11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * linux/kernel/workqueue.c 31da177e4SLinus Torvalds * 41da177e4SLinus Torvalds * Generic mechanism for defining kernel helper threads for running 51da177e4SLinus Torvalds * arbitrary tasks in process context. 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Started by Ingo Molnar, Copyright (C) 2002 81da177e4SLinus Torvalds * 91da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 101da177e4SLinus Torvalds * 111da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 12e1f8e874SFrancois Cami * Andrew Morton 131da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 141da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1589ada679SChristoph Lameter * 16cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 171da177e4SLinus Torvalds */ 181da177e4SLinus Torvalds 191da177e4SLinus Torvalds #include <linux/module.h> 201da177e4SLinus Torvalds #include <linux/kernel.h> 211da177e4SLinus Torvalds #include <linux/sched.h> 221da177e4SLinus Torvalds #include <linux/init.h> 231da177e4SLinus Torvalds #include <linux/signal.h> 241da177e4SLinus Torvalds #include <linux/completion.h> 251da177e4SLinus Torvalds #include <linux/workqueue.h> 261da177e4SLinus Torvalds #include <linux/slab.h> 271da177e4SLinus Torvalds #include <linux/cpu.h> 281da177e4SLinus Torvalds #include <linux/notifier.h> 291da177e4SLinus Torvalds #include <linux/kthread.h> 301fa44ecaSJames Bottomley #include <linux/hardirq.h> 3146934023SChristoph Lameter #include <linux/mempolicy.h> 32341a5958SRafael J. Wysocki #include <linux/freezer.h> 33d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 34d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 354e6045f1SJohannes Berg #include <linux/lockdep.h> 36c34056a3STejun Heo #include <linux/idr.h> 37e22bee78STejun Heo 38e22bee78STejun Heo #include "workqueue_sched.h" 391da177e4SLinus Torvalds 40c8e55f36STejun Heo enum { 41db7bccf4STejun Heo /* global_cwq flags */ 42e22bee78STejun Heo GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 43e22bee78STejun Heo GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */ 44e22bee78STejun Heo GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 45db7bccf4STejun Heo GCWQ_FREEZING = 1 << 3, /* freeze in progress */ 46db7bccf4STejun Heo 47c8e55f36STejun Heo /* worker flags */ 48c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 49c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 50c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 51e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 52db7bccf4STejun Heo WORKER_ROGUE = 1 << 4, /* not bound to any cpu */ 53e22bee78STejun Heo WORKER_REBIND = 1 << 5, /* mom is home, come back */ 54e22bee78STejun Heo 55e22bee78STejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND, 56db7bccf4STejun Heo 57db7bccf4STejun Heo /* gcwq->trustee_state */ 58db7bccf4STejun Heo TRUSTEE_START = 0, /* start */ 59db7bccf4STejun Heo TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */ 60db7bccf4STejun Heo TRUSTEE_BUTCHER = 2, /* butcher workers */ 61db7bccf4STejun Heo TRUSTEE_RELEASE = 3, /* release workers */ 62db7bccf4STejun Heo TRUSTEE_DONE = 4, /* trustee is done */ 63c8e55f36STejun Heo 64c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 65c8e55f36STejun Heo BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, 66c8e55f36STejun Heo BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, 67db7bccf4STejun Heo 68e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 69e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 70e22bee78STejun Heo 71e22bee78STejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */ 72e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 73e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 74db7bccf4STejun Heo TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */ 75e22bee78STejun Heo 76e22bee78STejun Heo /* 77e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 78e22bee78STejun Heo * all cpus. Give -20. 79e22bee78STejun Heo */ 80e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 81c8e55f36STejun Heo }; 82c8e55f36STejun Heo 831da177e4SLinus Torvalds /* 844690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 854690c4abSTejun Heo * 864690c4abSTejun Heo * I: Set during initialization and read-only afterwards. 874690c4abSTejun Heo * 88e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 89e22bee78STejun Heo * only be modified and accessed from the local cpu. 90e22bee78STejun Heo * 918b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 924690c4abSTejun Heo * 93e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 94e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 95e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 96e22bee78STejun Heo * While trustee is in charge, it's identical to L. 97e22bee78STejun Heo * 9873f53c4aSTejun Heo * F: wq->flush_mutex protected. 9973f53c4aSTejun Heo * 1004690c4abSTejun Heo * W: workqueue_lock protected. 1014690c4abSTejun Heo */ 1024690c4abSTejun Heo 1038b03ae3cSTejun Heo struct global_cwq; 104c34056a3STejun Heo 105e22bee78STejun Heo /* 106e22bee78STejun Heo * The poor guys doing the actual heavy lifting. All on-duty workers 107e22bee78STejun Heo * are either serving the manager role, on idle list or on busy hash. 108e22bee78STejun Heo */ 109c34056a3STejun Heo struct worker { 110c8e55f36STejun Heo /* on idle list while idle, on busy hash table while busy */ 111c8e55f36STejun Heo union { 112c8e55f36STejun Heo struct list_head entry; /* L: while idle */ 113c8e55f36STejun Heo struct hlist_node hentry; /* L: while busy */ 114c8e55f36STejun Heo }; 115c8e55f36STejun Heo 116c34056a3STejun Heo struct work_struct *current_work; /* L: work being processed */ 1178cca0eeaSTejun Heo struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ 118affee4b2STejun Heo struct list_head scheduled; /* L: scheduled works */ 119c34056a3STejun Heo struct task_struct *task; /* I: worker task */ 1208b03ae3cSTejun Heo struct global_cwq *gcwq; /* I: the associated gcwq */ 121e22bee78STejun Heo /* 64 bytes boundary on 64bit, 32 on 32bit */ 122e22bee78STejun Heo unsigned long last_active; /* L: last active timestamp */ 123e22bee78STejun Heo unsigned int flags; /* X: flags */ 124c34056a3STejun Heo int id; /* I: worker id */ 125e22bee78STejun Heo struct work_struct rebind_work; /* L: rebind worker to cpu */ 126c34056a3STejun Heo }; 127c34056a3STejun Heo 1284690c4abSTejun Heo /* 129e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 130e22bee78STejun Heo * and all works are queued and processed here regardless of their 131e22bee78STejun Heo * target workqueues. 1328b03ae3cSTejun Heo */ 1338b03ae3cSTejun Heo struct global_cwq { 1348b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1357e11629dSTejun Heo struct list_head worklist; /* L: list of pending works */ 1368b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 137db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 138c8e55f36STejun Heo 139c8e55f36STejun Heo int nr_workers; /* L: total number of workers */ 140c8e55f36STejun Heo int nr_idle; /* L: currently idle ones */ 141c8e55f36STejun Heo 142c8e55f36STejun Heo /* workers are chained either in the idle_list or busy_hash */ 143e22bee78STejun Heo struct list_head idle_list; /* X: list of idle workers */ 144c8e55f36STejun Heo struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; 145c8e55f36STejun Heo /* L: hash of busy workers */ 146c8e55f36STejun Heo 147e22bee78STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 148e22bee78STejun Heo struct timer_list mayday_timer; /* L: SOS timer for dworkers */ 149e22bee78STejun Heo 1508b03ae3cSTejun Heo struct ida worker_ida; /* L: for worker IDs */ 151db7bccf4STejun Heo 152db7bccf4STejun Heo struct task_struct *trustee; /* L: for gcwq shutdown */ 153db7bccf4STejun Heo unsigned int trustee_state; /* L: trustee state */ 154db7bccf4STejun Heo wait_queue_head_t trustee_wait; /* trustee wait */ 155e22bee78STejun Heo struct worker *first_idle; /* L: first idle worker */ 1568b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1578b03ae3cSTejun Heo 1588b03ae3cSTejun Heo /* 159502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 1600f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 1610f900049STejun Heo * aligned at two's power of the number of flag bits. 1621da177e4SLinus Torvalds */ 1631da177e4SLinus Torvalds struct cpu_workqueue_struct { 1648b03ae3cSTejun Heo struct global_cwq *gcwq; /* I: the associated gcwq */ 1654690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 16673f53c4aSTejun Heo int work_color; /* L: current color */ 16773f53c4aSTejun Heo int flush_color; /* L: flushing color */ 16873f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 16973f53c4aSTejun Heo /* L: nr of in_flight works */ 1701e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 171a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1721e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1730f900049STejun Heo }; 1741da177e4SLinus Torvalds 1751da177e4SLinus Torvalds /* 17673f53c4aSTejun Heo * Structure used to wait for workqueue flush. 17773f53c4aSTejun Heo */ 17873f53c4aSTejun Heo struct wq_flusher { 17973f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 18073f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 18173f53c4aSTejun Heo struct completion done; /* flush completion */ 18273f53c4aSTejun Heo }; 18373f53c4aSTejun Heo 18473f53c4aSTejun Heo /* 1851da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 1861da177e4SLinus Torvalds * per-CPU workqueues: 1871da177e4SLinus Torvalds */ 1881da177e4SLinus Torvalds struct workqueue_struct { 18997e37d7bSTejun Heo unsigned int flags; /* I: WQ_* flags */ 1904690c4abSTejun Heo struct cpu_workqueue_struct *cpu_wq; /* I: cwq's */ 1914690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 19273f53c4aSTejun Heo 19373f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 19473f53c4aSTejun Heo int work_color; /* F: current work color */ 19573f53c4aSTejun Heo int flush_color; /* F: current flush color */ 19673f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 19773f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 19873f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 19973f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 20073f53c4aSTejun Heo 201502ca9d8STejun Heo unsigned long single_cpu; /* cpu for single cpu wq */ 202502ca9d8STejun Heo 203e22bee78STejun Heo cpumask_var_t mayday_mask; /* cpus requesting rescue */ 204e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 205e22bee78STejun Heo 206dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2074690c4abSTejun Heo const char *name; /* I: workqueue name */ 2084e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2094e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2104e6045f1SJohannes Berg #endif 2111da177e4SLinus Torvalds }; 2121da177e4SLinus Torvalds 213d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 214d320c038STejun Heo struct workqueue_struct *system_long_wq __read_mostly; 215d320c038STejun Heo struct workqueue_struct *system_nrt_wq __read_mostly; 216d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 217d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 218d320c038STejun Heo EXPORT_SYMBOL_GPL(system_nrt_wq); 219d320c038STejun Heo 220db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 221db7bccf4STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ 222db7bccf4STejun Heo hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) 223db7bccf4STejun Heo 224dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 225dc186ad7SThomas Gleixner 226dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 227dc186ad7SThomas Gleixner 228dc186ad7SThomas Gleixner /* 229dc186ad7SThomas Gleixner * fixup_init is called when: 230dc186ad7SThomas Gleixner * - an active object is initialized 231dc186ad7SThomas Gleixner */ 232dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 233dc186ad7SThomas Gleixner { 234dc186ad7SThomas Gleixner struct work_struct *work = addr; 235dc186ad7SThomas Gleixner 236dc186ad7SThomas Gleixner switch (state) { 237dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 238dc186ad7SThomas Gleixner cancel_work_sync(work); 239dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 240dc186ad7SThomas Gleixner return 1; 241dc186ad7SThomas Gleixner default: 242dc186ad7SThomas Gleixner return 0; 243dc186ad7SThomas Gleixner } 244dc186ad7SThomas Gleixner } 245dc186ad7SThomas Gleixner 246dc186ad7SThomas Gleixner /* 247dc186ad7SThomas Gleixner * fixup_activate is called when: 248dc186ad7SThomas Gleixner * - an active object is activated 249dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 250dc186ad7SThomas Gleixner */ 251dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 252dc186ad7SThomas Gleixner { 253dc186ad7SThomas Gleixner struct work_struct *work = addr; 254dc186ad7SThomas Gleixner 255dc186ad7SThomas Gleixner switch (state) { 256dc186ad7SThomas Gleixner 257dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 258dc186ad7SThomas Gleixner /* 259dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 260dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 261dc186ad7SThomas Gleixner * is tracked in the object tracker. 262dc186ad7SThomas Gleixner */ 26322df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 264dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 265dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 266dc186ad7SThomas Gleixner return 0; 267dc186ad7SThomas Gleixner } 268dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 269dc186ad7SThomas Gleixner return 0; 270dc186ad7SThomas Gleixner 271dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 272dc186ad7SThomas Gleixner WARN_ON(1); 273dc186ad7SThomas Gleixner 274dc186ad7SThomas Gleixner default: 275dc186ad7SThomas Gleixner return 0; 276dc186ad7SThomas Gleixner } 277dc186ad7SThomas Gleixner } 278dc186ad7SThomas Gleixner 279dc186ad7SThomas Gleixner /* 280dc186ad7SThomas Gleixner * fixup_free is called when: 281dc186ad7SThomas Gleixner * - an active object is freed 282dc186ad7SThomas Gleixner */ 283dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 284dc186ad7SThomas Gleixner { 285dc186ad7SThomas Gleixner struct work_struct *work = addr; 286dc186ad7SThomas Gleixner 287dc186ad7SThomas Gleixner switch (state) { 288dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 289dc186ad7SThomas Gleixner cancel_work_sync(work); 290dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 291dc186ad7SThomas Gleixner return 1; 292dc186ad7SThomas Gleixner default: 293dc186ad7SThomas Gleixner return 0; 294dc186ad7SThomas Gleixner } 295dc186ad7SThomas Gleixner } 296dc186ad7SThomas Gleixner 297dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 298dc186ad7SThomas Gleixner .name = "work_struct", 299dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 300dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 301dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 302dc186ad7SThomas Gleixner }; 303dc186ad7SThomas Gleixner 304dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 305dc186ad7SThomas Gleixner { 306dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 307dc186ad7SThomas Gleixner } 308dc186ad7SThomas Gleixner 309dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 310dc186ad7SThomas Gleixner { 311dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 312dc186ad7SThomas Gleixner } 313dc186ad7SThomas Gleixner 314dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 315dc186ad7SThomas Gleixner { 316dc186ad7SThomas Gleixner if (onstack) 317dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 318dc186ad7SThomas Gleixner else 319dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 320dc186ad7SThomas Gleixner } 321dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 322dc186ad7SThomas Gleixner 323dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 324dc186ad7SThomas Gleixner { 325dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 326dc186ad7SThomas Gleixner } 327dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 328dc186ad7SThomas Gleixner 329dc186ad7SThomas Gleixner #else 330dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 331dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 332dc186ad7SThomas Gleixner #endif 333dc186ad7SThomas Gleixner 33495402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 33595402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 3361da177e4SLinus Torvalds static LIST_HEAD(workqueues); 337a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 338c34056a3STejun Heo 339e22bee78STejun Heo /* 340e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 341e22bee78STejun Heo * which is expected to be used frequently by other cpus via 342e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 343e22bee78STejun Heo */ 3448b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 345e22bee78STejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running); 3468b03ae3cSTejun Heo 347c34056a3STejun Heo static int worker_thread(void *__worker); 3481da177e4SLinus Torvalds 3498b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 3508b03ae3cSTejun Heo { 3518b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 3528b03ae3cSTejun Heo } 3538b03ae3cSTejun Heo 354e22bee78STejun Heo static atomic_t *get_gcwq_nr_running(unsigned int cpu) 355e22bee78STejun Heo { 356e22bee78STejun Heo return &per_cpu(gcwq_nr_running, cpu); 357e22bee78STejun Heo } 358e22bee78STejun Heo 3594690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 3604690c4abSTejun Heo struct workqueue_struct *wq) 361a848e3b6SOleg Nesterov { 362a848e3b6SOleg Nesterov return per_cpu_ptr(wq->cpu_wq, cpu); 363a848e3b6SOleg Nesterov } 364a848e3b6SOleg Nesterov 36573f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 36673f53c4aSTejun Heo { 36773f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 36873f53c4aSTejun Heo } 36973f53c4aSTejun Heo 37073f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 37173f53c4aSTejun Heo { 37273f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 37373f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 37473f53c4aSTejun Heo } 37573f53c4aSTejun Heo 37673f53c4aSTejun Heo static int work_next_color(int color) 37773f53c4aSTejun Heo { 37873f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 37973f53c4aSTejun Heo } 38073f53c4aSTejun Heo 3814594bf15SDavid Howells /* 3827a22ad75STejun Heo * Work data points to the cwq while a work is on queue. Once 3837a22ad75STejun Heo * execution starts, it points to the cpu the work was last on. This 3847a22ad75STejun Heo * can be distinguished by comparing the data value against 3857a22ad75STejun Heo * PAGE_OFFSET. 3867a22ad75STejun Heo * 3877a22ad75STejun Heo * set_work_{cwq|cpu}() and clear_work_data() can be used to set the 3887a22ad75STejun Heo * cwq, cpu or clear work->data. These functions should only be 3897a22ad75STejun Heo * called while the work is owned - ie. while the PENDING bit is set. 3907a22ad75STejun Heo * 3917a22ad75STejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq 3927a22ad75STejun Heo * corresponding to a work. gcwq is available once the work has been 3937a22ad75STejun Heo * queued anywhere after initialization. cwq is available only from 3947a22ad75STejun Heo * queueing until execution starts. 