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 82e22bee78STejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */ 83e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 84e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 85db7bccf4STejun Heo TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */ 861da177e4SLinus Torvalds 871da177e4SLinus Torvalds /* 88e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 89e22bee78STejun Heo * all cpus. Give -20. 90e22bee78STejun Heo */ 91e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 92c8e55f36STejun Heo }; 93c8e55f36STejun Heo 941da177e4SLinus Torvalds /* 954690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 964690c4abSTejun Heo * 97e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 98e41e704bSTejun Heo * everyone else. 994690c4abSTejun Heo * 100e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 101e22bee78STejun Heo * only be modified and accessed from the local cpu. 102e22bee78STejun Heo * 1038b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 1044690c4abSTejun Heo * 105e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 106e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 107e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 108f3421797STejun Heo * If GCWQ_DISASSOCIATED is set, it's identical to L. 109e22bee78STejun Heo * 11073f53c4aSTejun Heo * F: wq->flush_mutex protected. 11173f53c4aSTejun Heo * 1124690c4abSTejun Heo * W: workqueue_lock protected. 1134690c4abSTejun Heo */ 1144690c4abSTejun Heo 1158b03ae3cSTejun Heo struct global_cwq; 116c34056a3STejun Heo 117e22bee78STejun Heo /* 118e22bee78STejun Heo * The poor guys doing the actual heavy lifting. All on-duty workers 119e22bee78STejun Heo * are either serving the manager role, on idle list or on busy hash. 120e22bee78STejun Heo */ 121c34056a3STejun Heo struct worker { 122c8e55f36STejun Heo /* on idle list while idle, on busy hash table while busy */ 123c8e55f36STejun Heo union { 124c8e55f36STejun Heo struct list_head entry; /* L: while idle */ 125c8e55f36STejun Heo struct hlist_node hentry; /* L: while busy */ 126c8e55f36STejun Heo }; 127c8e55f36STejun Heo 128c34056a3STejun Heo struct work_struct *current_work; /* L: work being processed */ 1298cca0eeaSTejun Heo struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ 130affee4b2STejun Heo struct list_head scheduled; /* L: scheduled works */ 131c34056a3STejun Heo struct task_struct *task; /* I: worker task */ 1328b03ae3cSTejun Heo struct global_cwq *gcwq; /* I: the associated gcwq */ 133e22bee78STejun Heo /* 64 bytes boundary on 64bit, 32 on 32bit */ 134e22bee78STejun Heo unsigned long last_active; /* L: last active timestamp */ 135e22bee78STejun Heo unsigned int flags; /* X: flags */ 136c34056a3STejun Heo int id; /* I: worker id */ 137e22bee78STejun Heo struct work_struct rebind_work; /* L: rebind worker to cpu */ 138c34056a3STejun Heo }; 139c34056a3STejun Heo 1404690c4abSTejun Heo /* 141e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 142e22bee78STejun Heo * and all works are queued and processed here regardless of their 143e22bee78STejun Heo * target workqueues. 1448b03ae3cSTejun Heo */ 1458b03ae3cSTejun Heo struct global_cwq { 1468b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1477e11629dSTejun Heo struct list_head worklist; /* L: list of pending works */ 1488b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 149db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 150c8e55f36STejun Heo 151c8e55f36STejun Heo int nr_workers; /* L: total number of workers */ 152c8e55f36STejun Heo int nr_idle; /* L: currently idle ones */ 153c8e55f36STejun Heo 154c8e55f36STejun Heo /* workers are chained either in the idle_list or busy_hash */ 155e22bee78STejun Heo struct list_head idle_list; /* X: list of idle workers */ 156c8e55f36STejun Heo struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; 157c8e55f36STejun Heo /* L: hash of busy workers */ 158c8e55f36STejun Heo 159e22bee78STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 160e22bee78STejun Heo struct timer_list mayday_timer; /* L: SOS timer for dworkers */ 161e22bee78STejun Heo 1628b03ae3cSTejun Heo struct ida worker_ida; /* L: for worker IDs */ 163db7bccf4STejun Heo 164db7bccf4STejun Heo struct task_struct *trustee; /* L: for gcwq shutdown */ 165db7bccf4STejun Heo unsigned int trustee_state; /* L: trustee state */ 166db7bccf4STejun Heo wait_queue_head_t trustee_wait; /* trustee wait */ 167e22bee78STejun Heo struct worker *first_idle; /* L: first idle worker */ 1688b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1698b03ae3cSTejun Heo 1708b03ae3cSTejun Heo /* 171502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1720f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1730f900049STejun Heo * aligned at two's power of the number of flag bits. 1741da177e4SLinus Torvalds */ 1751da177e4SLinus Torvalds struct cpu_workqueue_struct { 1768b03ae3cSTejun Heo struct global_cwq *gcwq; /* I: the associated gcwq */ 1774690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 17873f53c4aSTejun Heo int work_color; /* L: current color */ 17973f53c4aSTejun Heo int flush_color; /* L: flushing color */ 18073f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 18173f53c4aSTejun Heo /* L: nr of in_flight works */ 1821e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 183a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1841e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1850f900049STejun Heo }; 1861da177e4SLinus Torvalds 1871da177e4SLinus Torvalds /* 18873f53c4aSTejun Heo * Structure used to wait for workqueue flush. 18973f53c4aSTejun Heo */ 19073f53c4aSTejun Heo struct wq_flusher { 19173f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 19273f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 19373f53c4aSTejun Heo struct completion done; /* flush completion */ 19473f53c4aSTejun Heo }; 1951da177e4SLinus Torvalds 19673f53c4aSTejun Heo /* 197f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 198f2e005aaSTejun Heo * used to determine whether there's something to be done. 199f2e005aaSTejun Heo */ 200f2e005aaSTejun Heo #ifdef CONFIG_SMP 201f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 202f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 203f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 204f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 205f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 2069c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 207f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 208f2e005aaSTejun Heo #else 209f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 210f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 211f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 212f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 213f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 214f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 215f2e005aaSTejun Heo #endif 2161da177e4SLinus Torvalds 2171da177e4SLinus Torvalds /* 2181da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2191da177e4SLinus Torvalds * per-CPU workqueues: 2201da177e4SLinus Torvalds */ 2211da177e4SLinus Torvalds struct workqueue_struct { 22297e37d7bSTejun Heo unsigned int flags; /* I: WQ_* flags */ 223bdbc5dd7STejun Heo union { 224bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 225bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 226bdbc5dd7STejun Heo unsigned long v; 227bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2284690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 22973f53c4aSTejun Heo 23073f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 23173f53c4aSTejun Heo int work_color; /* F: current work color */ 23273f53c4aSTejun Heo int flush_color; /* F: current flush color */ 23373f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 23473f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 23573f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 23673f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 23773f53c4aSTejun Heo 238f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 239e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 240e22bee78STejun Heo 241dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2424690c4abSTejun Heo const char *name; /* I: workqueue name */ 2434e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2444e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2454e6045f1SJohannes Berg #endif 2461da177e4SLinus Torvalds }; 2471da177e4SLinus Torvalds 248d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 249d320c038STejun Heo struct workqueue_struct *system_long_wq __read_mostly; 250d320c038STejun Heo struct workqueue_struct *system_nrt_wq __read_mostly; 251f3421797STejun Heo struct workqueue_struct *system_unbound_wq __read_mostly; 252d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 253d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 254d320c038STejun Heo EXPORT_SYMBOL_GPL(system_nrt_wq); 255f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 256d320c038STejun Heo 25797bd2347STejun Heo #define CREATE_TRACE_POINTS 25897bd2347STejun Heo #include <trace/events/workqueue.h> 25997bd2347STejun Heo 260db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 261db7bccf4STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ 262db7bccf4STejun Heo hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) 263db7bccf4STejun Heo 264f3421797STejun Heo static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, 265f3421797STejun Heo unsigned int sw) 266f3421797STejun Heo { 267f3421797STejun Heo if (cpu < nr_cpu_ids) { 268f3421797STejun Heo if (sw & 1) { 269f3421797STejun Heo cpu = cpumask_next(cpu, mask); 270f3421797STejun Heo if (cpu < nr_cpu_ids) 271f3421797STejun Heo return cpu; 272f3421797STejun Heo } 273f3421797STejun Heo if (sw & 2) 274f3421797STejun Heo return WORK_CPU_UNBOUND; 275f3421797STejun Heo } 276f3421797STejun Heo return WORK_CPU_NONE; 277f3421797STejun Heo } 278f3421797STejun Heo 279f3421797STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 280f3421797STejun Heo struct workqueue_struct *wq) 281f3421797STejun Heo { 282f3421797STejun Heo return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 283f3421797STejun Heo } 284f3421797STejun Heo 28509884951STejun Heo /* 28609884951STejun Heo * CPU iterators 28709884951STejun Heo * 28809884951STejun Heo * An extra gcwq is defined for an invalid cpu number 28909884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 29009884951STejun Heo * specific CPU. The following iterators are similar to 29109884951STejun Heo * for_each_*_cpu() iterators but also considers the unbound gcwq. 29209884951STejun Heo * 29309884951STejun Heo * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND 29409884951STejun Heo * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND 29509884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 29609884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 29709884951STejun Heo */ 298f3421797STejun Heo #define for_each_gcwq_cpu(cpu) \ 299f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ 300f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 301f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) 302f3421797STejun Heo 303f3421797STejun Heo #define for_each_online_gcwq_cpu(cpu) \ 304f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ 305f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 306f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) 307f3421797STejun Heo 308f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 309f3421797STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ 310f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 311f3421797STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) 312f3421797STejun Heo 313dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 314dc186ad7SThomas Gleixner 315dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 316dc186ad7SThomas Gleixner 317dc186ad7SThomas Gleixner /* 318dc186ad7SThomas Gleixner * fixup_init is called when: 319dc186ad7SThomas Gleixner * - an active object is initialized 320dc186ad7SThomas Gleixner */ 321dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 322dc186ad7SThomas Gleixner { 323dc186ad7SThomas Gleixner struct work_struct *work = addr; 324dc186ad7SThomas Gleixner 325dc186ad7SThomas Gleixner switch (state) { 326dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 327dc186ad7SThomas Gleixner cancel_work_sync(work); 328dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 329dc186ad7SThomas Gleixner return 1; 330dc186ad7SThomas Gleixner default: 331dc186ad7SThomas Gleixner return 0; 332dc186ad7SThomas Gleixner } 333dc186ad7SThomas Gleixner } 334dc186ad7SThomas Gleixner 335dc186ad7SThomas Gleixner /* 336dc186ad7SThomas Gleixner * fixup_activate is called when: 337dc186ad7SThomas Gleixner * - an active object is activated 338dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 339dc186ad7SThomas Gleixner */ 340dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 341dc186ad7SThomas Gleixner { 342dc186ad7SThomas Gleixner struct work_struct *work = addr; 343dc186ad7SThomas Gleixner 344dc186ad7SThomas Gleixner switch (state) { 345dc186ad7SThomas Gleixner 346dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 347dc186ad7SThomas Gleixner /* 348dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 349dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 350dc186ad7SThomas Gleixner * is tracked in the object tracker. 351dc186ad7SThomas Gleixner */ 35222df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 353dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 354dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 355dc186ad7SThomas Gleixner return 0; 356dc186ad7SThomas Gleixner } 357dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 358dc186ad7SThomas Gleixner return 0; 359dc186ad7SThomas Gleixner 360dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 361dc186ad7SThomas Gleixner WARN_ON(1); 362dc186ad7SThomas Gleixner 363dc186ad7SThomas Gleixner default: 364dc186ad7SThomas Gleixner return 0; 365dc186ad7SThomas Gleixner } 366dc186ad7SThomas Gleixner } 367dc186ad7SThomas Gleixner 368dc186ad7SThomas Gleixner /* 369dc186ad7SThomas Gleixner * fixup_free is called when: 370dc186ad7SThomas Gleixner * - an active object is freed 371dc186ad7SThomas Gleixner */ 372dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 373dc186ad7SThomas Gleixner { 374dc186ad7SThomas Gleixner struct work_struct *work = addr; 375dc186ad7SThomas Gleixner 376dc186ad7SThomas Gleixner switch (state) { 377dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 378dc186ad7SThomas Gleixner cancel_work_sync(work); 379dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 380dc186ad7SThomas Gleixner return 1; 381dc186ad7SThomas Gleixner default: 382dc186ad7SThomas Gleixner return 0; 383dc186ad7SThomas Gleixner } 384dc186ad7SThomas Gleixner } 385dc186ad7SThomas Gleixner 386dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 387dc186ad7SThomas Gleixner .name = "work_struct", 388dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 389dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 390dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 391dc186ad7SThomas Gleixner }; 392dc186ad7SThomas Gleixner 393dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 394dc186ad7SThomas Gleixner { 395dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 396dc186ad7SThomas Gleixner } 397dc186ad7SThomas Gleixner 398dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 399dc186ad7SThomas Gleixner { 400dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 401dc186ad7SThomas Gleixner } 402dc186ad7SThomas Gleixner 403dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 404dc186ad7SThomas Gleixner { 405dc186ad7SThomas Gleixner if (onstack) 406dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 407dc186ad7SThomas Gleixner else 408dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 409dc186ad7SThomas Gleixner } 410dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 411dc186ad7SThomas Gleixner 412dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 413dc186ad7SThomas Gleixner { 414dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 415dc186ad7SThomas Gleixner } 416dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 417dc186ad7SThomas Gleixner 418dc186ad7SThomas Gleixner #else 419dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 420dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 421dc186ad7SThomas Gleixner #endif 422dc186ad7SThomas Gleixner 42395402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 42495402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4251da177e4SLinus Torvalds static LIST_HEAD(workqueues); 426a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4271da177e4SLinus Torvalds 42814441960SOleg Nesterov /* 429e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 430e22bee78STejun Heo * which is expected to be used frequently by other cpus via 431e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 43214441960SOleg Nesterov */ 4338b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 434e22bee78STejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running); 435f756d5e2SNathan Lynch 436f3421797STejun Heo /* 437f3421797STejun Heo * Global cpu workqueue and nr_running counter for unbound gcwq. The 438f3421797STejun Heo * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its 439f3421797STejun Heo * workers have WORKER_UNBOUND set. 440f3421797STejun Heo */ 441f3421797STejun Heo static struct global_cwq unbound_global_cwq; 442f3421797STejun Heo static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */ 443f3421797STejun Heo 444c34056a3STejun Heo static int worker_thread(void *__worker); 4451da177e4SLinus Torvalds 4468b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 4471da177e4SLinus Torvalds { 448f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 4498b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 450f3421797STejun Heo else 451f3421797STejun Heo return &unbound_global_cwq; 4521da177e4SLinus Torvalds } 4531da177e4SLinus Torvalds 454e22bee78STejun Heo static atomic_t *get_gcwq_nr_running(unsigned int cpu) 455b1f4ec17SOleg Nesterov { 456f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 457e22bee78STejun Heo return &per_cpu(gcwq_nr_running, cpu); 458f3421797STejun Heo else 459f3421797STejun Heo return &unbound_gcwq_nr_running; 460b1f4ec17SOleg Nesterov } 461b1f4ec17SOleg Nesterov 4624690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 4634690c4abSTejun Heo struct workqueue_struct *wq) 464a848e3b6SOleg Nesterov { 465f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 466f3421797STejun Heo if (likely(cpu < nr_cpu_ids)) { 467f3421797STejun Heo #ifdef CONFIG_SMP 468bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 469f3421797STejun Heo #else 470f3421797STejun Heo return wq->cpu_wq.single; 471bdbc5dd7STejun Heo #endif 472a848e3b6SOleg Nesterov } 473f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 474f3421797STejun Heo return wq->cpu_wq.single; 475f3421797STejun Heo return NULL; 476f3421797STejun Heo } 477a848e3b6SOleg Nesterov 47873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 47973f53c4aSTejun Heo { 48073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 48173f53c4aSTejun Heo } 48273f53c4aSTejun Heo 48373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 48473f53c4aSTejun Heo { 48573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 48673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 48773f53c4aSTejun Heo } 48873f53c4aSTejun Heo 48973f53c4aSTejun Heo static int work_next_color(int color) 49073f53c4aSTejun Heo { 49173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 4921da177e4SLinus Torvalds } 4931da177e4SLinus Torvalds 4944594bf15SDavid Howells /* 495e120153dSTejun Heo * A work's data points to the cwq with WORK_STRUCT_CWQ set while the 496e120153dSTejun Heo * work is on queue. Once execution starts, WORK_STRUCT_CWQ is 497e120153dSTejun Heo * cleared and the work data contains the cpu number it was last on. 4987a22ad75STejun Heo * 4997a22ad75STejun Heo * set_work_{cwq|cpu}() and clear_work_data() can be used to set the 5007a22ad75STejun Heo * cwq, cpu or clear work->data. These functions should only be 5017a22ad75STejun Heo * called while the work is owned - ie. while the PENDING bit is set. 5027a22ad75STejun Heo * 5037a22ad75STejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq 5047a22ad75STejun Heo * corresponding to a work. gcwq is available once the work has been 5057a22ad75STejun Heo * queued anywhere after initialization. cwq is available only from 5067a22ad75STejun Heo * queueing until execution starts. 