NULL; 231673f53c4aSTejun Heo 231773f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 231873f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 231973f53c4aSTejun Heo 232073f53c4aSTejun Heo while (true) { 232173f53c4aSTejun Heo struct wq_flusher *next, *tmp; 232273f53c4aSTejun Heo 232373f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 232473f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 232573f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 232673f53c4aSTejun Heo break; 232773f53c4aSTejun Heo list_del_init(&next->list); 232873f53c4aSTejun Heo complete(&next->done); 232973f53c4aSTejun Heo } 233073f53c4aSTejun Heo 233173f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 233273f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 233373f53c4aSTejun Heo 233473f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 233573f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 233673f53c4aSTejun Heo 233773f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 233873f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 233973f53c4aSTejun Heo /* 234073f53c4aSTejun Heo * Assign the same color to all overflowed 234173f53c4aSTejun Heo * flushers, advance work_color and append to 234273f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 234373f53c4aSTejun Heo * phase for these overflowed flushers. 234473f53c4aSTejun Heo */ 234573f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 234673f53c4aSTejun Heo tmp->flush_color = wq->work_color; 234773f53c4aSTejun Heo 234873f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 234973f53c4aSTejun Heo 235073f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 235173f53c4aSTejun Heo &wq->flusher_queue); 235273f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 235373f53c4aSTejun Heo } 235473f53c4aSTejun Heo 235573f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 235673f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 235773f53c4aSTejun Heo break; 235873f53c4aSTejun Heo } 235973f53c4aSTejun Heo 236073f53c4aSTejun Heo /* 236173f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 236273f53c4aSTejun Heo * the new first flusher and arm cwqs. 236373f53c4aSTejun Heo */ 236473f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 236573f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 236673f53c4aSTejun Heo 236773f53c4aSTejun Heo list_del_init(&next->list); 236873f53c4aSTejun Heo wq->first_flusher = next; 236973f53c4aSTejun Heo 237073f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 237173f53c4aSTejun Heo break; 237273f53c4aSTejun Heo 237373f53c4aSTejun Heo /* 237473f53c4aSTejun Heo * Meh... this color is already done, clear first 237573f53c4aSTejun Heo * flusher and repeat cascading. 237673f53c4aSTejun Heo */ 237773f53c4aSTejun Heo wq->first_flusher = NULL; 237873f53c4aSTejun Heo } 237973f53c4aSTejun Heo 238073f53c4aSTejun Heo out_unlock: 238173f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 23821da177e4SLinus Torvalds } 2383ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 23841da177e4SLinus Torvalds 23859c5a2ba7STejun Heo /** 23869c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 23879c5a2ba7STejun Heo * @wq: workqueue to drain 23889c5a2ba7STejun Heo * 23899c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 23909c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 23919c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 23929c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 23939c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 23949c5a2ba7STejun Heo * takes too long. 23959c5a2ba7STejun Heo */ 23969c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 23979c5a2ba7STejun Heo { 23989c5a2ba7STejun Heo unsigned int flush_cnt = 0; 23999c5a2ba7STejun Heo unsigned int cpu; 24009c5a2ba7STejun Heo 24019c5a2ba7STejun Heo /* 24029c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 24039c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 24049c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 24059c5a2ba7STejun Heo */ 24069c5a2ba7STejun Heo spin_lock(&workqueue_lock); 24079c5a2ba7STejun Heo if (!wq->nr_drainers++) 24089c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 24099c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 24109c5a2ba7STejun Heo reflush: 24119c5a2ba7STejun Heo flush_workqueue(wq); 24129c5a2ba7STejun Heo 24139c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 24149c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 24159c5a2ba7STejun Heo 24169c5a2ba7STejun Heo if (!cwq->nr_active && list_empty(&cwq->delayed_works)) 24179c5a2ba7STejun Heo continue; 24189c5a2ba7STejun Heo 24199c5a2ba7STejun Heo if (++flush_cnt == 10 || 24209c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 24219c5a2ba7STejun Heo pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n", 24229c5a2ba7STejun Heo wq->name, flush_cnt); 24239c5a2ba7STejun Heo goto reflush; 24249c5a2ba7STejun Heo } 24259c5a2ba7STejun Heo 24269c5a2ba7STejun Heo spin_lock(&workqueue_lock); 24279c5a2ba7STejun Heo if (!--wq->nr_drainers) 24289c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 24299c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 24309c5a2ba7STejun Heo } 24319c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 24329c5a2ba7STejun Heo 2433baf59022STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2434baf59022STejun Heo bool wait_executing) 2435baf59022STejun Heo { 2436baf59022STejun Heo struct worker *worker = NULL; 2437baf59022STejun Heo struct global_cwq *gcwq; 2438baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2439baf59022STejun Heo 2440baf59022STejun Heo might_sleep(); 2441baf59022STejun Heo gcwq = get_work_gcwq(work); 2442baf59022STejun Heo if (!gcwq) 2443baf59022STejun Heo return false; 2444baf59022STejun Heo 2445baf59022STejun Heo spin_lock_irq(&gcwq->lock); 2446baf59022STejun Heo if (!list_empty(&work->entry)) { 2447baf59022STejun Heo /* 2448baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2449baf59022STejun Heo * If it was re-queued to a different gcwq under us, we 2450baf59022STejun Heo * are not going to wait. 2451baf59022STejun Heo */ 2452baf59022STejun Heo smp_rmb(); 2453baf59022STejun Heo cwq = get_work_cwq(work); 2454baf59022STejun Heo if (unlikely(!cwq || gcwq != cwq->gcwq)) 2455baf59022STejun Heo goto already_gone; 2456baf59022STejun Heo } else if (wait_executing) { 2457baf59022STejun Heo worker = find_worker_executing_work(gcwq, work); 2458baf59022STejun Heo if (!worker) 2459baf59022STejun Heo goto already_gone; 2460baf59022STejun Heo cwq = worker->current_cwq; 2461baf59022STejun Heo } else 2462baf59022STejun Heo goto already_gone; 2463baf59022STejun Heo 2464baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2465baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2466baf59022STejun Heo 2467e159489bSTejun Heo /* 2468e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2469e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2470e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2471e159489bSTejun Heo * access. 2472e159489bSTejun Heo */ 2473e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2474baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2475e159489bSTejun Heo else 2476e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2477baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2478e159489bSTejun Heo 2479baf59022STejun Heo return true; 2480baf59022STejun Heo already_gone: 2481baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2482baf59022STejun Heo return false; 2483baf59022STejun Heo } 2484baf59022STejun Heo 2485db700897SOleg Nesterov /** 2486401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2487401a8d04STejun Heo * @work: the work to flush 2488db700897SOleg Nesterov * 2489401a8d04STejun Heo * Wait until @work has finished execution. This function considers 2490401a8d04STejun Heo * only the last queueing instance of @work. If @work has been 2491401a8d04STejun Heo * enqueued across different CPUs on a non-reentrant workqueue or on 2492401a8d04STejun Heo * multiple workqueues, @work might still be executing on return on 2493401a8d04STejun Heo * some of the CPUs from earlier queueing. 2494a67da70dSOleg Nesterov * 2495401a8d04STejun Heo * If @work was queued only on a non-reentrant, ordered or unbound 2496401a8d04STejun Heo * workqueue, @work is guaranteed to be idle on return if it hasn't 2497401a8d04STejun Heo * been requeued since flush started. 