3954594bf15SDavid Howells */ 3967a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 3977a22ad75STejun Heo unsigned long flags) 3987a22ad75STejun Heo { 3997a22ad75STejun Heo BUG_ON(!work_pending(work)); 4007a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 4017a22ad75STejun Heo } 4027a22ad75STejun Heo 4037a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 4044690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 4054690c4abSTejun Heo unsigned long extra_flags) 406365970a1SDavid Howells { 4077a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 40822df02bbSTejun Heo WORK_STRUCT_PENDING | extra_flags); 409365970a1SDavid Howells } 410365970a1SDavid Howells 4117a22ad75STejun Heo static void set_work_cpu(struct work_struct *work, unsigned int cpu) 4124d707b9fSOleg Nesterov { 4137a22ad75STejun Heo set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING); 4144d707b9fSOleg Nesterov } 4154d707b9fSOleg Nesterov 4167a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 417365970a1SDavid Howells { 4187a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 4197a22ad75STejun Heo } 4207a22ad75STejun Heo 4217a22ad75STejun Heo static inline unsigned long get_work_data(struct work_struct *work) 4227a22ad75STejun Heo { 4237a22ad75STejun Heo return atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK; 4247a22ad75STejun Heo } 4257a22ad75STejun Heo 4267a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 4277a22ad75STejun Heo { 4287a22ad75STejun Heo unsigned long data = get_work_data(work); 4297a22ad75STejun Heo 4307a22ad75STejun Heo return data >= PAGE_OFFSET ? (void *)data : NULL; 4317a22ad75STejun Heo } 4327a22ad75STejun Heo 4337a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 4347a22ad75STejun Heo { 4357a22ad75STejun Heo unsigned long data = get_work_data(work); 4367a22ad75STejun Heo unsigned int cpu; 4377a22ad75STejun Heo 4387a22ad75STejun Heo if (data >= PAGE_OFFSET) 4397a22ad75STejun Heo return ((struct cpu_workqueue_struct *)data)->gcwq; 4407a22ad75STejun Heo 4417a22ad75STejun Heo cpu = data >> WORK_STRUCT_FLAG_BITS; 4427a22ad75STejun Heo if (cpu == NR_CPUS) 4437a22ad75STejun Heo return NULL; 4447a22ad75STejun Heo 4457a22ad75STejun Heo BUG_ON(cpu >= num_possible_cpus()); 4467a22ad75STejun Heo return get_gcwq(cpu); 447365970a1SDavid Howells } 448365970a1SDavid Howells 449e22bee78STejun Heo /* 450e22bee78STejun Heo * Policy functions. These define the policies on how the global 451e22bee78STejun Heo * worker pool is managed. Unless noted otherwise, these functions 452e22bee78STejun Heo * assume that they're being called with gcwq->lock held. 453e22bee78STejun Heo */ 454e22bee78STejun Heo 455e22bee78STejun Heo /* 456e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 457e22bee78STejun Heo * running workers. 458e22bee78STejun Heo */ 459e22bee78STejun Heo static bool need_more_worker(struct global_cwq *gcwq) 460e22bee78STejun Heo { 461e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 462e22bee78STejun Heo 463e22bee78STejun Heo return !list_empty(&gcwq->worklist) && !atomic_read(nr_running); 464e22bee78STejun Heo } 465e22bee78STejun Heo 466e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 467e22bee78STejun Heo static bool may_start_working(struct global_cwq *gcwq) 468e22bee78STejun Heo { 469e22bee78STejun Heo return gcwq->nr_idle; 470e22bee78STejun Heo } 471e22bee78STejun Heo 472e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 473e22bee78STejun Heo static bool keep_working(struct global_cwq *gcwq) 474e22bee78STejun Heo { 475e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 476e22bee78STejun Heo 477e22bee78STejun Heo return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1; 478e22bee78STejun Heo } 479e22bee78STejun Heo 480e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 481e22bee78STejun Heo static bool need_to_create_worker(struct global_cwq *gcwq) 482e22bee78STejun Heo { 483e22bee78STejun Heo return need_more_worker(gcwq) && !may_start_working(gcwq); 484e22bee78STejun Heo } 485e22bee78STejun Heo 486e22bee78STejun Heo /* Do I need to be the manager? */ 487e22bee78STejun Heo static bool need_to_manage_workers(struct global_cwq *gcwq) 488e22bee78STejun Heo { 489e22bee78STejun Heo return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS; 490e22bee78STejun Heo } 491e22bee78STejun Heo 492e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 493e22bee78STejun Heo static bool too_many_workers(struct global_cwq *gcwq) 494e22bee78STejun Heo { 495e22bee78STejun Heo bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS; 496e22bee78STejun Heo int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */ 497e22bee78STejun Heo int nr_busy = gcwq->nr_workers - nr_idle; 498e22bee78STejun Heo 499e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 500e22bee78STejun Heo } 501e22bee78STejun Heo 502e22bee78STejun Heo /* 503e22bee78STejun Heo * Wake up functions. 504e22bee78STejun Heo */ 505e22bee78STejun Heo 5067e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 5077e11629dSTejun Heo static struct worker *first_worker(struct global_cwq *gcwq) 5087e11629dSTejun Heo { 5097e11629dSTejun Heo if (unlikely(list_empty(&gcwq->idle_list))) 5107e11629dSTejun Heo return NULL; 5117e11629dSTejun Heo 5127e11629dSTejun Heo return list_first_entry(&gcwq->idle_list, struct worker, entry); 5137e11629dSTejun Heo } 5147e11629dSTejun Heo 5157e11629dSTejun Heo /** 5167e11629dSTejun Heo * wake_up_worker - wake up an idle worker 5177e11629dSTejun Heo * @gcwq: gcwq to wake worker for 5187e11629dSTejun Heo * 5197e11629dSTejun Heo * Wake up the first idle worker of @gcwq. 5207e11629dSTejun Heo * 5217e11629dSTejun Heo * CONTEXT: 5227e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 5237e11629dSTejun Heo */ 5247e11629dSTejun Heo static void wake_up_worker(struct global_cwq *gcwq) 5257e11629dSTejun Heo { 5267e11629dSTejun Heo struct worker *worker = first_worker(gcwq); 5277e11629dSTejun Heo 5287e11629dSTejun Heo if (likely(worker)) 5297e11629dSTejun Heo wake_up_process(worker->task); 5307e11629dSTejun Heo } 5317e11629dSTejun Heo 5324690c4abSTejun Heo /** 533e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 534e22bee78STejun Heo * @task: task waking up 535e22bee78STejun Heo * @cpu: CPU @task is waking up to 536e22bee78STejun Heo * 537e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 538e22bee78STejun Heo * being awoken. 539e22bee78STejun Heo * 540e22bee78STejun Heo * CONTEXT: 541e22bee78STejun Heo * spin_lock_irq(rq->lock) 542e22bee78STejun Heo */ 543e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 544e22bee78STejun Heo { 545e22bee78STejun Heo struct worker *worker = kthread_data(task); 546e22bee78STejun Heo 547e22bee78STejun Heo if (likely(!(worker->flags & WORKER_NOT_RUNNING))) 548e22bee78STejun Heo atomic_inc(get_gcwq_nr_running(cpu)); 549e22bee78STejun Heo } 550e22bee78STejun Heo 551e22bee78STejun Heo /** 552e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 553e22bee78STejun Heo * @task: task going to sleep 554e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 555e22bee78STejun Heo * 556e22bee78STejun Heo * This function is called during schedule() when a busy worker is 557e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 558e22bee78STejun Heo * returning pointer to its task. 559e22bee78STejun Heo * 560e22bee78STejun Heo * CONTEXT: 561e22bee78STejun Heo * spin_lock_irq(rq->lock) 562e22bee78STejun Heo * 563e22bee78STejun Heo * RETURNS: 564e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 565e22bee78STejun Heo */ 566e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 567e22bee78STejun Heo unsigned int cpu) 568e22bee78STejun Heo { 569e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 570e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 571e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(cpu); 572e22bee78STejun Heo 573e22bee78STejun Heo if (unlikely(worker->flags & WORKER_NOT_RUNNING)) 574e22bee78STejun Heo return NULL; 575e22bee78STejun Heo 576e22bee78STejun Heo /* this can only happen on the local cpu */ 577e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 578e22bee78STejun Heo 579e22bee78STejun Heo /* 580e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 581e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 582e22bee78STejun Heo * Please read comment there. 583e22bee78STejun Heo * 584e22bee78STejun Heo * NOT_RUNNING is clear. This means that trustee is not in 585e22bee78STejun Heo * charge and we're running on the local cpu w/ rq lock held 586e22bee78STejun Heo * and preemption disabled, which in turn means that none else 587e22bee78STejun Heo * could be manipulating idle_list, so dereferencing idle_list 588e22bee78STejun Heo * without gcwq lock is safe. 589e22bee78STejun Heo */ 590e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist)) 591e22bee78STejun Heo to_wakeup = first_worker(gcwq); 592e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 593e22bee78STejun Heo } 594e22bee78STejun Heo 595e22bee78STejun Heo /** 596e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 597d302f017STejun Heo * @worker: worker to set flags for 598d302f017STejun Heo * @flags: flags to set 599d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 600d302f017STejun Heo * 601e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 602e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 603e22bee78STejun Heo * woken up. 604d302f017STejun Heo * 605d302f017STejun Heo * LOCKING: 606d302f017STejun Heo * spin_lock_irq(gcwq->lock). 607d302f017STejun Heo */ 608d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 609d302f017STejun Heo bool wakeup) 610d302f017STejun Heo { 611e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 612e22bee78STejun Heo 613e22bee78STejun Heo /* 614e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 615e22bee78STejun Heo * wake up an idle worker as necessary if requested by 616e22bee78STejun Heo * @wakeup. 617e22bee78STejun Heo */ 618e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 619e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 620e22bee78STejun Heo atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); 621e22bee78STejun Heo 622e22bee78STejun Heo if (wakeup) { 623e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 624e22bee78STejun Heo !list_empty(&gcwq->worklist)) 625e22bee78STejun Heo wake_up_worker(gcwq); 626e22bee78STejun Heo } else 627e22bee78STejun Heo atomic_dec(nr_running); 628e22bee78STejun Heo } 629e22bee78STejun Heo 630d302f017STejun Heo worker->flags |= flags; 631d302f017STejun Heo } 632d302f017STejun Heo 633d302f017STejun Heo /** 634e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 635d302f017STejun Heo * @worker: worker to set flags for 636d302f017STejun Heo * @flags: flags to clear 637d302f017STejun Heo * 638e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 639d302f017STejun Heo * 640d302f017STejun Heo * LOCKING: 641d302f017STejun Heo * spin_lock_irq(gcwq->lock). 642d302f017STejun Heo */ 643d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 644d302f017STejun Heo { 645e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 646e22bee78STejun Heo unsigned int oflags = worker->flags; 647e22bee78STejun Heo 648d302f017STejun Heo worker->flags &= ~flags; 649e22bee78STejun Heo 650e22bee78STejun Heo /* if transitioning out of NOT_RUNNING, increment nr_running */ 651e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 652e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 653e22bee78STejun Heo atomic_inc(get_gcwq_nr_running(gcwq->cpu)); 654d302f017STejun Heo } 655d302f017STejun Heo 656d302f017STejun Heo /** 657c8e55f36STejun Heo * busy_worker_head - return the busy hash head for a work 658c8e55f36STejun Heo * @gcwq: gcwq of interest 659c8e55f36STejun Heo * @work: work to be hashed 660c8e55f36STejun Heo * 661c8e55f36STejun Heo * Return hash head of @gcwq for @work. 662c8e55f36STejun Heo * 663c8e55f36STejun Heo * CONTEXT: 664c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 665c8e55f36STejun Heo * 666c8e55f36STejun Heo * RETURNS: 667c8e55f36STejun Heo * Pointer to the hash head. 668c8e55f36STejun Heo */ 669c8e55f36STejun Heo static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, 670c8e55f36STejun Heo struct work_struct *work) 671c8e55f36STejun Heo { 672c8e55f36STejun Heo const int base_shift = ilog2(sizeof(struct work_struct)); 673c8e55f36STejun Heo unsigned long v = (unsigned long)work; 674c8e55f36STejun Heo 675c8e55f36STejun Heo /* simple shift and fold hash, do we need something better? */ 676c8e55f36STejun Heo v >>= base_shift; 677c8e55f36STejun Heo v += v >> BUSY_WORKER_HASH_ORDER; 678c8e55f36STejun Heo v &= BUSY_WORKER_HASH_MASK; 679c8e55f36STejun Heo 680c8e55f36STejun Heo return &gcwq->busy_hash[v]; 681c8e55f36STejun Heo } 682c8e55f36STejun Heo 683c8e55f36STejun Heo /** 6848cca0eeaSTejun Heo * __find_worker_executing_work - find worker which is executing a work 6858cca0eeaSTejun Heo * @gcwq: gcwq of interest 6868cca0eeaSTejun Heo * @bwh: hash head as returned by busy_worker_head() 6878cca0eeaSTejun Heo * @work: work to find worker for 6888cca0eeaSTejun Heo * 6898cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. @bwh should be 6908cca0eeaSTejun Heo * the hash head obtained by calling busy_worker_head() with the same 6918cca0eeaSTejun Heo * work. 6928cca0eeaSTejun Heo * 6938cca0eeaSTejun Heo * CONTEXT: 6948cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 6958cca0eeaSTejun Heo * 6968cca0eeaSTejun Heo * RETURNS: 6978cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 6988cca0eeaSTejun Heo * otherwise. 6998cca0eeaSTejun Heo */ 7008cca0eeaSTejun Heo static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, 7018cca0eeaSTejun Heo struct hlist_head *bwh, 7028cca0eeaSTejun Heo struct work_struct *work) 7038cca0eeaSTejun Heo { 7048cca0eeaSTejun Heo struct worker *worker; 7058cca0eeaSTejun Heo struct hlist_node *tmp; 7068cca0eeaSTejun Heo 7078cca0eeaSTejun Heo hlist_for_each_entry(worker, tmp, bwh, hentry) 7088cca0eeaSTejun Heo if (worker->current_work == work) 7098cca0eeaSTejun Heo return worker; 7108cca0eeaSTejun Heo return NULL; 7118cca0eeaSTejun Heo } 7128cca0eeaSTejun Heo 7138cca0eeaSTejun Heo /** 7148cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 7158cca0eeaSTejun Heo * @gcwq: gcwq of interest 7168cca0eeaSTejun Heo * @work: work to find worker for 7178cca0eeaSTejun Heo * 7188cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. This function is 7198cca0eeaSTejun Heo * identical to __find_worker_executing_work() except that this 7208cca0eeaSTejun Heo * function calculates @bwh itself. 7218cca0eeaSTejun Heo * 7228cca0eeaSTejun Heo * CONTEXT: 7238cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 7248cca0eeaSTejun Heo * 7258cca0eeaSTejun Heo * RETURNS: 7268cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 7278cca0eeaSTejun Heo * otherwise. 7288cca0eeaSTejun Heo */ 7298cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 7308cca0eeaSTejun Heo struct work_struct *work) 7318cca0eeaSTejun Heo { 7328cca0eeaSTejun Heo return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), 7338cca0eeaSTejun Heo work); 7348cca0eeaSTejun Heo } 7358cca0eeaSTejun Heo 7368cca0eeaSTejun Heo /** 7377e11629dSTejun Heo * insert_work - insert a work into gcwq 7384690c4abSTejun Heo * @cwq: cwq @work belongs to 7394690c4abSTejun Heo * @work: work to insert 7404690c4abSTejun Heo * @head: insertion point 7414690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 7424690c4abSTejun Heo * 7437e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 7447e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 7454690c4abSTejun Heo * 7464690c4abSTejun Heo * CONTEXT: 7478b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 7484690c4abSTejun Heo */ 749b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 7504690c4abSTejun Heo struct work_struct *work, struct list_head *head, 7514690c4abSTejun Heo unsigned int extra_flags) 752b89deed3SOleg Nesterov { 753e22bee78STejun Heo struct global_cwq *gcwq = cwq->gcwq; 754e22bee78STejun Heo 7554690c4abSTejun Heo /* we own @work, set data and link */ 7567a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 7574690c4abSTejun Heo 7586e84d644SOleg Nesterov /* 7596e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 7606e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 7616e84d644SOleg Nesterov */ 7626e84d644SOleg Nesterov smp_wmb(); 7634690c4abSTejun Heo 7641a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 765e22bee78STejun Heo 766e22bee78STejun Heo /* 767e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 768e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 769e22bee78STejun Heo * lying around lazily while there are works to be processed. 770e22bee78STejun Heo */ 771e22bee78STejun Heo smp_mb(); 772e22bee78STejun Heo 773e22bee78STejun Heo if (!atomic_read(get_gcwq_nr_running(gcwq->cpu))) 774e22bee78STejun Heo wake_up_worker(gcwq); 775b89deed3SOleg Nesterov } 776b89deed3SOleg Nesterov 777502ca9d8STejun Heo /** 778502ca9d8STejun Heo * cwq_unbind_single_cpu - unbind cwq from single cpu workqueue processing 779502ca9d8STejun Heo * @cwq: cwq to unbind 780502ca9d8STejun Heo * 781502ca9d8STejun Heo * Try to unbind @cwq from single cpu workqueue processing. If 782502ca9d8STejun Heo * @cwq->wq is frozen, unbind is delayed till the workqueue is thawed. 783502ca9d8STejun Heo * 784502ca9d8STejun Heo * CONTEXT: 785502ca9d8STejun Heo * spin_lock_irq(gcwq->lock). 786502ca9d8STejun Heo */ 787502ca9d8STejun Heo static void cwq_unbind_single_cpu(struct cpu_workqueue_struct *cwq) 788502ca9d8STejun Heo { 789502ca9d8STejun Heo struct workqueue_struct *wq = cwq->wq; 790502ca9d8STejun Heo struct global_cwq *gcwq = cwq->gcwq; 791502ca9d8STejun Heo 792502ca9d8STejun Heo BUG_ON(wq->single_cpu != gcwq->cpu); 793502ca9d8STejun Heo /* 794502ca9d8STejun Heo * Unbind from workqueue if @cwq is not frozen. If frozen, 795502ca9d8STejun Heo * thaw_workqueues() will either restart processing on this 796502ca9d8STejun Heo * cpu or unbind if empty. This keeps works queued while 797502ca9d8STejun Heo * frozen fully ordered and flushable. 798502ca9d8STejun Heo */ 799502ca9d8STejun Heo if (likely(!(gcwq->flags & GCWQ_FREEZING))) { 800502ca9d8STejun Heo smp_wmb(); /* paired with cmpxchg() in __queue_work() */ 801502ca9d8STejun Heo wq->single_cpu = NR_CPUS; 802502ca9d8STejun Heo } 803502ca9d8STejun Heo } 804502ca9d8STejun Heo 8054690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 8061da177e4SLinus Torvalds struct work_struct *work) 8071da177e4SLinus Torvalds { 808502ca9d8STejun Heo struct global_cwq *gcwq; 809502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 8101e19ffc6STejun Heo struct list_head *worklist; 8111da177e4SLinus Torvalds unsigned long flags; 812502ca9d8STejun Heo bool arbitrate; 8131da177e4SLinus Torvalds 814dc186ad7SThomas Gleixner debug_work_activate(work); 8151e19ffc6STejun Heo 81618aa9effSTejun Heo /* 81718aa9effSTejun Heo * Determine gcwq to use. SINGLE_CPU is inherently 81818aa9effSTejun Heo * NON_REENTRANT, so test it first. 81918aa9effSTejun Heo */ 820502ca9d8STejun Heo if (!(wq->flags & WQ_SINGLE_CPU)) { 82118aa9effSTejun Heo struct global_cwq *last_gcwq; 82218aa9effSTejun Heo 82318aa9effSTejun Heo /* 82418aa9effSTejun Heo * It's multi cpu. If @wq is non-reentrant and @work 82518aa9effSTejun Heo * was previously on a different cpu, it might still 82618aa9effSTejun Heo * be running there, in which case the work needs to 82718aa9effSTejun Heo * be queued on that cpu to guarantee non-reentrance. 82818aa9effSTejun Heo */ 829502ca9d8STejun Heo gcwq = get_gcwq(cpu); 83018aa9effSTejun Heo if (wq->flags & WQ_NON_REENTRANT && 83118aa9effSTejun Heo (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) { 83218aa9effSTejun Heo struct worker *worker; 83318aa9effSTejun Heo 83418aa9effSTejun Heo spin_lock_irqsave(&last_gcwq->lock, flags); 83518aa9effSTejun Heo 83618aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 83718aa9effSTejun Heo 83818aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 83918aa9effSTejun Heo gcwq = last_gcwq; 84018aa9effSTejun Heo else { 84118aa9effSTejun Heo /* meh... not running there, queue here */ 84218aa9effSTejun Heo spin_unlock_irqrestore(&last_gcwq->lock, flags); 84318aa9effSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 84418aa9effSTejun Heo } 84518aa9effSTejun Heo } else 8468b03ae3cSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 847502ca9d8STejun Heo } else { 848502ca9d8STejun Heo unsigned int req_cpu = cpu; 849502ca9d8STejun Heo 850502ca9d8STejun Heo /* 851502ca9d8STejun Heo * It's a bit more complex for single cpu workqueues. 