5074594bf15SDavid Howells */ 5087a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5097a22ad75STejun Heo unsigned long flags) 5107a22ad75STejun Heo { 5117a22ad75STejun Heo BUG_ON(!work_pending(work)); 5127a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5137a22ad75STejun Heo } 5147a22ad75STejun Heo 5157a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5164690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5174690c4abSTejun Heo unsigned long extra_flags) 518365970a1SDavid Howells { 5197a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 520e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 521365970a1SDavid Howells } 522365970a1SDavid Howells 5237a22ad75STejun Heo static void set_work_cpu(struct work_struct *work, unsigned int cpu) 5244d707b9fSOleg Nesterov { 5257a22ad75STejun Heo set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING); 5264d707b9fSOleg Nesterov } 5274d707b9fSOleg Nesterov 5287a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 529365970a1SDavid Howells { 5307a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 5317a22ad75STejun Heo } 5327a22ad75STejun Heo 5337a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5347a22ad75STejun Heo { 535e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5367a22ad75STejun Heo 537e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 538e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 539e120153dSTejun Heo else 540e120153dSTejun Heo return NULL; 5417a22ad75STejun Heo } 5427a22ad75STejun Heo 5437a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 5447a22ad75STejun Heo { 545e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5467a22ad75STejun Heo unsigned int cpu; 5477a22ad75STejun Heo 548e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 549e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 550e120153dSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq; 5517a22ad75STejun Heo 5527a22ad75STejun Heo cpu = data >> WORK_STRUCT_FLAG_BITS; 553bdbc5dd7STejun Heo if (cpu == WORK_CPU_NONE) 5547a22ad75STejun Heo return NULL; 5557a22ad75STejun Heo 556f3421797STejun Heo BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); 5577a22ad75STejun Heo return get_gcwq(cpu); 558365970a1SDavid Howells } 559365970a1SDavid Howells 560e22bee78STejun Heo /* 561e22bee78STejun Heo * Policy functions. These define the policies on how the global 562e22bee78STejun Heo * worker pool is managed. Unless noted otherwise, these functions 563e22bee78STejun Heo * assume that they're being called with gcwq->lock held. 564e22bee78STejun Heo */ 565e22bee78STejun Heo 566649027d7STejun Heo static bool __need_more_worker(struct global_cwq *gcwq) 567649027d7STejun Heo { 568649027d7STejun Heo return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) || 569649027d7STejun Heo gcwq->flags & GCWQ_HIGHPRI_PENDING; 570649027d7STejun Heo } 571649027d7STejun Heo 572e22bee78STejun Heo /* 573e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 574e22bee78STejun Heo * running workers. 575e22bee78STejun Heo */ 576e22bee78STejun Heo static bool need_more_worker(struct global_cwq *gcwq) 577e22bee78STejun Heo { 578649027d7STejun Heo return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq); 579e22bee78STejun Heo } 580e22bee78STejun Heo 581e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 582e22bee78STejun Heo static bool may_start_working(struct global_cwq *gcwq) 583e22bee78STejun Heo { 584e22bee78STejun Heo return gcwq->nr_idle; 585e22bee78STejun Heo } 586e22bee78STejun Heo 587e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 588e22bee78STejun Heo static bool keep_working(struct global_cwq *gcwq) 589e22bee78STejun Heo { 590e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 591e22bee78STejun Heo 59230310045STejun Heo return !list_empty(&gcwq->worklist) && 59330310045STejun Heo (atomic_read(nr_running) <= 1 || 59430310045STejun Heo gcwq->flags & GCWQ_HIGHPRI_PENDING); 595e22bee78STejun Heo } 596e22bee78STejun Heo 597e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 598e22bee78STejun Heo static bool need_to_create_worker(struct global_cwq *gcwq) 599e22bee78STejun Heo { 600e22bee78STejun Heo return need_more_worker(gcwq) && !may_start_working(gcwq); 601e22bee78STejun Heo } 602e22bee78STejun Heo 603e22bee78STejun Heo /* Do I need to be the manager? */ 604e22bee78STejun Heo static bool need_to_manage_workers(struct global_cwq *gcwq) 605e22bee78STejun Heo { 606e22bee78STejun Heo return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS; 607e22bee78STejun Heo } 608e22bee78STejun Heo 609e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 610e22bee78STejun Heo static bool too_many_workers(struct global_cwq *gcwq) 611e22bee78STejun Heo { 612e22bee78STejun Heo bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS; 613e22bee78STejun Heo int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */ 614e22bee78STejun Heo int nr_busy = gcwq->nr_workers - nr_idle; 615e22bee78STejun Heo 616e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 617e22bee78STejun Heo } 618e22bee78STejun Heo 619e22bee78STejun Heo /* 620e22bee78STejun Heo * Wake up functions. 621e22bee78STejun Heo */ 622e22bee78STejun Heo 6237e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 6247e11629dSTejun Heo static struct worker *first_worker(struct global_cwq *gcwq) 6257e11629dSTejun Heo { 6267e11629dSTejun Heo if (unlikely(list_empty(&gcwq->idle_list))) 6277e11629dSTejun Heo return NULL; 6287e11629dSTejun Heo 6297e11629dSTejun Heo return list_first_entry(&gcwq->idle_list, struct worker, entry); 6307e11629dSTejun Heo } 6317e11629dSTejun Heo 6327e11629dSTejun Heo /** 6337e11629dSTejun Heo * wake_up_worker - wake up an idle worker 6347e11629dSTejun Heo * @gcwq: gcwq to wake worker for 6357e11629dSTejun Heo * 6367e11629dSTejun Heo * Wake up the first idle worker of @gcwq. 6377e11629dSTejun Heo * 6387e11629dSTejun Heo * CONTEXT: 6397e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 6407e11629dSTejun Heo */ 6417e11629dSTejun Heo static void wake_up_worker(struct global_cwq *gcwq) 6427e11629dSTejun Heo { 6437e11629dSTejun Heo struct worker *worker = first_worker(gcwq); 6447e11629dSTejun Heo 6457e11629dSTejun Heo if (likely(worker)) 6467e11629dSTejun Heo wake_up_process(worker->task); 6477e11629dSTejun Heo } 6487e11629dSTejun Heo 6494690c4abSTejun Heo /** 650e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 651e22bee78STejun Heo * @task: task waking up 652e22bee78STejun Heo * @cpu: CPU @task is waking up to 653e22bee78STejun Heo * 654e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 655e22bee78STejun Heo * being awoken. 656e22bee78STejun Heo * 657e22bee78STejun Heo * CONTEXT: 658e22bee78STejun Heo * spin_lock_irq(rq->lock) 659e22bee78STejun Heo */ 660e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 661e22bee78STejun Heo { 662e22bee78STejun Heo struct worker *worker = kthread_data(task); 663e22bee78STejun Heo 6642d64672eSSteven Rostedt if (!(worker->flags & WORKER_NOT_RUNNING)) 665e22bee78STejun Heo atomic_inc(get_gcwq_nr_running(cpu)); 666e22bee78STejun Heo } 667e22bee78STejun Heo 668e22bee78STejun Heo /** 669e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 670e22bee78STejun Heo * @task: task going to sleep 671e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 672e22bee78STejun Heo * 673e22bee78STejun Heo * This function is called during schedule() when a busy worker is 674e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 675e22bee78STejun Heo * returning pointer to its task. 676e22bee78STejun Heo * 677e22bee78STejun Heo * CONTEXT: 678e22bee78STejun Heo * spin_lock_irq(rq->lock) 679e22bee78STejun Heo * 680e22bee78STejun Heo * RETURNS: 681e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 682e22bee78STejun Heo */ 683e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 684e22bee78STejun Heo unsigned int cpu) 685e22bee78STejun Heo { 686e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 687e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 688e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(cpu); 689e22bee78STejun Heo 6902d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 691e22bee78STejun Heo return NULL; 692e22bee78STejun Heo 693e22bee78STejun Heo /* this can only happen on the local cpu */ 694e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 695e22bee78STejun Heo 696e22bee78STejun Heo /* 697e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 698e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 699e22bee78STejun Heo * Please read comment there. 700e22bee78STejun Heo * 701e22bee78STejun Heo * NOT_RUNNING is clear. This means that trustee is not in 702e22bee78STejun Heo * charge and we're running on the local cpu w/ rq lock held 703e22bee78STejun Heo * and preemption disabled, which in turn means that none else 704e22bee78STejun Heo * could be manipulating idle_list, so dereferencing idle_list 705e22bee78STejun Heo * without gcwq lock is safe. 706e22bee78STejun Heo */ 707e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist)) 708e22bee78STejun Heo to_wakeup = first_worker(gcwq); 709e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 710e22bee78STejun Heo } 711e22bee78STejun Heo 712e22bee78STejun Heo /** 713e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 714cb444766STejun Heo * @worker: self 715d302f017STejun Heo * @flags: flags to set 716d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 717d302f017STejun Heo * 718e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 719e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 720e22bee78STejun Heo * woken up. 721d302f017STejun Heo * 722cb444766STejun Heo * CONTEXT: 723cb444766STejun Heo * spin_lock_irq(gcwq->lock) 724d302f017STejun Heo */ 725d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 726d302f017STejun Heo bool wakeup) 727d302f017STejun Heo { 728e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 729e22bee78STejun Heo 730cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 731cb444766STejun Heo 732e22bee78STejun Heo /* 733e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 734e22bee78STejun Heo * wake up an idle worker as necessary if requested by 735e22bee78STejun Heo * @wakeup. 736e22bee78STejun Heo */ 737e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 738e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 739e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 740e22bee78STejun Heo 741e22bee78STejun Heo if (wakeup) { 742e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 743e22bee78STejun Heo !list_empty(&gcwq->worklist)) 744e22bee78STejun Heo wake_up_worker(gcwq); 745e22bee78STejun Heo } else 746e22bee78STejun Heo atomic_dec(nr_running); 747e22bee78STejun Heo } 748e22bee78STejun Heo 749d302f017STejun Heo worker->flags |= flags; 750d302f017STejun Heo } 751d302f017STejun Heo 752d302f017STejun Heo /** 753e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 754cb444766STejun Heo * @worker: self 755d302f017STejun Heo * @flags: flags to clear 756d302f017STejun Heo * 757e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 758d302f017STejun Heo * 759cb444766STejun Heo * CONTEXT: 760cb444766STejun Heo * spin_lock_irq(gcwq->lock) 761d302f017STejun Heo */ 762d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 763d302f017STejun Heo { 764e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 765e22bee78STejun Heo unsigned int oflags = worker->flags; 766e22bee78STejun Heo 767cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 768cb444766STejun Heo 769d302f017STejun Heo worker->flags &= ~flags; 770e22bee78STejun Heo 771e22bee78STejun Heo /* if transitioning out of NOT_RUNNING, increment nr_running */ 772e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 773e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 774e22bee78STejun Heo atomic_inc(get_gcwq_nr_running(gcwq->cpu)); 775d302f017STejun Heo } 776d302f017STejun Heo 777d302f017STejun Heo /** 778c8e55f36STejun Heo * busy_worker_head - return the busy hash head for a work 779c8e55f36STejun Heo * @gcwq: gcwq of interest 780c8e55f36STejun Heo * @work: work to be hashed 781c8e55f36STejun Heo * 782c8e55f36STejun Heo * Return hash head of @gcwq for @work. 783c8e55f36STejun Heo * 784c8e55f36STejun Heo * CONTEXT: 785c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 786c8e55f36STejun Heo * 787c8e55f36STejun Heo * RETURNS: 788c8e55f36STejun Heo * Pointer to the hash head. 789c8e55f36STejun Heo */ 790c8e55f36STejun Heo static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, 791c8e55f36STejun Heo struct work_struct *work) 792c8e55f36STejun Heo { 793c8e55f36STejun Heo const int base_shift = ilog2(sizeof(struct work_struct)); 794c8e55f36STejun Heo unsigned long v = (unsigned long)work; 795c8e55f36STejun Heo 796c8e55f36STejun Heo /* simple shift and fold hash, do we need something better? */ 797c8e55f36STejun Heo v >>= base_shift; 798c8e55f36STejun Heo v += v >> BUSY_WORKER_HASH_ORDER; 799c8e55f36STejun Heo v &= BUSY_WORKER_HASH_MASK; 800c8e55f36STejun Heo 801c8e55f36STejun Heo return &gcwq->busy_hash[v]; 802c8e55f36STejun Heo } 803c8e55f36STejun Heo 804c8e55f36STejun Heo /** 8058cca0eeaSTejun Heo * __find_worker_executing_work - find worker which is executing a work 8068cca0eeaSTejun Heo * @gcwq: gcwq of interest 8078cca0eeaSTejun Heo * @bwh: hash head as returned by busy_worker_head() 8088cca0eeaSTejun Heo * @work: work to find worker for 8098cca0eeaSTejun Heo * 8108cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. @bwh should be 8118cca0eeaSTejun Heo * the hash head obtained by calling busy_worker_head() with the same 8128cca0eeaSTejun Heo * work. 8138cca0eeaSTejun Heo * 8148cca0eeaSTejun Heo * CONTEXT: 8158cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8168cca0eeaSTejun Heo * 8178cca0eeaSTejun Heo * RETURNS: 8188cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8198cca0eeaSTejun Heo * otherwise. 8208cca0eeaSTejun Heo */ 8218cca0eeaSTejun Heo static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, 8228cca0eeaSTejun Heo struct hlist_head *bwh, 8238cca0eeaSTejun Heo struct work_struct *work) 8248cca0eeaSTejun Heo { 8258cca0eeaSTejun Heo struct worker *worker; 8268cca0eeaSTejun Heo struct hlist_node *tmp; 8278cca0eeaSTejun Heo 8288cca0eeaSTejun Heo hlist_for_each_entry(worker, tmp, bwh, hentry) 8298cca0eeaSTejun Heo if (worker->current_work == work) 8308cca0eeaSTejun Heo return worker; 8318cca0eeaSTejun Heo return NULL; 8328cca0eeaSTejun Heo } 8338cca0eeaSTejun Heo 8348cca0eeaSTejun Heo /** 8358cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 8368cca0eeaSTejun Heo * @gcwq: gcwq of interest 8378cca0eeaSTejun Heo * @work: work to find worker for 8388cca0eeaSTejun Heo * 8398cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. This function is 8408cca0eeaSTejun Heo * identical to __find_worker_executing_work() except that this 8418cca0eeaSTejun Heo * function calculates @bwh itself. 8428cca0eeaSTejun Heo * 8438cca0eeaSTejun Heo * CONTEXT: 8448cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8458cca0eeaSTejun Heo * 8468cca0eeaSTejun Heo * RETURNS: 8478cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8488cca0eeaSTejun Heo * otherwise. 8498cca0eeaSTejun Heo */ 8508cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 8518cca0eeaSTejun Heo struct work_struct *work) 8528cca0eeaSTejun Heo { 8538cca0eeaSTejun Heo return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), 8548cca0eeaSTejun Heo work); 8558cca0eeaSTejun Heo } 8568cca0eeaSTejun Heo 8578cca0eeaSTejun Heo /** 858649027d7STejun Heo * gcwq_determine_ins_pos - find insertion position 859649027d7STejun Heo * @gcwq: gcwq of interest 860649027d7STejun Heo * @cwq: cwq a work is being queued for 861649027d7STejun Heo * 862649027d7STejun Heo * A work for @cwq is about to be queued on @gcwq, determine insertion 863649027d7STejun Heo * position for the work. If @cwq is for HIGHPRI wq, the work is 864649027d7STejun Heo * queued at the head of the queue but in FIFO order with respect to 865649027d7STejun Heo * other HIGHPRI works; otherwise, at the end of the queue. This 866649027d7STejun Heo * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that 867649027d7STejun Heo * there are HIGHPRI works pending. 868649027d7STejun Heo * 869649027d7STejun Heo * CONTEXT: 870649027d7STejun Heo * spin_lock_irq(gcwq->lock). 871649027d7STejun Heo * 872649027d7STejun Heo * RETURNS: 873649027d7STejun Heo * Pointer to inserstion position. 874649027d7STejun Heo */ 875649027d7STejun Heo static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq, 8761da177e4SLinus Torvalds struct cpu_workqueue_struct *cwq) 8771da177e4SLinus Torvalds { 878649027d7STejun Heo struct work_struct *twork; 8791da177e4SLinus Torvalds 880649027d7STejun Heo if (likely(!(cwq->wq->flags & WQ_HIGHPRI))) 881649027d7STejun Heo return &gcwq->worklist; 8821da177e4SLinus Torvalds 883649027d7STejun Heo list_for_each_entry(twork, &gcwq->worklist, entry) { 884649027d7STejun Heo struct cpu_workqueue_struct *tcwq = get_work_cwq(twork); 885649027d7STejun Heo 886649027d7STejun Heo if (!(tcwq->wq->flags & WQ_HIGHPRI)) 887649027d7STejun Heo break; 8881da177e4SLinus Torvalds } 8891da177e4SLinus Torvalds 890649027d7STejun Heo gcwq->flags |= GCWQ_HIGHPRI_PENDING; 891649027d7STejun Heo return &twork->entry; 892649027d7STejun Heo } 893649027d7STejun Heo 894649027d7STejun Heo /** 8957e11629dSTejun Heo * insert_work - insert a work into gcwq 8964690c4abSTejun Heo * @cwq: cwq @work belongs to 8974690c4abSTejun Heo * @work: work to insert 8984690c4abSTejun Heo * @head: insertion point 8994690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 9004690c4abSTejun Heo * 9017e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 9027e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 9034690c4abSTejun Heo * 9044690c4abSTejun Heo * CONTEXT: 9058b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 9061da177e4SLinus Torvalds */ 907b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 9084690c4abSTejun Heo struct work_struct *work, struct list_head *head, 9094690c4abSTejun Heo unsigned int extra_flags) 910b89deed3SOleg Nesterov { 911e22bee78STejun Heo struct global_cwq *gcwq = cwq->gcwq; 912e1d8aa9fSFrederic Weisbecker 9134690c4abSTejun Heo /* we own @work, set data and link */ 9147a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 9154690c4abSTejun Heo 9166e84d644SOleg Nesterov /* 9176e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 9186e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 9196e84d644SOleg Nesterov */ 9206e84d644SOleg Nesterov smp_wmb(); 9214690c4abSTejun Heo 9221a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 923e22bee78STejun Heo 924e22bee78STejun Heo /* 925e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 926e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 927e22bee78STejun Heo * lying around lazily while there are works to be processed. 928e22bee78STejun Heo */ 929e22bee78STejun Heo smp_mb(); 930e22bee78STejun Heo 931649027d7STejun Heo if (__need_more_worker(gcwq)) 932e22bee78STejun Heo wake_up_worker(gcwq); 933b89deed3SOleg Nesterov } 934b89deed3SOleg Nesterov 935c8efcc25STejun Heo /* 936c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 937c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 938c8efcc25STejun Heo * cold paths. 939c8efcc25STejun Heo */ 940c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 941c8efcc25STejun Heo { 942c8efcc25STejun Heo unsigned long flags; 943c8efcc25STejun Heo unsigned int cpu; 944c8efcc25STejun Heo 945c8efcc25STejun Heo for_each_gcwq_cpu(cpu) { 946c8efcc25STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 947c8efcc25STejun Heo struct worker *worker; 948c8efcc25STejun Heo struct hlist_node *pos; 949c8efcc25STejun Heo int i; 950c8efcc25STejun Heo 951c8efcc25STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 952c8efcc25STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 953c8efcc25STejun Heo if (worker->task != current) 954c8efcc25STejun Heo continue; 955c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 956c8efcc25STejun Heo /* 957c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 958c8efcc25STejun Heo * is headed to the same workqueue. 959c8efcc25STejun Heo */ 960c8efcc25STejun Heo return worker->current_cwq->wq == wq; 961c8efcc25STejun Heo } 962c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 963c8efcc25STejun Heo } 964c8efcc25STejun Heo return false; 965c8efcc25STejun Heo } 966c8efcc25STejun Heo 9674690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 9681da177e4SLinus Torvalds struct work_struct *work) 9691da177e4SLinus Torvalds { 970502ca9d8STejun Heo struct global_cwq *gcwq; 971502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 9721e19ffc6STejun Heo struct list_head *worklist; 9738a2e8e5dSTejun Heo unsigned int work_flags; 9741da177e4SLinus Torvalds unsigned long flags; 9751da177e4SLinus Torvalds 976dc186ad7SThomas Gleixner debug_work_activate(work); 9771e19ffc6STejun Heo 978c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 979c8efcc25STejun Heo if (unlikely(wq->flags & WQ_DYING) && 980c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 981e41e704bSTejun Heo return; 982e41e704bSTejun Heo 983c7fc77f7STejun Heo /* determine gcwq to use */ 984c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 985c7fc77f7STejun Heo struct global_cwq *last_gcwq; 986c7fc77f7STejun Heo 987f3421797STejun Heo if (unlikely(cpu == WORK_CPU_UNBOUND)) 988f3421797STejun Heo cpu = raw_smp_processor_id(); 989f3421797STejun Heo 99018aa9effSTejun Heo /* 99118aa9effSTejun Heo * It's multi cpu. If @wq is non-reentrant and @work 99218aa9effSTejun Heo * was previously on a different cpu, it might still 99318aa9effSTejun Heo * be running there, in which case the work needs to 99418aa9effSTejun Heo * be queued on that cpu to guarantee non-reentrance. 