11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4429c91e99STejun Heo #include <linux/jhash.h> 4542f8570fSSasha Levin #include <linux/hashtable.h> 4676af4d93STejun Heo #include <linux/rculist.h> 47e22bee78STejun Heo 48ea138446STejun Heo #include "workqueue_internal.h" 491da177e4SLinus Torvalds 50c8e55f36STejun Heo enum { 51bc2ae0f5STejun Heo /* 5224647570STejun Heo * worker_pool flags 53bc2ae0f5STejun Heo * 5424647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 55bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 56bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 57bc2ae0f5STejun Heo * is in effect. 58bc2ae0f5STejun Heo * 59bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 60bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6124647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 62bc2ae0f5STejun Heo * 63bc2ae0f5STejun Heo * Note that DISASSOCIATED can be flipped only while holding 6424647570STejun Heo * assoc_mutex to avoid changing binding state while 6524647570STejun Heo * create_worker() is in progress. 66bc2ae0f5STejun Heo */ 6711ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 6824647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 6935b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 70db7bccf4STejun Heo 71c8e55f36STejun Heo /* worker flags */ 72c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 73c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 74c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 75e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 76fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 77f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 78e22bee78STejun Heo 795f7dabfdSLai Jiangshan WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND | 80403c821dSTejun Heo WORKER_CPU_INTENSIVE, 81db7bccf4STejun Heo 82e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 834ce62e9eSTejun Heo 8429c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 85c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 86db7bccf4STejun Heo 87e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 88e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 89e22bee78STejun Heo 903233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 913233cdbdSTejun Heo /* call for help after 10ms 923233cdbdSTejun Heo (min two ticks) */ 93e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 94e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 951da177e4SLinus Torvalds 961da177e4SLinus Torvalds /* 97e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 98e22bee78STejun Heo * all cpus. Give -20. 99e22bee78STejun Heo */ 100e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1013270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 102c8e55f36STejun Heo }; 103c8e55f36STejun Heo 1041da177e4SLinus Torvalds /* 1054690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1064690c4abSTejun Heo * 107e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 108e41e704bSTejun Heo * everyone else. 1094690c4abSTejun Heo * 110e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 111e22bee78STejun Heo * only be modified and accessed from the local cpu. 112e22bee78STejun Heo * 113d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1144690c4abSTejun Heo * 115d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 116d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 117d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 118d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 119e22bee78STejun Heo * 12073f53c4aSTejun Heo * F: wq->flush_mutex protected. 12173f53c4aSTejun Heo * 1224690c4abSTejun Heo * W: workqueue_lock protected. 12376af4d93STejun Heo * 12476af4d93STejun Heo * R: workqueue_lock protected for writes. Sched-RCU protected for reads. 12575ccf595STejun Heo * 12675ccf595STejun Heo * FR: wq->flush_mutex and workqueue_lock protected for writes. Sched-RCU 12775ccf595STejun Heo * protected for reads. 1284690c4abSTejun Heo */ 1294690c4abSTejun Heo 1302eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 131c34056a3STejun Heo 132bd7bdd43STejun Heo struct worker_pool { 133d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 134d84ff051STejun Heo int cpu; /* I: the associated cpu */ 1359daf9e67STejun Heo int id; /* I: pool ID */ 13611ebea50STejun Heo unsigned int flags; /* X: flags */ 137bd7bdd43STejun Heo 138bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 139bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 140ea1abd61SLai Jiangshan 141ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 142bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 143bd7bdd43STejun Heo 144bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 145bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 146bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 147bd7bdd43STejun Heo 148c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 149c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 150c9e7cf27STejun Heo /* L: hash of busy workers */ 151c9e7cf27STejun Heo 15234a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 15324647570STejun Heo struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */ 154bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 155e19e397aSTejun Heo 1567a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 157c5aa87bbSTejun Heo struct hlist_node hash_node; /* W: unbound_pool_hash node */ 158c5aa87bbSTejun Heo int refcnt; /* W: refcnt for unbound pools */ 1597a4e344cSTejun Heo 160e19e397aSTejun Heo /* 161e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 162e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 163e19e397aSTejun Heo * cacheline. 164e19e397aSTejun Heo */ 165e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 16629c91e99STejun Heo 16729c91e99STejun Heo /* 16829c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 16929c91e99STejun Heo * from get_work_pool(). 17029c91e99STejun Heo */ 17129c91e99STejun Heo struct rcu_head rcu; 1728b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1738b03ae3cSTejun Heo 1748b03ae3cSTejun Heo /* 175112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 176112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 177112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 178112202d9STejun Heo * number of flag bits. 1791da177e4SLinus Torvalds */ 180112202d9STejun Heo struct pool_workqueue { 181bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1824690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 18373f53c4aSTejun Heo int work_color; /* L: current color */ 18473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1858864b4e5STejun Heo int refcnt; /* L: reference count */ 18673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 18773f53c4aSTejun Heo /* L: nr of in_flight works */ 1881e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 189a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1901e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 19175ccf595STejun Heo struct list_head pwqs_node; /* FR: node on wq->pwqs */ 192493a1724STejun Heo struct list_head mayday_node; /* W: node on wq->maydays */ 1938864b4e5STejun Heo 1948864b4e5STejun Heo /* 1958864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 1968864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 1978864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 1988864b4e5STejun Heo * determined without grabbing workqueue_lock. 1998864b4e5STejun Heo */ 2008864b4e5STejun Heo struct work_struct unbound_release_work; 2018864b4e5STejun Heo struct rcu_head rcu; 202e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2031da177e4SLinus Torvalds 2041da177e4SLinus Torvalds /* 20573f53c4aSTejun Heo * Structure used to wait for workqueue flush. 20673f53c4aSTejun Heo */ 20773f53c4aSTejun Heo struct wq_flusher { 20873f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 20973f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 21073f53c4aSTejun Heo struct completion done; /* flush completion */ 21173f53c4aSTejun Heo }; 2121da177e4SLinus Torvalds 213226223abSTejun Heo struct wq_device; 214226223abSTejun Heo 21573f53c4aSTejun Heo /* 216c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 217c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2181da177e4SLinus Torvalds */ 2191da177e4SLinus Torvalds struct workqueue_struct { 2209c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 221420c0ddbSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ 22275ccf595STejun Heo struct list_head pwqs; /* FR: all pwqs of this wq */ 2234690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 22473f53c4aSTejun Heo 22573f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 22673f53c4aSTejun Heo int work_color; /* F: current work color */ 22773f53c4aSTejun Heo int flush_color; /* F: current flush color */ 228112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 22973f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 23073f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 23173f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 23273f53c4aSTejun Heo 233493a1724STejun Heo struct list_head maydays; /* W: pwqs requesting rescue */ 234e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 235e22bee78STejun Heo 2369c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 237112202d9STejun Heo int saved_max_active; /* W: saved pwq max_active */ 238226223abSTejun Heo 239226223abSTejun Heo #ifdef CONFIG_SYSFS 240226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 241226223abSTejun Heo #endif 2424e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2434e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2444e6045f1SJohannes Berg #endif 245b196be89STejun Heo char name[]; /* I: workqueue name */ 2461da177e4SLinus Torvalds }; 2471da177e4SLinus Torvalds 248e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 249e904e6c2STejun Heo 250c5aa87bbSTejun Heo /* W: hash of all unbound pools keyed by pool->attrs */ 25129c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 25229c91e99STejun Heo 253c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 25429c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 25529c91e99STejun Heo 256d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 257d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 258044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2591aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 260044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 261d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 262044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 263f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 264044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 26524d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 266d320c038STejun Heo 26797bd2347STejun Heo #define CREATE_TRACE_POINTS 26897bd2347STejun Heo #include <trace/events/workqueue.h> 26997bd2347STejun Heo 27076af4d93STejun Heo #define assert_rcu_or_wq_lock() \ 27176af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 27276af4d93STejun Heo lockdep_is_held(&workqueue_lock), \ 27376af4d93STejun Heo "sched RCU or workqueue lock should be held") 27476af4d93STejun Heo 275f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 276f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 277f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 2787a62c2c8STejun Heo (pool)++) 2794ce62e9eSTejun Heo 280b67bfe0dSSasha Levin #define for_each_busy_worker(worker, i, pool) \ 281b67bfe0dSSasha Levin hash_for_each(pool->busy_hash, i, worker, hentry) 282db7bccf4STejun Heo 28349e3cf44STejun Heo /** 28417116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 28517116969STejun Heo * @pool: iteration cursor 28617116969STejun Heo * @id: integer used for iteration 287fa1b54e6STejun Heo * 288fa1b54e6STejun Heo * This must be called either with workqueue_lock held or sched RCU read 289fa1b54e6STejun Heo * locked. If the pool needs to be used beyond the locking in effect, the 290fa1b54e6STejun Heo * caller is responsible for guaranteeing that the pool stays online. 291fa1b54e6STejun Heo * 292fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 293fa1b54e6STejun Heo * ignored. 29417116969STejun Heo */ 29517116969STejun Heo #define for_each_pool(pool, id) \ 296fa1b54e6STejun Heo idr_for_each_entry(&worker_pool_idr, pool, id) \ 297fa1b54e6STejun Heo if (({ assert_rcu_or_wq_lock(); false; })) { } \ 298fa1b54e6STejun Heo else 29917116969STejun Heo 30017116969STejun Heo /** 30149e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 30249e3cf44STejun Heo * @pwq: iteration cursor 30349e3cf44STejun Heo * @wq: the target workqueue 30476af4d93STejun Heo * 30576af4d93STejun Heo * This must be called either with workqueue_lock held or sched RCU read 30676af4d93STejun Heo * locked. If the pwq needs to be used beyond the locking in effect, the 30776af4d93STejun Heo * caller is responsible for guaranteeing that the pwq stays online. 30876af4d93STejun Heo * 30976af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 31076af4d93STejun Heo * ignored. 31149e3cf44STejun Heo */ 31249e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 31376af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 31476af4d93STejun Heo if (({ assert_rcu_or_wq_lock(); false; })) { } \ 31576af4d93STejun Heo else 316f3421797STejun Heo 317dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 318dc186ad7SThomas Gleixner 319dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 320dc186ad7SThomas Gleixner 32199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 32299777288SStanislaw Gruszka { 32399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 32499777288SStanislaw Gruszka } 32599777288SStanislaw Gruszka 326dc186ad7SThomas Gleixner /* 327dc186ad7SThomas Gleixner * fixup_init is called when: 328dc186ad7SThomas Gleixner * - an active object is initialized 329dc186ad7SThomas Gleixner */ 330dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 331dc186ad7SThomas Gleixner { 332dc186ad7SThomas Gleixner struct work_struct *work = addr; 333dc186ad7SThomas Gleixner 334dc186ad7SThomas Gleixner switch (state) { 335dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 336dc186ad7SThomas Gleixner cancel_work_sync(work); 337dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 338dc186ad7SThomas Gleixner return 1; 339dc186ad7SThomas Gleixner default: 340dc186ad7SThomas Gleixner return 0; 341dc186ad7SThomas Gleixner } 342dc186ad7SThomas Gleixner } 343dc186ad7SThomas Gleixner 344dc186ad7SThomas Gleixner /* 345dc186ad7SThomas Gleixner * fixup_activate is called when: 346dc186ad7SThomas Gleixner * - an active object is activated 347dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 348dc186ad7SThomas Gleixner */ 349dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 350dc186ad7SThomas Gleixner { 351dc186ad7SThomas Gleixner struct work_struct *work = addr; 352dc186ad7SThomas Gleixner 353dc186ad7SThomas Gleixner switch (state) { 354dc186ad7SThomas Gleixner 355dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 356dc186ad7SThomas Gleixner /* 357dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 358dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 359dc186ad7SThomas Gleixner * is tracked in the object tracker. 360dc186ad7SThomas Gleixner */ 36122df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 362dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 363dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 364dc186ad7SThomas Gleixner return 0; 365dc186ad7SThomas Gleixner } 366dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 367dc186ad7SThomas Gleixner return 0; 368dc186ad7SThomas Gleixner 369dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 370dc186ad7SThomas Gleixner WARN_ON(1); 371dc186ad7SThomas Gleixner 372dc186ad7SThomas Gleixner default: 373dc186ad7SThomas Gleixner return 0; 374dc186ad7SThomas Gleixner } 375dc186ad7SThomas Gleixner } 376dc186ad7SThomas Gleixner 377dc186ad7SThomas Gleixner /* 378dc186ad7SThomas Gleixner * fixup_free is called when: 379dc186ad7SThomas Gleixner * - an active object is freed 380dc186ad7SThomas Gleixner */ 381dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 382dc186ad7SThomas Gleixner { 383dc186ad7SThomas Gleixner struct work_struct *work = addr; 384dc186ad7SThomas Gleixner 385dc186ad7SThomas Gleixner switch (state) { 386dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 387dc186ad7SThomas Gleixner cancel_work_sync(work); 388dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 389dc186ad7SThomas Gleixner return 1; 390dc186ad7SThomas Gleixner default: 391dc186ad7SThomas Gleixner return 0; 392dc186ad7SThomas Gleixner } 393dc186ad7SThomas Gleixner } 394dc186ad7SThomas Gleixner 395dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 396dc186ad7SThomas Gleixner .name = "work_struct", 39799777288SStanislaw Gruszka .debug_hint = work_debug_hint, 398dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 399dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 400dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 401dc186ad7SThomas Gleixner }; 402dc186ad7SThomas Gleixner 403dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 404dc186ad7SThomas Gleixner { 405dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 406dc186ad7SThomas Gleixner } 407dc186ad7SThomas Gleixner 408dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 409dc186ad7SThomas Gleixner { 410dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 411dc186ad7SThomas Gleixner } 412dc186ad7SThomas Gleixner 413dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 414dc186ad7SThomas Gleixner { 415dc186ad7SThomas Gleixner if (onstack) 416dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 417dc186ad7SThomas Gleixner else 418dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 419dc186ad7SThomas Gleixner } 420dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 421dc186ad7SThomas Gleixner 422dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 423dc186ad7SThomas Gleixner { 424dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 425dc186ad7SThomas Gleixner } 426dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 427dc186ad7SThomas Gleixner 428dc186ad7SThomas Gleixner #else 429dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 430dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 431dc186ad7SThomas Gleixner #endif 432dc186ad7SThomas Gleixner 43395402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 43495402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4351da177e4SLinus Torvalds static LIST_HEAD(workqueues); 436a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4371da177e4SLinus Torvalds 438c5aa87bbSTejun Heo /* the per-cpu worker pools */ 439e19e397aSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 440f02ae73aSTejun Heo cpu_worker_pools); 441f3421797STejun Heo 442fa1b54e6STejun Heo /* 443c5aa87bbSTejun Heo * R: idr of all pools. Modifications are protected by workqueue_lock. 444c5aa87bbSTejun Heo * Read accesses are protected by sched-RCU protected. 445fa1b54e6STejun Heo */ 4469daf9e67STejun Heo static DEFINE_IDR(worker_pool_idr); 4479daf9e67STejun Heo 448c34056a3STejun Heo static int worker_thread(void *__worker); 449226223abSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 450226223abSTejun Heo const struct workqueue_attrs *from); 4511da177e4SLinus Torvalds 4529daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4539daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4549daf9e67STejun Heo { 4559daf9e67STejun Heo int ret; 4569daf9e67STejun Heo 457fa1b54e6STejun Heo do { 458fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 459fa1b54e6STejun Heo return -ENOMEM; 460fa1b54e6STejun Heo 461fa1b54e6STejun Heo spin_lock_irq(&workqueue_lock); 4629daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 463fa1b54e6STejun Heo spin_unlock_irq(&workqueue_lock); 464fa1b54e6STejun Heo } while (ret == -EAGAIN); 4659daf9e67STejun Heo 4669daf9e67STejun Heo return ret; 4679daf9e67STejun Heo } 4689daf9e67STejun Heo 46976af4d93STejun Heo /** 47076af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 47176af4d93STejun Heo * @wq: the target workqueue 47276af4d93STejun Heo * 47376af4d93STejun Heo * This must be called either with workqueue_lock held or sched RCU read 47476af4d93STejun Heo * locked. If the pwq needs to be used beyond the locking in effect, the 47576af4d93STejun Heo * caller is responsible for guaranteeing that the pwq stays online. 47676af4d93STejun Heo */ 4777fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 478a848e3b6SOleg Nesterov { 47976af4d93STejun Heo assert_rcu_or_wq_lock(); 48076af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 48176af4d93STejun Heo pwqs_node); 482f3421797STejun Heo } 483a848e3b6SOleg Nesterov 48473f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 48573f53c4aSTejun Heo { 48673f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 48773f53c4aSTejun Heo } 48873f53c4aSTejun Heo 48973f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 49073f53c4aSTejun Heo { 49173f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 49273f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 49373f53c4aSTejun Heo } 49473f53c4aSTejun Heo 49573f53c4aSTejun Heo static int work_next_color(int color) 49673f53c4aSTejun Heo { 49773f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 498b1f4ec17SOleg Nesterov } 499b1f4ec17SOleg Nesterov 5004594bf15SDavid Howells /* 501112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 502112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5037c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5047a22ad75STejun Heo * 505112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 506112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 507bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 508bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5097a22ad75STejun Heo * 510112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5117c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 512112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5137c3eed5cSTejun Heo * available only while the work item is queued. 514bbb68dfaSTejun Heo * 515bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 516bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 517bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 518bbb68dfaSTejun Heo * try to steal the PENDING bit. 5194594bf15SDavid Howells */ 5207a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5217a22ad75STejun Heo unsigned long flags) 5227a22ad75STejun Heo { 5236183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5247a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5257a22ad75STejun Heo } 5267a22ad75STejun Heo 527112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5284690c4abSTejun Heo unsigned long extra_flags) 529365970a1SDavid Howells { 530112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 531112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 532365970a1SDavid Howells } 533365970a1SDavid Howells 5344468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5354468a00fSLai Jiangshan int pool_id) 5364468a00fSLai Jiangshan { 5374468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5384468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5394468a00fSLai Jiangshan } 5404468a00fSLai Jiangshan 5417c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5427c3eed5cSTejun Heo int pool_id) 5434d707b9fSOleg Nesterov { 54423657bb1STejun Heo /* 54523657bb1STejun Heo * The following wmb is paired with the implied mb in 54623657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 54723657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 54823657bb1STejun Heo * owner. 54923657bb1STejun Heo */ 55023657bb1STejun Heo smp_wmb(); 5517c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5524d707b9fSOleg Nesterov } 5534d707b9fSOleg Nesterov 5547a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 555365970a1SDavid Howells { 5567c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5577c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5587a22ad75STejun Heo } 5597a22ad75STejun Heo 560112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5617a22ad75STejun Heo { 562e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5637a22ad75STejun Heo 564112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 565e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 566e120153dSTejun Heo else 567e120153dSTejun Heo return NULL; 5687a22ad75STejun Heo } 5697a22ad75STejun Heo 5707c3eed5cSTejun Heo /** 5717c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 5727c3eed5cSTejun Heo * @work: the work item of interest 5737c3eed5cSTejun Heo * 5747c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 575fa1b54e6STejun Heo * 576fa1b54e6STejun Heo * Pools are created and destroyed under workqueue_lock, and allows read 577fa1b54e6STejun Heo * access under sched-RCU read lock. As such, this function should be 578fa1b54e6STejun Heo * called under workqueue_lock or with preemption disabled. 579fa1b54e6STejun Heo * 580fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 581fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 582fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 583fa1b54e6STejun Heo * returned pool is and stays online. 