2498401a8d04STejun Heo * 2499401a8d04STejun Heo * RETURNS: 2500401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2501401a8d04STejun Heo * %false if it was already idle. 2502db700897SOleg Nesterov */ 2503401a8d04STejun Heo bool flush_work(struct work_struct *work) 2504db700897SOleg Nesterov { 2505db700897SOleg Nesterov struct wq_barrier barr; 2506db700897SOleg Nesterov 2507baf59022STejun Heo if (start_flush_work(work, &barr, true)) { 2508db700897SOleg Nesterov wait_for_completion(&barr.done); 2509dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2510401a8d04STejun Heo return true; 2511baf59022STejun Heo } else 2512401a8d04STejun Heo return false; 2513db700897SOleg Nesterov } 2514db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2515db700897SOleg Nesterov 2516401a8d04STejun Heo static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work) 2517401a8d04STejun Heo { 2518401a8d04STejun Heo struct wq_barrier barr; 2519401a8d04STejun Heo struct worker *worker; 2520401a8d04STejun Heo 2521401a8d04STejun Heo spin_lock_irq(&gcwq->lock); 2522401a8d04STejun Heo 2523401a8d04STejun Heo worker = find_worker_executing_work(gcwq, work); 2524401a8d04STejun Heo if (unlikely(worker)) 2525401a8d04STejun Heo insert_wq_barrier(worker->current_cwq, &barr, work, worker); 2526401a8d04STejun Heo 2527401a8d04STejun Heo spin_unlock_irq(&gcwq->lock); 2528401a8d04STejun Heo 2529401a8d04STejun Heo if (unlikely(worker)) { 2530401a8d04STejun Heo wait_for_completion(&barr.done); 2531401a8d04STejun Heo destroy_work_on_stack(&barr.work); 2532401a8d04STejun Heo return true; 2533401a8d04STejun Heo } else 2534401a8d04STejun Heo return false; 2535401a8d04STejun Heo } 2536401a8d04STejun Heo 2537401a8d04STejun Heo static bool wait_on_work(struct work_struct *work) 2538401a8d04STejun Heo { 2539401a8d04STejun Heo bool ret = false; 2540401a8d04STejun Heo int cpu; 2541401a8d04STejun Heo 2542401a8d04STejun Heo might_sleep(); 2543401a8d04STejun Heo 2544401a8d04STejun Heo lock_map_acquire(&work->lockdep_map); 2545401a8d04STejun Heo lock_map_release(&work->lockdep_map); 2546401a8d04STejun Heo 2547401a8d04STejun Heo for_each_gcwq_cpu(cpu) 2548401a8d04STejun Heo ret |= wait_on_cpu_work(get_gcwq(cpu), work); 2549401a8d04STejun Heo return ret; 2550401a8d04STejun Heo } 2551401a8d04STejun Heo 255209383498STejun Heo /** 255309383498STejun Heo * flush_work_sync - wait until a work has finished execution 255409383498STejun Heo * @work: the work to flush 255509383498STejun Heo * 255609383498STejun Heo * Wait until @work has finished execution. On return, it's 255709383498STejun Heo * guaranteed that all queueing instances of @work which happened 255809383498STejun Heo * before this function is called are finished. In other words, if 255909383498STejun Heo * @work hasn't been requeued since this function was called, @work is 256009383498STejun Heo * guaranteed to be idle on return. 256109383498STejun Heo * 256209383498STejun Heo * RETURNS: 256309383498STejun Heo * %true if flush_work_sync() waited for the work to finish execution, 256409383498STejun Heo * %false if it was already idle. 256509383498STejun Heo */ 256609383498STejun Heo bool flush_work_sync(struct work_struct *work) 256709383498STejun Heo { 256809383498STejun Heo struct wq_barrier barr; 256909383498STejun Heo bool pending, waited; 257009383498STejun Heo 257109383498STejun Heo /* we'll wait for executions separately, queue barr only if pending */ 257209383498STejun Heo pending = start_flush_work(work, &barr, false); 257309383498STejun Heo 257409383498STejun Heo /* wait for executions to finish */ 257509383498STejun Heo waited = wait_on_work(work); 257609383498STejun Heo 257709383498STejun Heo /* wait for the pending one */ 257809383498STejun Heo if (pending) { 257909383498STejun Heo wait_for_completion(&barr.done); 258009383498STejun Heo destroy_work_on_stack(&barr.work); 258109383498STejun Heo } 258209383498STejun Heo 258309383498STejun Heo return pending || waited; 258409383498STejun Heo } 258509383498STejun Heo EXPORT_SYMBOL_GPL(flush_work_sync); 258609383498STejun Heo 25876e84d644SOleg Nesterov /* 25881f1f642eSOleg Nesterov * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit, 25896e84d644SOleg Nesterov * so this work can't be re-armed in any way. 25906e84d644SOleg Nesterov */ 25916e84d644SOleg Nesterov static int try_to_grab_pending(struct work_struct *work) 25926e84d644SOleg Nesterov { 25938b03ae3cSTejun Heo struct global_cwq *gcwq; 25941f1f642eSOleg Nesterov int ret = -1; 25956e84d644SOleg Nesterov 259622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 25971f1f642eSOleg Nesterov return 0; 25986e84d644SOleg Nesterov 25996e84d644SOleg Nesterov /* 26006e84d644SOleg Nesterov * The queueing is in progress, or it is already queued. Try to 26016e84d644SOleg Nesterov * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 26026e84d644SOleg Nesterov */ 26037a22ad75STejun Heo gcwq = get_work_gcwq(work); 26047a22ad75STejun Heo if (!gcwq) 26056e84d644SOleg Nesterov return ret; 26066e84d644SOleg Nesterov 26078b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 26086e84d644SOleg Nesterov if (!list_empty(&work->entry)) { 26096e84d644SOleg Nesterov /* 26107a22ad75STejun Heo * This work is queued, but perhaps we locked the wrong gcwq. 26116e84d644SOleg Nesterov * In that case we must see the new value after rmb(), see 26126e84d644SOleg Nesterov * insert_work()->wmb(). 26136e84d644SOleg Nesterov */ 26146e84d644SOleg Nesterov smp_rmb(); 26157a22ad75STejun Heo if (gcwq == get_work_gcwq(work)) { 2616dc186ad7SThomas Gleixner debug_work_deactivate(work); 26176e84d644SOleg Nesterov list_del_init(&work->entry); 26187a22ad75STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 26198a2e8e5dSTejun Heo get_work_color(work), 26208a2e8e5dSTejun Heo *work_data_bits(work) & WORK_STRUCT_DELAYED); 26216e84d644SOleg Nesterov ret = 1; 26226e84d644SOleg Nesterov } 26236e84d644SOleg Nesterov } 26248b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 26256e84d644SOleg Nesterov 26266e84d644SOleg Nesterov return ret; 26276e84d644SOleg Nesterov } 26286e84d644SOleg Nesterov 2629401a8d04STejun Heo static bool __cancel_work_timer(struct work_struct *work, 26301f1f642eSOleg Nesterov struct timer_list* timer) 26311f1f642eSOleg Nesterov { 26321f1f642eSOleg Nesterov int ret; 26331f1f642eSOleg Nesterov 26341f1f642eSOleg Nesterov do { 26351f1f642eSOleg Nesterov ret = (timer && likely(del_timer(timer))); 26361f1f642eSOleg Nesterov if (!ret) 26371f1f642eSOleg Nesterov ret = try_to_grab_pending(work); 26381f1f642eSOleg Nesterov wait_on_work(work); 26391f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 26401f1f642eSOleg Nesterov 26417a22ad75STejun Heo clear_work_data(work); 26421f1f642eSOleg Nesterov return ret; 26431f1f642eSOleg Nesterov } 26441f1f642eSOleg Nesterov 26456e84d644SOleg Nesterov /** 2646401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2647401a8d04STejun Heo * @work: the work to cancel 26486e84d644SOleg Nesterov * 2649401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2650401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2651401a8d04STejun Heo * another workqueue. On return from this function, @work is 2652401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 26531f1f642eSOleg Nesterov * 2654401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2655401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 26566e84d644SOleg Nesterov * 2657401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 26586e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2659401a8d04STejun Heo * 2660401a8d04STejun Heo * RETURNS: 2661401a8d04STejun Heo * %true if @work was pending, %false otherwise. 26626e84d644SOleg Nesterov */ 2663401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 26646e84d644SOleg Nesterov { 26651f1f642eSOleg Nesterov return __cancel_work_timer(work, NULL); 2666b89deed3SOleg Nesterov } 266728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2668b89deed3SOleg Nesterov 26696e84d644SOleg Nesterov /** 2670401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2671401a8d04STejun Heo * @dwork: the delayed work to flush 26726e84d644SOleg Nesterov * 2673401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2674401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2675401a8d04STejun Heo * considers the last queueing instance of @dwork. 26761f1f642eSOleg Nesterov * 2677401a8d04STejun Heo * RETURNS: 2678401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2679401a8d04STejun Heo * %false if it was already idle. 26806e84d644SOleg Nesterov */ 2681401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2682401a8d04STejun Heo { 2683401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 2684401a8d04STejun Heo __queue_work(raw_smp_processor_id(), 2685401a8d04STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 2686401a8d04STejun Heo return flush_work(&dwork->work); 2687401a8d04STejun Heo } 2688401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2689401a8d04STejun Heo 2690401a8d04STejun Heo /** 269109383498STejun Heo * flush_delayed_work_sync - wait for a dwork to finish 269209383498STejun Heo * @dwork: the delayed work to flush 269309383498STejun Heo * 269409383498STejun Heo * Delayed timer is cancelled and the pending work is queued for 269509383498STejun Heo * execution immediately. Other than timer handling, its behavior 269609383498STejun Heo * is identical to flush_work_sync(). 