852502ca9d8STejun Heo * We first need to determine which cpu is going to be 853502ca9d8STejun Heo * used. If no cpu is currently serving this 854502ca9d8STejun Heo * workqueue, arbitrate using atomic accesses to 855502ca9d8STejun Heo * wq->single_cpu; otherwise, use the current one. 856502ca9d8STejun Heo */ 857502ca9d8STejun Heo retry: 858502ca9d8STejun Heo cpu = wq->single_cpu; 859502ca9d8STejun Heo arbitrate = cpu == NR_CPUS; 860502ca9d8STejun Heo if (arbitrate) 861502ca9d8STejun Heo cpu = req_cpu; 862502ca9d8STejun Heo 863502ca9d8STejun Heo gcwq = get_gcwq(cpu); 864502ca9d8STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 865502ca9d8STejun Heo 866502ca9d8STejun Heo /* 867502ca9d8STejun Heo * The following cmpxchg() is a full barrier paired 868502ca9d8STejun Heo * with smp_wmb() in cwq_unbind_single_cpu() and 869502ca9d8STejun Heo * guarantees that all changes to wq->st_* fields are 870502ca9d8STejun Heo * visible on the new cpu after this point. 871502ca9d8STejun Heo */ 872502ca9d8STejun Heo if (arbitrate) 873502ca9d8STejun Heo cmpxchg(&wq->single_cpu, NR_CPUS, cpu); 874502ca9d8STejun Heo 875502ca9d8STejun Heo if (unlikely(wq->single_cpu != cpu)) { 876502ca9d8STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 877502ca9d8STejun Heo goto retry; 878502ca9d8STejun Heo } 879502ca9d8STejun Heo } 880502ca9d8STejun Heo 881502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 882502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 883502ca9d8STejun Heo 8844690c4abSTejun Heo BUG_ON(!list_empty(&work->entry)); 8851e19ffc6STejun Heo 88673f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 8871e19ffc6STejun Heo 8881e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 8891e19ffc6STejun Heo cwq->nr_active++; 8907e11629dSTejun Heo worklist = &gcwq->worklist; 8911e19ffc6STejun Heo } else 8921e19ffc6STejun Heo worklist = &cwq->delayed_works; 8931e19ffc6STejun Heo 8941e19ffc6STejun Heo insert_work(cwq, work, worklist, work_color_to_flags(cwq->work_color)); 8951e19ffc6STejun Heo 8968b03ae3cSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 8971da177e4SLinus Torvalds } 8981da177e4SLinus Torvalds 8990fcb78c2SRolf Eike Beer /** 9000fcb78c2SRolf Eike Beer * queue_work - queue work on a workqueue 9010fcb78c2SRolf Eike Beer * @wq: workqueue to use 9020fcb78c2SRolf Eike Beer * @work: work to queue 9030fcb78c2SRolf Eike Beer * 904057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 9051da177e4SLinus Torvalds * 90600dfcaf7SOleg Nesterov * We queue the work to the CPU on which it was submitted, but if the CPU dies 90700dfcaf7SOleg Nesterov * it can be processed by another CPU. 9081da177e4SLinus Torvalds */ 9097ad5b3a5SHarvey Harrison int queue_work(struct workqueue_struct *wq, struct work_struct *work) 9101da177e4SLinus Torvalds { 911ef1ca236SOleg Nesterov int ret; 9121da177e4SLinus Torvalds 913ef1ca236SOleg Nesterov ret = queue_work_on(get_cpu(), wq, work); 914a848e3b6SOleg Nesterov put_cpu(); 915ef1ca236SOleg Nesterov 9161da177e4SLinus Torvalds return ret; 9171da177e4SLinus Torvalds } 918ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_work); 9191da177e4SLinus Torvalds 920c1a220e7SZhang Rui /** 921c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 922c1a220e7SZhang Rui * @cpu: CPU number to execute work on 923c1a220e7SZhang Rui * @wq: workqueue to use 924c1a220e7SZhang Rui * @work: work to queue 925c1a220e7SZhang Rui * 926c1a220e7SZhang Rui * Returns 0 if @work was already on a queue, non-zero otherwise. 927c1a220e7SZhang Rui * 928c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 929c1a220e7SZhang Rui * can't go away. 930c1a220e7SZhang Rui */ 931c1a220e7SZhang Rui int 932c1a220e7SZhang Rui queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work) 933c1a220e7SZhang Rui { 934c1a220e7SZhang Rui int ret = 0; 935c1a220e7SZhang Rui 93622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 9374690c4abSTejun Heo __queue_work(cpu, wq, work); 938c1a220e7SZhang Rui ret = 1; 939c1a220e7SZhang Rui } 940c1a220e7SZhang Rui return ret; 941c1a220e7SZhang Rui } 942c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 943c1a220e7SZhang Rui 9446d141c3fSLi Zefan static void delayed_work_timer_fn(unsigned long __data) 9451da177e4SLinus Torvalds { 94652bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 9477a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 9481da177e4SLinus Torvalds 9494690c4abSTejun Heo __queue_work(smp_processor_id(), cwq->wq, &dwork->work); 9501da177e4SLinus Torvalds } 9511da177e4SLinus Torvalds 9520fcb78c2SRolf Eike Beer /** 9530fcb78c2SRolf Eike Beer * queue_delayed_work - queue work on a workqueue after delay 9540fcb78c2SRolf Eike Beer * @wq: workqueue to use 955af9997e4SRandy Dunlap * @dwork: delayable work to queue 9560fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 9570fcb78c2SRolf Eike Beer * 958057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 9590fcb78c2SRolf Eike Beer */ 9607ad5b3a5SHarvey Harrison int queue_delayed_work(struct workqueue_struct *wq, 96152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 9621da177e4SLinus Torvalds { 96352bad64dSDavid Howells if (delay == 0) 96463bc0362SOleg Nesterov return queue_work(wq, &dwork->work); 9651da177e4SLinus Torvalds 96663bc0362SOleg Nesterov return queue_delayed_work_on(-1, wq, dwork, delay); 9671da177e4SLinus Torvalds } 968ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work); 9691da177e4SLinus Torvalds 9700fcb78c2SRolf Eike Beer /** 9710fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 9720fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 9730fcb78c2SRolf Eike Beer * @wq: workqueue to use 974af9997e4SRandy Dunlap * @dwork: work to queue 9750fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 9760fcb78c2SRolf Eike Beer * 977057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 9780fcb78c2SRolf Eike Beer */ 9797a6bc1cdSVenkatesh Pallipadi int queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 98052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 9817a6bc1cdSVenkatesh Pallipadi { 9827a6bc1cdSVenkatesh Pallipadi int ret = 0; 98352bad64dSDavid Howells struct timer_list *timer = &dwork->timer; 98452bad64dSDavid Howells struct work_struct *work = &dwork->work; 9857a6bc1cdSVenkatesh Pallipadi 98622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 9877a22ad75STejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 9887a22ad75STejun Heo unsigned int lcpu = gcwq ? gcwq->cpu : raw_smp_processor_id(); 9897a22ad75STejun Heo 9907a6bc1cdSVenkatesh Pallipadi BUG_ON(timer_pending(timer)); 9917a6bc1cdSVenkatesh Pallipadi BUG_ON(!list_empty(&work->entry)); 9927a6bc1cdSVenkatesh Pallipadi 9938a3e77ccSAndrew Liu timer_stats_timer_set_start_info(&dwork->timer); 9947a22ad75STejun Heo /* 9957a22ad75STejun Heo * This stores cwq for the moment, for the timer_fn. 9967a22ad75STejun Heo * Note that the work's gcwq is preserved to allow 9977a22ad75STejun Heo * reentrance detection for delayed works. 9987a22ad75STejun Heo */ 9997a22ad75STejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 10007a6bc1cdSVenkatesh Pallipadi timer->expires = jiffies + delay; 100152bad64dSDavid Howells timer->data = (unsigned long)dwork; 10027a6bc1cdSVenkatesh Pallipadi timer->function = delayed_work_timer_fn; 100363bc0362SOleg Nesterov 100463bc0362SOleg Nesterov if (unlikely(cpu >= 0)) 10057a6bc1cdSVenkatesh Pallipadi add_timer_on(timer, cpu); 100663bc0362SOleg Nesterov else 100763bc0362SOleg Nesterov add_timer(timer); 10087a6bc1cdSVenkatesh Pallipadi ret = 1; 10097a6bc1cdSVenkatesh Pallipadi } 10107a6bc1cdSVenkatesh Pallipadi return ret; 10117a6bc1cdSVenkatesh Pallipadi } 1012ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 10131da177e4SLinus Torvalds 1014c8e55f36STejun Heo /** 1015c8e55f36STejun Heo * worker_enter_idle - enter idle state 1016c8e55f36STejun Heo * @worker: worker which is entering idle state 1017c8e55f36STejun Heo * 1018c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1019c8e55f36STejun Heo * necessary. 1020c8e55f36STejun Heo * 1021c8e55f36STejun Heo * LOCKING: 1022c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1023c8e55f36STejun Heo */ 1024c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 1025c8e55f36STejun Heo { 1026c8e55f36STejun Heo struct global_cwq *gcwq = worker->gcwq; 1027c8e55f36STejun Heo 1028c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1029c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1030c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1031c8e55f36STejun Heo 1032d302f017STejun Heo worker_set_flags(worker, WORKER_IDLE, false); 1033c8e55f36STejun Heo gcwq->nr_idle++; 1034e22bee78STejun Heo worker->last_active = jiffies; 1035c8e55f36STejun Heo 1036c8e55f36STejun Heo /* idle_list is LIFO */ 1037c8e55f36STejun Heo list_add(&worker->entry, &gcwq->idle_list); 1038db7bccf4STejun Heo 1039e22bee78STejun Heo if (likely(!(worker->flags & WORKER_ROGUE))) { 1040e22bee78STejun Heo if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer)) 1041e22bee78STejun Heo mod_timer(&gcwq->idle_timer, 1042e22bee78STejun Heo jiffies + IDLE_WORKER_TIMEOUT); 1043e22bee78STejun Heo } else 1044db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 1045c8e55f36STejun Heo } 1046c8e55f36STejun Heo 1047c8e55f36STejun Heo /** 1048c8e55f36STejun Heo * worker_leave_idle - leave idle state 1049c8e55f36STejun Heo * @worker: worker which is leaving idle state 1050c8e55f36STejun Heo * 1051c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1052c8e55f36STejun Heo * 1053c8e55f36STejun Heo * LOCKING: 1054c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1055c8e55f36STejun Heo */ 1056c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1057c8e55f36STejun Heo { 1058c8e55f36STejun Heo struct global_cwq *gcwq = worker->gcwq; 1059c8e55f36STejun Heo 1060c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1061d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1062c8e55f36STejun Heo gcwq->nr_idle--; 1063c8e55f36STejun Heo list_del_init(&worker->entry); 1064c8e55f36STejun Heo } 1065c8e55f36STejun Heo 1066e22bee78STejun Heo /** 1067e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1068e22bee78STejun Heo * @worker: self 1069e22bee78STejun Heo * 1070e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1071e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1072e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1073e22bee78STejun Heo * guaranteed to execute on the cpu. 1074e22bee78STejun Heo * 1075e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1076e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1077e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1078e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1079e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1080e22bee78STejun Heo * [dis]associated in the meantime. 1081e22bee78STejun Heo * 1082e22bee78STejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies 1083e22bee78STejun Heo * the binding against GCWQ_DISASSOCIATED which is set during 1084e22bee78STejun Heo * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters 1085e22bee78STejun Heo * idle state or fetches works without dropping lock, it can guarantee 1086e22bee78STejun Heo * the scheduling requirement described in the first paragraph. 1087e22bee78STejun Heo * 1088e22bee78STejun Heo * CONTEXT: 1089e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1090e22bee78STejun Heo * held. 1091e22bee78STejun Heo * 1092e22bee78STejun Heo * RETURNS: 1093e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1094e22bee78STejun Heo * bound), %false if offline. 1095e22bee78STejun Heo */ 1096e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1097e22bee78STejun Heo { 1098e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1099e22bee78STejun Heo struct task_struct *task = worker->task; 1100e22bee78STejun Heo 1101e22bee78STejun Heo while (true) { 1102e22bee78STejun Heo /* 1103e22bee78STejun Heo * The following call may fail, succeed or succeed 1104e22bee78STejun Heo * without actually migrating the task to the cpu if 1105e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1106e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1107e22bee78STejun Heo */ 1108e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1109e22bee78STejun Heo 1110e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1111e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1112e22bee78STejun Heo return false; 1113e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1114e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1115e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1116e22bee78STejun Heo return true; 1117e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1118e22bee78STejun Heo 1119e22bee78STejun Heo /* CPU has come up inbetween, retry migration */ 1120e22bee78STejun Heo cpu_relax(); 1121e22bee78STejun Heo } 1122e22bee78STejun Heo } 1123e22bee78STejun Heo 1124e22bee78STejun Heo /* 1125e22bee78STejun Heo * Function for worker->rebind_work used to rebind rogue busy workers 1126e22bee78STejun Heo * to the associated cpu which is coming back online. This is 1127e22bee78STejun Heo * scheduled by cpu up but can race with other cpu hotplug operations 1128e22bee78STejun Heo * and may be executed twice without intervening cpu down. 1129e22bee78STejun Heo */ 1130e22bee78STejun Heo static void worker_rebind_fn(struct work_struct *work) 1131e22bee78STejun Heo { 1132e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1133e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1134e22bee78STejun Heo 1135e22bee78STejun Heo if (worker_maybe_bind_and_lock(worker)) 1136e22bee78STejun Heo worker_clr_flags(worker, WORKER_REBIND); 1137e22bee78STejun Heo 1138e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1139e22bee78STejun Heo } 1140e22bee78STejun Heo 1141c34056a3STejun Heo static struct worker *alloc_worker(void) 1142c34056a3STejun Heo { 1143c34056a3STejun Heo struct worker *worker; 1144c34056a3STejun Heo 1145c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1146c8e55f36STejun Heo if (worker) { 1147c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1148affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1149e22bee78STejun Heo INIT_WORK(&worker->rebind_work, worker_rebind_fn); 1150e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1151e22bee78STejun Heo worker->flags = WORKER_PREP; 1152c8e55f36STejun Heo } 1153c34056a3STejun Heo return worker; 1154c34056a3STejun Heo } 1155c34056a3STejun Heo 1156c34056a3STejun Heo /** 1157c34056a3STejun Heo * create_worker - create a new workqueue worker 11587e11629dSTejun Heo * @gcwq: gcwq the new worker will belong to 1159c34056a3STejun Heo * @bind: whether to set affinity to @cpu or not 1160c34056a3STejun Heo * 11617e11629dSTejun Heo * Create a new worker which is bound to @gcwq. The returned worker 1162c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1163c34056a3STejun Heo * destroy_worker(). 1164c34056a3STejun Heo * 1165c34056a3STejun Heo * CONTEXT: 1166c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1167c34056a3STejun Heo * 1168c34056a3STejun Heo * RETURNS: 1169c34056a3STejun Heo * Pointer to the newly created worker. 1170c34056a3STejun Heo */ 11717e11629dSTejun Heo static struct worker *create_worker(struct global_cwq *gcwq, bool bind) 1172c34056a3STejun Heo { 1173c34056a3STejun Heo int id = -1; 1174c34056a3STejun Heo struct worker *worker = NULL; 1175c34056a3STejun Heo 11768b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 11778b03ae3cSTejun Heo while (ida_get_new(&gcwq->worker_ida, &id)) { 11788b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 11798b03ae3cSTejun Heo if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL)) 1180c34056a3STejun Heo goto fail; 11818b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1182c34056a3STejun Heo } 11838b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1184c34056a3STejun Heo 1185c34056a3STejun Heo worker = alloc_worker(); 1186c34056a3STejun Heo if (!worker) 1187c34056a3STejun Heo goto fail; 1188c34056a3STejun Heo 11898b03ae3cSTejun Heo worker->gcwq = gcwq; 1190c34056a3STejun Heo worker->id = id; 1191c34056a3STejun Heo 1192c34056a3STejun Heo worker->task = kthread_create(worker_thread, worker, "kworker/%u:%d", 11938b03ae3cSTejun Heo gcwq->cpu, id); 1194c34056a3STejun Heo if (IS_ERR(worker->task)) 1195c34056a3STejun Heo goto fail; 1196c34056a3STejun Heo 1197db7bccf4STejun Heo /* 1198db7bccf4STejun Heo * A rogue worker will become a regular one if CPU comes 1199db7bccf4STejun Heo * online later on. Make sure every worker has 1200db7bccf4STejun Heo * PF_THREAD_BOUND set. 1201db7bccf4STejun Heo */ 1202c34056a3STejun Heo if (bind) 12038b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1204db7bccf4STejun Heo else 1205db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1206c34056a3STejun Heo 1207c34056a3STejun Heo return worker; 1208c34056a3STejun Heo fail: 1209c34056a3STejun Heo if (id >= 0) { 12108b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 12118b03ae3cSTejun Heo ida_remove(&gcwq->worker_ida, id); 12128b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1213c34056a3STejun Heo } 1214c34056a3STejun Heo kfree(worker); 1215c34056a3STejun Heo return NULL; 1216c34056a3STejun Heo } 1217c34056a3STejun Heo 1218c34056a3STejun Heo /** 1219c34056a3STejun Heo * start_worker - start a newly created worker 1220c34056a3STejun Heo * @worker: worker to start 1221c34056a3STejun Heo * 1222c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1223c34056a3STejun Heo * 1224c34056a3STejun Heo * CONTEXT: 12258b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1226c34056a3STejun Heo */ 1227c34056a3STejun Heo static void start_worker(struct worker *worker) 1228c34056a3STejun Heo { 1229d302f017STejun Heo worker_set_flags(worker, WORKER_STARTED, false); 1230c8e55f36STejun Heo worker->gcwq->nr_workers++; 1231c8e55f36STejun Heo worker_enter_idle(worker); 1232c34056a3STejun Heo wake_up_process(worker->task); 1233c34056a3STejun Heo } 1234c34056a3STejun Heo 1235c34056a3STejun Heo /** 1236c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1237c34056a3STejun Heo * @worker: worker to be destroyed 1238c34056a3STejun Heo * 1239c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1240c8e55f36STejun Heo * 1241c8e55f36STejun Heo * CONTEXT: 1242c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1243c34056a3STejun Heo */ 1244c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1245c34056a3STejun Heo { 12468b03ae3cSTejun Heo struct global_cwq *gcwq = worker->gcwq; 1247c34056a3STejun Heo int id = worker->id; 1248c34056a3STejun Heo 1249c34056a3STejun Heo /* sanity check frenzy */ 1250c34056a3STejun Heo BUG_ON(worker->current_work); 1251affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1252c34056a3STejun Heo 1253c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1254c8e55f36STejun Heo gcwq->nr_workers--; 1255c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1256c8e55f36STejun Heo gcwq->nr_idle--; 1257c8e55f36STejun Heo 1258c8e55f36STejun Heo list_del_init(&worker->entry); 1259d302f017STejun Heo worker_set_flags(worker, WORKER_DIE, false); 1260c8e55f36STejun Heo 1261c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1262c8e55f36STejun Heo 1263c34056a3STejun Heo kthread_stop(worker->task); 1264c34056a3STejun Heo kfree(worker); 1265c34056a3STejun Heo 12668b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 12678b03ae3cSTejun Heo ida_remove(&gcwq->worker_ida, id); 1268c34056a3STejun Heo } 1269c34056a3STejun Heo 1270e22bee78STejun Heo static void idle_worker_timeout(unsigned long __gcwq) 1271e22bee78STejun Heo { 1272e22bee78STejun Heo struct global_cwq *gcwq = (void *)__gcwq; 1273e22bee78STejun Heo 1274e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1275e22bee78STejun Heo 1276e22bee78STejun Heo if (too_many_workers(gcwq)) { 1277e22bee78STejun Heo struct worker *worker; 1278e22bee78STejun Heo unsigned long expires; 1279e22bee78STejun Heo 1280e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 1281e22bee78STejun Heo worker = list_entry(gcwq->idle_list.prev, struct worker, entry); 1282e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1283e22bee78STejun Heo 1284e22bee78STejun Heo if (time_before(jiffies, expires)) 1285e22bee78STejun Heo mod_timer(&gcwq->idle_timer, expires); 1286e22bee78STejun Heo else { 1287e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 1288e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGE_WORKERS; 1289e22bee78STejun Heo wake_up_worker(gcwq); 1290e22bee78STejun Heo } 1291e22bee78STejun Heo } 1292e22bee78STejun Heo 1293e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1294e22bee78STejun Heo } 1295e22bee78STejun Heo 1296e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1297e22bee78STejun Heo { 1298e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1299e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1300e22bee78STejun Heo 1301e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1302e22bee78STejun Heo return false; 1303e22bee78STejun Heo 1304e22bee78STejun Heo /* mayday mayday mayday */ 1305e22bee78STejun Heo if (!cpumask_test_and_set_cpu(cwq->gcwq->cpu, wq->mayday_mask)) 1306e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1307e22bee78STejun Heo return true; 1308e22bee78STejun Heo } 1309e22bee78STejun Heo 1310e22bee78STejun Heo static void gcwq_mayday_timeout(unsigned long __gcwq) 1311e22bee78STejun Heo { 1312e22bee78STejun Heo struct global_cwq *gcwq = (void *)__gcwq; 1313e22bee78STejun Heo struct work_struct *work; 1314e22bee78STejun Heo 1315e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1316e22bee78STejun Heo 1317e22bee78STejun Heo if (need_to_create_worker(gcwq)) { 1318e22bee78STejun Heo /* 1319e22bee78STejun Heo * We've been trying to create a new worker but 1320e22bee78STejun Heo * haven't been successful. We might be hitting an 1321e22bee78STejun Heo * allocation deadlock. Send distress signals to 1322e22bee78STejun Heo * rescuers. 