5847c3eed5cSTejun Heo */ 5857c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 5867a22ad75STejun Heo { 587e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5887c3eed5cSTejun Heo int pool_id; 5897a22ad75STejun Heo 590fa1b54e6STejun Heo assert_rcu_or_wq_lock(); 591fa1b54e6STejun Heo 592112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 593112202d9STejun Heo return ((struct pool_workqueue *) 5947c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 5957a22ad75STejun Heo 5967c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 5977c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 5987a22ad75STejun Heo return NULL; 5997a22ad75STejun Heo 600fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6017c3eed5cSTejun Heo } 6027c3eed5cSTejun Heo 6037c3eed5cSTejun Heo /** 6047c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6057c3eed5cSTejun Heo * @work: the work item of interest 6067c3eed5cSTejun Heo * 6077c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6087c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6097c3eed5cSTejun Heo */ 6107c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6117c3eed5cSTejun Heo { 61254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6137c3eed5cSTejun Heo 614112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 615112202d9STejun Heo return ((struct pool_workqueue *) 61654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 61754d5b7d0SLai Jiangshan 61854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6197c3eed5cSTejun Heo } 6207c3eed5cSTejun Heo 621bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 622bbb68dfaSTejun Heo { 6237c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 624bbb68dfaSTejun Heo 6257c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6267c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 627bbb68dfaSTejun Heo } 628bbb68dfaSTejun Heo 629bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 630bbb68dfaSTejun Heo { 631bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 632bbb68dfaSTejun Heo 633112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 634bbb68dfaSTejun Heo } 635bbb68dfaSTejun Heo 636e22bee78STejun Heo /* 6373270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6383270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 639d565ed63STejun Heo * they're being called with pool->lock held. 640e22bee78STejun Heo */ 641e22bee78STejun Heo 64263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 643649027d7STejun Heo { 644e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 645649027d7STejun Heo } 646649027d7STejun Heo 647e22bee78STejun Heo /* 648e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 649e22bee78STejun Heo * running workers. 650974271c4STejun Heo * 651974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 652706026c2STejun Heo * function will always return %true for unbound pools as long as the 653974271c4STejun Heo * worklist isn't empty. 654e22bee78STejun Heo */ 65563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 656e22bee78STejun Heo { 65763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 658e22bee78STejun Heo } 659e22bee78STejun Heo 660e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 66163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 662e22bee78STejun Heo { 66363d95a91STejun Heo return pool->nr_idle; 664e22bee78STejun Heo } 665e22bee78STejun Heo 666e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 66763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 668e22bee78STejun Heo { 669e19e397aSTejun Heo return !list_empty(&pool->worklist) && 670e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 671e22bee78STejun Heo } 672e22bee78STejun Heo 673e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 67463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 675e22bee78STejun Heo { 67663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 677e22bee78STejun Heo } 678e22bee78STejun Heo 679e22bee78STejun Heo /* Do I need to be the manager? */ 68063d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 681e22bee78STejun Heo { 68263d95a91STejun Heo return need_to_create_worker(pool) || 68311ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 684e22bee78STejun Heo } 685e22bee78STejun Heo 686e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 68763d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 688e22bee78STejun Heo { 68934a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 69063d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 69163d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 692e22bee78STejun Heo 693ea1abd61SLai Jiangshan /* 694ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 695ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 696ea1abd61SLai Jiangshan */ 697ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 698ea1abd61SLai Jiangshan return false; 699ea1abd61SLai Jiangshan 700e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 701e22bee78STejun Heo } 702e22bee78STejun Heo 703e22bee78STejun Heo /* 704e22bee78STejun Heo * Wake up functions. 705e22bee78STejun Heo */ 706e22bee78STejun Heo 7077e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 70863d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7097e11629dSTejun Heo { 71063d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7117e11629dSTejun Heo return NULL; 7127e11629dSTejun Heo 71363d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7147e11629dSTejun Heo } 7157e11629dSTejun Heo 7167e11629dSTejun Heo /** 7177e11629dSTejun Heo * wake_up_worker - wake up an idle worker 71863d95a91STejun Heo * @pool: worker pool to wake worker from 7197e11629dSTejun Heo * 72063d95a91STejun Heo * Wake up the first idle worker of @pool. 7217e11629dSTejun Heo * 7227e11629dSTejun Heo * CONTEXT: 723d565ed63STejun Heo * spin_lock_irq(pool->lock). 7247e11629dSTejun Heo */ 72563d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7267e11629dSTejun Heo { 72763d95a91STejun Heo struct worker *worker = first_worker(pool); 7287e11629dSTejun Heo 7297e11629dSTejun Heo if (likely(worker)) 7307e11629dSTejun Heo wake_up_process(worker->task); 7317e11629dSTejun Heo } 7327e11629dSTejun Heo 7334690c4abSTejun Heo /** 734e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 735e22bee78STejun Heo * @task: task waking up 736e22bee78STejun Heo * @cpu: CPU @task is waking up to 737e22bee78STejun Heo * 738e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 739e22bee78STejun Heo * being awoken. 740e22bee78STejun Heo * 741e22bee78STejun Heo * CONTEXT: 742e22bee78STejun Heo * spin_lock_irq(rq->lock) 743e22bee78STejun Heo */ 744d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 745e22bee78STejun Heo { 746e22bee78STejun Heo struct worker *worker = kthread_data(task); 747e22bee78STejun Heo 74836576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 749ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 750e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 751e22bee78STejun Heo } 75236576000SJoonsoo Kim } 753e22bee78STejun Heo 754e22bee78STejun Heo /** 755e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 756e22bee78STejun Heo * @task: task going to sleep 757e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 758e22bee78STejun Heo * 759e22bee78STejun Heo * This function is called during schedule() when a busy worker is 760e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 761e22bee78STejun Heo * returning pointer to its task. 762e22bee78STejun Heo * 763e22bee78STejun Heo * CONTEXT: 764e22bee78STejun Heo * spin_lock_irq(rq->lock) 765e22bee78STejun Heo * 766e22bee78STejun Heo * RETURNS: 767e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 768e22bee78STejun Heo */ 769d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 770e22bee78STejun Heo { 771e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 772111c225aSTejun Heo struct worker_pool *pool; 773e22bee78STejun Heo 774111c225aSTejun Heo /* 775111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 776111c225aSTejun Heo * workers, also reach here, let's not access anything before 777111c225aSTejun Heo * checking NOT_RUNNING. 778111c225aSTejun Heo */ 7792d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 780e22bee78STejun Heo return NULL; 781e22bee78STejun Heo 782111c225aSTejun Heo pool = worker->pool; 783111c225aSTejun Heo 784e22bee78STejun Heo /* this can only happen on the local cpu */ 7856183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 7866183c009STejun Heo return NULL; 787e22bee78STejun Heo 788e22bee78STejun Heo /* 789e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 790e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 791e22bee78STejun Heo * Please read comment there. 792e22bee78STejun Heo * 793628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 794628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 795628c78e7STejun Heo * disabled, which in turn means that none else could be 796d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 797628c78e7STejun Heo * lock is safe. 798e22bee78STejun Heo */ 799e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 800e19e397aSTejun Heo !list_empty(&pool->worklist)) 80163d95a91STejun Heo to_wakeup = first_worker(pool); 802e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 803e22bee78STejun Heo } 804e22bee78STejun Heo 805e22bee78STejun Heo /** 806e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 807cb444766STejun Heo * @worker: self 808d302f017STejun Heo * @flags: flags to set 809d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 810d302f017STejun Heo * 811e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 812e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 813e22bee78STejun Heo * woken up. 814d302f017STejun Heo * 815cb444766STejun Heo * CONTEXT: 816d565ed63STejun Heo * spin_lock_irq(pool->lock) 817d302f017STejun Heo */ 818d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 819d302f017STejun Heo bool wakeup) 820d302f017STejun Heo { 821bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 822e22bee78STejun Heo 823cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 824cb444766STejun Heo 825e22bee78STejun Heo /* 826e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 827e22bee78STejun Heo * wake up an idle worker as necessary if requested by 828e22bee78STejun Heo * @wakeup. 829e22bee78STejun Heo */ 830e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 831e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 832e22bee78STejun Heo if (wakeup) { 833e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 834bd7bdd43STejun Heo !list_empty(&pool->worklist)) 83563d95a91STejun Heo wake_up_worker(pool); 836e22bee78STejun Heo } else 837e19e397aSTejun Heo atomic_dec(&pool->nr_running); 838e22bee78STejun Heo } 839e22bee78STejun Heo 840d302f017STejun Heo worker->flags |= flags; 841d302f017STejun Heo } 842d302f017STejun Heo 843d302f017STejun Heo /** 844e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 845cb444766STejun Heo * @worker: self 846d302f017STejun Heo * @flags: flags to clear 847d302f017STejun Heo * 848e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 849d302f017STejun Heo * 850cb444766STejun Heo * CONTEXT: 851d565ed63STejun Heo * spin_lock_irq(pool->lock) 852d302f017STejun Heo */ 853d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 854d302f017STejun Heo { 85563d95a91STejun Heo struct worker_pool *pool = worker->pool; 856e22bee78STejun Heo unsigned int oflags = worker->flags; 857e22bee78STejun Heo 858cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 859cb444766STejun Heo 860d302f017STejun Heo worker->flags &= ~flags; 861e22bee78STejun Heo 86242c025f3STejun Heo /* 86342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 86442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 86542c025f3STejun Heo * of multiple flags, not a single flag. 86642c025f3STejun Heo */ 867e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 868e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 869e19e397aSTejun Heo atomic_inc(&pool->nr_running); 870d302f017STejun Heo } 871d302f017STejun Heo 872d302f017STejun Heo /** 8738cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 874c9e7cf27STejun Heo * @pool: pool of interest 8758cca0eeaSTejun Heo * @work: work to find worker for 8768cca0eeaSTejun Heo * 877c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 878c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 879a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 880a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 881a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 882a2c1c57bSTejun Heo * being executed. 883a2c1c57bSTejun Heo * 884a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 885a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 886a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 887a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 888a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 889a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 890a2c1c57bSTejun Heo * 891c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 892c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 893c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 894c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 895c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 896c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 8978cca0eeaSTejun Heo * 8988cca0eeaSTejun Heo * CONTEXT: 899d565ed63STejun Heo * spin_lock_irq(pool->lock). 9008cca0eeaSTejun Heo * 9018cca0eeaSTejun Heo * RETURNS: 9028cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9038cca0eeaSTejun Heo * otherwise. 9048cca0eeaSTejun Heo */ 905c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9068cca0eeaSTejun Heo struct work_struct *work) 9078cca0eeaSTejun Heo { 90842f8570fSSasha Levin struct worker *worker; 90942f8570fSSasha Levin 910b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 911a2c1c57bSTejun Heo (unsigned long)work) 912a2c1c57bSTejun Heo if (worker->current_work == work && 913a2c1c57bSTejun Heo worker->current_func == work->func) 91442f8570fSSasha Levin return worker; 91542f8570fSSasha Levin 91642f8570fSSasha Levin return NULL; 9178cca0eeaSTejun Heo } 9188cca0eeaSTejun Heo 9198cca0eeaSTejun Heo /** 920bf4ede01STejun Heo * move_linked_works - move linked works to a list 921bf4ede01STejun Heo * @work: start of series of works to be scheduled 922bf4ede01STejun Heo * @head: target list to append @work to 923bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 924bf4ede01STejun Heo * 925bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 926bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 927bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 928bf4ede01STejun Heo * 929bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 930bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 931bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 932bf4ede01STejun Heo * 933bf4ede01STejun Heo * CONTEXT: 934d565ed63STejun Heo * spin_lock_irq(pool->lock). 935bf4ede01STejun Heo */ 936bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 937bf4ede01STejun Heo struct work_struct **nextp) 938bf4ede01STejun Heo { 939bf4ede01STejun Heo struct work_struct *n; 940bf4ede01STejun Heo 941bf4ede01STejun Heo /* 942bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 943bf4ede01STejun Heo * use NULL for list head. 944bf4ede01STejun Heo */ 945bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 946bf4ede01STejun Heo list_move_tail(&work->entry, head); 947bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 948bf4ede01STejun Heo break; 949bf4ede01STejun Heo } 950bf4ede01STejun Heo 951bf4ede01STejun Heo /* 952bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 953bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 954bf4ede01STejun Heo * needs to be updated. 955bf4ede01STejun Heo */ 956bf4ede01STejun Heo if (nextp) 957bf4ede01STejun Heo *nextp = n; 958bf4ede01STejun Heo } 959bf4ede01STejun Heo 9608864b4e5STejun Heo /** 9618864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 9628864b4e5STejun Heo * @pwq: pool_workqueue to get 9638864b4e5STejun Heo * 9648864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 9658864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 9668864b4e5STejun Heo */ 9678864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 9688864b4e5STejun Heo { 9698864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 9708864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 9718864b4e5STejun Heo pwq->refcnt++; 9728864b4e5STejun Heo } 9738864b4e5STejun Heo 9748864b4e5STejun Heo /** 9758864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 9768864b4e5STejun Heo * @pwq: pool_workqueue to put 9778864b4e5STejun Heo * 9788864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 9798864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 9808864b4e5STejun Heo */ 9818864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 9828864b4e5STejun Heo { 9838864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 9848864b4e5STejun Heo if (likely(--pwq->refcnt)) 9858864b4e5STejun Heo return; 9868864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 9878864b4e5STejun Heo return; 9888864b4e5STejun Heo /* 9898864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 9908864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 9918864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 9928864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 9938864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 9948864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 9958864b4e5STejun Heo */ 9968864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 9978864b4e5STejun Heo } 9988864b4e5STejun Heo 999112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1000bf4ede01STejun Heo { 1001112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1002bf4ede01STejun Heo 1003bf4ede01STejun Heo trace_workqueue_activate_work(work); 1004112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1005bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1006112202d9STejun Heo pwq->nr_active++; 1007bf4ede01STejun Heo } 1008bf4ede01STejun Heo 1009112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10103aa62497SLai Jiangshan { 1011112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10123aa62497SLai Jiangshan struct work_struct, entry); 10133aa62497SLai Jiangshan 1014112202d9STejun Heo pwq_activate_delayed_work(work); 10153aa62497SLai Jiangshan } 10163aa62497SLai Jiangshan 1017bf4ede01STejun Heo /** 1018112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1019112202d9STejun Heo * @pwq: pwq of interest 1020bf4ede01STejun Heo * @color: color of work which left the queue 1021bf4ede01STejun Heo * 1022bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1023112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1024bf4ede01STejun Heo * 1025bf4ede01STejun Heo * CONTEXT: 1026d565ed63STejun Heo * spin_lock_irq(pool->lock). 1027bf4ede01STejun Heo */ 1028112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1029bf4ede01STejun Heo { 10308864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1031bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10328864b4e5STejun Heo goto out_put; 1033bf4ede01STejun Heo 1034112202d9STejun Heo pwq->nr_in_flight[color]--; 1035bf4ede01STejun Heo 1036112202d9STejun Heo pwq->nr_active--; 1037112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1038bf4ede01STejun Heo /* one down, submit a delayed one */ 1039112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1040112202d9STejun Heo pwq_activate_first_delayed(pwq); 1041bf4ede01STejun Heo } 1042bf4ede01STejun Heo 1043bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1044112202d9STejun Heo if (likely(pwq->flush_color != color)) 10458864b4e5STejun Heo goto out_put; 1046bf4ede01STejun Heo 1047bf4ede01STejun Heo /* are there still in-flight works? */ 1048112202d9STejun Heo if (pwq->nr_in_flight[color]) 10498864b4e5STejun Heo goto out_put; 1050bf4ede01STejun Heo 1051112202d9STejun Heo /* this pwq is done, clear flush_color */ 1052112202d9STejun Heo pwq->flush_color = -1; 1053bf4ede01STejun Heo 1054bf4ede01STejun Heo /* 1055112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1056bf4ede01STejun Heo * will handle the rest. 1057bf4ede01STejun Heo */ 1058112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1059112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 10608864b4e5STejun Heo out_put: 10618864b4e5STejun Heo put_pwq(pwq); 1062bf4ede01STejun Heo } 1063bf4ede01STejun Heo 106436e227d2STejun Heo /** 1065bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 106636e227d2STejun Heo * @work: work item to steal 106736e227d2STejun Heo * @is_dwork: @work is a delayed_work 1068bbb68dfaSTejun Heo * @flags: place to store irq state 106936e227d2STejun Heo * 107036e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 107136e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 107236e227d2STejun Heo * 107336e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 107436e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 107536e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1076bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1077bbb68dfaSTejun Heo * for arbitrarily long 107836e227d2STejun Heo * 1079bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1080e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1081e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1082e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1083bbb68dfaSTejun Heo * 1084bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1085bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1086bbb68dfaSTejun Heo * 1087e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1088bf4ede01STejun Heo */ 1089bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1090bbb68dfaSTejun Heo unsigned long *flags) 1091bf4ede01STejun Heo { 1092d565ed63STejun Heo struct worker_pool *pool; 1093112202d9STejun Heo struct pool_workqueue *pwq; 1094bf4ede01STejun Heo 1095bbb68dfaSTejun Heo local_irq_save(*flags); 1096bbb68dfaSTejun Heo 109736e227d2STejun Heo /* try to steal the timer if it exists */ 109836e227d2STejun Heo if (is_dwork) { 109936e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 110036e227d2STejun Heo 1101e0aecdd8STejun Heo /* 1102e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1103e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1104e0aecdd8STejun Heo * running on the local CPU. 1105e0aecdd8STejun Heo */ 110636e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 110736e227d2STejun Heo return 1; 110836e227d2STejun Heo } 110936e227d2STejun Heo 111036e227d2STejun Heo /* try to claim PENDING the normal way */ 1111bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1112bf4ede01STejun Heo return 0; 1113bf4ede01STejun Heo 1114bf4ede01STejun Heo /* 1115bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1116bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1117bf4ede01STejun Heo */ 1118d565ed63STejun Heo pool = get_work_pool(work); 1119d565ed63STejun Heo if (!pool) 1120bbb68dfaSTejun Heo goto fail; 1121bf4ede01STejun Heo 1122d565ed63STejun Heo spin_lock(&pool->lock); 1123bf4ede01STejun Heo /* 1124112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1125112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1126112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1127112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1128112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11290b3dae68SLai Jiangshan * item is currently queued on that pool. 1130bf4ede01STejun Heo */ 1131112202d9STejun Heo pwq = get_work_pwq(work); 1132112202d9STejun Heo if (pwq && pwq->pool == pool) { 1133bf4ede01STejun Heo debug_work_deactivate(work); 11343aa62497SLai Jiangshan 11353aa62497SLai Jiangshan /* 113616062836STejun Heo * A delayed work item cannot be grabbed directly because 113716062836STejun Heo * it might have linked NO_COLOR work items which, if left 1138112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 113916062836STejun Heo * management later on and cause stall. Make sure the work 114016062836STejun Heo * item is activated before grabbing. 11413aa62497SLai Jiangshan */ 11423aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1143112202d9STejun Heo pwq_activate_delayed_work(work); 11443aa62497SLai Jiangshan 1145bf4ede01STejun Heo list_del_init(&work->entry); 1146112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 114736e227d2STejun Heo 1148112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11494468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11504468a00fSLai Jiangshan 1151d565ed63STejun Heo spin_unlock(&pool->lock); 115236e227d2STejun Heo return 1; 1153bf4ede01STejun Heo } 1154d565ed63STejun Heo spin_unlock(&pool->lock); 1155bbb68dfaSTejun Heo fail: 1156bbb68dfaSTejun Heo local_irq_restore(*flags); 1157bbb68dfaSTejun Heo if (work_is_canceling(work)) 1158bbb68dfaSTejun Heo return -ENOENT; 1159bbb68dfaSTejun Heo cpu_relax(); 116036e227d2STejun Heo return -EAGAIN; 1161bf4ede01STejun Heo } 1162bf4ede01STejun Heo 1163bf4ede01STejun Heo /** 1164706026c2STejun Heo * insert_work - insert a work into a pool 1165112202d9STejun Heo * @pwq: pwq @work belongs to 11664690c4abSTejun Heo * @work: work to insert 11674690c4abSTejun Heo * @head: insertion point 11684690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11694690c4abSTejun Heo * 1170112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1171706026c2STejun Heo * work_struct flags. 