99518aa9effSTejun Heo */ 996502ca9d8STejun Heo gcwq = get_gcwq(cpu); 99718aa9effSTejun Heo if (wq->flags & WQ_NON_REENTRANT && 99818aa9effSTejun Heo (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) { 99918aa9effSTejun Heo struct worker *worker; 100018aa9effSTejun Heo 100118aa9effSTejun Heo spin_lock_irqsave(&last_gcwq->lock, flags); 100218aa9effSTejun Heo 100318aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 100418aa9effSTejun Heo 100518aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 100618aa9effSTejun Heo gcwq = last_gcwq; 100718aa9effSTejun Heo else { 100818aa9effSTejun Heo /* meh... not running there, queue here */ 100918aa9effSTejun Heo spin_unlock_irqrestore(&last_gcwq->lock, flags); 101018aa9effSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 101118aa9effSTejun Heo } 101218aa9effSTejun Heo } else 10138b03ae3cSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1014f3421797STejun Heo } else { 1015f3421797STejun Heo gcwq = get_gcwq(WORK_CPU_UNBOUND); 1016f3421797STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1017502ca9d8STejun Heo } 1018502ca9d8STejun Heo 1019502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 1020502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 1021cdadf009STejun Heo trace_workqueue_queue_work(cpu, cwq, work); 1022502ca9d8STejun Heo 10234690c4abSTejun Heo BUG_ON(!list_empty(&work->entry)); 10241e19ffc6STejun Heo 102573f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 10268a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 10271e19ffc6STejun Heo 10281e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1029cdadf009STejun Heo trace_workqueue_activate_work(work); 10301e19ffc6STejun Heo cwq->nr_active++; 1031649027d7STejun Heo worklist = gcwq_determine_ins_pos(gcwq, cwq); 10328a2e8e5dSTejun Heo } else { 10338a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 10341e19ffc6STejun Heo worklist = &cwq->delayed_works; 10358a2e8e5dSTejun Heo } 10361e19ffc6STejun Heo 10378a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 10381e19ffc6STejun Heo 10398b03ae3cSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 10401da177e4SLinus Torvalds } 10411da177e4SLinus Torvalds 10420fcb78c2SRolf Eike Beer /** 10430fcb78c2SRolf Eike Beer * queue_work - queue work on a workqueue 10440fcb78c2SRolf Eike Beer * @wq: workqueue to use 10450fcb78c2SRolf Eike Beer * @work: work to queue 10460fcb78c2SRolf Eike Beer * 1047057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 10481da177e4SLinus Torvalds * 104900dfcaf7SOleg Nesterov * We queue the work to the CPU on which it was submitted, but if the CPU dies 105000dfcaf7SOleg Nesterov * it can be processed by another CPU. 10511da177e4SLinus Torvalds */ 10527ad5b3a5SHarvey Harrison int queue_work(struct workqueue_struct *wq, struct work_struct *work) 10531da177e4SLinus Torvalds { 1054ef1ca236SOleg Nesterov int ret; 10551da177e4SLinus Torvalds 1056ef1ca236SOleg Nesterov ret = queue_work_on(get_cpu(), wq, work); 1057a848e3b6SOleg Nesterov put_cpu(); 1058ef1ca236SOleg Nesterov 10591da177e4SLinus Torvalds return ret; 10601da177e4SLinus Torvalds } 1061ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_work); 10621da177e4SLinus Torvalds 1063c1a220e7SZhang Rui /** 1064c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1065c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1066c1a220e7SZhang Rui * @wq: workqueue to use 1067c1a220e7SZhang Rui * @work: work to queue 1068c1a220e7SZhang Rui * 1069c1a220e7SZhang Rui * Returns 0 if @work was already on a queue, non-zero otherwise. 1070c1a220e7SZhang Rui * 1071c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1072c1a220e7SZhang Rui * can't go away. 1073c1a220e7SZhang Rui */ 1074c1a220e7SZhang Rui int 1075c1a220e7SZhang Rui queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work) 1076c1a220e7SZhang Rui { 1077c1a220e7SZhang Rui int ret = 0; 1078c1a220e7SZhang Rui 107922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 10804690c4abSTejun Heo __queue_work(cpu, wq, work); 1081c1a220e7SZhang Rui ret = 1; 1082c1a220e7SZhang Rui } 1083c1a220e7SZhang Rui return ret; 1084c1a220e7SZhang Rui } 1085c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1086c1a220e7SZhang Rui 10876d141c3fSLi Zefan static void delayed_work_timer_fn(unsigned long __data) 10881da177e4SLinus Torvalds { 108952bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 10907a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 10911da177e4SLinus Torvalds 10924690c4abSTejun Heo __queue_work(smp_processor_id(), cwq->wq, &dwork->work); 10931da177e4SLinus Torvalds } 10941da177e4SLinus Torvalds 10950fcb78c2SRolf Eike Beer /** 10960fcb78c2SRolf Eike Beer * queue_delayed_work - queue work on a workqueue after delay 10970fcb78c2SRolf Eike Beer * @wq: workqueue to use 1098af9997e4SRandy Dunlap * @dwork: delayable work to queue 10990fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 11000fcb78c2SRolf Eike Beer * 1101057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 11020fcb78c2SRolf Eike Beer */ 11037ad5b3a5SHarvey Harrison int queue_delayed_work(struct workqueue_struct *wq, 110452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 11051da177e4SLinus Torvalds { 110652bad64dSDavid Howells if (delay == 0) 110763bc0362SOleg Nesterov return queue_work(wq, &dwork->work); 11081da177e4SLinus Torvalds 110963bc0362SOleg Nesterov return queue_delayed_work_on(-1, wq, dwork, delay); 11101da177e4SLinus Torvalds } 1111ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work); 11121da177e4SLinus Torvalds 11130fcb78c2SRolf Eike Beer /** 11140fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 11150fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 11160fcb78c2SRolf Eike Beer * @wq: workqueue to use 1117af9997e4SRandy Dunlap * @dwork: work to queue 11180fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 11190fcb78c2SRolf Eike Beer * 1120057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 11210fcb78c2SRolf Eike Beer */ 11227a6bc1cdSVenkatesh Pallipadi int queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 112352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 11247a6bc1cdSVenkatesh Pallipadi { 11257a6bc1cdSVenkatesh Pallipadi int ret = 0; 112652bad64dSDavid Howells struct timer_list *timer = &dwork->timer; 112752bad64dSDavid Howells struct work_struct *work = &dwork->work; 11287a6bc1cdSVenkatesh Pallipadi 112922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1130c7fc77f7STejun Heo unsigned int lcpu; 11317a22ad75STejun Heo 11327a6bc1cdSVenkatesh Pallipadi BUG_ON(timer_pending(timer)); 11337a6bc1cdSVenkatesh Pallipadi BUG_ON(!list_empty(&work->entry)); 11347a6bc1cdSVenkatesh Pallipadi 11358a3e77ccSAndrew Liu timer_stats_timer_set_start_info(&dwork->timer); 11368a3e77ccSAndrew Liu 11377a22ad75STejun Heo /* 11387a22ad75STejun Heo * This stores cwq for the moment, for the timer_fn. 11397a22ad75STejun Heo * Note that the work's gcwq is preserved to allow 11407a22ad75STejun Heo * reentrance detection for delayed works. 11417a22ad75STejun Heo */ 1142c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1143c7fc77f7STejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 1144c7fc77f7STejun Heo 1145c7fc77f7STejun Heo if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND) 1146c7fc77f7STejun Heo lcpu = gcwq->cpu; 1147c7fc77f7STejun Heo else 1148c7fc77f7STejun Heo lcpu = raw_smp_processor_id(); 1149c7fc77f7STejun Heo } else 1150c7fc77f7STejun Heo lcpu = WORK_CPU_UNBOUND; 1151c7fc77f7STejun Heo 11527a22ad75STejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 1153c7fc77f7STejun Heo 11547a6bc1cdSVenkatesh Pallipadi timer->expires = jiffies + delay; 115552bad64dSDavid Howells timer->data = (unsigned long)dwork; 11567a6bc1cdSVenkatesh Pallipadi timer->function = delayed_work_timer_fn; 115763bc0362SOleg Nesterov 115863bc0362SOleg Nesterov if (unlikely(cpu >= 0)) 11597a6bc1cdSVenkatesh Pallipadi add_timer_on(timer, cpu); 116063bc0362SOleg Nesterov else 116163bc0362SOleg Nesterov add_timer(timer); 11627a6bc1cdSVenkatesh Pallipadi ret = 1; 11637a6bc1cdSVenkatesh Pallipadi } 11647a6bc1cdSVenkatesh Pallipadi return ret; 11657a6bc1cdSVenkatesh Pallipadi } 1166ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 11671da177e4SLinus Torvalds 1168c8e55f36STejun Heo /** 1169c8e55f36STejun Heo * worker_enter_idle - enter idle state 1170c8e55f36STejun Heo * @worker: worker which is entering idle state 1171c8e55f36STejun Heo * 1172c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1173c8e55f36STejun Heo * necessary. 1174c8e55f36STejun Heo * 1175c8e55f36STejun Heo * LOCKING: 1176c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1177c8e55f36STejun Heo */ 1178c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 11791da177e4SLinus Torvalds { 1180c8e55f36STejun Heo struct global_cwq *gcwq = worker->gcwq; 1181c8e55f36STejun Heo 1182c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1183c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1184c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1185c8e55f36STejun Heo 1186cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1187cb444766STejun Heo worker->flags |= WORKER_IDLE; 1188c8e55f36STejun Heo gcwq->nr_idle++; 1189e22bee78STejun Heo worker->last_active = jiffies; 1190c8e55f36STejun Heo 1191c8e55f36STejun Heo /* idle_list is LIFO */ 1192c8e55f36STejun Heo list_add(&worker->entry, &gcwq->idle_list); 1193db7bccf4STejun Heo 1194e22bee78STejun Heo if (likely(!(worker->flags & WORKER_ROGUE))) { 1195e22bee78STejun Heo if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer)) 1196e22bee78STejun Heo mod_timer(&gcwq->idle_timer, 1197e22bee78STejun Heo jiffies + IDLE_WORKER_TIMEOUT); 1198e22bee78STejun Heo } else 1199db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 1200cb444766STejun Heo 1201cb444766STejun Heo /* sanity check nr_running */ 1202cb444766STejun Heo WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle && 1203cb444766STejun Heo atomic_read(get_gcwq_nr_running(gcwq->cpu))); 1204c8e55f36STejun Heo } 1205c8e55f36STejun Heo 1206c8e55f36STejun Heo /** 1207c8e55f36STejun Heo * worker_leave_idle - leave idle state 1208c8e55f36STejun Heo * @worker: worker which is leaving idle state 1209c8e55f36STejun Heo * 1210c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1211c8e55f36STejun Heo * 1212c8e55f36STejun Heo * LOCKING: 1213c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1214c8e55f36STejun Heo */ 1215c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1216c8e55f36STejun Heo { 1217c8e55f36STejun Heo struct global_cwq *gcwq = worker->gcwq; 1218c8e55f36STejun Heo 1219c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1220d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1221c8e55f36STejun Heo gcwq->nr_idle--; 1222c8e55f36STejun Heo list_del_init(&worker->entry); 1223c8e55f36STejun Heo } 1224c8e55f36STejun Heo 1225e22bee78STejun Heo /** 1226e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1227e22bee78STejun Heo * @worker: self 1228e22bee78STejun Heo * 1229e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1230e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1231e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1232e22bee78STejun Heo * guaranteed to execute on the cpu. 1233e22bee78STejun Heo * 1234e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1235e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1236e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1237e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1238e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1239e22bee78STejun Heo * [dis]associated in the meantime. 1240e22bee78STejun Heo * 1241e22bee78STejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies 1242e22bee78STejun Heo * the binding against GCWQ_DISASSOCIATED which is set during 1243e22bee78STejun Heo * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters 1244e22bee78STejun Heo * idle state or fetches works without dropping lock, it can guarantee 1245e22bee78STejun Heo * the scheduling requirement described in the first paragraph. 1246e22bee78STejun Heo * 1247e22bee78STejun Heo * CONTEXT: 1248e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1249e22bee78STejun Heo * held. 1250e22bee78STejun Heo * 1251e22bee78STejun Heo * RETURNS: 1252e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1253e22bee78STejun Heo * bound), %false if offline. 1254e22bee78STejun Heo */ 1255e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1256972fa1c5SNamhyung Kim __acquires(&gcwq->lock) 1257e22bee78STejun Heo { 1258e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1259e22bee78STejun Heo struct task_struct *task = worker->task; 1260e22bee78STejun Heo 1261e22bee78STejun Heo while (true) { 1262e22bee78STejun Heo /* 1263e22bee78STejun Heo * The following call may fail, succeed or succeed 1264e22bee78STejun Heo * without actually migrating the task to the cpu if 1265e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1266e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1267e22bee78STejun Heo */ 1268f3421797STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) 1269e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1270e22bee78STejun Heo 1271e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1272e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1273e22bee78STejun Heo return false; 1274e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1275e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1276e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1277e22bee78STejun Heo return true; 1278e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1279e22bee78STejun Heo 1280e22bee78STejun Heo /* CPU has come up inbetween, retry migration */ 1281e22bee78STejun Heo cpu_relax(); 1282e22bee78STejun Heo } 1283e22bee78STejun Heo } 1284e22bee78STejun Heo 1285e22bee78STejun Heo /* 1286e22bee78STejun Heo * Function for worker->rebind_work used to rebind rogue busy workers 1287e22bee78STejun Heo * to the associated cpu which is coming back online. This is 1288e22bee78STejun Heo * scheduled by cpu up but can race with other cpu hotplug operations 1289e22bee78STejun Heo * and may be executed twice without intervening cpu down. 1290e22bee78STejun Heo */ 1291e22bee78STejun Heo static void worker_rebind_fn(struct work_struct *work) 1292e22bee78STejun Heo { 1293e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1294e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1295e22bee78STejun Heo 1296e22bee78STejun Heo if (worker_maybe_bind_and_lock(worker)) 1297e22bee78STejun Heo worker_clr_flags(worker, WORKER_REBIND); 1298e22bee78STejun Heo 1299e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1300e22bee78STejun Heo } 1301e22bee78STejun Heo 1302c34056a3STejun Heo static struct worker *alloc_worker(void) 1303c34056a3STejun Heo { 1304c34056a3STejun Heo struct worker *worker; 1305c34056a3STejun Heo 1306c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1307c8e55f36STejun Heo if (worker) { 1308c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1309affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1310e22bee78STejun Heo INIT_WORK(&worker->rebind_work, worker_rebind_fn); 1311e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1312e22bee78STejun Heo worker->flags = WORKER_PREP; 1313c8e55f36STejun Heo } 1314c34056a3STejun Heo return worker; 1315c34056a3STejun Heo } 1316c34056a3STejun Heo 1317c34056a3STejun Heo /** 1318c34056a3STejun Heo * create_worker - create a new workqueue worker 13197e11629dSTejun Heo * @gcwq: gcwq the new worker will belong to 1320c34056a3STejun Heo * @bind: whether to set affinity to @cpu or not 1321c34056a3STejun Heo * 13227e11629dSTejun Heo * Create a new worker which is bound to @gcwq. The returned worker 1323c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1324c34056a3STejun Heo * destroy_worker(). 1325c34056a3STejun Heo * 1326c34056a3STejun Heo * CONTEXT: 1327c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1328c34056a3STejun Heo * 1329c34056a3STejun Heo * RETURNS: 1330c34056a3STejun Heo * Pointer to the newly created worker. 1331c34056a3STejun Heo */ 13327e11629dSTejun Heo static struct worker *create_worker(struct global_cwq *gcwq, bool bind) 1333c34056a3STejun Heo { 1334f3421797STejun Heo bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND; 1335c34056a3STejun Heo struct worker *worker = NULL; 1336f3421797STejun Heo int id = -1; 1337c34056a3STejun Heo 13388b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 13398b03ae3cSTejun Heo while (ida_get_new(&gcwq->worker_ida, &id)) { 13408b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 13418b03ae3cSTejun Heo if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL)) 1342c34056a3STejun Heo goto fail; 13438b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1344c34056a3STejun Heo } 13458b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1346c34056a3STejun Heo 1347c34056a3STejun Heo worker = alloc_worker(); 1348c34056a3STejun Heo if (!worker) 1349c34056a3STejun Heo goto fail; 1350c34056a3STejun Heo 13518b03ae3cSTejun Heo worker->gcwq = gcwq; 1352c34056a3STejun Heo worker->id = id; 1353c34056a3STejun Heo 1354f3421797STejun Heo if (!on_unbound_cpu) 1355f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1356f3421797STejun Heo "kworker/%u:%d", gcwq->cpu, id); 1357f3421797STejun Heo else 1358f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1359f3421797STejun Heo "kworker/u:%d", id); 1360c34056a3STejun Heo if (IS_ERR(worker->task)) 1361c34056a3STejun Heo goto fail; 1362c34056a3STejun Heo 1363db7bccf4STejun Heo /* 1364db7bccf4STejun Heo * A rogue worker will become a regular one if CPU comes 1365db7bccf4STejun Heo * online later on. Make sure every worker has 1366db7bccf4STejun Heo * PF_THREAD_BOUND set. 1367db7bccf4STejun Heo */ 1368f3421797STejun Heo if (bind && !on_unbound_cpu) 13698b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1370f3421797STejun Heo else { 1371db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1372f3421797STejun Heo if (on_unbound_cpu) 1373f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1374f3421797STejun Heo } 1375c34056a3STejun Heo 1376c34056a3STejun Heo return worker; 1377c34056a3STejun Heo fail: 1378c34056a3STejun Heo if (id >= 0) { 13798b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 13808b03ae3cSTejun Heo ida_remove(&gcwq->worker_ida, id); 13818b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1382c34056a3STejun Heo } 1383c34056a3STejun Heo kfree(worker); 1384c34056a3STejun Heo return NULL; 1385c34056a3STejun Heo } 1386c34056a3STejun Heo 1387c34056a3STejun Heo /** 1388c34056a3STejun Heo * start_worker - start a newly created worker 1389c34056a3STejun Heo * @worker: worker to start 1390c34056a3STejun Heo * 1391c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1392c34056a3STejun Heo * 1393c34056a3STejun Heo * CONTEXT: 13948b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1395c34056a3STejun Heo */ 1396c34056a3STejun Heo static void start_worker(struct worker *worker) 1397c34056a3STejun Heo { 1398cb444766STejun Heo worker->flags |= WORKER_STARTED; 1399c8e55f36STejun Heo worker->gcwq->nr_workers++; 1400c8e55f36STejun Heo worker_enter_idle(worker); 1401c34056a3STejun Heo wake_up_process(worker->task); 1402c34056a3STejun Heo } 1403c34056a3STejun Heo 1404c34056a3STejun Heo /** 1405c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1406c34056a3STejun Heo * @worker: worker to be destroyed 1407c34056a3STejun Heo * 1408c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1409c8e55f36STejun Heo * 1410c8e55f36STejun Heo * CONTEXT: 1411c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1412c34056a3STejun Heo */ 1413c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1414c34056a3STejun Heo { 14158b03ae3cSTejun Heo struct global_cwq *gcwq = worker->gcwq; 1416c34056a3STejun Heo int id = worker->id; 1417c34056a3STejun Heo 1418c34056a3STejun Heo /* sanity check frenzy */ 1419c34056a3STejun Heo BUG_ON(worker->current_work); 1420affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1421c34056a3STejun Heo 1422c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1423c8e55f36STejun Heo gcwq->nr_workers--; 1424c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1425c8e55f36STejun Heo gcwq->nr_idle--; 1426c8e55f36STejun Heo 1427c8e55f36STejun Heo list_del_init(&worker->entry); 1428cb444766STejun Heo worker->flags |= WORKER_DIE; 1429c8e55f36STejun Heo 1430c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1431c8e55f36STejun Heo 1432c34056a3STejun Heo kthread_stop(worker->task); 1433c34056a3STejun Heo kfree(worker); 1434c34056a3STejun Heo 14358b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 14368b03ae3cSTejun Heo ida_remove(&gcwq->worker_ida, id); 1437c34056a3STejun Heo } 1438c34056a3STejun Heo 1439e22bee78STejun Heo static void idle_worker_timeout(unsigned long __gcwq) 1440e22bee78STejun Heo { 1441e22bee78STejun Heo struct global_cwq *gcwq = (void *)__gcwq; 1442e22bee78STejun Heo 1443e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1444e22bee78STejun Heo 1445e22bee78STejun Heo if (too_many_workers(gcwq)) { 1446e22bee78STejun Heo struct worker *worker; 1447e22bee78STejun Heo unsigned long expires; 1448e22bee78STejun Heo 1449e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 1450e22bee78STejun Heo worker = list_entry(gcwq->idle_list.prev, struct worker, entry); 1451e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1452e22bee78STejun Heo 1453e22bee78STejun Heo if (time_before(jiffies, expires)) 1454e22bee78STejun Heo mod_timer(&gcwq->idle_timer, expires); 1455e22bee78STejun Heo else { 1456e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 1457e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGE_WORKERS; 1458e22bee78STejun Heo wake_up_worker(gcwq); 1459e22bee78STejun Heo } 1460e22bee78STejun Heo } 1461e22bee78STejun Heo 1462e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1463e22bee78STejun Heo } 1464e22bee78STejun Heo 1465e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1466e22bee78STejun Heo { 1467e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1468e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1469f3421797STejun Heo unsigned int cpu; 1470e22bee78STejun Heo 1471e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1472e22bee78STejun Heo return false; 1473e22bee78STejun Heo 1474e22bee78STejun Heo /* mayday mayday mayday */ 1475f3421797STejun Heo cpu = cwq->gcwq->cpu; 1476f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1477f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1478f3421797STejun Heo cpu = 0; 1479f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1480e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1481e22bee78STejun Heo return true; 1482e22bee78STejun Heo } 1483e22bee78STejun Heo 1484e22bee78STejun Heo static void gcwq_mayday_timeout(unsigned long __gcwq) 1485e22bee78STejun Heo { 1486e22bee78STejun Heo struct global_cwq *gcwq = (void *)__gcwq; 1487e22bee78STejun Heo struct work_struct *work; 1488e22bee78STejun Heo 1489e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1490e22bee78STejun Heo 1491e22bee78STejun Heo if (need_to_create_worker(gcwq)) { 1492e22bee78STejun Heo /* 1493e22bee78STejun Heo * We've been trying to create a new worker but 1494e22bee78STejun Heo * haven't been successful. We might be hitting an 1495e22bee78STejun Heo * allocation deadlock. Send distress signals to 1496e22bee78STejun Heo * rescuers. 