269709383498STejun Heo * 269809383498STejun Heo * RETURNS: 269909383498STejun Heo * %true if flush_work_sync() waited for the work to finish execution, 270009383498STejun Heo * %false if it was already idle. 270109383498STejun Heo */ 270209383498STejun Heo bool flush_delayed_work_sync(struct delayed_work *dwork) 270309383498STejun Heo { 270409383498STejun Heo if (del_timer_sync(&dwork->timer)) 270509383498STejun Heo __queue_work(raw_smp_processor_id(), 270609383498STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 270709383498STejun Heo return flush_work_sync(&dwork->work); 270809383498STejun Heo } 270909383498STejun Heo EXPORT_SYMBOL(flush_delayed_work_sync); 271009383498STejun Heo 271109383498STejun Heo /** 2712401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2713401a8d04STejun Heo * @dwork: the delayed work cancel 2714401a8d04STejun Heo * 2715401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2716401a8d04STejun Heo * 2717401a8d04STejun Heo * RETURNS: 2718401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2719401a8d04STejun Heo */ 2720401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 27216e84d644SOleg Nesterov { 27221f1f642eSOleg Nesterov return __cancel_work_timer(&dwork->work, &dwork->timer); 27236e84d644SOleg Nesterov } 2724f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 27251da177e4SLinus Torvalds 27260fcb78c2SRolf Eike Beer /** 27270fcb78c2SRolf Eike Beer * schedule_work - put work task in global workqueue 27280fcb78c2SRolf Eike Beer * @work: job to be done 27290fcb78c2SRolf Eike Beer * 27305b0f437dSBart Van Assche * Returns zero if @work was already on the kernel-global workqueue and 27315b0f437dSBart Van Assche * non-zero otherwise. 27325b0f437dSBart Van Assche * 27335b0f437dSBart Van Assche * This puts a job in the kernel-global workqueue if it was not already 27345b0f437dSBart Van Assche * queued and leaves it in the same position on the kernel-global 27355b0f437dSBart Van Assche * workqueue otherwise. 27360fcb78c2SRolf Eike Beer */ 27377ad5b3a5SHarvey Harrison int schedule_work(struct work_struct *work) 27381da177e4SLinus Torvalds { 2739d320c038STejun Heo return queue_work(system_wq, work); 27401da177e4SLinus Torvalds } 2741ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_work); 27421da177e4SLinus Torvalds 2743c1a220e7SZhang Rui /* 2744c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2745c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2746c1a220e7SZhang Rui * @work: job to be done 2747c1a220e7SZhang Rui * 2748c1a220e7SZhang Rui * This puts a job on a specific cpu 2749c1a220e7SZhang Rui */ 2750c1a220e7SZhang Rui int schedule_work_on(int cpu, struct work_struct *work) 2751c1a220e7SZhang Rui { 2752d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2753c1a220e7SZhang Rui } 2754c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2755c1a220e7SZhang Rui 27560fcb78c2SRolf Eike Beer /** 27570fcb78c2SRolf Eike Beer * schedule_delayed_work - put work task in global workqueue after delay 275852bad64dSDavid Howells * @dwork: job to be done 275952bad64dSDavid Howells * @delay: number of jiffies to wait or 0 for immediate execution 27600fcb78c2SRolf Eike Beer * 27610fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 27620fcb78c2SRolf Eike Beer * workqueue. 27630fcb78c2SRolf Eike Beer */ 27647ad5b3a5SHarvey Harrison int schedule_delayed_work(struct delayed_work *dwork, 276582f67cd9SIngo Molnar unsigned long delay) 27661da177e4SLinus Torvalds { 2767d320c038STejun Heo return queue_delayed_work(system_wq, dwork, delay); 27681da177e4SLinus Torvalds } 2769ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work); 27701da177e4SLinus Torvalds 27710fcb78c2SRolf Eike Beer /** 27720fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 27730fcb78c2SRolf Eike Beer * @cpu: cpu to use 277452bad64dSDavid Howells * @dwork: job to be done 27750fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 27760fcb78c2SRolf Eike Beer * 27770fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 27780fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 27790fcb78c2SRolf Eike Beer */ 27801da177e4SLinus Torvalds int schedule_delayed_work_on(int cpu, 278152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 27821da177e4SLinus Torvalds { 2783d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 27841da177e4SLinus Torvalds } 2785ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 27861da177e4SLinus Torvalds 2787b6136773SAndrew Morton /** 278831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2789b6136773SAndrew Morton * @func: the function to call 2790b6136773SAndrew Morton * 279131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 279231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2793b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 279431ddd871STejun Heo * 279531ddd871STejun Heo * RETURNS: 279631ddd871STejun Heo * 0 on success, -errno on failure. 2797b6136773SAndrew Morton */ 279865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 279915316ba8SChristoph Lameter { 280015316ba8SChristoph Lameter int cpu; 280138f51568SNamhyung Kim struct work_struct __percpu *works; 280215316ba8SChristoph Lameter 2803b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2804b6136773SAndrew Morton if (!works) 280515316ba8SChristoph Lameter return -ENOMEM; 2806b6136773SAndrew Morton 280795402b38SGautham R Shenoy get_online_cpus(); 280893981800STejun Heo 280915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 28109bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 28119bfb1839SIngo Molnar 28129bfb1839SIngo Molnar INIT_WORK(work, func); 28138de6d308SOleg Nesterov schedule_work_on(cpu, work); 281415316ba8SChristoph Lameter } 281593981800STejun Heo 281693981800STejun Heo for_each_online_cpu(cpu) 28178616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 281893981800STejun Heo 281995402b38SGautham R Shenoy put_online_cpus(); 2820b6136773SAndrew Morton free_percpu(works); 282115316ba8SChristoph Lameter return 0; 282215316ba8SChristoph Lameter } 282315316ba8SChristoph Lameter 2824eef6a7d5SAlan Stern /** 2825eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2826eef6a7d5SAlan Stern * 2827eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2828eef6a7d5SAlan Stern * completion. 2829eef6a7d5SAlan Stern * 2830eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2831eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2832eef6a7d5SAlan Stern * will lead to deadlock: 2833eef6a7d5SAlan Stern * 2834eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2835eef6a7d5SAlan Stern * a lock held by your code or its caller. 2836eef6a7d5SAlan Stern * 2837eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2838eef6a7d5SAlan Stern * 2839eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2840eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2841eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2842eef6a7d5SAlan Stern * 2843eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2844eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2845eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2846eef6a7d5SAlan Stern * cancel_work_sync() instead. 2847eef6a7d5SAlan Stern */ 28481da177e4SLinus Torvalds void flush_scheduled_work(void) 28491da177e4SLinus Torvalds { 2850d320c038STejun Heo flush_workqueue(system_wq); 28511da177e4SLinus Torvalds } 2852ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 28531da177e4SLinus Torvalds 28541da177e4SLinus Torvalds /** 28551fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 28561fa44ecaSJames Bottomley * @fn: the function to execute 28571fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 28581fa44ecaSJames Bottomley * be available when the work executes) 28591fa44ecaSJames Bottomley * 28601fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 28611fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 28621fa44ecaSJames Bottomley * 28631fa44ecaSJames Bottomley * Returns: 0 - function was executed 28641fa44ecaSJames Bottomley * 1 - function was scheduled for execution 28651fa44ecaSJames Bottomley */ 286665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 28671fa44ecaSJames Bottomley { 28681fa44ecaSJames Bottomley if (!in_interrupt()) { 286965f27f38SDavid Howells fn(&ew->work); 28701fa44ecaSJames Bottomley return 0; 28711fa44ecaSJames Bottomley } 28721fa44ecaSJames Bottomley 287365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 28741fa44ecaSJames Bottomley schedule_work(&ew->work); 28751fa44ecaSJames Bottomley 28761fa44ecaSJames Bottomley return 1; 28771fa44ecaSJames Bottomley } 28781fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 28791fa44ecaSJames Bottomley 28801da177e4SLinus Torvalds int keventd_up(void) 28811da177e4SLinus Torvalds { 2882d320c038STejun Heo return system_wq != NULL; 28831da177e4SLinus Torvalds } 28841da177e4SLinus Torvalds 2885bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 28861da177e4SLinus Torvalds { 28873af24433SOleg Nesterov /* 28880f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 28890f900049STejun Heo * Make sure that the alignment isn't lower than that of 28900f900049STejun Heo * unsigned long long. 28913af24433SOleg Nesterov */ 28920f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 28930f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 28940f900049STejun Heo __alignof__(unsigned long long)); 2895931ac77eSTejun Heo #ifdef CONFIG_SMP 2896931ac77eSTejun Heo bool percpu = !