11724690c4abSTejun Heo * 11734690c4abSTejun Heo * CONTEXT: 1174d565ed63STejun Heo * spin_lock_irq(pool->lock). 1175365970a1SDavid Howells */ 1176112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1177112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1178b89deed3SOleg Nesterov { 1179112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1180e1d8aa9fSFrederic Weisbecker 11814690c4abSTejun Heo /* we own @work, set data and link */ 1182112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 11831a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 11848864b4e5STejun Heo get_pwq(pwq); 1185e22bee78STejun Heo 1186e22bee78STejun Heo /* 1187c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1188c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1189c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1190e22bee78STejun Heo */ 1191e22bee78STejun Heo smp_mb(); 1192e22bee78STejun Heo 119363d95a91STejun Heo if (__need_more_worker(pool)) 119463d95a91STejun Heo wake_up_worker(pool); 1195b89deed3SOleg Nesterov } 1196b89deed3SOleg Nesterov 1197c8efcc25STejun Heo /* 1198c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 11998d03ecfeSTejun Heo * same workqueue. 1200c8efcc25STejun Heo */ 1201c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1202c8efcc25STejun Heo { 1203c8efcc25STejun Heo struct worker *worker; 1204c8efcc25STejun Heo 12058d03ecfeSTejun Heo worker = current_wq_worker(); 1206c8efcc25STejun Heo /* 12078d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12088d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1209c8efcc25STejun Heo */ 1210112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1211c8efcc25STejun Heo } 1212c8efcc25STejun Heo 1213d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12141da177e4SLinus Torvalds struct work_struct *work) 12151da177e4SLinus Torvalds { 1216112202d9STejun Heo struct pool_workqueue *pwq; 1217c9178087STejun Heo struct worker_pool *last_pool; 12181e19ffc6STejun Heo struct list_head *worklist; 12198a2e8e5dSTejun Heo unsigned int work_flags; 1220b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12218930cabaSTejun Heo 12228930cabaSTejun Heo /* 12238930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12248930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12258930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12268930cabaSTejun Heo * happen with IRQ disabled. 12278930cabaSTejun Heo */ 12288930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12291da177e4SLinus Torvalds 1230dc186ad7SThomas Gleixner debug_work_activate(work); 12311e19ffc6STejun Heo 1232c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1233618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1234c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1235e41e704bSTejun Heo return; 12369e8cd2f5STejun Heo retry: 1237c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1238c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 123957469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1240f3421797STejun Heo cpu = raw_smp_processor_id(); 1241c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1242c9178087STejun Heo } else { 1243c9178087STejun Heo pwq = first_pwq(wq); 1244c9178087STejun Heo } 1245f3421797STejun Heo 124618aa9effSTejun Heo /* 1247c9178087STejun Heo * If @work was previously on a different pool, it might still be 1248c9178087STejun Heo * running there, in which case the work needs to be queued on that 1249c9178087STejun Heo * pool to guarantee non-reentrancy. 125018aa9effSTejun Heo */ 1251c9e7cf27STejun Heo last_pool = get_work_pool(work); 1252112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 125318aa9effSTejun Heo struct worker *worker; 125418aa9effSTejun Heo 1255d565ed63STejun Heo spin_lock(&last_pool->lock); 125618aa9effSTejun Heo 1257c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 125818aa9effSTejun Heo 1259112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1260c9178087STejun Heo pwq = worker->current_pwq; 12618594fadeSLai Jiangshan } else { 126218aa9effSTejun Heo /* meh... not running there, queue here */ 1263d565ed63STejun Heo spin_unlock(&last_pool->lock); 1264112202d9STejun Heo spin_lock(&pwq->pool->lock); 126518aa9effSTejun Heo } 12668930cabaSTejun Heo } else { 1267112202d9STejun Heo spin_lock(&pwq->pool->lock); 12688930cabaSTejun Heo } 1269502ca9d8STejun Heo 12709e8cd2f5STejun Heo /* 12719e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 12729e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 12739e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 12749e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 12759e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 12769e8cd2f5STejun Heo * make forward-progress. 12779e8cd2f5STejun Heo */ 12789e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 12799e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 12809e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 12819e8cd2f5STejun Heo cpu_relax(); 12829e8cd2f5STejun Heo goto retry; 12839e8cd2f5STejun Heo } 12849e8cd2f5STejun Heo /* oops */ 12859e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 12869e8cd2f5STejun Heo wq->name, cpu); 12879e8cd2f5STejun Heo } 12889e8cd2f5STejun Heo 1289112202d9STejun Heo /* pwq determined, queue */ 1290112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1291502ca9d8STejun Heo 1292f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1293112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1294f5b2552bSDan Carpenter return; 1295f5b2552bSDan Carpenter } 12961e19ffc6STejun Heo 1297112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1298112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 12991e19ffc6STejun Heo 1300112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1301cdadf009STejun Heo trace_workqueue_activate_work(work); 1302112202d9STejun Heo pwq->nr_active++; 1303112202d9STejun Heo worklist = &pwq->pool->worklist; 13048a2e8e5dSTejun Heo } else { 13058a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1306112202d9STejun Heo worklist = &pwq->delayed_works; 13078a2e8e5dSTejun Heo } 13081e19ffc6STejun Heo 1309112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13101e19ffc6STejun Heo 1311112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13121da177e4SLinus Torvalds } 13131da177e4SLinus Torvalds 13140fcb78c2SRolf Eike Beer /** 1315c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1316c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1317c1a220e7SZhang Rui * @wq: workqueue to use 1318c1a220e7SZhang Rui * @work: work to queue 1319c1a220e7SZhang Rui * 1320d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1321c1a220e7SZhang Rui * 1322c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1323c1a220e7SZhang Rui * can't go away. 1324c1a220e7SZhang Rui */ 1325d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1326d4283e93STejun Heo struct work_struct *work) 1327c1a220e7SZhang Rui { 1328d4283e93STejun Heo bool ret = false; 13298930cabaSTejun Heo unsigned long flags; 13308930cabaSTejun Heo 13318930cabaSTejun Heo local_irq_save(flags); 1332c1a220e7SZhang Rui 133322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13344690c4abSTejun Heo __queue_work(cpu, wq, work); 1335d4283e93STejun Heo ret = true; 1336c1a220e7SZhang Rui } 13378930cabaSTejun Heo 13388930cabaSTejun Heo local_irq_restore(flags); 1339c1a220e7SZhang Rui return ret; 1340c1a220e7SZhang Rui } 1341c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1342c1a220e7SZhang Rui 13430a13c00eSTejun Heo /** 13440a13c00eSTejun Heo * queue_work - queue work on a workqueue 13450a13c00eSTejun Heo * @wq: workqueue to use 13460a13c00eSTejun Heo * @work: work to queue 13470a13c00eSTejun Heo * 1348d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 13490a13c00eSTejun Heo * 13500a13c00eSTejun Heo * We queue the work to the CPU on which it was submitted, but if the CPU dies 13510a13c00eSTejun Heo * it can be processed by another CPU. 13520a13c00eSTejun Heo */ 1353d4283e93STejun Heo bool queue_work(struct workqueue_struct *wq, struct work_struct *work) 13540a13c00eSTejun Heo { 135557469821STejun Heo return queue_work_on(WORK_CPU_UNBOUND, wq, work); 13560a13c00eSTejun Heo } 13570a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_work); 13580a13c00eSTejun Heo 1359d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13601da177e4SLinus Torvalds { 136152bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13621da177e4SLinus Torvalds 1363e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 136460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13651da177e4SLinus Torvalds } 13661438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13671da177e4SLinus Torvalds 13687beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 136952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13701da177e4SLinus Torvalds { 13717beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13727beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13731da177e4SLinus Torvalds 13747beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13757beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1376fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1377fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13787beb2edfSTejun Heo 13798852aac2STejun Heo /* 13808852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13818852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13828852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13838852aac2STejun Heo * on that there's no such delay when @delay is 0. 13848852aac2STejun Heo */ 13858852aac2STejun Heo if (!delay) { 13868852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 13878852aac2STejun Heo return; 13888852aac2STejun Heo } 13898852aac2STejun Heo 13907beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 13917beb2edfSTejun Heo 139260c057bcSLai Jiangshan dwork->wq = wq; 13931265057fSTejun Heo dwork->cpu = cpu; 13947beb2edfSTejun Heo timer->expires = jiffies + delay; 13957beb2edfSTejun Heo 13967beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 13977beb2edfSTejun Heo add_timer_on(timer, cpu); 13987beb2edfSTejun Heo else 13997beb2edfSTejun Heo add_timer(timer); 14007beb2edfSTejun Heo } 14011da177e4SLinus Torvalds 14020fcb78c2SRolf Eike Beer /** 14030fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14040fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14050fcb78c2SRolf Eike Beer * @wq: workqueue to use 1406af9997e4SRandy Dunlap * @dwork: work to queue 14070fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14080fcb78c2SRolf Eike Beer * 1409715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1410715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1411715f1300STejun Heo * execution. 14120fcb78c2SRolf Eike Beer */ 1413d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 141452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14157a6bc1cdSVenkatesh Pallipadi { 141652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1417d4283e93STejun Heo bool ret = false; 14188930cabaSTejun Heo unsigned long flags; 14198930cabaSTejun Heo 14208930cabaSTejun Heo /* read the comment in __queue_work() */ 14218930cabaSTejun Heo local_irq_save(flags); 14227a6bc1cdSVenkatesh Pallipadi 142322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14247beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1425d4283e93STejun Heo ret = true; 14267a6bc1cdSVenkatesh Pallipadi } 14278930cabaSTejun Heo 14288930cabaSTejun Heo local_irq_restore(flags); 14297a6bc1cdSVenkatesh Pallipadi return ret; 14307a6bc1cdSVenkatesh Pallipadi } 1431ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14321da177e4SLinus Torvalds 1433c8e55f36STejun Heo /** 14340a13c00eSTejun Heo * queue_delayed_work - queue work on a workqueue after delay 14350a13c00eSTejun Heo * @wq: workqueue to use 14360a13c00eSTejun Heo * @dwork: delayable work to queue 14370a13c00eSTejun Heo * @delay: number of jiffies to wait before queueing 14380a13c00eSTejun Heo * 1439715f1300STejun Heo * Equivalent to queue_delayed_work_on() but tries to use the local CPU. 14400a13c00eSTejun Heo */ 1441d4283e93STejun Heo bool queue_delayed_work(struct workqueue_struct *wq, 14420a13c00eSTejun Heo struct delayed_work *dwork, unsigned long delay) 14430a13c00eSTejun Heo { 144457469821STejun Heo return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14450a13c00eSTejun Heo } 14460a13c00eSTejun Heo EXPORT_SYMBOL_GPL(queue_delayed_work); 14470a13c00eSTejun Heo 14480a13c00eSTejun Heo /** 14498376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14508376fe22STejun Heo * @cpu: CPU number to execute work on 14518376fe22STejun Heo * @wq: workqueue to use 14528376fe22STejun Heo * @dwork: work to queue 14538376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14548376fe22STejun Heo * 14558376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14568376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14578376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14588376fe22STejun Heo * current state. 14598376fe22STejun Heo * 14608376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14618376fe22STejun Heo * pending and its timer was modified. 14628376fe22STejun Heo * 1463e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14648376fe22STejun Heo * See try_to_grab_pending() for details. 14658376fe22STejun Heo */ 14668376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14678376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14688376fe22STejun Heo { 14698376fe22STejun Heo unsigned long flags; 14708376fe22STejun Heo int ret; 14718376fe22STejun Heo 14728376fe22STejun Heo do { 14738376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14748376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14758376fe22STejun Heo 14768376fe22STejun Heo if (likely(ret >= 0)) { 14778376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14788376fe22STejun Heo local_irq_restore(flags); 14798376fe22STejun Heo } 14808376fe22STejun Heo 14818376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14828376fe22STejun Heo return ret; 14838376fe22STejun Heo } 14848376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14858376fe22STejun Heo 14868376fe22STejun Heo /** 14878376fe22STejun Heo * mod_delayed_work - modify delay of or queue a delayed work 14888376fe22STejun Heo * @wq: workqueue to use 14898376fe22STejun Heo * @dwork: work to queue 14908376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14918376fe22STejun Heo * 14928376fe22STejun Heo * mod_delayed_work_on() on local CPU. 14938376fe22STejun Heo */ 14948376fe22STejun Heo bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork, 14958376fe22STejun Heo unsigned long delay) 14968376fe22STejun Heo { 14978376fe22STejun Heo return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay); 14988376fe22STejun Heo } 14998376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work); 15008376fe22STejun Heo 15018376fe22STejun Heo /** 1502c8e55f36STejun Heo * worker_enter_idle - enter idle state 1503c8e55f36STejun Heo * @worker: worker which is entering idle state 1504c8e55f36STejun Heo * 1505c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1506c8e55f36STejun Heo * necessary. 1507c8e55f36STejun Heo * 1508c8e55f36STejun Heo * LOCKING: 1509d565ed63STejun Heo * spin_lock_irq(pool->lock). 1510c8e55f36STejun Heo */ 1511c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15121da177e4SLinus Torvalds { 1513bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1514c8e55f36STejun Heo 15156183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15166183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15176183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15186183c009STejun Heo return; 1519c8e55f36STejun Heo 1520cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1521cb444766STejun Heo worker->flags |= WORKER_IDLE; 1522bd7bdd43STejun Heo pool->nr_idle++; 1523e22bee78STejun Heo worker->last_active = jiffies; 1524c8e55f36STejun Heo 1525c8e55f36STejun Heo /* idle_list is LIFO */ 1526bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1527db7bccf4STejun Heo 152863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1529628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1530cb444766STejun Heo 1531544ecf31STejun Heo /* 1532706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1533d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1534628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1535628c78e7STejun Heo * unbind is not in progress. 1536544ecf31STejun Heo */ 153724647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1538bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1539e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1540c8e55f36STejun Heo } 1541c8e55f36STejun Heo 1542c8e55f36STejun Heo /** 1543c8e55f36STejun Heo * worker_leave_idle - leave idle state 1544c8e55f36STejun Heo * @worker: worker which is leaving idle state 1545c8e55f36STejun Heo * 1546c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1547c8e55f36STejun Heo * 1548c8e55f36STejun Heo * LOCKING: 1549d565ed63STejun Heo * spin_lock_irq(pool->lock). 1550c8e55f36STejun Heo */ 1551c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1552c8e55f36STejun Heo { 1553bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1554c8e55f36STejun Heo 15556183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15566183c009STejun Heo return; 1557d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1558bd7bdd43STejun Heo pool->nr_idle--; 1559c8e55f36STejun Heo list_del_init(&worker->entry); 1560c8e55f36STejun Heo } 1561c8e55f36STejun Heo 1562e22bee78STejun Heo /** 1563f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1564f36dc67bSLai Jiangshan * @pool: target worker_pool 1565f36dc67bSLai Jiangshan * 1566f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1567e22bee78STejun Heo * 1568e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1569e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1570e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1571e22bee78STejun Heo * guaranteed to execute on the cpu. 1572e22bee78STejun Heo * 1573f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1574e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1575e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1576e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1577706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1578e22bee78STejun Heo * [dis]associated in the meantime. 1579e22bee78STejun Heo * 1580706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 158124647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1582f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1583f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1584f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1585e22bee78STejun Heo * 1586e22bee78STejun Heo * CONTEXT: 1587d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1588e22bee78STejun Heo * held. 1589e22bee78STejun Heo * 1590e22bee78STejun Heo * RETURNS: 1591706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1592e22bee78STejun Heo * bound), %false if offline. 1593e22bee78STejun Heo */ 1594f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1595d565ed63STejun Heo __acquires(&pool->lock) 1596e22bee78STejun Heo { 1597e22bee78STejun Heo while (true) { 1598e22bee78STejun Heo /* 1599e22bee78STejun Heo * The following call may fail, succeed or succeed 1600e22bee78STejun Heo * without actually migrating the task to the cpu if 1601e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 160224647570STejun Heo * against POOL_DISASSOCIATED. 1603e22bee78STejun Heo */ 160424647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16057a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1606e22bee78STejun Heo 1607d565ed63STejun Heo spin_lock_irq(&pool->lock); 160824647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1609e22bee78STejun Heo return false; 1610f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16117a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1612e22bee78STejun Heo return true; 1613d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1614e22bee78STejun Heo 16155035b20fSTejun Heo /* 16165035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16175035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16185035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16195035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16205035b20fSTejun Heo */ 1621e22bee78STejun Heo cpu_relax(); 16225035b20fSTejun Heo cond_resched(); 1623e22bee78STejun Heo } 1624e22bee78STejun Heo } 1625e22bee78STejun Heo 1626e22bee78STejun Heo /* 1627ea1abd61SLai Jiangshan * Rebind an idle @worker to its CPU. worker_thread() will test 16285f7dabfdSLai Jiangshan * list_empty(@worker->entry) before leaving idle and call this function. 162925511a47STejun Heo */ 163025511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 163125511a47STejun Heo { 16325f7dabfdSLai Jiangshan /* CPU may go down again inbetween, clear UNBOUND only on success */ 1633f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 16345f7dabfdSLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 163525511a47STejun Heo 1636ea1abd61SLai Jiangshan /* rebind complete, become available again */ 1637ea1abd61SLai Jiangshan list_add(&worker->entry, &worker->pool->idle_list); 1638d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 163925511a47STejun Heo } 164025511a47STejun Heo 164125511a47STejun Heo /* 164225511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1643403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1644403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1645403c821dSTejun Heo * executed twice without intervening cpu down. 1646e22bee78STejun Heo */ 164725511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1648e22bee78STejun Heo { 1649e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1650e22bee78STejun Heo 1651f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(worker->pool)) 1652eab6d828SLai Jiangshan worker_clr_flags(worker, WORKER_UNBOUND); 1653e22bee78STejun Heo 1654d565ed63STejun Heo spin_unlock_irq(&worker->pool->lock); 1655e22bee78STejun Heo } 1656e22bee78STejun Heo 165725511a47STejun Heo /** 165894cf58bbSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 165994cf58bbSTejun Heo * @pool: pool of interest 166025511a47STejun Heo * 166194cf58bbSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding 166225511a47STejun Heo * is different for idle and busy ones. 166325511a47STejun Heo * 1664ea1abd61SLai Jiangshan * Idle ones will be removed from the idle_list and woken up. They will 1665ea1abd61SLai Jiangshan * add themselves back after completing rebind. This ensures that the 1666ea1abd61SLai Jiangshan * idle_list doesn't contain any unbound workers when re-bound busy workers 1667ea1abd61SLai Jiangshan * try to perform local wake-ups for concurrency management. 166825511a47STejun Heo * 1669ea1abd61SLai Jiangshan * Busy workers can rebind after they finish their current work items. 1670ea1abd61SLai Jiangshan * Queueing the rebind work item at the head of the scheduled list is 1671ea1abd61SLai Jiangshan * enough. Note that nr_running will be properly bumped as busy workers 1672ea1abd61SLai Jiangshan * rebind. 167325511a47STejun Heo * 1674ea1abd61SLai Jiangshan * On return, all non-manager workers are scheduled for rebind - see 1675ea1abd61SLai Jiangshan * manage_workers() for the manager special case. Any idle worker 1676ea1abd61SLai Jiangshan * including the manager will not appear on @idle_list until rebind is 1677ea1abd61SLai Jiangshan * complete, making local wake-ups safe. 167825511a47STejun Heo */ 167994cf58bbSTejun Heo static void rebind_workers(struct worker_pool *pool) 168025511a47STejun Heo { 1681ea1abd61SLai Jiangshan struct worker *worker, *n; 168225511a47STejun Heo int i; 168325511a47STejun Heo 1684b2eb83d1SLai Jiangshan lockdep_assert_held(&pool->assoc_mutex); 1685d565ed63STejun Heo lockdep_assert_held(&pool->lock); 168625511a47STejun Heo 16875f7dabfdSLai Jiangshan /* dequeue and kick idle ones */ 1688ea1abd61SLai Jiangshan list_for_each_entry_safe(worker, n, &pool->idle_list, entry) { 1689ea1abd61SLai Jiangshan /* 169094cf58bbSTejun Heo * idle workers should be off @pool->idle_list until rebind 169194cf58bbSTejun Heo * is complete to avoid receiving premature local wake-ups. 1692ea1abd61SLai Jiangshan */ 1693ea1abd61SLai Jiangshan list_del_init(&worker->entry); 169425511a47STejun Heo 169525511a47STejun Heo /* 169694cf58bbSTejun Heo * worker_thread() will see the above dequeuing and call 169794cf58bbSTejun Heo * idle_worker_rebind(). 