1497e22bee78STejun Heo */ 1498e22bee78STejun Heo list_for_each_entry(work, &gcwq->worklist, entry) 1499e22bee78STejun Heo send_mayday(work); 1500e22bee78STejun Heo } 1501e22bee78STejun Heo 1502e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1503e22bee78STejun Heo 1504e22bee78STejun Heo mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL); 1505e22bee78STejun Heo } 1506e22bee78STejun Heo 1507e22bee78STejun Heo /** 1508e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 1509e22bee78STejun Heo * @gcwq: gcwq to create a new worker for 1510e22bee78STejun Heo * 1511e22bee78STejun Heo * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to 1512e22bee78STejun Heo * have at least one idle worker on return from this function. If 1513e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 1514e22bee78STejun Heo * sent to all rescuers with works scheduled on @gcwq to resolve 1515e22bee78STejun Heo * possible allocation deadlock. 1516e22bee78STejun Heo * 1517e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1518e22bee78STejun Heo * may_start_working() true. 1519e22bee78STejun Heo * 1520e22bee78STejun Heo * LOCKING: 1521e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1522e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1523e22bee78STejun Heo * manager. 1524e22bee78STejun Heo * 1525e22bee78STejun Heo * RETURNS: 1526e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1527e22bee78STejun Heo * otherwise. 1528e22bee78STejun Heo */ 1529e22bee78STejun Heo static bool maybe_create_worker(struct global_cwq *gcwq) 153006bd6ebfSNamhyung Kim __releases(&gcwq->lock) 153106bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 1532e22bee78STejun Heo { 1533e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1534e22bee78STejun Heo return false; 1535e22bee78STejun Heo restart: 15369f9c2364STejun Heo spin_unlock_irq(&gcwq->lock); 15379f9c2364STejun Heo 1538e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 1539e22bee78STejun Heo mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1540e22bee78STejun Heo 1541e22bee78STejun Heo while (true) { 1542e22bee78STejun Heo struct worker *worker; 1543e22bee78STejun Heo 1544e22bee78STejun Heo worker = create_worker(gcwq, true); 1545e22bee78STejun Heo if (worker) { 1546e22bee78STejun Heo del_timer_sync(&gcwq->mayday_timer); 1547e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1548e22bee78STejun Heo start_worker(worker); 1549e22bee78STejun Heo BUG_ON(need_to_create_worker(gcwq)); 1550e22bee78STejun Heo return true; 1551e22bee78STejun Heo } 1552e22bee78STejun Heo 1553e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1554e22bee78STejun Heo break; 1555e22bee78STejun Heo 1556e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1557e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 15589f9c2364STejun Heo 1559e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1560e22bee78STejun Heo break; 1561e22bee78STejun Heo } 1562e22bee78STejun Heo 1563e22bee78STejun Heo del_timer_sync(&gcwq->mayday_timer); 1564e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1565e22bee78STejun Heo if (need_to_create_worker(gcwq)) 1566e22bee78STejun Heo goto restart; 1567e22bee78STejun Heo return true; 1568e22bee78STejun Heo } 1569e22bee78STejun Heo 1570e22bee78STejun Heo /** 1571e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 1572e22bee78STejun Heo * @gcwq: gcwq to destroy workers for 1573e22bee78STejun Heo * 1574e22bee78STejun Heo * Destroy @gcwq workers which have been idle for longer than 1575e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1576e22bee78STejun Heo * 1577e22bee78STejun Heo * LOCKING: 1578e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1579e22bee78STejun Heo * multiple times. Called only from manager. 1580e22bee78STejun Heo * 1581e22bee78STejun Heo * RETURNS: 1582e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1583e22bee78STejun Heo * otherwise. 1584e22bee78STejun Heo */ 1585e22bee78STejun Heo static bool maybe_destroy_workers(struct global_cwq *gcwq) 1586e22bee78STejun Heo { 1587e22bee78STejun Heo bool ret = false; 1588e22bee78STejun Heo 1589e22bee78STejun Heo while (too_many_workers(gcwq)) { 1590e22bee78STejun Heo struct worker *worker; 1591e22bee78STejun Heo unsigned long expires; 1592e22bee78STejun Heo 1593e22bee78STejun Heo worker = list_entry(gcwq->idle_list.prev, struct worker, entry); 1594e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1595e22bee78STejun Heo 1596e22bee78STejun Heo if (time_before(jiffies, expires)) { 1597e22bee78STejun Heo mod_timer(&gcwq->idle_timer, expires); 1598e22bee78STejun Heo break; 1599e22bee78STejun Heo } 1600e22bee78STejun Heo 1601e22bee78STejun Heo destroy_worker(worker); 1602e22bee78STejun Heo ret = true; 1603e22bee78STejun Heo } 1604e22bee78STejun Heo 1605e22bee78STejun Heo return ret; 1606e22bee78STejun Heo } 1607e22bee78STejun Heo 1608e22bee78STejun Heo /** 1609e22bee78STejun Heo * manage_workers - manage worker pool 1610e22bee78STejun Heo * @worker: self 1611e22bee78STejun Heo * 1612e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 1613e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1614e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 1615e22bee78STejun Heo * 1616e22bee78STejun Heo * The caller can safely start processing works on false return. On 1617e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1618e22bee78STejun Heo * and may_start_working() is true. 1619e22bee78STejun Heo * 1620e22bee78STejun Heo * CONTEXT: 1621e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1622e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1623e22bee78STejun Heo * 1624e22bee78STejun Heo * RETURNS: 1625e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 1626e22bee78STejun Heo * some action was taken. 1627e22bee78STejun Heo */ 1628e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1629e22bee78STejun Heo { 1630e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1631e22bee78STejun Heo bool ret = false; 1632e22bee78STejun Heo 1633e22bee78STejun Heo if (gcwq->flags & GCWQ_MANAGING_WORKERS) 1634e22bee78STejun Heo return ret; 1635e22bee78STejun Heo 1636e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGE_WORKERS; 1637e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGING_WORKERS; 1638e22bee78STejun Heo 1639e22bee78STejun Heo /* 1640e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 1641e22bee78STejun Heo * on return. 1642e22bee78STejun Heo */ 1643e22bee78STejun Heo ret |= maybe_destroy_workers(gcwq); 1644e22bee78STejun Heo ret |= maybe_create_worker(gcwq); 1645e22bee78STejun Heo 1646e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGING_WORKERS; 1647e22bee78STejun Heo 1648e22bee78STejun Heo /* 1649e22bee78STejun Heo * The trustee might be waiting to take over the manager 1650e22bee78STejun Heo * position, tell it we're done. 1651e22bee78STejun Heo */ 1652e22bee78STejun Heo if (unlikely(gcwq->trustee)) 1653e22bee78STejun Heo wake_up_all(&gcwq->trustee_wait); 1654e22bee78STejun Heo 1655e22bee78STejun Heo return ret; 1656e22bee78STejun Heo } 1657e22bee78STejun Heo 1658a62428c0STejun Heo /** 1659affee4b2STejun Heo * move_linked_works - move linked works to a list 1660affee4b2STejun Heo * @work: start of series of works to be scheduled 1661affee4b2STejun Heo * @head: target list to append @work to 1662affee4b2STejun Heo * @nextp: out paramter for nested worklist walking 1663affee4b2STejun Heo * 1664affee4b2STejun Heo * Schedule linked works starting from @work to @head. Work series to 1665affee4b2STejun Heo * be scheduled starts at @work and includes any consecutive work with 1666affee4b2STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1667affee4b2STejun Heo * 1668affee4b2STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1669affee4b2STejun Heo * the last scheduled work. This allows move_linked_works() to be 1670affee4b2STejun Heo * nested inside outer list_for_each_entry_safe(). 1671affee4b2STejun Heo * 1672affee4b2STejun Heo * CONTEXT: 16738b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1674affee4b2STejun Heo */ 1675affee4b2STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1676affee4b2STejun Heo struct work_struct **nextp) 1677affee4b2STejun Heo { 1678affee4b2STejun Heo struct work_struct *n; 1679affee4b2STejun Heo 1680affee4b2STejun Heo /* 1681affee4b2STejun Heo * Linked worklist will always end before the end of the list, 1682affee4b2STejun Heo * use NULL for list head. 1683affee4b2STejun Heo */ 1684affee4b2STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1685affee4b2STejun Heo list_move_tail(&work->entry, head); 1686affee4b2STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1687affee4b2STejun Heo break; 1688affee4b2STejun Heo } 1689affee4b2STejun Heo 1690affee4b2STejun Heo /* 1691affee4b2STejun Heo * If we're already inside safe list traversal and have moved 1692affee4b2STejun Heo * multiple works to the scheduled queue, the next position 1693affee4b2STejun Heo * needs to be updated. 1694affee4b2STejun Heo */ 1695affee4b2STejun Heo if (nextp) 1696affee4b2STejun Heo *nextp = n; 1697affee4b2STejun Heo } 1698affee4b2STejun Heo 16991e19ffc6STejun Heo static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 17001e19ffc6STejun Heo { 17011e19ffc6STejun Heo struct work_struct *work = list_first_entry(&cwq->delayed_works, 17021da177e4SLinus Torvalds struct work_struct, entry); 1703649027d7STejun Heo struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq); 17041e19ffc6STejun Heo 1705cdadf009STejun Heo trace_workqueue_activate_work(work); 1706649027d7STejun Heo move_linked_works(work, pos, NULL); 17078a2e8e5dSTejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 17081e19ffc6STejun Heo cwq->nr_active++; 17091e19ffc6STejun Heo } 17101e19ffc6STejun Heo 1711affee4b2STejun Heo /** 171273f53c4aSTejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 171373f53c4aSTejun Heo * @cwq: cwq of interest 171473f53c4aSTejun Heo * @color: color of work which left the queue 17158a2e8e5dSTejun Heo * @delayed: for a delayed work 171673f53c4aSTejun Heo * 171773f53c4aSTejun Heo * A work either has completed or is removed from pending queue, 171873f53c4aSTejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 171973f53c4aSTejun Heo * 172073f53c4aSTejun Heo * CONTEXT: 17218b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 172273f53c4aSTejun Heo */ 17238a2e8e5dSTejun Heo static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color, 17248a2e8e5dSTejun Heo bool delayed) 172573f53c4aSTejun Heo { 172673f53c4aSTejun Heo /* ignore uncolored works */ 172773f53c4aSTejun Heo if (color == WORK_NO_COLOR) 172873f53c4aSTejun Heo return; 172973f53c4aSTejun Heo 173073f53c4aSTejun Heo cwq->nr_in_flight[color]--; 17311e19ffc6STejun Heo 17328a2e8e5dSTejun Heo if (!delayed) { 17338a2e8e5dSTejun Heo cwq->nr_active--; 1734502ca9d8STejun Heo if (!list_empty(&cwq->delayed_works)) { 17351e19ffc6STejun Heo /* one down, submit a delayed one */ 1736502ca9d8STejun Heo if (cwq->nr_active < cwq->max_active) 17371e19ffc6STejun Heo cwq_activate_first_delayed(cwq); 1738502ca9d8STejun Heo } 17398a2e8e5dSTejun Heo } 174073f53c4aSTejun Heo 174173f53c4aSTejun Heo /* is flush in progress and are we at the flushing tip? */ 174273f53c4aSTejun Heo if (likely(cwq->flush_color != color)) 174373f53c4aSTejun Heo return; 174473f53c4aSTejun Heo 174573f53c4aSTejun Heo /* are there still in-flight works? */ 174673f53c4aSTejun Heo if (cwq->nr_in_flight[color]) 174773f53c4aSTejun Heo return; 174873f53c4aSTejun Heo 174973f53c4aSTejun Heo /* this cwq is done, clear flush_color */ 175073f53c4aSTejun Heo cwq->flush_color = -1; 175173f53c4aSTejun Heo 175273f53c4aSTejun Heo /* 175373f53c4aSTejun Heo * If this was the last cwq, wake up the first flusher. It 175473f53c4aSTejun Heo * will handle the rest. 175573f53c4aSTejun Heo */ 175673f53c4aSTejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 175773f53c4aSTejun Heo complete(&cwq->wq->first_flusher->done); 175873f53c4aSTejun Heo } 175973f53c4aSTejun Heo 176073f53c4aSTejun Heo /** 1761a62428c0STejun Heo * process_one_work - process single work 1762c34056a3STejun Heo * @worker: self 1763a62428c0STejun Heo * @work: work to process 1764a62428c0STejun Heo * 1765a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1766a62428c0STejun Heo * process a single work including synchronization against and 1767a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1768a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1769a62428c0STejun Heo * call this function to process a work. 1770a62428c0STejun Heo * 1771a62428c0STejun Heo * CONTEXT: 17728b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1773a62428c0STejun Heo */ 1774c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 177506bd6ebfSNamhyung Kim __releases(&gcwq->lock) 177606bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 17771da177e4SLinus Torvalds { 17787e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 17798b03ae3cSTejun Heo struct global_cwq *gcwq = cwq->gcwq; 1780c8e55f36STejun Heo struct hlist_head *bwh = busy_worker_head(gcwq, work); 1781fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 17826bb49e59SDavid Howells work_func_t f = work->func; 178373f53c4aSTejun Heo int work_color; 17847e11629dSTejun Heo struct worker *collision; 17854e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 17864e6045f1SJohannes Berg /* 1787a62428c0STejun Heo * It is permissible to free the struct work_struct from 1788a62428c0STejun Heo * inside the function that is called from it, this we need to 1789a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1790a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1791a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 17924e6045f1SJohannes Berg */ 17934e6045f1SJohannes Berg struct lockdep_map lockdep_map = work->lockdep_map; 17944e6045f1SJohannes Berg #endif 17957e11629dSTejun Heo /* 17967e11629dSTejun Heo * A single work shouldn't be executed concurrently by 17977e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 17987e11629dSTejun Heo * already processing the work. If so, defer the work to the 17997e11629dSTejun Heo * currently executing one. 18007e11629dSTejun Heo */ 18017e11629dSTejun Heo collision = __find_worker_executing_work(gcwq, bwh, work); 18027e11629dSTejun Heo if (unlikely(collision)) { 18037e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 18047e11629dSTejun Heo return; 18057e11629dSTejun Heo } 18061da177e4SLinus Torvalds 1807a62428c0STejun Heo /* claim and process */ 18081da177e4SLinus Torvalds debug_work_deactivate(work); 1809c8e55f36STejun Heo hlist_add_head(&worker->hentry, bwh); 1810c34056a3STejun Heo worker->current_work = work; 18118cca0eeaSTejun Heo worker->current_cwq = cwq; 181273f53c4aSTejun Heo work_color = get_work_color(work); 18137a22ad75STejun Heo 18147a22ad75STejun Heo /* record the current cpu number in the work data and dequeue */ 18157a22ad75STejun Heo set_work_cpu(work, gcwq->cpu); 1816a62428c0STejun Heo list_del_init(&work->entry); 1817a62428c0STejun Heo 1818649027d7STejun Heo /* 1819649027d7STejun Heo * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI, 1820649027d7STejun Heo * wake up another worker; otherwise, clear HIGHPRI_PENDING. 1821649027d7STejun Heo */ 1822649027d7STejun Heo if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) { 1823649027d7STejun Heo struct work_struct *nwork = list_first_entry(&gcwq->worklist, 1824649027d7STejun Heo struct work_struct, entry); 1825649027d7STejun Heo 1826649027d7STejun Heo if (!list_empty(&gcwq->worklist) && 1827649027d7STejun Heo get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI) 1828649027d7STejun Heo wake_up_worker(gcwq); 1829649027d7STejun Heo else 1830649027d7STejun Heo gcwq->flags &= ~GCWQ_HIGHPRI_PENDING; 1831649027d7STejun Heo } 1832649027d7STejun Heo 1833fb0e7bebSTejun Heo /* 1834fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 1835fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 1836fb0e7bebSTejun Heo */ 1837fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1838fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 1839fb0e7bebSTejun Heo 18408b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 18411da177e4SLinus Torvalds 184223b2e599SOleg Nesterov work_clear_pending(work); 1843e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 18443295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 1845e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 184665f27f38SDavid Howells f(work); 1847e36c886aSArjan van de Ven /* 1848e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 1849e36c886aSArjan van de Ven * point will only record its address. 1850e36c886aSArjan van de Ven */ 1851e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 18523295f0efSIngo Molnar lock_map_release(&lockdep_map); 18533295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 18541da177e4SLinus Torvalds 1855d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 1856d5abe669SPeter Zijlstra printk(KERN_ERR "BUG: workqueue leaked lock or atomic: " 1857d5abe669SPeter Zijlstra "%s/0x%08x/%d\n", 1858a62428c0STejun Heo current->comm, preempt_count(), task_pid_nr(current)); 1859d5abe669SPeter Zijlstra printk(KERN_ERR " last function: "); 1860d5abe669SPeter Zijlstra print_symbol("%s\n", (unsigned long)f); 1861d5abe669SPeter Zijlstra debug_show_held_locks(current); 1862d5abe669SPeter Zijlstra dump_stack(); 1863d5abe669SPeter Zijlstra } 1864d5abe669SPeter Zijlstra 18658b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1866a62428c0STejun Heo 1867fb0e7bebSTejun Heo /* clear cpu intensive status */ 1868fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1869fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 1870fb0e7bebSTejun Heo 1871a62428c0STejun Heo /* we're done with it, release */ 1872c8e55f36STejun Heo hlist_del_init(&worker->hentry); 1873c34056a3STejun Heo worker->current_work = NULL; 18748cca0eeaSTejun Heo worker->current_cwq = NULL; 18758a2e8e5dSTejun Heo cwq_dec_nr_in_flight(cwq, work_color, false); 18761da177e4SLinus Torvalds } 18771da177e4SLinus Torvalds 1878affee4b2STejun Heo /** 1879affee4b2STejun Heo * process_scheduled_works - process scheduled works 1880affee4b2STejun Heo * @worker: self 1881affee4b2STejun Heo * 1882affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 1883affee4b2STejun Heo * may change while processing a work, so this function repeatedly 1884affee4b2STejun Heo * fetches a work from the top and executes it. 1885affee4b2STejun Heo * 1886affee4b2STejun Heo * CONTEXT: 18878b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1888affee4b2STejun Heo * multiple times. 1889affee4b2STejun Heo */ 1890affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 18911da177e4SLinus Torvalds { 1892affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 1893affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 1894a62428c0STejun Heo struct work_struct, entry); 1895c34056a3STejun Heo process_one_work(worker, work); 1896a62428c0STejun Heo } 18971da177e4SLinus Torvalds } 18981da177e4SLinus Torvalds 18994690c4abSTejun Heo /** 19004690c4abSTejun Heo * worker_thread - the worker thread function 1901c34056a3STejun Heo * @__worker: self 19024690c4abSTejun Heo * 1903e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 1904e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 1905e22bee78STejun Heo * their specific target workqueue. The only exception is works which 1906e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 1907e22bee78STejun Heo * rescuer_thread(). 