(wq->flags & WQ_UNBOUND); 2897931ac77eSTejun Heo #else 2898931ac77eSTejun Heo bool percpu = false; 2899931ac77eSTejun Heo #endif 29003af24433SOleg Nesterov 2901931ac77eSTejun Heo if (percpu) 2902f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 2903931ac77eSTejun Heo else { 29040f900049STejun Heo void *ptr; 2905e1d8aa9fSFrederic Weisbecker 29060f900049STejun Heo /* 2907f3421797STejun Heo * Allocate enough room to align cwq and put an extra 2908f3421797STejun Heo * pointer at the end pointing back to the originally 2909f3421797STejun Heo * allocated pointer which will be used for free. 29100f900049STejun Heo */ 2911bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 2912bdbc5dd7STejun Heo if (ptr) { 2913bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 2914bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 2915bdbc5dd7STejun Heo } 29163af24433SOleg Nesterov } 29173af24433SOleg Nesterov 291852605627SDavid Howells /* just in case, make sure it's actually aligned 291952605627SDavid Howells * - this is affected by PERCPU() alignment in vmlinux.lds.S 292052605627SDavid Howells */ 2921bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 2922bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 29230f900049STejun Heo } 29240f900049STejun Heo 2925bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 292606ba38a9SOleg Nesterov { 2927931ac77eSTejun Heo #ifdef CONFIG_SMP 2928931ac77eSTejun Heo bool percpu = !(wq->flags & WQ_UNBOUND); 2929931ac77eSTejun Heo #else 2930931ac77eSTejun Heo bool percpu = false; 2931931ac77eSTejun Heo #endif 293206ba38a9SOleg Nesterov 2933931ac77eSTejun Heo if (percpu) 2934bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 2935f3421797STejun Heo else if (wq->cpu_wq.single) { 2936f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 2937f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 293806ba38a9SOleg Nesterov } 293906ba38a9SOleg Nesterov } 294006ba38a9SOleg Nesterov 2941f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 2942f3421797STejun Heo const char *name) 2943b71ab8c2STejun Heo { 2944f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 2945f3421797STejun Heo 2946f3421797STejun Heo if (max_active < 1 || max_active > lim) 2947b71ab8c2STejun Heo printk(KERN_WARNING "workqueue: max_active %d requested for %s " 2948b71ab8c2STejun Heo "is out of range, clamping between %d and %d\n", 2949f3421797STejun Heo max_active, name, 1, lim); 2950b71ab8c2STejun Heo 2951f3421797STejun Heo return clamp_val(max_active, 1, lim); 2952b71ab8c2STejun Heo } 2953b71ab8c2STejun Heo 2954d320c038STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *name, 295597e37d7bSTejun Heo unsigned int flags, 29561e19ffc6STejun Heo int max_active, 2957eb13ba87SJohannes Berg struct lock_class_key *key, 2958eb13ba87SJohannes Berg const char *lock_name) 29593af24433SOleg Nesterov { 29603af24433SOleg Nesterov struct workqueue_struct *wq; 2961c34056a3STejun Heo unsigned int cpu; 29623af24433SOleg Nesterov 2963f3421797STejun Heo /* 29646370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 29656370a6adSTejun Heo * have a rescuer to guarantee forward progress. 29666370a6adSTejun Heo */ 29676370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 29686370a6adSTejun Heo flags |= WQ_RESCUER; 29696370a6adSTejun Heo 29706370a6adSTejun Heo /* 2971f3421797STejun Heo * Unbound workqueues aren't concurrency managed and should be 2972f3421797STejun Heo * dispatched to workers immediately. 2973f3421797STejun Heo */ 2974f3421797STejun Heo if (flags & WQ_UNBOUND) 2975f3421797STejun Heo flags |= WQ_HIGHPRI; 2976f3421797STejun Heo 2977d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 2978f3421797STejun Heo max_active = wq_clamp_max_active(max_active, flags, name); 29793af24433SOleg Nesterov 29803af24433SOleg Nesterov wq = kzalloc(sizeof(*wq), GFP_KERNEL); 29813af24433SOleg Nesterov if (!wq) 29824690c4abSTejun Heo goto err; 29833af24433SOleg Nesterov 298497e37d7bSTejun Heo wq->flags = flags; 2985a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 298673f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 298773f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 298873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 298973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 29903af24433SOleg Nesterov 29913af24433SOleg Nesterov wq->name = name; 2992eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 2993cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 29943af24433SOleg Nesterov 2995bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 2996bdbc5dd7STejun Heo goto err; 2997bdbc5dd7STejun Heo 2998f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 29991537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 30008b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 30011537663fSTejun Heo 30020f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 30038b03ae3cSTejun Heo cwq->gcwq = gcwq; 3004c34056a3STejun Heo cwq->wq = wq; 300573f53c4aSTejun Heo cwq->flush_color = -1; 30061e19ffc6STejun Heo cwq->max_active = max_active; 30071e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3008e22bee78STejun Heo } 30091537663fSTejun Heo 3010e22bee78STejun Heo if (flags & WQ_RESCUER) { 3011e22bee78STejun Heo struct worker *rescuer; 3012e22bee78STejun Heo 3013f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3014e22bee78STejun Heo goto err; 3015e22bee78STejun Heo 3016e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3017e22bee78STejun Heo if (!rescuer) 3018e22bee78STejun Heo goto err; 3019e22bee78STejun Heo 3020e22bee78STejun Heo rescuer->task = kthread_create(rescuer_thread, wq, "%s", name); 3021e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3022e22bee78STejun Heo goto err; 3023e22bee78STejun Heo 3024e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3025e22bee78STejun Heo wake_up_process(rescuer->task); 30263af24433SOleg Nesterov } 30271537663fSTejun Heo 30283af24433SOleg Nesterov /* 3029a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3030a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3031a0a1a5fdSTejun Heo * workqueue to workqueues list. 30323af24433SOleg Nesterov */ 30333af24433SOleg Nesterov spin_lock(&workqueue_lock); 3034a0a1a5fdSTejun Heo 303558a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3036f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3037a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3038a0a1a5fdSTejun Heo 30393af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3040a0a1a5fdSTejun Heo 30413af24433SOleg Nesterov spin_unlock(&workqueue_lock); 30423af24433SOleg Nesterov 30433af24433SOleg Nesterov return wq; 30444690c4abSTejun Heo err: 30454690c4abSTejun Heo if (wq) { 3046bdbc5dd7STejun Heo free_cwqs(wq); 3047f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3048e22bee78STejun Heo kfree(wq->rescuer); 30494690c4abSTejun Heo kfree(wq); 30503af24433SOleg Nesterov } 30514690c4abSTejun Heo return NULL; 30521da177e4SLinus Torvalds } 3053d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 30541da177e4SLinus Torvalds 30553af24433SOleg Nesterov /** 30563af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 30573af24433SOleg Nesterov * @wq: target workqueue 30583af24433SOleg Nesterov * 30593af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 30603af24433SOleg Nesterov */ 30613af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 30623af24433SOleg Nesterov { 3063c8e55f36STejun Heo unsigned int cpu; 30643af24433SOleg Nesterov 30659c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 30669c5a2ba7STejun Heo drain_workqueue(wq); 3067c8efcc25STejun Heo 3068a0a1a5fdSTejun Heo /* 3069a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3070a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3071a0a1a5fdSTejun Heo */ 307295402b38SGautham R Shenoy spin_lock(&workqueue_lock); 30733af24433SOleg Nesterov list_del(&wq->list); 307495402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 30753af24433SOleg Nesterov 3076e22bee78STejun Heo /* sanity check */ 3077f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 307873f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 307973f53c4aSTejun Heo int i; 30803af24433SOleg Nesterov 308173f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 308273f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 30831e19ffc6STejun Heo BUG_ON(cwq->nr_active); 30841e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 308573f53c4aSTejun Heo } 30861537663fSTejun Heo 3087e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3088e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3089f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 30908d9df9f0SXiaotian Feng kfree(wq->rescuer); 3091e22bee78STejun Heo } 3092e22bee78STejun Heo 3093bdbc5dd7STejun Heo free_cwqs(wq); 30943af24433SOleg Nesterov kfree(wq); 30953af24433SOleg Nesterov } 30963af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 30973af24433SOleg Nesterov 3098dcd989cbSTejun Heo /** 3099dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3100dcd989cbSTejun Heo * @wq: target workqueue 3101dcd989cbSTejun Heo * @max_active: new max_active value. 