169825511a47STejun Heo */ 169925511a47STejun Heo wake_up_process(worker->task); 170025511a47STejun Heo } 170125511a47STejun Heo 1702ea1abd61SLai Jiangshan /* rebind busy workers */ 1703b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) { 170425511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 1705e2b6a6d5SJoonsoo Kim struct workqueue_struct *wq; 170625511a47STejun Heo 170725511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 170825511a47STejun Heo work_data_bits(rebind_work))) 170925511a47STejun Heo continue; 171025511a47STejun Heo 171125511a47STejun Heo debug_work_activate(rebind_work); 171290beca5dSTejun Heo 171390beca5dSTejun Heo /* 171494cf58bbSTejun Heo * wq doesn't really matter but let's keep @worker->pool 1715112202d9STejun Heo * and @pwq->pool consistent for sanity. 171690beca5dSTejun Heo */ 17177a4e344cSTejun Heo if (worker->pool->attrs->nice < 0) 1718e2b6a6d5SJoonsoo Kim wq = system_highpri_wq; 1719e2b6a6d5SJoonsoo Kim else 1720e2b6a6d5SJoonsoo Kim wq = system_wq; 1721ec58815aSTejun Heo 17227fb98ea7STejun Heo insert_work(per_cpu_ptr(wq->cpu_pwqs, pool->cpu), rebind_work, 172325511a47STejun Heo worker->scheduled.next, 172425511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 1725ec58815aSTejun Heo } 172625511a47STejun Heo } 172725511a47STejun Heo 1728c34056a3STejun Heo static struct worker *alloc_worker(void) 1729c34056a3STejun Heo { 1730c34056a3STejun Heo struct worker *worker; 1731c34056a3STejun Heo 1732c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1733c8e55f36STejun Heo if (worker) { 1734c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1735affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 173625511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1737e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1738e22bee78STejun Heo worker->flags = WORKER_PREP; 1739c8e55f36STejun Heo } 1740c34056a3STejun Heo return worker; 1741c34056a3STejun Heo } 1742c34056a3STejun Heo 1743c34056a3STejun Heo /** 1744c34056a3STejun Heo * create_worker - create a new workqueue worker 174563d95a91STejun Heo * @pool: pool the new worker will belong to 1746c34056a3STejun Heo * 174763d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1748c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1749c34056a3STejun Heo * destroy_worker(). 1750c34056a3STejun Heo * 1751c34056a3STejun Heo * CONTEXT: 1752c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1753c34056a3STejun Heo * 1754c34056a3STejun Heo * RETURNS: 1755c34056a3STejun Heo * Pointer to the newly created worker. 1756c34056a3STejun Heo */ 1757bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1758c34056a3STejun Heo { 17597a4e344cSTejun Heo const char *pri = pool->attrs->nice < 0 ? "H" : ""; 1760c34056a3STejun Heo struct worker *worker = NULL; 1761f3421797STejun Heo int id = -1; 1762c34056a3STejun Heo 1763d565ed63STejun Heo spin_lock_irq(&pool->lock); 1764bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 1765d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1766bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1767c34056a3STejun Heo goto fail; 1768d565ed63STejun Heo spin_lock_irq(&pool->lock); 1769c34056a3STejun Heo } 1770d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1771c34056a3STejun Heo 1772c34056a3STejun Heo worker = alloc_worker(); 1773c34056a3STejun Heo if (!worker) 1774c34056a3STejun Heo goto fail; 1775c34056a3STejun Heo 1776bd7bdd43STejun Heo worker->pool = pool; 1777c34056a3STejun Heo worker->id = id; 1778c34056a3STejun Heo 177929c91e99STejun Heo if (pool->cpu >= 0) 178094dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1781ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1782d84ff051STejun Heo "kworker/%d:%d%s", pool->cpu, id, pri); 1783f3421797STejun Heo else 1784f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1785ac6104cdSTejun Heo "kworker/u%d:%d%s", 1786ac6104cdSTejun Heo pool->id, id, pri); 1787c34056a3STejun Heo if (IS_ERR(worker->task)) 1788c34056a3STejun Heo goto fail; 1789c34056a3STejun Heo 1790c5aa87bbSTejun Heo /* 1791c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1792c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1793c5aa87bbSTejun Heo */ 17947a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 17957a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17963270476aSTejun Heo 1797db7bccf4STejun Heo /* 17987a4e344cSTejun Heo * %PF_THREAD_BOUND is used to prevent userland from meddling with 17997a4e344cSTejun Heo * cpumask of workqueue workers. This is an abuse. We need 18007a4e344cSTejun Heo * %PF_NO_SETAFFINITY. 1801db7bccf4STejun Heo */ 1802db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 18037a4e344cSTejun Heo 18047a4e344cSTejun Heo /* 18057a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 18067a4e344cSTejun Heo * remains stable across this function. See the comments above the 18077a4e344cSTejun Heo * flag definition for details. 18087a4e344cSTejun Heo */ 18097a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1810f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1811c34056a3STejun Heo 1812c34056a3STejun Heo return worker; 1813c34056a3STejun Heo fail: 1814c34056a3STejun Heo if (id >= 0) { 1815d565ed63STejun Heo spin_lock_irq(&pool->lock); 1816bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1817d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1818c34056a3STejun Heo } 1819c34056a3STejun Heo kfree(worker); 1820c34056a3STejun Heo return NULL; 1821c34056a3STejun Heo } 1822c34056a3STejun Heo 1823c34056a3STejun Heo /** 1824c34056a3STejun Heo * start_worker - start a newly created worker 1825c34056a3STejun Heo * @worker: worker to start 1826c34056a3STejun Heo * 1827706026c2STejun Heo * Make the pool aware of @worker and start it. 1828c34056a3STejun Heo * 1829c34056a3STejun Heo * CONTEXT: 1830d565ed63STejun Heo * spin_lock_irq(pool->lock). 1831c34056a3STejun Heo */ 1832c34056a3STejun Heo static void start_worker(struct worker *worker) 1833c34056a3STejun Heo { 1834cb444766STejun Heo worker->flags |= WORKER_STARTED; 1835bd7bdd43STejun Heo worker->pool->nr_workers++; 1836c8e55f36STejun Heo worker_enter_idle(worker); 1837c34056a3STejun Heo wake_up_process(worker->task); 1838c34056a3STejun Heo } 1839c34056a3STejun Heo 1840c34056a3STejun Heo /** 1841c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1842c34056a3STejun Heo * @worker: worker to be destroyed 1843c34056a3STejun Heo * 1844706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1845c8e55f36STejun Heo * 1846c8e55f36STejun Heo * CONTEXT: 1847d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1848c34056a3STejun Heo */ 1849c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1850c34056a3STejun Heo { 1851bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1852c34056a3STejun Heo int id = worker->id; 1853c34056a3STejun Heo 1854c34056a3STejun Heo /* sanity check frenzy */ 18556183c009STejun Heo if (WARN_ON(worker->current_work) || 18566183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18576183c009STejun Heo return; 1858c34056a3STejun Heo 1859c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1860bd7bdd43STejun Heo pool->nr_workers--; 1861c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1862bd7bdd43STejun Heo pool->nr_idle--; 1863c8e55f36STejun Heo 1864c8e55f36STejun Heo list_del_init(&worker->entry); 1865cb444766STejun Heo worker->flags |= WORKER_DIE; 1866c8e55f36STejun Heo 1867d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1868c8e55f36STejun Heo 1869c34056a3STejun Heo kthread_stop(worker->task); 1870c34056a3STejun Heo kfree(worker); 1871c34056a3STejun Heo 1872d565ed63STejun Heo spin_lock_irq(&pool->lock); 1873bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1874c34056a3STejun Heo } 1875c34056a3STejun Heo 187663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1877e22bee78STejun Heo { 187863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1879e22bee78STejun Heo 1880d565ed63STejun Heo spin_lock_irq(&pool->lock); 1881e22bee78STejun Heo 188263d95a91STejun Heo if (too_many_workers(pool)) { 1883e22bee78STejun Heo struct worker *worker; 1884e22bee78STejun Heo unsigned long expires; 1885e22bee78STejun Heo 1886e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 188763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1888e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1889e22bee78STejun Heo 1890e22bee78STejun Heo if (time_before(jiffies, expires)) 189163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1892e22bee78STejun Heo else { 1893e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 189411ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 189563d95a91STejun Heo wake_up_worker(pool); 1896e22bee78STejun Heo } 1897e22bee78STejun Heo } 1898e22bee78STejun Heo 1899d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1900e22bee78STejun Heo } 1901e22bee78STejun Heo 1902493a1724STejun Heo static void send_mayday(struct work_struct *work) 1903e22bee78STejun Heo { 1904112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1905112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1906493a1724STejun Heo 1907493a1724STejun Heo lockdep_assert_held(&workqueue_lock); 1908e22bee78STejun Heo 1909493008a8STejun Heo if (!wq->rescuer) 1910493a1724STejun Heo return; 1911e22bee78STejun Heo 1912e22bee78STejun Heo /* mayday mayday mayday */ 1913493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1914493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1915e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1916493a1724STejun Heo } 1917e22bee78STejun Heo } 1918e22bee78STejun Heo 1919706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1920e22bee78STejun Heo { 192163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1922e22bee78STejun Heo struct work_struct *work; 1923e22bee78STejun Heo 1924493a1724STejun Heo spin_lock_irq(&workqueue_lock); /* for wq->maydays */ 1925493a1724STejun Heo spin_lock(&pool->lock); 1926e22bee78STejun Heo 192763d95a91STejun Heo if (need_to_create_worker(pool)) { 1928e22bee78STejun Heo /* 1929e22bee78STejun Heo * We've been trying to create a new worker but 1930e22bee78STejun Heo * haven't been successful. We might be hitting an 1931e22bee78STejun Heo * allocation deadlock. Send distress signals to 1932e22bee78STejun Heo * rescuers. 1933e22bee78STejun Heo */ 193463d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1935e22bee78STejun Heo send_mayday(work); 1936e22bee78STejun Heo } 1937e22bee78STejun Heo 1938493a1724STejun Heo spin_unlock(&pool->lock); 1939493a1724STejun Heo spin_unlock_irq(&workqueue_lock); 1940e22bee78STejun Heo 194163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1942e22bee78STejun Heo } 1943e22bee78STejun Heo 1944e22bee78STejun Heo /** 1945e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 194663d95a91STejun Heo * @pool: pool to create a new worker for 1947e22bee78STejun Heo * 194863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1949e22bee78STejun Heo * have at least one idle worker on return from this function. If 1950e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 195163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1952e22bee78STejun Heo * possible allocation deadlock. 1953e22bee78STejun Heo * 1954c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1955c5aa87bbSTejun Heo * may_start_working() %true. 1956e22bee78STejun Heo * 1957e22bee78STejun Heo * LOCKING: 1958d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1959e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1960e22bee78STejun Heo * manager. 1961e22bee78STejun Heo * 1962e22bee78STejun Heo * RETURNS: 1963c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1964e22bee78STejun Heo * otherwise. 1965e22bee78STejun Heo */ 196663d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1967d565ed63STejun Heo __releases(&pool->lock) 1968d565ed63STejun Heo __acquires(&pool->lock) 1969e22bee78STejun Heo { 197063d95a91STejun Heo if (!need_to_create_worker(pool)) 1971e22bee78STejun Heo return false; 1972e22bee78STejun Heo restart: 1973d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19749f9c2364STejun Heo 1975e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 197663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1977e22bee78STejun Heo 1978e22bee78STejun Heo while (true) { 1979e22bee78STejun Heo struct worker *worker; 1980e22bee78STejun Heo 1981bc2ae0f5STejun Heo worker = create_worker(pool); 1982e22bee78STejun Heo if (worker) { 198363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1984d565ed63STejun Heo spin_lock_irq(&pool->lock); 1985e22bee78STejun Heo start_worker(worker); 19866183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19876183c009STejun Heo goto restart; 1988e22bee78STejun Heo return true; 1989e22bee78STejun Heo } 1990e22bee78STejun Heo 199163d95a91STejun Heo if (!need_to_create_worker(pool)) 1992e22bee78STejun Heo break; 1993e22bee78STejun Heo 1994e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1995e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19969f9c2364STejun Heo 199763d95a91STejun Heo if (!need_to_create_worker(pool)) 1998e22bee78STejun Heo break; 1999e22bee78STejun Heo } 2000e22bee78STejun Heo 200163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2002d565ed63STejun Heo spin_lock_irq(&pool->lock); 200363d95a91STejun Heo if (need_to_create_worker(pool)) 2004e22bee78STejun Heo goto restart; 2005e22bee78STejun Heo return true; 2006e22bee78STejun Heo } 2007e22bee78STejun Heo 2008e22bee78STejun Heo /** 2009e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 201063d95a91STejun Heo * @pool: pool to destroy workers for 2011e22bee78STejun Heo * 201263d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 2013e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 2014e22bee78STejun Heo * 2015e22bee78STejun Heo * LOCKING: 2016d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2017e22bee78STejun Heo * multiple times. Called only from manager. 2018e22bee78STejun Heo * 2019e22bee78STejun Heo * RETURNS: 2020c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 2021e22bee78STejun Heo * otherwise. 2022e22bee78STejun Heo */ 202363d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 2024e22bee78STejun Heo { 2025e22bee78STejun Heo bool ret = false; 2026e22bee78STejun Heo 202763d95a91STejun Heo while (too_many_workers(pool)) { 2028e22bee78STejun Heo struct worker *worker; 2029e22bee78STejun Heo unsigned long expires; 2030e22bee78STejun Heo 203163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2032e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2033e22bee78STejun Heo 2034e22bee78STejun Heo if (time_before(jiffies, expires)) { 203563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2036e22bee78STejun Heo break; 2037e22bee78STejun Heo } 2038e22bee78STejun Heo 2039e22bee78STejun Heo destroy_worker(worker); 2040e22bee78STejun Heo ret = true; 2041e22bee78STejun Heo } 2042e22bee78STejun Heo 2043e22bee78STejun Heo return ret; 2044e22bee78STejun Heo } 2045e22bee78STejun Heo 2046e22bee78STejun Heo /** 2047e22bee78STejun Heo * manage_workers - manage worker pool 2048e22bee78STejun Heo * @worker: self 2049e22bee78STejun Heo * 2050706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2051e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2052706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2053e22bee78STejun Heo * 2054e22bee78STejun Heo * The caller can safely start processing works on false return. On 2055e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2056e22bee78STejun Heo * and may_start_working() is true. 2057e22bee78STejun Heo * 2058e22bee78STejun Heo * CONTEXT: 2059d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2060e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2061e22bee78STejun Heo * 2062e22bee78STejun Heo * RETURNS: 2063d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2064d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2065e22bee78STejun Heo */ 2066e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2067e22bee78STejun Heo { 206863d95a91STejun Heo struct worker_pool *pool = worker->pool; 2069e22bee78STejun Heo bool ret = false; 2070e22bee78STejun Heo 207134a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2072e22bee78STejun Heo return ret; 2073e22bee78STejun Heo 2074ee378aa4SLai Jiangshan /* 2075ee378aa4SLai Jiangshan * To simplify both worker management and CPU hotplug, hold off 2076ee378aa4SLai Jiangshan * management while hotplug is in progress. CPU hotplug path can't 207734a06bd6STejun Heo * grab @pool->manager_arb to achieve this because that can lead to 207834a06bd6STejun Heo * idle worker depletion (all become busy thinking someone else is 207934a06bd6STejun Heo * managing) which in turn can result in deadlock under extreme 208034a06bd6STejun Heo * circumstances. Use @pool->assoc_mutex to synchronize manager 208134a06bd6STejun Heo * against CPU hotplug. 2082ee378aa4SLai Jiangshan * 2083b2eb83d1SLai Jiangshan * assoc_mutex would always be free unless CPU hotplug is in 2084d565ed63STejun Heo * progress. trylock first without dropping @pool->lock. 2085ee378aa4SLai Jiangshan */ 2086b2eb83d1SLai Jiangshan if (unlikely(!mutex_trylock(&pool->assoc_mutex))) { 2087d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2088b2eb83d1SLai Jiangshan mutex_lock(&pool->assoc_mutex); 2089ee378aa4SLai Jiangshan /* 2090ee378aa4SLai Jiangshan * CPU hotplug could have happened while we were waiting 2091b2eb83d1SLai Jiangshan * for assoc_mutex. Hotplug itself can't handle us 2092ee378aa4SLai Jiangshan * because manager isn't either on idle or busy list, and 2093706026c2STejun Heo * @pool's state and ours could have deviated. 2094ee378aa4SLai Jiangshan * 2095b2eb83d1SLai Jiangshan * As hotplug is now excluded via assoc_mutex, we can 2096ee378aa4SLai Jiangshan * simply try to bind. It will succeed or fail depending 2097706026c2STejun Heo * on @pool's current state. Try it and adjust 2098ee378aa4SLai Jiangshan * %WORKER_UNBOUND accordingly. 2099ee378aa4SLai Jiangshan */ 2100f36dc67bSLai Jiangshan if (worker_maybe_bind_and_lock(pool)) 2101ee378aa4SLai Jiangshan worker->flags &= ~WORKER_UNBOUND; 2102ee378aa4SLai Jiangshan else 2103ee378aa4SLai Jiangshan worker->flags |= WORKER_UNBOUND; 2104ee378aa4SLai Jiangshan 2105ee378aa4SLai Jiangshan ret = true; 2106ee378aa4SLai Jiangshan } 2107ee378aa4SLai Jiangshan 210811ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2109e22bee78STejun Heo 2110e22bee78STejun Heo /* 2111e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2112e22bee78STejun Heo * on return. 2113e22bee78STejun Heo */ 211463d95a91STejun Heo ret |= maybe_destroy_workers(pool); 211563d95a91STejun Heo ret |= maybe_create_worker(pool); 2116e22bee78STejun Heo 2117b2eb83d1SLai Jiangshan mutex_unlock(&pool->assoc_mutex); 211834a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2119e22bee78STejun Heo return ret; 2120e22bee78STejun Heo } 2121e22bee78STejun Heo 2122a62428c0STejun Heo /** 2123a62428c0STejun Heo * process_one_work - process single work 2124c34056a3STejun Heo * @worker: self 2125a62428c0STejun Heo * @work: work to process 2126a62428c0STejun Heo * 2127a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2128a62428c0STejun Heo * process a single work including synchronization against and 2129a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2130a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2131a62428c0STejun Heo * call this function to process a work. 2132a62428c0STejun Heo * 2133a62428c0STejun Heo * CONTEXT: 2134d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2135a62428c0STejun Heo */ 2136c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2137d565ed63STejun Heo __releases(&pool->lock) 2138d565ed63STejun Heo __acquires(&pool->lock) 21391da177e4SLinus Torvalds { 2140112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2141bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2142112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 214373f53c4aSTejun Heo int work_color; 21447e11629dSTejun Heo struct worker *collision; 21454e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21464e6045f1SJohannes Berg /* 2147a62428c0STejun Heo * It is permissible to free the struct work_struct from 2148a62428c0STejun Heo * inside the function that is called from it, this we need to 2149a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2150a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2151a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21524e6045f1SJohannes Berg */ 21534d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21544d82a1deSPeter Zijlstra 21554d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21564e6045f1SJohannes Berg #endif 21576fec10a1STejun Heo /* 21586fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21596fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 216024647570STejun Heo * unbound or a disassociated pool. 21616fec10a1STejun Heo */ 21625f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 216324647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2164ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 216525511a47STejun Heo 21667e11629dSTejun Heo /* 21677e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21687e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21697e11629dSTejun Heo * already processing the work. If so, defer the work to the 21707e11629dSTejun Heo * currently executing one. 21717e11629dSTejun Heo */ 2172c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21737e11629dSTejun Heo if (unlikely(collision)) { 21747e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21757e11629dSTejun Heo return; 21767e11629dSTejun Heo } 21771da177e4SLinus Torvalds 21788930cabaSTejun Heo /* claim and dequeue */ 21791da177e4SLinus Torvalds debug_work_deactivate(work); 2180c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2181c34056a3STejun Heo worker->current_work = work; 2182a2c1c57bSTejun Heo worker->current_func = work->func; 2183112202d9STejun Heo worker->current_pwq = pwq; 218473f53c4aSTejun Heo work_color = get_work_color(work); 21857a22ad75STejun Heo 2186a62428c0STejun Heo list_del_init(&work->entry); 2187a62428c0STejun Heo 2188649027d7STejun Heo /* 2189fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2190fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2191fb0e7bebSTejun Heo */ 2192fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2193fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2194fb0e7bebSTejun Heo 2195974271c4STejun Heo /* 2196d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2197974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2198974271c4STejun Heo */ 219963d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 220063d95a91STejun Heo wake_up_worker(pool); 2201974271c4STejun Heo 22028930cabaSTejun Heo /* 22037c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2204d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 220523657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 220623657bb1STejun Heo * disabled. 22078930cabaSTejun Heo */ 22087c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22091da177e4SLinus Torvalds 2210d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2211365970a1SDavid Howells 2212112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 22133295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2214e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2215a2c1c57bSTejun Heo worker->current_func(work); 2216e36c886aSArjan van de Ven /* 2217e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2218e36c886aSArjan van de Ven * point will only record its address. 2219e36c886aSArjan van de Ven */ 2220e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 22213295f0efSIngo Molnar lock_map_release(&lockdep_map); 2222112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22231da177e4SLinus Torvalds 2224d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2225044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2226044c782cSValentin Ilie " last function: %pf\n", 2227a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2228a2c1c57bSTejun Heo worker->current_func); 2229d5abe669SPeter Zijlstra debug_show_held_locks(current); 2230d5abe669SPeter Zijlstra dump_stack(); 2231d5abe669SPeter Zijlstra } 2232d5abe669SPeter Zijlstra 2233d565ed63STejun Heo spin_lock_irq(&pool->lock); 2234a62428c0STejun Heo 2235fb0e7bebSTejun Heo /* clear cpu intensive status */ 2236fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2237fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2238fb0e7bebSTejun Heo 2239a62428c0STejun Heo /* we're done with it, release */ 224042f8570fSSasha Levin hash_del(&worker->hentry); 2241c34056a3STejun Heo worker->current_work = NULL; 2242a2c1c57bSTejun Heo worker->current_func = NULL; 2243112202d9STejun Heo worker->current_pwq = NULL; 2244112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22451da177e4SLinus Torvalds } 22461da177e4SLinus Torvalds 2247affee4b2STejun Heo /** 2248affee4b2STejun Heo * process_scheduled_works - process scheduled works 2249affee4b2STejun Heo * @worker: self 2250affee4b2STejun Heo * 2251affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2252affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2253affee4b2STejun Heo * fetches a work from the top and executes it. 2254affee4b2STejun Heo * 2255affee4b2STejun Heo * CONTEXT: 2256d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2257affee4b2STejun Heo * multiple times. 