1323e22bee78STejun Heo */ 1324e22bee78STejun Heo list_for_each_entry(work, &gcwq->worklist, entry) 1325e22bee78STejun Heo send_mayday(work); 1326e22bee78STejun Heo } 1327e22bee78STejun Heo 1328e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1329e22bee78STejun Heo 1330e22bee78STejun Heo mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL); 1331e22bee78STejun Heo } 1332e22bee78STejun Heo 1333e22bee78STejun Heo /** 1334e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 1335e22bee78STejun Heo * @gcwq: gcwq to create a new worker for 1336e22bee78STejun Heo * 1337e22bee78STejun Heo * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to 1338e22bee78STejun Heo * have at least one idle worker on return from this function. If 1339e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 1340e22bee78STejun Heo * sent to all rescuers with works scheduled on @gcwq to resolve 1341e22bee78STejun Heo * possible allocation deadlock. 1342e22bee78STejun Heo * 1343e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1344e22bee78STejun Heo * may_start_working() true. 1345e22bee78STejun Heo * 1346e22bee78STejun Heo * LOCKING: 1347e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1348e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1349e22bee78STejun Heo * manager. 1350e22bee78STejun Heo * 1351e22bee78STejun Heo * RETURNS: 1352e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1353e22bee78STejun Heo * otherwise. 1354e22bee78STejun Heo */ 1355e22bee78STejun Heo static bool maybe_create_worker(struct global_cwq *gcwq) 1356e22bee78STejun Heo { 1357e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1358e22bee78STejun Heo return false; 1359e22bee78STejun Heo restart: 1360e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 1361e22bee78STejun Heo mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1362e22bee78STejun Heo 1363e22bee78STejun Heo while (true) { 1364e22bee78STejun Heo struct worker *worker; 1365e22bee78STejun Heo 1366e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1367e22bee78STejun Heo 1368e22bee78STejun Heo worker = create_worker(gcwq, true); 1369e22bee78STejun Heo if (worker) { 1370e22bee78STejun Heo del_timer_sync(&gcwq->mayday_timer); 1371e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1372e22bee78STejun Heo start_worker(worker); 1373e22bee78STejun Heo BUG_ON(need_to_create_worker(gcwq)); 1374e22bee78STejun Heo return true; 1375e22bee78STejun Heo } 1376e22bee78STejun Heo 1377e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1378e22bee78STejun Heo break; 1379e22bee78STejun Heo 1380e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1381e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1382e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 1383e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1384e22bee78STejun Heo if (!need_to_create_worker(gcwq)) 1385e22bee78STejun Heo break; 1386e22bee78STejun Heo } 1387e22bee78STejun Heo 1388e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1389e22bee78STejun Heo del_timer_sync(&gcwq->mayday_timer); 1390e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1391e22bee78STejun Heo if (need_to_create_worker(gcwq)) 1392e22bee78STejun Heo goto restart; 1393e22bee78STejun Heo return true; 1394e22bee78STejun Heo } 1395e22bee78STejun Heo 1396e22bee78STejun Heo /** 1397e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 1398e22bee78STejun Heo * @gcwq: gcwq to destroy workers for 1399e22bee78STejun Heo * 1400e22bee78STejun Heo * Destroy @gcwq workers which have been idle for longer than 1401e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1402e22bee78STejun Heo * 1403e22bee78STejun Heo * LOCKING: 1404e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1405e22bee78STejun Heo * multiple times. Called only from manager. 1406e22bee78STejun Heo * 1407e22bee78STejun Heo * RETURNS: 1408e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1409e22bee78STejun Heo * otherwise. 1410e22bee78STejun Heo */ 1411e22bee78STejun Heo static bool maybe_destroy_workers(struct global_cwq *gcwq) 1412e22bee78STejun Heo { 1413e22bee78STejun Heo bool ret = false; 1414e22bee78STejun Heo 1415e22bee78STejun Heo while (too_many_workers(gcwq)) { 1416e22bee78STejun Heo struct worker *worker; 1417e22bee78STejun Heo unsigned long expires; 1418e22bee78STejun Heo 1419e22bee78STejun Heo worker = list_entry(gcwq->idle_list.prev, struct worker, entry); 1420e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1421e22bee78STejun Heo 1422e22bee78STejun Heo if (time_before(jiffies, expires)) { 1423e22bee78STejun Heo mod_timer(&gcwq->idle_timer, expires); 1424e22bee78STejun Heo break; 1425e22bee78STejun Heo } 1426e22bee78STejun Heo 1427e22bee78STejun Heo destroy_worker(worker); 1428e22bee78STejun Heo ret = true; 1429e22bee78STejun Heo } 1430e22bee78STejun Heo 1431e22bee78STejun Heo return ret; 1432e22bee78STejun Heo } 1433e22bee78STejun Heo 1434e22bee78STejun Heo /** 1435e22bee78STejun Heo * manage_workers - manage worker pool 1436e22bee78STejun Heo * @worker: self 1437e22bee78STejun Heo * 1438e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 1439e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1440e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 1441e22bee78STejun Heo * 1442e22bee78STejun Heo * The caller can safely start processing works on false return. On 1443e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1444e22bee78STejun Heo * and may_start_working() is true. 1445e22bee78STejun Heo * 1446e22bee78STejun Heo * CONTEXT: 1447e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1448e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1449e22bee78STejun Heo * 1450e22bee78STejun Heo * RETURNS: 1451e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 1452e22bee78STejun Heo * some action was taken. 1453e22bee78STejun Heo */ 1454e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1455e22bee78STejun Heo { 1456e22bee78STejun Heo struct global_cwq *gcwq = worker->gcwq; 1457e22bee78STejun Heo bool ret = false; 1458e22bee78STejun Heo 1459e22bee78STejun Heo if (gcwq->flags & GCWQ_MANAGING_WORKERS) 1460e22bee78STejun Heo return ret; 1461e22bee78STejun Heo 1462e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGE_WORKERS; 1463e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGING_WORKERS; 1464e22bee78STejun Heo 1465e22bee78STejun Heo /* 1466e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 1467e22bee78STejun Heo * on return. 1468e22bee78STejun Heo */ 1469e22bee78STejun Heo ret |= maybe_destroy_workers(gcwq); 1470e22bee78STejun Heo ret |= maybe_create_worker(gcwq); 1471e22bee78STejun Heo 1472e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGING_WORKERS; 1473e22bee78STejun Heo 1474e22bee78STejun Heo /* 1475e22bee78STejun Heo * The trustee might be waiting to take over the manager 1476e22bee78STejun Heo * position, tell it we're done. 1477e22bee78STejun Heo */ 1478e22bee78STejun Heo if (unlikely(gcwq->trustee)) 1479e22bee78STejun Heo wake_up_all(&gcwq->trustee_wait); 1480e22bee78STejun Heo 1481e22bee78STejun Heo return ret; 1482e22bee78STejun Heo } 1483e22bee78STejun Heo 1484a62428c0STejun Heo /** 1485affee4b2STejun Heo * move_linked_works - move linked works to a list 1486affee4b2STejun Heo * @work: start of series of works to be scheduled 1487affee4b2STejun Heo * @head: target list to append @work to 1488affee4b2STejun Heo * @nextp: out paramter for nested worklist walking 1489affee4b2STejun Heo * 1490affee4b2STejun Heo * Schedule linked works starting from @work to @head. Work series to 1491affee4b2STejun Heo * be scheduled starts at @work and includes any consecutive work with 1492affee4b2STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1493affee4b2STejun Heo * 1494affee4b2STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1495affee4b2STejun Heo * the last scheduled work. This allows move_linked_works() to be 1496affee4b2STejun Heo * nested inside outer list_for_each_entry_safe(). 1497affee4b2STejun Heo * 1498affee4b2STejun Heo * CONTEXT: 14998b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1500affee4b2STejun Heo */ 1501affee4b2STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1502affee4b2STejun Heo struct work_struct **nextp) 1503affee4b2STejun Heo { 1504affee4b2STejun Heo struct work_struct *n; 1505affee4b2STejun Heo 1506affee4b2STejun Heo /* 1507affee4b2STejun Heo * Linked worklist will always end before the end of the list, 1508affee4b2STejun Heo * use NULL for list head. 1509affee4b2STejun Heo */ 1510affee4b2STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1511affee4b2STejun Heo list_move_tail(&work->entry, head); 1512affee4b2STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1513affee4b2STejun Heo break; 1514affee4b2STejun Heo } 1515affee4b2STejun Heo 1516affee4b2STejun Heo /* 1517affee4b2STejun Heo * If we're already inside safe list traversal and have moved 1518affee4b2STejun Heo * multiple works to the scheduled queue, the next position 1519affee4b2STejun Heo * needs to be updated. 1520affee4b2STejun Heo */ 1521affee4b2STejun Heo if (nextp) 1522affee4b2STejun Heo *nextp = n; 1523affee4b2STejun Heo } 1524affee4b2STejun Heo 15251e19ffc6STejun Heo static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 15261e19ffc6STejun Heo { 15271e19ffc6STejun Heo struct work_struct *work = list_first_entry(&cwq->delayed_works, 15281e19ffc6STejun Heo struct work_struct, entry); 15291e19ffc6STejun Heo 15307e11629dSTejun Heo move_linked_works(work, &cwq->gcwq->worklist, NULL); 15311e19ffc6STejun Heo cwq->nr_active++; 15321e19ffc6STejun Heo } 15331e19ffc6STejun Heo 1534affee4b2STejun Heo /** 153573f53c4aSTejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 153673f53c4aSTejun Heo * @cwq: cwq of interest 153773f53c4aSTejun Heo * @color: color of work which left the queue 153873f53c4aSTejun Heo * 153973f53c4aSTejun Heo * A work either has completed or is removed from pending queue, 154073f53c4aSTejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 154173f53c4aSTejun Heo * 154273f53c4aSTejun Heo * CONTEXT: 15438b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 154473f53c4aSTejun Heo */ 154573f53c4aSTejun Heo static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color) 154673f53c4aSTejun Heo { 154773f53c4aSTejun Heo /* ignore uncolored works */ 154873f53c4aSTejun Heo if (color == WORK_NO_COLOR) 154973f53c4aSTejun Heo return; 155073f53c4aSTejun Heo 155173f53c4aSTejun Heo cwq->nr_in_flight[color]--; 15521e19ffc6STejun Heo cwq->nr_active--; 15531e19ffc6STejun Heo 1554502ca9d8STejun Heo if (!list_empty(&cwq->delayed_works)) { 15551e19ffc6STejun Heo /* one down, submit a delayed one */ 1556502ca9d8STejun Heo if (cwq->nr_active < cwq->max_active) 15571e19ffc6STejun Heo cwq_activate_first_delayed(cwq); 1558502ca9d8STejun Heo } else if (!cwq->nr_active && cwq->wq->flags & WQ_SINGLE_CPU) { 1559502ca9d8STejun Heo /* this was the last work, unbind from single cpu */ 1560502ca9d8STejun Heo cwq_unbind_single_cpu(cwq); 1561502ca9d8STejun Heo } 156273f53c4aSTejun Heo 156373f53c4aSTejun Heo /* is flush in progress and are we at the flushing tip? */ 156473f53c4aSTejun Heo if (likely(cwq->flush_color != color)) 156573f53c4aSTejun Heo return; 156673f53c4aSTejun Heo 156773f53c4aSTejun Heo /* are there still in-flight works? */ 156873f53c4aSTejun Heo if (cwq->nr_in_flight[color]) 156973f53c4aSTejun Heo return; 157073f53c4aSTejun Heo 157173f53c4aSTejun Heo /* this cwq is done, clear flush_color */ 157273f53c4aSTejun Heo cwq->flush_color = -1; 157373f53c4aSTejun Heo 157473f53c4aSTejun Heo /* 157573f53c4aSTejun Heo * If this was the last cwq, wake up the first flusher. It 157673f53c4aSTejun Heo * will handle the rest. 157773f53c4aSTejun Heo */ 157873f53c4aSTejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 157973f53c4aSTejun Heo complete(&cwq->wq->first_flusher->done); 158073f53c4aSTejun Heo } 158173f53c4aSTejun Heo 158273f53c4aSTejun Heo /** 1583a62428c0STejun Heo * process_one_work - process single work 1584c34056a3STejun Heo * @worker: self 1585a62428c0STejun Heo * @work: work to process 1586a62428c0STejun Heo * 1587a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1588a62428c0STejun Heo * process a single work including synchronization against and 1589a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1590a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1591a62428c0STejun Heo * call this function to process a work. 1592a62428c0STejun Heo * 1593a62428c0STejun Heo * CONTEXT: 15948b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1595a62428c0STejun Heo */ 1596c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 15971da177e4SLinus Torvalds { 15987e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 15998b03ae3cSTejun Heo struct global_cwq *gcwq = cwq->gcwq; 1600c8e55f36STejun Heo struct hlist_head *bwh = busy_worker_head(gcwq, work); 16016bb49e59SDavid Howells work_func_t f = work->func; 160273f53c4aSTejun Heo int work_color; 16037e11629dSTejun Heo struct worker *collision; 16044e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 16054e6045f1SJohannes Berg /* 1606a62428c0STejun Heo * It is permissible to free the struct work_struct from 1607a62428c0STejun Heo * inside the function that is called from it, this we need to 1608a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1609a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1610a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 16114e6045f1SJohannes Berg */ 16124e6045f1SJohannes Berg struct lockdep_map lockdep_map = work->lockdep_map; 16134e6045f1SJohannes Berg #endif 16147e11629dSTejun Heo /* 16157e11629dSTejun Heo * A single work shouldn't be executed concurrently by 16167e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 16177e11629dSTejun Heo * already processing the work. If so, defer the work to the 16187e11629dSTejun Heo * currently executing one. 16197e11629dSTejun Heo */ 16207e11629dSTejun Heo collision = __find_worker_executing_work(gcwq, bwh, work); 16217e11629dSTejun Heo if (unlikely(collision)) { 16227e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 16237e11629dSTejun Heo return; 16247e11629dSTejun Heo } 16257e11629dSTejun Heo 1626a62428c0STejun Heo /* claim and process */ 1627dc186ad7SThomas Gleixner debug_work_deactivate(work); 1628c8e55f36STejun Heo hlist_add_head(&worker->hentry, bwh); 1629c34056a3STejun Heo worker->current_work = work; 16308cca0eeaSTejun Heo worker->current_cwq = cwq; 163173f53c4aSTejun Heo work_color = get_work_color(work); 16327a22ad75STejun Heo 16337a22ad75STejun Heo /* record the current cpu number in the work data and dequeue */ 16347a22ad75STejun Heo set_work_cpu(work, gcwq->cpu); 1635a62428c0STejun Heo list_del_init(&work->entry); 1636a62428c0STejun Heo 16378b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 16381da177e4SLinus Torvalds 163923b2e599SOleg Nesterov work_clear_pending(work); 16403295f0efSIngo Molnar lock_map_acquire(&cwq->wq->lockdep_map); 16413295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 164265f27f38SDavid Howells f(work); 16433295f0efSIngo Molnar lock_map_release(&lockdep_map); 16443295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 16451da177e4SLinus Torvalds 1646d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 1647d5abe669SPeter Zijlstra printk(KERN_ERR "BUG: workqueue leaked lock or atomic: " 1648d5abe669SPeter Zijlstra "%s/0x%08x/%d\n", 1649a62428c0STejun Heo current->comm, preempt_count(), task_pid_nr(current)); 1650d5abe669SPeter Zijlstra printk(KERN_ERR " last function: "); 1651d5abe669SPeter Zijlstra print_symbol("%s\n", (unsigned long)f); 1652d5abe669SPeter Zijlstra debug_show_held_locks(current); 1653d5abe669SPeter Zijlstra dump_stack(); 1654d5abe669SPeter Zijlstra } 1655d5abe669SPeter Zijlstra 16568b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1657a62428c0STejun Heo 1658a62428c0STejun Heo /* we're done with it, release */ 1659c8e55f36STejun Heo hlist_del_init(&worker->hentry); 1660c34056a3STejun Heo worker->current_work = NULL; 16618cca0eeaSTejun Heo worker->current_cwq = NULL; 166273f53c4aSTejun Heo cwq_dec_nr_in_flight(cwq, work_color); 16631da177e4SLinus Torvalds } 1664a62428c0STejun Heo 1665affee4b2STejun Heo /** 1666affee4b2STejun Heo * process_scheduled_works - process scheduled works 1667affee4b2STejun Heo * @worker: self 1668affee4b2STejun Heo * 1669affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 1670affee4b2STejun Heo * may change while processing a work, so this function repeatedly 1671affee4b2STejun Heo * fetches a work from the top and executes it. 1672affee4b2STejun Heo * 1673affee4b2STejun Heo * CONTEXT: 16748b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1675affee4b2STejun Heo * multiple times. 1676affee4b2STejun Heo */ 1677affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 1678a62428c0STejun Heo { 1679affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 1680affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 1681a62428c0STejun Heo struct work_struct, entry); 1682c34056a3STejun Heo process_one_work(worker, work); 1683a62428c0STejun Heo } 16841da177e4SLinus Torvalds } 16851da177e4SLinus Torvalds 16864690c4abSTejun Heo /** 16874690c4abSTejun Heo * worker_thread - the worker thread function 1688c34056a3STejun Heo * @__worker: self 16894690c4abSTejun Heo * 1690e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 1691e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 1692e22bee78STejun Heo * their specific target workqueue. The only exception is works which 1693e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 1694e22bee78STejun Heo * rescuer_thread(). 