3102dcd989cbSTejun Heo * 3103dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3104dcd989cbSTejun Heo * 3105dcd989cbSTejun Heo * CONTEXT: 3106dcd989cbSTejun Heo * Don't call from IRQ context. 3107dcd989cbSTejun Heo */ 3108dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3109dcd989cbSTejun Heo { 3110dcd989cbSTejun Heo unsigned int cpu; 3111dcd989cbSTejun Heo 3112f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3113dcd989cbSTejun Heo 3114dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3115dcd989cbSTejun Heo 3116dcd989cbSTejun Heo wq->saved_max_active = max_active; 3117dcd989cbSTejun Heo 3118f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 3119dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3120dcd989cbSTejun Heo 3121dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 3122dcd989cbSTejun Heo 312358a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 3124dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 3125dcd989cbSTejun Heo get_cwq(gcwq->cpu, wq)->max_active = max_active; 3126dcd989cbSTejun Heo 3127dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 3128dcd989cbSTejun Heo } 3129dcd989cbSTejun Heo 3130dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3131dcd989cbSTejun Heo } 3132dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3133dcd989cbSTejun Heo 3134dcd989cbSTejun Heo /** 3135dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3136dcd989cbSTejun Heo * @cpu: CPU in question 3137dcd989cbSTejun Heo * @wq: target workqueue 3138dcd989cbSTejun Heo * 3139dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3140dcd989cbSTejun Heo * no synchronization around this function and the test result is 3141dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3142dcd989cbSTejun Heo * 3143dcd989cbSTejun Heo * RETURNS: 3144dcd989cbSTejun Heo * %true if congested, %false otherwise. 3145dcd989cbSTejun Heo */ 3146dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3147dcd989cbSTejun Heo { 3148dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3149dcd989cbSTejun Heo 3150dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3151dcd989cbSTejun Heo } 3152dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3153dcd989cbSTejun Heo 3154dcd989cbSTejun Heo /** 3155dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 3156dcd989cbSTejun Heo * @work: the work of interest 3157dcd989cbSTejun Heo * 3158dcd989cbSTejun Heo * RETURNS: 3159bdbc5dd7STejun Heo * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. 3160dcd989cbSTejun Heo */ 3161dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 3162dcd989cbSTejun Heo { 3163dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3164dcd989cbSTejun Heo 3165bdbc5dd7STejun Heo return gcwq ? gcwq->cpu : WORK_CPU_NONE; 3166dcd989cbSTejun Heo } 3167dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 3168dcd989cbSTejun Heo 3169dcd989cbSTejun Heo /** 3170dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3171dcd989cbSTejun Heo * @work: the work to be tested 3172dcd989cbSTejun Heo * 3173dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3174dcd989cbSTejun Heo * synchronization around this function and the test result is 3175dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3176dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 3177dcd989cbSTejun Heo * running state. 3178dcd989cbSTejun Heo * 3179dcd989cbSTejun Heo * RETURNS: 3180dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3181dcd989cbSTejun Heo */ 3182dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3183dcd989cbSTejun Heo { 3184dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3185dcd989cbSTejun Heo unsigned long flags; 3186dcd989cbSTejun Heo unsigned int ret = 0; 3187dcd989cbSTejun Heo 3188dcd989cbSTejun Heo if (!gcwq) 3189dcd989cbSTejun Heo return false; 3190dcd989cbSTejun Heo 3191dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 3192dcd989cbSTejun Heo 3193dcd989cbSTejun Heo if (work_pending(work)) 3194dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3195dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 3196dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3197dcd989cbSTejun Heo 3198dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 3199dcd989cbSTejun Heo 3200dcd989cbSTejun Heo return ret; 3201dcd989cbSTejun Heo } 3202dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3203dcd989cbSTejun Heo 3204db7bccf4STejun Heo /* 3205db7bccf4STejun Heo * CPU hotplug. 3206db7bccf4STejun Heo * 3207e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3208e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3209e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 3210e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 3211e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 3212e22bee78STejun Heo * blocked draining impractical. 3213e22bee78STejun Heo * 3214e22bee78STejun Heo * This is solved by allowing a gcwq to be detached from CPU, running 3215e22bee78STejun Heo * it with unbound (rogue) workers and allowing it to be reattached 3216e22bee78STejun Heo * later if the cpu comes back online. A separate thread is created 3217e22bee78STejun Heo * to govern a gcwq in such state and is called the trustee of the 3218e22bee78STejun Heo * gcwq. 3219db7bccf4STejun Heo * 3220db7bccf4STejun Heo * Trustee states and their descriptions. 3221db7bccf4STejun Heo * 3222db7bccf4STejun Heo * START Command state used on startup. On CPU_DOWN_PREPARE, a 3223db7bccf4STejun Heo * new trustee is started with this state. 3224db7bccf4STejun Heo * 3225db7bccf4STejun Heo * IN_CHARGE Once started, trustee will enter this state after 3226e22bee78STejun Heo * assuming the manager role and making all existing 3227e22bee78STejun Heo * workers rogue. DOWN_PREPARE waits for trustee to 3228e22bee78STejun Heo * enter this state. After reaching IN_CHARGE, trustee 3229e22bee78STejun Heo * tries to execute the pending worklist until it's empty 3230e22bee78STejun Heo * and the state is set to BUTCHER, or the state is set 3231e22bee78STejun Heo * to RELEASE. 3232db7bccf4STejun Heo * 3233db7bccf4STejun Heo * BUTCHER Command state which is set by the cpu callback after 3234db7bccf4STejun Heo * the cpu has went down. Once this state is set trustee 3235db7bccf4STejun Heo * knows that there will be no new works on the worklist 3236db7bccf4STejun Heo * and once the worklist is empty it can proceed to 3237db7bccf4STejun Heo * killing idle workers. 3238db7bccf4STejun Heo * 3239db7bccf4STejun Heo * RELEASE Command state which is set by the cpu callback if the 3240db7bccf4STejun Heo * cpu down has been canceled or it has come online 3241db7bccf4STejun Heo * again. After recognizing this state, trustee stops 3242e22bee78STejun Heo * trying to drain or butcher and clears ROGUE, rebinds 3243e22bee78STejun Heo * all remaining workers back to the cpu and releases 3244e22bee78STejun Heo * manager role. 3245db7bccf4STejun Heo * 3246db7bccf4STejun Heo * DONE Trustee will enter this state after BUTCHER or RELEASE 3247db7bccf4STejun Heo * is complete. 3248db7bccf4STejun Heo * 3249db7bccf4STejun Heo * trustee CPU draining 3250db7bccf4STejun Heo * took over down complete 3251db7bccf4STejun Heo * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE 3252db7bccf4STejun Heo * | | ^ 3253db7bccf4STejun Heo * | CPU is back online v return workers | 3254db7bccf4STejun Heo * ----------------> RELEASE -------------- 3255db7bccf4STejun Heo */ 3256db7bccf4STejun Heo 3257db7bccf4STejun Heo /** 3258db7bccf4STejun Heo * trustee_wait_event_timeout - timed event wait for trustee 3259db7bccf4STejun Heo * @cond: condition to wait for 3260db7bccf4STejun Heo * @timeout: timeout in jiffies 3261db7bccf4STejun Heo * 3262db7bccf4STejun Heo * wait_event_timeout() for trustee to use. Handles locking and 3263db7bccf4STejun Heo * checks for RELEASE request. 3264db7bccf4STejun Heo * 3265db7bccf4STejun Heo * CONTEXT: 3266db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3267db7bccf4STejun Heo * multiple times. To be used by trustee. 