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 * 63bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 64bc3a1afcSTejun Heo * manager_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 */ 78a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 79e22bee78STejun Heo 80a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 81a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 82db7bccf4STejun Heo 83e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 844ce62e9eSTejun Heo 8529c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 86c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 87db7bccf4STejun Heo 88e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 89e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 90e22bee78STejun Heo 913233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 923233cdbdSTejun Heo /* call for help after 10ms 933233cdbdSTejun Heo (min two ticks) */ 94e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 95e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 961da177e4SLinus Torvalds 971da177e4SLinus Torvalds /* 98e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 99e22bee78STejun Heo * all cpus. Give -20. 100e22bee78STejun Heo */ 101e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1023270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 103c8e55f36STejun Heo }; 104c8e55f36STejun Heo 1051da177e4SLinus Torvalds /* 1064690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1074690c4abSTejun Heo * 108e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 109e41e704bSTejun Heo * everyone else. 1104690c4abSTejun Heo * 111e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 112e22bee78STejun Heo * only be modified and accessed from the local cpu. 113e22bee78STejun Heo * 114d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1154690c4abSTejun Heo * 116d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 117d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 118d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 119d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 120e22bee78STejun Heo * 12173f53c4aSTejun Heo * F: wq->flush_mutex protected. 12273f53c4aSTejun Heo * 123822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 124822d8405STejun Heo * locks. Reads can happen under either lock. 125822d8405STejun Heo * 12668e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12776af4d93STejun Heo * 12868e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1295bcab335STejun Heo * 130794b18bcSTejun Heo * PW: pwq_lock protected. 131794b18bcSTejun Heo * 132794b18bcSTejun Heo * FR: wq->flush_mutex and pwq_lock protected for writes. Sched-RCU 13375ccf595STejun Heo * protected for reads. 1342e109a28STejun Heo * 1352e109a28STejun Heo * MD: wq_mayday_lock protected. 1364690c4abSTejun Heo */ 1374690c4abSTejun Heo 1382eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 139c34056a3STejun Heo 140bd7bdd43STejun Heo struct worker_pool { 141d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 142d84ff051STejun Heo int cpu; /* I: the associated cpu */ 1439daf9e67STejun Heo int id; /* I: pool ID */ 14411ebea50STejun Heo unsigned int flags; /* X: flags */ 145bd7bdd43STejun Heo 146bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 147bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 148ea1abd61SLai Jiangshan 149ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 150bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 151bd7bdd43STejun Heo 152bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 153bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 154bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 155bd7bdd43STejun Heo 156c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 157c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 158c9e7cf27STejun Heo /* L: hash of busy workers */ 159c9e7cf27STejun Heo 160bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16134a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 162bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 163822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 164e19e397aSTejun Heo 1657a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16668e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 16768e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1687a4e344cSTejun Heo 169e19e397aSTejun Heo /* 170e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 171e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 172e19e397aSTejun Heo * cacheline. 173e19e397aSTejun Heo */ 174e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17529c91e99STejun Heo 17629c91e99STejun Heo /* 17729c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 17829c91e99STejun Heo * from get_work_pool(). 17929c91e99STejun Heo */ 18029c91e99STejun Heo struct rcu_head rcu; 1818b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1828b03ae3cSTejun Heo 1838b03ae3cSTejun Heo /* 184112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 185112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 186112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 187112202d9STejun Heo * number of flag bits. 1881da177e4SLinus Torvalds */ 189112202d9STejun Heo struct pool_workqueue { 190bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1914690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19273f53c4aSTejun Heo int work_color; /* L: current color */ 19373f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1948864b4e5STejun Heo int refcnt; /* L: reference count */ 19573f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19673f53c4aSTejun Heo /* L: nr of in_flight works */ 1971e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 198a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1991e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 20075ccf595STejun Heo struct list_head pwqs_node; /* FR: node on wq->pwqs */ 2012e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2028864b4e5STejun Heo 2038864b4e5STejun Heo /* 2048864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2058864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2068864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 207794b18bcSTejun Heo * determined without grabbing pwq_lock. 2088864b4e5STejun Heo */ 2098864b4e5STejun Heo struct work_struct unbound_release_work; 2108864b4e5STejun Heo struct rcu_head rcu; 211e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2121da177e4SLinus Torvalds 2131da177e4SLinus Torvalds /* 21473f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21573f53c4aSTejun Heo */ 21673f53c4aSTejun Heo struct wq_flusher { 21773f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 21873f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 21973f53c4aSTejun Heo struct completion done; /* flush completion */ 22073f53c4aSTejun Heo }; 2211da177e4SLinus Torvalds 222226223abSTejun Heo struct wq_device; 223226223abSTejun Heo 22473f53c4aSTejun Heo /* 225c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 226c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2271da177e4SLinus Torvalds */ 2281da177e4SLinus Torvalds struct workqueue_struct { 22968e13a67SLai Jiangshan unsigned int flags; /* PL: WQ_* flags */ 230420c0ddbSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ 23175ccf595STejun Heo struct list_head pwqs; /* FR: all pwqs of this wq */ 23268e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23373f53c4aSTejun Heo 23473f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 23573f53c4aSTejun Heo int work_color; /* F: current work color */ 23673f53c4aSTejun Heo int flush_color; /* F: current flush color */ 237112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 23873f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 23973f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 24073f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 24173f53c4aSTejun Heo 2422e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 243e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 244e22bee78STejun Heo 24568e13a67SLai Jiangshan int nr_drainers; /* PL: drain in progress */ 246794b18bcSTejun Heo int saved_max_active; /* PW: saved pwq max_active */ 247226223abSTejun Heo 248226223abSTejun Heo #ifdef CONFIG_SYSFS 249226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 250226223abSTejun Heo #endif 2514e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2524e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2534e6045f1SJohannes Berg #endif 254b196be89STejun Heo char name[]; /* I: workqueue name */ 2551da177e4SLinus Torvalds }; 2561da177e4SLinus Torvalds 257e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 258e904e6c2STejun Heo 25968e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 260794b18bcSTejun Heo static DEFINE_SPINLOCK(pwq_lock); /* protects pool_workqueues */ 2612e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2625bcab335STejun Heo 26368e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 26468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2657d19c5ceSTejun Heo 2667d19c5ceSTejun Heo /* the per-cpu worker pools */ 2677d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2687d19c5ceSTejun Heo cpu_worker_pools); 2697d19c5ceSTejun Heo 27068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2717d19c5ceSTejun Heo 27268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 27329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 27429c91e99STejun Heo 275c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 27629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 27729c91e99STejun Heo 278d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 279d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 280044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2811aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 282044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 283d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 284044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 285f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 286044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 28724d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 288d320c038STejun Heo 2897d19c5ceSTejun Heo static int worker_thread(void *__worker); 2907d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 2917d19c5ceSTejun Heo const struct workqueue_attrs *from); 2927d19c5ceSTejun Heo 29397bd2347STejun Heo #define CREATE_TRACE_POINTS 29497bd2347STejun Heo #include <trace/events/workqueue.h> 29597bd2347STejun Heo 29668e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 2975bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 29868e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 29968e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3005bcab335STejun Heo 301794b18bcSTejun Heo #define assert_rcu_or_pwq_lock() \ 30276af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 303794b18bcSTejun Heo lockdep_is_held(&pwq_lock), \ 304794b18bcSTejun Heo "sched RCU or pwq_lock should be held") 30576af4d93STejun Heo 306822d8405STejun Heo #ifdef CONFIG_LOCKDEP 307822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 308519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 309519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 310822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 311822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 312822d8405STejun Heo #else 313822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 314822d8405STejun Heo #endif 315822d8405STejun Heo 316f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 317f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 318f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3197a62c2c8STejun Heo (pool)++) 3204ce62e9eSTejun Heo 32149e3cf44STejun Heo /** 32217116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 32317116969STejun Heo * @pool: iteration cursor 324611c92a0STejun Heo * @pi: integer used for iteration 325fa1b54e6STejun Heo * 32668e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 32768e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 32868e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 329fa1b54e6STejun Heo * 330fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 331fa1b54e6STejun Heo * ignored. 33217116969STejun Heo */ 333611c92a0STejun Heo #define for_each_pool(pool, pi) \ 334611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 33568e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 336fa1b54e6STejun Heo else 33717116969STejun Heo 33817116969STejun Heo /** 339822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 340822d8405STejun Heo * @worker: iteration cursor 341822d8405STejun Heo * @wi: integer used for iteration 342822d8405STejun Heo * @pool: worker_pool to iterate workers of 343822d8405STejun Heo * 344822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 345822d8405STejun Heo * 346822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 347822d8405STejun Heo * ignored. 348822d8405STejun Heo */ 349822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 350822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 351822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 352822d8405STejun Heo else 353822d8405STejun Heo 354822d8405STejun Heo /** 35549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 35649e3cf44STejun Heo * @pwq: iteration cursor 35749e3cf44STejun Heo * @wq: the target workqueue 35876af4d93STejun Heo * 359794b18bcSTejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 360794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 361794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 36276af4d93STejun Heo * 36376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 36476af4d93STejun Heo * ignored. 36549e3cf44STejun Heo */ 36649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 36776af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 368794b18bcSTejun Heo if (({ assert_rcu_or_pwq_lock(); false; })) { } \ 36976af4d93STejun Heo else 370f3421797STejun Heo 371dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 372dc186ad7SThomas Gleixner 373dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 374dc186ad7SThomas Gleixner 37599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 37699777288SStanislaw Gruszka { 37799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 37899777288SStanislaw Gruszka } 37999777288SStanislaw Gruszka 380dc186ad7SThomas Gleixner /* 381dc186ad7SThomas Gleixner * fixup_init is called when: 382dc186ad7SThomas Gleixner * - an active object is initialized 383dc186ad7SThomas Gleixner */ 384dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 385dc186ad7SThomas Gleixner { 386dc186ad7SThomas Gleixner struct work_struct *work = addr; 387dc186ad7SThomas Gleixner 388dc186ad7SThomas Gleixner switch (state) { 389dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 390dc186ad7SThomas Gleixner cancel_work_sync(work); 391dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 392dc186ad7SThomas Gleixner return 1; 393dc186ad7SThomas Gleixner default: 394dc186ad7SThomas Gleixner return 0; 395dc186ad7SThomas Gleixner } 396dc186ad7SThomas Gleixner } 397dc186ad7SThomas Gleixner 398dc186ad7SThomas Gleixner /* 399dc186ad7SThomas Gleixner * fixup_activate is called when: 400dc186ad7SThomas Gleixner * - an active object is activated 401dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 402dc186ad7SThomas Gleixner */ 403dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 404dc186ad7SThomas Gleixner { 405dc186ad7SThomas Gleixner struct work_struct *work = addr; 406dc186ad7SThomas Gleixner 407dc186ad7SThomas Gleixner switch (state) { 408dc186ad7SThomas Gleixner 409dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 410dc186ad7SThomas Gleixner /* 411dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 412dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 413dc186ad7SThomas Gleixner * is tracked in the object tracker. 414dc186ad7SThomas Gleixner */ 41522df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 416dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 417dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 418dc186ad7SThomas Gleixner return 0; 419dc186ad7SThomas Gleixner } 420dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 421dc186ad7SThomas Gleixner return 0; 422dc186ad7SThomas Gleixner 423dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 424dc186ad7SThomas Gleixner WARN_ON(1); 425dc186ad7SThomas Gleixner 426dc186ad7SThomas Gleixner default: 427dc186ad7SThomas Gleixner return 0; 428dc186ad7SThomas Gleixner } 429dc186ad7SThomas Gleixner } 430dc186ad7SThomas Gleixner 431dc186ad7SThomas Gleixner /* 432dc186ad7SThomas Gleixner * fixup_free is called when: 433dc186ad7SThomas Gleixner * - an active object is freed 434dc186ad7SThomas Gleixner */ 435dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 436dc186ad7SThomas Gleixner { 437dc186ad7SThomas Gleixner struct work_struct *work = addr; 438dc186ad7SThomas Gleixner 439dc186ad7SThomas Gleixner switch (state) { 440dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 441dc186ad7SThomas Gleixner cancel_work_sync(work); 442dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 443dc186ad7SThomas Gleixner return 1; 444dc186ad7SThomas Gleixner default: 445dc186ad7SThomas Gleixner return 0; 446dc186ad7SThomas Gleixner } 447dc186ad7SThomas Gleixner } 448dc186ad7SThomas Gleixner 449dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 450dc186ad7SThomas Gleixner .name = "work_struct", 45199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 452dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 453dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 454dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 455dc186ad7SThomas Gleixner }; 456dc186ad7SThomas Gleixner 457dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 458dc186ad7SThomas Gleixner { 459dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner 462dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 463dc186ad7SThomas Gleixner { 464dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 465dc186ad7SThomas Gleixner } 466dc186ad7SThomas Gleixner 467dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner if (onstack) 470dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 471dc186ad7SThomas Gleixner else 472dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 473dc186ad7SThomas Gleixner } 474dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 475dc186ad7SThomas Gleixner 476dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 477dc186ad7SThomas Gleixner { 478dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner #else 483dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 484dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 485dc186ad7SThomas Gleixner #endif 486dc186ad7SThomas Gleixner 4879daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4889daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4899daf9e67STejun Heo { 4909daf9e67STejun Heo int ret; 4919daf9e67STejun Heo 49268e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4935bcab335STejun Heo 494fa1b54e6STejun Heo do { 495fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 496fa1b54e6STejun Heo return -ENOMEM; 4979daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 498fa1b54e6STejun Heo } while (ret == -EAGAIN); 4999daf9e67STejun Heo 5009daf9e67STejun Heo return ret; 5019daf9e67STejun Heo } 5029daf9e67STejun Heo 50376af4d93STejun Heo /** 50476af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 50576af4d93STejun Heo * @wq: the target workqueue 50676af4d93STejun Heo * 507794b18bcSTejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 508794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 509794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 51076af4d93STejun Heo */ 5117fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 512a848e3b6SOleg Nesterov { 513794b18bcSTejun Heo assert_rcu_or_pwq_lock(); 51476af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 51576af4d93STejun Heo pwqs_node); 516f3421797STejun Heo } 517a848e3b6SOleg Nesterov 51873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 51973f53c4aSTejun Heo { 52073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 52173f53c4aSTejun Heo } 52273f53c4aSTejun Heo 52373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 52473f53c4aSTejun Heo { 52573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 52673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 52773f53c4aSTejun Heo } 52873f53c4aSTejun Heo 52973f53c4aSTejun Heo static int work_next_color(int color) 53073f53c4aSTejun Heo { 53173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 532b1f4ec17SOleg Nesterov } 533b1f4ec17SOleg Nesterov 5344594bf15SDavid Howells /* 535112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 536112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5377c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5387a22ad75STejun Heo * 539112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 540112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 541bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 542bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5437a22ad75STejun Heo * 544112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5457c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 546112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5477c3eed5cSTejun Heo * available only while the work item is queued. 