19084690c4abSTejun Heo */ 1909c34056a3STejun Heo static int worker_thread(void *__worker) 19101da177e4SLinus Torvalds { 1911c34056a3STejun Heo struct worker *worker = __worker; 19128b03ae3cSTejun Heo struct global_cwq *gcwq = worker->gcwq; 19131da177e4SLinus Torvalds 1914e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 1915e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 1916c8e55f36STejun Heo woke_up: 19178b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1918affee4b2STejun Heo 1919c8e55f36STejun Heo /* DIE can be set only while we're idle, checking here is enough */ 1920c8e55f36STejun Heo if (worker->flags & WORKER_DIE) { 1921c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1922e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 1923c8e55f36STejun Heo return 0; 1924c8e55f36STejun Heo } 1925c8e55f36STejun Heo 1926c8e55f36STejun Heo worker_leave_idle(worker); 1927db7bccf4STejun Heo recheck: 1928e22bee78STejun Heo /* no more worker necessary? */ 1929e22bee78STejun Heo if (!need_more_worker(gcwq)) 1930e22bee78STejun Heo goto sleep; 1931e22bee78STejun Heo 1932e22bee78STejun Heo /* do we need to manage? */ 1933e22bee78STejun Heo if (unlikely(!may_start_working(gcwq)) && manage_workers(worker)) 1934e22bee78STejun Heo goto recheck; 1935e22bee78STejun Heo 1936c8e55f36STejun Heo /* 1937c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 1938c8e55f36STejun Heo * preparing to process a work or actually processing it. 1939c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 1940c8e55f36STejun Heo */ 1941c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1942c8e55f36STejun Heo 1943e22bee78STejun Heo /* 1944e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 1945e22bee78STejun Heo * at least one idle worker or that someone else has already 1946e22bee78STejun Heo * assumed the manager role. 1947e22bee78STejun Heo */ 1948e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 1949e22bee78STejun Heo 1950e22bee78STejun Heo do { 1951affee4b2STejun Heo struct work_struct *work = 19527e11629dSTejun Heo list_first_entry(&gcwq->worklist, 1953affee4b2STejun Heo struct work_struct, entry); 1954affee4b2STejun Heo 1955c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 1956affee4b2STejun Heo /* optimization path, not strictly necessary */ 1957affee4b2STejun Heo process_one_work(worker, work); 1958affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 1959affee4b2STejun Heo process_scheduled_works(worker); 1960affee4b2STejun Heo } else { 1961c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 1962affee4b2STejun Heo process_scheduled_works(worker); 1963affee4b2STejun Heo } 1964e22bee78STejun Heo } while (keep_working(gcwq)); 1965affee4b2STejun Heo 1966e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 1967d313dd85STejun Heo sleep: 1968e22bee78STejun Heo if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker)) 1969e22bee78STejun Heo goto recheck; 1970d313dd85STejun Heo 1971c8e55f36STejun Heo /* 1972e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 1973e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 1974e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 1975e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 1976e22bee78STejun Heo * prevent losing any event. 1977c8e55f36STejun Heo */ 1978c8e55f36STejun Heo worker_enter_idle(worker); 1979c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 19808b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 19811da177e4SLinus Torvalds schedule(); 1982c8e55f36STejun Heo goto woke_up; 19831da177e4SLinus Torvalds } 19841da177e4SLinus Torvalds 1985e22bee78STejun Heo /** 1986e22bee78STejun Heo * rescuer_thread - the rescuer thread function 1987e22bee78STejun Heo * @__wq: the associated workqueue 1988e22bee78STejun Heo * 1989e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 1990e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 1991e22bee78STejun Heo * 1992e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 1993e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 1994e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 1995e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 1996e22bee78STejun Heo * the problem rescuer solves. 1997e22bee78STejun Heo * 1998e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 1999e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 2000e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2001e22bee78STejun Heo * 2002e22bee78STejun Heo * This should happen rarely. 2003e22bee78STejun Heo */ 2004e22bee78STejun Heo static int rescuer_thread(void *__wq) 2005e22bee78STejun Heo { 2006e22bee78STejun Heo struct workqueue_struct *wq = __wq; 2007e22bee78STejun Heo struct worker *rescuer = wq->rescuer; 2008e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2009f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2010e22bee78STejun Heo unsigned int cpu; 2011e22bee78STejun Heo 2012e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2013e22bee78STejun Heo repeat: 2014e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 20151da177e4SLinus Torvalds 20161da177e4SLinus Torvalds if (kthread_should_stop()) 2017e22bee78STejun Heo return 0; 20181da177e4SLinus Torvalds 2019f3421797STejun Heo /* 2020f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2021f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2022f3421797STejun Heo */ 2023f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2024f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2025f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2026e22bee78STejun Heo struct global_cwq *gcwq = cwq->gcwq; 2027e22bee78STejun Heo struct work_struct *work, *n; 2028e22bee78STejun Heo 2029e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2030f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2031e22bee78STejun Heo 2032e22bee78STejun Heo /* migrate to the target cpu if possible */ 2033e22bee78STejun Heo rescuer->gcwq = gcwq; 2034e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2035e22bee78STejun Heo 2036e22bee78STejun Heo /* 2037e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2038e22bee78STejun Heo * process'em. 2039e22bee78STejun Heo */ 2040e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2041e22bee78STejun Heo list_for_each_entry_safe(work, n, &gcwq->worklist, entry) 2042e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2043e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2044e22bee78STejun Heo 2045e22bee78STejun Heo process_scheduled_works(rescuer); 2046e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 20471da177e4SLinus Torvalds } 20481da177e4SLinus Torvalds 2049e22bee78STejun Heo schedule(); 2050e22bee78STejun Heo goto repeat; 20511da177e4SLinus Torvalds } 20521da177e4SLinus Torvalds 2053fc2e4d70SOleg Nesterov struct wq_barrier { 2054fc2e4d70SOleg Nesterov struct work_struct work; 2055fc2e4d70SOleg Nesterov struct completion done; 2056fc2e4d70SOleg Nesterov }; 2057fc2e4d70SOleg Nesterov 2058fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2059fc2e4d70SOleg Nesterov { 2060fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2061fc2e4d70SOleg Nesterov complete(&barr->done); 2062fc2e4d70SOleg Nesterov } 2063fc2e4d70SOleg Nesterov 20644690c4abSTejun Heo /** 20654690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 20664690c4abSTejun Heo * @cwq: cwq to insert barrier into 20674690c4abSTejun Heo * @barr: wq_barrier to insert 2068affee4b2STejun Heo * @target: target work to attach @barr to 2069affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 20704690c4abSTejun Heo * 2071affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2072affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2073affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2074affee4b2STejun Heo * cpu. 2075affee4b2STejun Heo * 2076affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2077affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2078affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2079affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2080affee4b2STejun Heo * after a work with LINKED flag set. 2081affee4b2STejun Heo * 2082affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2083affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 20844690c4abSTejun Heo * 20854690c4abSTejun Heo * CONTEXT: 20868b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 20874690c4abSTejun Heo */ 208883c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2089affee4b2STejun Heo struct wq_barrier *barr, 2090affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2091fc2e4d70SOleg Nesterov { 2092affee4b2STejun Heo struct list_head *head; 2093affee4b2STejun Heo unsigned int linked = 0; 2094affee4b2STejun Heo 2095dc186ad7SThomas Gleixner /* 20968b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 2097dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2098dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2099dc186ad7SThomas Gleixner * might deadlock. 2100dc186ad7SThomas Gleixner */ 2101ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 210222df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2103fc2e4d70SOleg Nesterov init_completion(&barr->done); 210483c22520SOleg Nesterov 2105affee4b2STejun Heo /* 2106affee4b2STejun Heo * If @target is currently being executed, schedule the 2107affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2108affee4b2STejun Heo */ 2109affee4b2STejun Heo if (worker) 2110affee4b2STejun Heo head = worker->scheduled.next; 2111affee4b2STejun Heo else { 2112affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2113affee4b2STejun Heo 2114affee4b2STejun Heo head = target->entry.next; 2115affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2116affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2117affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2118affee4b2STejun Heo } 2119affee4b2STejun Heo 2120dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2121affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2122affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2123fc2e4d70SOleg Nesterov } 2124fc2e4d70SOleg Nesterov 212573f53c4aSTejun Heo /** 212673f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 212773f53c4aSTejun Heo * @wq: workqueue being flushed 212873f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 212973f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 213073f53c4aSTejun Heo * 213173f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 213273f53c4aSTejun Heo * 213373f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 213473f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 213573f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 213673f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 213773f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 213873f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 213973f53c4aSTejun Heo * 214073f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 214173f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 214273f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 214373f53c4aSTejun Heo * is returned. 214473f53c4aSTejun Heo * 214573f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 214673f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 214773f53c4aSTejun Heo * advanced to @work_color. 214873f53c4aSTejun Heo * 214973f53c4aSTejun Heo * CONTEXT: 215073f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 215173f53c4aSTejun Heo * 215273f53c4aSTejun Heo * RETURNS: 215373f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 215473f53c4aSTejun Heo * otherwise. 215573f53c4aSTejun Heo */ 215673f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 215773f53c4aSTejun Heo int flush_color, int work_color) 21581da177e4SLinus Torvalds { 215973f53c4aSTejun Heo bool wait = false; 216073f53c4aSTejun Heo unsigned int cpu; 21611da177e4SLinus Torvalds 216273f53c4aSTejun Heo if (flush_color >= 0) { 216373f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 216473f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2165dc186ad7SThomas Gleixner } 216614441960SOleg Nesterov 2167f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 216873f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 21698b03ae3cSTejun Heo struct global_cwq *gcwq = cwq->gcwq; 21701da177e4SLinus Torvalds 21718b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 217273f53c4aSTejun Heo 217373f53c4aSTejun Heo if (flush_color >= 0) { 217473f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 217573f53c4aSTejun Heo 217673f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 217773f53c4aSTejun Heo cwq->flush_color = flush_color; 217873f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 217973f53c4aSTejun Heo wait = true; 21801da177e4SLinus Torvalds } 218173f53c4aSTejun Heo } 218273f53c4aSTejun Heo 218373f53c4aSTejun Heo if (work_color >= 0) { 218473f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 218573f53c4aSTejun Heo cwq->work_color = work_color; 218673f53c4aSTejun Heo } 218773f53c4aSTejun Heo 21888b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 21891da177e4SLinus Torvalds } 21901da177e4SLinus Torvalds 219173f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 219273f53c4aSTejun Heo complete(&wq->first_flusher->done); 219373f53c4aSTejun Heo 219473f53c4aSTejun Heo return wait; 219583c22520SOleg Nesterov } 21961da177e4SLinus Torvalds 21970fcb78c2SRolf Eike Beer /** 21981da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 21990fcb78c2SRolf Eike Beer * @wq: workqueue to flush 22001da177e4SLinus Torvalds * 22011da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 22021da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 22031da177e4SLinus Torvalds * 2204fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2205fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 22061da177e4SLinus Torvalds */ 22077ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 22081da177e4SLinus Torvalds { 220973f53c4aSTejun Heo struct wq_flusher this_flusher = { 221073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 221173f53c4aSTejun Heo .flush_color = -1, 221273f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 221373f53c4aSTejun Heo }; 221473f53c4aSTejun Heo int next_color; 2215b1f4ec17SOleg Nesterov 22163295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 22173295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 221873f53c4aSTejun Heo 221973f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 222073f53c4aSTejun Heo 222173f53c4aSTejun Heo /* 222273f53c4aSTejun Heo * Start-to-wait phase 222373f53c4aSTejun Heo */ 222473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 222573f53c4aSTejun Heo 222673f53c4aSTejun Heo if (next_color != wq->flush_color) { 222773f53c4aSTejun Heo /* 222873f53c4aSTejun Heo * Color space is not full. The current work_color 222973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 223073f53c4aSTejun Heo * by one. 223173f53c4aSTejun Heo */ 223273f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 223373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 223473f53c4aSTejun Heo wq->work_color = next_color; 223573f53c4aSTejun Heo 223673f53c4aSTejun Heo if (!wq->first_flusher) { 223773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 223873f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 223973f53c4aSTejun Heo 224073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 224173f53c4aSTejun Heo 224273f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 224373f53c4aSTejun Heo wq->work_color)) { 224473f53c4aSTejun Heo /* nothing to flush, done */ 224573f53c4aSTejun Heo wq->flush_color = next_color; 224673f53c4aSTejun Heo wq->first_flusher = NULL; 224773f53c4aSTejun Heo goto out_unlock; 224873f53c4aSTejun Heo } 224973f53c4aSTejun Heo } else { 225073f53c4aSTejun Heo /* wait in queue */ 225173f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 225273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 225373f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 225473f53c4aSTejun Heo } 225573f53c4aSTejun Heo } else { 225673f53c4aSTejun Heo /* 225773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 225873f53c4aSTejun Heo * The next flush completion will assign us 225973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 226073f53c4aSTejun Heo */ 226173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 226273f53c4aSTejun Heo } 226373f53c4aSTejun Heo 226473f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 226573f53c4aSTejun Heo 226673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 226773f53c4aSTejun Heo 226873f53c4aSTejun Heo /* 226973f53c4aSTejun Heo * Wake-up-and-cascade phase 227073f53c4aSTejun Heo * 227173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 227273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 227373f53c4aSTejun Heo */ 227473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 227573f53c4aSTejun Heo return; 227673f53c4aSTejun Heo 227773f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 227873f53c4aSTejun Heo 22794ce48b37STejun Heo /* we might have raced, check again with mutex held */ 22804ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 22814ce48b37STejun Heo goto out_unlock; 22824ce48b37STejun Heo 228373f53c4aSTejun Heo wq->first_flusher = NULL; 228473f53c4aSTejun Heo 228573f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 228673f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 228773f53c4aSTejun Heo 228873f53c4aSTejun Heo while (true) { 228973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 229073f53c4aSTejun Heo 229173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 229273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 229373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 229473f53c4aSTejun Heo break; 229573f53c4aSTejun Heo list_del_init(&next->list); 229673f53c4aSTejun Heo complete(&next->done); 229773f53c4aSTejun Heo } 229873f53c4aSTejun Heo 229973f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 230073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 230173f53c4aSTejun Heo 230273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 230373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 230473f53c4aSTejun Heo 230573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 230673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 230773f53c4aSTejun Heo /* 230873f53c4aSTejun Heo * Assign the same color to all overflowed 230973f53c4aSTejun Heo * flushers, advance work_color and append to 231073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 231173f53c4aSTejun Heo * phase for these overflowed flushers. 231273f53c4aSTejun Heo */ 231373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 231473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 231573f53c4aSTejun Heo 231673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 231773f53c4aSTejun Heo 231873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 231973f53c4aSTejun Heo &wq->flusher_queue); 232073f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 232173f53c4aSTejun Heo } 232273f53c4aSTejun Heo 232373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 232473f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 232573f53c4aSTejun Heo break; 232673f53c4aSTejun Heo } 232773f53c4aSTejun Heo 232873f53c4aSTejun Heo /* 232973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 233073f53c4aSTejun Heo * the new first flusher and arm cwqs. 233173f53c4aSTejun Heo */ 233273f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 233373f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 233473f53c4aSTejun Heo 233573f53c4aSTejun Heo list_del_init(&next->list); 233673f53c4aSTejun Heo wq->first_flusher = next; 233773f53c4aSTejun Heo 233873f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 233973f53c4aSTejun Heo break; 234073f53c4aSTejun Heo 234173f53c4aSTejun Heo /* 234273f53c4aSTejun Heo * Meh... this color is already done, clear first 234373f53c4aSTejun Heo * flusher and repeat cascading. 234473f53c4aSTejun Heo */ 234573f53c4aSTejun Heo wq->first_flusher = NULL; 234673f53c4aSTejun Heo } 234773f53c4aSTejun Heo 234873f53c4aSTejun Heo out_unlock: 234973f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 23501da177e4SLinus Torvalds } 2351ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 23521da177e4SLinus Torvalds 2353baf59022STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2354baf59022STejun Heo bool wait_executing) 2355baf59022STejun Heo { 2356baf59022STejun Heo struct worker *worker = NULL; 2357baf59022STejun Heo struct global_cwq *gcwq; 2358baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2359baf59022STejun Heo 2360baf59022STejun Heo might_sleep(); 2361baf59022STejun Heo gcwq = get_work_gcwq(work); 2362baf59022STejun Heo if (!gcwq) 2363baf59022STejun Heo return false; 2364baf59022STejun Heo 2365baf59022STejun Heo spin_lock_irq(&gcwq->lock); 2366baf59022STejun Heo if (!list_empty(&work->entry)) { 2367baf59022STejun Heo /* 2368baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2369baf59022STejun Heo * If it was re-queued to a different gcwq under us, we 2370baf59022STejun Heo * are not going to wait. 2371baf59022STejun Heo */ 2372baf59022STejun Heo smp_rmb(); 2373baf59022STejun Heo cwq = get_work_cwq(work); 2374baf59022STejun Heo if (unlikely(!cwq || gcwq != cwq->gcwq)) 2375baf59022STejun Heo goto already_gone; 2376baf59022STejun Heo } else if (wait_executing) { 2377baf59022STejun Heo worker = find_worker_executing_work(gcwq, work); 2378baf59022STejun Heo if (!worker) 2379baf59022STejun Heo goto already_gone; 2380baf59022STejun Heo cwq = worker->current_cwq; 2381baf59022STejun Heo } else 2382baf59022STejun Heo goto already_gone; 2383baf59022STejun Heo 2384baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2385baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2386baf59022STejun Heo 2387e159489bSTejun Heo /* 2388e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2389e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2390e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2391e159489bSTejun Heo * access. 