3268db7bccf4STejun Heo * 3269db7bccf4STejun Heo * RETURNS: 3270db7bccf4STejun Heo * Positive indicating left time if @cond is satisfied, 0 if timed 3271db7bccf4STejun Heo * out, -1 if canceled. 3272db7bccf4STejun Heo */ 3273db7bccf4STejun Heo #define trustee_wait_event_timeout(cond, timeout) ({ \ 3274db7bccf4STejun Heo long __ret = (timeout); \ 3275db7bccf4STejun Heo while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \ 3276db7bccf4STejun Heo __ret) { \ 3277db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); \ 3278db7bccf4STejun Heo __wait_event_timeout(gcwq->trustee_wait, (cond) || \ 3279db7bccf4STejun Heo (gcwq->trustee_state == TRUSTEE_RELEASE), \ 3280db7bccf4STejun Heo __ret); \ 3281db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); \ 3282db7bccf4STejun Heo } \ 3283db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \ 3284db7bccf4STejun Heo }) 3285db7bccf4STejun Heo 3286db7bccf4STejun Heo /** 3287db7bccf4STejun Heo * trustee_wait_event - event wait for trustee 3288db7bccf4STejun Heo * @cond: condition to wait for 3289db7bccf4STejun Heo * 3290db7bccf4STejun Heo * wait_event() for trustee to use. Automatically handles locking and 3291db7bccf4STejun Heo * checks for CANCEL request. 3292db7bccf4STejun Heo * 3293db7bccf4STejun Heo * CONTEXT: 3294db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3295db7bccf4STejun Heo * multiple times. To be used by trustee. 3296db7bccf4STejun Heo * 3297db7bccf4STejun Heo * RETURNS: 3298db7bccf4STejun Heo * 0 if @cond is satisfied, -1 if canceled. 3299db7bccf4STejun Heo */ 3300db7bccf4STejun Heo #define trustee_wait_event(cond) ({ \ 3301db7bccf4STejun Heo long __ret1; \ 3302db7bccf4STejun Heo __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\ 3303db7bccf4STejun Heo __ret1 < 0 ? -1 : 0; \ 3304db7bccf4STejun Heo }) 3305db7bccf4STejun Heo 3306db7bccf4STejun Heo static int __cpuinit trustee_thread(void *__gcwq) 3307db7bccf4STejun Heo { 3308db7bccf4STejun Heo struct global_cwq *gcwq = __gcwq; 3309db7bccf4STejun Heo struct worker *worker; 3310e22bee78STejun Heo struct work_struct *work; 3311db7bccf4STejun Heo struct hlist_node *pos; 3312e22bee78STejun Heo long rc; 3313db7bccf4STejun Heo int i; 3314db7bccf4STejun Heo 3315db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3316db7bccf4STejun Heo 3317db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3318db7bccf4STejun Heo /* 3319e22bee78STejun Heo * Claim the manager position and make all workers rogue. 3320e22bee78STejun Heo * Trustee must be bound to the target cpu and can't be 3321e22bee78STejun Heo * cancelled. 3322db7bccf4STejun Heo */ 3323db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3324e22bee78STejun Heo rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS)); 3325e22bee78STejun Heo BUG_ON(rc < 0); 3326e22bee78STejun Heo 3327e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGING_WORKERS; 3328db7bccf4STejun Heo 3329db7bccf4STejun Heo list_for_each_entry(worker, &gcwq->idle_list, entry) 3330cb444766STejun Heo worker->flags |= WORKER_ROGUE; 3331db7bccf4STejun Heo 3332db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 3333cb444766STejun Heo worker->flags |= WORKER_ROGUE; 3334db7bccf4STejun Heo 3335db7bccf4STejun Heo /* 3336e22bee78STejun Heo * Call schedule() so that we cross rq->lock and thus can 3337e22bee78STejun Heo * guarantee sched callbacks see the rogue flag. This is 3338e22bee78STejun Heo * necessary as scheduler callbacks may be invoked from other 3339e22bee78STejun Heo * cpus. 3340e22bee78STejun Heo */ 3341e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3342e22bee78STejun Heo schedule(); 3343e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3344e22bee78STejun Heo 3345e22bee78STejun Heo /* 3346cb444766STejun Heo * Sched callbacks are disabled now. Zap nr_running. After 3347cb444766STejun Heo * this, nr_running stays zero and need_more_worker() and 3348cb444766STejun Heo * keep_working() are always true as long as the worklist is 3349cb444766STejun Heo * not empty. 3350e22bee78STejun Heo */ 3351cb444766STejun Heo atomic_set(get_gcwq_nr_running(gcwq->cpu), 0); 3352e22bee78STejun Heo 3353e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3354e22bee78STejun Heo del_timer_sync(&gcwq->idle_timer); 3355e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3356e22bee78STejun Heo 3357e22bee78STejun Heo /* 3358db7bccf4STejun Heo * We're now in charge. Notify and proceed to drain. We need 3359db7bccf4STejun Heo * to keep the gcwq running during the whole CPU down 3360db7bccf4STejun Heo * procedure as other cpu hotunplug callbacks may need to 3361db7bccf4STejun Heo * flush currently running tasks. 3362db7bccf4STejun Heo */ 3363db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_IN_CHARGE; 3364db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3365db7bccf4STejun Heo 3366db7bccf4STejun Heo /* 3367db7bccf4STejun Heo * The original cpu is in the process of dying and may go away 3368db7bccf4STejun Heo * anytime now. When that happens, we and all workers would 3369e22bee78STejun Heo * be migrated to other cpus. Try draining any left work. We 3370e22bee78STejun Heo * want to get it over with ASAP - spam rescuers, wake up as 3371e22bee78STejun Heo * many idlers as necessary and create new ones till the 3372e22bee78STejun Heo * worklist is empty. Note that if the gcwq is frozen, there 337358a69cb4STejun Heo * may be frozen works in freezable cwqs. Don't declare 3374e22bee78STejun Heo * completion while frozen. 3375db7bccf4STejun Heo */ 3376db7bccf4STejun Heo while (gcwq->nr_workers != gcwq->nr_idle || 3377db7bccf4STejun Heo gcwq->flags & GCWQ_FREEZING || 3378db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_IN_CHARGE) { 3379e22bee78STejun Heo int nr_works = 0; 3380e22bee78STejun Heo 3381e22bee78STejun Heo list_for_each_entry(work, &gcwq->worklist, entry) { 3382e22bee78STejun Heo send_mayday(work); 3383e22bee78STejun Heo nr_works++; 3384e22bee78STejun Heo } 3385e22bee78STejun Heo 3386e22bee78STejun Heo list_for_each_entry(worker, &gcwq->idle_list, entry) { 3387e22bee78STejun Heo if (!nr_works--) 3388e22bee78STejun Heo break; 3389e22bee78STejun Heo wake_up_process(worker->task); 3390e22bee78STejun Heo } 3391e22bee78STejun Heo 3392e22bee78STejun Heo if (need_to_create_worker(gcwq)) { 3393e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3394e22bee78STejun Heo worker = create_worker(gcwq, false); 3395e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3396e22bee78STejun Heo if (worker) { 3397cb444766STejun Heo worker->flags |= WORKER_ROGUE; 3398e22bee78STejun Heo start_worker(worker); 3399e22bee78STejun Heo } 3400e22bee78STejun Heo } 3401e22bee78STejun Heo 3402db7bccf4STejun Heo /* give a breather */ 3403db7bccf4STejun Heo if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0) 3404db7bccf4STejun Heo break; 3405db7bccf4STejun Heo } 3406db7bccf4STejun Heo 3407e22bee78STejun Heo /* 3408e22bee78STejun Heo * Either all works have been scheduled and cpu is down, or 3409e22bee78STejun Heo * cpu down has already been canceled. Wait for and butcher 3410e22bee78STejun Heo * all workers till we're canceled. 3411e22bee78STejun Heo */ 3412e22bee78STejun Heo do { 3413e22bee78STejun Heo rc = trustee_wait_event(!list_empty(&gcwq->idle_list)); 3414e22bee78STejun Heo while (!list_empty(&gcwq->idle_list)) 3415e22bee78STejun Heo destroy_worker(list_first_entry(&gcwq->idle_list, 3416e22bee78STejun Heo struct worker, entry)); 3417e22bee78STejun Heo } while (gcwq->nr_workers && rc >= 0); 3418e22bee78STejun Heo 3419e22bee78STejun Heo /* 3420e22bee78STejun Heo * At this point, either draining has completed and no worker 3421e22bee78STejun Heo * is left, or cpu down has been canceled or the cpu is being 3422e22bee78STejun Heo * brought back up. There shouldn't be any idle one left. 3423e22bee78STejun Heo * Tell the remaining busy ones to rebind once it finishes the 3424e22bee78STejun Heo * currently scheduled works by scheduling the rebind_work. 3425e22bee78STejun Heo */ 3426e22bee78STejun Heo WARN_ON(!list_empty(&gcwq->idle_list)); 3427e22bee78STejun Heo 3428e22bee78STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 3429e22bee78STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 3430e22bee78STejun Heo 3431e22bee78STejun Heo /* 3432e22bee78STejun Heo * Rebind_work may race with future cpu hotplug 3433e22bee78STejun Heo * operations. Use a separate flag to mark that 3434e22bee78STejun Heo * rebinding is scheduled. 