548bbb68dfaSTejun Heo * 549bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 550bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 551bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 552bbb68dfaSTejun Heo * try to steal the PENDING bit. 5534594bf15SDavid Howells */ 5547a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5557a22ad75STejun Heo unsigned long flags) 5567a22ad75STejun Heo { 5576183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5587a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5597a22ad75STejun Heo } 5607a22ad75STejun Heo 561112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5624690c4abSTejun Heo unsigned long extra_flags) 563365970a1SDavid Howells { 564112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 565112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 566365970a1SDavid Howells } 567365970a1SDavid Howells 5684468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5694468a00fSLai Jiangshan int pool_id) 5704468a00fSLai Jiangshan { 5714468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5724468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5734468a00fSLai Jiangshan } 5744468a00fSLai Jiangshan 5757c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5767c3eed5cSTejun Heo int pool_id) 5774d707b9fSOleg Nesterov { 57823657bb1STejun Heo /* 57923657bb1STejun Heo * The following wmb is paired with the implied mb in 58023657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 58123657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 58223657bb1STejun Heo * owner. 58323657bb1STejun Heo */ 58423657bb1STejun Heo smp_wmb(); 5857c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5864d707b9fSOleg Nesterov } 5874d707b9fSOleg Nesterov 5887a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 589365970a1SDavid Howells { 5907c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5917c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5927a22ad75STejun Heo } 5937a22ad75STejun Heo 594112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5957a22ad75STejun Heo { 596e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5977a22ad75STejun Heo 598112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 599e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 600e120153dSTejun Heo else 601e120153dSTejun Heo return NULL; 6027a22ad75STejun Heo } 6037a22ad75STejun Heo 6047c3eed5cSTejun Heo /** 6057c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6067c3eed5cSTejun Heo * @work: the work item of interest 6077c3eed5cSTejun Heo * 6087c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 609fa1b54e6STejun Heo * 61068e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 61168e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 61268e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 613fa1b54e6STejun Heo * 614fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 615fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 616fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 617fa1b54e6STejun Heo * returned pool is and stays online. 6187c3eed5cSTejun Heo */ 6197c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6207a22ad75STejun Heo { 621e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6227c3eed5cSTejun Heo int pool_id; 6237a22ad75STejun Heo 62468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 625fa1b54e6STejun Heo 626112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 627112202d9STejun Heo return ((struct pool_workqueue *) 6287c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6297a22ad75STejun Heo 6307c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6317c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6327a22ad75STejun Heo return NULL; 6337a22ad75STejun Heo 634fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6357c3eed5cSTejun Heo } 6367c3eed5cSTejun Heo 6377c3eed5cSTejun Heo /** 6387c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6397c3eed5cSTejun Heo * @work: the work item of interest 6407c3eed5cSTejun Heo * 6417c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6427c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6437c3eed5cSTejun Heo */ 6447c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6457c3eed5cSTejun Heo { 64654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6477c3eed5cSTejun Heo 648112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 649112202d9STejun Heo return ((struct pool_workqueue *) 65054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 65154d5b7d0SLai Jiangshan 65254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6537c3eed5cSTejun Heo } 6547c3eed5cSTejun Heo 655bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 656bbb68dfaSTejun Heo { 6577c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 658bbb68dfaSTejun Heo 6597c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6607c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 661bbb68dfaSTejun Heo } 662bbb68dfaSTejun Heo 663bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 664bbb68dfaSTejun Heo { 665bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 666bbb68dfaSTejun Heo 667112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 668bbb68dfaSTejun Heo } 669bbb68dfaSTejun Heo 670e22bee78STejun Heo /* 6713270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6723270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 673d565ed63STejun Heo * they're being called with pool->lock held. 674e22bee78STejun Heo */ 675e22bee78STejun Heo 67663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 677649027d7STejun Heo { 678e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 679649027d7STejun Heo } 680649027d7STejun Heo 681e22bee78STejun Heo /* 682e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 683e22bee78STejun Heo * running workers. 684974271c4STejun Heo * 685974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 686706026c2STejun Heo * function will always return %true for unbound pools as long as the 687974271c4STejun Heo * worklist isn't empty. 688e22bee78STejun Heo */ 68963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 690e22bee78STejun Heo { 69163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 692e22bee78STejun Heo } 693e22bee78STejun Heo 694e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 69563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 696e22bee78STejun Heo { 69763d95a91STejun Heo return pool->nr_idle; 698e22bee78STejun Heo } 699e22bee78STejun Heo 700e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 70163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 702e22bee78STejun Heo { 703e19e397aSTejun Heo return !list_empty(&pool->worklist) && 704e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 705e22bee78STejun Heo } 706e22bee78STejun Heo 707e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 70863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 709e22bee78STejun Heo { 71063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 711e22bee78STejun Heo } 712e22bee78STejun Heo 713e22bee78STejun Heo /* Do I need to be the manager? */ 71463d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 715e22bee78STejun Heo { 71663d95a91STejun Heo return need_to_create_worker(pool) || 71711ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 718e22bee78STejun Heo } 719e22bee78STejun Heo 720e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 72163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 722e22bee78STejun Heo { 72334a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 72463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 72563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 726e22bee78STejun Heo 727ea1abd61SLai Jiangshan /* 728ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 729ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 730ea1abd61SLai Jiangshan */ 731ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 732ea1abd61SLai Jiangshan return false; 733ea1abd61SLai Jiangshan 734e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 735e22bee78STejun Heo } 736e22bee78STejun Heo 737e22bee78STejun Heo /* 738e22bee78STejun Heo * Wake up functions. 739e22bee78STejun Heo */ 740e22bee78STejun Heo 7417e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 74263d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7437e11629dSTejun Heo { 74463d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7457e11629dSTejun Heo return NULL; 7467e11629dSTejun Heo 74763d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7487e11629dSTejun Heo } 7497e11629dSTejun Heo 7507e11629dSTejun Heo /** 7517e11629dSTejun Heo * wake_up_worker - wake up an idle worker 75263d95a91STejun Heo * @pool: worker pool to wake worker from 7537e11629dSTejun Heo * 75463d95a91STejun Heo * Wake up the first idle worker of @pool. 7557e11629dSTejun Heo * 7567e11629dSTejun Heo * CONTEXT: 757d565ed63STejun Heo * spin_lock_irq(pool->lock). 7587e11629dSTejun Heo */ 75963d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7607e11629dSTejun Heo { 76163d95a91STejun Heo struct worker *worker = first_worker(pool); 7627e11629dSTejun Heo 7637e11629dSTejun Heo if (likely(worker)) 7647e11629dSTejun Heo wake_up_process(worker->task); 7657e11629dSTejun Heo } 7667e11629dSTejun Heo 7674690c4abSTejun Heo /** 768e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 769e22bee78STejun Heo * @task: task waking up 770e22bee78STejun Heo * @cpu: CPU @task is waking up to 771e22bee78STejun Heo * 772e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 773e22bee78STejun Heo * being awoken. 774e22bee78STejun Heo * 775e22bee78STejun Heo * CONTEXT: 776e22bee78STejun Heo * spin_lock_irq(rq->lock) 777e22bee78STejun Heo */ 778d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 779e22bee78STejun Heo { 780e22bee78STejun Heo struct worker *worker = kthread_data(task); 781e22bee78STejun Heo 78236576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 783ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 784e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 785e22bee78STejun Heo } 78636576000SJoonsoo Kim } 787e22bee78STejun Heo 788e22bee78STejun Heo /** 789e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 790e22bee78STejun Heo * @task: task going to sleep 791e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 792e22bee78STejun Heo * 793e22bee78STejun Heo * This function is called during schedule() when a busy worker is 794e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 795e22bee78STejun Heo * returning pointer to its task. 796e22bee78STejun Heo * 797e22bee78STejun Heo * CONTEXT: 798e22bee78STejun Heo * spin_lock_irq(rq->lock) 799e22bee78STejun Heo * 800e22bee78STejun Heo * RETURNS: 801e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 802e22bee78STejun Heo */ 803d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 804e22bee78STejun Heo { 805e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 806111c225aSTejun Heo struct worker_pool *pool; 807e22bee78STejun Heo 808111c225aSTejun Heo /* 809111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 810111c225aSTejun Heo * workers, also reach here, let's not access anything before 811111c225aSTejun Heo * checking NOT_RUNNING. 812111c225aSTejun Heo */ 8132d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 814e22bee78STejun Heo return NULL; 815e22bee78STejun Heo 816111c225aSTejun Heo pool = worker->pool; 817111c225aSTejun Heo 818e22bee78STejun Heo /* this can only happen on the local cpu */ 8196183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8206183c009STejun Heo return NULL; 821e22bee78STejun Heo 822e22bee78STejun Heo /* 823e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 824e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 825e22bee78STejun Heo * Please read comment there. 826e22bee78STejun Heo * 827628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 828628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 829628c78e7STejun Heo * disabled, which in turn means that none else could be 830d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 831628c78e7STejun Heo * lock is safe. 832e22bee78STejun Heo */ 833e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 834e19e397aSTejun Heo !list_empty(&pool->worklist)) 83563d95a91STejun Heo to_wakeup = first_worker(pool); 836e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 837e22bee78STejun Heo } 838e22bee78STejun Heo 839e22bee78STejun Heo /** 840e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 841cb444766STejun Heo * @worker: self 842d302f017STejun Heo * @flags: flags to set 843d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 844d302f017STejun Heo * 845e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 846e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 847e22bee78STejun Heo * woken up. 848d302f017STejun Heo * 849cb444766STejun Heo * CONTEXT: 850d565ed63STejun Heo * spin_lock_irq(pool->lock) 851d302f017STejun Heo */ 852d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 853d302f017STejun Heo bool wakeup) 854d302f017STejun Heo { 855bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 856e22bee78STejun Heo 857cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 858cb444766STejun Heo 859e22bee78STejun Heo /* 860e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 861e22bee78STejun Heo * wake up an idle worker as necessary if requested by 862e22bee78STejun Heo * @wakeup. 863e22bee78STejun Heo */ 864e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 865e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 866e22bee78STejun Heo if (wakeup) { 867e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 868bd7bdd43STejun Heo !list_empty(&pool->worklist)) 86963d95a91STejun Heo wake_up_worker(pool); 870e22bee78STejun Heo } else 871e19e397aSTejun Heo atomic_dec(&pool->nr_running); 872e22bee78STejun Heo } 873e22bee78STejun Heo 874d302f017STejun Heo worker->flags |= flags; 875d302f017STejun Heo } 876d302f017STejun Heo 877d302f017STejun Heo /** 878e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 879cb444766STejun Heo * @worker: self 880d302f017STejun Heo * @flags: flags to clear 881d302f017STejun Heo * 882e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 883d302f017STejun Heo * 884cb444766STejun Heo * CONTEXT: 885d565ed63STejun Heo * spin_lock_irq(pool->lock) 886d302f017STejun Heo */ 887d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 888d302f017STejun Heo { 88963d95a91STejun Heo struct worker_pool *pool = worker->pool; 890e22bee78STejun Heo unsigned int oflags = worker->flags; 891e22bee78STejun Heo 892cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 893cb444766STejun Heo 894d302f017STejun Heo worker->flags &= ~flags; 895e22bee78STejun Heo 89642c025f3STejun Heo /* 89742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 89842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 89942c025f3STejun Heo * of multiple flags, not a single flag. 90042c025f3STejun Heo */ 901e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 902e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 903e19e397aSTejun Heo atomic_inc(&pool->nr_running); 904d302f017STejun Heo } 905d302f017STejun Heo 906d302f017STejun Heo /** 9078cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 908c9e7cf27STejun Heo * @pool: pool of interest 9098cca0eeaSTejun Heo * @work: work to find worker for 9108cca0eeaSTejun Heo * 911c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 912c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 913a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 914a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 915a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 916a2c1c57bSTejun Heo * being executed. 917a2c1c57bSTejun Heo * 918a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 919a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 920a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 921a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 922a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 923a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 924a2c1c57bSTejun Heo * 925c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 926c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 927c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 928c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 929c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 930c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9318cca0eeaSTejun Heo * 9328cca0eeaSTejun Heo * CONTEXT: 933d565ed63STejun Heo * spin_lock_irq(pool->lock). 