2258affee4b2STejun Heo */ 2259affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22601da177e4SLinus Torvalds { 2261affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2262affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2263a62428c0STejun Heo struct work_struct, entry); 2264c34056a3STejun Heo process_one_work(worker, work); 2265a62428c0STejun Heo } 22661da177e4SLinus Torvalds } 22671da177e4SLinus Torvalds 22684690c4abSTejun Heo /** 22694690c4abSTejun Heo * worker_thread - the worker thread function 2270c34056a3STejun Heo * @__worker: self 22714690c4abSTejun Heo * 2272c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2273c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2274c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2275c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2276c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22774690c4abSTejun Heo */ 2278c34056a3STejun Heo static int worker_thread(void *__worker) 22791da177e4SLinus Torvalds { 2280c34056a3STejun Heo struct worker *worker = __worker; 2281bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22821da177e4SLinus Torvalds 2283e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2284e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2285c8e55f36STejun Heo woke_up: 2286d565ed63STejun Heo spin_lock_irq(&pool->lock); 2287affee4b2STejun Heo 22885f7dabfdSLai Jiangshan /* we are off idle list if destruction or rebind is requested */ 22895f7dabfdSLai Jiangshan if (unlikely(list_empty(&worker->entry))) { 2290d565ed63STejun Heo spin_unlock_irq(&pool->lock); 229125511a47STejun Heo 22925f7dabfdSLai Jiangshan /* if DIE is set, destruction is requested */ 229325511a47STejun Heo if (worker->flags & WORKER_DIE) { 2294e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2295c8e55f36STejun Heo return 0; 2296c8e55f36STejun Heo } 2297c8e55f36STejun Heo 22985f7dabfdSLai Jiangshan /* otherwise, rebind */ 229925511a47STejun Heo idle_worker_rebind(worker); 230025511a47STejun Heo goto woke_up; 230125511a47STejun Heo } 230225511a47STejun Heo 2303c8e55f36STejun Heo worker_leave_idle(worker); 2304db7bccf4STejun Heo recheck: 2305e22bee78STejun Heo /* no more worker necessary? */ 230663d95a91STejun Heo if (!need_more_worker(pool)) 2307e22bee78STejun Heo goto sleep; 2308e22bee78STejun Heo 2309e22bee78STejun Heo /* do we need to manage? */ 231063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2311e22bee78STejun Heo goto recheck; 2312e22bee78STejun Heo 2313c8e55f36STejun Heo /* 2314c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2315c8e55f36STejun Heo * preparing to process a work or actually processing it. 2316c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2317c8e55f36STejun Heo */ 23186183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2319c8e55f36STejun Heo 2320e22bee78STejun Heo /* 2321e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2322e22bee78STejun Heo * at least one idle worker or that someone else has already 2323e22bee78STejun Heo * assumed the manager role. 2324e22bee78STejun Heo */ 2325e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2326e22bee78STejun Heo 2327e22bee78STejun Heo do { 2328affee4b2STejun Heo struct work_struct *work = 2329bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2330affee4b2STejun Heo struct work_struct, entry); 2331affee4b2STejun Heo 2332c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2333affee4b2STejun Heo /* optimization path, not strictly necessary */ 2334affee4b2STejun Heo process_one_work(worker, work); 2335affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2336affee4b2STejun Heo process_scheduled_works(worker); 2337affee4b2STejun Heo } else { 2338c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2339affee4b2STejun Heo process_scheduled_works(worker); 2340affee4b2STejun Heo } 234163d95a91STejun Heo } while (keep_working(pool)); 2342affee4b2STejun Heo 2343e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2344d313dd85STejun Heo sleep: 234563d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2346e22bee78STejun Heo goto recheck; 2347d313dd85STejun Heo 2348c8e55f36STejun Heo /* 2349d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2350d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2351d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2352d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2353d565ed63STejun Heo * event. 2354c8e55f36STejun Heo */ 2355c8e55f36STejun Heo worker_enter_idle(worker); 2356c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2357d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23581da177e4SLinus Torvalds schedule(); 2359c8e55f36STejun Heo goto woke_up; 23601da177e4SLinus Torvalds } 23611da177e4SLinus Torvalds 2362e22bee78STejun Heo /** 2363e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2364111c225aSTejun Heo * @__rescuer: self 2365e22bee78STejun Heo * 2366e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2367493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2368e22bee78STejun Heo * 2369706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2370e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2371e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2372e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2373e22bee78STejun Heo * the problem rescuer solves. 2374e22bee78STejun Heo * 2375706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2376706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2377e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2378e22bee78STejun Heo * 2379e22bee78STejun Heo * This should happen rarely. 2380e22bee78STejun Heo */ 2381111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2382e22bee78STejun Heo { 2383111c225aSTejun Heo struct worker *rescuer = __rescuer; 2384111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2385e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2386e22bee78STejun Heo 2387e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2388111c225aSTejun Heo 2389111c225aSTejun Heo /* 2390111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2391111c225aSTejun Heo * doesn't participate in concurrency management. 2392111c225aSTejun Heo */ 2393111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2394e22bee78STejun Heo repeat: 2395e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23961da177e4SLinus Torvalds 2397412d32e6SMike Galbraith if (kthread_should_stop()) { 2398412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2399111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2400e22bee78STejun Heo return 0; 2401412d32e6SMike Galbraith } 24021da177e4SLinus Torvalds 2403493a1724STejun Heo /* see whether any pwq is asking for help */ 2404493a1724STejun Heo spin_lock_irq(&workqueue_lock); 2405493a1724STejun Heo 2406493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2407493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2408493a1724STejun Heo struct pool_workqueue, mayday_node); 2409112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2410e22bee78STejun Heo struct work_struct *work, *n; 2411e22bee78STejun Heo 2412e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2413493a1724STejun Heo list_del_init(&pwq->mayday_node); 2414493a1724STejun Heo 2415493a1724STejun Heo spin_unlock_irq(&workqueue_lock); 2416e22bee78STejun Heo 2417e22bee78STejun Heo /* migrate to the target cpu if possible */ 2418f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2419b3104104SLai Jiangshan rescuer->pool = pool; 2420e22bee78STejun Heo 2421e22bee78STejun Heo /* 2422e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2423e22bee78STejun Heo * process'em. 2424e22bee78STejun Heo */ 24256183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2426bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2427112202d9STejun Heo if (get_work_pwq(work) == pwq) 2428e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2429e22bee78STejun Heo 2430e22bee78STejun Heo process_scheduled_works(rescuer); 24317576958aSTejun Heo 24327576958aSTejun Heo /* 2433d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 24347576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24357576958aSTejun Heo * and stalling the execution. 24367576958aSTejun Heo */ 243763d95a91STejun Heo if (keep_working(pool)) 243863d95a91STejun Heo wake_up_worker(pool); 24397576958aSTejun Heo 2440b3104104SLai Jiangshan rescuer->pool = NULL; 2441493a1724STejun Heo spin_unlock(&pool->lock); 2442493a1724STejun Heo spin_lock(&workqueue_lock); 24431da177e4SLinus Torvalds } 24441da177e4SLinus Torvalds 2445493a1724STejun Heo spin_unlock_irq(&workqueue_lock); 2446493a1724STejun Heo 2447111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2448111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2449e22bee78STejun Heo schedule(); 2450e22bee78STejun Heo goto repeat; 24511da177e4SLinus Torvalds } 24521da177e4SLinus Torvalds 2453fc2e4d70SOleg Nesterov struct wq_barrier { 2454fc2e4d70SOleg Nesterov struct work_struct work; 2455fc2e4d70SOleg Nesterov struct completion done; 2456fc2e4d70SOleg Nesterov }; 2457fc2e4d70SOleg Nesterov 2458fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2459fc2e4d70SOleg Nesterov { 2460fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2461fc2e4d70SOleg Nesterov complete(&barr->done); 2462fc2e4d70SOleg Nesterov } 2463fc2e4d70SOleg Nesterov 24644690c4abSTejun Heo /** 24654690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2466112202d9STejun Heo * @pwq: pwq to insert barrier into 24674690c4abSTejun Heo * @barr: wq_barrier to insert 2468affee4b2STejun Heo * @target: target work to attach @barr to 2469affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24704690c4abSTejun Heo * 2471affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2472affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2473affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2474affee4b2STejun Heo * cpu. 2475affee4b2STejun Heo * 2476affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2477affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2478affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2479affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2480affee4b2STejun Heo * after a work with LINKED flag set. 2481affee4b2STejun Heo * 2482affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2483112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24844690c4abSTejun Heo * 24854690c4abSTejun Heo * CONTEXT: 2486d565ed63STejun Heo * spin_lock_irq(pool->lock). 24874690c4abSTejun Heo */ 2488112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2489affee4b2STejun Heo struct wq_barrier *barr, 2490affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2491fc2e4d70SOleg Nesterov { 2492affee4b2STejun Heo struct list_head *head; 2493affee4b2STejun Heo unsigned int linked = 0; 2494affee4b2STejun Heo 2495dc186ad7SThomas Gleixner /* 2496d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2497dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2498dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2499dc186ad7SThomas Gleixner * might deadlock. 2500dc186ad7SThomas Gleixner */ 2501ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 250222df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2503fc2e4d70SOleg Nesterov init_completion(&barr->done); 250483c22520SOleg Nesterov 2505affee4b2STejun Heo /* 2506affee4b2STejun Heo * If @target is currently being executed, schedule the 2507affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2508affee4b2STejun Heo */ 2509affee4b2STejun Heo if (worker) 2510affee4b2STejun Heo head = worker->scheduled.next; 2511affee4b2STejun Heo else { 2512affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2513affee4b2STejun Heo 2514affee4b2STejun Heo head = target->entry.next; 2515affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2516affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2517affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2518affee4b2STejun Heo } 2519affee4b2STejun Heo 2520dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2521112202d9STejun Heo insert_work(pwq, &barr->work, head, 2522affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2523fc2e4d70SOleg Nesterov } 2524fc2e4d70SOleg Nesterov 252573f53c4aSTejun Heo /** 2526112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 252773f53c4aSTejun Heo * @wq: workqueue being flushed 252873f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 252973f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 253073f53c4aSTejun Heo * 2531112202d9STejun Heo * Prepare pwqs for workqueue flushing. 253273f53c4aSTejun Heo * 2533112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2534112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2535112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2536112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2537112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 253873f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 253973f53c4aSTejun Heo * 254073f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 254173f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 254273f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 254373f53c4aSTejun Heo * is returned. 254473f53c4aSTejun Heo * 2545112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 254673f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 254773f53c4aSTejun Heo * advanced to @work_color. 254873f53c4aSTejun Heo * 254973f53c4aSTejun Heo * CONTEXT: 255073f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 255173f53c4aSTejun Heo * 255273f53c4aSTejun Heo * RETURNS: 255373f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 255473f53c4aSTejun Heo * otherwise. 255573f53c4aSTejun Heo */ 2556112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 255773f53c4aSTejun Heo int flush_color, int work_color) 25581da177e4SLinus Torvalds { 255973f53c4aSTejun Heo bool wait = false; 256049e3cf44STejun Heo struct pool_workqueue *pwq; 25611da177e4SLinus Torvalds 256273f53c4aSTejun Heo if (flush_color >= 0) { 25636183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2564112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2565dc186ad7SThomas Gleixner } 256614441960SOleg Nesterov 256776af4d93STejun Heo local_irq_disable(); 256876af4d93STejun Heo 256949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2570112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25711da177e4SLinus Torvalds 257276af4d93STejun Heo spin_lock(&pool->lock); 257373f53c4aSTejun Heo 257473f53c4aSTejun Heo if (flush_color >= 0) { 25756183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 257673f53c4aSTejun Heo 2577112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2578112202d9STejun Heo pwq->flush_color = flush_color; 2579112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 258073f53c4aSTejun Heo wait = true; 25811da177e4SLinus Torvalds } 258273f53c4aSTejun Heo } 258373f53c4aSTejun Heo 258473f53c4aSTejun Heo if (work_color >= 0) { 25856183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2586112202d9STejun Heo pwq->work_color = work_color; 258773f53c4aSTejun Heo } 258873f53c4aSTejun Heo 258976af4d93STejun Heo spin_unlock(&pool->lock); 25901da177e4SLinus Torvalds } 25911da177e4SLinus Torvalds 259276af4d93STejun Heo local_irq_enable(); 259376af4d93STejun Heo 2594112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 259573f53c4aSTejun Heo complete(&wq->first_flusher->done); 259673f53c4aSTejun Heo 259773f53c4aSTejun Heo return wait; 259883c22520SOleg Nesterov } 25991da177e4SLinus Torvalds 26000fcb78c2SRolf Eike Beer /** 26011da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26020fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26031da177e4SLinus Torvalds * 2604c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2605c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26061da177e4SLinus Torvalds */ 26077ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26081da177e4SLinus Torvalds { 260973f53c4aSTejun Heo struct wq_flusher this_flusher = { 261073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 261173f53c4aSTejun Heo .flush_color = -1, 261273f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 261373f53c4aSTejun Heo }; 261473f53c4aSTejun Heo int next_color; 2615b1f4ec17SOleg Nesterov 26163295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 26173295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 261873f53c4aSTejun Heo 261973f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 262073f53c4aSTejun Heo 262173f53c4aSTejun Heo /* 262273f53c4aSTejun Heo * Start-to-wait phase 262373f53c4aSTejun Heo */ 262473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 262573f53c4aSTejun Heo 262673f53c4aSTejun Heo if (next_color != wq->flush_color) { 262773f53c4aSTejun Heo /* 262873f53c4aSTejun Heo * Color space is not full. The current work_color 262973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 263073f53c4aSTejun Heo * by one. 263173f53c4aSTejun Heo */ 26326183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 263373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 263473f53c4aSTejun Heo wq->work_color = next_color; 263573f53c4aSTejun Heo 263673f53c4aSTejun Heo if (!wq->first_flusher) { 263773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26386183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 264173f53c4aSTejun Heo 2642112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 264373f53c4aSTejun Heo wq->work_color)) { 264473f53c4aSTejun Heo /* nothing to flush, done */ 264573f53c4aSTejun Heo wq->flush_color = next_color; 264673f53c4aSTejun Heo wq->first_flusher = NULL; 264773f53c4aSTejun Heo goto out_unlock; 264873f53c4aSTejun Heo } 264973f53c4aSTejun Heo } else { 265073f53c4aSTejun Heo /* wait in queue */ 26516183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 265273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2653112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 265473f53c4aSTejun Heo } 265573f53c4aSTejun Heo } else { 265673f53c4aSTejun Heo /* 265773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 265873f53c4aSTejun Heo * The next flush completion will assign us 265973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 266073f53c4aSTejun Heo */ 266173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 266273f53c4aSTejun Heo } 266373f53c4aSTejun Heo 266473f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 266573f53c4aSTejun Heo 266673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 266773f53c4aSTejun Heo 266873f53c4aSTejun Heo /* 266973f53c4aSTejun Heo * Wake-up-and-cascade phase 267073f53c4aSTejun Heo * 267173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 267273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 267373f53c4aSTejun Heo */ 267473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 267573f53c4aSTejun Heo return; 267673f53c4aSTejun Heo 267773f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 267873f53c4aSTejun Heo 26794ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26804ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26814ce48b37STejun Heo goto out_unlock; 26824ce48b37STejun Heo 268373f53c4aSTejun Heo wq->first_flusher = NULL; 268473f53c4aSTejun Heo 26856183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26866183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 268773f53c4aSTejun Heo 268873f53c4aSTejun Heo while (true) { 268973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 269273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 269373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 269473f53c4aSTejun Heo break; 269573f53c4aSTejun Heo list_del_init(&next->list); 269673f53c4aSTejun Heo complete(&next->done); 269773f53c4aSTejun Heo } 269873f53c4aSTejun Heo 26996183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 270073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 270173f53c4aSTejun Heo 270273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 270373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 270473f53c4aSTejun Heo 270573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 270673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 270773f53c4aSTejun Heo /* 270873f53c4aSTejun Heo * Assign the same color to all overflowed 270973f53c4aSTejun Heo * flushers, advance work_color and append to 271073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 271173f53c4aSTejun Heo * phase for these overflowed flushers. 271273f53c4aSTejun Heo */ 271373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 271473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 271573f53c4aSTejun Heo 271673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 271773f53c4aSTejun Heo 271873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 271973f53c4aSTejun Heo &wq->flusher_queue); 2720112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 272173f53c4aSTejun Heo } 272273f53c4aSTejun Heo 272373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27246183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 272573f53c4aSTejun Heo break; 272673f53c4aSTejun Heo } 272773f53c4aSTejun Heo 272873f53c4aSTejun Heo /* 272973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2730112202d9STejun Heo * the new first flusher and arm pwqs. 273173f53c4aSTejun Heo */ 27326183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27336183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 273473f53c4aSTejun Heo 273573f53c4aSTejun Heo list_del_init(&next->list); 273673f53c4aSTejun Heo wq->first_flusher = next; 273773f53c4aSTejun Heo 2738112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 273973f53c4aSTejun Heo break; 274073f53c4aSTejun Heo 274173f53c4aSTejun Heo /* 274273f53c4aSTejun Heo * Meh... this color is already done, clear first 274373f53c4aSTejun Heo * flusher and repeat cascading. 274473f53c4aSTejun Heo */ 274573f53c4aSTejun Heo wq->first_flusher = NULL; 274673f53c4aSTejun Heo } 274773f53c4aSTejun Heo 274873f53c4aSTejun Heo out_unlock: 274973f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 27501da177e4SLinus Torvalds } 2751ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27521da177e4SLinus Torvalds 27539c5a2ba7STejun Heo /** 27549c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27559c5a2ba7STejun Heo * @wq: workqueue to drain 27569c5a2ba7STejun Heo * 27579c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27589c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27599c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27609c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27619c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27629c5a2ba7STejun Heo * takes too long. 27639c5a2ba7STejun Heo */ 27649c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27659c5a2ba7STejun Heo { 27669c5a2ba7STejun Heo unsigned int flush_cnt = 0; 276749e3cf44STejun Heo struct pool_workqueue *pwq; 27689c5a2ba7STejun Heo 27699c5a2ba7STejun Heo /* 27709c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27719c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2772618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27739c5a2ba7STejun Heo */ 2774e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 27759c5a2ba7STejun Heo if (!wq->nr_drainers++) 2776618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 2777e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 27789c5a2ba7STejun Heo reflush: 27799c5a2ba7STejun Heo flush_workqueue(wq); 27809c5a2ba7STejun Heo 278176af4d93STejun Heo local_irq_disable(); 278276af4d93STejun Heo 278349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2784fa2563e4SThomas Tuttle bool drained; 27859c5a2ba7STejun Heo 278676af4d93STejun Heo spin_lock(&pwq->pool->lock); 2787112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 278876af4d93STejun Heo spin_unlock(&pwq->pool->lock); 2789fa2563e4SThomas Tuttle 2790fa2563e4SThomas Tuttle if (drained) 27919c5a2ba7STejun Heo continue; 27929c5a2ba7STejun Heo 27939c5a2ba7STejun Heo if (++flush_cnt == 10 || 27949c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2795c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27969c5a2ba7STejun Heo wq->name, flush_cnt); 279776af4d93STejun Heo 279876af4d93STejun Heo local_irq_enable(); 27999c5a2ba7STejun Heo goto reflush; 28009c5a2ba7STejun Heo } 28019c5a2ba7STejun Heo 280276af4d93STejun Heo spin_lock(&workqueue_lock); 28039c5a2ba7STejun Heo if (!