2392e159489bSTejun Heo */ 2393e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2394baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2395e159489bSTejun Heo else 2396e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2397baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2398e159489bSTejun Heo 2399baf59022STejun Heo return true; 2400baf59022STejun Heo already_gone: 2401baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2402baf59022STejun Heo return false; 2403baf59022STejun Heo } 2404baf59022STejun Heo 2405db700897SOleg Nesterov /** 2406401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2407401a8d04STejun Heo * @work: the work to flush 2408db700897SOleg Nesterov * 2409401a8d04STejun Heo * Wait until @work has finished execution. This function considers 2410401a8d04STejun Heo * only the last queueing instance of @work. If @work has been 2411401a8d04STejun Heo * enqueued across different CPUs on a non-reentrant workqueue or on 2412401a8d04STejun Heo * multiple workqueues, @work might still be executing on return on 2413401a8d04STejun Heo * some of the CPUs from earlier queueing. 2414a67da70dSOleg Nesterov * 2415401a8d04STejun Heo * If @work was queued only on a non-reentrant, ordered or unbound 2416401a8d04STejun Heo * workqueue, @work is guaranteed to be idle on return if it hasn't 2417401a8d04STejun Heo * been requeued since flush started. 2418401a8d04STejun Heo * 2419401a8d04STejun Heo * RETURNS: 2420401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2421401a8d04STejun Heo * %false if it was already idle. 2422db700897SOleg Nesterov */ 2423401a8d04STejun Heo bool flush_work(struct work_struct *work) 2424db700897SOleg Nesterov { 2425db700897SOleg Nesterov struct wq_barrier barr; 2426db700897SOleg Nesterov 2427baf59022STejun Heo if (start_flush_work(work, &barr, true)) { 2428db700897SOleg Nesterov wait_for_completion(&barr.done); 2429dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2430401a8d04STejun Heo return true; 2431baf59022STejun Heo } else 2432401a8d04STejun Heo return false; 2433db700897SOleg Nesterov } 2434db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2435db700897SOleg Nesterov 2436401a8d04STejun Heo static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work) 2437401a8d04STejun Heo { 2438401a8d04STejun Heo struct wq_barrier barr; 2439401a8d04STejun Heo struct worker *worker; 2440401a8d04STejun Heo 2441401a8d04STejun Heo spin_lock_irq(&gcwq->lock); 2442401a8d04STejun Heo 2443401a8d04STejun Heo worker = find_worker_executing_work(gcwq, work); 2444401a8d04STejun Heo if (unlikely(worker)) 2445401a8d04STejun Heo insert_wq_barrier(worker->current_cwq, &barr, work, worker); 2446401a8d04STejun Heo 2447401a8d04STejun Heo spin_unlock_irq(&gcwq->lock); 2448401a8d04STejun Heo 2449401a8d04STejun Heo if (unlikely(worker)) { 2450401a8d04STejun Heo wait_for_completion(&barr.done); 2451401a8d04STejun Heo destroy_work_on_stack(&barr.work); 2452401a8d04STejun Heo return true; 2453401a8d04STejun Heo } else 2454401a8d04STejun Heo return false; 2455401a8d04STejun Heo } 2456401a8d04STejun Heo 2457401a8d04STejun Heo static bool wait_on_work(struct work_struct *work) 2458401a8d04STejun Heo { 2459401a8d04STejun Heo bool ret = false; 2460401a8d04STejun Heo int cpu; 2461401a8d04STejun Heo 2462401a8d04STejun Heo might_sleep(); 2463401a8d04STejun Heo 2464401a8d04STejun Heo lock_map_acquire(&work->lockdep_map); 2465401a8d04STejun Heo lock_map_release(&work->lockdep_map); 2466401a8d04STejun Heo 2467401a8d04STejun Heo for_each_gcwq_cpu(cpu) 2468401a8d04STejun Heo ret |= wait_on_cpu_work(get_gcwq(cpu), work); 2469401a8d04STejun Heo return ret; 2470401a8d04STejun Heo } 2471401a8d04STejun Heo 247209383498STejun Heo /** 247309383498STejun Heo * flush_work_sync - wait until a work has finished execution 247409383498STejun Heo * @work: the work to flush 247509383498STejun Heo * 247609383498STejun Heo * Wait until @work has finished execution. On return, it's 247709383498STejun Heo * guaranteed that all queueing instances of @work which happened 247809383498STejun Heo * before this function is called are finished. In other words, if 247909383498STejun Heo * @work hasn't been requeued since this function was called, @work is 248009383498STejun Heo * guaranteed to be idle on return. 248109383498STejun Heo * 248209383498STejun Heo * RETURNS: 248309383498STejun Heo * %true if flush_work_sync() waited for the work to finish execution, 248409383498STejun Heo * %false if it was already idle. 248509383498STejun Heo */ 248609383498STejun Heo bool flush_work_sync(struct work_struct *work) 248709383498STejun Heo { 248809383498STejun Heo struct wq_barrier barr; 248909383498STejun Heo bool pending, waited; 249009383498STejun Heo 249109383498STejun Heo /* we'll wait for executions separately, queue barr only if pending */ 249209383498STejun Heo pending = start_flush_work(work, &barr, false); 249309383498STejun Heo 249409383498STejun Heo /* wait for executions to finish */ 249509383498STejun Heo waited = wait_on_work(work); 249609383498STejun Heo 249709383498STejun Heo /* wait for the pending one */ 249809383498STejun Heo if (pending) { 249909383498STejun Heo wait_for_completion(&barr.done); 250009383498STejun Heo destroy_work_on_stack(&barr.work); 250109383498STejun Heo } 250209383498STejun Heo 250309383498STejun Heo return pending || waited; 250409383498STejun Heo } 250509383498STejun Heo EXPORT_SYMBOL_GPL(flush_work_sync); 250609383498STejun Heo 25076e84d644SOleg Nesterov /* 25081f1f642eSOleg Nesterov * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit, 25096e84d644SOleg Nesterov * so this work can't be re-armed in any way. 25106e84d644SOleg Nesterov */ 25116e84d644SOleg Nesterov static int try_to_grab_pending(struct work_struct *work) 25126e84d644SOleg Nesterov { 25138b03ae3cSTejun Heo struct global_cwq *gcwq; 25141f1f642eSOleg Nesterov int ret = -1; 25156e84d644SOleg Nesterov 251622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 25171f1f642eSOleg Nesterov return 0; 25186e84d644SOleg Nesterov 25196e84d644SOleg Nesterov /* 25206e84d644SOleg Nesterov * The queueing is in progress, or it is already queued. Try to 25216e84d644SOleg Nesterov * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 25226e84d644SOleg Nesterov */ 25237a22ad75STejun Heo gcwq = get_work_gcwq(work); 25247a22ad75STejun Heo if (!gcwq) 25256e84d644SOleg Nesterov return ret; 25266e84d644SOleg Nesterov 25278b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 25286e84d644SOleg Nesterov if (!list_empty(&work->entry)) { 25296e84d644SOleg Nesterov /* 25307a22ad75STejun Heo * This work is queued, but perhaps we locked the wrong gcwq. 25316e84d644SOleg Nesterov * In that case we must see the new value after rmb(), see 25326e84d644SOleg Nesterov * insert_work()->wmb(). 25336e84d644SOleg Nesterov */ 25346e84d644SOleg Nesterov smp_rmb(); 25357a22ad75STejun Heo if (gcwq == get_work_gcwq(work)) { 2536dc186ad7SThomas Gleixner debug_work_deactivate(work); 25376e84d644SOleg Nesterov list_del_init(&work->entry); 25387a22ad75STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 25398a2e8e5dSTejun Heo get_work_color(work), 25408a2e8e5dSTejun Heo *work_data_bits(work) & WORK_STRUCT_DELAYED); 25416e84d644SOleg Nesterov ret = 1; 25426e84d644SOleg Nesterov } 25436e84d644SOleg Nesterov } 25448b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 25456e84d644SOleg Nesterov 25466e84d644SOleg Nesterov return ret; 25476e84d644SOleg Nesterov } 25486e84d644SOleg Nesterov 2549401a8d04STejun Heo static bool __cancel_work_timer(struct work_struct *work, 25501f1f642eSOleg Nesterov struct timer_list* timer) 25511f1f642eSOleg Nesterov { 25521f1f642eSOleg Nesterov int ret; 25531f1f642eSOleg Nesterov 25541f1f642eSOleg Nesterov do { 25551f1f642eSOleg Nesterov ret = (timer && likely(del_timer(timer))); 25561f1f642eSOleg Nesterov if (!ret) 25571f1f642eSOleg Nesterov ret = try_to_grab_pending(work); 25581f1f642eSOleg Nesterov wait_on_work(work); 25591f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 25601f1f642eSOleg Nesterov 25617a22ad75STejun Heo clear_work_data(work); 25621f1f642eSOleg Nesterov return ret; 25631f1f642eSOleg Nesterov } 25641f1f642eSOleg Nesterov 25656e84d644SOleg Nesterov /** 2566401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2567401a8d04STejun Heo * @work: the work to cancel 25686e84d644SOleg Nesterov * 2569401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2570401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2571401a8d04STejun Heo * another workqueue. On return from this function, @work is 2572401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 25731f1f642eSOleg Nesterov * 2574401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2575401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 25766e84d644SOleg Nesterov * 2577401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 25786e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2579401a8d04STejun Heo * 2580401a8d04STejun Heo * RETURNS: 2581401a8d04STejun Heo * %true if @work was pending, %false otherwise. 25826e84d644SOleg Nesterov */ 2583401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 25846e84d644SOleg Nesterov { 25851f1f642eSOleg Nesterov return __cancel_work_timer(work, NULL); 2586b89deed3SOleg Nesterov } 258728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2588b89deed3SOleg Nesterov 25896e84d644SOleg Nesterov /** 2590401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2591401a8d04STejun Heo * @dwork: the delayed work to flush 25926e84d644SOleg Nesterov * 2593401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2594401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2595401a8d04STejun Heo * considers the last queueing instance of @dwork. 25961f1f642eSOleg Nesterov * 2597401a8d04STejun Heo * RETURNS: 2598401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2599401a8d04STejun Heo * %false if it was already idle. 26006e84d644SOleg Nesterov */ 2601401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2602401a8d04STejun Heo { 2603401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 2604401a8d04STejun Heo __queue_work(raw_smp_processor_id(), 2605401a8d04STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 2606401a8d04STejun Heo return flush_work(&dwork->work); 2607401a8d04STejun Heo } 2608401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2609401a8d04STejun Heo 2610401a8d04STejun Heo /** 261109383498STejun Heo * flush_delayed_work_sync - wait for a dwork to finish 261209383498STejun Heo * @dwork: the delayed work to flush 261309383498STejun Heo * 261409383498STejun Heo * Delayed timer is cancelled and the pending work is queued for 261509383498STejun Heo * execution immediately. Other than timer handling, its behavior 261609383498STejun Heo * is identical to flush_work_sync(). 261709383498STejun Heo * 261809383498STejun Heo * RETURNS: 261909383498STejun Heo * %true if flush_work_sync() waited for the work to finish execution, 262009383498STejun Heo * %false if it was already idle. 262109383498STejun Heo */ 262209383498STejun Heo bool flush_delayed_work_sync(struct delayed_work *dwork) 262309383498STejun Heo { 262409383498STejun Heo if (del_timer_sync(&dwork->timer)) 262509383498STejun Heo __queue_work(raw_smp_processor_id(), 262609383498STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 262709383498STejun Heo return flush_work_sync(&dwork->work); 262809383498STejun Heo } 262909383498STejun Heo EXPORT_SYMBOL(flush_delayed_work_sync); 263009383498STejun Heo 263109383498STejun Heo /** 2632401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2633401a8d04STejun Heo * @dwork: the delayed work cancel 2634401a8d04STejun Heo * 2635401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2636401a8d04STejun Heo * 2637401a8d04STejun Heo * RETURNS: 2638401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2639401a8d04STejun Heo */ 2640401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 26416e84d644SOleg Nesterov { 26421f1f642eSOleg Nesterov return __cancel_work_timer(&dwork->work, &dwork->timer); 26436e84d644SOleg Nesterov } 2644f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 26451da177e4SLinus Torvalds 26460fcb78c2SRolf Eike Beer /** 26470fcb78c2SRolf Eike Beer * schedule_work - put work task in global workqueue 26480fcb78c2SRolf Eike Beer * @work: job to be done 26490fcb78c2SRolf Eike Beer * 26505b0f437dSBart Van Assche * Returns zero if @work was already on the kernel-global workqueue and 26515b0f437dSBart Van Assche * non-zero otherwise. 26525b0f437dSBart Van Assche * 26535b0f437dSBart Van Assche * This puts a job in the kernel-global workqueue if it was not already 26545b0f437dSBart Van Assche * queued and leaves it in the same position on the kernel-global 26555b0f437dSBart Van Assche * workqueue otherwise. 26560fcb78c2SRolf Eike Beer */ 26577ad5b3a5SHarvey Harrison int schedule_work(struct work_struct *work) 26581da177e4SLinus Torvalds { 2659d320c038STejun Heo return queue_work(system_wq, work); 26601da177e4SLinus Torvalds } 2661ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_work); 26621da177e4SLinus Torvalds 2663c1a220e7SZhang Rui /* 2664c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2665c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2666c1a220e7SZhang Rui * @work: job to be done 2667c1a220e7SZhang Rui * 2668c1a220e7SZhang Rui * This puts a job on a specific cpu 2669c1a220e7SZhang Rui */ 2670c1a220e7SZhang Rui int schedule_work_on(int cpu, struct work_struct *work) 2671c1a220e7SZhang Rui { 2672d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2673c1a220e7SZhang Rui } 2674c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2675c1a220e7SZhang Rui 26760fcb78c2SRolf Eike Beer /** 26770fcb78c2SRolf Eike Beer * schedule_delayed_work - put work task in global workqueue after delay 267852bad64dSDavid Howells * @dwork: job to be done 267952bad64dSDavid Howells * @delay: number of jiffies to wait or 0 for immediate execution 26800fcb78c2SRolf Eike Beer * 26810fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 26820fcb78c2SRolf Eike Beer * workqueue. 26830fcb78c2SRolf Eike Beer */ 26847ad5b3a5SHarvey Harrison int schedule_delayed_work(struct delayed_work *dwork, 268582f67cd9SIngo Molnar unsigned long delay) 26861da177e4SLinus Torvalds { 2687d320c038STejun Heo return queue_delayed_work(system_wq, dwork, delay); 26881da177e4SLinus Torvalds } 2689ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work); 26901da177e4SLinus Torvalds 26910fcb78c2SRolf Eike Beer /** 26920fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 26930fcb78c2SRolf Eike Beer * @cpu: cpu to use 269452bad64dSDavid Howells * @dwork: job to be done 26950fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 26960fcb78c2SRolf Eike Beer * 26970fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 26980fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 26990fcb78c2SRolf Eike Beer */ 27001da177e4SLinus Torvalds int schedule_delayed_work_on(int cpu, 270152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 27021da177e4SLinus Torvalds { 2703d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 27041da177e4SLinus Torvalds } 2705ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 27061da177e4SLinus Torvalds 2707b6136773SAndrew Morton /** 270831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2709b6136773SAndrew Morton * @func: the function to call 2710b6136773SAndrew Morton * 271131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 271231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2713b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 271431ddd871STejun Heo * 271531ddd871STejun Heo * RETURNS: 271631ddd871STejun Heo * 0 on success, -errno on failure. 2717b6136773SAndrew Morton */ 271865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 271915316ba8SChristoph Lameter { 272015316ba8SChristoph Lameter int cpu; 272138f51568SNamhyung Kim struct work_struct __percpu *works; 272215316ba8SChristoph Lameter 2723b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2724b6136773SAndrew Morton if (!works) 272515316ba8SChristoph Lameter return -ENOMEM; 2726b6136773SAndrew Morton 272795402b38SGautham R Shenoy get_online_cpus(); 272893981800STejun Heo 272915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 27309bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 27319bfb1839SIngo Molnar 27329bfb1839SIngo Molnar INIT_WORK(work, func); 27338de6d308SOleg Nesterov schedule_work_on(cpu, work); 273415316ba8SChristoph Lameter } 273593981800STejun Heo 273693981800STejun Heo for_each_online_cpu(cpu) 27378616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 273893981800STejun Heo 273995402b38SGautham R Shenoy put_online_cpus(); 2740b6136773SAndrew Morton free_percpu(works); 274115316ba8SChristoph Lameter return 0; 274215316ba8SChristoph Lameter } 274315316ba8SChristoph Lameter 2744eef6a7d5SAlan Stern /** 2745eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2746eef6a7d5SAlan Stern * 2747eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2748eef6a7d5SAlan Stern * completion. 2749eef6a7d5SAlan Stern * 2750eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2751eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2752eef6a7d5SAlan Stern * will lead to deadlock: 2753eef6a7d5SAlan Stern * 2754eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2755eef6a7d5SAlan Stern * a lock held by your code or its caller. 2756eef6a7d5SAlan Stern * 2757eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2758eef6a7d5SAlan Stern * 2759eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2760eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2761eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2762eef6a7d5SAlan Stern * 2763eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2764eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2765eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2766eef6a7d5SAlan Stern * cancel_work_sync() instead. 2767eef6a7d5SAlan Stern */ 27681da177e4SLinus Torvalds void flush_scheduled_work(void) 27691da177e4SLinus Torvalds { 2770d320c038STejun Heo flush_workqueue(system_wq); 27711da177e4SLinus Torvalds } 2772ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 27731da177e4SLinus Torvalds 27741da177e4SLinus Torvalds /** 27751fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 27761fa44ecaSJames Bottomley * @fn: the function to execute 27771fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 27781fa44ecaSJames Bottomley * be available when the work executes) 27791fa44ecaSJames Bottomley * 27801fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 27811fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 27821fa44ecaSJames Bottomley * 27831fa44ecaSJames Bottomley * Returns: 0 - function was executed 27841fa44ecaSJames Bottomley * 1 - function was scheduled for execution 27851fa44ecaSJames Bottomley */ 278665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 27871fa44ecaSJames Bottomley { 27881fa44ecaSJames Bottomley if (!in_interrupt()) { 278965f27f38SDavid Howells fn(&ew->work); 27901fa44ecaSJames Bottomley return 0; 27911fa44ecaSJames Bottomley } 27921fa44ecaSJames Bottomley 279365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 27941fa44ecaSJames Bottomley schedule_work(&ew->work); 27951fa44ecaSJames Bottomley 27961fa44ecaSJames Bottomley return 1; 27971fa44ecaSJames Bottomley } 27981fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 27991fa44ecaSJames Bottomley 28001da177e4SLinus Torvalds int keventd_up(void) 28011da177e4SLinus Torvalds { 2802d320c038STejun Heo return system_wq != NULL; 28031da177e4SLinus Torvalds } 28041da177e4SLinus Torvalds 2805bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 28061da177e4SLinus Torvalds { 28073af24433SOleg Nesterov /* 28080f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 28090f900049STejun Heo * Make sure that the alignment isn't lower than that of 28100f900049STejun Heo * unsigned long long. 28113af24433SOleg Nesterov */ 28120f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 28130f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 28140f900049STejun Heo __alignof__(unsigned long long)); 2815931ac77eSTejun Heo #ifdef CONFIG_SMP 2816931ac77eSTejun Heo bool percpu = !(wq->flags & WQ_UNBOUND); 2817931ac77eSTejun Heo #else 2818931ac77eSTejun Heo bool percpu = false; 2819931ac77eSTejun Heo #endif 28203af24433SOleg Nesterov 2821931ac77eSTejun Heo if (percpu) 2822f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 2823931ac77eSTejun Heo else { 28240f900049STejun Heo void *ptr; 2825e1d8aa9fSFrederic Weisbecker 28260f900049STejun Heo /* 2827f3421797STejun Heo * Allocate enough room to align cwq and put an extra 2828f3421797STejun Heo * pointer at the end pointing back to the originally 2829f3421797STejun Heo * allocated pointer which will be used for free. 28300f900049STejun Heo */ 2831bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 2832bdbc5dd7STejun Heo if (ptr) { 2833bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 2834bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 2835bdbc5dd7STejun Heo } 28363af24433SOleg Nesterov } 28373af24433SOleg Nesterov 283852605627SDavid Howells /* just in case, make sure it's actually aligned 283952605627SDavid Howells * - this is affected by PERCPU() alignment in vmlinux.lds.S 284052605627SDavid Howells */ 2841bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 2842bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 28430f900049STejun Heo } 28440f900049STejun Heo 2845bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 284606ba38a9SOleg Nesterov { 2847931ac77eSTejun Heo #ifdef CONFIG_SMP 2848931ac77eSTejun Heo bool percpu = !