9348cca0eeaSTejun Heo * 9358cca0eeaSTejun Heo * RETURNS: 9368cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9378cca0eeaSTejun Heo * otherwise. 9388cca0eeaSTejun Heo */ 939c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9408cca0eeaSTejun Heo struct work_struct *work) 9418cca0eeaSTejun Heo { 94242f8570fSSasha Levin struct worker *worker; 94342f8570fSSasha Levin 944b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 945a2c1c57bSTejun Heo (unsigned long)work) 946a2c1c57bSTejun Heo if (worker->current_work == work && 947a2c1c57bSTejun Heo worker->current_func == work->func) 94842f8570fSSasha Levin return worker; 94942f8570fSSasha Levin 95042f8570fSSasha Levin return NULL; 9518cca0eeaSTejun Heo } 9528cca0eeaSTejun Heo 9538cca0eeaSTejun Heo /** 954bf4ede01STejun Heo * move_linked_works - move linked works to a list 955bf4ede01STejun Heo * @work: start of series of works to be scheduled 956bf4ede01STejun Heo * @head: target list to append @work to 957bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 958bf4ede01STejun Heo * 959bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 960bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 961bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 962bf4ede01STejun Heo * 963bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 964bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 965bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 966bf4ede01STejun Heo * 967bf4ede01STejun Heo * CONTEXT: 968d565ed63STejun Heo * spin_lock_irq(pool->lock). 969bf4ede01STejun Heo */ 970bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 971bf4ede01STejun Heo struct work_struct **nextp) 972bf4ede01STejun Heo { 973bf4ede01STejun Heo struct work_struct *n; 974bf4ede01STejun Heo 975bf4ede01STejun Heo /* 976bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 977bf4ede01STejun Heo * use NULL for list head. 978bf4ede01STejun Heo */ 979bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 980bf4ede01STejun Heo list_move_tail(&work->entry, head); 981bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 982bf4ede01STejun Heo break; 983bf4ede01STejun Heo } 984bf4ede01STejun Heo 985bf4ede01STejun Heo /* 986bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 987bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 988bf4ede01STejun Heo * needs to be updated. 989bf4ede01STejun Heo */ 990bf4ede01STejun Heo if (nextp) 991bf4ede01STejun Heo *nextp = n; 992bf4ede01STejun Heo } 993bf4ede01STejun Heo 9948864b4e5STejun Heo /** 9958864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 9968864b4e5STejun Heo * @pwq: pool_workqueue to get 9978864b4e5STejun Heo * 9988864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 9998864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10008864b4e5STejun Heo */ 10018864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10028864b4e5STejun Heo { 10038864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10048864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10058864b4e5STejun Heo pwq->refcnt++; 10068864b4e5STejun Heo } 10078864b4e5STejun Heo 10088864b4e5STejun Heo /** 10098864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10108864b4e5STejun Heo * @pwq: pool_workqueue to put 10118864b4e5STejun Heo * 10128864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10138864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10148864b4e5STejun Heo */ 10158864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10168864b4e5STejun Heo { 10178864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10188864b4e5STejun Heo if (likely(--pwq->refcnt)) 10198864b4e5STejun Heo return; 10208864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10218864b4e5STejun Heo return; 10228864b4e5STejun Heo /* 10238864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10248864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10258864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10268864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10278864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10288864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10298864b4e5STejun Heo */ 10308864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10318864b4e5STejun Heo } 10328864b4e5STejun Heo 1033112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1034bf4ede01STejun Heo { 1035112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1036bf4ede01STejun Heo 1037bf4ede01STejun Heo trace_workqueue_activate_work(work); 1038112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1039bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1040112202d9STejun Heo pwq->nr_active++; 1041bf4ede01STejun Heo } 1042bf4ede01STejun Heo 1043112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10443aa62497SLai Jiangshan { 1045112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10463aa62497SLai Jiangshan struct work_struct, entry); 10473aa62497SLai Jiangshan 1048112202d9STejun Heo pwq_activate_delayed_work(work); 10493aa62497SLai Jiangshan } 10503aa62497SLai Jiangshan 1051bf4ede01STejun Heo /** 1052112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1053112202d9STejun Heo * @pwq: pwq of interest 1054bf4ede01STejun Heo * @color: color of work which left the queue 1055bf4ede01STejun Heo * 1056bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1057112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1058bf4ede01STejun Heo * 1059bf4ede01STejun Heo * CONTEXT: 1060d565ed63STejun Heo * spin_lock_irq(pool->lock). 1061bf4ede01STejun Heo */ 1062112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1063bf4ede01STejun Heo { 10648864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1065bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10668864b4e5STejun Heo goto out_put; 1067bf4ede01STejun Heo 1068112202d9STejun Heo pwq->nr_in_flight[color]--; 1069bf4ede01STejun Heo 1070112202d9STejun Heo pwq->nr_active--; 1071112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1072bf4ede01STejun Heo /* one down, submit a delayed one */ 1073112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1074112202d9STejun Heo pwq_activate_first_delayed(pwq); 1075bf4ede01STejun Heo } 1076bf4ede01STejun Heo 1077bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1078112202d9STejun Heo if (likely(pwq->flush_color != color)) 10798864b4e5STejun Heo goto out_put; 1080bf4ede01STejun Heo 1081bf4ede01STejun Heo /* are there still in-flight works? */ 1082112202d9STejun Heo if (pwq->nr_in_flight[color]) 10838864b4e5STejun Heo goto out_put; 1084bf4ede01STejun Heo 1085112202d9STejun Heo /* this pwq is done, clear flush_color */ 1086112202d9STejun Heo pwq->flush_color = -1; 1087bf4ede01STejun Heo 1088bf4ede01STejun Heo /* 1089112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1090bf4ede01STejun Heo * will handle the rest. 1091bf4ede01STejun Heo */ 1092112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1093112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 10948864b4e5STejun Heo out_put: 10958864b4e5STejun Heo put_pwq(pwq); 1096bf4ede01STejun Heo } 1097bf4ede01STejun Heo 109836e227d2STejun Heo /** 1099bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 110036e227d2STejun Heo * @work: work item to steal 110136e227d2STejun Heo * @is_dwork: @work is a delayed_work 1102bbb68dfaSTejun Heo * @flags: place to store irq state 110336e227d2STejun Heo * 110436e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 110536e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 110636e227d2STejun Heo * 110736e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 110836e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 110936e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1110bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1111bbb68dfaSTejun Heo * for arbitrarily long 111236e227d2STejun Heo * 1113bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1114e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1115e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1116e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1117bbb68dfaSTejun Heo * 1118bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1119bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1120bbb68dfaSTejun Heo * 1121e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1122bf4ede01STejun Heo */ 1123bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1124bbb68dfaSTejun Heo unsigned long *flags) 1125bf4ede01STejun Heo { 1126d565ed63STejun Heo struct worker_pool *pool; 1127112202d9STejun Heo struct pool_workqueue *pwq; 1128bf4ede01STejun Heo 1129bbb68dfaSTejun Heo local_irq_save(*flags); 1130bbb68dfaSTejun Heo 113136e227d2STejun Heo /* try to steal the timer if it exists */ 113236e227d2STejun Heo if (is_dwork) { 113336e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 113436e227d2STejun Heo 1135e0aecdd8STejun Heo /* 1136e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1137e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1138e0aecdd8STejun Heo * running on the local CPU. 1139e0aecdd8STejun Heo */ 114036e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 114136e227d2STejun Heo return 1; 114236e227d2STejun Heo } 114336e227d2STejun Heo 114436e227d2STejun Heo /* try to claim PENDING the normal way */ 1145bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1146bf4ede01STejun Heo return 0; 1147bf4ede01STejun Heo 1148bf4ede01STejun Heo /* 1149bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1150bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1151bf4ede01STejun Heo */ 1152d565ed63STejun Heo pool = get_work_pool(work); 1153d565ed63STejun Heo if (!pool) 1154bbb68dfaSTejun Heo goto fail; 1155bf4ede01STejun Heo 1156d565ed63STejun Heo spin_lock(&pool->lock); 1157bf4ede01STejun Heo /* 1158112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1159112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1160112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1161112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1162112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11630b3dae68SLai Jiangshan * item is currently queued on that pool. 1164bf4ede01STejun Heo */ 1165112202d9STejun Heo pwq = get_work_pwq(work); 1166112202d9STejun Heo if (pwq && pwq->pool == pool) { 1167bf4ede01STejun Heo debug_work_deactivate(work); 11683aa62497SLai Jiangshan 11693aa62497SLai Jiangshan /* 117016062836STejun Heo * A delayed work item cannot be grabbed directly because 117116062836STejun Heo * it might have linked NO_COLOR work items which, if left 1172112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 117316062836STejun Heo * management later on and cause stall. Make sure the work 117416062836STejun Heo * item is activated before grabbing. 11753aa62497SLai Jiangshan */ 11763aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1177112202d9STejun Heo pwq_activate_delayed_work(work); 11783aa62497SLai Jiangshan 1179bf4ede01STejun Heo list_del_init(&work->entry); 1180112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 118136e227d2STejun Heo 1182112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11834468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11844468a00fSLai Jiangshan 1185d565ed63STejun Heo spin_unlock(&pool->lock); 118636e227d2STejun Heo return 1; 1187bf4ede01STejun Heo } 1188d565ed63STejun Heo spin_unlock(&pool->lock); 1189bbb68dfaSTejun Heo fail: 1190bbb68dfaSTejun Heo local_irq_restore(*flags); 1191bbb68dfaSTejun Heo if (work_is_canceling(work)) 1192bbb68dfaSTejun Heo return -ENOENT; 1193bbb68dfaSTejun Heo cpu_relax(); 119436e227d2STejun Heo return -EAGAIN; 1195bf4ede01STejun Heo } 1196bf4ede01STejun Heo 1197bf4ede01STejun Heo /** 1198706026c2STejun Heo * insert_work - insert a work into a pool 1199112202d9STejun Heo * @pwq: pwq @work belongs to 12004690c4abSTejun Heo * @work: work to insert 12014690c4abSTejun Heo * @head: insertion point 12024690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12034690c4abSTejun Heo * 1204112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1205706026c2STejun Heo * work_struct flags. 12064690c4abSTejun Heo * 12074690c4abSTejun Heo * CONTEXT: 1208d565ed63STejun Heo * spin_lock_irq(pool->lock). 1209365970a1SDavid Howells */ 1210112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1211112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1212b89deed3SOleg Nesterov { 1213112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1214e1d8aa9fSFrederic Weisbecker 12154690c4abSTejun Heo /* we own @work, set data and link */ 1216112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12171a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12188864b4e5STejun Heo get_pwq(pwq); 1219e22bee78STejun Heo 1220e22bee78STejun Heo /* 1221c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1222c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1223c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1224e22bee78STejun Heo */ 1225e22bee78STejun Heo smp_mb(); 1226e22bee78STejun Heo 122763d95a91STejun Heo if (__need_more_worker(pool)) 122863d95a91STejun Heo wake_up_worker(pool); 1229b89deed3SOleg Nesterov } 1230b89deed3SOleg Nesterov 1231c8efcc25STejun Heo /* 1232c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12338d03ecfeSTejun Heo * same workqueue. 1234c8efcc25STejun Heo */ 1235c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1236c8efcc25STejun Heo { 1237c8efcc25STejun Heo struct worker *worker; 1238c8efcc25STejun Heo 12398d03ecfeSTejun Heo worker = current_wq_worker(); 1240c8efcc25STejun Heo /* 12418d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12428d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1243c8efcc25STejun Heo */ 1244112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1245c8efcc25STejun Heo } 1246c8efcc25STejun Heo 1247d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12481da177e4SLinus Torvalds struct work_struct *work) 12491da177e4SLinus Torvalds { 1250112202d9STejun Heo struct pool_workqueue *pwq; 1251c9178087STejun Heo struct worker_pool *last_pool; 12521e19ffc6STejun Heo struct list_head *worklist; 12538a2e8e5dSTejun Heo unsigned int work_flags; 1254b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12558930cabaSTejun Heo 12568930cabaSTejun Heo /* 12578930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12588930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12598930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12608930cabaSTejun Heo * happen with IRQ disabled. 12618930cabaSTejun Heo */ 12628930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12631da177e4SLinus Torvalds 1264dc186ad7SThomas Gleixner debug_work_activate(work); 12651e19ffc6STejun Heo 1266c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1267618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1268c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1269e41e704bSTejun Heo return; 12709e8cd2f5STejun Heo retry: 1271c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1272c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 127357469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1274f3421797STejun Heo cpu = raw_smp_processor_id(); 1275c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1276c9178087STejun Heo } else { 1277c9178087STejun Heo pwq = first_pwq(wq); 1278c9178087STejun Heo } 1279f3421797STejun Heo 128018aa9effSTejun Heo /* 1281c9178087STejun Heo * If @work was previously on a different pool, it might still be 1282c9178087STejun Heo * running there, in which case the work needs to be queued on that 1283c9178087STejun Heo * pool to guarantee non-reentrancy. 128418aa9effSTejun Heo */ 1285c9e7cf27STejun Heo last_pool = get_work_pool(work); 1286112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 128718aa9effSTejun Heo struct worker *worker; 128818aa9effSTejun Heo 1289d565ed63STejun Heo spin_lock(&last_pool->lock); 129018aa9effSTejun Heo 1291c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 129218aa9effSTejun Heo 1293112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1294c9178087STejun Heo pwq = worker->current_pwq; 12958594fadeSLai Jiangshan } else { 129618aa9effSTejun Heo /* meh... not running there, queue here */ 1297d565ed63STejun Heo spin_unlock(&last_pool->lock); 1298112202d9STejun Heo spin_lock(&pwq->pool->lock); 129918aa9effSTejun Heo } 13008930cabaSTejun Heo } else { 1301112202d9STejun Heo spin_lock(&pwq->pool->lock); 13028930cabaSTejun Heo } 1303502ca9d8STejun Heo 13049e8cd2f5STejun Heo /* 13059e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13069e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13079e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 13089e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 13099e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 13109e8cd2f5STejun Heo * make forward-progress. 