16954690c4abSTejun Heo */ 1696c34056a3STejun Heo static int worker_thread(void *__worker) 16971da177e4SLinus Torvalds { 1698c34056a3STejun Heo struct worker *worker = __worker; 16998b03ae3cSTejun Heo struct global_cwq *gcwq = worker->gcwq; 17001da177e4SLinus Torvalds 1701e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 1702e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 1703c8e55f36STejun Heo woke_up: 17048b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1705affee4b2STejun Heo 1706c8e55f36STejun Heo /* DIE can be set only while we're idle, checking here is enough */ 1707c8e55f36STejun Heo if (worker->flags & WORKER_DIE) { 1708c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1709e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 1710c8e55f36STejun Heo return 0; 1711c8e55f36STejun Heo } 1712c8e55f36STejun Heo 1713c8e55f36STejun Heo worker_leave_idle(worker); 1714db7bccf4STejun Heo recheck: 1715e22bee78STejun Heo /* no more worker necessary? */ 1716e22bee78STejun Heo if (!need_more_worker(gcwq)) 1717e22bee78STejun Heo goto sleep; 1718e22bee78STejun Heo 1719e22bee78STejun Heo /* do we need to manage? */ 1720e22bee78STejun Heo if (unlikely(!may_start_working(gcwq)) && manage_workers(worker)) 1721e22bee78STejun Heo goto recheck; 1722e22bee78STejun Heo 1723c8e55f36STejun Heo /* 1724c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 1725c8e55f36STejun Heo * preparing to process a work or actually processing it. 1726c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 1727c8e55f36STejun Heo */ 1728c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1729c8e55f36STejun Heo 1730e22bee78STejun Heo /* 1731e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 1732e22bee78STejun Heo * at least one idle worker or that someone else has already 1733e22bee78STejun Heo * assumed the manager role. 1734e22bee78STejun Heo */ 1735e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 1736e22bee78STejun Heo 1737e22bee78STejun Heo do { 1738affee4b2STejun Heo struct work_struct *work = 17397e11629dSTejun Heo list_first_entry(&gcwq->worklist, 1740affee4b2STejun Heo struct work_struct, entry); 1741affee4b2STejun Heo 1742c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 1743affee4b2STejun Heo /* optimization path, not strictly necessary */ 1744affee4b2STejun Heo process_one_work(worker, work); 1745affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 1746affee4b2STejun Heo process_scheduled_works(worker); 1747affee4b2STejun Heo } else { 1748c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 1749affee4b2STejun Heo process_scheduled_works(worker); 1750affee4b2STejun Heo } 1751e22bee78STejun Heo } while (keep_working(gcwq)); 1752affee4b2STejun Heo 1753e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 1754e22bee78STejun Heo 1755e22bee78STejun Heo if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker)) 1756e22bee78STejun Heo goto recheck; 1757e22bee78STejun Heo sleep: 1758c8e55f36STejun Heo /* 1759e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 1760e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 1761e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 1762e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 1763e22bee78STejun Heo * prevent losing any event. 1764c8e55f36STejun Heo */ 1765c8e55f36STejun Heo worker_enter_idle(worker); 1766c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 17678b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1768c8e55f36STejun Heo schedule(); 1769c8e55f36STejun Heo goto woke_up; 17701da177e4SLinus Torvalds } 17711da177e4SLinus Torvalds 1772e22bee78STejun Heo /** 1773e22bee78STejun Heo * rescuer_thread - the rescuer thread function 1774e22bee78STejun Heo * @__wq: the associated workqueue 1775e22bee78STejun Heo * 1776e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 1777e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 1778e22bee78STejun Heo * 1779e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 1780e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 1781e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 1782e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 1783e22bee78STejun Heo * the problem rescuer solves. 1784e22bee78STejun Heo * 1785e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 1786e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 1787e22bee78STejun Heo * those works so that forward progress can be guaranteed. 1788e22bee78STejun Heo * 1789e22bee78STejun Heo * This should happen rarely. 1790e22bee78STejun Heo */ 1791e22bee78STejun Heo static int rescuer_thread(void *__wq) 1792e22bee78STejun Heo { 1793e22bee78STejun Heo struct workqueue_struct *wq = __wq; 1794e22bee78STejun Heo struct worker *rescuer = wq->rescuer; 1795e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 1796e22bee78STejun Heo unsigned int cpu; 1797e22bee78STejun Heo 1798e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 1799e22bee78STejun Heo repeat: 1800e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 1801e22bee78STejun Heo 1802e22bee78STejun Heo if (kthread_should_stop()) 1803e22bee78STejun Heo return 0; 1804e22bee78STejun Heo 1805e22bee78STejun Heo for_each_cpu(cpu, wq->mayday_mask) { 1806e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 1807e22bee78STejun Heo struct global_cwq *gcwq = cwq->gcwq; 1808e22bee78STejun Heo struct work_struct *work, *n; 1809e22bee78STejun Heo 1810e22bee78STejun Heo __set_current_state(TASK_RUNNING); 1811e22bee78STejun Heo cpumask_clear_cpu(cpu, wq->mayday_mask); 1812e22bee78STejun Heo 1813e22bee78STejun Heo /* migrate to the target cpu if possible */ 1814e22bee78STejun Heo rescuer->gcwq = gcwq; 1815e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 1816e22bee78STejun Heo 1817e22bee78STejun Heo /* 1818e22bee78STejun Heo * Slurp in all works issued via this workqueue and 1819e22bee78STejun Heo * process'em. 1820e22bee78STejun Heo */ 1821e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 1822e22bee78STejun Heo list_for_each_entry_safe(work, n, &gcwq->worklist, entry) 1823e22bee78STejun Heo if (get_work_cwq(work) == cwq) 1824e22bee78STejun Heo move_linked_works(work, scheduled, &n); 1825e22bee78STejun Heo 1826e22bee78STejun Heo process_scheduled_works(rescuer); 1827e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1828e22bee78STejun Heo } 1829e22bee78STejun Heo 1830e22bee78STejun Heo schedule(); 1831e22bee78STejun Heo goto repeat; 1832e22bee78STejun Heo } 1833e22bee78STejun Heo 1834fc2e4d70SOleg Nesterov struct wq_barrier { 1835fc2e4d70SOleg Nesterov struct work_struct work; 1836fc2e4d70SOleg Nesterov struct completion done; 1837fc2e4d70SOleg Nesterov }; 1838fc2e4d70SOleg Nesterov 1839fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 1840fc2e4d70SOleg Nesterov { 1841fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 1842fc2e4d70SOleg Nesterov complete(&barr->done); 1843fc2e4d70SOleg Nesterov } 1844fc2e4d70SOleg Nesterov 18454690c4abSTejun Heo /** 18464690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 18474690c4abSTejun Heo * @cwq: cwq to insert barrier into 18484690c4abSTejun Heo * @barr: wq_barrier to insert 1849affee4b2STejun Heo * @target: target work to attach @barr to 1850affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 18514690c4abSTejun Heo * 1852affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 1853affee4b2STejun Heo * @target finishes execution. Please note that the ordering 1854affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 1855affee4b2STejun Heo * cpu. 1856affee4b2STejun Heo * 1857affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 1858affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 1859affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 1860affee4b2STejun Heo * flag of the previous work while there must be a valid next work 1861affee4b2STejun Heo * after a work with LINKED flag set. 1862affee4b2STejun Heo * 1863affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 1864affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 18654690c4abSTejun Heo * 18664690c4abSTejun Heo * CONTEXT: 18678b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 18684690c4abSTejun Heo */ 186983c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 1870affee4b2STejun Heo struct wq_barrier *barr, 1871affee4b2STejun Heo struct work_struct *target, struct worker *worker) 1872fc2e4d70SOleg Nesterov { 1873affee4b2STejun Heo struct list_head *head; 1874affee4b2STejun Heo unsigned int linked = 0; 1875affee4b2STejun Heo 1876dc186ad7SThomas Gleixner /* 18778b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 1878dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 1879dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 1880dc186ad7SThomas Gleixner * might deadlock. 1881dc186ad7SThomas Gleixner */ 1882dc186ad7SThomas Gleixner INIT_WORK_ON_STACK(&barr->work, wq_barrier_func); 188322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 1884fc2e4d70SOleg Nesterov init_completion(&barr->done); 188583c22520SOleg Nesterov 1886affee4b2STejun Heo /* 1887affee4b2STejun Heo * If @target is currently being executed, schedule the 1888affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 1889affee4b2STejun Heo */ 1890affee4b2STejun Heo if (worker) 1891affee4b2STejun Heo head = worker->scheduled.next; 1892affee4b2STejun Heo else { 1893affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 1894affee4b2STejun Heo 1895affee4b2STejun Heo head = target->entry.next; 1896affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 1897affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 1898affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 1899affee4b2STejun Heo } 1900affee4b2STejun Heo 1901dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 1902affee4b2STejun Heo insert_work(cwq, &barr->work, head, 1903affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 1904fc2e4d70SOleg Nesterov } 1905fc2e4d70SOleg Nesterov 190673f53c4aSTejun Heo /** 190773f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 190873f53c4aSTejun Heo * @wq: workqueue being flushed 190973f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 191073f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 191173f53c4aSTejun Heo * 191273f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 191373f53c4aSTejun Heo * 191473f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 191573f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 191673f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 191773f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 191873f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 191973f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 192073f53c4aSTejun Heo * 192173f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 192273f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 192373f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 192473f53c4aSTejun Heo * is returned. 192573f53c4aSTejun Heo * 192673f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 192773f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 192873f53c4aSTejun Heo * advanced to @work_color. 192973f53c4aSTejun Heo * 193073f53c4aSTejun Heo * CONTEXT: 193173f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 193273f53c4aSTejun Heo * 193373f53c4aSTejun Heo * RETURNS: 193473f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 193573f53c4aSTejun Heo * otherwise. 193673f53c4aSTejun Heo */ 193773f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 193873f53c4aSTejun Heo int flush_color, int work_color) 19391da177e4SLinus Torvalds { 194073f53c4aSTejun Heo bool wait = false; 194173f53c4aSTejun Heo unsigned int cpu; 19421da177e4SLinus Torvalds 194373f53c4aSTejun Heo if (flush_color >= 0) { 194473f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 194573f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 194673f53c4aSTejun Heo } 194773f53c4aSTejun Heo 194873f53c4aSTejun Heo for_each_possible_cpu(cpu) { 194973f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 19508b03ae3cSTejun Heo struct global_cwq *gcwq = cwq->gcwq; 19512355b70fSLai Jiangshan 19528b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 195373f53c4aSTejun Heo 195473f53c4aSTejun Heo if (flush_color >= 0) { 195573f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 195673f53c4aSTejun Heo 195773f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 195873f53c4aSTejun Heo cwq->flush_color = flush_color; 195973f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 196073f53c4aSTejun Heo wait = true; 196183c22520SOleg Nesterov } 196273f53c4aSTejun Heo } 196373f53c4aSTejun Heo 196473f53c4aSTejun Heo if (work_color >= 0) { 196573f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 196673f53c4aSTejun Heo cwq->work_color = work_color; 196773f53c4aSTejun Heo } 196873f53c4aSTejun Heo 19698b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1970dc186ad7SThomas Gleixner } 197114441960SOleg Nesterov 197273f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 197373f53c4aSTejun Heo complete(&wq->first_flusher->done); 197473f53c4aSTejun Heo 197573f53c4aSTejun Heo return wait; 197683c22520SOleg Nesterov } 19771da177e4SLinus Torvalds 19780fcb78c2SRolf Eike Beer /** 19791da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 19800fcb78c2SRolf Eike Beer * @wq: workqueue to flush 19811da177e4SLinus Torvalds * 19821da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 19831da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 19841da177e4SLinus Torvalds * 1985fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 1986fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 19871da177e4SLinus Torvalds */ 19887ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 19891da177e4SLinus Torvalds { 199073f53c4aSTejun Heo struct wq_flusher this_flusher = { 199173f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 199273f53c4aSTejun Heo .flush_color = -1, 199373f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 199473f53c4aSTejun Heo }; 199573f53c4aSTejun Heo int next_color; 1996b1f4ec17SOleg Nesterov 19973295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 19983295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 199973f53c4aSTejun Heo 200073f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 200173f53c4aSTejun Heo 200273f53c4aSTejun Heo /* 200373f53c4aSTejun Heo * Start-to-wait phase 200473f53c4aSTejun Heo */ 200573f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 200673f53c4aSTejun Heo 200773f53c4aSTejun Heo if (next_color != wq->flush_color) { 200873f53c4aSTejun Heo /* 200973f53c4aSTejun Heo * Color space is not full. The current work_color 201073f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 201173f53c4aSTejun Heo * by one. 201273f53c4aSTejun Heo */ 201373f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 201473f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 201573f53c4aSTejun Heo wq->work_color = next_color; 201673f53c4aSTejun Heo 201773f53c4aSTejun Heo if (!wq->first_flusher) { 201873f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 201973f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 202073f53c4aSTejun Heo 202173f53c4aSTejun Heo wq->first_flusher = &this_flusher; 202273f53c4aSTejun Heo 202373f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 202473f53c4aSTejun Heo wq->work_color)) { 202573f53c4aSTejun Heo /* nothing to flush, done */ 202673f53c4aSTejun Heo wq->flush_color = next_color; 202773f53c4aSTejun Heo wq->first_flusher = NULL; 202873f53c4aSTejun Heo goto out_unlock; 202973f53c4aSTejun Heo } 203073f53c4aSTejun Heo } else { 203173f53c4aSTejun Heo /* wait in queue */ 203273f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 203373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 203473f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 203573f53c4aSTejun Heo } 203673f53c4aSTejun Heo } else { 203773f53c4aSTejun Heo /* 203873f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 203973f53c4aSTejun Heo * The next flush completion will assign us 204073f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 204173f53c4aSTejun Heo */ 204273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 204373f53c4aSTejun Heo } 204473f53c4aSTejun Heo 204573f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 204673f53c4aSTejun Heo 204773f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 204873f53c4aSTejun Heo 204973f53c4aSTejun Heo /* 205073f53c4aSTejun Heo * Wake-up-and-cascade phase 205173f53c4aSTejun Heo * 205273f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 205373f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 205473f53c4aSTejun Heo */ 205573f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 205673f53c4aSTejun Heo return; 205773f53c4aSTejun Heo 205873f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 205973f53c4aSTejun Heo 206073f53c4aSTejun Heo wq->first_flusher = NULL; 206173f53c4aSTejun Heo 206273f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 206373f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 206473f53c4aSTejun Heo 206573f53c4aSTejun Heo while (true) { 206673f53c4aSTejun Heo struct wq_flusher *next, *tmp; 206773f53c4aSTejun Heo 206873f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 206973f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 207073f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 207173f53c4aSTejun Heo break; 207273f53c4aSTejun Heo list_del_init(&next->list); 207373f53c4aSTejun Heo complete(&next->done); 207473f53c4aSTejun Heo } 207573f53c4aSTejun Heo 207673f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 207773f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 207873f53c4aSTejun Heo 207973f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 208073f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 208173f53c4aSTejun Heo 208273f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 208373f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 208473f53c4aSTejun Heo /* 208573f53c4aSTejun Heo * Assign the same color to all overflowed 208673f53c4aSTejun Heo * flushers, advance work_color and append to 208773f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 208873f53c4aSTejun Heo * phase for these overflowed flushers. 208973f53c4aSTejun Heo */ 209073f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 209173f53c4aSTejun Heo tmp->flush_color = wq->work_color; 209273f53c4aSTejun Heo 209373f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 209473f53c4aSTejun Heo 209573f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 209673f53c4aSTejun Heo &wq->flusher_queue); 209773f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 209873f53c4aSTejun Heo } 209973f53c4aSTejun Heo 210073f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 210173f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 210273f53c4aSTejun Heo break; 210373f53c4aSTejun Heo } 210473f53c4aSTejun Heo 210573f53c4aSTejun Heo /* 210673f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 210773f53c4aSTejun Heo * the new first flusher and arm cwqs. 210873f53c4aSTejun Heo */ 210973f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 211073f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 211173f53c4aSTejun Heo 211273f53c4aSTejun Heo list_del_init(&next->list); 211373f53c4aSTejun Heo wq->first_flusher = next; 211473f53c4aSTejun Heo 211573f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 211673f53c4aSTejun Heo break; 211773f53c4aSTejun Heo 211873f53c4aSTejun Heo /* 211973f53c4aSTejun Heo * Meh... this color is already done, clear first 212073f53c4aSTejun Heo * flusher and repeat cascading. 212173f53c4aSTejun Heo */ 212273f53c4aSTejun Heo wq->first_flusher = NULL; 212373f53c4aSTejun Heo } 212473f53c4aSTejun Heo 212573f53c4aSTejun Heo out_unlock: 212673f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 21271da177e4SLinus Torvalds } 2128ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 21291da177e4SLinus Torvalds 2130db700897SOleg Nesterov /** 2131db700897SOleg Nesterov * flush_work - block until a work_struct's callback has terminated 2132db700897SOleg Nesterov * @work: the work which is to be flushed 2133db700897SOleg Nesterov * 2134a67da70dSOleg Nesterov * Returns false if @work has already terminated. 2135a67da70dSOleg Nesterov * 2136db700897SOleg Nesterov * It is expected that, prior to calling flush_work(), the caller has 2137db700897SOleg Nesterov * arranged for the work to not be requeued, otherwise it doesn't make 2138db700897SOleg Nesterov * sense to use this function. 2139db700897SOleg Nesterov */ 2140db700897SOleg Nesterov int flush_work(struct work_struct *work) 2141db700897SOleg Nesterov { 2142affee4b2STejun Heo struct worker *worker = NULL; 21438b03ae3cSTejun Heo struct global_cwq *gcwq; 21447a22ad75STejun Heo struct cpu_workqueue_struct *cwq; 2145db700897SOleg Nesterov struct wq_barrier barr; 2146db700897SOleg Nesterov 2147db700897SOleg Nesterov might_sleep(); 21487a22ad75STejun Heo gcwq = get_work_gcwq(work); 21497a22ad75STejun Heo if (!gcwq) 2150db700897SOleg Nesterov return 0; 2151a67da70dSOleg Nesterov 21528b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2153db700897SOleg Nesterov if (!list_empty(&work->entry)) { 2154db700897SOleg Nesterov /* 2155db700897SOleg Nesterov * See the comment near try_to_grab_pending()->smp_rmb(). 