(wq->flags & WQ_UNBOUND); 2849931ac77eSTejun Heo #else 2850931ac77eSTejun Heo bool percpu = false; 2851931ac77eSTejun Heo #endif 285206ba38a9SOleg Nesterov 2853931ac77eSTejun Heo if (percpu) 2854bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 2855f3421797STejun Heo else if (wq->cpu_wq.single) { 2856f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 2857f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 285806ba38a9SOleg Nesterov } 285906ba38a9SOleg Nesterov } 286006ba38a9SOleg Nesterov 2861f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 2862f3421797STejun Heo const char *name) 2863b71ab8c2STejun Heo { 2864f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 2865f3421797STejun Heo 2866f3421797STejun Heo if (max_active < 1 || max_active > lim) 2867b71ab8c2STejun Heo printk(KERN_WARNING "workqueue: max_active %d requested for %s " 2868b71ab8c2STejun Heo "is out of range, clamping between %d and %d\n", 2869f3421797STejun Heo max_active, name, 1, lim); 2870b71ab8c2STejun Heo 2871f3421797STejun Heo return clamp_val(max_active, 1, lim); 2872b71ab8c2STejun Heo } 2873b71ab8c2STejun Heo 2874d320c038STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *name, 287597e37d7bSTejun Heo unsigned int flags, 28761e19ffc6STejun Heo int max_active, 2877eb13ba87SJohannes Berg struct lock_class_key *key, 2878eb13ba87SJohannes Berg const char *lock_name) 28793af24433SOleg Nesterov { 28803af24433SOleg Nesterov struct workqueue_struct *wq; 2881c34056a3STejun Heo unsigned int cpu; 28823af24433SOleg Nesterov 2883f3421797STejun Heo /* 28846370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 28856370a6adSTejun Heo * have a rescuer to guarantee forward progress. 28866370a6adSTejun Heo */ 28876370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 28886370a6adSTejun Heo flags |= WQ_RESCUER; 28896370a6adSTejun Heo 28906370a6adSTejun Heo /* 2891f3421797STejun Heo * Unbound workqueues aren't concurrency managed and should be 2892f3421797STejun Heo * dispatched to workers immediately. 2893f3421797STejun Heo */ 2894f3421797STejun Heo if (flags & WQ_UNBOUND) 2895f3421797STejun Heo flags |= WQ_HIGHPRI; 2896f3421797STejun Heo 2897d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 2898f3421797STejun Heo max_active = wq_clamp_max_active(max_active, flags, name); 28993af24433SOleg Nesterov 29003af24433SOleg Nesterov wq = kzalloc(sizeof(*wq), GFP_KERNEL); 29013af24433SOleg Nesterov if (!wq) 29024690c4abSTejun Heo goto err; 29033af24433SOleg Nesterov 290497e37d7bSTejun Heo wq->flags = flags; 2905a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 290673f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 290773f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 290873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 290973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 29103af24433SOleg Nesterov 29113af24433SOleg Nesterov wq->name = name; 2912eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 2913cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 29143af24433SOleg Nesterov 2915bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 2916bdbc5dd7STejun Heo goto err; 2917bdbc5dd7STejun Heo 2918f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 29191537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 29208b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 29211537663fSTejun Heo 29220f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 29238b03ae3cSTejun Heo cwq->gcwq = gcwq; 2924c34056a3STejun Heo cwq->wq = wq; 292573f53c4aSTejun Heo cwq->flush_color = -1; 29261e19ffc6STejun Heo cwq->max_active = max_active; 29271e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 2928e22bee78STejun Heo } 29291537663fSTejun Heo 2930e22bee78STejun Heo if (flags & WQ_RESCUER) { 2931e22bee78STejun Heo struct worker *rescuer; 2932e22bee78STejun Heo 2933f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 2934e22bee78STejun Heo goto err; 2935e22bee78STejun Heo 2936e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 2937e22bee78STejun Heo if (!rescuer) 2938e22bee78STejun Heo goto err; 2939e22bee78STejun Heo 2940e22bee78STejun Heo rescuer->task = kthread_create(rescuer_thread, wq, "%s", name); 2941e22bee78STejun Heo if (IS_ERR(rescuer->task)) 2942e22bee78STejun Heo goto err; 2943e22bee78STejun Heo 2944e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 2945e22bee78STejun Heo wake_up_process(rescuer->task); 29463af24433SOleg Nesterov } 29471537663fSTejun Heo 29483af24433SOleg Nesterov /* 2949a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 2950a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 2951a0a1a5fdSTejun Heo * workqueue to workqueues list. 29523af24433SOleg Nesterov */ 29533af24433SOleg Nesterov spin_lock(&workqueue_lock); 2954a0a1a5fdSTejun Heo 2955a0a1a5fdSTejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZEABLE) 2956f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 2957a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 2958a0a1a5fdSTejun Heo 29593af24433SOleg Nesterov list_add(&wq->list, &workqueues); 2960a0a1a5fdSTejun Heo 29613af24433SOleg Nesterov spin_unlock(&workqueue_lock); 29623af24433SOleg Nesterov 29633af24433SOleg Nesterov return wq; 29644690c4abSTejun Heo err: 29654690c4abSTejun Heo if (wq) { 2966bdbc5dd7STejun Heo free_cwqs(wq); 2967f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 2968e22bee78STejun Heo kfree(wq->rescuer); 29694690c4abSTejun Heo kfree(wq); 29703af24433SOleg Nesterov } 29714690c4abSTejun Heo return NULL; 29721da177e4SLinus Torvalds } 2973d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 29741da177e4SLinus Torvalds 29753af24433SOleg Nesterov /** 29763af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 29773af24433SOleg Nesterov * @wq: target workqueue 29783af24433SOleg Nesterov * 29793af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 29803af24433SOleg Nesterov */ 29813af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 29823af24433SOleg Nesterov { 2983c8efcc25STejun Heo unsigned int flush_cnt = 0; 2984c8e55f36STejun Heo unsigned int cpu; 29853af24433SOleg Nesterov 2986c8efcc25STejun Heo /* 2987c8efcc25STejun Heo * Mark @wq dying and drain all pending works. Once WQ_DYING is 2988c8efcc25STejun Heo * set, only chain queueing is allowed. IOW, only currently 2989c8efcc25STejun Heo * pending or running work items on @wq can queue further work 2990c8efcc25STejun Heo * items on it. @wq is flushed repeatedly until it becomes empty. 2991c8efcc25STejun Heo * The number of flushing is detemined by the depth of chaining and 2992c8efcc25STejun Heo * should be relatively short. Whine if it takes too long. 2993c8efcc25STejun Heo */ 2994e41e704bSTejun Heo wq->flags |= WQ_DYING; 2995c8efcc25STejun Heo reflush: 2996a0a1a5fdSTejun Heo flush_workqueue(wq); 2997a0a1a5fdSTejun Heo 2998c8efcc25STejun Heo for_each_cwq_cpu(cpu, wq) { 2999c8efcc25STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3000c8efcc25STejun Heo 3001c8efcc25STejun Heo if (!cwq->nr_active && list_empty(&cwq->delayed_works)) 3002c8efcc25STejun Heo continue; 3003c8efcc25STejun Heo 3004c8efcc25STejun Heo if (++flush_cnt == 10 || 3005c8efcc25STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3006c8efcc25STejun Heo printk(KERN_WARNING "workqueue %s: flush on " 3007c8efcc25STejun Heo "destruction isn't complete after %u tries\n", 3008c8efcc25STejun Heo wq->name, flush_cnt); 3009c8efcc25STejun Heo goto reflush; 3010c8efcc25STejun Heo } 3011c8efcc25STejun Heo 3012a0a1a5fdSTejun Heo /* 3013a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3014a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3015a0a1a5fdSTejun Heo */ 301695402b38SGautham R Shenoy spin_lock(&workqueue_lock); 30173af24433SOleg Nesterov list_del(&wq->list); 301895402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 30193af24433SOleg Nesterov 3020e22bee78STejun Heo /* sanity check */ 3021f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 302273f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 302373f53c4aSTejun Heo int i; 30243af24433SOleg Nesterov 302573f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 302673f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 30271e19ffc6STejun Heo BUG_ON(cwq->nr_active); 30281e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 302973f53c4aSTejun Heo } 30301537663fSTejun Heo 3031e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3032e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3033f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 30348d9df9f0SXiaotian Feng kfree(wq->rescuer); 3035e22bee78STejun Heo } 3036e22bee78STejun Heo 3037bdbc5dd7STejun Heo free_cwqs(wq); 30383af24433SOleg Nesterov kfree(wq); 30393af24433SOleg Nesterov } 30403af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 30413af24433SOleg Nesterov 3042dcd989cbSTejun Heo /** 3043dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3044dcd989cbSTejun Heo * @wq: target workqueue 3045dcd989cbSTejun Heo * @max_active: new max_active value. 3046dcd989cbSTejun Heo * 3047dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3048dcd989cbSTejun Heo * 3049dcd989cbSTejun Heo * CONTEXT: 3050dcd989cbSTejun Heo * Don't call from IRQ context. 3051dcd989cbSTejun Heo */ 3052dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3053dcd989cbSTejun Heo { 3054dcd989cbSTejun Heo unsigned int cpu; 3055dcd989cbSTejun Heo 3056f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3057dcd989cbSTejun Heo 3058dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3059dcd989cbSTejun Heo 3060dcd989cbSTejun Heo wq->saved_max_active = max_active; 3061dcd989cbSTejun Heo 3062f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 3063dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3064dcd989cbSTejun Heo 3065dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 3066dcd989cbSTejun Heo 3067dcd989cbSTejun Heo if (!(wq->flags & WQ_FREEZEABLE) || 3068dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 3069dcd989cbSTejun Heo get_cwq(gcwq->cpu, wq)->max_active = max_active; 3070dcd989cbSTejun Heo 3071dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 3072dcd989cbSTejun Heo } 3073dcd989cbSTejun Heo 3074dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3075dcd989cbSTejun Heo } 3076dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3077dcd989cbSTejun Heo 3078dcd989cbSTejun Heo /** 3079dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3080dcd989cbSTejun Heo * @cpu: CPU in question 3081dcd989cbSTejun Heo * @wq: target workqueue 3082dcd989cbSTejun Heo * 3083dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3084dcd989cbSTejun Heo * no synchronization around this function and the test result is 3085dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3086dcd989cbSTejun Heo * 3087dcd989cbSTejun Heo * RETURNS: 3088dcd989cbSTejun Heo * %true if congested, %false otherwise. 3089dcd989cbSTejun Heo */ 3090dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3091dcd989cbSTejun Heo { 3092dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3093dcd989cbSTejun Heo 3094dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3095dcd989cbSTejun Heo } 3096dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3097dcd989cbSTejun Heo 3098dcd989cbSTejun Heo /** 3099dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 3100dcd989cbSTejun Heo * @work: the work of interest 3101dcd989cbSTejun Heo * 3102dcd989cbSTejun Heo * RETURNS: 3103bdbc5dd7STejun Heo * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. 3104dcd989cbSTejun Heo */ 3105dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 3106dcd989cbSTejun Heo { 3107dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3108dcd989cbSTejun Heo 3109bdbc5dd7STejun Heo return gcwq ? gcwq->cpu : WORK_CPU_NONE; 3110dcd989cbSTejun Heo } 3111dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 3112dcd989cbSTejun Heo 3113dcd989cbSTejun Heo /** 3114dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3115dcd989cbSTejun Heo * @work: the work to be tested 3116dcd989cbSTejun Heo * 3117dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3118dcd989cbSTejun Heo * synchronization around this function and the test result is 3119dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3120dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 3121dcd989cbSTejun Heo * running state. 3122dcd989cbSTejun Heo * 3123dcd989cbSTejun Heo * RETURNS: 3124dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3125dcd989cbSTejun Heo */ 3126dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3127dcd989cbSTejun Heo { 3128dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3129dcd989cbSTejun Heo unsigned long flags; 3130dcd989cbSTejun Heo unsigned int ret = 0; 3131dcd989cbSTejun Heo 3132dcd989cbSTejun Heo if (!gcwq) 3133dcd989cbSTejun Heo return false; 3134dcd989cbSTejun Heo 3135dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 3136dcd989cbSTejun Heo 3137dcd989cbSTejun Heo if (work_pending(work)) 3138dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3139dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 3140dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3141dcd989cbSTejun Heo 3142dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 3143dcd989cbSTejun Heo 3144dcd989cbSTejun Heo return ret; 3145dcd989cbSTejun Heo } 3146dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3147dcd989cbSTejun Heo 3148db7bccf4STejun Heo /* 3149db7bccf4STejun Heo * CPU hotplug. 3150db7bccf4STejun Heo * 3151e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3152e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3153e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 3154e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 3155e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 3156e22bee78STejun Heo * blocked draining impractical. 3157e22bee78STejun Heo * 3158e22bee78STejun Heo * This is solved by allowing a gcwq to be detached from CPU, running 3159e22bee78STejun Heo * it with unbound (rogue) workers and allowing it to be reattached 3160e22bee78STejun Heo * later if the cpu comes back online. A separate thread is created 3161e22bee78STejun Heo * to govern a gcwq in such state and is called the trustee of the 3162e22bee78STejun Heo * gcwq. 3163db7bccf4STejun Heo * 3164db7bccf4STejun Heo * Trustee states and their descriptions. 3165db7bccf4STejun Heo * 3166db7bccf4STejun Heo * START Command state used on startup. On CPU_DOWN_PREPARE, a 3167db7bccf4STejun Heo * new trustee is started with this state. 3168db7bccf4STejun Heo * 3169db7bccf4STejun Heo * IN_CHARGE Once started, trustee will enter this state after 3170e22bee78STejun Heo * assuming the manager role and making all existing 3171e22bee78STejun Heo * workers rogue. DOWN_PREPARE waits for trustee to 3172e22bee78STejun Heo * enter this state. After reaching IN_CHARGE, trustee 3173e22bee78STejun Heo * tries to execute the pending worklist until it's empty 3174e22bee78STejun Heo * and the state is set to BUTCHER, or the state is set 3175e22bee78STejun Heo * to RELEASE. 3176db7bccf4STejun Heo * 3177db7bccf4STejun Heo * BUTCHER Command state which is set by the cpu callback after 3178db7bccf4STejun Heo * the cpu has went down. Once this state is set trustee 3179db7bccf4STejun Heo * knows that there will be no new works on the worklist 3180db7bccf4STejun Heo * and once the worklist is empty it can proceed to 3181db7bccf4STejun Heo * killing idle workers. 3182db7bccf4STejun Heo * 3183db7bccf4STejun Heo * RELEASE Command state which is set by the cpu callback if the 3184db7bccf4STejun Heo * cpu down has been canceled or it has come online 3185db7bccf4STejun Heo * again. After recognizing this state, trustee stops 3186e22bee78STejun Heo * trying to drain or butcher and clears ROGUE, rebinds 3187e22bee78STejun Heo * all remaining workers back to the cpu and releases 3188e22bee78STejun Heo * manager role. 3189db7bccf4STejun Heo * 3190db7bccf4STejun Heo * DONE Trustee will enter this state after BUTCHER or RELEASE 3191db7bccf4STejun Heo * is complete. 3192db7bccf4STejun Heo * 3193db7bccf4STejun Heo * trustee CPU draining 3194db7bccf4STejun Heo * took over down complete 3195db7bccf4STejun Heo * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE 3196db7bccf4STejun Heo * | | ^ 3197db7bccf4STejun Heo * | CPU is back online v return workers | 3198db7bccf4STejun Heo * ----------------> RELEASE -------------- 3199db7bccf4STejun Heo */ 3200db7bccf4STejun Heo 3201db7bccf4STejun Heo /** 3202db7bccf4STejun Heo * trustee_wait_event_timeout - timed event wait for trustee 3203db7bccf4STejun Heo * @cond: condition to wait for 3204db7bccf4STejun Heo * @timeout: timeout in jiffies 3205db7bccf4STejun Heo * 3206db7bccf4STejun Heo * wait_event_timeout() for trustee to use. Handles locking and 3207db7bccf4STejun Heo * checks for RELEASE request. 3208db7bccf4STejun Heo * 3209db7bccf4STejun Heo * CONTEXT: 3210db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3211db7bccf4STejun Heo * multiple times. To be used by trustee. 3212db7bccf4STejun Heo * 3213db7bccf4STejun Heo * RETURNS: 3214db7bccf4STejun Heo * Positive indicating left time if @cond is satisfied, 0 if timed 3215db7bccf4STejun Heo * out, -1 if canceled. 3216db7bccf4STejun Heo */ 3217db7bccf4STejun Heo #define trustee_wait_event_timeout(cond, timeout) ({ \ 3218db7bccf4STejun Heo long __ret = (timeout); \ 3219db7bccf4STejun Heo while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \ 3220db7bccf4STejun Heo __ret) { \ 3221db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); \ 3222db7bccf4STejun Heo __wait_event_timeout(gcwq->trustee_wait, (cond) || \ 3223db7bccf4STejun Heo (gcwq->trustee_state == TRUSTEE_RELEASE), \ 3224db7bccf4STejun Heo __ret); \ 3225db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); \ 3226db7bccf4STejun Heo } \ 3227db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \ 3228db7bccf4STejun Heo }) 3229db7bccf4STejun Heo 3230db7bccf4STejun Heo /** 3231db7bccf4STejun Heo * trustee_wait_event - event wait for trustee 3232db7bccf4STejun Heo * @cond: condition to wait for 3233db7bccf4STejun Heo * 3234db7bccf4STejun Heo * wait_event() for trustee to use. Automatically handles locking and 3235db7bccf4STejun Heo * checks for CANCEL request. 3236db7bccf4STejun Heo * 3237db7bccf4STejun Heo * CONTEXT: 3238db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3239db7bccf4STejun Heo * multiple times. To be used by trustee. 3240db7bccf4STejun Heo * 3241db7bccf4STejun Heo * RETURNS: 3242db7bccf4STejun Heo * 0 if @cond is satisfied, -1 if canceled. 3243db7bccf4STejun Heo */ 3244db7bccf4STejun Heo #define trustee_wait_event(cond) ({ \ 3245db7bccf4STejun Heo long __ret1; \ 3246db7bccf4STejun Heo __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\ 3247db7bccf4STejun Heo __ret1 < 0 ? -1 : 0; \ 3248db7bccf4STejun Heo }) 3249db7bccf4STejun Heo 3250db7bccf4STejun Heo static int __cpuinit trustee_thread(void *__gcwq) 3251db7bccf4STejun Heo { 3252db7bccf4STejun Heo struct global_cwq *gcwq = __gcwq; 3253db7bccf4STejun Heo struct worker *worker; 3254e22bee78STejun Heo struct work_struct *work; 3255db7bccf4STejun Heo struct hlist_node *pos; 3256e22bee78STejun Heo long rc; 3257db7bccf4STejun Heo int i; 3258db7bccf4STejun Heo 3259db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3260db7bccf4STejun Heo 3261db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3262db7bccf4STejun Heo /* 3263e22bee78STejun Heo * Claim the manager position and make all workers rogue. 3264e22bee78STejun Heo * Trustee must be bound to the target cpu and can't be 3265e22bee78STejun Heo * cancelled. 3266db7bccf4STejun Heo */ 3267db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3268e22bee78STejun Heo rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS)); 3269e22bee78STejun Heo BUG_ON(rc < 0); 3270e22bee78STejun Heo 3271e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGING_WORKERS; 3272db7bccf4STejun Heo 3273db7bccf4STejun Heo list_for_each_entry(worker, &gcwq->idle_list, entry) 3274cb444766STejun Heo worker->flags |= WORKER_ROGUE; 3275db7bccf4STejun Heo 3276db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 3277cb444766STejun Heo worker->flags |= WORKER_ROGUE; 3278db7bccf4STejun Heo 3279db7bccf4STejun Heo /* 3280e22bee78STejun Heo * Call schedule() so that we cross rq->lock and thus can 3281e22bee78STejun Heo * guarantee sched callbacks see the rogue flag. This is 3282e22bee78STejun Heo * necessary as scheduler callbacks may be invoked from other 3283e22bee78STejun Heo * cpus. 3284e22bee78STejun Heo */ 3285e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3286e22bee78STejun Heo schedule(); 3287e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3288e22bee78STejun Heo 3289e22bee78STejun Heo /* 3290cb444766STejun Heo * Sched callbacks are disabled now. Zap nr_running. After 3291cb444766STejun Heo * this, nr_running stays zero and need_more_worker() and 3292cb444766STejun Heo * keep_working() are always true as long as the worklist is 3293cb444766STejun Heo * not empty. 3294e22bee78STejun Heo */ 3295cb444766STejun Heo atomic_set(get_gcwq_nr_running(gcwq->cpu), 0); 3296e22bee78STejun Heo 3297e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3298e22bee78STejun Heo del_timer_sync(&gcwq->idle_timer); 3299e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3300e22bee78STejun Heo 3301e22bee78STejun Heo /* 3302db7bccf4STejun Heo * We're now in charge. Notify and proceed to drain. We need 3303db7bccf4STejun Heo * to keep the gcwq running during the whole CPU down 3304db7bccf4STejun Heo * procedure as other cpu hotunplug callbacks may need to 3305db7bccf4STejun Heo * flush currently running tasks. 