13119e8cd2f5STejun Heo */ 13129e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13139e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13149e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13159e8cd2f5STejun Heo cpu_relax(); 13169e8cd2f5STejun Heo goto retry; 13179e8cd2f5STejun Heo } 13189e8cd2f5STejun Heo /* oops */ 13199e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13209e8cd2f5STejun Heo wq->name, cpu); 13219e8cd2f5STejun Heo } 13229e8cd2f5STejun Heo 1323112202d9STejun Heo /* pwq determined, queue */ 1324112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1325502ca9d8STejun Heo 1326f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1327112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1328f5b2552bSDan Carpenter return; 1329f5b2552bSDan Carpenter } 13301e19ffc6STejun Heo 1331112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1332112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13331e19ffc6STejun Heo 1334112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1335cdadf009STejun Heo trace_workqueue_activate_work(work); 1336112202d9STejun Heo pwq->nr_active++; 1337112202d9STejun Heo worklist = &pwq->pool->worklist; 13388a2e8e5dSTejun Heo } else { 13398a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1340112202d9STejun Heo worklist = &pwq->delayed_works; 13418a2e8e5dSTejun Heo } 13421e19ffc6STejun Heo 1343112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13441e19ffc6STejun Heo 1345112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13461da177e4SLinus Torvalds } 13471da177e4SLinus Torvalds 13480fcb78c2SRolf Eike Beer /** 1349c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1350c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1351c1a220e7SZhang Rui * @wq: workqueue to use 1352c1a220e7SZhang Rui * @work: work to queue 1353c1a220e7SZhang Rui * 1354d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1355c1a220e7SZhang Rui * 1356c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1357c1a220e7SZhang Rui * can't go away. 1358c1a220e7SZhang Rui */ 1359d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1360d4283e93STejun Heo struct work_struct *work) 1361c1a220e7SZhang Rui { 1362d4283e93STejun Heo bool ret = false; 13638930cabaSTejun Heo unsigned long flags; 13648930cabaSTejun Heo 13658930cabaSTejun Heo local_irq_save(flags); 1366c1a220e7SZhang Rui 136722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13684690c4abSTejun Heo __queue_work(cpu, wq, work); 1369d4283e93STejun Heo ret = true; 1370c1a220e7SZhang Rui } 13718930cabaSTejun Heo 13728930cabaSTejun Heo local_irq_restore(flags); 1373c1a220e7SZhang Rui return ret; 1374c1a220e7SZhang Rui } 1375c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1376c1a220e7SZhang Rui 1377d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13781da177e4SLinus Torvalds { 137952bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13801da177e4SLinus Torvalds 1381e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 138260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13831da177e4SLinus Torvalds } 13841438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13851da177e4SLinus Torvalds 13867beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 138752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13881da177e4SLinus Torvalds { 13897beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13907beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13911da177e4SLinus Torvalds 13927beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13937beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1394fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1395fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13967beb2edfSTejun Heo 13978852aac2STejun Heo /* 13988852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13998852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14008852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14018852aac2STejun Heo * on that there's no such delay when @delay is 0. 14028852aac2STejun Heo */ 14038852aac2STejun Heo if (!delay) { 14048852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14058852aac2STejun Heo return; 14068852aac2STejun Heo } 14078852aac2STejun Heo 14087beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14097beb2edfSTejun Heo 141060c057bcSLai Jiangshan dwork->wq = wq; 14111265057fSTejun Heo dwork->cpu = cpu; 14127beb2edfSTejun Heo timer->expires = jiffies + delay; 14137beb2edfSTejun Heo 14147beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14157beb2edfSTejun Heo add_timer_on(timer, cpu); 14167beb2edfSTejun Heo else 14177beb2edfSTejun Heo add_timer(timer); 14187beb2edfSTejun Heo } 14191da177e4SLinus Torvalds 14200fcb78c2SRolf Eike Beer /** 14210fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14220fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14230fcb78c2SRolf Eike Beer * @wq: workqueue to use 1424af9997e4SRandy Dunlap * @dwork: work to queue 14250fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14260fcb78c2SRolf Eike Beer * 1427715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1428715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1429715f1300STejun Heo * execution. 14300fcb78c2SRolf Eike Beer */ 1431d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 143252bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14337a6bc1cdSVenkatesh Pallipadi { 143452bad64dSDavid Howells struct work_struct *work = &dwork->work; 1435d4283e93STejun Heo bool ret = false; 14368930cabaSTejun Heo unsigned long flags; 14378930cabaSTejun Heo 14388930cabaSTejun Heo /* read the comment in __queue_work() */ 14398930cabaSTejun Heo local_irq_save(flags); 14407a6bc1cdSVenkatesh Pallipadi 144122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14427beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1443d4283e93STejun Heo ret = true; 14447a6bc1cdSVenkatesh Pallipadi } 14458930cabaSTejun Heo 14468930cabaSTejun Heo local_irq_restore(flags); 14477a6bc1cdSVenkatesh Pallipadi return ret; 14487a6bc1cdSVenkatesh Pallipadi } 1449ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14501da177e4SLinus Torvalds 1451c8e55f36STejun Heo /** 14528376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14538376fe22STejun Heo * @cpu: CPU number to execute work on 14548376fe22STejun Heo * @wq: workqueue to use 14558376fe22STejun Heo * @dwork: work to queue 14568376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14578376fe22STejun Heo * 14588376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14598376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14608376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14618376fe22STejun Heo * current state. 14628376fe22STejun Heo * 14638376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14648376fe22STejun Heo * pending and its timer was modified. 14658376fe22STejun Heo * 1466e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14678376fe22STejun Heo * See try_to_grab_pending() for details. 14688376fe22STejun Heo */ 14698376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14708376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14718376fe22STejun Heo { 14728376fe22STejun Heo unsigned long flags; 14738376fe22STejun Heo int ret; 14748376fe22STejun Heo 14758376fe22STejun Heo do { 14768376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14778376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14788376fe22STejun Heo 14798376fe22STejun Heo if (likely(ret >= 0)) { 14808376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14818376fe22STejun Heo local_irq_restore(flags); 14828376fe22STejun Heo } 14838376fe22STejun Heo 14848376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14858376fe22STejun Heo return ret; 14868376fe22STejun Heo } 14878376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14888376fe22STejun Heo 14898376fe22STejun Heo /** 1490c8e55f36STejun Heo * worker_enter_idle - enter idle state 1491c8e55f36STejun Heo * @worker: worker which is entering idle state 1492c8e55f36STejun Heo * 1493c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1494c8e55f36STejun Heo * necessary. 1495c8e55f36STejun Heo * 1496c8e55f36STejun Heo * LOCKING: 1497d565ed63STejun Heo * spin_lock_irq(pool->lock). 1498c8e55f36STejun Heo */ 1499c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15001da177e4SLinus Torvalds { 1501bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1502c8e55f36STejun Heo 15036183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15046183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15056183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15066183c009STejun Heo return; 1507c8e55f36STejun Heo 1508cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1509cb444766STejun Heo worker->flags |= WORKER_IDLE; 1510bd7bdd43STejun Heo pool->nr_idle++; 1511e22bee78STejun Heo worker->last_active = jiffies; 1512c8e55f36STejun Heo 1513c8e55f36STejun Heo /* idle_list is LIFO */ 1514bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1515db7bccf4STejun Heo 151663d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1517628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1518cb444766STejun Heo 1519544ecf31STejun Heo /* 1520706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1521d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1522628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1523628c78e7STejun Heo * unbind is not in progress. 1524544ecf31STejun Heo */ 152524647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1526bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1527e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1528c8e55f36STejun Heo } 1529c8e55f36STejun Heo 1530c8e55f36STejun Heo /** 1531c8e55f36STejun Heo * worker_leave_idle - leave idle state 1532c8e55f36STejun Heo * @worker: worker which is leaving idle state 1533c8e55f36STejun Heo * 1534c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1535c8e55f36STejun Heo * 1536c8e55f36STejun Heo * LOCKING: 1537d565ed63STejun Heo * spin_lock_irq(pool->lock). 1538c8e55f36STejun Heo */ 1539c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1540c8e55f36STejun Heo { 1541bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1542c8e55f36STejun Heo 15436183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15446183c009STejun Heo return; 1545d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1546bd7bdd43STejun Heo pool->nr_idle--; 1547c8e55f36STejun Heo list_del_init(&worker->entry); 1548c8e55f36STejun Heo } 1549c8e55f36STejun Heo 1550e22bee78STejun Heo /** 1551f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1552f36dc67bSLai Jiangshan * @pool: target worker_pool 1553f36dc67bSLai Jiangshan * 1554f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1555e22bee78STejun Heo * 1556e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1557e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1558e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1559e22bee78STejun Heo * guaranteed to execute on the cpu. 1560e22bee78STejun Heo * 1561f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1562e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1563e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1564e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1565706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1566e22bee78STejun Heo * [dis]associated in the meantime. 1567e22bee78STejun Heo * 1568706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 156924647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1570f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1571f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1572f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1573e22bee78STejun Heo * 1574e22bee78STejun Heo * CONTEXT: 1575d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1576e22bee78STejun Heo * held. 1577e22bee78STejun Heo * 1578e22bee78STejun Heo * RETURNS: 1579706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1580e22bee78STejun Heo * bound), %false if offline. 1581e22bee78STejun Heo */ 1582f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1583d565ed63STejun Heo __acquires(&pool->lock) 1584e22bee78STejun Heo { 1585e22bee78STejun Heo while (true) { 1586e22bee78STejun Heo /* 1587e22bee78STejun Heo * The following call may fail, succeed or succeed 1588e22bee78STejun Heo * without actually migrating the task to the cpu if 1589e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 159024647570STejun Heo * against POOL_DISASSOCIATED. 1591e22bee78STejun Heo */ 159224647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 15937a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1594e22bee78STejun Heo 1595d565ed63STejun Heo spin_lock_irq(&pool->lock); 159624647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1597e22bee78STejun Heo return false; 1598f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 15997a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1600e22bee78STejun Heo return true; 1601d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1602e22bee78STejun Heo 16035035b20fSTejun Heo /* 16045035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16055035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16065035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16075035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16085035b20fSTejun Heo */ 1609e22bee78STejun Heo cpu_relax(); 16105035b20fSTejun Heo cond_resched(); 1611e22bee78STejun Heo } 1612e22bee78STejun Heo } 1613e22bee78STejun Heo 1614c34056a3STejun Heo static struct worker *alloc_worker(void) 1615c34056a3STejun Heo { 1616c34056a3STejun Heo struct worker *worker; 1617c34056a3STejun Heo 1618c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1619c8e55f36STejun Heo if (worker) { 1620c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1621affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1622e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1623e22bee78STejun Heo worker->flags = WORKER_PREP; 1624c8e55f36STejun Heo } 1625c34056a3STejun Heo return worker; 1626c34056a3STejun Heo } 1627c34056a3STejun Heo 1628c34056a3STejun Heo /** 1629c34056a3STejun Heo * create_worker - create a new workqueue worker 163063d95a91STejun Heo * @pool: pool the new worker will belong to 1631c34056a3STejun Heo * 163263d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1633c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1634c34056a3STejun Heo * destroy_worker(). 1635c34056a3STejun Heo * 1636c34056a3STejun Heo * CONTEXT: 1637c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1638c34056a3STejun Heo * 1639c34056a3STejun Heo * RETURNS: 1640c34056a3STejun Heo * Pointer to the newly created worker. 1641c34056a3STejun Heo */ 1642bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1643c34056a3STejun Heo { 16447a4e344cSTejun Heo const char *pri = pool->attrs->nice < 0 ? "H" : ""; 1645c34056a3STejun Heo struct worker *worker = NULL; 1646f3421797STejun Heo int id = -1; 1647c34056a3STejun Heo 1648cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1649cd549687STejun Heo 1650822d8405STejun Heo /* 1651822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1652822d8405STejun Heo * without installing the pointer. 1653822d8405STejun Heo */ 1654822d8405STejun Heo idr_preload(GFP_KERNEL); 1655d565ed63STejun Heo spin_lock_irq(&pool->lock); 1656822d8405STejun Heo 1657822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1658822d8405STejun Heo 1659d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1660822d8405STejun Heo idr_preload_end(); 1661822d8405STejun Heo if (id < 0) 1662c34056a3STejun Heo goto fail; 1663c34056a3STejun Heo 1664c34056a3STejun Heo worker = alloc_worker(); 1665c34056a3STejun Heo if (!worker) 1666c34056a3STejun Heo goto fail; 1667c34056a3STejun Heo 1668bd7bdd43STejun Heo worker->pool = pool; 1669c34056a3STejun Heo worker->id = id; 1670c34056a3STejun Heo 167129c91e99STejun Heo if (pool->cpu >= 0) 167294dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1673ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1674d84ff051STejun Heo "kworker/%d:%d%s", pool->cpu, id, pri); 1675f3421797STejun Heo else 1676f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1677ac6104cdSTejun Heo "kworker/u%d:%d%s", 1678ac6104cdSTejun Heo pool->id, id, pri); 1679c34056a3STejun Heo if (IS_ERR(worker->task)) 1680c34056a3STejun Heo goto fail; 1681c34056a3STejun Heo 1682c5aa87bbSTejun Heo /* 1683c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1684c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1685c5aa87bbSTejun Heo */ 16867a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 16877a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16883270476aSTejun Heo 168914a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 169014a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 16917a4e344cSTejun Heo 16927a4e344cSTejun Heo /* 16937a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 16947a4e344cSTejun Heo * remains stable across this function. See the comments above the 16957a4e344cSTejun Heo * flag definition for details. 16967a4e344cSTejun Heo */ 16977a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1698f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1699c34056a3STejun Heo 1700822d8405STejun Heo /* successful, commit the pointer to idr */ 1701822d8405STejun Heo spin_lock_irq(&pool->lock); 1702822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1703822d8405STejun Heo spin_unlock_irq(&pool->lock); 1704822d8405STejun Heo 1705c34056a3STejun Heo return worker; 1706822d8405STejun Heo 1707c34056a3STejun Heo fail: 1708c34056a3STejun Heo if (id >= 0) { 1709d565ed63STejun Heo spin_lock_irq(&pool->lock); 1710822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1711d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1712c34056a3STejun Heo } 1713c34056a3STejun Heo kfree(worker); 1714c34056a3STejun Heo return NULL; 1715c34056a3STejun Heo } 1716c34056a3STejun Heo 1717c34056a3STejun Heo /** 1718c34056a3STejun Heo * start_worker - start a newly created worker 1719c34056a3STejun Heo * @worker: worker to start 1720c34056a3STejun Heo * 1721706026c2STejun Heo * Make the pool aware of @worker and start it. 1722c34056a3STejun Heo * 1723c34056a3STejun Heo * CONTEXT: 1724d565ed63STejun Heo * spin_lock_irq(pool->lock). 