--wq->nr_drainers) 2804618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 280576af4d93STejun Heo spin_unlock(&workqueue_lock); 280676af4d93STejun Heo 280776af4d93STejun Heo local_irq_enable(); 28089c5a2ba7STejun Heo } 28099c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28109c5a2ba7STejun Heo 2811606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2812baf59022STejun Heo { 2813baf59022STejun Heo struct worker *worker = NULL; 2814c9e7cf27STejun Heo struct worker_pool *pool; 2815112202d9STejun Heo struct pool_workqueue *pwq; 2816baf59022STejun Heo 2817baf59022STejun Heo might_sleep(); 2818baf59022STejun Heo 2819fa1b54e6STejun Heo local_irq_disable(); 2820fa1b54e6STejun Heo pool = get_work_pool(work); 2821fa1b54e6STejun Heo if (!pool) { 2822fa1b54e6STejun Heo local_irq_enable(); 2823fa1b54e6STejun Heo return false; 2824fa1b54e6STejun Heo } 2825fa1b54e6STejun Heo 2826fa1b54e6STejun Heo spin_lock(&pool->lock); 28270b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2828112202d9STejun Heo pwq = get_work_pwq(work); 2829112202d9STejun Heo if (pwq) { 2830112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2831baf59022STejun Heo goto already_gone; 2832606a5020STejun Heo } else { 2833c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2834baf59022STejun Heo if (!worker) 2835baf59022STejun Heo goto already_gone; 2836112202d9STejun Heo pwq = worker->current_pwq; 2837606a5020STejun Heo } 2838baf59022STejun Heo 2839112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2840d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2841baf59022STejun Heo 2842e159489bSTejun Heo /* 2843e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2844e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2845e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2846e159489bSTejun Heo * access. 2847e159489bSTejun Heo */ 2848493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2849112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2850e159489bSTejun Heo else 2851112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2852112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2853e159489bSTejun Heo 2854baf59022STejun Heo return true; 2855baf59022STejun Heo already_gone: 2856d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2857baf59022STejun Heo return false; 2858baf59022STejun Heo } 2859baf59022STejun Heo 2860db700897SOleg Nesterov /** 2861401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2862401a8d04STejun Heo * @work: the work to flush 2863db700897SOleg Nesterov * 2864606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2865606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2866401a8d04STejun Heo * 2867401a8d04STejun Heo * RETURNS: 2868401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2869401a8d04STejun Heo * %false if it was already idle. 2870db700897SOleg Nesterov */ 2871401a8d04STejun Heo bool flush_work(struct work_struct *work) 2872db700897SOleg Nesterov { 2873db700897SOleg Nesterov struct wq_barrier barr; 2874db700897SOleg Nesterov 28750976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28760976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28770976dfc1SStephen Boyd 2878606a5020STejun Heo if (start_flush_work(work, &barr)) { 2879db700897SOleg Nesterov wait_for_completion(&barr.done); 2880dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2881401a8d04STejun Heo return true; 2882606a5020STejun Heo } else { 2883401a8d04STejun Heo return false; 2884db700897SOleg Nesterov } 2885606a5020STejun Heo } 2886db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2887db700897SOleg Nesterov 288836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2889401a8d04STejun Heo { 2890bbb68dfaSTejun Heo unsigned long flags; 28911f1f642eSOleg Nesterov int ret; 28921f1f642eSOleg Nesterov 28931f1f642eSOleg Nesterov do { 2894bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2895bbb68dfaSTejun Heo /* 2896bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2897bbb68dfaSTejun Heo * would be waiting for before retrying. 2898bbb68dfaSTejun Heo */ 2899bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2900606a5020STejun Heo flush_work(work); 29011f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29021f1f642eSOleg Nesterov 2903bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2904bbb68dfaSTejun Heo mark_work_canceling(work); 2905bbb68dfaSTejun Heo local_irq_restore(flags); 2906bbb68dfaSTejun Heo 2907606a5020STejun Heo flush_work(work); 29087a22ad75STejun Heo clear_work_data(work); 29091f1f642eSOleg Nesterov return ret; 29101f1f642eSOleg Nesterov } 29111f1f642eSOleg Nesterov 29126e84d644SOleg Nesterov /** 2913401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2914401a8d04STejun Heo * @work: the work to cancel 29156e84d644SOleg Nesterov * 2916401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2917401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2918401a8d04STejun Heo * another workqueue. On return from this function, @work is 2919401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 29201f1f642eSOleg Nesterov * 2921401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2922401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 29236e84d644SOleg Nesterov * 2924401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 29256e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2926401a8d04STejun Heo * 2927401a8d04STejun Heo * RETURNS: 2928401a8d04STejun Heo * %true if @work was pending, %false otherwise. 29296e84d644SOleg Nesterov */ 2930401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 29316e84d644SOleg Nesterov { 293236e227d2STejun Heo return __cancel_work_timer(work, false); 2933b89deed3SOleg Nesterov } 293428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2935b89deed3SOleg Nesterov 29366e84d644SOleg Nesterov /** 2937401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2938401a8d04STejun Heo * @dwork: the delayed work to flush 29396e84d644SOleg Nesterov * 2940401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2941401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2942401a8d04STejun Heo * considers the last queueing instance of @dwork. 29431f1f642eSOleg Nesterov * 2944401a8d04STejun Heo * RETURNS: 2945401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2946401a8d04STejun Heo * %false if it was already idle. 29476e84d644SOleg Nesterov */ 2948401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2949401a8d04STejun Heo { 29508930cabaSTejun Heo local_irq_disable(); 2951401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 295260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29538930cabaSTejun Heo local_irq_enable(); 2954401a8d04STejun Heo return flush_work(&dwork->work); 2955401a8d04STejun Heo } 2956401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2957401a8d04STejun Heo 2958401a8d04STejun Heo /** 295957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 296057b30ae7STejun Heo * @dwork: delayed_work to cancel 296109383498STejun Heo * 296257b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 296357b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 296457b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 296557b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 296657b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 296709383498STejun Heo * 296857b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 296909383498STejun Heo */ 297057b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 297109383498STejun Heo { 297257b30ae7STejun Heo unsigned long flags; 297357b30ae7STejun Heo int ret; 297457b30ae7STejun Heo 297557b30ae7STejun Heo do { 297657b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 297757b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 297857b30ae7STejun Heo 297957b30ae7STejun Heo if (unlikely(ret < 0)) 298057b30ae7STejun Heo return false; 298157b30ae7STejun Heo 29827c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29837c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 298457b30ae7STejun Heo local_irq_restore(flags); 2985c0158ca6SDan Magenheimer return ret; 298609383498STejun Heo } 298757b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 298809383498STejun Heo 298909383498STejun Heo /** 2990401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2991401a8d04STejun Heo * @dwork: the delayed work cancel 2992401a8d04STejun Heo * 2993401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2994401a8d04STejun Heo * 2995401a8d04STejun Heo * RETURNS: 2996401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2997401a8d04STejun Heo */ 2998401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29996e84d644SOleg Nesterov { 300036e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 30016e84d644SOleg Nesterov } 3002f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 30031da177e4SLinus Torvalds 30040fcb78c2SRolf Eike Beer /** 3005c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 3006c1a220e7SZhang Rui * @cpu: cpu to put the work task on 3007c1a220e7SZhang Rui * @work: job to be done 3008c1a220e7SZhang Rui * 3009c1a220e7SZhang Rui * This puts a job on a specific cpu 3010c1a220e7SZhang Rui */ 3011d4283e93STejun Heo bool schedule_work_on(int cpu, struct work_struct *work) 3012c1a220e7SZhang Rui { 3013d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 3014c1a220e7SZhang Rui } 3015c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 3016c1a220e7SZhang Rui 30170fcb78c2SRolf Eike Beer /** 3018ae90dd5dSDave Jones * schedule_work - put work task in global workqueue 30191da177e4SLinus Torvalds * @work: job to be done 30201da177e4SLinus Torvalds * 3021d4283e93STejun Heo * Returns %false if @work was already on the kernel-global workqueue and 3022d4283e93STejun Heo * %true otherwise. 302352bad64dSDavid Howells * 30240fcb78c2SRolf Eike Beer * This puts a job in the kernel-global workqueue if it was not already 30250fcb78c2SRolf Eike Beer * queued and leaves it in the same position on the kernel-global 30260fcb78c2SRolf Eike Beer * workqueue otherwise. 30271da177e4SLinus Torvalds */ 3028d4283e93STejun Heo bool schedule_work(struct work_struct *work) 30291da177e4SLinus Torvalds { 30300fcb78c2SRolf Eike Beer return queue_work(system_wq, work); 30311da177e4SLinus Torvalds } 30320fcb78c2SRolf Eike Beer EXPORT_SYMBOL(schedule_work); 30331da177e4SLinus Torvalds 30340fcb78c2SRolf Eike Beer /** 30350fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 30360fcb78c2SRolf Eike Beer * @cpu: cpu to use 30370fcb78c2SRolf Eike Beer * @dwork: job to be done 30380fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 30390fcb78c2SRolf Eike Beer * 30400fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 30410fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 30420fcb78c2SRolf Eike Beer */ 3043d4283e93STejun Heo bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork, 3044d4283e93STejun Heo unsigned long delay) 30451da177e4SLinus Torvalds { 3046d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 30471da177e4SLinus Torvalds } 3048ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 30491da177e4SLinus Torvalds 3050b6136773SAndrew Morton /** 30510a13c00eSTejun Heo * schedule_delayed_work - put work task in global workqueue after delay 30520a13c00eSTejun Heo * @dwork: job to be done 30530a13c00eSTejun Heo * @delay: number of jiffies to wait or 0 for immediate execution 30540a13c00eSTejun Heo * 30550a13c00eSTejun Heo * After waiting for a given time this puts a job in the kernel-global 30560a13c00eSTejun Heo * workqueue. 30570a13c00eSTejun Heo */ 3058d4283e93STejun Heo bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay) 30590a13c00eSTejun Heo { 30600a13c00eSTejun Heo return queue_delayed_work(system_wq, dwork, delay); 30610a13c00eSTejun Heo } 30620a13c00eSTejun Heo EXPORT_SYMBOL(schedule_delayed_work); 30630a13c00eSTejun Heo 30640a13c00eSTejun Heo /** 306531ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3066b6136773SAndrew Morton * @func: the function to call 3067b6136773SAndrew Morton * 306831ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 306931ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3070b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 307131ddd871STejun Heo * 307231ddd871STejun Heo * RETURNS: 307331ddd871STejun Heo * 0 on success, -errno on failure. 3074b6136773SAndrew Morton */ 307565f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 307615316ba8SChristoph Lameter { 307715316ba8SChristoph Lameter int cpu; 307838f51568SNamhyung Kim struct work_struct __percpu *works; 307915316ba8SChristoph Lameter 3080b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3081b6136773SAndrew Morton if (!works) 308215316ba8SChristoph Lameter return -ENOMEM; 3083b6136773SAndrew Morton 308495402b38SGautham R Shenoy get_online_cpus(); 308593981800STejun Heo 308615316ba8SChristoph Lameter for_each_online_cpu(cpu) { 30879bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 30889bfb1839SIngo Molnar 30899bfb1839SIngo Molnar INIT_WORK(work, func); 30908de6d308SOleg Nesterov schedule_work_on(cpu, work); 309115316ba8SChristoph Lameter } 309293981800STejun Heo 309393981800STejun Heo for_each_online_cpu(cpu) 30948616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 309593981800STejun Heo 309695402b38SGautham R Shenoy put_online_cpus(); 3097b6136773SAndrew Morton free_percpu(works); 309815316ba8SChristoph Lameter return 0; 309915316ba8SChristoph Lameter } 310015316ba8SChristoph Lameter 3101eef6a7d5SAlan Stern /** 3102eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3103eef6a7d5SAlan Stern * 3104eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3105eef6a7d5SAlan Stern * completion. 3106eef6a7d5SAlan Stern * 3107eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3108eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3109eef6a7d5SAlan Stern * will lead to deadlock: 3110eef6a7d5SAlan Stern * 3111eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3112eef6a7d5SAlan Stern * a lock held by your code or its caller. 3113eef6a7d5SAlan Stern * 3114eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3115eef6a7d5SAlan Stern * 3116eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3117eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3118eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3119eef6a7d5SAlan Stern * 3120eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3121eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3122eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3123eef6a7d5SAlan Stern * cancel_work_sync() instead. 3124eef6a7d5SAlan Stern */ 31251da177e4SLinus Torvalds void flush_scheduled_work(void) 31261da177e4SLinus Torvalds { 3127d320c038STejun Heo flush_workqueue(system_wq); 31281da177e4SLinus Torvalds } 3129ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 31301da177e4SLinus Torvalds 31311da177e4SLinus Torvalds /** 31321fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31331fa44ecaSJames Bottomley * @fn: the function to execute 31341fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31351fa44ecaSJames Bottomley * be available when the work executes) 31361fa44ecaSJames Bottomley * 31371fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31381fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31391fa44ecaSJames Bottomley * 31401fa44ecaSJames Bottomley * Returns: 0 - function was executed 31411fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31421fa44ecaSJames Bottomley */ 314365f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31441fa44ecaSJames Bottomley { 31451fa44ecaSJames Bottomley if (!in_interrupt()) { 314665f27f38SDavid Howells fn(&ew->work); 31471fa44ecaSJames Bottomley return 0; 31481fa44ecaSJames Bottomley } 31491fa44ecaSJames Bottomley 315065f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31511fa44ecaSJames Bottomley schedule_work(&ew->work); 31521fa44ecaSJames Bottomley 31531fa44ecaSJames Bottomley return 1; 31541fa44ecaSJames Bottomley } 31551fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31561fa44ecaSJames Bottomley 31571da177e4SLinus Torvalds int keventd_up(void) 31581da177e4SLinus Torvalds { 3159d320c038STejun Heo return system_wq != NULL; 31601da177e4SLinus Torvalds } 31611da177e4SLinus Torvalds 3162226223abSTejun Heo #ifdef CONFIG_SYSFS 3163226223abSTejun Heo /* 3164226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3165226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3166226223abSTejun Heo * following attributes. 3167226223abSTejun Heo * 3168226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3169226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3170226223abSTejun Heo * 3171226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3172226223abSTejun Heo * 3173226223abSTejun Heo * id RO int : the associated pool ID 3174226223abSTejun Heo * nice RW int : nice value of the workers 3175226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3176226223abSTejun Heo */ 3177226223abSTejun Heo struct wq_device { 3178226223abSTejun Heo struct workqueue_struct *wq; 3179226223abSTejun Heo struct device dev; 3180226223abSTejun Heo }; 3181226223abSTejun Heo 3182226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3183226223abSTejun Heo { 3184226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3185226223abSTejun Heo 3186226223abSTejun Heo return wq_dev->wq; 3187226223abSTejun Heo } 3188226223abSTejun Heo 3189226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3190226223abSTejun Heo struct device_attribute *attr, char *buf) 3191226223abSTejun Heo { 3192226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3193226223abSTejun Heo 3194226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3195226223abSTejun Heo } 3196226223abSTejun Heo 3197226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3198226223abSTejun Heo struct device_attribute *attr, char *buf) 3199226223abSTejun Heo { 3200226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3201226223abSTejun Heo 3202226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3203226223abSTejun Heo } 3204226223abSTejun Heo 3205226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3206226223abSTejun Heo struct device_attribute *attr, 3207226223abSTejun Heo const char *buf, size_t count) 3208226223abSTejun Heo { 3209226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3210226223abSTejun Heo int val; 3211226223abSTejun Heo 3212226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3213226223abSTejun Heo return -EINVAL; 3214226223abSTejun Heo 3215226223abSTejun Heo workqueue_set_max_active(wq, val); 3216226223abSTejun Heo return count; 3217226223abSTejun Heo } 3218226223abSTejun Heo 3219226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3220226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3221226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3222226223abSTejun Heo __ATTR_NULL, 3223226223abSTejun Heo }; 3224226223abSTejun Heo 3225226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3226226223abSTejun Heo struct device_attribute *attr, char *buf) 3227226223abSTejun Heo { 3228226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3229226223abSTejun Heo struct worker_pool *pool; 3230226223abSTejun Heo int written; 3231226223abSTejun Heo 3232226223abSTejun Heo rcu_read_lock_sched(); 3233226223abSTejun Heo pool = first_pwq(wq)->pool; 3234226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3235226223abSTejun Heo rcu_read_unlock_sched(); 3236226223abSTejun Heo 3237226223abSTejun Heo return written; 3238226223abSTejun Heo } 3239226223abSTejun Heo 3240226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3241226223abSTejun Heo char *buf) 3242226223abSTejun Heo { 3243226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3244226223abSTejun Heo int written; 3245226223abSTejun Heo 3246226223abSTejun Heo rcu_read_lock_sched(); 3247226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3248226223abSTejun Heo first_pwq(wq)->pool->attrs->nice); 3249226223abSTejun Heo rcu_read_unlock_sched(); 3250226223abSTejun Heo 3251226223abSTejun Heo return written; 3252226223abSTejun Heo } 3253226223abSTejun Heo 3254226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3255226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3256226223abSTejun Heo { 3257226223abSTejun Heo struct workqueue_attrs *attrs; 3258226223abSTejun Heo 3259226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3260226223abSTejun Heo if (!attrs) 3261226223abSTejun Heo return NULL; 3262226223abSTejun Heo 3263226223abSTejun Heo rcu_read_lock_sched(); 3264226223abSTejun Heo copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs); 3265226223abSTejun Heo rcu_read_unlock_sched(); 3266226223abSTejun Heo return attrs; 3267226223abSTejun Heo } 3268226223abSTejun Heo 3269226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3270226223abSTejun Heo const char *buf, size_t count) 3271226223abSTejun Heo { 3272226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3273226223abSTejun Heo struct workqueue_attrs *attrs; 3274226223abSTejun Heo int ret; 3275226223abSTejun Heo 3276226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3277226223abSTejun Heo if (!attrs) 3278226223abSTejun Heo return -ENOMEM; 3279226223abSTejun Heo 3280226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3281226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3282226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3283226223abSTejun Heo else 3284226223abSTejun Heo ret = -EINVAL; 3285226223abSTejun Heo 3286226223abSTejun Heo free_workqueue_attrs(attrs); 3287226223abSTejun Heo return ret ?: count; 3288226223abSTejun Heo } 3289226223abSTejun Heo 3290226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3291226223abSTejun Heo struct device_attribute *attr, char *buf) 3292226223abSTejun Heo { 3293226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3294226223abSTejun Heo int written; 3295226223abSTejun Heo 3296226223abSTejun Heo rcu_read_lock_sched(); 3297226223abSTejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, 3298226223abSTejun Heo first_pwq(wq)->pool->attrs->cpumask); 3299226223abSTejun Heo rcu_read_unlock_sched(); 3300226223abSTejun Heo 3301226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3302226223abSTejun Heo return written; 3303226223abSTejun Heo } 3304226223abSTejun Heo 3305226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3306226223abSTejun Heo struct device_attribute *attr, 3307226223abSTejun Heo const char *buf, size_t count) 3308226223abSTejun Heo { 3309226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3310226223abSTejun Heo struct workqueue_attrs *attrs; 3311226223abSTejun Heo int ret; 3312226223abSTejun Heo 3313226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3314226223abSTejun Heo if (!attrs) 3315226223abSTejun Heo return -ENOMEM; 3316226223abSTejun Heo 3317226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3318226223abSTejun Heo if (!ret) 3319226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3320226223abSTejun Heo 3321226223abSTejun Heo free_workqueue_attrs(attrs); 3322226223abSTejun Heo return ret ?: count; 3323226223abSTejun Heo } 3324226223abSTejun Heo 3325226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3326226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3327226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3328226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3329226223abSTejun Heo __ATTR_NULL, 3330226223abSTejun Heo }; 3331226223abSTejun Heo 3332226223abSTejun Heo static struct bus_type wq_subsys = { 3333226223abSTejun Heo .name = "workqueue", 3334226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3335226223abSTejun Heo }; 3336226223abSTejun Heo 3337226223abSTejun Heo static int __init wq_sysfs_init(void) 3338226223abSTejun Heo { 3339226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3340226223abSTejun Heo } 3341226223abSTejun Heo core_initcall(wq_sysfs_init); 3342226223abSTejun Heo 3343226223abSTejun Heo static void wq_device_release(struct device *dev) 3344226223abSTejun Heo { 3345226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3346226223abSTejun Heo 3347226223abSTejun Heo kfree(wq_dev); 3348226223abSTejun Heo } 3349226223abSTejun Heo 3350226223abSTejun Heo /** 3351226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3352226223abSTejun Heo * @wq: the workqueue to register 3353226223abSTejun Heo * 3354226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3355226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3356226223abSTejun Heo * which is the preferred method. 3357226223abSTejun Heo * 3358226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3359226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3360226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3361226223abSTejun Heo * attributes. 3362226223abSTejun Heo * 3363226223abSTejun Heo * Returns 0 on success, -errno on failure. 3364226223abSTejun Heo */ 3365226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3366226223abSTejun Heo { 3367226223abSTejun Heo struct wq_device *wq_dev; 3368226223abSTejun Heo int ret; 3369226223abSTejun Heo 3370226223abSTejun Heo /* 3371226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3372226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3373226223abSTejun Heo * workqueues. 3374226223abSTejun Heo */ 3375226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3376226223abSTejun Heo return -EINVAL; 3377226223abSTejun Heo 3378226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3379226223abSTejun Heo if (!wq_dev) 3380226223abSTejun Heo return -ENOMEM; 3381226223abSTejun Heo 3382226223abSTejun Heo wq_dev->wq = wq; 3383226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3384226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3385226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3386226223abSTejun Heo 3387226223abSTejun Heo /* 3388226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3389226223abSTejun Heo * everything is ready. 