3306db7bccf4STejun Heo */ 3307db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_IN_CHARGE; 3308db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3309db7bccf4STejun Heo 3310db7bccf4STejun Heo /* 3311db7bccf4STejun Heo * The original cpu is in the process of dying and may go away 3312db7bccf4STejun Heo * anytime now. When that happens, we and all workers would 3313e22bee78STejun Heo * be migrated to other cpus. Try draining any left work. We 3314e22bee78STejun Heo * want to get it over with ASAP - spam rescuers, wake up as 3315e22bee78STejun Heo * many idlers as necessary and create new ones till the 3316e22bee78STejun Heo * worklist is empty. Note that if the gcwq is frozen, there 3317e22bee78STejun Heo * may be frozen works in freezeable cwqs. Don't declare 3318e22bee78STejun Heo * completion while frozen. 3319db7bccf4STejun Heo */ 3320db7bccf4STejun Heo while (gcwq->nr_workers != gcwq->nr_idle || 3321db7bccf4STejun Heo gcwq->flags & GCWQ_FREEZING || 3322db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_IN_CHARGE) { 3323e22bee78STejun Heo int nr_works = 0; 3324e22bee78STejun Heo 3325e22bee78STejun Heo list_for_each_entry(work, &gcwq->worklist, entry) { 3326e22bee78STejun Heo send_mayday(work); 3327e22bee78STejun Heo nr_works++; 3328e22bee78STejun Heo } 3329e22bee78STejun Heo 3330e22bee78STejun Heo list_for_each_entry(worker, &gcwq->idle_list, entry) { 3331e22bee78STejun Heo if (!nr_works--) 3332e22bee78STejun Heo break; 3333e22bee78STejun Heo wake_up_process(worker->task); 3334e22bee78STejun Heo } 3335e22bee78STejun Heo 3336e22bee78STejun Heo if (need_to_create_worker(gcwq)) { 3337e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3338e22bee78STejun Heo worker = create_worker(gcwq, false); 3339e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3340e22bee78STejun Heo if (worker) { 3341cb444766STejun Heo worker->flags |= WORKER_ROGUE; 3342e22bee78STejun Heo start_worker(worker); 3343e22bee78STejun Heo } 3344e22bee78STejun Heo } 3345e22bee78STejun Heo 3346db7bccf4STejun Heo /* give a breather */ 3347db7bccf4STejun Heo if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0) 3348db7bccf4STejun Heo break; 3349db7bccf4STejun Heo } 3350db7bccf4STejun Heo 3351e22bee78STejun Heo /* 3352e22bee78STejun Heo * Either all works have been scheduled and cpu is down, or 3353e22bee78STejun Heo * cpu down has already been canceled. Wait for and butcher 3354e22bee78STejun Heo * all workers till we're canceled. 3355e22bee78STejun Heo */ 3356e22bee78STejun Heo do { 3357e22bee78STejun Heo rc = trustee_wait_event(!list_empty(&gcwq->idle_list)); 3358e22bee78STejun Heo while (!list_empty(&gcwq->idle_list)) 3359e22bee78STejun Heo destroy_worker(list_first_entry(&gcwq->idle_list, 3360e22bee78STejun Heo struct worker, entry)); 3361e22bee78STejun Heo } while (gcwq->nr_workers && rc >= 0); 3362e22bee78STejun Heo 3363e22bee78STejun Heo /* 3364e22bee78STejun Heo * At this point, either draining has completed and no worker 3365e22bee78STejun Heo * is left, or cpu down has been canceled or the cpu is being 3366e22bee78STejun Heo * brought back up. There shouldn't be any idle one left. 3367e22bee78STejun Heo * Tell the remaining busy ones to rebind once it finishes the 3368e22bee78STejun Heo * currently scheduled works by scheduling the rebind_work. 3369e22bee78STejun Heo */ 3370e22bee78STejun Heo WARN_ON(!list_empty(&gcwq->idle_list)); 3371e22bee78STejun Heo 3372e22bee78STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 3373e22bee78STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 3374e22bee78STejun Heo 3375e22bee78STejun Heo /* 3376e22bee78STejun Heo * Rebind_work may race with future cpu hotplug 3377e22bee78STejun Heo * operations. Use a separate flag to mark that 3378e22bee78STejun Heo * rebinding is scheduled. 3379e22bee78STejun Heo */ 3380cb444766STejun Heo worker->flags |= WORKER_REBIND; 3381cb444766STejun Heo worker->flags &= ~WORKER_ROGUE; 3382e22bee78STejun Heo 3383e22bee78STejun Heo /* queue rebind_work, wq doesn't matter, use the default one */ 3384e22bee78STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 3385e22bee78STejun Heo work_data_bits(rebind_work))) 3386e22bee78STejun Heo continue; 3387e22bee78STejun Heo 3388e22bee78STejun Heo debug_work_activate(rebind_work); 3389d320c038STejun Heo insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work, 3390e22bee78STejun Heo worker->scheduled.next, 3391e22bee78STejun Heo work_color_to_flags(WORK_NO_COLOR)); 3392e22bee78STejun Heo } 3393e22bee78STejun Heo 3394e22bee78STejun Heo /* relinquish manager role */ 3395e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGING_WORKERS; 3396e22bee78STejun Heo 3397db7bccf4STejun Heo /* notify completion */ 3398db7bccf4STejun Heo gcwq->trustee = NULL; 3399db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3400db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3401db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3402db7bccf4STejun Heo return 0; 3403db7bccf4STejun Heo } 3404db7bccf4STejun Heo 3405db7bccf4STejun Heo /** 3406db7bccf4STejun Heo * wait_trustee_state - wait for trustee to enter the specified state 3407db7bccf4STejun Heo * @gcwq: gcwq the trustee of interest belongs to 3408db7bccf4STejun Heo * @state: target state to wait for 3409db7bccf4STejun Heo * 3410db7bccf4STejun Heo * Wait for the trustee to reach @state. DONE is already matched. 3411db7bccf4STejun Heo * 3412db7bccf4STejun Heo * CONTEXT: 3413db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3414db7bccf4STejun Heo * multiple times. To be used by cpu_callback. 3415db7bccf4STejun Heo */ 3416db7bccf4STejun Heo static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state) 341706bd6ebfSNamhyung Kim __releases(&gcwq->lock) 341806bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 3419db7bccf4STejun Heo { 3420db7bccf4STejun Heo if (!(gcwq->trustee_state == state || 3421db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE)) { 3422db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3423db7bccf4STejun Heo __wait_event(gcwq->trustee_wait, 3424db7bccf4STejun Heo gcwq->trustee_state == state || 3425db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE); 3426db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3427db7bccf4STejun Heo } 3428db7bccf4STejun Heo } 3429db7bccf4STejun Heo 34309c7b216dSChandra Seetharaman static int __devinit workqueue_cpu_callback(struct notifier_block *nfb, 34311da177e4SLinus Torvalds unsigned long action, 34321da177e4SLinus Torvalds void *hcpu) 34331da177e4SLinus Torvalds { 34343af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3435db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3436db7bccf4STejun Heo struct task_struct *new_trustee = NULL; 3437e22bee78STejun Heo struct worker *uninitialized_var(new_worker); 3438db7bccf4STejun Heo unsigned long flags; 34391da177e4SLinus Torvalds 34408bb78442SRafael J. Wysocki action &= ~CPU_TASKS_FROZEN; 34418bb78442SRafael J. Wysocki 34421da177e4SLinus Torvalds switch (action) { 3443db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3444db7bccf4STejun Heo new_trustee = kthread_create(trustee_thread, gcwq, 3445db7bccf4STejun Heo "workqueue_trustee/%d\n", cpu); 3446db7bccf4STejun Heo if (IS_ERR(new_trustee)) 3447db7bccf4STejun Heo return notifier_from_errno(PTR_ERR(new_trustee)); 3448db7bccf4STejun Heo kthread_bind(new_trustee, cpu); 3449e22bee78STejun Heo /* fall through */ 34503af24433SOleg Nesterov case CPU_UP_PREPARE: 3451e22bee78STejun Heo BUG_ON(gcwq->first_idle); 3452e22bee78STejun Heo new_worker = create_worker(gcwq, false); 3453e22bee78STejun Heo if (!new_worker) { 3454e22bee78STejun Heo if (new_trustee) 3455e22bee78STejun Heo kthread_stop(new_trustee); 3456e22bee78STejun Heo return NOTIFY_BAD; 34573af24433SOleg Nesterov } 3458db7bccf4STejun Heo } 34591537663fSTejun Heo 3460db7bccf4STejun Heo /* some are called w/ irq disabled, don't disturb irq status */ 3461db7bccf4STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 34623af24433SOleg Nesterov 34633af24433SOleg Nesterov switch (action) { 3464db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3465db7bccf4STejun Heo /* initialize trustee and tell it to acquire the gcwq */ 3466db7bccf4STejun Heo BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE); 3467db7bccf4STejun Heo gcwq->trustee = new_trustee; 3468db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_START; 3469db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3470db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE); 3471e22bee78STejun Heo /* fall through */ 34723af24433SOleg Nesterov case CPU_UP_PREPARE: 3473e22bee78STejun Heo BUG_ON(gcwq->first_idle); 3474e22bee78STejun Heo gcwq->first_idle = new_worker; 34751da177e4SLinus Torvalds break; 34761da177e4SLinus Torvalds 3477e22bee78STejun Heo case CPU_DYING: 3478e22bee78STejun Heo /* 3479e22bee78STejun Heo * Before this, the trustee and all workers except for 3480e22bee78STejun Heo * the ones which are still executing works from 3481e22bee78STejun Heo * before the last CPU down must be on the cpu. After 3482e22bee78STejun Heo * this, they'll all be diasporas. 3483e22bee78STejun Heo */ 3484e22bee78STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3485db7bccf4STejun Heo break; 3486db7bccf4STejun Heo 34873da1c84cSOleg Nesterov case CPU_POST_DEAD: 3488db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_BUTCHER; 3489e22bee78STejun Heo /* fall through */ 3490e22bee78STejun Heo case CPU_UP_CANCELED: 3491e22bee78STejun Heo destroy_worker(gcwq->first_idle); 3492e22bee78STejun Heo gcwq->first_idle = NULL; 3493db7bccf4STejun Heo break; 3494db7bccf4STejun Heo 3495db7bccf4STejun Heo case CPU_DOWN_FAILED: 34961da177e4SLinus Torvalds case CPU_ONLINE: 3497e22bee78STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 3498db7bccf4STejun Heo if (gcwq->trustee_state != TRUSTEE_DONE) { 3499db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_RELEASE; 3500db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3501db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_DONE); 3502db7bccf4STejun Heo } 35031da177e4SLinus Torvalds 3504e22bee78STejun Heo /* 3505e22bee78STejun Heo * Trustee is done and there might be no worker left. 3506e22bee78STejun Heo * Put the first_idle in and request a real manager to 3507e22bee78STejun Heo * take a look. 3508e22bee78STejun Heo */ 3509e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3510e22bee78STejun Heo kthread_bind(gcwq->first_idle->task, cpu); 3511e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3512e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGE_WORKERS; 3513e22bee78STejun Heo start_worker(gcwq->first_idle); 3514e22bee78STejun Heo gcwq->first_idle = NULL; 35151da177e4SLinus Torvalds break; 35161da177e4SLinus Torvalds } 35171da177e4SLinus Torvalds 3518db7bccf4STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 351900dfcaf7SOleg Nesterov 35201537663fSTejun Heo return notifier_from_errno(0); 35211da177e4SLinus Torvalds } 35221da177e4SLinus Torvalds 35232d3854a3SRusty Russell #ifdef CONFIG_SMP 35248ccad40dSRusty Russell 35252d3854a3SRusty Russell struct work_for_cpu { 35266b44003eSAndrew Morton struct completion completion; 35272d3854a3SRusty Russell long (*fn)(void *); 35282d3854a3SRusty Russell void *arg; 35292d3854a3SRusty Russell long ret; 35302d3854a3SRusty Russell }; 35312d3854a3SRusty Russell 35326b44003eSAndrew Morton static int do_work_for_cpu(void *_wfc) 35332d3854a3SRusty Russell { 35346b44003eSAndrew Morton struct work_for_cpu *wfc = _wfc; 35352d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 35366b44003eSAndrew Morton complete(&wfc->completion); 35376b44003eSAndrew Morton return 0; 35382d3854a3SRusty Russell } 35392d3854a3SRusty Russell 35402d3854a3SRusty Russell /** 35412d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 35422d3854a3SRusty Russell * @cpu: the cpu to run on 35432d3854a3SRusty Russell * @fn: the function to run 35442d3854a3SRusty Russell * @arg: the function arg 35452d3854a3SRusty Russell * 354631ad9081SRusty Russell * This will return the value @fn returns. 354731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 35486b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 35492d3854a3SRusty Russell */ 35502d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 35512d3854a3SRusty Russell { 35526b44003eSAndrew Morton struct task_struct *sub_thread; 35536b44003eSAndrew Morton struct work_for_cpu wfc = { 35546b44003eSAndrew Morton .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), 35556b44003eSAndrew Morton .fn = fn, 35566b44003eSAndrew Morton .arg = arg, 35576b44003eSAndrew Morton }; 35582d3854a3SRusty Russell 35596b44003eSAndrew Morton sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); 35606b44003eSAndrew Morton if (IS_ERR(sub_thread)) 35616b44003eSAndrew Morton return PTR_ERR(sub_thread); 35626b44003eSAndrew Morton kthread_bind(sub_thread, cpu); 35636b44003eSAndrew Morton wake_up_process(sub_thread); 35646b44003eSAndrew Morton wait_for_completion(&wfc.completion); 35652d3854a3SRusty Russell return wfc.ret; 35662d3854a3SRusty Russell } 35672d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 35682d3854a3SRusty Russell #endif /* CONFIG_SMP */ 35692d3854a3SRusty Russell 3570a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3571e7577c50SRusty Russell 3572a0a1a5fdSTejun Heo /** 3573a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3574a0a1a5fdSTejun Heo * 3575a0a1a5fdSTejun Heo * Start freezing workqueues. After this function returns, all 3576a0a1a5fdSTejun Heo * freezeable workqueues will queue new works to their frozen_works 35777e11629dSTejun Heo * list instead of gcwq->worklist. 3578a0a1a5fdSTejun Heo * 3579a0a1a5fdSTejun Heo * CONTEXT: 35808b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3581a0a1a5fdSTejun Heo */ 3582a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3583a0a1a5fdSTejun Heo { 3584a0a1a5fdSTejun Heo unsigned int cpu; 3585a0a1a5fdSTejun Heo 3586a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3587a0a1a5fdSTejun Heo 3588a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3589a0a1a5fdSTejun Heo workqueue_freezing = true; 3590a0a1a5fdSTejun Heo 3591f3421797STejun Heo for_each_gcwq_cpu(cpu) { 35928b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3593bdbc5dd7STejun Heo struct workqueue_struct *wq; 35948b03ae3cSTejun Heo 35958b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 35968b03ae3cSTejun Heo 3597db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3598db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3599db7bccf4STejun Heo 3600a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3601a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3602a0a1a5fdSTejun Heo 3603f3421797STejun Heo if (cwq && wq->flags & WQ_FREEZEABLE) 3604a0a1a5fdSTejun Heo cwq->max_active = 0; 36051da177e4SLinus Torvalds } 36068b03ae3cSTejun Heo 36078b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3608a0a1a5fdSTejun Heo } 3609a0a1a5fdSTejun Heo 3610a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3611a0a1a5fdSTejun Heo } 3612a0a1a5fdSTejun Heo 3613a0a1a5fdSTejun Heo /** 3614a0a1a5fdSTejun Heo * freeze_workqueues_busy - are freezeable workqueues still busy? 3615a0a1a5fdSTejun Heo * 3616a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3617a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3618a0a1a5fdSTejun Heo * 3619a0a1a5fdSTejun Heo * CONTEXT: 3620a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3621a0a1a5fdSTejun Heo * 3622a0a1a5fdSTejun Heo * RETURNS: 3623a0a1a5fdSTejun Heo * %true if some freezeable workqueues are still busy. %false if 3624a0a1a5fdSTejun Heo * freezing is complete. 3625a0a1a5fdSTejun Heo */ 3626a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3627a0a1a5fdSTejun Heo { 3628a0a1a5fdSTejun Heo unsigned int cpu; 3629a0a1a5fdSTejun Heo bool busy = false; 3630a0a1a5fdSTejun Heo 3631a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3632a0a1a5fdSTejun Heo 3633a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3634a0a1a5fdSTejun Heo 3635f3421797STejun Heo for_each_gcwq_cpu(cpu) { 3636bdbc5dd7STejun Heo struct workqueue_struct *wq; 3637a0a1a5fdSTejun Heo /* 3638a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3639a0a1a5fdSTejun Heo * to peek without lock. 3640a0a1a5fdSTejun Heo */ 3641a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3642a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3643a0a1a5fdSTejun Heo 3644f3421797STejun Heo if (!cwq || !(wq->flags & WQ_FREEZEABLE)) 3645a0a1a5fdSTejun Heo continue; 3646a0a1a5fdSTejun Heo 3647a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3648a0a1a5fdSTejun Heo if (cwq->nr_active) { 3649a0a1a5fdSTejun Heo busy = true; 3650a0a1a5fdSTejun Heo goto out_unlock; 3651a0a1a5fdSTejun Heo } 3652a0a1a5fdSTejun Heo } 3653a0a1a5fdSTejun Heo } 3654a0a1a5fdSTejun Heo out_unlock: 3655a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3656a0a1a5fdSTejun Heo return busy; 3657a0a1a5fdSTejun Heo } 3658a0a1a5fdSTejun Heo 3659a0a1a5fdSTejun Heo /** 3660a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3661a0a1a5fdSTejun Heo * 3662a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 36637e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3664a0a1a5fdSTejun Heo * 3665a0a1a5fdSTejun Heo * CONTEXT: 36668b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3667a0a1a5fdSTejun Heo */ 3668a0a1a5fdSTejun Heo void thaw_workqueues(void) 3669a0a1a5fdSTejun Heo { 3670a0a1a5fdSTejun Heo unsigned int cpu; 3671a0a1a5fdSTejun Heo 3672a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3673a0a1a5fdSTejun Heo 3674a0a1a5fdSTejun Heo if (!workqueue_freezing) 3675a0a1a5fdSTejun Heo goto out_unlock; 3676a0a1a5fdSTejun Heo 3677f3421797STejun Heo for_each_gcwq_cpu(cpu) { 36788b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3679bdbc5dd7STejun Heo struct workqueue_struct *wq; 36808b03ae3cSTejun Heo 36818b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 36828b03ae3cSTejun Heo 3683db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3684db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3685db7bccf4STejun Heo 3686a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3687a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3688a0a1a5fdSTejun Heo 3689f3421797STejun Heo if (!cwq || !(wq->flags & WQ_FREEZEABLE)) 3690a0a1a5fdSTejun Heo continue; 3691a0a1a5fdSTejun Heo 3692a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3693a0a1a5fdSTejun Heo cwq->max_active = wq->saved_max_active; 3694a0a1a5fdSTejun Heo 3695a0a1a5fdSTejun Heo while (!list_empty(&cwq->delayed_works) && 3696a0a1a5fdSTejun Heo cwq->nr_active < cwq->max_active) 3697a0a1a5fdSTejun Heo cwq_activate_first_delayed(cwq); 3698a0a1a5fdSTejun Heo } 36998b03ae3cSTejun Heo 3700e22bee78STejun Heo wake_up_worker(gcwq); 3701e22bee78STejun Heo 37028b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3703a0a1a5fdSTejun Heo } 3704a0a1a5fdSTejun Heo 3705a0a1a5fdSTejun Heo workqueue_freezing = false; 3706a0a1a5fdSTejun Heo out_unlock: 3707a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3708a0a1a5fdSTejun Heo } 3709a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3710a0a1a5fdSTejun Heo 37116ee0578bSSuresh Siddha static int __init init_workqueues(void) 37121da177e4SLinus Torvalds { 3713c34056a3STejun Heo unsigned int cpu; 3714c8e55f36STejun Heo int i; 3715c34056a3STejun Heo 3716f6500947STejun Heo cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE); 37178b03ae3cSTejun Heo 37188b03ae3cSTejun Heo /* initialize gcwqs */ 3719f3421797STejun Heo for_each_gcwq_cpu(cpu) { 37208b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 37218b03ae3cSTejun Heo 37228b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 37237e11629dSTejun Heo INIT_LIST_HEAD(&gcwq->worklist); 37248b03ae3cSTejun Heo gcwq->cpu = cpu; 3725f3421797STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 37268b03ae3cSTejun Heo 3727c8e55f36STejun Heo INIT_LIST_HEAD(&gcwq->idle_list); 3728c8e55f36STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) 3729c8e55f36STejun Heo INIT_HLIST_HEAD(&gcwq->busy_hash[i]); 3730c8e55f36STejun Heo 3731e22bee78STejun Heo init_timer_deferrable(&gcwq->idle_timer); 3732e22bee78STejun Heo gcwq->idle_timer.function = idle_worker_timeout; 3733e22bee78STejun Heo gcwq->idle_timer.data = (unsigned long)gcwq; 3734e22bee78STejun Heo 3735e22bee78STejun Heo setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout, 3736e22bee78STejun Heo (unsigned long)gcwq); 3737e22bee78STejun Heo 37388b03ae3cSTejun Heo ida_init(&gcwq->worker_ida); 3739db7bccf4STejun Heo 3740db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3741db7bccf4STejun Heo init_waitqueue_head(&gcwq->trustee_wait); 37428b03ae3cSTejun Heo } 37438b03ae3cSTejun Heo 3744e22bee78STejun Heo /* create the initial worker */ 3745f3421797STejun Heo for_each_online_gcwq_cpu(cpu) { 3746e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3747e22bee78STejun Heo struct worker *worker; 3748e22bee78STejun Heo 3749477a3c33STejun Heo if (cpu != WORK_CPU_UNBOUND) 3750477a3c33STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 3751e22bee78STejun Heo worker = create_worker(gcwq, true); 3752e22bee78STejun Heo BUG_ON(!worker); 3753e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3754e22bee78STejun Heo start_worker(worker); 3755e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3756e22bee78STejun Heo } 3757e22bee78STejun Heo 3758d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 3759d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3760d320c038STejun Heo system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0); 3761f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3762f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 3763e5cba24eSHitoshi Mitake BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq || 3764e5cba24eSHitoshi Mitake !system_unbound_wq); 37656ee0578bSSuresh Siddha return 0; 37661da177e4SLinus Torvalds } 37676ee0578bSSuresh Siddha early_initcall(init_workqueues); 3768