1725c34056a3STejun Heo */ 1726c34056a3STejun Heo static void start_worker(struct worker *worker) 1727c34056a3STejun Heo { 1728cb444766STejun Heo worker->flags |= WORKER_STARTED; 1729bd7bdd43STejun Heo worker->pool->nr_workers++; 1730c8e55f36STejun Heo worker_enter_idle(worker); 1731c34056a3STejun Heo wake_up_process(worker->task); 1732c34056a3STejun Heo } 1733c34056a3STejun Heo 1734c34056a3STejun Heo /** 1735ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1736ebf44d16STejun Heo * @pool: the target pool 1737ebf44d16STejun Heo * 1738cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1739ebf44d16STejun Heo */ 1740ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1741ebf44d16STejun Heo { 1742ebf44d16STejun Heo struct worker *worker; 1743ebf44d16STejun Heo 1744cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1745cd549687STejun Heo 1746ebf44d16STejun Heo worker = create_worker(pool); 1747ebf44d16STejun Heo if (worker) { 1748ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1749ebf44d16STejun Heo start_worker(worker); 1750ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1751ebf44d16STejun Heo } 1752ebf44d16STejun Heo 1753cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1754cd549687STejun Heo 1755ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1756ebf44d16STejun Heo } 1757ebf44d16STejun Heo 1758ebf44d16STejun Heo /** 1759c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1760c34056a3STejun Heo * @worker: worker to be destroyed 1761c34056a3STejun Heo * 1762706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1763c8e55f36STejun Heo * 1764c8e55f36STejun Heo * CONTEXT: 1765d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1766c34056a3STejun Heo */ 1767c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1768c34056a3STejun Heo { 1769bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1770c34056a3STejun Heo 1771cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1772cd549687STejun Heo lockdep_assert_held(&pool->lock); 1773cd549687STejun Heo 1774c34056a3STejun Heo /* sanity check frenzy */ 17756183c009STejun Heo if (WARN_ON(worker->current_work) || 17766183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 17776183c009STejun Heo return; 1778c34056a3STejun Heo 1779c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1780bd7bdd43STejun Heo pool->nr_workers--; 1781c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1782bd7bdd43STejun Heo pool->nr_idle--; 1783c8e55f36STejun Heo 1784c8e55f36STejun Heo list_del_init(&worker->entry); 1785cb444766STejun Heo worker->flags |= WORKER_DIE; 1786c8e55f36STejun Heo 1787822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1788822d8405STejun Heo 1789d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1790c8e55f36STejun Heo 1791c34056a3STejun Heo kthread_stop(worker->task); 1792c34056a3STejun Heo kfree(worker); 1793c34056a3STejun Heo 1794d565ed63STejun Heo spin_lock_irq(&pool->lock); 1795c34056a3STejun Heo } 1796c34056a3STejun Heo 179763d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1798e22bee78STejun Heo { 179963d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1800e22bee78STejun Heo 1801d565ed63STejun Heo spin_lock_irq(&pool->lock); 1802e22bee78STejun Heo 180363d95a91STejun Heo if (too_many_workers(pool)) { 1804e22bee78STejun Heo struct worker *worker; 1805e22bee78STejun Heo unsigned long expires; 1806e22bee78STejun Heo 1807e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 180863d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1809e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1810e22bee78STejun Heo 1811e22bee78STejun Heo if (time_before(jiffies, expires)) 181263d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1813e22bee78STejun Heo else { 1814e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 181511ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 181663d95a91STejun Heo wake_up_worker(pool); 1817e22bee78STejun Heo } 1818e22bee78STejun Heo } 1819e22bee78STejun Heo 1820d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1821e22bee78STejun Heo } 1822e22bee78STejun Heo 1823493a1724STejun Heo static void send_mayday(struct work_struct *work) 1824e22bee78STejun Heo { 1825112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1826112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1827493a1724STejun Heo 18282e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1829e22bee78STejun Heo 1830493008a8STejun Heo if (!wq->rescuer) 1831493a1724STejun Heo return; 1832e22bee78STejun Heo 1833e22bee78STejun Heo /* mayday mayday mayday */ 1834493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1835493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1836e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1837493a1724STejun Heo } 1838e22bee78STejun Heo } 1839e22bee78STejun Heo 1840706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1841e22bee78STejun Heo { 184263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1843e22bee78STejun Heo struct work_struct *work; 1844e22bee78STejun Heo 18452e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1846493a1724STejun Heo spin_lock(&pool->lock); 1847e22bee78STejun Heo 184863d95a91STejun Heo if (need_to_create_worker(pool)) { 1849e22bee78STejun Heo /* 1850e22bee78STejun Heo * We've been trying to create a new worker but 1851e22bee78STejun Heo * haven't been successful. We might be hitting an 1852e22bee78STejun Heo * allocation deadlock. Send distress signals to 1853e22bee78STejun Heo * rescuers. 1854e22bee78STejun Heo */ 185563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1856e22bee78STejun Heo send_mayday(work); 1857e22bee78STejun Heo } 1858e22bee78STejun Heo 1859493a1724STejun Heo spin_unlock(&pool->lock); 18602e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1861e22bee78STejun Heo 186263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1863e22bee78STejun Heo } 1864e22bee78STejun Heo 1865e22bee78STejun Heo /** 1866e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 186763d95a91STejun Heo * @pool: pool to create a new worker for 1868e22bee78STejun Heo * 186963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1870e22bee78STejun Heo * have at least one idle worker on return from this function. If 1871e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 187263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1873e22bee78STejun Heo * possible allocation deadlock. 1874e22bee78STejun Heo * 1875c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1876c5aa87bbSTejun Heo * may_start_working() %true. 1877e22bee78STejun Heo * 1878e22bee78STejun Heo * LOCKING: 1879d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1880e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1881e22bee78STejun Heo * manager. 1882e22bee78STejun Heo * 1883e22bee78STejun Heo * RETURNS: 1884c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1885e22bee78STejun Heo * otherwise. 1886e22bee78STejun Heo */ 188763d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1888d565ed63STejun Heo __releases(&pool->lock) 1889d565ed63STejun Heo __acquires(&pool->lock) 1890e22bee78STejun Heo { 189163d95a91STejun Heo if (!need_to_create_worker(pool)) 1892e22bee78STejun Heo return false; 1893e22bee78STejun Heo restart: 1894d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18959f9c2364STejun Heo 1896e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 189763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1898e22bee78STejun Heo 1899e22bee78STejun Heo while (true) { 1900e22bee78STejun Heo struct worker *worker; 1901e22bee78STejun Heo 1902bc2ae0f5STejun Heo worker = create_worker(pool); 1903e22bee78STejun Heo if (worker) { 190463d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1905d565ed63STejun Heo spin_lock_irq(&pool->lock); 1906e22bee78STejun Heo start_worker(worker); 19076183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19086183c009STejun Heo goto restart; 1909e22bee78STejun Heo return true; 1910e22bee78STejun Heo } 1911e22bee78STejun Heo 191263d95a91STejun Heo if (!need_to_create_worker(pool)) 1913e22bee78STejun Heo break; 1914e22bee78STejun Heo 1915e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1916e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19179f9c2364STejun Heo 191863d95a91STejun Heo if (!need_to_create_worker(pool)) 1919e22bee78STejun Heo break; 1920e22bee78STejun Heo } 1921e22bee78STejun Heo 192263d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1923d565ed63STejun Heo spin_lock_irq(&pool->lock); 192463d95a91STejun Heo if (need_to_create_worker(pool)) 1925e22bee78STejun Heo goto restart; 1926e22bee78STejun Heo return true; 1927e22bee78STejun Heo } 1928e22bee78STejun Heo 1929e22bee78STejun Heo /** 1930e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 193163d95a91STejun Heo * @pool: pool to destroy workers for 1932e22bee78STejun Heo * 193363d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1934e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1935e22bee78STejun Heo * 1936e22bee78STejun Heo * LOCKING: 1937d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1938e22bee78STejun Heo * multiple times. Called only from manager. 1939e22bee78STejun Heo * 1940e22bee78STejun Heo * RETURNS: 1941c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1942e22bee78STejun Heo * otherwise. 1943e22bee78STejun Heo */ 194463d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1945e22bee78STejun Heo { 1946e22bee78STejun Heo bool ret = false; 1947e22bee78STejun Heo 194863d95a91STejun Heo while (too_many_workers(pool)) { 1949e22bee78STejun Heo struct worker *worker; 1950e22bee78STejun Heo unsigned long expires; 1951e22bee78STejun Heo 195263d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1953e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1954e22bee78STejun Heo 1955e22bee78STejun Heo if (time_before(jiffies, expires)) { 195663d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1957e22bee78STejun Heo break; 1958e22bee78STejun Heo } 1959e22bee78STejun Heo 1960e22bee78STejun Heo destroy_worker(worker); 1961e22bee78STejun Heo ret = true; 1962e22bee78STejun Heo } 1963e22bee78STejun Heo 1964e22bee78STejun Heo return ret; 1965e22bee78STejun Heo } 1966e22bee78STejun Heo 1967e22bee78STejun Heo /** 1968e22bee78STejun Heo * manage_workers - manage worker pool 1969e22bee78STejun Heo * @worker: self 1970e22bee78STejun Heo * 1971706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1972e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1973706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1974e22bee78STejun Heo * 1975e22bee78STejun Heo * The caller can safely start processing works on false return. On 1976e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1977e22bee78STejun Heo * and may_start_working() is true. 1978e22bee78STejun Heo * 1979e22bee78STejun Heo * CONTEXT: 1980d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1981e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1982e22bee78STejun Heo * 1983e22bee78STejun Heo * RETURNS: 1984d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1985d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 1986e22bee78STejun Heo */ 1987e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1988e22bee78STejun Heo { 198963d95a91STejun Heo struct worker_pool *pool = worker->pool; 1990e22bee78STejun Heo bool ret = false; 1991e22bee78STejun Heo 1992bc3a1afcSTejun Heo /* 1993bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 1994bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 1995bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1996bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1997bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1998bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1999bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2000bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2001bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2002bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2003bc3a1afcSTejun Heo * 2004bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2005bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2006bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2007bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2008bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2009bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2010bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2011bc3a1afcSTejun Heo * manager_mutex. 2012bc3a1afcSTejun Heo */ 201334a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2014e22bee78STejun Heo return ret; 2015e22bee78STejun Heo 2016ee378aa4SLai Jiangshan /* 2017bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2018bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2019ee378aa4SLai Jiangshan */ 2020bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2021d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2022bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2023ee378aa4SLai Jiangshan ret = true; 2024ee378aa4SLai Jiangshan } 2025ee378aa4SLai Jiangshan 202611ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2027e22bee78STejun Heo 2028e22bee78STejun Heo /* 2029e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2030e22bee78STejun Heo * on return. 2031e22bee78STejun Heo */ 203263d95a91STejun Heo ret |= maybe_destroy_workers(pool); 203363d95a91STejun Heo ret |= maybe_create_worker(pool); 2034e22bee78STejun Heo 2035bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 203634a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2037e22bee78STejun Heo return ret; 2038e22bee78STejun Heo } 2039e22bee78STejun Heo 2040a62428c0STejun Heo /** 2041a62428c0STejun Heo * process_one_work - process single work 2042c34056a3STejun Heo * @worker: self 2043a62428c0STejun Heo * @work: work to process 2044a62428c0STejun Heo * 2045a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2046a62428c0STejun Heo * process a single work including synchronization against and 2047a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2048a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2049a62428c0STejun Heo * call this function to process a work. 2050a62428c0STejun Heo * 2051a62428c0STejun Heo * CONTEXT: 2052d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2053a62428c0STejun Heo */ 2054c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2055d565ed63STejun Heo __releases(&pool->lock) 2056d565ed63STejun Heo __acquires(&pool->lock) 20571da177e4SLinus Torvalds { 2058112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2059bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2060112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 206173f53c4aSTejun Heo int work_color; 20627e11629dSTejun Heo struct worker *collision; 20634e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20644e6045f1SJohannes Berg /* 2065a62428c0STejun Heo * It is permissible to free the struct work_struct from 2066a62428c0STejun Heo * inside the function that is called from it, this we need to 2067a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2068a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2069a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20704e6045f1SJohannes Berg */ 20714d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20724d82a1deSPeter Zijlstra 20734d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20744e6045f1SJohannes Berg #endif 20756fec10a1STejun Heo /* 20766fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20776fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 207824647570STejun Heo * unbound or a disassociated pool. 20796fec10a1STejun Heo */ 20805f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 208124647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2082ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 208325511a47STejun Heo 20847e11629dSTejun Heo /* 20857e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20867e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20877e11629dSTejun Heo * already processing the work. If so, defer the work to the 20887e11629dSTejun Heo * currently executing one. 20897e11629dSTejun Heo */ 2090c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20917e11629dSTejun Heo if (unlikely(collision)) { 20927e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20937e11629dSTejun Heo return; 20947e11629dSTejun Heo } 20951da177e4SLinus Torvalds 20968930cabaSTejun Heo /* claim and dequeue */ 20971da177e4SLinus Torvalds debug_work_deactivate(work); 2098c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2099c34056a3STejun Heo worker->current_work = work; 2100a2c1c57bSTejun Heo worker->current_func = work->func; 2101112202d9STejun Heo worker->current_pwq = pwq; 210273f53c4aSTejun Heo work_color = get_work_color(work); 21037a22ad75STejun Heo 2104a62428c0STejun Heo list_del_init(&work->entry); 2105a62428c0STejun Heo 2106649027d7STejun Heo /* 2107fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2108fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2109fb0e7bebSTejun Heo */ 2110fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2111fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2112fb0e7bebSTejun Heo 2113974271c4STejun Heo /* 2114d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2115974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2116974271c4STejun Heo */ 211763d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 211863d95a91STejun Heo wake_up_worker(pool); 2119974271c4STejun Heo 21208930cabaSTejun Heo /* 21217c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2122d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 212323657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 212423657bb1STejun Heo * disabled. 21258930cabaSTejun Heo */ 21267c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21271da177e4SLinus Torvalds 2128d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2129365970a1SDavid Howells 2130112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21313295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2132e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2133a2c1c57bSTejun Heo worker->current_func(work); 2134e36c886aSArjan van de Ven /* 2135e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2136e36c886aSArjan van de Ven * point will only record its address. 2137e36c886aSArjan van de Ven */ 2138e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21393295f0efSIngo Molnar lock_map_release(&lockdep_map); 2140112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21411da177e4SLinus Torvalds 2142d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2143044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2144044c782cSValentin Ilie " last function: %pf\n", 2145a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2146a2c1c57bSTejun Heo worker->current_func); 2147d5abe669SPeter Zijlstra debug_show_held_locks(current); 2148d5abe669SPeter Zijlstra dump_stack(); 2149d5abe669SPeter Zijlstra } 2150d5abe669SPeter Zijlstra 2151d565ed63STejun Heo spin_lock_irq(&pool->lock); 2152a62428c0STejun Heo 2153fb0e7bebSTejun Heo /* clear cpu intensive status */ 2154fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2155fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2156fb0e7bebSTejun Heo 2157a62428c0STejun Heo /* we're done with it, release */ 215842f8570fSSasha Levin hash_del(&worker->hentry); 2159c34056a3STejun Heo worker->current_work = NULL; 2160a2c1c57bSTejun Heo worker->current_func = NULL; 2161112202d9STejun Heo worker->current_pwq = NULL; 2162112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21631da177e4SLinus Torvalds } 21641da177e4SLinus Torvalds 2165affee4b2STejun Heo /** 2166affee4b2STejun Heo * process_scheduled_works - process scheduled works 2167affee4b2STejun Heo * @worker: self 2168affee4b2STejun Heo * 2169affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2170affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2171affee4b2STejun Heo * fetches a work from the top and executes it. 2172affee4b2STejun Heo * 2173affee4b2STejun Heo * CONTEXT: 2174d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2175affee4b2STejun Heo * multiple times. 