21567a22ad75STejun Heo * If it was re-queued to a different gcwq under us, we 21577a22ad75STejun Heo * are not going to wait. 2158db700897SOleg Nesterov */ 2159db700897SOleg Nesterov smp_rmb(); 21607a22ad75STejun Heo cwq = get_work_cwq(work); 21617a22ad75STejun Heo if (unlikely(!cwq || gcwq != cwq->gcwq)) 21624690c4abSTejun Heo goto already_gone; 2163db700897SOleg Nesterov } else { 21647a22ad75STejun Heo worker = find_worker_executing_work(gcwq, work); 2165affee4b2STejun Heo if (!worker) 21664690c4abSTejun Heo goto already_gone; 21677a22ad75STejun Heo cwq = worker->current_cwq; 2168db700897SOleg Nesterov } 2169db700897SOleg Nesterov 2170affee4b2STejun Heo insert_wq_barrier(cwq, &barr, work, worker); 21718b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 21727a22ad75STejun Heo 21737a22ad75STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 21747a22ad75STejun Heo lock_map_release(&cwq->wq->lockdep_map); 21757a22ad75STejun Heo 2176db700897SOleg Nesterov wait_for_completion(&barr.done); 2177dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2178db700897SOleg Nesterov return 1; 21794690c4abSTejun Heo already_gone: 21808b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 21814690c4abSTejun Heo return 0; 2182db700897SOleg Nesterov } 2183db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2184db700897SOleg Nesterov 21856e84d644SOleg Nesterov /* 21861f1f642eSOleg Nesterov * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit, 21876e84d644SOleg Nesterov * so this work can't be re-armed in any way. 21886e84d644SOleg Nesterov */ 21896e84d644SOleg Nesterov static int try_to_grab_pending(struct work_struct *work) 21906e84d644SOleg Nesterov { 21918b03ae3cSTejun Heo struct global_cwq *gcwq; 21921f1f642eSOleg Nesterov int ret = -1; 21936e84d644SOleg Nesterov 219422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 21951f1f642eSOleg Nesterov return 0; 21966e84d644SOleg Nesterov 21976e84d644SOleg Nesterov /* 21986e84d644SOleg Nesterov * The queueing is in progress, or it is already queued. Try to 21996e84d644SOleg Nesterov * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 22006e84d644SOleg Nesterov */ 22017a22ad75STejun Heo gcwq = get_work_gcwq(work); 22027a22ad75STejun Heo if (!gcwq) 22036e84d644SOleg Nesterov return ret; 22046e84d644SOleg Nesterov 22058b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 22066e84d644SOleg Nesterov if (!list_empty(&work->entry)) { 22076e84d644SOleg Nesterov /* 22087a22ad75STejun Heo * This work is queued, but perhaps we locked the wrong gcwq. 22096e84d644SOleg Nesterov * In that case we must see the new value after rmb(), see 22106e84d644SOleg Nesterov * insert_work()->wmb(). 22116e84d644SOleg Nesterov */ 22126e84d644SOleg Nesterov smp_rmb(); 22137a22ad75STejun Heo if (gcwq == get_work_gcwq(work)) { 2214dc186ad7SThomas Gleixner debug_work_deactivate(work); 22156e84d644SOleg Nesterov list_del_init(&work->entry); 22167a22ad75STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 22177a22ad75STejun Heo get_work_color(work)); 22186e84d644SOleg Nesterov ret = 1; 22196e84d644SOleg Nesterov } 22206e84d644SOleg Nesterov } 22218b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 22226e84d644SOleg Nesterov 22236e84d644SOleg Nesterov return ret; 22246e84d644SOleg Nesterov } 22256e84d644SOleg Nesterov 22267a22ad75STejun Heo static void wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work) 2227b89deed3SOleg Nesterov { 2228b89deed3SOleg Nesterov struct wq_barrier barr; 2229affee4b2STejun Heo struct worker *worker; 2230b89deed3SOleg Nesterov 22318b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2232affee4b2STejun Heo 22337a22ad75STejun Heo worker = find_worker_executing_work(gcwq, work); 22347a22ad75STejun Heo if (unlikely(worker)) 22357a22ad75STejun Heo insert_wq_barrier(worker->current_cwq, &barr, work, worker); 2236affee4b2STejun Heo 22378b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 2238b89deed3SOleg Nesterov 2239affee4b2STejun Heo if (unlikely(worker)) { 2240b89deed3SOleg Nesterov wait_for_completion(&barr.done); 2241dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2242dc186ad7SThomas Gleixner } 2243b89deed3SOleg Nesterov } 2244b89deed3SOleg Nesterov 22456e84d644SOleg Nesterov static void wait_on_work(struct work_struct *work) 2246b89deed3SOleg Nesterov { 2247b1f4ec17SOleg Nesterov int cpu; 2248b89deed3SOleg Nesterov 2249f293ea92SOleg Nesterov might_sleep(); 2250f293ea92SOleg Nesterov 22513295f0efSIngo Molnar lock_map_acquire(&work->lockdep_map); 22523295f0efSIngo Molnar lock_map_release(&work->lockdep_map); 22534e6045f1SJohannes Berg 22541537663fSTejun Heo for_each_possible_cpu(cpu) 22557a22ad75STejun Heo wait_on_cpu_work(get_gcwq(cpu), work); 22566e84d644SOleg Nesterov } 22576e84d644SOleg Nesterov 22581f1f642eSOleg Nesterov static int __cancel_work_timer(struct work_struct *work, 22591f1f642eSOleg Nesterov struct timer_list* timer) 22601f1f642eSOleg Nesterov { 22611f1f642eSOleg Nesterov int ret; 22621f1f642eSOleg Nesterov 22631f1f642eSOleg Nesterov do { 22641f1f642eSOleg Nesterov ret = (timer && likely(del_timer(timer))); 22651f1f642eSOleg Nesterov if (!ret) 22661f1f642eSOleg Nesterov ret = try_to_grab_pending(work); 22671f1f642eSOleg Nesterov wait_on_work(work); 22681f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 22691f1f642eSOleg Nesterov 22707a22ad75STejun Heo clear_work_data(work); 22711f1f642eSOleg Nesterov return ret; 22721f1f642eSOleg Nesterov } 22731f1f642eSOleg Nesterov 22746e84d644SOleg Nesterov /** 22756e84d644SOleg Nesterov * cancel_work_sync - block until a work_struct's callback has terminated 22766e84d644SOleg Nesterov * @work: the work which is to be flushed 22776e84d644SOleg Nesterov * 22781f1f642eSOleg Nesterov * Returns true if @work was pending. 22791f1f642eSOleg Nesterov * 22806e84d644SOleg Nesterov * cancel_work_sync() will cancel the work if it is queued. If the work's 22816e84d644SOleg Nesterov * callback appears to be running, cancel_work_sync() will block until it 22826e84d644SOleg Nesterov * has completed. 22836e84d644SOleg Nesterov * 22846e84d644SOleg Nesterov * It is possible to use this function if the work re-queues itself. It can 22856e84d644SOleg Nesterov * cancel the work even if it migrates to another workqueue, however in that 22866e84d644SOleg Nesterov * case it only guarantees that work->func() has completed on the last queued 22876e84d644SOleg Nesterov * workqueue. 22886e84d644SOleg Nesterov * 22896e84d644SOleg Nesterov * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not 22906e84d644SOleg Nesterov * pending, otherwise it goes into a busy-wait loop until the timer expires. 22916e84d644SOleg Nesterov * 22926e84d644SOleg Nesterov * The caller must ensure that workqueue_struct on which this work was last 22936e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 22946e84d644SOleg Nesterov */ 22951f1f642eSOleg Nesterov int cancel_work_sync(struct work_struct *work) 22966e84d644SOleg Nesterov { 22971f1f642eSOleg Nesterov return __cancel_work_timer(work, NULL); 2298b89deed3SOleg Nesterov } 229928e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2300b89deed3SOleg Nesterov 23016e84d644SOleg Nesterov /** 2302f5a421a4SOleg Nesterov * cancel_delayed_work_sync - reliably kill off a delayed work. 23036e84d644SOleg Nesterov * @dwork: the delayed work struct 23046e84d644SOleg Nesterov * 23051f1f642eSOleg Nesterov * Returns true if @dwork was pending. 23061f1f642eSOleg Nesterov * 23076e84d644SOleg Nesterov * It is possible to use this function if @dwork rearms itself via queue_work() 23086e84d644SOleg Nesterov * or queue_delayed_work(). See also the comment for cancel_work_sync(). 23096e84d644SOleg Nesterov */ 23101f1f642eSOleg Nesterov int cancel_delayed_work_sync(struct delayed_work *dwork) 23116e84d644SOleg Nesterov { 23121f1f642eSOleg Nesterov return __cancel_work_timer(&dwork->work, &dwork->timer); 23136e84d644SOleg Nesterov } 2314f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 23151da177e4SLinus Torvalds 23160fcb78c2SRolf Eike Beer /** 23170fcb78c2SRolf Eike Beer * schedule_work - put work task in global workqueue 23180fcb78c2SRolf Eike Beer * @work: job to be done 23190fcb78c2SRolf Eike Beer * 23205b0f437dSBart Van Assche * Returns zero if @work was already on the kernel-global workqueue and 23215b0f437dSBart Van Assche * non-zero otherwise. 23225b0f437dSBart Van Assche * 23235b0f437dSBart Van Assche * This puts a job in the kernel-global workqueue if it was not already 23245b0f437dSBart Van Assche * queued and leaves it in the same position on the kernel-global 23255b0f437dSBart Van Assche * workqueue otherwise. 23260fcb78c2SRolf Eike Beer */ 23277ad5b3a5SHarvey Harrison int schedule_work(struct work_struct *work) 23281da177e4SLinus Torvalds { 2329d320c038STejun Heo return queue_work(system_wq, work); 23301da177e4SLinus Torvalds } 2331ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_work); 23321da177e4SLinus Torvalds 2333c1a220e7SZhang Rui /* 2334c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2335c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2336c1a220e7SZhang Rui * @work: job to be done 2337c1a220e7SZhang Rui * 2338c1a220e7SZhang Rui * This puts a job on a specific cpu 2339c1a220e7SZhang Rui */ 2340c1a220e7SZhang Rui int schedule_work_on(int cpu, struct work_struct *work) 2341c1a220e7SZhang Rui { 2342d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2343c1a220e7SZhang Rui } 2344c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2345c1a220e7SZhang Rui 23460fcb78c2SRolf Eike Beer /** 23470fcb78c2SRolf Eike Beer * schedule_delayed_work - put work task in global workqueue after delay 234852bad64dSDavid Howells * @dwork: job to be done 234952bad64dSDavid Howells * @delay: number of jiffies to wait or 0 for immediate execution 23500fcb78c2SRolf Eike Beer * 23510fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 23520fcb78c2SRolf Eike Beer * workqueue. 23530fcb78c2SRolf Eike Beer */ 23547ad5b3a5SHarvey Harrison int schedule_delayed_work(struct delayed_work *dwork, 235582f67cd9SIngo Molnar unsigned long delay) 23561da177e4SLinus Torvalds { 2357d320c038STejun Heo return queue_delayed_work(system_wq, dwork, delay); 23581da177e4SLinus Torvalds } 2359ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work); 23601da177e4SLinus Torvalds 23610fcb78c2SRolf Eike Beer /** 23628c53e463SLinus Torvalds * flush_delayed_work - block until a dwork_struct's callback has terminated 23638c53e463SLinus Torvalds * @dwork: the delayed work which is to be flushed 23648c53e463SLinus Torvalds * 23658c53e463SLinus Torvalds * Any timeout is cancelled, and any pending work is run immediately. 23668c53e463SLinus Torvalds */ 23678c53e463SLinus Torvalds void flush_delayed_work(struct delayed_work *dwork) 23688c53e463SLinus Torvalds { 23698c53e463SLinus Torvalds if (del_timer_sync(&dwork->timer)) { 23707a22ad75STejun Heo __queue_work(get_cpu(), get_work_cwq(&dwork->work)->wq, 23714690c4abSTejun Heo &dwork->work); 23728c53e463SLinus Torvalds put_cpu(); 23738c53e463SLinus Torvalds } 23748c53e463SLinus Torvalds flush_work(&dwork->work); 23758c53e463SLinus Torvalds } 23768c53e463SLinus Torvalds EXPORT_SYMBOL(flush_delayed_work); 23778c53e463SLinus Torvalds 23788c53e463SLinus Torvalds /** 23790fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 23800fcb78c2SRolf Eike Beer * @cpu: cpu to use 238152bad64dSDavid Howells * @dwork: job to be done 23820fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 23830fcb78c2SRolf Eike Beer * 23840fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 23850fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 23860fcb78c2SRolf Eike Beer */ 23871da177e4SLinus Torvalds int schedule_delayed_work_on(int cpu, 238852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 23891da177e4SLinus Torvalds { 2390d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 23911da177e4SLinus Torvalds } 2392ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 23931da177e4SLinus Torvalds 2394b6136773SAndrew Morton /** 2395b6136773SAndrew Morton * schedule_on_each_cpu - call a function on each online CPU from keventd 2396b6136773SAndrew Morton * @func: the function to call 2397b6136773SAndrew Morton * 2398b6136773SAndrew Morton * Returns zero on success. 2399b6136773SAndrew Morton * Returns -ve errno on failure. 2400b6136773SAndrew Morton * 2401b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 2402b6136773SAndrew Morton */ 240365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 240415316ba8SChristoph Lameter { 240515316ba8SChristoph Lameter int cpu; 2406b6136773SAndrew Morton struct work_struct *works; 240715316ba8SChristoph Lameter 2408b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2409b6136773SAndrew Morton if (!works) 241015316ba8SChristoph Lameter return -ENOMEM; 2411b6136773SAndrew Morton 241295402b38SGautham R Shenoy get_online_cpus(); 241393981800STejun Heo 241415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 24159bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 24169bfb1839SIngo Molnar 24179bfb1839SIngo Molnar INIT_WORK(work, func); 24188de6d308SOleg Nesterov schedule_work_on(cpu, work); 241915316ba8SChristoph Lameter } 242093981800STejun Heo 242193981800STejun Heo for_each_online_cpu(cpu) 24228616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 242393981800STejun Heo 242495402b38SGautham R Shenoy put_online_cpus(); 2425b6136773SAndrew Morton free_percpu(works); 242615316ba8SChristoph Lameter return 0; 242715316ba8SChristoph Lameter } 242815316ba8SChristoph Lameter 2429eef6a7d5SAlan Stern /** 2430eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2431eef6a7d5SAlan Stern * 2432eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2433eef6a7d5SAlan Stern * completion. 2434eef6a7d5SAlan Stern * 2435eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2436eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2437eef6a7d5SAlan Stern * will lead to deadlock: 2438eef6a7d5SAlan Stern * 2439eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2440eef6a7d5SAlan Stern * a lock held by your code or its caller. 2441eef6a7d5SAlan Stern * 2442eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2443eef6a7d5SAlan Stern * 2444eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2445eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2446eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2447eef6a7d5SAlan Stern * 2448eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2449eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2450eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2451eef6a7d5SAlan Stern * cancel_work_sync() instead. 2452eef6a7d5SAlan Stern */ 24531da177e4SLinus Torvalds void flush_scheduled_work(void) 24541da177e4SLinus Torvalds { 2455d320c038STejun Heo flush_workqueue(system_wq); 24561da177e4SLinus Torvalds } 2457ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 24581da177e4SLinus Torvalds 24591da177e4SLinus Torvalds /** 24601fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 24611fa44ecaSJames Bottomley * @fn: the function to execute 24621fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 24631fa44ecaSJames Bottomley * be available when the work executes) 24641fa44ecaSJames Bottomley * 24651fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 24661fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 24671fa44ecaSJames Bottomley * 24681fa44ecaSJames Bottomley * Returns: 0 - function was executed 24691fa44ecaSJames Bottomley * 1 - function was scheduled for execution 24701fa44ecaSJames Bottomley */ 247165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 24721fa44ecaSJames Bottomley { 24731fa44ecaSJames Bottomley if (!in_interrupt()) { 247465f27f38SDavid Howells fn(&ew->work); 24751fa44ecaSJames Bottomley return 0; 24761fa44ecaSJames Bottomley } 24771fa44ecaSJames Bottomley 247865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 24791fa44ecaSJames Bottomley schedule_work(&ew->work); 24801fa44ecaSJames Bottomley 24811fa44ecaSJames Bottomley return 1; 24821fa44ecaSJames Bottomley } 24831fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 24841fa44ecaSJames Bottomley 24851da177e4SLinus Torvalds int keventd_up(void) 24861da177e4SLinus Torvalds { 2487d320c038STejun Heo return system_wq != NULL; 24881da177e4SLinus Torvalds } 24891da177e4SLinus Torvalds 24900f900049STejun Heo static struct cpu_workqueue_struct *alloc_cwqs(void) 24910f900049STejun Heo { 24920f900049STejun Heo /* 24930f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 24940f900049STejun Heo * Make sure that the alignment isn't lower than that of 24950f900049STejun Heo * unsigned long long. 24960f900049STejun Heo */ 24970f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 24980f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 24990f900049STejun Heo __alignof__(unsigned long long)); 25000f900049STejun Heo struct cpu_workqueue_struct *cwqs; 25010f900049STejun Heo #ifndef CONFIG_SMP 25020f900049STejun Heo void *ptr; 25030f900049STejun Heo 25040f900049STejun Heo /* 25050f900049STejun Heo * On UP, percpu allocator doesn't honor alignment parameter 25060f900049STejun Heo * and simply uses arch-dependent default. Allocate enough 25070f900049STejun Heo * room to align cwq and put an extra pointer at the end 25080f900049STejun Heo * pointing back to the originally allocated pointer which 25090f900049STejun Heo * will be used for free. 25100f900049STejun Heo * 25110f900049STejun Heo * FIXME: This really belongs to UP percpu code. Update UP 25120f900049STejun Heo * percpu code to honor alignment and remove this ugliness. 