3435e22bee78STejun Heo */ 3436cb444766STejun Heo worker->flags |= WORKER_REBIND; 3437cb444766STejun Heo worker->flags &= ~WORKER_ROGUE; 3438e22bee78STejun Heo 3439e22bee78STejun Heo /* queue rebind_work, wq doesn't matter, use the default one */ 3440e22bee78STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 3441e22bee78STejun Heo work_data_bits(rebind_work))) 3442e22bee78STejun Heo continue; 3443e22bee78STejun Heo 3444e22bee78STejun Heo debug_work_activate(rebind_work); 3445d320c038STejun Heo insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work, 3446e22bee78STejun Heo worker->scheduled.next, 3447e22bee78STejun Heo work_color_to_flags(WORK_NO_COLOR)); 3448e22bee78STejun Heo } 3449e22bee78STejun Heo 3450e22bee78STejun Heo /* relinquish manager role */ 3451e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGING_WORKERS; 3452e22bee78STejun Heo 3453db7bccf4STejun Heo /* notify completion */ 3454db7bccf4STejun Heo gcwq->trustee = NULL; 3455db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3456db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3457db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3458db7bccf4STejun Heo return 0; 3459db7bccf4STejun Heo } 3460db7bccf4STejun Heo 3461db7bccf4STejun Heo /** 3462db7bccf4STejun Heo * wait_trustee_state - wait for trustee to enter the specified state 3463db7bccf4STejun Heo * @gcwq: gcwq the trustee of interest belongs to 3464db7bccf4STejun Heo * @state: target state to wait for 3465db7bccf4STejun Heo * 3466db7bccf4STejun Heo * Wait for the trustee to reach @state. DONE is already matched. 3467db7bccf4STejun Heo * 3468db7bccf4STejun Heo * CONTEXT: 3469db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3470db7bccf4STejun Heo * multiple times. To be used by cpu_callback. 3471db7bccf4STejun Heo */ 3472db7bccf4STejun Heo static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state) 347306bd6ebfSNamhyung Kim __releases(&gcwq->lock) 347406bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 3475db7bccf4STejun Heo { 3476db7bccf4STejun Heo if (!(gcwq->trustee_state == state || 3477db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE)) { 3478db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3479db7bccf4STejun Heo __wait_event(gcwq->trustee_wait, 3480db7bccf4STejun Heo gcwq->trustee_state == state || 3481db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE); 3482db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3483db7bccf4STejun Heo } 3484db7bccf4STejun Heo } 3485db7bccf4STejun Heo 34869c7b216dSChandra Seetharaman static int __devinit workqueue_cpu_callback(struct notifier_block *nfb, 34871da177e4SLinus Torvalds unsigned long action, 34881da177e4SLinus Torvalds void *hcpu) 34891da177e4SLinus Torvalds { 34903af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3491db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3492db7bccf4STejun Heo struct task_struct *new_trustee = NULL; 3493e22bee78STejun Heo struct worker *uninitialized_var(new_worker); 3494db7bccf4STejun Heo unsigned long flags; 34951da177e4SLinus Torvalds 34968bb78442SRafael J. Wysocki action &= ~CPU_TASKS_FROZEN; 34978bb78442SRafael J. Wysocki 34981da177e4SLinus Torvalds switch (action) { 3499db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3500db7bccf4STejun Heo new_trustee = kthread_create(trustee_thread, gcwq, 3501db7bccf4STejun Heo "workqueue_trustee/%d\n", cpu); 3502db7bccf4STejun Heo if (IS_ERR(new_trustee)) 3503db7bccf4STejun Heo return notifier_from_errno(PTR_ERR(new_trustee)); 3504db7bccf4STejun Heo kthread_bind(new_trustee, cpu); 3505e22bee78STejun Heo /* fall through */ 35063af24433SOleg Nesterov case CPU_UP_PREPARE: 3507e22bee78STejun Heo BUG_ON(gcwq->first_idle); 3508e22bee78STejun Heo new_worker = create_worker(gcwq, false); 3509e22bee78STejun Heo if (!new_worker) { 3510e22bee78STejun Heo if (new_trustee) 3511e22bee78STejun Heo kthread_stop(new_trustee); 3512e22bee78STejun Heo return NOTIFY_BAD; 35133af24433SOleg Nesterov } 3514db7bccf4STejun Heo } 35151537663fSTejun Heo 3516db7bccf4STejun Heo /* some are called w/ irq disabled, don't disturb irq status */ 3517db7bccf4STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 35183af24433SOleg Nesterov 35193af24433SOleg Nesterov switch (action) { 3520db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3521db7bccf4STejun Heo /* initialize trustee and tell it to acquire the gcwq */ 3522db7bccf4STejun Heo BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE); 3523db7bccf4STejun Heo gcwq->trustee = new_trustee; 3524db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_START; 3525db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3526db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE); 3527e22bee78STejun Heo /* fall through */ 35283af24433SOleg Nesterov case CPU_UP_PREPARE: 3529e22bee78STejun Heo BUG_ON(gcwq->first_idle); 3530e22bee78STejun Heo gcwq->first_idle = new_worker; 35311da177e4SLinus Torvalds break; 35321da177e4SLinus Torvalds 3533e22bee78STejun Heo case CPU_DYING: 3534e22bee78STejun Heo /* 3535e22bee78STejun Heo * Before this, the trustee and all workers except for 3536e22bee78STejun Heo * the ones which are still executing works from 3537e22bee78STejun Heo * before the last CPU down must be on the cpu. After 3538e22bee78STejun Heo * this, they'll all be diasporas. 3539e22bee78STejun Heo */ 3540e22bee78STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3541db7bccf4STejun Heo break; 3542db7bccf4STejun Heo 35433da1c84cSOleg Nesterov case CPU_POST_DEAD: 3544db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_BUTCHER; 3545e22bee78STejun Heo /* fall through */ 3546e22bee78STejun Heo case CPU_UP_CANCELED: 3547e22bee78STejun Heo destroy_worker(gcwq->first_idle); 3548e22bee78STejun Heo gcwq->first_idle = NULL; 3549db7bccf4STejun Heo break; 3550db7bccf4STejun Heo 3551db7bccf4STejun Heo case CPU_DOWN_FAILED: 35521da177e4SLinus Torvalds case CPU_ONLINE: 3553e22bee78STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 3554db7bccf4STejun Heo if (gcwq->trustee_state != TRUSTEE_DONE) { 3555db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_RELEASE; 3556db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3557db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_DONE); 3558db7bccf4STejun Heo } 35591da177e4SLinus Torvalds 3560e22bee78STejun Heo /* 3561e22bee78STejun Heo * Trustee is done and there might be no worker left. 3562e22bee78STejun Heo * Put the first_idle in and request a real manager to 3563e22bee78STejun Heo * take a look. 3564e22bee78STejun Heo */ 3565e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3566e22bee78STejun Heo kthread_bind(gcwq->first_idle->task, cpu); 3567e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3568e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGE_WORKERS; 3569e22bee78STejun Heo start_worker(gcwq->first_idle); 3570e22bee78STejun Heo gcwq->first_idle = NULL; 35711da177e4SLinus Torvalds break; 35721da177e4SLinus Torvalds } 35731da177e4SLinus Torvalds 3574db7bccf4STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 357500dfcaf7SOleg Nesterov 35761537663fSTejun Heo return notifier_from_errno(0); 35771da177e4SLinus Torvalds } 35781da177e4SLinus Torvalds 35792d3854a3SRusty Russell #ifdef CONFIG_SMP 35808ccad40dSRusty Russell 35812d3854a3SRusty Russell struct work_for_cpu { 35826b44003eSAndrew Morton struct completion completion; 35832d3854a3SRusty Russell long (*fn)(void *); 35842d3854a3SRusty Russell void *arg; 35852d3854a3SRusty Russell long ret; 35862d3854a3SRusty Russell }; 35872d3854a3SRusty Russell 35886b44003eSAndrew Morton static int do_work_for_cpu(void *_wfc) 35892d3854a3SRusty Russell { 35906b44003eSAndrew Morton struct work_for_cpu *wfc = _wfc; 35912d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 35926b44003eSAndrew Morton complete(&wfc->completion); 35936b44003eSAndrew Morton return 0; 35942d3854a3SRusty Russell } 35952d3854a3SRusty Russell 35962d3854a3SRusty Russell /** 35972d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 35982d3854a3SRusty Russell * @cpu: the cpu to run on 35992d3854a3SRusty Russell * @fn: the function to run 36002d3854a3SRusty Russell * @arg: the function arg 36012d3854a3SRusty Russell * 360231ad9081SRusty Russell * This will return the value @fn returns. 360331ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 36046b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 36052d3854a3SRusty Russell */ 36062d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 36072d3854a3SRusty Russell { 36086b44003eSAndrew Morton struct task_struct *sub_thread; 36096b44003eSAndrew Morton struct work_for_cpu wfc = { 36106b44003eSAndrew Morton .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), 36116b44003eSAndrew Morton .fn = fn, 36126b44003eSAndrew Morton .arg = arg, 36136b44003eSAndrew Morton }; 36142d3854a3SRusty Russell 36156b44003eSAndrew Morton sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); 36166b44003eSAndrew Morton if (IS_ERR(sub_thread)) 36176b44003eSAndrew Morton return PTR_ERR(sub_thread); 36186b44003eSAndrew Morton kthread_bind(sub_thread, cpu); 36196b44003eSAndrew Morton wake_up_process(sub_thread); 36206b44003eSAndrew Morton wait_for_completion(&wfc.completion); 36212d3854a3SRusty Russell return wfc.ret; 36222d3854a3SRusty Russell } 36232d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 36242d3854a3SRusty Russell #endif /* CONFIG_SMP */ 36252d3854a3SRusty Russell 3626a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3627e7577c50SRusty Russell 3628a0a1a5fdSTejun Heo /** 3629a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3630a0a1a5fdSTejun Heo * 363158a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 363258a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 363358a69cb4STejun Heo * gcwq->worklist. 