2176affee4b2STejun Heo */ 2177affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21781da177e4SLinus Torvalds { 2179affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2180affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2181a62428c0STejun Heo struct work_struct, entry); 2182c34056a3STejun Heo process_one_work(worker, work); 2183a62428c0STejun Heo } 21841da177e4SLinus Torvalds } 21851da177e4SLinus Torvalds 21864690c4abSTejun Heo /** 21874690c4abSTejun Heo * worker_thread - the worker thread function 2188c34056a3STejun Heo * @__worker: self 21894690c4abSTejun Heo * 2190c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2191c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2192c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2193c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2194c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 21954690c4abSTejun Heo */ 2196c34056a3STejun Heo static int worker_thread(void *__worker) 21971da177e4SLinus Torvalds { 2198c34056a3STejun Heo struct worker *worker = __worker; 2199bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22001da177e4SLinus Torvalds 2201e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2202e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2203c8e55f36STejun Heo woke_up: 2204d565ed63STejun Heo spin_lock_irq(&pool->lock); 2205affee4b2STejun Heo 2206a9ab775bSTejun Heo /* am I supposed to die? */ 2207a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2208d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2209a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2210e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2211c8e55f36STejun Heo return 0; 2212c8e55f36STejun Heo } 2213c8e55f36STejun Heo 2214c8e55f36STejun Heo worker_leave_idle(worker); 2215db7bccf4STejun Heo recheck: 2216e22bee78STejun Heo /* no more worker necessary? */ 221763d95a91STejun Heo if (!need_more_worker(pool)) 2218e22bee78STejun Heo goto sleep; 2219e22bee78STejun Heo 2220e22bee78STejun Heo /* do we need to manage? */ 222163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2222e22bee78STejun Heo goto recheck; 2223e22bee78STejun Heo 2224c8e55f36STejun Heo /* 2225c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2226c8e55f36STejun Heo * preparing to process a work or actually processing it. 2227c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2228c8e55f36STejun Heo */ 22296183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2230c8e55f36STejun Heo 2231e22bee78STejun Heo /* 2232a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2233a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2234a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2235a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2236a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2237e22bee78STejun Heo */ 2238a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2239e22bee78STejun Heo 2240e22bee78STejun Heo do { 2241affee4b2STejun Heo struct work_struct *work = 2242bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2243affee4b2STejun Heo struct work_struct, entry); 2244affee4b2STejun Heo 2245c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2246affee4b2STejun Heo /* optimization path, not strictly necessary */ 2247affee4b2STejun Heo process_one_work(worker, work); 2248affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2249affee4b2STejun Heo process_scheduled_works(worker); 2250affee4b2STejun Heo } else { 2251c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2252affee4b2STejun Heo process_scheduled_works(worker); 2253affee4b2STejun Heo } 225463d95a91STejun Heo } while (keep_working(pool)); 2255affee4b2STejun Heo 2256e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2257d313dd85STejun Heo sleep: 225863d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2259e22bee78STejun Heo goto recheck; 2260d313dd85STejun Heo 2261c8e55f36STejun Heo /* 2262d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2263d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2264d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2265d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2266d565ed63STejun Heo * event. 2267c8e55f36STejun Heo */ 2268c8e55f36STejun Heo worker_enter_idle(worker); 2269c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2270d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22711da177e4SLinus Torvalds schedule(); 2272c8e55f36STejun Heo goto woke_up; 22731da177e4SLinus Torvalds } 22741da177e4SLinus Torvalds 2275e22bee78STejun Heo /** 2276e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2277111c225aSTejun Heo * @__rescuer: self 2278e22bee78STejun Heo * 2279e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2280493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2281e22bee78STejun Heo * 2282706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2283e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2284e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2285e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2286e22bee78STejun Heo * the problem rescuer solves. 2287e22bee78STejun Heo * 2288706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2289706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2290e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2291e22bee78STejun Heo * 2292e22bee78STejun Heo * This should happen rarely. 2293e22bee78STejun Heo */ 2294111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2295e22bee78STejun Heo { 2296111c225aSTejun Heo struct worker *rescuer = __rescuer; 2297111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2298e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2299e22bee78STejun Heo 2300e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2301111c225aSTejun Heo 2302111c225aSTejun Heo /* 2303111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2304111c225aSTejun Heo * doesn't participate in concurrency management. 2305111c225aSTejun Heo */ 2306111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2307e22bee78STejun Heo repeat: 2308e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23091da177e4SLinus Torvalds 2310412d32e6SMike Galbraith if (kthread_should_stop()) { 2311412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2312111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2313e22bee78STejun Heo return 0; 2314412d32e6SMike Galbraith } 23151da177e4SLinus Torvalds 2316493a1724STejun Heo /* see whether any pwq is asking for help */ 23172e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2318493a1724STejun Heo 2319493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2320493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2321493a1724STejun Heo struct pool_workqueue, mayday_node); 2322112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2323e22bee78STejun Heo struct work_struct *work, *n; 2324e22bee78STejun Heo 2325e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2326493a1724STejun Heo list_del_init(&pwq->mayday_node); 2327493a1724STejun Heo 23282e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2329e22bee78STejun Heo 2330e22bee78STejun Heo /* migrate to the target cpu if possible */ 2331f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2332b3104104SLai Jiangshan rescuer->pool = pool; 2333e22bee78STejun Heo 2334e22bee78STejun Heo /* 2335e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2336e22bee78STejun Heo * process'em. 2337e22bee78STejun Heo */ 23386183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2339bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2340112202d9STejun Heo if (get_work_pwq(work) == pwq) 2341e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2342e22bee78STejun Heo 2343e22bee78STejun Heo process_scheduled_works(rescuer); 23447576958aSTejun Heo 23457576958aSTejun Heo /* 2346d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23477576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23487576958aSTejun Heo * and stalling the execution. 23497576958aSTejun Heo */ 235063d95a91STejun Heo if (keep_working(pool)) 235163d95a91STejun Heo wake_up_worker(pool); 23527576958aSTejun Heo 2353b3104104SLai Jiangshan rescuer->pool = NULL; 2354493a1724STejun Heo spin_unlock(&pool->lock); 23552e109a28STejun Heo spin_lock(&wq_mayday_lock); 23561da177e4SLinus Torvalds } 23571da177e4SLinus Torvalds 23582e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2359493a1724STejun Heo 2360111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2361111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2362e22bee78STejun Heo schedule(); 2363e22bee78STejun Heo goto repeat; 23641da177e4SLinus Torvalds } 23651da177e4SLinus Torvalds 2366fc2e4d70SOleg Nesterov struct wq_barrier { 2367fc2e4d70SOleg Nesterov struct work_struct work; 2368fc2e4d70SOleg Nesterov struct completion done; 2369fc2e4d70SOleg Nesterov }; 2370fc2e4d70SOleg Nesterov 2371fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2372fc2e4d70SOleg Nesterov { 2373fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2374fc2e4d70SOleg Nesterov complete(&barr->done); 2375fc2e4d70SOleg Nesterov } 2376fc2e4d70SOleg Nesterov 23774690c4abSTejun Heo /** 23784690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2379112202d9STejun Heo * @pwq: pwq to insert barrier into 23804690c4abSTejun Heo * @barr: wq_barrier to insert 2381affee4b2STejun Heo * @target: target work to attach @barr to 2382affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23834690c4abSTejun Heo * 2384affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2385affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2386affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2387affee4b2STejun Heo * cpu. 2388affee4b2STejun Heo * 2389affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2390affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2391affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2392affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2393affee4b2STejun Heo * after a work with LINKED flag set. 2394affee4b2STejun Heo * 2395affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2396112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23974690c4abSTejun Heo * 23984690c4abSTejun Heo * CONTEXT: 2399d565ed63STejun Heo * spin_lock_irq(pool->lock). 24004690c4abSTejun Heo */ 2401112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2402affee4b2STejun Heo struct wq_barrier *barr, 2403affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2404fc2e4d70SOleg Nesterov { 2405affee4b2STejun Heo struct list_head *head; 2406affee4b2STejun Heo unsigned int linked = 0; 2407affee4b2STejun Heo 2408dc186ad7SThomas Gleixner /* 2409d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2410dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2411dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2412dc186ad7SThomas Gleixner * might deadlock. 2413dc186ad7SThomas Gleixner */ 2414ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 241522df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2416fc2e4d70SOleg Nesterov init_completion(&barr->done); 241783c22520SOleg Nesterov 2418affee4b2STejun Heo /* 2419affee4b2STejun Heo * If @target is currently being executed, schedule the 2420affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2421affee4b2STejun Heo */ 2422affee4b2STejun Heo if (worker) 2423affee4b2STejun Heo head = worker->scheduled.next; 2424affee4b2STejun Heo else { 2425affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2426affee4b2STejun Heo 2427affee4b2STejun Heo head = target->entry.next; 2428affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2429affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2430affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2431affee4b2STejun Heo } 2432affee4b2STejun Heo 2433dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2434112202d9STejun Heo insert_work(pwq, &barr->work, head, 2435affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2436fc2e4d70SOleg Nesterov } 2437fc2e4d70SOleg Nesterov 243873f53c4aSTejun Heo /** 2439112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 244073f53c4aSTejun Heo * @wq: workqueue being flushed 244173f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 244273f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 244373f53c4aSTejun Heo * 2444112202d9STejun Heo * Prepare pwqs for workqueue flushing. 244573f53c4aSTejun Heo * 2446112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2447112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2448112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2449112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2450112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 245173f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 245273f53c4aSTejun Heo * 245373f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 245473f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 245573f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 245673f53c4aSTejun Heo * is returned. 245773f53c4aSTejun Heo * 2458112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 245973f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 246073f53c4aSTejun Heo * advanced to @work_color. 246173f53c4aSTejun Heo * 246273f53c4aSTejun Heo * CONTEXT: 246373f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 246473f53c4aSTejun Heo * 246573f53c4aSTejun Heo * RETURNS: 246673f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 246773f53c4aSTejun Heo * otherwise. 246873f53c4aSTejun Heo */ 2469112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 247073f53c4aSTejun Heo int flush_color, int work_color) 24711da177e4SLinus Torvalds { 247273f53c4aSTejun Heo bool wait = false; 247349e3cf44STejun Heo struct pool_workqueue *pwq; 24741da177e4SLinus Torvalds 247573f53c4aSTejun Heo if (flush_color >= 0) { 24766183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2477112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2478dc186ad7SThomas Gleixner } 247914441960SOleg Nesterov 248076af4d93STejun Heo local_irq_disable(); 248176af4d93STejun Heo 248249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2483112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24841da177e4SLinus Torvalds 248576af4d93STejun Heo spin_lock(&pool->lock); 248673f53c4aSTejun Heo 248773f53c4aSTejun Heo if (flush_color >= 0) { 24886183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 248973f53c4aSTejun Heo 2490112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2491112202d9STejun Heo pwq->flush_color = flush_color; 2492112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 249373f53c4aSTejun Heo wait = true; 24941da177e4SLinus Torvalds } 249573f53c4aSTejun Heo } 249673f53c4aSTejun Heo 249773f53c4aSTejun Heo if (work_color >= 0) { 24986183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2499112202d9STejun Heo pwq->work_color = work_color; 250073f53c4aSTejun Heo } 250173f53c4aSTejun Heo 250276af4d93STejun Heo spin_unlock(&pool->lock); 25031da177e4SLinus Torvalds } 25041da177e4SLinus Torvalds 250576af4d93STejun Heo local_irq_enable(); 250676af4d93STejun Heo 2507112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 250873f53c4aSTejun Heo complete(&wq->first_flusher->done); 250973f53c4aSTejun Heo 251073f53c4aSTejun Heo return wait; 251183c22520SOleg Nesterov } 25121da177e4SLinus Torvalds 25130fcb78c2SRolf Eike Beer /** 25141da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25150fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25161da177e4SLinus Torvalds * 2517c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2518c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25191da177e4SLinus Torvalds */ 25207ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25211da177e4SLinus Torvalds { 252273f53c4aSTejun Heo struct wq_flusher this_flusher = { 252373f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 252473f53c4aSTejun Heo .flush_color = -1, 252573f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 252673f53c4aSTejun Heo }; 252773f53c4aSTejun Heo int next_color; 2528b1f4ec17SOleg Nesterov 25293295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25303295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 253173f53c4aSTejun Heo 253273f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 253373f53c4aSTejun Heo 253473f53c4aSTejun Heo /* 253573f53c4aSTejun Heo * Start-to-wait phase 253673f53c4aSTejun Heo */ 253773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 253873f53c4aSTejun Heo 253973f53c4aSTejun Heo if (next_color != wq->flush_color) { 254073f53c4aSTejun Heo /* 254173f53c4aSTejun Heo * Color space is not full. The current work_color 254273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 254373f53c4aSTejun Heo * by one. 254473f53c4aSTejun Heo */ 25456183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 254673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 254773f53c4aSTejun Heo wq->work_color = next_color; 254873f53c4aSTejun Heo 254973f53c4aSTejun Heo if (!wq->first_flusher) { 255073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25516183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 255273f53c4aSTejun Heo 255373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 255473f53c4aSTejun Heo 2555112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 255673f53c4aSTejun Heo wq->work_color)) { 255773f53c4aSTejun Heo /* nothing to flush, done */ 255873f53c4aSTejun Heo wq->flush_color = next_color; 255973f53c4aSTejun Heo wq->first_flusher = NULL; 256073f53c4aSTejun Heo goto out_unlock; 256173f53c4aSTejun Heo } 256273f53c4aSTejun Heo } else { 256373f53c4aSTejun Heo /* wait in queue */ 25646183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 256573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2566112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 256773f53c4aSTejun Heo } 256873f53c4aSTejun Heo } else { 256973f53c4aSTejun Heo /* 257073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 257173f53c4aSTejun Heo * The next flush completion will assign us 257273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 257373f53c4aSTejun Heo */ 257473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 257573f53c4aSTejun Heo } 257673f53c4aSTejun Heo 257773f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 257873f53c4aSTejun Heo 257973f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 258073f53c4aSTejun Heo 258173f53c4aSTejun Heo /* 258273f53c4aSTejun Heo * Wake-up-and-cascade phase 258373f53c4aSTejun Heo * 258473f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 258573f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 258673f53c4aSTejun Heo */ 258773f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 258873f53c4aSTejun Heo return; 258973f53c4aSTejun Heo 259073f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 259173f53c4aSTejun Heo 25924ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25934ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25944ce48b37STejun Heo goto out_unlock; 25954ce48b37STejun Heo 259673f53c4aSTejun Heo wq->first_flusher = NULL; 259773f53c4aSTejun Heo 25986183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25996183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 260073f53c4aSTejun Heo 260173f53c4aSTejun Heo while (true) { 260273f53c4aSTejun Heo struct wq_flusher *next, *tmp; 260373f53c4aSTejun Heo 260473f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 260573f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 260673f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 260773f53c4aSTejun Heo break; 260873f53c4aSTejun Heo list_del_init(&next->list); 260973f53c4aSTejun Heo complete(&next->done); 261073f53c4aSTejun Heo } 261173f53c4aSTejun Heo 26126183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 261373f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 261473f53c4aSTejun Heo 261573f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 261673f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 261773f53c4aSTejun Heo 261873f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 261973f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 262073f53c4aSTejun Heo /* 262173f53c4aSTejun Heo * Assign the same color to all overflowed 262273f53c4aSTejun Heo * flushers, advance work_color and append to 262373f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 262473f53c4aSTejun Heo * phase for these overflowed flushers. 