25130f900049STejun Heo */ 25140f900049STejun Heo ptr = __alloc_percpu(size + align + sizeof(void *), 1); 25150f900049STejun Heo cwqs = PTR_ALIGN(ptr, align); 25160f900049STejun Heo *(void **)per_cpu_ptr(cwqs + 1, 0) = ptr; 25170f900049STejun Heo #else 25180f900049STejun Heo /* On SMP, percpu allocator can do it itself */ 25190f900049STejun Heo cwqs = __alloc_percpu(size, align); 25200f900049STejun Heo #endif 25210f900049STejun Heo /* just in case, make sure it's actually aligned */ 25220f900049STejun Heo BUG_ON(!IS_ALIGNED((unsigned long)cwqs, align)); 25230f900049STejun Heo return cwqs; 25240f900049STejun Heo } 25250f900049STejun Heo 25260f900049STejun Heo static void free_cwqs(struct cpu_workqueue_struct *cwqs) 25270f900049STejun Heo { 25280f900049STejun Heo #ifndef CONFIG_SMP 25290f900049STejun Heo /* on UP, the pointer to free is stored right after the cwq */ 25300f900049STejun Heo if (cwqs) 25310f900049STejun Heo free_percpu(*(void **)per_cpu_ptr(cwqs + 1, 0)); 25320f900049STejun Heo #else 25330f900049STejun Heo free_percpu(cwqs); 25340f900049STejun Heo #endif 25350f900049STejun Heo } 25360f900049STejun Heo 2537b71ab8c2STejun Heo static int wq_clamp_max_active(int max_active, const char *name) 2538b71ab8c2STejun Heo { 2539b71ab8c2STejun Heo if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 2540b71ab8c2STejun Heo printk(KERN_WARNING "workqueue: max_active %d requested for %s " 2541b71ab8c2STejun Heo "is out of range, clamping between %d and %d\n", 2542b71ab8c2STejun Heo max_active, name, 1, WQ_MAX_ACTIVE); 2543b71ab8c2STejun Heo 2544b71ab8c2STejun Heo return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 2545b71ab8c2STejun Heo } 2546b71ab8c2STejun Heo 2547d320c038STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *name, 254897e37d7bSTejun Heo unsigned int flags, 25491e19ffc6STejun Heo int max_active, 2550eb13ba87SJohannes Berg struct lock_class_key *key, 2551eb13ba87SJohannes Berg const char *lock_name) 25523af24433SOleg Nesterov { 25533af24433SOleg Nesterov struct workqueue_struct *wq; 2554c34056a3STejun Heo unsigned int cpu; 25553af24433SOleg Nesterov 2556d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 2557b71ab8c2STejun Heo max_active = wq_clamp_max_active(max_active, name); 25581e19ffc6STejun Heo 25593af24433SOleg Nesterov wq = kzalloc(sizeof(*wq), GFP_KERNEL); 25603af24433SOleg Nesterov if (!wq) 25614690c4abSTejun Heo goto err; 25623af24433SOleg Nesterov 25630f900049STejun Heo wq->cpu_wq = alloc_cwqs(); 25644690c4abSTejun Heo if (!wq->cpu_wq) 25654690c4abSTejun Heo goto err; 25663af24433SOleg Nesterov 256797e37d7bSTejun Heo wq->flags = flags; 2568a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 256973f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 257073f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 257173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 257273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 2573502ca9d8STejun Heo wq->single_cpu = NR_CPUS; 2574502ca9d8STejun Heo 25753af24433SOleg Nesterov wq->name = name; 2576eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 2577cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 25783af24433SOleg Nesterov 25793af24433SOleg Nesterov for_each_possible_cpu(cpu) { 25801537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 25818b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 25821537663fSTejun Heo 25830f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 25848b03ae3cSTejun Heo cwq->gcwq = gcwq; 2585c34056a3STejun Heo cwq->wq = wq; 258673f53c4aSTejun Heo cwq->flush_color = -1; 25871e19ffc6STejun Heo cwq->max_active = max_active; 25881e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 2589e22bee78STejun Heo } 25901537663fSTejun Heo 2591e22bee78STejun Heo if (flags & WQ_RESCUER) { 2592e22bee78STejun Heo struct worker *rescuer; 2593e22bee78STejun Heo 2594e22bee78STejun Heo if (!alloc_cpumask_var(&wq->mayday_mask, GFP_KERNEL)) 2595e22bee78STejun Heo goto err; 2596e22bee78STejun Heo 2597e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 2598e22bee78STejun Heo if (!rescuer) 2599e22bee78STejun Heo goto err; 2600e22bee78STejun Heo 2601e22bee78STejun Heo rescuer->task = kthread_create(rescuer_thread, wq, "%s", name); 2602e22bee78STejun Heo if (IS_ERR(rescuer->task)) 2603e22bee78STejun Heo goto err; 2604e22bee78STejun Heo 2605e22bee78STejun Heo wq->rescuer = rescuer; 2606e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 2607e22bee78STejun Heo wake_up_process(rescuer->task); 26083af24433SOleg Nesterov } 26091537663fSTejun Heo 2610a0a1a5fdSTejun Heo /* 2611a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 2612a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 2613a0a1a5fdSTejun Heo * workqueue to workqueues list. 2614a0a1a5fdSTejun Heo */ 26151537663fSTejun Heo spin_lock(&workqueue_lock); 2616a0a1a5fdSTejun Heo 2617a0a1a5fdSTejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZEABLE) 2618a0a1a5fdSTejun Heo for_each_possible_cpu(cpu) 2619a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 2620a0a1a5fdSTejun Heo 26211537663fSTejun Heo list_add(&wq->list, &workqueues); 2622a0a1a5fdSTejun Heo 26231537663fSTejun Heo spin_unlock(&workqueue_lock); 26241537663fSTejun Heo 26253af24433SOleg Nesterov return wq; 26264690c4abSTejun Heo err: 26274690c4abSTejun Heo if (wq) { 26280f900049STejun Heo free_cwqs(wq->cpu_wq); 2629e22bee78STejun Heo free_cpumask_var(wq->mayday_mask); 2630e22bee78STejun Heo kfree(wq->rescuer); 26314690c4abSTejun Heo kfree(wq); 26324690c4abSTejun Heo } 26334690c4abSTejun Heo return NULL; 26343af24433SOleg Nesterov } 2635d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 26363af24433SOleg Nesterov 26373af24433SOleg Nesterov /** 26383af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 26393af24433SOleg Nesterov * @wq: target workqueue 26403af24433SOleg Nesterov * 26413af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 26423af24433SOleg Nesterov */ 26433af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 26443af24433SOleg Nesterov { 2645c8e55f36STejun Heo unsigned int cpu; 26463af24433SOleg Nesterov 2647a0a1a5fdSTejun Heo flush_workqueue(wq); 2648a0a1a5fdSTejun Heo 2649a0a1a5fdSTejun Heo /* 2650a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 2651a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 2652a0a1a5fdSTejun Heo */ 265395402b38SGautham R Shenoy spin_lock(&workqueue_lock); 26543af24433SOleg Nesterov list_del(&wq->list); 265595402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 26563af24433SOleg Nesterov 2657e22bee78STejun Heo /* sanity check */ 265873f53c4aSTejun Heo for_each_possible_cpu(cpu) { 265973f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 266073f53c4aSTejun Heo int i; 266173f53c4aSTejun Heo 266273f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 266373f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 26641e19ffc6STejun Heo BUG_ON(cwq->nr_active); 26651e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 266673f53c4aSTejun Heo } 26671537663fSTejun Heo 2668e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 2669e22bee78STejun Heo kthread_stop(wq->rescuer->task); 2670e22bee78STejun Heo free_cpumask_var(wq->mayday_mask); 2671e22bee78STejun Heo } 2672e22bee78STejun Heo 26730f900049STejun Heo free_cwqs(wq->cpu_wq); 26743af24433SOleg Nesterov kfree(wq); 26753af24433SOleg Nesterov } 26763af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 26773af24433SOleg Nesterov 2678dcd989cbSTejun Heo /** 2679dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 2680dcd989cbSTejun Heo * @wq: target workqueue 2681dcd989cbSTejun Heo * @max_active: new max_active value. 2682dcd989cbSTejun Heo * 2683dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 2684dcd989cbSTejun Heo * 2685dcd989cbSTejun Heo * CONTEXT: 2686dcd989cbSTejun Heo * Don't call from IRQ context. 2687dcd989cbSTejun Heo */ 2688dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 2689dcd989cbSTejun Heo { 2690dcd989cbSTejun Heo unsigned int cpu; 2691dcd989cbSTejun Heo 2692dcd989cbSTejun Heo max_active = wq_clamp_max_active(max_active, wq->name); 2693dcd989cbSTejun Heo 2694dcd989cbSTejun Heo spin_lock(&workqueue_lock); 2695dcd989cbSTejun Heo 2696dcd989cbSTejun Heo wq->saved_max_active = max_active; 2697dcd989cbSTejun Heo 2698dcd989cbSTejun Heo for_each_possible_cpu(cpu) { 2699dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 2700dcd989cbSTejun Heo 2701dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 2702dcd989cbSTejun Heo 2703dcd989cbSTejun Heo if (!(wq->flags & WQ_FREEZEABLE) || 2704dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 2705dcd989cbSTejun Heo get_cwq(gcwq->cpu, wq)->max_active = max_active; 2706dcd989cbSTejun Heo 2707dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 2708dcd989cbSTejun Heo } 2709dcd989cbSTejun Heo 2710dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 2711dcd989cbSTejun Heo } 2712dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 2713dcd989cbSTejun Heo 2714dcd989cbSTejun Heo /** 2715dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 2716dcd989cbSTejun Heo * @cpu: CPU in question 2717dcd989cbSTejun Heo * @wq: target workqueue 2718dcd989cbSTejun Heo * 2719dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 2720dcd989cbSTejun Heo * no synchronization around this function and the test result is 2721dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 2722dcd989cbSTejun Heo * 2723dcd989cbSTejun Heo * RETURNS: 2724dcd989cbSTejun Heo * %true if congested, %false otherwise. 2725dcd989cbSTejun Heo */ 2726dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 2727dcd989cbSTejun Heo { 2728dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2729dcd989cbSTejun Heo 2730dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 2731dcd989cbSTejun Heo } 2732dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 2733dcd989cbSTejun Heo 2734dcd989cbSTejun Heo /** 2735dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 2736dcd989cbSTejun Heo * @work: the work of interest 2737dcd989cbSTejun Heo * 2738dcd989cbSTejun Heo * RETURNS: 2739dcd989cbSTejun Heo * CPU number if @work was ever queued. NR_CPUS otherwise. 2740dcd989cbSTejun Heo */ 2741dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 2742dcd989cbSTejun Heo { 2743dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 2744dcd989cbSTejun Heo 2745dcd989cbSTejun Heo return gcwq ? gcwq->cpu : NR_CPUS; 2746dcd989cbSTejun Heo } 2747dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 2748dcd989cbSTejun Heo 2749dcd989cbSTejun Heo /** 2750dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 2751dcd989cbSTejun Heo * @work: the work to be tested 2752dcd989cbSTejun Heo * 2753dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 2754dcd989cbSTejun Heo * synchronization around this function and the test result is 2755dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 2756dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 2757dcd989cbSTejun Heo * running state. 2758dcd989cbSTejun Heo * 2759dcd989cbSTejun Heo * RETURNS: 2760dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 2761dcd989cbSTejun Heo */ 2762dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 2763dcd989cbSTejun Heo { 2764dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 2765dcd989cbSTejun Heo unsigned long flags; 2766dcd989cbSTejun Heo unsigned int ret = 0; 2767dcd989cbSTejun Heo 2768dcd989cbSTejun Heo if (!gcwq) 2769dcd989cbSTejun Heo return false; 2770dcd989cbSTejun Heo 2771dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 2772dcd989cbSTejun Heo 2773dcd989cbSTejun Heo if (work_pending(work)) 2774dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 2775dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 2776dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 2777dcd989cbSTejun Heo 2778dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 2779dcd989cbSTejun Heo 2780dcd989cbSTejun Heo return ret; 2781dcd989cbSTejun Heo } 2782dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 2783dcd989cbSTejun Heo 2784db7bccf4STejun Heo /* 2785db7bccf4STejun Heo * CPU hotplug. 2786db7bccf4STejun Heo * 2787e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 2788e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 2789e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 2790e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 2791e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 2792e22bee78STejun Heo * blocked draining impractical. 2793e22bee78STejun Heo * 2794e22bee78STejun Heo * This is solved by allowing a gcwq to be detached from CPU, running 2795e22bee78STejun Heo * it with unbound (rogue) workers and allowing it to be reattached 2796e22bee78STejun Heo * later if the cpu comes back online. A separate thread is created 2797e22bee78STejun Heo * to govern a gcwq in such state and is called the trustee of the 2798e22bee78STejun Heo * gcwq. 2799db7bccf4STejun Heo * 2800db7bccf4STejun Heo * Trustee states and their descriptions. 2801db7bccf4STejun Heo * 2802db7bccf4STejun Heo * START Command state used on startup. On CPU_DOWN_PREPARE, a 2803db7bccf4STejun Heo * new trustee is started with this state. 2804db7bccf4STejun Heo * 2805db7bccf4STejun Heo * IN_CHARGE Once started, trustee will enter this state after 2806e22bee78STejun Heo * assuming the manager role and making all existing 2807e22bee78STejun Heo * workers rogue. DOWN_PREPARE waits for trustee to 2808e22bee78STejun Heo * enter this state. After reaching IN_CHARGE, trustee 2809e22bee78STejun Heo * tries to execute the pending worklist until it's empty 2810e22bee78STejun Heo * and the state is set to BUTCHER, or the state is set 2811e22bee78STejun Heo * to RELEASE. 2812db7bccf4STejun Heo * 2813db7bccf4STejun Heo * BUTCHER Command state which is set by the cpu callback after 2814db7bccf4STejun Heo * the cpu has went down. Once this state is set trustee 2815db7bccf4STejun Heo * knows that there will be no new works on the worklist 2816db7bccf4STejun Heo * and once the worklist is empty it can proceed to 2817db7bccf4STejun Heo * killing idle workers. 2818db7bccf4STejun Heo * 2819db7bccf4STejun Heo * RELEASE Command state which is set by the cpu callback if the 2820db7bccf4STejun Heo * cpu down has been canceled or it has come online 2821db7bccf4STejun Heo * again. After recognizing this state, trustee stops 2822e22bee78STejun Heo * trying to drain or butcher and clears ROGUE, rebinds 2823e22bee78STejun Heo * all remaining workers back to the cpu and releases 2824e22bee78STejun Heo * manager role. 2825db7bccf4STejun Heo * 2826db7bccf4STejun Heo * DONE Trustee will enter this state after BUTCHER or RELEASE 2827db7bccf4STejun Heo * is complete. 2828db7bccf4STejun Heo * 2829db7bccf4STejun Heo * trustee CPU draining 2830db7bccf4STejun Heo * took over down complete 2831db7bccf4STejun Heo * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE 2832db7bccf4STejun Heo * | | ^ 2833db7bccf4STejun Heo * | CPU is back online v return workers | 2834db7bccf4STejun Heo * ----------------> RELEASE -------------- 2835db7bccf4STejun Heo */ 2836db7bccf4STejun Heo 2837db7bccf4STejun Heo /** 2838db7bccf4STejun Heo * trustee_wait_event_timeout - timed event wait for trustee 2839db7bccf4STejun Heo * @cond: condition to wait for 2840db7bccf4STejun Heo * @timeout: timeout in jiffies 2841db7bccf4STejun Heo * 2842db7bccf4STejun Heo * wait_event_timeout() for trustee to use. Handles locking and 2843db7bccf4STejun Heo * checks for RELEASE request. 2844db7bccf4STejun Heo * 2845db7bccf4STejun Heo * CONTEXT: 2846db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2847db7bccf4STejun Heo * multiple times. To be used by trustee. 2848db7bccf4STejun Heo * 2849db7bccf4STejun Heo * RETURNS: 2850db7bccf4STejun Heo * Positive indicating left time if @cond is satisfied, 0 if timed 2851db7bccf4STejun Heo * out, -1 if canceled. 2852db7bccf4STejun Heo */ 2853db7bccf4STejun Heo #define trustee_wait_event_timeout(cond, timeout) ({ \ 2854db7bccf4STejun Heo long __ret = (timeout); \ 2855db7bccf4STejun Heo while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \ 2856db7bccf4STejun Heo __ret) { \ 2857db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); \ 2858db7bccf4STejun Heo __wait_event_timeout(gcwq->trustee_wait, (cond) || \ 2859db7bccf4STejun Heo (gcwq->trustee_state == TRUSTEE_RELEASE), \ 2860db7bccf4STejun Heo __ret); \ 2861db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); \ 2862db7bccf4STejun Heo } \ 2863db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \ 2864db7bccf4STejun Heo }) 2865db7bccf4STejun Heo 2866db7bccf4STejun Heo /** 2867db7bccf4STejun Heo * trustee_wait_event - event wait for trustee 2868db7bccf4STejun Heo * @cond: condition to wait for 2869db7bccf4STejun Heo * 2870db7bccf4STejun Heo * wait_event() for trustee to use. Automatically handles locking and 2871db7bccf4STejun Heo * checks for CANCEL request. 2872db7bccf4STejun Heo * 2873db7bccf4STejun Heo * CONTEXT: 2874db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2875db7bccf4STejun Heo * multiple times. To be used by trustee. 2876db7bccf4STejun Heo * 2877db7bccf4STejun Heo * RETURNS: 2878db7bccf4STejun Heo * 0 if @cond is satisfied, -1 if canceled. 2879db7bccf4STejun Heo */ 2880db7bccf4STejun Heo #define trustee_wait_event(cond) ({ \ 2881db7bccf4STejun Heo long __ret1; \ 2882db7bccf4STejun Heo __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\ 2883db7bccf4STejun Heo __ret1 < 0 ? -1 : 0; \ 2884db7bccf4STejun Heo }) 2885db7bccf4STejun Heo 2886db7bccf4STejun Heo static int __cpuinit trustee_thread(void *__gcwq) 2887db7bccf4STejun Heo { 2888db7bccf4STejun Heo struct global_cwq *gcwq = __gcwq; 2889db7bccf4STejun Heo struct worker *worker; 2890e22bee78STejun Heo struct work_struct *work; 2891db7bccf4STejun Heo struct hlist_node *pos; 2892e22bee78STejun Heo long rc; 2893db7bccf4STejun Heo int i; 2894db7bccf4STejun Heo 2895db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 2896db7bccf4STejun Heo 2897db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 2898db7bccf4STejun Heo /* 2899e22bee78STejun Heo * Claim the manager position and make all workers rogue. 2900e22bee78STejun Heo * Trustee must be bound to the target cpu and can't be 2901e22bee78STejun Heo * cancelled. 2902db7bccf4STejun Heo */ 2903db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 2904e22bee78STejun Heo rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS)); 2905e22bee78STejun Heo BUG_ON(rc < 0); 2906e22bee78STejun Heo 2907e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGING_WORKERS; 2908db7bccf4STejun Heo 2909db7bccf4STejun Heo list_for_each_entry(worker, &gcwq->idle_list, entry) 2910d302f017STejun Heo worker_set_flags(worker, WORKER_ROGUE, false); 2911db7bccf4STejun Heo 2912db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 2913d302f017STejun Heo worker_set_flags(worker, WORKER_ROGUE, false); 2914db7bccf4STejun Heo 2915db7bccf4STejun Heo /* 2916e22bee78STejun Heo * Call schedule() so that we cross rq->lock and thus can 2917e22bee78STejun Heo * guarantee sched callbacks see the rogue flag. This is 2918e22bee78STejun Heo * necessary as scheduler callbacks may be invoked from other 2919e22bee78STejun Heo * cpus. 2920e22bee78STejun Heo */ 2921e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 2922e22bee78STejun Heo schedule(); 2923e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 2924e22bee78STejun Heo 2925e22bee78STejun Heo /* 2926e22bee78STejun Heo * Sched callbacks are disabled now. gcwq->nr_running should 2927e22bee78STejun Heo * be zero and will stay that way, making need_more_worker() 2928e22bee78STejun Heo * and keep_working() always return true as long as the 2929e22bee78STejun Heo * worklist is not empty. 2930e22bee78STejun Heo */ 2931e22bee78STejun Heo WARN_ON_ONCE(atomic_read(get_gcwq_nr_running(gcwq->cpu)) != 0); 2932e22bee78STejun Heo 2933e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 2934e22bee78STejun Heo del_timer_sync(&gcwq->idle_timer); 2935e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 2936e22bee78STejun Heo 2937e22bee78STejun Heo /* 2938db7bccf4STejun Heo * We're now in charge. Notify and proceed to drain. We need 2939db7bccf4STejun Heo * to keep the gcwq running during the whole CPU down 2940db7bccf4STejun Heo * procedure as other cpu hotunplug callbacks may need to 2941db7bccf4STejun Heo * flush currently running tasks. 2942db7bccf4STejun Heo */ 2943db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_IN_CHARGE; 2944db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 2945db7bccf4STejun Heo 2946db7bccf4STejun Heo /* 2947db7bccf4STejun Heo * The original cpu is in the process of dying and may go away 2948db7bccf4STejun Heo * anytime now. When that happens, we and all workers would 2949e22bee78STejun Heo * be migrated to other cpus. Try draining any left work. We 2950e22bee78STejun Heo * want to get it over with ASAP - spam rescuers, wake up as 2951e22bee78STejun Heo * many idlers as necessary and create new ones till the 2952e22bee78STejun Heo * worklist is empty. Note that if the gcwq is frozen, there 2953e22bee78STejun Heo * may be frozen works in freezeable cwqs. Don't declare 2954e22bee78STejun Heo * completion while frozen. 2955db7bccf4STejun Heo */ 2956db7bccf4STejun Heo while (gcwq->nr_workers != gcwq->nr_idle || 2957db7bccf4STejun Heo gcwq->flags & GCWQ_FREEZING || 2958db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_IN_CHARGE) { 2959e22bee78STejun Heo int nr_works = 0; 2960e22bee78STejun Heo 2961e22bee78STejun Heo list_for_each_entry(work, &gcwq->worklist, entry) { 2962e22bee78STejun Heo send_mayday(work); 2963e22bee78STejun Heo nr_works++; 2964e22bee78STejun Heo } 2965e22bee78STejun Heo 2966e22bee78STejun Heo list_for_each_entry(worker, &gcwq->idle_list, entry) { 2967e22bee78STejun Heo if (!nr_works--) 2968e22bee78STejun Heo break; 2969e22bee78STejun Heo wake_up_process(worker->task); 2970e22bee78STejun Heo } 2971e22bee78STejun Heo 2972e22bee78STejun Heo if (need_to_create_worker(gcwq)) { 2973e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 2974e22bee78STejun Heo worker = create_worker(gcwq, false); 2975e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 2976e22bee78STejun Heo if (worker) { 2977e22bee78STejun Heo worker_set_flags(worker, WORKER_ROGUE, false); 2978e22bee78STejun Heo start_worker(worker); 2979e22bee78STejun Heo } 2980e22bee78STejun Heo } 2981e22bee78STejun Heo 2982db7bccf4STejun Heo /* give a breather */ 2983db7bccf4STejun Heo if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0) 2984db7bccf4STejun Heo break; 2985db7bccf4STejun Heo } 2986db7bccf4STejun Heo 2987e22bee78STejun Heo /* 2988e22bee78STejun Heo * Either all works have been scheduled and cpu is down, or 2989e22bee78STejun Heo * cpu down has already been canceled. Wait for and butcher 2990e22bee78STejun Heo * all workers till we're canceled. 