3390226223abSTejun Heo */ 3391226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3392226223abSTejun Heo 3393226223abSTejun Heo ret = device_register(&wq_dev->dev); 3394226223abSTejun Heo if (ret) { 3395226223abSTejun Heo kfree(wq_dev); 3396226223abSTejun Heo wq->wq_dev = NULL; 3397226223abSTejun Heo return ret; 3398226223abSTejun Heo } 3399226223abSTejun Heo 3400226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3401226223abSTejun Heo struct device_attribute *attr; 3402226223abSTejun Heo 3403226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3404226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3405226223abSTejun Heo if (ret) { 3406226223abSTejun Heo device_unregister(&wq_dev->dev); 3407226223abSTejun Heo wq->wq_dev = NULL; 3408226223abSTejun Heo return ret; 3409226223abSTejun Heo } 3410226223abSTejun Heo } 3411226223abSTejun Heo } 3412226223abSTejun Heo 3413226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3414226223abSTejun Heo return 0; 3415226223abSTejun Heo } 3416226223abSTejun Heo 3417226223abSTejun Heo /** 3418226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3419226223abSTejun Heo * @wq: the workqueue to unregister 3420226223abSTejun Heo * 3421226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3422226223abSTejun Heo */ 3423226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3424226223abSTejun Heo { 3425226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3426226223abSTejun Heo 3427226223abSTejun Heo if (!wq->wq_dev) 3428226223abSTejun Heo return; 3429226223abSTejun Heo 3430226223abSTejun Heo wq->wq_dev = NULL; 3431226223abSTejun Heo device_unregister(&wq_dev->dev); 3432226223abSTejun Heo } 3433226223abSTejun Heo #else /* CONFIG_SYSFS */ 3434226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3435226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3436226223abSTejun Heo 34377a4e344cSTejun Heo /** 34387a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 34397a4e344cSTejun Heo * @attrs: workqueue_attrs to free 34407a4e344cSTejun Heo * 34417a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 34427a4e344cSTejun Heo */ 34437a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 34447a4e344cSTejun Heo { 34457a4e344cSTejun Heo if (attrs) { 34467a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 34477a4e344cSTejun Heo kfree(attrs); 34487a4e344cSTejun Heo } 34497a4e344cSTejun Heo } 34507a4e344cSTejun Heo 34517a4e344cSTejun Heo /** 34527a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 34537a4e344cSTejun Heo * @gfp_mask: allocation mask to use 34547a4e344cSTejun Heo * 34557a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 34567a4e344cSTejun Heo * return it. Returns NULL on failure. 34577a4e344cSTejun Heo */ 34587a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 34597a4e344cSTejun Heo { 34607a4e344cSTejun Heo struct workqueue_attrs *attrs; 34617a4e344cSTejun Heo 34627a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 34637a4e344cSTejun Heo if (!attrs) 34647a4e344cSTejun Heo goto fail; 34657a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 34667a4e344cSTejun Heo goto fail; 34677a4e344cSTejun Heo 34687a4e344cSTejun Heo cpumask_setall(attrs->cpumask); 34697a4e344cSTejun Heo return attrs; 34707a4e344cSTejun Heo fail: 34717a4e344cSTejun Heo free_workqueue_attrs(attrs); 34727a4e344cSTejun Heo return NULL; 34737a4e344cSTejun Heo } 34747a4e344cSTejun Heo 347529c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 347629c91e99STejun Heo const struct workqueue_attrs *from) 347729c91e99STejun Heo { 347829c91e99STejun Heo to->nice = from->nice; 347929c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 348029c91e99STejun Heo } 348129c91e99STejun Heo 348229c91e99STejun Heo /* 348329c91e99STejun Heo * Hacky implementation of jhash of bitmaps which only considers the 348429c91e99STejun Heo * specified number of bits. We probably want a proper implementation in 348529c91e99STejun Heo * include/linux/jhash.h. 348629c91e99STejun Heo */ 348729c91e99STejun Heo static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash) 348829c91e99STejun Heo { 348929c91e99STejun Heo int nr_longs = bits / BITS_PER_LONG; 349029c91e99STejun Heo int nr_leftover = bits % BITS_PER_LONG; 349129c91e99STejun Heo unsigned long leftover = 0; 349229c91e99STejun Heo 349329c91e99STejun Heo if (nr_longs) 349429c91e99STejun Heo hash = jhash(bitmap, nr_longs * sizeof(long), hash); 349529c91e99STejun Heo if (nr_leftover) { 349629c91e99STejun Heo bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover); 349729c91e99STejun Heo hash = jhash(&leftover, sizeof(long), hash); 349829c91e99STejun Heo } 349929c91e99STejun Heo return hash; 350029c91e99STejun Heo } 350129c91e99STejun Heo 350229c91e99STejun Heo /* hash value of the content of @attr */ 350329c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 350429c91e99STejun Heo { 350529c91e99STejun Heo u32 hash = 0; 350629c91e99STejun Heo 350729c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 350829c91e99STejun Heo hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash); 350929c91e99STejun Heo return hash; 351029c91e99STejun Heo } 351129c91e99STejun Heo 351229c91e99STejun Heo /* content equality test */ 351329c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 351429c91e99STejun Heo const struct workqueue_attrs *b) 351529c91e99STejun Heo { 351629c91e99STejun Heo if (a->nice != b->nice) 351729c91e99STejun Heo return false; 351829c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 351929c91e99STejun Heo return false; 352029c91e99STejun Heo return true; 352129c91e99STejun Heo } 352229c91e99STejun Heo 35237a4e344cSTejun Heo /** 35247a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 35257a4e344cSTejun Heo * @pool: worker_pool to initialize 35267a4e344cSTejun Heo * 35277a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 352829c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 352929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 353029c91e99STejun Heo * on @pool safely to release it. 35317a4e344cSTejun Heo */ 35327a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 35334e1a1f9aSTejun Heo { 35344e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 353529c91e99STejun Heo pool->id = -1; 353629c91e99STejun Heo pool->cpu = -1; 35374e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 35384e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 35394e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 35404e1a1f9aSTejun Heo hash_init(pool->busy_hash); 35414e1a1f9aSTejun Heo 35424e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 35434e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 35444e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 35454e1a1f9aSTejun Heo 35464e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 35474e1a1f9aSTejun Heo (unsigned long)pool); 35484e1a1f9aSTejun Heo 35494e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 35504e1a1f9aSTejun Heo mutex_init(&pool->assoc_mutex); 35514e1a1f9aSTejun Heo ida_init(&pool->worker_ida); 35527a4e344cSTejun Heo 355329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 355429c91e99STejun Heo pool->refcnt = 1; 355529c91e99STejun Heo 355629c91e99STejun Heo /* shouldn't fail above this point */ 35577a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 35587a4e344cSTejun Heo if (!pool->attrs) 35597a4e344cSTejun Heo return -ENOMEM; 35607a4e344cSTejun Heo return 0; 35614e1a1f9aSTejun Heo } 35624e1a1f9aSTejun Heo 356329c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 356429c91e99STejun Heo { 356529c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 356629c91e99STejun Heo 356729c91e99STejun Heo ida_destroy(&pool->worker_ida); 356829c91e99STejun Heo free_workqueue_attrs(pool->attrs); 356929c91e99STejun Heo kfree(pool); 357029c91e99STejun Heo } 357129c91e99STejun Heo 357229c91e99STejun Heo /** 357329c91e99STejun Heo * put_unbound_pool - put a worker_pool 357429c91e99STejun Heo * @pool: worker_pool to put 357529c91e99STejun Heo * 357629c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3577c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3578c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3579c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 358029c91e99STejun Heo */ 358129c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 358229c91e99STejun Heo { 358329c91e99STejun Heo struct worker *worker; 358429c91e99STejun Heo 358529c91e99STejun Heo spin_lock_irq(&workqueue_lock); 358629c91e99STejun Heo if (--pool->refcnt) { 358729c91e99STejun Heo spin_unlock_irq(&workqueue_lock); 358829c91e99STejun Heo return; 358929c91e99STejun Heo } 359029c91e99STejun Heo 359129c91e99STejun Heo /* sanity checks */ 359229c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 359329c91e99STejun Heo WARN_ON(!list_empty(&pool->worklist))) { 359429c91e99STejun Heo spin_unlock_irq(&workqueue_lock); 359529c91e99STejun Heo return; 359629c91e99STejun Heo } 359729c91e99STejun Heo 359829c91e99STejun Heo /* release id and unhash */ 359929c91e99STejun Heo if (pool->id >= 0) 360029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 360129c91e99STejun Heo hash_del(&pool->hash_node); 360229c91e99STejun Heo 360329c91e99STejun Heo spin_unlock_irq(&workqueue_lock); 360429c91e99STejun Heo 3605c5aa87bbSTejun Heo /* 3606c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3607c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3608c5aa87bbSTejun Heo * manager_mutex. 3609c5aa87bbSTejun Heo */ 361029c91e99STejun Heo mutex_lock(&pool->manager_arb); 361129c91e99STejun Heo spin_lock_irq(&pool->lock); 361229c91e99STejun Heo 361329c91e99STejun Heo while ((worker = first_worker(pool))) 361429c91e99STejun Heo destroy_worker(worker); 361529c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 361629c91e99STejun Heo 361729c91e99STejun Heo spin_unlock_irq(&pool->lock); 361829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 361929c91e99STejun Heo 362029c91e99STejun Heo /* shut down the timers */ 362129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 362229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 362329c91e99STejun Heo 362429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 362529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 362629c91e99STejun Heo } 362729c91e99STejun Heo 362829c91e99STejun Heo /** 362929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 363029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 363129c91e99STejun Heo * 363229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 363329c91e99STejun Heo * reference count and return it. If there already is a matching 363429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 363529c91e99STejun Heo * create a new one. On failure, returns NULL. 363629c91e99STejun Heo */ 363729c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 363829c91e99STejun Heo { 363929c91e99STejun Heo static DEFINE_MUTEX(create_mutex); 364029c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 364129c91e99STejun Heo struct worker_pool *pool; 364229c91e99STejun Heo struct worker *worker; 364329c91e99STejun Heo 364429c91e99STejun Heo mutex_lock(&create_mutex); 364529c91e99STejun Heo 364629c91e99STejun Heo /* do we already have a matching pool? */ 364729c91e99STejun Heo spin_lock_irq(&workqueue_lock); 364829c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 364929c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 365029c91e99STejun Heo pool->refcnt++; 365129c91e99STejun Heo goto out_unlock; 365229c91e99STejun Heo } 365329c91e99STejun Heo } 365429c91e99STejun Heo spin_unlock_irq(&workqueue_lock); 365529c91e99STejun Heo 365629c91e99STejun Heo /* nope, create a new one */ 365729c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 365829c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 365929c91e99STejun Heo goto fail; 366029c91e99STejun Heo 36618864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 366229c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 366329c91e99STejun Heo 366429c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 366529c91e99STejun Heo goto fail; 366629c91e99STejun Heo 366729c91e99STejun Heo /* create and start the initial worker */ 366829c91e99STejun Heo worker = create_worker(pool); 366929c91e99STejun Heo if (!worker) 367029c91e99STejun Heo goto fail; 367129c91e99STejun Heo 367229c91e99STejun Heo spin_lock_irq(&pool->lock); 367329c91e99STejun Heo start_worker(worker); 367429c91e99STejun Heo spin_unlock_irq(&pool->lock); 367529c91e99STejun Heo 367629c91e99STejun Heo /* install */ 367729c91e99STejun Heo spin_lock_irq(&workqueue_lock); 367829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 367929c91e99STejun Heo out_unlock: 368029c91e99STejun Heo spin_unlock_irq(&workqueue_lock); 368129c91e99STejun Heo mutex_unlock(&create_mutex); 368229c91e99STejun Heo return pool; 368329c91e99STejun Heo fail: 368429c91e99STejun Heo mutex_unlock(&create_mutex); 368529c91e99STejun Heo if (pool) 368629c91e99STejun Heo put_unbound_pool(pool); 368729c91e99STejun Heo return NULL; 368829c91e99STejun Heo } 368929c91e99STejun Heo 36908864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36918864b4e5STejun Heo { 36928864b4e5STejun Heo kmem_cache_free(pwq_cache, 36938864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36948864b4e5STejun Heo } 36958864b4e5STejun Heo 36968864b4e5STejun Heo /* 36978864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36988864b4e5STejun Heo * and needs to be destroyed. 36998864b4e5STejun Heo */ 37008864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 37018864b4e5STejun Heo { 37028864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 37038864b4e5STejun Heo unbound_release_work); 37048864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 37058864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 37068864b4e5STejun Heo 37078864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 37088864b4e5STejun Heo return; 37098864b4e5STejun Heo 371075ccf595STejun Heo /* 371175ccf595STejun Heo * Unlink @pwq. Synchronization against flush_mutex isn't strictly 371275ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 371375ccf595STejun Heo * and consistent with the linking path. 371475ccf595STejun Heo */ 371575ccf595STejun Heo mutex_lock(&wq->flush_mutex); 37168864b4e5STejun Heo spin_lock_irq(&workqueue_lock); 37178864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 37188864b4e5STejun Heo spin_unlock_irq(&workqueue_lock); 371975ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 37208864b4e5STejun Heo 37218864b4e5STejun Heo put_unbound_pool(pool); 37228864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 37238864b4e5STejun Heo 37248864b4e5STejun Heo /* 37258864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 37268864b4e5STejun Heo * is gonna access it anymore. Free it. 37278864b4e5STejun Heo */ 37288864b4e5STejun Heo if (list_empty(&wq->pwqs)) 37298864b4e5STejun Heo kfree(wq); 37308864b4e5STejun Heo } 37318864b4e5STejun Heo 37320fbd95aaSTejun Heo /** 3733699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37340fbd95aaSTejun Heo * @pwq: target pool_workqueue 37350fbd95aaSTejun Heo * 3736699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3737699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3738699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37390fbd95aaSTejun Heo */ 3740699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37410fbd95aaSTejun Heo { 3742699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3743699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3744699ce097STejun Heo 3745699ce097STejun Heo /* for @wq->saved_max_active */ 3746699ce097STejun Heo lockdep_assert_held(&workqueue_lock); 3747699ce097STejun Heo 3748699ce097STejun Heo /* fast exit for non-freezable wqs */ 3749699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3750699ce097STejun Heo return; 3751699ce097STejun Heo 3752699ce097STejun Heo spin_lock(&pwq->pool->lock); 3753699ce097STejun Heo 3754699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3755699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37560fbd95aaSTejun Heo 37570fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 37580fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 37590fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3760699ce097STejun Heo } else { 3761699ce097STejun Heo pwq->max_active = 0; 3762699ce097STejun Heo } 3763699ce097STejun Heo 3764699ce097STejun Heo spin_unlock(&pwq->pool->lock); 37650fbd95aaSTejun Heo } 37660fbd95aaSTejun Heo 3767d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3768d2c1d404STejun Heo struct workqueue_struct *wq, 37699e8cd2f5STejun Heo struct worker_pool *pool, 37709e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3771d2c1d404STejun Heo { 3772d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3773d2c1d404STejun Heo 3774d2c1d404STejun Heo pwq->pool = pool; 3775d2c1d404STejun Heo pwq->wq = wq; 3776d2c1d404STejun Heo pwq->flush_color = -1; 37778864b4e5STejun Heo pwq->refcnt = 1; 3778d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3779d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37808864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3781d2c1d404STejun Heo 378275ccf595STejun Heo mutex_lock(&wq->flush_mutex); 378375ccf595STejun Heo spin_lock_irq(&workqueue_lock); 378475ccf595STejun Heo 3785983ca25eSTejun Heo /* 3786983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 3787983ca25eSTejun Heo * flush_mutex to avoid confusing flush_workqueue(). 3788983ca25eSTejun Heo */ 37899e8cd2f5STejun Heo if (p_last_pwq) 37909e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 379175ccf595STejun Heo pwq->work_color = wq->work_color; 3792983ca25eSTejun Heo 3793983ca25eSTejun Heo /* sync max_active to the current setting */ 3794983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3795983ca25eSTejun Heo 3796983ca25eSTejun Heo /* link in @pwq */ 37979e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 379875ccf595STejun Heo 379975ccf595STejun Heo spin_unlock_irq(&workqueue_lock); 380075ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 3801d2c1d404STejun Heo } 3802d2c1d404STejun Heo 38039e8cd2f5STejun Heo /** 38049e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38059e8cd2f5STejun Heo * @wq: the target workqueue 38069e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38079e8cd2f5STejun Heo * 38089e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 38099e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 38109e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 38119e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 38129e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 38139e8cd2f5STejun Heo * 38149e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 38159e8cd2f5STejun Heo * failure. 38169e8cd2f5STejun Heo */ 38179e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38189e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38199e8cd2f5STejun Heo { 38209e8cd2f5STejun Heo struct pool_workqueue *pwq, *last_pwq; 38219e8cd2f5STejun Heo struct worker_pool *pool; 38229e8cd2f5STejun Heo 38238719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38249e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38259e8cd2f5STejun Heo return -EINVAL; 38269e8cd2f5STejun Heo 38278719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38288719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38298719dceaSTejun Heo return -EINVAL; 38308719dceaSTejun Heo 38319e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 38329e8cd2f5STejun Heo if (!pwq) 38339e8cd2f5STejun Heo return -ENOMEM; 38349e8cd2f5STejun Heo 38359e8cd2f5STejun Heo pool = get_unbound_pool(attrs); 38369e8cd2f5STejun Heo if (!pool) { 38379e8cd2f5STejun Heo kmem_cache_free(pwq_cache, pwq); 38389e8cd2f5STejun Heo return -ENOMEM; 38399e8cd2f5STejun Heo } 38409e8cd2f5STejun Heo 38419e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 38429e8cd2f5STejun Heo if (last_pwq) { 38439e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 38449e8cd2f5STejun Heo put_pwq(last_pwq); 38459e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 38469e8cd2f5STejun Heo } 38479e8cd2f5STejun Heo 38489e8cd2f5STejun Heo return 0; 38499e8cd2f5STejun Heo } 38509e8cd2f5STejun Heo 385130cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 38521da177e4SLinus Torvalds { 385349e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 385430cdf249STejun Heo int cpu; 3855e1d8aa9fSFrederic Weisbecker 385630cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3857420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3858420c0ddbSTejun Heo if (!wq->cpu_pwqs) 385930cdf249STejun Heo return -ENOMEM; 386030cdf249STejun Heo 386130cdf249STejun Heo for_each_possible_cpu(cpu) { 38627fb98ea7STejun Heo struct pool_workqueue *pwq = 38637fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 38647a62c2c8STejun Heo struct worker_pool *cpu_pools = 3865f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 386630cdf249STejun Heo 38679e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 386830cdf249STejun Heo } 386930cdf249STejun Heo return 0; 38709e8cd2f5STejun Heo } else { 38719e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38729e8cd2f5STejun Heo } 38730f900049STejun Heo } 38740f900049STejun Heo 3875f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3876f3421797STejun Heo const char *name) 3877b71ab8c2STejun Heo { 3878f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3879f3421797STejun Heo 3880f3421797STejun Heo if (max_active < 1 || max_active > lim) 3881044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3882f3421797STejun Heo max_active, name, 1, lim); 3883b71ab8c2STejun Heo 3884f3421797STejun Heo return clamp_val(max_active, 1, lim); 3885b71ab8c2STejun Heo } 3886b71ab8c2STejun Heo 3887b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 388897e37d7bSTejun Heo unsigned int flags, 38891e19ffc6STejun Heo int max_active, 3890eb13ba87SJohannes Berg struct lock_class_key *key, 3891b196be89STejun Heo const char *lock_name, ...) 38923af24433SOleg Nesterov { 3893b196be89STejun Heo va_list args, args1; 38943af24433SOleg Nesterov struct workqueue_struct *wq; 389549e3cf44STejun Heo struct pool_workqueue *pwq; 3896b196be89STejun Heo size_t namelen; 3897b196be89STejun Heo 3898b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3899b196be89STejun Heo va_start(args, lock_name); 3900b196be89STejun Heo va_copy(args1, args); 3901b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3902b196be89STejun Heo 3903b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3904b196be89STejun Heo if (!wq) 3905d2c1d404STejun Heo return NULL; 3906b196be89STejun Heo 3907b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3908b196be89STejun Heo va_end(args); 3909b196be89STejun Heo va_end(args1); 39103af24433SOleg Nesterov 3911d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3912b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 39133af24433SOleg Nesterov 3914b196be89STejun Heo /* init wq */ 391597e37d7bSTejun Heo wq->flags = flags; 3916a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 391773f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3918112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 391930cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 392073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 392173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3922493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 39233af24433SOleg Nesterov 3924eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3925cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 39263af24433SOleg Nesterov 392730cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3928d2c1d404STejun Heo goto err_free_wq; 39291537663fSTejun Heo 3930493008a8STejun Heo /* 3931493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3932493008a8STejun Heo * have a rescuer to guarantee forward progress. 