262573f53c4aSTejun Heo */ 262673f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 262773f53c4aSTejun Heo tmp->flush_color = wq->work_color; 262873f53c4aSTejun Heo 262973f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 263073f53c4aSTejun Heo 263173f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 263273f53c4aSTejun Heo &wq->flusher_queue); 2633112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 263473f53c4aSTejun Heo } 263573f53c4aSTejun Heo 263673f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26376183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 263873f53c4aSTejun Heo break; 263973f53c4aSTejun Heo } 264073f53c4aSTejun Heo 264173f53c4aSTejun Heo /* 264273f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2643112202d9STejun Heo * the new first flusher and arm pwqs. 264473f53c4aSTejun Heo */ 26456183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26466183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 264773f53c4aSTejun Heo 264873f53c4aSTejun Heo list_del_init(&next->list); 264973f53c4aSTejun Heo wq->first_flusher = next; 265073f53c4aSTejun Heo 2651112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 265273f53c4aSTejun Heo break; 265373f53c4aSTejun Heo 265473f53c4aSTejun Heo /* 265573f53c4aSTejun Heo * Meh... this color is already done, clear first 265673f53c4aSTejun Heo * flusher and repeat cascading. 265773f53c4aSTejun Heo */ 265873f53c4aSTejun Heo wq->first_flusher = NULL; 265973f53c4aSTejun Heo } 266073f53c4aSTejun Heo 266173f53c4aSTejun Heo out_unlock: 266273f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 26631da177e4SLinus Torvalds } 2664ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26651da177e4SLinus Torvalds 26669c5a2ba7STejun Heo /** 26679c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26689c5a2ba7STejun Heo * @wq: workqueue to drain 26699c5a2ba7STejun Heo * 26709c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26719c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26729c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26739c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26749c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26759c5a2ba7STejun Heo * takes too long. 26769c5a2ba7STejun Heo */ 26779c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26789c5a2ba7STejun Heo { 26799c5a2ba7STejun Heo unsigned int flush_cnt = 0; 268049e3cf44STejun Heo struct pool_workqueue *pwq; 26819c5a2ba7STejun Heo 26829c5a2ba7STejun Heo /* 26839c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26849c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2685618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26869c5a2ba7STejun Heo */ 268768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 26889c5a2ba7STejun Heo if (!wq->nr_drainers++) 2689618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 269068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 26919c5a2ba7STejun Heo reflush: 26929c5a2ba7STejun Heo flush_workqueue(wq); 26939c5a2ba7STejun Heo 269476af4d93STejun Heo local_irq_disable(); 269576af4d93STejun Heo 269649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2697fa2563e4SThomas Tuttle bool drained; 26989c5a2ba7STejun Heo 269976af4d93STejun Heo spin_lock(&pwq->pool->lock); 2700112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 270176af4d93STejun Heo spin_unlock(&pwq->pool->lock); 2702fa2563e4SThomas Tuttle 2703fa2563e4SThomas Tuttle if (drained) 27049c5a2ba7STejun Heo continue; 27059c5a2ba7STejun Heo 27069c5a2ba7STejun Heo if (++flush_cnt == 10 || 27079c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2708c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27099c5a2ba7STejun Heo wq->name, flush_cnt); 271076af4d93STejun Heo 271176af4d93STejun Heo local_irq_enable(); 27129c5a2ba7STejun Heo goto reflush; 27139c5a2ba7STejun Heo } 27149c5a2ba7STejun Heo 27155bcab335STejun Heo local_irq_enable(); 27165bcab335STejun Heo 271768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 27189c5a2ba7STejun Heo if (!--wq->nr_drainers) 2719618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 272068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 27219c5a2ba7STejun Heo } 27229c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27239c5a2ba7STejun Heo 2724606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2725baf59022STejun Heo { 2726baf59022STejun Heo struct worker *worker = NULL; 2727c9e7cf27STejun Heo struct worker_pool *pool; 2728112202d9STejun Heo struct pool_workqueue *pwq; 2729baf59022STejun Heo 2730baf59022STejun Heo might_sleep(); 2731baf59022STejun Heo 2732fa1b54e6STejun Heo local_irq_disable(); 2733fa1b54e6STejun Heo pool = get_work_pool(work); 2734fa1b54e6STejun Heo if (!pool) { 2735fa1b54e6STejun Heo local_irq_enable(); 2736fa1b54e6STejun Heo return false; 2737fa1b54e6STejun Heo } 2738fa1b54e6STejun Heo 2739fa1b54e6STejun Heo spin_lock(&pool->lock); 27400b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2741112202d9STejun Heo pwq = get_work_pwq(work); 2742112202d9STejun Heo if (pwq) { 2743112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2744baf59022STejun Heo goto already_gone; 2745606a5020STejun Heo } else { 2746c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2747baf59022STejun Heo if (!worker) 2748baf59022STejun Heo goto already_gone; 2749112202d9STejun Heo pwq = worker->current_pwq; 2750606a5020STejun Heo } 2751baf59022STejun Heo 2752112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2753d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2754baf59022STejun Heo 2755e159489bSTejun Heo /* 2756e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2757e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2758e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2759e159489bSTejun Heo * access. 2760e159489bSTejun Heo */ 2761493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2762112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2763e159489bSTejun Heo else 2764112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2765112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2766e159489bSTejun Heo 2767baf59022STejun Heo return true; 2768baf59022STejun Heo already_gone: 2769d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2770baf59022STejun Heo return false; 2771baf59022STejun Heo } 2772baf59022STejun Heo 2773db700897SOleg Nesterov /** 2774401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2775401a8d04STejun Heo * @work: the work to flush 2776db700897SOleg Nesterov * 2777606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2778606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2779401a8d04STejun Heo * 2780401a8d04STejun Heo * RETURNS: 2781401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2782401a8d04STejun Heo * %false if it was already idle. 2783db700897SOleg Nesterov */ 2784401a8d04STejun Heo bool flush_work(struct work_struct *work) 2785db700897SOleg Nesterov { 2786db700897SOleg Nesterov struct wq_barrier barr; 2787db700897SOleg Nesterov 27880976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27890976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27900976dfc1SStephen Boyd 2791606a5020STejun Heo if (start_flush_work(work, &barr)) { 2792db700897SOleg Nesterov wait_for_completion(&barr.done); 2793dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2794401a8d04STejun Heo return true; 2795606a5020STejun Heo } else { 2796401a8d04STejun Heo return false; 2797db700897SOleg Nesterov } 2798606a5020STejun Heo } 2799db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2800db700897SOleg Nesterov 280136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2802401a8d04STejun Heo { 2803bbb68dfaSTejun Heo unsigned long flags; 28041f1f642eSOleg Nesterov int ret; 28051f1f642eSOleg Nesterov 28061f1f642eSOleg Nesterov do { 2807bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2808bbb68dfaSTejun Heo /* 2809bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2810bbb68dfaSTejun Heo * would be waiting for before retrying. 2811bbb68dfaSTejun Heo */ 2812bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2813606a5020STejun Heo flush_work(work); 28141f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28151f1f642eSOleg Nesterov 2816bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2817bbb68dfaSTejun Heo mark_work_canceling(work); 2818bbb68dfaSTejun Heo local_irq_restore(flags); 2819bbb68dfaSTejun Heo 2820606a5020STejun Heo flush_work(work); 28217a22ad75STejun Heo clear_work_data(work); 28221f1f642eSOleg Nesterov return ret; 28231f1f642eSOleg Nesterov } 28241f1f642eSOleg Nesterov 28256e84d644SOleg Nesterov /** 2826401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2827401a8d04STejun Heo * @work: the work to cancel 28286e84d644SOleg Nesterov * 2829401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2830401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2831401a8d04STejun Heo * another workqueue. On return from this function, @work is 2832401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28331f1f642eSOleg Nesterov * 2834401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2835401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28366e84d644SOleg Nesterov * 2837401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28386e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2839401a8d04STejun Heo * 2840401a8d04STejun Heo * RETURNS: 2841401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28426e84d644SOleg Nesterov */ 2843401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28446e84d644SOleg Nesterov { 284536e227d2STejun Heo return __cancel_work_timer(work, false); 2846b89deed3SOleg Nesterov } 284728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2848b89deed3SOleg Nesterov 28496e84d644SOleg Nesterov /** 2850401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2851401a8d04STejun Heo * @dwork: the delayed work to flush 28526e84d644SOleg Nesterov * 2853401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2854401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2855401a8d04STejun Heo * considers the last queueing instance of @dwork. 28561f1f642eSOleg Nesterov * 2857401a8d04STejun Heo * RETURNS: 2858401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2859401a8d04STejun Heo * %false if it was already idle. 28606e84d644SOleg Nesterov */ 2861401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2862401a8d04STejun Heo { 28638930cabaSTejun Heo local_irq_disable(); 2864401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 286560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28668930cabaSTejun Heo local_irq_enable(); 2867401a8d04STejun Heo return flush_work(&dwork->work); 2868401a8d04STejun Heo } 2869401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2870401a8d04STejun Heo 2871401a8d04STejun Heo /** 287257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 287357b30ae7STejun Heo * @dwork: delayed_work to cancel 287409383498STejun Heo * 287557b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 287657b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 287757b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 287857b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 287957b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 288009383498STejun Heo * 288157b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 288209383498STejun Heo */ 288357b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 288409383498STejun Heo { 288557b30ae7STejun Heo unsigned long flags; 288657b30ae7STejun Heo int ret; 288757b30ae7STejun Heo 288857b30ae7STejun Heo do { 288957b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 289057b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 289157b30ae7STejun Heo 289257b30ae7STejun Heo if (unlikely(ret < 0)) 289357b30ae7STejun Heo return false; 289457b30ae7STejun Heo 28957c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28967c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289757b30ae7STejun Heo local_irq_restore(flags); 2898c0158ca6SDan Magenheimer return ret; 289909383498STejun Heo } 290057b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 290109383498STejun Heo 290209383498STejun Heo /** 2903401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2904401a8d04STejun Heo * @dwork: the delayed work cancel 2905401a8d04STejun Heo * 2906401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2907401a8d04STejun Heo * 2908401a8d04STejun Heo * RETURNS: 2909401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2910401a8d04STejun Heo */ 2911401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29126e84d644SOleg Nesterov { 291336e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29146e84d644SOleg Nesterov } 2915f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29161da177e4SLinus Torvalds 29170fcb78c2SRolf Eike Beer /** 291831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2919b6136773SAndrew Morton * @func: the function to call 2920b6136773SAndrew Morton * 292131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 292231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2923b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 292431ddd871STejun Heo * 292531ddd871STejun Heo * RETURNS: 292631ddd871STejun Heo * 0 on success, -errno on failure. 2927b6136773SAndrew Morton */ 292865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 292915316ba8SChristoph Lameter { 293015316ba8SChristoph Lameter int cpu; 293138f51568SNamhyung Kim struct work_struct __percpu *works; 293215316ba8SChristoph Lameter 2933b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2934b6136773SAndrew Morton if (!works) 293515316ba8SChristoph Lameter return -ENOMEM; 2936b6136773SAndrew Morton 293795402b38SGautham R Shenoy get_online_cpus(); 293893981800STejun Heo 293915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29409bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29419bfb1839SIngo Molnar 29429bfb1839SIngo Molnar INIT_WORK(work, func); 29438de6d308SOleg Nesterov schedule_work_on(cpu, work); 294415316ba8SChristoph Lameter } 294593981800STejun Heo 294693981800STejun Heo for_each_online_cpu(cpu) 29478616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 294893981800STejun Heo 294995402b38SGautham R Shenoy put_online_cpus(); 2950b6136773SAndrew Morton free_percpu(works); 295115316ba8SChristoph Lameter return 0; 295215316ba8SChristoph Lameter } 295315316ba8SChristoph Lameter 2954eef6a7d5SAlan Stern /** 2955eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2956eef6a7d5SAlan Stern * 2957eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2958eef6a7d5SAlan Stern * completion. 2959eef6a7d5SAlan Stern * 2960eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2961eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2962eef6a7d5SAlan Stern * will lead to deadlock: 2963eef6a7d5SAlan Stern * 2964eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2965eef6a7d5SAlan Stern * a lock held by your code or its caller. 2966eef6a7d5SAlan Stern * 2967eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2968eef6a7d5SAlan Stern * 2969eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2970eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2971eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2972eef6a7d5SAlan Stern * 2973eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2974eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2975eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2976eef6a7d5SAlan Stern * cancel_work_sync() instead. 2977eef6a7d5SAlan Stern */ 29781da177e4SLinus Torvalds void flush_scheduled_work(void) 29791da177e4SLinus Torvalds { 2980d320c038STejun Heo flush_workqueue(system_wq); 29811da177e4SLinus Torvalds } 2982ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29831da177e4SLinus Torvalds 29841da177e4SLinus Torvalds /** 29851fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29861fa44ecaSJames Bottomley * @fn: the function to execute 29871fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29881fa44ecaSJames Bottomley * be available when the work executes) 29891fa44ecaSJames Bottomley * 29901fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29911fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29921fa44ecaSJames Bottomley * 29931fa44ecaSJames Bottomley * Returns: 0 - function was executed 29941fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29951fa44ecaSJames Bottomley */ 299665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29971fa44ecaSJames Bottomley { 29981fa44ecaSJames Bottomley if (!in_interrupt()) { 299965f27f38SDavid Howells fn(&ew->work); 30001fa44ecaSJames Bottomley return 0; 30011fa44ecaSJames Bottomley } 30021fa44ecaSJames Bottomley 300365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30041fa44ecaSJames Bottomley schedule_work(&ew->work); 30051fa44ecaSJames Bottomley 30061fa44ecaSJames Bottomley return 1; 30071fa44ecaSJames Bottomley } 30081fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30091fa44ecaSJames Bottomley 3010226223abSTejun Heo #ifdef CONFIG_SYSFS 3011226223abSTejun Heo /* 3012226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3013226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3014226223abSTejun Heo * following attributes. 