3634a0a1a5fdSTejun Heo * 3635a0a1a5fdSTejun Heo * CONTEXT: 36368b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3637a0a1a5fdSTejun Heo */ 3638a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3639a0a1a5fdSTejun Heo { 3640a0a1a5fdSTejun Heo unsigned int cpu; 3641a0a1a5fdSTejun Heo 3642a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3643a0a1a5fdSTejun Heo 3644a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3645a0a1a5fdSTejun Heo workqueue_freezing = true; 3646a0a1a5fdSTejun Heo 3647f3421797STejun Heo for_each_gcwq_cpu(cpu) { 36488b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3649bdbc5dd7STejun Heo struct workqueue_struct *wq; 36508b03ae3cSTejun Heo 36518b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 36528b03ae3cSTejun Heo 3653db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3654db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3655db7bccf4STejun Heo 3656a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3657a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3658a0a1a5fdSTejun Heo 365958a69cb4STejun Heo if (cwq && wq->flags & WQ_FREEZABLE) 3660a0a1a5fdSTejun Heo cwq->max_active = 0; 36611da177e4SLinus Torvalds } 36628b03ae3cSTejun Heo 36638b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3664a0a1a5fdSTejun Heo } 3665a0a1a5fdSTejun Heo 3666a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3667a0a1a5fdSTejun Heo } 3668a0a1a5fdSTejun Heo 3669a0a1a5fdSTejun Heo /** 367058a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3671a0a1a5fdSTejun Heo * 3672a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3673a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3674a0a1a5fdSTejun Heo * 3675a0a1a5fdSTejun Heo * CONTEXT: 3676a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3677a0a1a5fdSTejun Heo * 3678a0a1a5fdSTejun Heo * RETURNS: 367958a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 368058a69cb4STejun Heo * is complete. 3681a0a1a5fdSTejun Heo */ 3682a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3683a0a1a5fdSTejun Heo { 3684a0a1a5fdSTejun Heo unsigned int cpu; 3685a0a1a5fdSTejun Heo bool busy = false; 3686a0a1a5fdSTejun Heo 3687a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3688a0a1a5fdSTejun Heo 3689a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3690a0a1a5fdSTejun Heo 3691f3421797STejun Heo for_each_gcwq_cpu(cpu) { 3692bdbc5dd7STejun Heo struct workqueue_struct *wq; 3693a0a1a5fdSTejun Heo /* 3694a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3695a0a1a5fdSTejun Heo * to peek without lock. 3696a0a1a5fdSTejun Heo */ 3697a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3698a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3699a0a1a5fdSTejun Heo 370058a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3701a0a1a5fdSTejun Heo continue; 3702a0a1a5fdSTejun Heo 3703a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3704a0a1a5fdSTejun Heo if (cwq->nr_active) { 3705a0a1a5fdSTejun Heo busy = true; 3706a0a1a5fdSTejun Heo goto out_unlock; 3707a0a1a5fdSTejun Heo } 3708a0a1a5fdSTejun Heo } 3709a0a1a5fdSTejun Heo } 3710a0a1a5fdSTejun Heo out_unlock: 3711a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3712a0a1a5fdSTejun Heo return busy; 3713a0a1a5fdSTejun Heo } 3714a0a1a5fdSTejun Heo 3715a0a1a5fdSTejun Heo /** 3716a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3717a0a1a5fdSTejun Heo * 3718a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 37197e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3720a0a1a5fdSTejun Heo * 3721a0a1a5fdSTejun Heo * CONTEXT: 37228b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3723a0a1a5fdSTejun Heo */ 3724a0a1a5fdSTejun Heo void thaw_workqueues(void) 3725a0a1a5fdSTejun Heo { 3726a0a1a5fdSTejun Heo unsigned int cpu; 3727a0a1a5fdSTejun Heo 3728a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3729a0a1a5fdSTejun Heo 3730a0a1a5fdSTejun Heo if (!workqueue_freezing) 3731a0a1a5fdSTejun Heo goto out_unlock; 3732a0a1a5fdSTejun Heo 3733f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37348b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3735bdbc5dd7STejun Heo struct workqueue_struct *wq; 37368b03ae3cSTejun Heo 37378b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 37388b03ae3cSTejun Heo 3739db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3740db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3741db7bccf4STejun Heo 3742a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3743a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3744a0a1a5fdSTejun Heo 374558a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3746a0a1a5fdSTejun Heo continue; 3747a0a1a5fdSTejun Heo 3748a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3749a0a1a5fdSTejun Heo cwq->max_active = wq->saved_max_active; 3750a0a1a5fdSTejun Heo 3751a0a1a5fdSTejun Heo while (!list_empty(&cwq->delayed_works) && 3752a0a1a5fdSTejun Heo cwq->nr_active < cwq->max_active) 3753a0a1a5fdSTejun Heo cwq_activate_first_delayed(cwq); 3754a0a1a5fdSTejun Heo } 37558b03ae3cSTejun Heo 3756e22bee78STejun Heo wake_up_worker(gcwq); 3757e22bee78STejun Heo 37588b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3759a0a1a5fdSTejun Heo } 3760a0a1a5fdSTejun Heo 3761a0a1a5fdSTejun Heo workqueue_freezing = false; 3762a0a1a5fdSTejun Heo out_unlock: 3763a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3764a0a1a5fdSTejun Heo } 3765a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3766a0a1a5fdSTejun Heo 37676ee0578bSSuresh Siddha static int __init init_workqueues(void) 37681da177e4SLinus Torvalds { 3769c34056a3STejun Heo unsigned int cpu; 3770c8e55f36STejun Heo int i; 3771c34056a3STejun Heo 3772f6500947STejun Heo cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE); 37738b03ae3cSTejun Heo 37748b03ae3cSTejun Heo /* initialize gcwqs */ 3775f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37768b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 37778b03ae3cSTejun Heo 37788b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 37797e11629dSTejun Heo INIT_LIST_HEAD(&gcwq->worklist); 37808b03ae3cSTejun Heo gcwq->cpu = cpu; 3781f3421797STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 37828b03ae3cSTejun Heo 3783c8e55f36STejun Heo INIT_LIST_HEAD(&gcwq->idle_list); 3784c8e55f36STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) 3785c8e55f36STejun Heo INIT_HLIST_HEAD(&gcwq->busy_hash[i]); 3786c8e55f36STejun Heo 3787e22bee78STejun Heo init_timer_deferrable(&gcwq->idle_timer); 3788e22bee78STejun Heo gcwq->idle_timer.function = idle_worker_timeout; 3789e22bee78STejun Heo gcwq->idle_timer.data = (unsigned long)gcwq; 3790e22bee78STejun Heo 3791e22bee78STejun Heo setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout, 3792e22bee78STejun Heo (unsigned long)gcwq); 3793e22bee78STejun Heo 37948b03ae3cSTejun Heo ida_init(&gcwq->worker_ida); 3795db7bccf4STejun Heo 3796db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3797db7bccf4STejun Heo init_waitqueue_head(&gcwq->trustee_wait); 37988b03ae3cSTejun Heo } 37998b03ae3cSTejun Heo 3800e22bee78STejun Heo /* create the initial worker */ 3801f3421797STejun Heo for_each_online_gcwq_cpu(cpu) { 3802e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3803e22bee78STejun Heo struct worker *worker; 3804e22bee78STejun Heo 3805477a3c33STejun Heo if (cpu != WORK_CPU_UNBOUND) 3806477a3c33STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 3807e22bee78STejun Heo worker = create_worker(gcwq, true); 3808e22bee78STejun Heo BUG_ON(!worker); 3809e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3810e22bee78STejun Heo start_worker(worker); 3811e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3812e22bee78STejun Heo } 3813e22bee78STejun Heo 3814d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 3815d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3816d320c038STejun Heo system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0); 3817f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3818f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 381924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 382024d51addSTejun Heo WQ_FREEZABLE, 0); 3821e5cba24eSHitoshi Mitake BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq || 382224d51addSTejun Heo !system_unbound_wq || !system_freezable_wq); 38236ee0578bSSuresh Siddha return 0; 38241da177e4SLinus Torvalds } 38256ee0578bSSuresh Siddha early_initcall(init_workqueues); 3826