3933493008a8STejun Heo */ 3934493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3935e22bee78STejun Heo struct worker *rescuer; 3936e22bee78STejun Heo 3937d2c1d404STejun Heo rescuer = alloc_worker(); 3938e22bee78STejun Heo if (!rescuer) 3939d2c1d404STejun Heo goto err_destroy; 3940e22bee78STejun Heo 3941111c225aSTejun Heo rescuer->rescue_wq = wq; 3942111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3943b196be89STejun Heo wq->name); 3944d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3945d2c1d404STejun Heo kfree(rescuer); 3946d2c1d404STejun Heo goto err_destroy; 3947d2c1d404STejun Heo } 3948e22bee78STejun Heo 3949d2c1d404STejun Heo wq->rescuer = rescuer; 3950e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3951e22bee78STejun Heo wake_up_process(rescuer->task); 39523af24433SOleg Nesterov } 39531537663fSTejun Heo 3954226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3955226223abSTejun Heo goto err_destroy; 3956226223abSTejun Heo 39573af24433SOleg Nesterov /* 3958699ce097STejun Heo * workqueue_lock protects global freeze state and workqueues list. 3959699ce097STejun Heo * Grab it, adjust max_active and add the new workqueue to 3960699ce097STejun Heo * workqueues list. 39613af24433SOleg Nesterov */ 3962e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 3963a0a1a5fdSTejun Heo 396449e3cf44STejun Heo for_each_pwq(pwq, wq) 3965699ce097STejun Heo pwq_adjust_max_active(pwq); 3966a0a1a5fdSTejun Heo 39673af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3968a0a1a5fdSTejun Heo 3969e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 39703af24433SOleg Nesterov 39713af24433SOleg Nesterov return wq; 3972d2c1d404STejun Heo 3973d2c1d404STejun Heo err_free_wq: 39744690c4abSTejun Heo kfree(wq); 3975d2c1d404STejun Heo return NULL; 3976d2c1d404STejun Heo err_destroy: 3977d2c1d404STejun Heo destroy_workqueue(wq); 39784690c4abSTejun Heo return NULL; 39791da177e4SLinus Torvalds } 3980d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 39811da177e4SLinus Torvalds 39823af24433SOleg Nesterov /** 39833af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 39843af24433SOleg Nesterov * @wq: target workqueue 39853af24433SOleg Nesterov * 39863af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 39873af24433SOleg Nesterov */ 39883af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 39893af24433SOleg Nesterov { 399049e3cf44STejun Heo struct pool_workqueue *pwq; 39913af24433SOleg Nesterov 39929c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 39939c5a2ba7STejun Heo drain_workqueue(wq); 3994c8efcc25STejun Heo 399576af4d93STejun Heo spin_lock_irq(&workqueue_lock); 399676af4d93STejun Heo 39976183c009STejun Heo /* sanity checks */ 399849e3cf44STejun Heo for_each_pwq(pwq, wq) { 39996183c009STejun Heo int i; 40006183c009STejun Heo 400176af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 400276af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 400376af4d93STejun Heo spin_unlock_irq(&workqueue_lock); 40046183c009STejun Heo return; 400576af4d93STejun Heo } 400676af4d93STejun Heo } 400776af4d93STejun Heo 40088864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 40098864b4e5STejun Heo WARN_ON(pwq->nr_active) || 401076af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 401176af4d93STejun Heo spin_unlock_irq(&workqueue_lock); 40126183c009STejun Heo return; 40136183c009STejun Heo } 401476af4d93STejun Heo } 40156183c009STejun Heo 4016a0a1a5fdSTejun Heo /* 4017a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4018a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4019a0a1a5fdSTejun Heo */ 4020d2c1d404STejun Heo list_del_init(&wq->list); 402176af4d93STejun Heo 4022e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 40233af24433SOleg Nesterov 4024226223abSTejun Heo workqueue_sysfs_unregister(wq); 4025226223abSTejun Heo 4026493008a8STejun Heo if (wq->rescuer) { 4027e22bee78STejun Heo kthread_stop(wq->rescuer->task); 40288d9df9f0SXiaotian Feng kfree(wq->rescuer); 4029493008a8STejun Heo wq->rescuer = NULL; 4030e22bee78STejun Heo } 4031e22bee78STejun Heo 40328864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 403329c91e99STejun Heo /* 40348864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 40358864b4e5STejun Heo * free the pwqs and wq. 403629c91e99STejun Heo */ 40378864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 40388864b4e5STejun Heo kfree(wq); 40398864b4e5STejun Heo } else { 40408864b4e5STejun Heo /* 40418864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 40428864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 40438864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 40448864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 40458864b4e5STejun Heo * released. 40468864b4e5STejun Heo */ 404729c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 404829c91e99STejun Heo pwqs_node); 40498864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 40508864b4e5STejun Heo put_pwq(pwq); 40518864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 405229c91e99STejun Heo } 40533af24433SOleg Nesterov } 40543af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 40553af24433SOleg Nesterov 4056dcd989cbSTejun Heo /** 4057dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4058dcd989cbSTejun Heo * @wq: target workqueue 4059dcd989cbSTejun Heo * @max_active: new max_active value. 4060dcd989cbSTejun Heo * 4061dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4062dcd989cbSTejun Heo * 4063dcd989cbSTejun Heo * CONTEXT: 4064dcd989cbSTejun Heo * Don't call from IRQ context. 4065dcd989cbSTejun Heo */ 4066dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4067dcd989cbSTejun Heo { 406849e3cf44STejun Heo struct pool_workqueue *pwq; 4069dcd989cbSTejun Heo 40708719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 40718719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 40728719dceaSTejun Heo return; 40738719dceaSTejun Heo 4074f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4075dcd989cbSTejun Heo 4076e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 4077dcd989cbSTejun Heo 4078dcd989cbSTejun Heo wq->saved_max_active = max_active; 4079dcd989cbSTejun Heo 4080699ce097STejun Heo for_each_pwq(pwq, wq) 4081699ce097STejun Heo pwq_adjust_max_active(pwq); 4082dcd989cbSTejun Heo 4083e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 4084dcd989cbSTejun Heo } 4085dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4086dcd989cbSTejun Heo 4087dcd989cbSTejun Heo /** 4088e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4089e6267616STejun Heo * 4090e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4091e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4092e6267616STejun Heo */ 4093e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4094e6267616STejun Heo { 4095e6267616STejun Heo struct worker *worker = current_wq_worker(); 4096e6267616STejun Heo 4097e6267616STejun Heo return worker && worker == worker->current_pwq->wq->rescuer; 4098e6267616STejun Heo } 4099e6267616STejun Heo 4100e6267616STejun Heo /** 4101dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4102dcd989cbSTejun Heo * @cpu: CPU in question 4103dcd989cbSTejun Heo * @wq: target workqueue 4104dcd989cbSTejun Heo * 4105dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4106dcd989cbSTejun Heo * no synchronization around this function and the test result is 4107dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4108dcd989cbSTejun Heo * 4109dcd989cbSTejun Heo * RETURNS: 4110dcd989cbSTejun Heo * %true if congested, %false otherwise. 4111dcd989cbSTejun Heo */ 4112d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4113dcd989cbSTejun Heo { 41147fb98ea7STejun Heo struct pool_workqueue *pwq; 411576af4d93STejun Heo bool ret; 411676af4d93STejun Heo 411776af4d93STejun Heo preempt_disable(); 41187fb98ea7STejun Heo 41197fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 41207fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 41217fb98ea7STejun Heo else 41227fb98ea7STejun Heo pwq = first_pwq(wq); 4123dcd989cbSTejun Heo 412476af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 412576af4d93STejun Heo preempt_enable(); 412676af4d93STejun Heo 412776af4d93STejun Heo return ret; 4128dcd989cbSTejun Heo } 4129dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4130dcd989cbSTejun Heo 4131dcd989cbSTejun Heo /** 4132dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4133dcd989cbSTejun Heo * @work: the work to be tested 4134dcd989cbSTejun Heo * 4135dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4136dcd989cbSTejun Heo * synchronization around this function and the test result is 4137dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4138dcd989cbSTejun Heo * 4139dcd989cbSTejun Heo * RETURNS: 4140dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4141dcd989cbSTejun Heo */ 4142dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4143dcd989cbSTejun Heo { 4144fa1b54e6STejun Heo struct worker_pool *pool; 4145dcd989cbSTejun Heo unsigned long flags; 4146dcd989cbSTejun Heo unsigned int ret = 0; 4147dcd989cbSTejun Heo 4148dcd989cbSTejun Heo if (work_pending(work)) 4149dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4150038366c5SLai Jiangshan 4151fa1b54e6STejun Heo local_irq_save(flags); 4152fa1b54e6STejun Heo pool = get_work_pool(work); 4153038366c5SLai Jiangshan if (pool) { 4154fa1b54e6STejun Heo spin_lock(&pool->lock); 4155c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4156dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4157fa1b54e6STejun Heo spin_unlock(&pool->lock); 4158038366c5SLai Jiangshan } 4159fa1b54e6STejun Heo local_irq_restore(flags); 4160dcd989cbSTejun Heo 4161dcd989cbSTejun Heo return ret; 4162dcd989cbSTejun Heo } 4163dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4164dcd989cbSTejun Heo 4165db7bccf4STejun Heo /* 4166db7bccf4STejun Heo * CPU hotplug. 4167db7bccf4STejun Heo * 4168e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4169112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4170706026c2STejun Heo * pool which make migrating pending and scheduled works very 4171e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 417294cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4173e22bee78STejun Heo * blocked draining impractical. 4174e22bee78STejun Heo * 417524647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4176628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4177628c78e7STejun Heo * cpu comes back online. 4178db7bccf4STejun Heo */ 4179db7bccf4STejun Heo 4180706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4181db7bccf4STejun Heo { 418238db41d9STejun Heo int cpu = smp_processor_id(); 41834ce62e9eSTejun Heo struct worker_pool *pool; 4184db7bccf4STejun Heo struct worker *worker; 4185db7bccf4STejun Heo int i; 4186db7bccf4STejun Heo 4187f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41886183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4189db7bccf4STejun Heo 419094cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 419194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4192e22bee78STejun Heo 4193f2d5a0eeSTejun Heo /* 419494cf58bbSTejun Heo * We've claimed all manager positions. Make all workers 419594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 419694cf58bbSTejun Heo * except for the ones which are still executing works from 419794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 419894cf58bbSTejun Heo * this, they may become diasporas. 4199f2d5a0eeSTejun Heo */ 42004ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 4201403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4202db7bccf4STejun Heo 4203b67bfe0dSSasha Levin for_each_busy_worker(worker, i, pool) 4204403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4205db7bccf4STejun Heo 420624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4207f2d5a0eeSTejun Heo 420894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 420994cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 421094cf58bbSTejun Heo } 4211e22bee78STejun Heo 4212e22bee78STejun Heo /* 4213628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4214628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4215628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4216628c78e7STejun Heo */ 4217628c78e7STejun Heo schedule(); 4218628c78e7STejun Heo 4219628c78e7STejun Heo /* 4220628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4221628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 422238db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 422338db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 422438db41d9STejun Heo * pools which are served by workers tied to the CPU. 4225628c78e7STejun Heo * 4226628c78e7STejun Heo * On return from this function, the current worker would trigger 4227628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4228628c78e7STejun Heo * didn't already. 4229e22bee78STejun Heo */ 4230f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4231e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4232db7bccf4STejun Heo } 4233db7bccf4STejun Heo 42348db25e78STejun Heo /* 42358db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 42368db25e78STejun Heo * This will be registered high priority CPU notifier. 42378db25e78STejun Heo */ 42389fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 42391da177e4SLinus Torvalds unsigned long action, 42401da177e4SLinus Torvalds void *hcpu) 42411da177e4SLinus Torvalds { 4242d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42434ce62e9eSTejun Heo struct worker_pool *pool; 42441da177e4SLinus Torvalds 42458db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 42463af24433SOleg Nesterov case CPU_UP_PREPARE: 4247f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 42483ce63377STejun Heo struct worker *worker; 42493ce63377STejun Heo 42503ce63377STejun Heo if (pool->nr_workers) 42513ce63377STejun Heo continue; 42523ce63377STejun Heo 42533ce63377STejun Heo worker = create_worker(pool); 42543ce63377STejun Heo if (!worker) 42553ce63377STejun Heo return NOTIFY_BAD; 42563ce63377STejun Heo 4257d565ed63STejun Heo spin_lock_irq(&pool->lock); 42583ce63377STejun Heo start_worker(worker); 4259d565ed63STejun Heo spin_unlock_irq(&pool->lock); 42603af24433SOleg Nesterov } 42611da177e4SLinus Torvalds break; 42621da177e4SLinus Torvalds 426365758202STejun Heo case CPU_DOWN_FAILED: 426465758202STejun Heo case CPU_ONLINE: 4265f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 426694cf58bbSTejun Heo mutex_lock(&pool->assoc_mutex); 426794cf58bbSTejun Heo spin_lock_irq(&pool->lock); 426894cf58bbSTejun Heo 426924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 427094cf58bbSTejun Heo rebind_workers(pool); 427194cf58bbSTejun Heo 427294cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 427394cf58bbSTejun Heo mutex_unlock(&pool->assoc_mutex); 427494cf58bbSTejun Heo } 42758db25e78STejun Heo break; 427665758202STejun Heo } 427765758202STejun Heo return NOTIFY_OK; 427865758202STejun Heo } 427965758202STejun Heo 428065758202STejun Heo /* 428165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 428265758202STejun Heo * This will be registered as low priority CPU notifier. 428365758202STejun Heo */ 42849fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 428565758202STejun Heo unsigned long action, 428665758202STejun Heo void *hcpu) 428765758202STejun Heo { 4288d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42898db25e78STejun Heo struct work_struct unbind_work; 42908db25e78STejun Heo 429165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 429265758202STejun Heo case CPU_DOWN_PREPARE: 42938db25e78STejun Heo /* unbinding should happen on the local CPU */ 4294706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42957635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42968db25e78STejun Heo flush_work(&unbind_work); 42978db25e78STejun Heo break; 429865758202STejun Heo } 429965758202STejun Heo return NOTIFY_OK; 430065758202STejun Heo } 430165758202STejun Heo 43022d3854a3SRusty Russell #ifdef CONFIG_SMP 43038ccad40dSRusty Russell 43042d3854a3SRusty Russell struct work_for_cpu { 4305ed48ece2STejun Heo struct work_struct work; 43062d3854a3SRusty Russell long (*fn)(void *); 43072d3854a3SRusty Russell void *arg; 43082d3854a3SRusty Russell long ret; 43092d3854a3SRusty Russell }; 43102d3854a3SRusty Russell 4311ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 43122d3854a3SRusty Russell { 4313ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4314ed48ece2STejun Heo 43152d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 43162d3854a3SRusty Russell } 43172d3854a3SRusty Russell 43182d3854a3SRusty Russell /** 43192d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 43202d3854a3SRusty Russell * @cpu: the cpu to run on 43212d3854a3SRusty Russell * @fn: the function to run 43222d3854a3SRusty Russell * @arg: the function arg 43232d3854a3SRusty Russell * 432431ad9081SRusty Russell * This will return the value @fn returns. 432531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 43266b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 43272d3854a3SRusty Russell */ 4328d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 43292d3854a3SRusty Russell { 4330ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 43312d3854a3SRusty Russell 4332ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4333ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4334ed48ece2STejun Heo flush_work(&wfc.work); 43352d3854a3SRusty Russell return wfc.ret; 43362d3854a3SRusty Russell } 43372d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 43382d3854a3SRusty Russell #endif /* CONFIG_SMP */ 43392d3854a3SRusty Russell 4340a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4341e7577c50SRusty Russell 4342a0a1a5fdSTejun Heo /** 4343a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4344a0a1a5fdSTejun Heo * 434558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4346c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4347706026c2STejun Heo * pool->worklist. 4348a0a1a5fdSTejun Heo * 4349a0a1a5fdSTejun Heo * CONTEXT: 4350d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 4351a0a1a5fdSTejun Heo */ 4352a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4353a0a1a5fdSTejun Heo { 435417116969STejun Heo struct worker_pool *pool; 435524b8a847STejun Heo struct workqueue_struct *wq; 435624b8a847STejun Heo struct pool_workqueue *pwq; 435717116969STejun Heo int id; 4358a0a1a5fdSTejun Heo 4359e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 4360a0a1a5fdSTejun Heo 43616183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4362a0a1a5fdSTejun Heo workqueue_freezing = true; 4363a0a1a5fdSTejun Heo 436424b8a847STejun Heo /* set FREEZING */ 436517116969STejun Heo for_each_pool(pool, id) { 4366e98d5b16STejun Heo spin_lock(&pool->lock); 436735b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 436835b6bb63STejun Heo pool->flags |= POOL_FREEZING; 436924b8a847STejun Heo spin_unlock(&pool->lock); 43701da177e4SLinus Torvalds } 43718b03ae3cSTejun Heo 437224b8a847STejun Heo /* suppress further executions by setting max_active to zero */ 437324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4374699ce097STejun Heo for_each_pwq(pwq, wq) 4375699ce097STejun Heo pwq_adjust_max_active(pwq); 4376a1056305STejun Heo } 4377a0a1a5fdSTejun Heo 4378e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 4379a0a1a5fdSTejun Heo } 4380a0a1a5fdSTejun Heo 4381a0a1a5fdSTejun Heo /** 438258a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4383a0a1a5fdSTejun Heo * 4384a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4385a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4386a0a1a5fdSTejun Heo * 4387a0a1a5fdSTejun Heo * CONTEXT: 4388a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 4389a0a1a5fdSTejun Heo * 4390a0a1a5fdSTejun Heo * RETURNS: 439158a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 439258a69cb4STejun Heo * is complete. 4393a0a1a5fdSTejun Heo */ 4394a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4395a0a1a5fdSTejun Heo { 4396a0a1a5fdSTejun Heo bool busy = false; 439724b8a847STejun Heo struct workqueue_struct *wq; 439824b8a847STejun Heo struct pool_workqueue *pwq; 4399a0a1a5fdSTejun Heo 4400e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 4401a0a1a5fdSTejun Heo 44026183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4403a0a1a5fdSTejun Heo 440424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 440524b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 440624b8a847STejun Heo continue; 4407a0a1a5fdSTejun Heo /* 4408a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4409a0a1a5fdSTejun Heo * to peek without lock. 4410a0a1a5fdSTejun Heo */ 441124b8a847STejun Heo for_each_pwq(pwq, wq) { 44126183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4413112202d9STejun Heo if (pwq->nr_active) { 4414a0a1a5fdSTejun Heo busy = true; 4415a0a1a5fdSTejun Heo goto out_unlock; 4416a0a1a5fdSTejun Heo } 4417a0a1a5fdSTejun Heo } 4418a0a1a5fdSTejun Heo } 4419a0a1a5fdSTejun Heo out_unlock: 4420e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 4421a0a1a5fdSTejun Heo return busy; 4422a0a1a5fdSTejun Heo } 4423a0a1a5fdSTejun Heo 4424a0a1a5fdSTejun Heo /** 4425a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4426a0a1a5fdSTejun Heo * 4427a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4428706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4429a0a1a5fdSTejun Heo * 4430a0a1a5fdSTejun Heo * CONTEXT: 4431d565ed63STejun Heo * Grabs and releases workqueue_lock and pool->lock's. 4432a0a1a5fdSTejun Heo */ 4433a0a1a5fdSTejun Heo void thaw_workqueues(void) 4434a0a1a5fdSTejun Heo { 443524b8a847STejun Heo struct workqueue_struct *wq; 443624b8a847STejun Heo struct pool_workqueue *pwq; 443724b8a847STejun Heo struct worker_pool *pool; 443824b8a847STejun Heo int id; 4439a0a1a5fdSTejun Heo 4440e98d5b16STejun Heo spin_lock_irq(&workqueue_lock); 4441a0a1a5fdSTejun Heo 4442a0a1a5fdSTejun Heo if (!workqueue_freezing) 4443a0a1a5fdSTejun Heo goto out_unlock; 4444a0a1a5fdSTejun Heo 444524b8a847STejun Heo /* clear FREEZING */ 444624b8a847STejun Heo for_each_pool(pool, id) { 4447e98d5b16STejun Heo spin_lock(&pool->lock); 444835b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 444935b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 4450e98d5b16STejun Heo spin_unlock(&pool->lock); 4451d565ed63STejun Heo } 445224b8a847STejun Heo 445324b8a847STejun Heo /* restore max_active and repopulate worklist */ 445424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4455699ce097STejun Heo for_each_pwq(pwq, wq) 4456699ce097STejun Heo pwq_adjust_max_active(pwq); 445724b8a847STejun Heo } 445824b8a847STejun Heo 445924b8a847STejun Heo /* kick workers */ 446024b8a847STejun Heo for_each_pool(pool, id) { 446124b8a847STejun Heo spin_lock(&pool->lock); 446224b8a847STejun Heo wake_up_worker(pool); 446324b8a847STejun Heo spin_unlock(&pool->lock); 4464a0a1a5fdSTejun Heo } 4465a0a1a5fdSTejun Heo 4466a0a1a5fdSTejun Heo workqueue_freezing = false; 4467a0a1a5fdSTejun Heo out_unlock: 4468e98d5b16STejun Heo spin_unlock_irq(&workqueue_lock); 4469a0a1a5fdSTejun Heo } 4470a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4471a0a1a5fdSTejun Heo 44726ee0578bSSuresh Siddha static int __init init_workqueues(void) 44731da177e4SLinus Torvalds { 44747a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44757a4e344cSTejun Heo int i, cpu; 4476c34056a3STejun Heo 44777c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44787c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44796be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4480b5490077STejun Heo 4481e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4482e904e6c2STejun Heo 4483e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4484e904e6c2STejun Heo 448565758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4486a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44878b03ae3cSTejun Heo 4488706026c2STejun Heo /* initialize CPU pools */ 448929c91e99STejun Heo for_each_possible_cpu(cpu) { 44904ce62e9eSTejun Heo struct worker_pool *pool; 44918b03ae3cSTejun Heo 44927a4e344cSTejun Heo i = 0; 4493f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44947a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4495ec22ca5eSTejun Heo pool->cpu = cpu; 44967a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 44977a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 44987a4e344cSTejun Heo 44999daf9e67STejun Heo /* alloc pool ID */ 45009daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 45014ce62e9eSTejun Heo } 45028b03ae3cSTejun Heo } 45038b03ae3cSTejun Heo 4504e22bee78STejun Heo /* create the initial worker */ 450529c91e99STejun Heo for_each_online_cpu(cpu) { 45064ce62e9eSTejun Heo struct worker_pool *pool; 4507e22bee78STejun Heo 4508f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45094ce62e9eSTejun Heo struct worker *worker; 45104ce62e9eSTejun Heo 451124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 451224647570STejun Heo 4513bc2ae0f5STejun Heo worker = create_worker(pool); 4514e22bee78STejun Heo BUG_ON(!worker); 4515d565ed63STejun Heo spin_lock_irq(&pool->lock); 4516e22bee78STejun Heo start_worker(worker); 4517d565ed63STejun Heo spin_unlock_irq(&pool->lock); 4518e22bee78STejun Heo } 45194ce62e9eSTejun Heo } 4520e22bee78STejun Heo 452129c91e99STejun Heo /* create default unbound wq attrs */ 452229c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 452329c91e99STejun Heo struct workqueue_attrs *attrs; 452429c91e99STejun Heo 452529c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 452629c91e99STejun Heo 452729c91e99STejun Heo attrs->nice = std_nice[i]; 452829c91e99STejun Heo cpumask_setall(attrs->cpumask); 452929c91e99STejun Heo 453029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 453129c91e99STejun Heo } 453229c91e99STejun Heo 4533d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 45341aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4535d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4536f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4537f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 453824d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 453924d51addSTejun Heo WQ_FREEZABLE, 0); 45401aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4541ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 45426ee0578bSSuresh Siddha return 0; 45431da177e4SLinus Torvalds } 45446ee0578bSSuresh Siddha early_initcall(init_workqueues); 4545