2991e22bee78STejun Heo */ 2992e22bee78STejun Heo do { 2993e22bee78STejun Heo rc = trustee_wait_event(!list_empty(&gcwq->idle_list)); 2994e22bee78STejun Heo while (!list_empty(&gcwq->idle_list)) 2995e22bee78STejun Heo destroy_worker(list_first_entry(&gcwq->idle_list, 2996e22bee78STejun Heo struct worker, entry)); 2997e22bee78STejun Heo } while (gcwq->nr_workers && rc >= 0); 2998e22bee78STejun Heo 2999e22bee78STejun Heo /* 3000e22bee78STejun Heo * At this point, either draining has completed and no worker 3001e22bee78STejun Heo * is left, or cpu down has been canceled or the cpu is being 3002e22bee78STejun Heo * brought back up. There shouldn't be any idle one left. 3003e22bee78STejun Heo * Tell the remaining busy ones to rebind once it finishes the 3004e22bee78STejun Heo * currently scheduled works by scheduling the rebind_work. 3005e22bee78STejun Heo */ 3006e22bee78STejun Heo WARN_ON(!list_empty(&gcwq->idle_list)); 3007e22bee78STejun Heo 3008e22bee78STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 3009e22bee78STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 3010e22bee78STejun Heo 3011e22bee78STejun Heo /* 3012e22bee78STejun Heo * Rebind_work may race with future cpu hotplug 3013e22bee78STejun Heo * operations. Use a separate flag to mark that 3014e22bee78STejun Heo * rebinding is scheduled. 3015e22bee78STejun Heo */ 3016e22bee78STejun Heo worker_set_flags(worker, WORKER_REBIND, false); 3017e22bee78STejun Heo worker_clr_flags(worker, WORKER_ROGUE); 3018e22bee78STejun Heo 3019e22bee78STejun Heo /* queue rebind_work, wq doesn't matter, use the default one */ 3020e22bee78STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 3021e22bee78STejun Heo work_data_bits(rebind_work))) 3022e22bee78STejun Heo continue; 3023e22bee78STejun Heo 3024e22bee78STejun Heo debug_work_activate(rebind_work); 3025d320c038STejun Heo insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work, 3026e22bee78STejun Heo worker->scheduled.next, 3027e22bee78STejun Heo work_color_to_flags(WORK_NO_COLOR)); 3028e22bee78STejun Heo } 3029e22bee78STejun Heo 3030e22bee78STejun Heo /* relinquish manager role */ 3031e22bee78STejun Heo gcwq->flags &= ~GCWQ_MANAGING_WORKERS; 3032e22bee78STejun Heo 3033db7bccf4STejun Heo /* notify completion */ 3034db7bccf4STejun Heo gcwq->trustee = NULL; 3035db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3036db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3037db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3038db7bccf4STejun Heo return 0; 3039db7bccf4STejun Heo } 3040db7bccf4STejun Heo 3041db7bccf4STejun Heo /** 3042db7bccf4STejun Heo * wait_trustee_state - wait for trustee to enter the specified state 3043db7bccf4STejun Heo * @gcwq: gcwq the trustee of interest belongs to 3044db7bccf4STejun Heo * @state: target state to wait for 3045db7bccf4STejun Heo * 3046db7bccf4STejun Heo * Wait for the trustee to reach @state. DONE is already matched. 3047db7bccf4STejun Heo * 3048db7bccf4STejun Heo * CONTEXT: 3049db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3050db7bccf4STejun Heo * multiple times. To be used by cpu_callback. 3051db7bccf4STejun Heo */ 3052db7bccf4STejun Heo static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state) 3053db7bccf4STejun Heo { 3054db7bccf4STejun Heo if (!(gcwq->trustee_state == state || 3055db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE)) { 3056db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3057db7bccf4STejun Heo __wait_event(gcwq->trustee_wait, 3058db7bccf4STejun Heo gcwq->trustee_state == state || 3059db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE); 3060db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3061db7bccf4STejun Heo } 3062db7bccf4STejun Heo } 3063db7bccf4STejun Heo 30649c7b216dSChandra Seetharaman static int __devinit workqueue_cpu_callback(struct notifier_block *nfb, 30651da177e4SLinus Torvalds unsigned long action, 30661da177e4SLinus Torvalds void *hcpu) 30671da177e4SLinus Torvalds { 30683af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3069db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3070db7bccf4STejun Heo struct task_struct *new_trustee = NULL; 3071e22bee78STejun Heo struct worker *uninitialized_var(new_worker); 3072db7bccf4STejun Heo unsigned long flags; 30731da177e4SLinus Torvalds 30748bb78442SRafael J. Wysocki action &= ~CPU_TASKS_FROZEN; 30758bb78442SRafael J. Wysocki 3076db7bccf4STejun Heo switch (action) { 3077db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3078db7bccf4STejun Heo new_trustee = kthread_create(trustee_thread, gcwq, 3079db7bccf4STejun Heo "workqueue_trustee/%d\n", cpu); 3080db7bccf4STejun Heo if (IS_ERR(new_trustee)) 3081db7bccf4STejun Heo return notifier_from_errno(PTR_ERR(new_trustee)); 3082db7bccf4STejun Heo kthread_bind(new_trustee, cpu); 3083e22bee78STejun Heo /* fall through */ 3084e22bee78STejun Heo case CPU_UP_PREPARE: 3085e22bee78STejun Heo BUG_ON(gcwq->first_idle); 3086e22bee78STejun Heo new_worker = create_worker(gcwq, false); 3087e22bee78STejun Heo if (!new_worker) { 3088e22bee78STejun Heo if (new_trustee) 3089e22bee78STejun Heo kthread_stop(new_trustee); 3090e22bee78STejun Heo return NOTIFY_BAD; 3091e22bee78STejun Heo } 3092db7bccf4STejun Heo } 30931537663fSTejun Heo 3094db7bccf4STejun Heo /* some are called w/ irq disabled, don't disturb irq status */ 3095db7bccf4STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 30963af24433SOleg Nesterov 30973af24433SOleg Nesterov switch (action) { 3098db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3099db7bccf4STejun Heo /* initialize trustee and tell it to acquire the gcwq */ 3100db7bccf4STejun Heo BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE); 3101db7bccf4STejun Heo gcwq->trustee = new_trustee; 3102db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_START; 3103db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3104db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE); 3105e22bee78STejun Heo /* fall through */ 3106e22bee78STejun Heo case CPU_UP_PREPARE: 3107e22bee78STejun Heo BUG_ON(gcwq->first_idle); 3108e22bee78STejun Heo gcwq->first_idle = new_worker; 3109e22bee78STejun Heo break; 3110e22bee78STejun Heo 3111e22bee78STejun Heo case CPU_DYING: 3112e22bee78STejun Heo /* 3113e22bee78STejun Heo * Before this, the trustee and all workers except for 3114e22bee78STejun Heo * the ones which are still executing works from 3115e22bee78STejun Heo * before the last CPU down must be on the cpu. After 3116e22bee78STejun Heo * this, they'll all be diasporas. 3117e22bee78STejun Heo */ 3118e22bee78STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3119db7bccf4STejun Heo break; 3120db7bccf4STejun Heo 31213da1c84cSOleg Nesterov case CPU_POST_DEAD: 3122db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_BUTCHER; 3123e22bee78STejun Heo /* fall through */ 3124e22bee78STejun Heo case CPU_UP_CANCELED: 3125e22bee78STejun Heo destroy_worker(gcwq->first_idle); 3126e22bee78STejun Heo gcwq->first_idle = NULL; 3127db7bccf4STejun Heo break; 3128db7bccf4STejun Heo 3129db7bccf4STejun Heo case CPU_DOWN_FAILED: 3130db7bccf4STejun Heo case CPU_ONLINE: 3131e22bee78STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 3132db7bccf4STejun Heo if (gcwq->trustee_state != TRUSTEE_DONE) { 3133db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_RELEASE; 3134db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3135db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_DONE); 3136db7bccf4STejun Heo } 3137db7bccf4STejun Heo 3138e22bee78STejun Heo /* 3139e22bee78STejun Heo * Trustee is done and there might be no worker left. 3140e22bee78STejun Heo * Put the first_idle in and request a real manager to 3141e22bee78STejun Heo * take a look. 3142e22bee78STejun Heo */ 3143e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3144e22bee78STejun Heo kthread_bind(gcwq->first_idle->task, cpu); 3145e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3146e22bee78STejun Heo gcwq->flags |= GCWQ_MANAGE_WORKERS; 3147e22bee78STejun Heo start_worker(gcwq->first_idle); 3148e22bee78STejun Heo gcwq->first_idle = NULL; 31491da177e4SLinus Torvalds break; 31501da177e4SLinus Torvalds } 3151db7bccf4STejun Heo 3152db7bccf4STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 31531da177e4SLinus Torvalds 31541537663fSTejun Heo return notifier_from_errno(0); 31551da177e4SLinus Torvalds } 31561da177e4SLinus Torvalds 31572d3854a3SRusty Russell #ifdef CONFIG_SMP 31588ccad40dSRusty Russell 31592d3854a3SRusty Russell struct work_for_cpu { 31606b44003eSAndrew Morton struct completion completion; 31612d3854a3SRusty Russell long (*fn)(void *); 31622d3854a3SRusty Russell void *arg; 31632d3854a3SRusty Russell long ret; 31642d3854a3SRusty Russell }; 31652d3854a3SRusty Russell 31666b44003eSAndrew Morton static int do_work_for_cpu(void *_wfc) 31672d3854a3SRusty Russell { 31686b44003eSAndrew Morton struct work_for_cpu *wfc = _wfc; 31692d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 31706b44003eSAndrew Morton complete(&wfc->completion); 31716b44003eSAndrew Morton return 0; 31722d3854a3SRusty Russell } 31732d3854a3SRusty Russell 31742d3854a3SRusty Russell /** 31752d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 31762d3854a3SRusty Russell * @cpu: the cpu to run on 31772d3854a3SRusty Russell * @fn: the function to run 31782d3854a3SRusty Russell * @arg: the function arg 31792d3854a3SRusty Russell * 318031ad9081SRusty Russell * This will return the value @fn returns. 318131ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 31826b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 31832d3854a3SRusty Russell */ 31842d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 31852d3854a3SRusty Russell { 31866b44003eSAndrew Morton struct task_struct *sub_thread; 31876b44003eSAndrew Morton struct work_for_cpu wfc = { 31886b44003eSAndrew Morton .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), 31896b44003eSAndrew Morton .fn = fn, 31906b44003eSAndrew Morton .arg = arg, 31916b44003eSAndrew Morton }; 31922d3854a3SRusty Russell 31936b44003eSAndrew Morton sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); 31946b44003eSAndrew Morton if (IS_ERR(sub_thread)) 31956b44003eSAndrew Morton return PTR_ERR(sub_thread); 31966b44003eSAndrew Morton kthread_bind(sub_thread, cpu); 31976b44003eSAndrew Morton wake_up_process(sub_thread); 31986b44003eSAndrew Morton wait_for_completion(&wfc.completion); 31992d3854a3SRusty Russell return wfc.ret; 32002d3854a3SRusty Russell } 32012d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 32022d3854a3SRusty Russell #endif /* CONFIG_SMP */ 32032d3854a3SRusty Russell 3204a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3205a0a1a5fdSTejun Heo 3206a0a1a5fdSTejun Heo /** 3207a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3208a0a1a5fdSTejun Heo * 3209a0a1a5fdSTejun Heo * Start freezing workqueues. After this function returns, all 3210a0a1a5fdSTejun Heo * freezeable workqueues will queue new works to their frozen_works 32117e11629dSTejun Heo * list instead of gcwq->worklist. 3212a0a1a5fdSTejun Heo * 3213a0a1a5fdSTejun Heo * CONTEXT: 32148b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3215a0a1a5fdSTejun Heo */ 3216a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3217a0a1a5fdSTejun Heo { 3218a0a1a5fdSTejun Heo struct workqueue_struct *wq; 3219a0a1a5fdSTejun Heo unsigned int cpu; 3220a0a1a5fdSTejun Heo 3221a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3222a0a1a5fdSTejun Heo 3223a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3224a0a1a5fdSTejun Heo workqueue_freezing = true; 3225a0a1a5fdSTejun Heo 3226a0a1a5fdSTejun Heo for_each_possible_cpu(cpu) { 32278b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 32288b03ae3cSTejun Heo 32298b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 32308b03ae3cSTejun Heo 3231db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3232db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3233db7bccf4STejun Heo 3234a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3235a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3236a0a1a5fdSTejun Heo 3237a0a1a5fdSTejun Heo if (wq->flags & WQ_FREEZEABLE) 3238a0a1a5fdSTejun Heo cwq->max_active = 0; 3239a0a1a5fdSTejun Heo } 32408b03ae3cSTejun Heo 32418b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3242a0a1a5fdSTejun Heo } 3243a0a1a5fdSTejun Heo 3244a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3245a0a1a5fdSTejun Heo } 3246a0a1a5fdSTejun Heo 3247a0a1a5fdSTejun Heo /** 3248a0a1a5fdSTejun Heo * freeze_workqueues_busy - are freezeable workqueues still busy? 3249a0a1a5fdSTejun Heo * 3250a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3251a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3252a0a1a5fdSTejun Heo * 3253a0a1a5fdSTejun Heo * CONTEXT: 3254a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3255a0a1a5fdSTejun Heo * 3256a0a1a5fdSTejun Heo * RETURNS: 3257a0a1a5fdSTejun Heo * %true if some freezeable workqueues are still busy. %false if 3258a0a1a5fdSTejun Heo * freezing is complete. 3259a0a1a5fdSTejun Heo */ 3260a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3261a0a1a5fdSTejun Heo { 3262a0a1a5fdSTejun Heo struct workqueue_struct *wq; 3263a0a1a5fdSTejun Heo unsigned int cpu; 3264a0a1a5fdSTejun Heo bool busy = false; 3265a0a1a5fdSTejun Heo 3266a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3267a0a1a5fdSTejun Heo 3268a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3269a0a1a5fdSTejun Heo 3270a0a1a5fdSTejun Heo for_each_possible_cpu(cpu) { 3271a0a1a5fdSTejun Heo /* 3272a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3273a0a1a5fdSTejun Heo * to peek without lock. 3274a0a1a5fdSTejun Heo */ 3275a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3276a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3277a0a1a5fdSTejun Heo 3278a0a1a5fdSTejun Heo if (!(wq->flags & WQ_FREEZEABLE)) 3279a0a1a5fdSTejun Heo continue; 3280a0a1a5fdSTejun Heo 3281a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3282a0a1a5fdSTejun Heo if (cwq->nr_active) { 3283a0a1a5fdSTejun Heo busy = true; 3284a0a1a5fdSTejun Heo goto out_unlock; 3285a0a1a5fdSTejun Heo } 3286a0a1a5fdSTejun Heo } 3287a0a1a5fdSTejun Heo } 3288a0a1a5fdSTejun Heo out_unlock: 3289a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3290a0a1a5fdSTejun Heo return busy; 3291a0a1a5fdSTejun Heo } 3292a0a1a5fdSTejun Heo 3293a0a1a5fdSTejun Heo /** 3294a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3295a0a1a5fdSTejun Heo * 3296a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 32977e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3298a0a1a5fdSTejun Heo * 3299a0a1a5fdSTejun Heo * CONTEXT: 33008b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3301a0a1a5fdSTejun Heo */ 3302a0a1a5fdSTejun Heo void thaw_workqueues(void) 3303a0a1a5fdSTejun Heo { 3304a0a1a5fdSTejun Heo struct workqueue_struct *wq; 3305a0a1a5fdSTejun Heo unsigned int cpu; 3306a0a1a5fdSTejun Heo 3307a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3308a0a1a5fdSTejun Heo 3309a0a1a5fdSTejun Heo if (!workqueue_freezing) 3310a0a1a5fdSTejun Heo goto out_unlock; 3311a0a1a5fdSTejun Heo 3312a0a1a5fdSTejun Heo for_each_possible_cpu(cpu) { 33138b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 33148b03ae3cSTejun Heo 33158b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 33168b03ae3cSTejun Heo 3317db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3318db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3319db7bccf4STejun Heo 3320a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3321a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3322a0a1a5fdSTejun Heo 3323a0a1a5fdSTejun Heo if (!(wq->flags & WQ_FREEZEABLE)) 3324a0a1a5fdSTejun Heo continue; 3325a0a1a5fdSTejun Heo 3326a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3327a0a1a5fdSTejun Heo cwq->max_active = wq->saved_max_active; 3328a0a1a5fdSTejun Heo 3329a0a1a5fdSTejun Heo while (!list_empty(&cwq->delayed_works) && 3330a0a1a5fdSTejun Heo cwq->nr_active < cwq->max_active) 3331a0a1a5fdSTejun Heo cwq_activate_first_delayed(cwq); 3332a0a1a5fdSTejun Heo 3333502ca9d8STejun Heo /* perform delayed unbind from single cpu if empty */ 3334502ca9d8STejun Heo if (wq->single_cpu == gcwq->cpu && 3335502ca9d8STejun Heo !cwq->nr_active && list_empty(&cwq->delayed_works)) 3336502ca9d8STejun Heo cwq_unbind_single_cpu(cwq); 3337a0a1a5fdSTejun Heo } 33388b03ae3cSTejun Heo 3339e22bee78STejun Heo wake_up_worker(gcwq); 3340e22bee78STejun Heo 33418b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3342a0a1a5fdSTejun Heo } 3343a0a1a5fdSTejun Heo 3344a0a1a5fdSTejun Heo workqueue_freezing = false; 3345a0a1a5fdSTejun Heo out_unlock: 3346a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3347a0a1a5fdSTejun Heo } 3348a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3349a0a1a5fdSTejun Heo 3350c12920d1SOleg Nesterov void __init init_workqueues(void) 33511da177e4SLinus Torvalds { 3352c34056a3STejun Heo unsigned int cpu; 3353c8e55f36STejun Heo int i; 3354c34056a3STejun Heo 33557a22ad75STejun Heo /* 33567a22ad75STejun Heo * The pointer part of work->data is either pointing to the 33577a22ad75STejun Heo * cwq or contains the cpu number the work ran last on. Make 33587a22ad75STejun Heo * sure cpu number won't overflow into kernel pointer area so 33597a22ad75STejun Heo * that they can be distinguished. 33607a22ad75STejun Heo */ 33617a22ad75STejun Heo BUILD_BUG_ON(NR_CPUS << WORK_STRUCT_FLAG_BITS >= PAGE_OFFSET); 33627a22ad75STejun Heo 3363db7bccf4STejun Heo hotcpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE); 33648b03ae3cSTejun Heo 33658b03ae3cSTejun Heo /* initialize gcwqs */ 33668b03ae3cSTejun Heo for_each_possible_cpu(cpu) { 33678b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 33688b03ae3cSTejun Heo 33698b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 33707e11629dSTejun Heo INIT_LIST_HEAD(&gcwq->worklist); 33718b03ae3cSTejun Heo gcwq->cpu = cpu; 33728b03ae3cSTejun Heo 3373c8e55f36STejun Heo INIT_LIST_HEAD(&gcwq->idle_list); 3374c8e55f36STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) 3375c8e55f36STejun Heo INIT_HLIST_HEAD(&gcwq->busy_hash[i]); 3376c8e55f36STejun Heo 3377e22bee78STejun Heo init_timer_deferrable(&gcwq->idle_timer); 3378e22bee78STejun Heo gcwq->idle_timer.function = idle_worker_timeout; 3379e22bee78STejun Heo gcwq->idle_timer.data = (unsigned long)gcwq; 3380e22bee78STejun Heo 3381e22bee78STejun Heo setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout, 3382e22bee78STejun Heo (unsigned long)gcwq); 3383e22bee78STejun Heo 33848b03ae3cSTejun Heo ida_init(&gcwq->worker_ida); 3385db7bccf4STejun Heo 3386db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3387db7bccf4STejun Heo init_waitqueue_head(&gcwq->trustee_wait); 33888b03ae3cSTejun Heo } 33898b03ae3cSTejun Heo 3390e22bee78STejun Heo /* create the initial worker */ 3391e22bee78STejun Heo for_each_online_cpu(cpu) { 3392e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3393e22bee78STejun Heo struct worker *worker; 3394e22bee78STejun Heo 3395e22bee78STejun Heo worker = create_worker(gcwq, true); 3396e22bee78STejun Heo BUG_ON(!worker); 3397e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3398e22bee78STejun Heo start_worker(worker); 3399e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3400e22bee78STejun Heo } 3401e22bee78STejun Heo 3402d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 3403d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3404d320c038STejun Heo system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0); 3405d320c038STejun Heo BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq); 34061da177e4SLinus Torvalds } 3407