3015226223abSTejun Heo * 3016226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3017226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3018226223abSTejun Heo * 3019226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3020226223abSTejun Heo * 3021226223abSTejun Heo * id RO int : the associated pool ID 3022226223abSTejun Heo * nice RW int : nice value of the workers 3023226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3024226223abSTejun Heo */ 3025226223abSTejun Heo struct wq_device { 3026226223abSTejun Heo struct workqueue_struct *wq; 3027226223abSTejun Heo struct device dev; 3028226223abSTejun Heo }; 3029226223abSTejun Heo 3030226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3031226223abSTejun Heo { 3032226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3033226223abSTejun Heo 3034226223abSTejun Heo return wq_dev->wq; 3035226223abSTejun Heo } 3036226223abSTejun Heo 3037226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3038226223abSTejun Heo struct device_attribute *attr, char *buf) 3039226223abSTejun Heo { 3040226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3041226223abSTejun Heo 3042226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3043226223abSTejun Heo } 3044226223abSTejun Heo 3045226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3046226223abSTejun Heo struct device_attribute *attr, char *buf) 3047226223abSTejun Heo { 3048226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3049226223abSTejun Heo 3050226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3051226223abSTejun Heo } 3052226223abSTejun Heo 3053226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3054226223abSTejun Heo struct device_attribute *attr, 3055226223abSTejun Heo const char *buf, size_t count) 3056226223abSTejun Heo { 3057226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3058226223abSTejun Heo int val; 3059226223abSTejun Heo 3060226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3061226223abSTejun Heo return -EINVAL; 3062226223abSTejun Heo 3063226223abSTejun Heo workqueue_set_max_active(wq, val); 3064226223abSTejun Heo return count; 3065226223abSTejun Heo } 3066226223abSTejun Heo 3067226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3068226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3069226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3070226223abSTejun Heo __ATTR_NULL, 3071226223abSTejun Heo }; 3072226223abSTejun Heo 3073226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3074226223abSTejun Heo struct device_attribute *attr, char *buf) 3075226223abSTejun Heo { 3076226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3077226223abSTejun Heo struct worker_pool *pool; 3078226223abSTejun Heo int written; 3079226223abSTejun Heo 3080226223abSTejun Heo rcu_read_lock_sched(); 3081226223abSTejun Heo pool = first_pwq(wq)->pool; 3082226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3083226223abSTejun Heo rcu_read_unlock_sched(); 3084226223abSTejun Heo 3085226223abSTejun Heo return written; 3086226223abSTejun Heo } 3087226223abSTejun Heo 3088226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3089226223abSTejun Heo char *buf) 3090226223abSTejun Heo { 3091226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3092226223abSTejun Heo int written; 3093226223abSTejun Heo 3094226223abSTejun Heo rcu_read_lock_sched(); 3095226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3096226223abSTejun Heo first_pwq(wq)->pool->attrs->nice); 3097226223abSTejun Heo rcu_read_unlock_sched(); 3098226223abSTejun Heo 3099226223abSTejun Heo return written; 3100226223abSTejun Heo } 3101226223abSTejun Heo 3102226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3103226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3104226223abSTejun Heo { 3105226223abSTejun Heo struct workqueue_attrs *attrs; 3106226223abSTejun Heo 3107226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3108226223abSTejun Heo if (!attrs) 3109226223abSTejun Heo return NULL; 3110226223abSTejun Heo 3111226223abSTejun Heo rcu_read_lock_sched(); 3112226223abSTejun Heo copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs); 3113226223abSTejun Heo rcu_read_unlock_sched(); 3114226223abSTejun Heo return attrs; 3115226223abSTejun Heo } 3116226223abSTejun Heo 3117226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3118226223abSTejun Heo const char *buf, size_t count) 3119226223abSTejun Heo { 3120226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3121226223abSTejun Heo struct workqueue_attrs *attrs; 3122226223abSTejun Heo int ret; 3123226223abSTejun Heo 3124226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3125226223abSTejun Heo if (!attrs) 3126226223abSTejun Heo return -ENOMEM; 3127226223abSTejun Heo 3128226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3129226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3130226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3131226223abSTejun Heo else 3132226223abSTejun Heo ret = -EINVAL; 3133226223abSTejun Heo 3134226223abSTejun Heo free_workqueue_attrs(attrs); 3135226223abSTejun Heo return ret ?: count; 3136226223abSTejun Heo } 3137226223abSTejun Heo 3138226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3139226223abSTejun Heo struct device_attribute *attr, char *buf) 3140226223abSTejun Heo { 3141226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3142226223abSTejun Heo int written; 3143226223abSTejun Heo 3144226223abSTejun Heo rcu_read_lock_sched(); 3145226223abSTejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, 3146226223abSTejun Heo first_pwq(wq)->pool->attrs->cpumask); 3147226223abSTejun Heo rcu_read_unlock_sched(); 3148226223abSTejun Heo 3149226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3150226223abSTejun Heo return written; 3151226223abSTejun Heo } 3152226223abSTejun Heo 3153226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3154226223abSTejun Heo struct device_attribute *attr, 3155226223abSTejun Heo const char *buf, size_t count) 3156226223abSTejun Heo { 3157226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3158226223abSTejun Heo struct workqueue_attrs *attrs; 3159226223abSTejun Heo int ret; 3160226223abSTejun Heo 3161226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3162226223abSTejun Heo if (!attrs) 3163226223abSTejun Heo return -ENOMEM; 3164226223abSTejun Heo 3165226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3166226223abSTejun Heo if (!ret) 3167226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3168226223abSTejun Heo 3169226223abSTejun Heo free_workqueue_attrs(attrs); 3170226223abSTejun Heo return ret ?: count; 3171226223abSTejun Heo } 3172226223abSTejun Heo 3173226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3174226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3175226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3176226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3177226223abSTejun Heo __ATTR_NULL, 3178226223abSTejun Heo }; 3179226223abSTejun Heo 3180226223abSTejun Heo static struct bus_type wq_subsys = { 3181226223abSTejun Heo .name = "workqueue", 3182226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3183226223abSTejun Heo }; 3184226223abSTejun Heo 3185226223abSTejun Heo static int __init wq_sysfs_init(void) 3186226223abSTejun Heo { 3187226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3188226223abSTejun Heo } 3189226223abSTejun Heo core_initcall(wq_sysfs_init); 3190226223abSTejun Heo 3191226223abSTejun Heo static void wq_device_release(struct device *dev) 3192226223abSTejun Heo { 3193226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3194226223abSTejun Heo 3195226223abSTejun Heo kfree(wq_dev); 3196226223abSTejun Heo } 3197226223abSTejun Heo 3198226223abSTejun Heo /** 3199226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3200226223abSTejun Heo * @wq: the workqueue to register 3201226223abSTejun Heo * 3202226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3203226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3204226223abSTejun Heo * which is the preferred method. 3205226223abSTejun Heo * 3206226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3207226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3208226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3209226223abSTejun Heo * attributes. 3210226223abSTejun Heo * 3211226223abSTejun Heo * Returns 0 on success, -errno on failure. 3212226223abSTejun Heo */ 3213226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3214226223abSTejun Heo { 3215226223abSTejun Heo struct wq_device *wq_dev; 3216226223abSTejun Heo int ret; 3217226223abSTejun Heo 3218226223abSTejun Heo /* 3219226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3220226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3221226223abSTejun Heo * workqueues. 3222226223abSTejun Heo */ 3223226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3224226223abSTejun Heo return -EINVAL; 3225226223abSTejun Heo 3226226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3227226223abSTejun Heo if (!wq_dev) 3228226223abSTejun Heo return -ENOMEM; 3229226223abSTejun Heo 3230226223abSTejun Heo wq_dev->wq = wq; 3231226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3232226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3233226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3234226223abSTejun Heo 3235226223abSTejun Heo /* 3236226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3237226223abSTejun Heo * everything is ready. 3238226223abSTejun Heo */ 3239226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3240226223abSTejun Heo 3241226223abSTejun Heo ret = device_register(&wq_dev->dev); 3242226223abSTejun Heo if (ret) { 3243226223abSTejun Heo kfree(wq_dev); 3244226223abSTejun Heo wq->wq_dev = NULL; 3245226223abSTejun Heo return ret; 3246226223abSTejun Heo } 3247226223abSTejun Heo 3248226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3249226223abSTejun Heo struct device_attribute *attr; 3250226223abSTejun Heo 3251226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3252226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3253226223abSTejun Heo if (ret) { 3254226223abSTejun Heo device_unregister(&wq_dev->dev); 3255226223abSTejun Heo wq->wq_dev = NULL; 3256226223abSTejun Heo return ret; 3257226223abSTejun Heo } 3258226223abSTejun Heo } 3259226223abSTejun Heo } 3260226223abSTejun Heo 3261226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3262226223abSTejun Heo return 0; 3263226223abSTejun Heo } 3264226223abSTejun Heo 3265226223abSTejun Heo /** 3266226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3267226223abSTejun Heo * @wq: the workqueue to unregister 3268226223abSTejun Heo * 3269226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3270226223abSTejun Heo */ 3271226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3272226223abSTejun Heo { 3273226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3274226223abSTejun Heo 3275226223abSTejun Heo if (!wq->wq_dev) 3276226223abSTejun Heo return; 3277226223abSTejun Heo 3278226223abSTejun Heo wq->wq_dev = NULL; 3279226223abSTejun Heo device_unregister(&wq_dev->dev); 3280226223abSTejun Heo } 3281226223abSTejun Heo #else /* CONFIG_SYSFS */ 3282226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3283226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3284226223abSTejun Heo 32857a4e344cSTejun Heo /** 32867a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32877a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32887a4e344cSTejun Heo * 32897a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32907a4e344cSTejun Heo */ 32917a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32927a4e344cSTejun Heo { 32937a4e344cSTejun Heo if (attrs) { 32947a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32957a4e344cSTejun Heo kfree(attrs); 32967a4e344cSTejun Heo } 32977a4e344cSTejun Heo } 32987a4e344cSTejun Heo 32997a4e344cSTejun Heo /** 33007a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33017a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33027a4e344cSTejun Heo * 33037a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33047a4e344cSTejun Heo * return it. Returns NULL on failure. 33057a4e344cSTejun Heo */ 33067a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33077a4e344cSTejun Heo { 33087a4e344cSTejun Heo struct workqueue_attrs *attrs; 33097a4e344cSTejun Heo 33107a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33117a4e344cSTejun Heo if (!attrs) 33127a4e344cSTejun Heo goto fail; 33137a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33147a4e344cSTejun Heo goto fail; 33157a4e344cSTejun Heo 33167a4e344cSTejun Heo cpumask_setall(attrs->cpumask); 33177a4e344cSTejun Heo return attrs; 33187a4e344cSTejun Heo fail: 33197a4e344cSTejun Heo free_workqueue_attrs(attrs); 33207a4e344cSTejun Heo return NULL; 33217a4e344cSTejun Heo } 33227a4e344cSTejun Heo 332329c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 332429c91e99STejun Heo const struct workqueue_attrs *from) 332529c91e99STejun Heo { 332629c91e99STejun Heo to->nice = from->nice; 332729c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 332829c91e99STejun Heo } 332929c91e99STejun Heo 333029c91e99STejun Heo /* 333129c91e99STejun Heo * Hacky implementation of jhash of bitmaps which only considers the 333229c91e99STejun Heo * specified number of bits. We probably want a proper implementation in 333329c91e99STejun Heo * include/linux/jhash.h. 333429c91e99STejun Heo */ 333529c91e99STejun Heo static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash) 333629c91e99STejun Heo { 333729c91e99STejun Heo int nr_longs = bits / BITS_PER_LONG; 333829c91e99STejun Heo int nr_leftover = bits % BITS_PER_LONG; 333929c91e99STejun Heo unsigned long leftover = 0; 334029c91e99STejun Heo 334129c91e99STejun Heo if (nr_longs) 334229c91e99STejun Heo hash = jhash(bitmap, nr_longs * sizeof(long), hash); 334329c91e99STejun Heo if (nr_leftover) { 334429c91e99STejun Heo bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover); 334529c91e99STejun Heo hash = jhash(&leftover, sizeof(long), hash); 334629c91e99STejun Heo } 334729c91e99STejun Heo return hash; 334829c91e99STejun Heo } 334929c91e99STejun Heo 335029c91e99STejun Heo /* hash value of the content of @attr */ 335129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 335229c91e99STejun Heo { 335329c91e99STejun Heo u32 hash = 0; 335429c91e99STejun Heo 335529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 335629c91e99STejun Heo hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash); 335729c91e99STejun Heo return hash; 335829c91e99STejun Heo } 335929c91e99STejun Heo 336029c91e99STejun Heo /* content equality test */ 336129c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 336229c91e99STejun Heo const struct workqueue_attrs *b) 336329c91e99STejun Heo { 336429c91e99STejun Heo if (a->nice != b->nice) 336529c91e99STejun Heo return false; 336629c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 336729c91e99STejun Heo return false; 336829c91e99STejun Heo return true; 336929c91e99STejun Heo } 337029c91e99STejun Heo 33717a4e344cSTejun Heo /** 33727a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33737a4e344cSTejun Heo * @pool: worker_pool to initialize 33747a4e344cSTejun Heo * 33757a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 337629c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 337729c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 337829c91e99STejun Heo * on @pool safely to release it. 33797a4e344cSTejun Heo */ 33807a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33814e1a1f9aSTejun Heo { 33824e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 338329c91e99STejun Heo pool->id = -1; 338429c91e99STejun Heo pool->cpu = -1; 33854e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33864e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33874e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33884e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33894e1a1f9aSTejun Heo 33904e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33914e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33924e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33934e1a1f9aSTejun Heo 33944e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33954e1a1f9aSTejun Heo (unsigned long)pool); 33964e1a1f9aSTejun Heo 33974e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3398bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3399822d8405STejun Heo idr_init(&pool->worker_idr); 34007a4e344cSTejun Heo 340129c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 340229c91e99STejun Heo pool->refcnt = 1; 340329c91e99STejun Heo 340429c91e99STejun Heo /* shouldn't fail above this point */ 34057a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34067a4e344cSTejun Heo if (!pool->attrs) 34077a4e344cSTejun Heo return -ENOMEM; 34087a4e344cSTejun Heo return 0; 34094e1a1f9aSTejun Heo } 34104e1a1f9aSTejun Heo 341129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 341229c91e99STejun Heo { 341329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 341429c91e99STejun Heo 3415822d8405STejun Heo idr_destroy(&pool->worker_idr); 341629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 341729c91e99STejun Heo kfree(pool); 341829c91e99STejun Heo } 341929c91e99STejun Heo 342029c91e99STejun Heo /** 342129c91e99STejun Heo * put_unbound_pool - put a worker_pool 342229c91e99STejun Heo * @pool: worker_pool to put 342329c91e99STejun Heo * 342429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3425c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3426c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3427c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 342829c91e99STejun Heo */ 342929c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 343029c91e99STejun Heo { 343129c91e99STejun Heo struct worker *worker; 343229c91e99STejun Heo 343368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 343429c91e99STejun Heo if (--pool->refcnt) { 343568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 343629c91e99STejun Heo return; 343729c91e99STejun Heo } 343829c91e99STejun Heo 343929c91e99STejun Heo /* sanity checks */ 344029c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 344129c91e99STejun Heo WARN_ON(!list_empty(&pool->worklist))) { 344268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 344329c91e99STejun Heo return; 344429c91e99STejun Heo } 344529c91e99STejun Heo 344629c91e99STejun Heo /* release id and unhash */ 344729c91e99STejun Heo if (pool->id >= 0) 344829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 344929c91e99STejun Heo hash_del(&pool->hash_node); 345029c91e99STejun Heo 345168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 345229c91e99STejun Heo 3453c5aa87bbSTejun Heo /* 3454c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3455c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3456c5aa87bbSTejun Heo * manager_mutex. 3457c5aa87bbSTejun Heo */ 345829c91e99STejun Heo mutex_lock(&pool->manager_arb); 3459cd549687STejun Heo mutex_lock(&pool->manager_mutex); 346029c91e99STejun Heo spin_lock_irq(&pool->lock); 346129c91e99STejun Heo 346229c91e99STejun Heo while ((worker = first_worker(pool))) 346329c91e99STejun Heo destroy_worker(worker); 346429c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 346529c91e99STejun Heo 346629c91e99STejun Heo spin_unlock_irq(&pool->lock); 3467cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 346829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 346929c91e99STejun Heo 347029c91e99STejun Heo /* shut down the timers */ 347129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 347229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 347329c91e99STejun Heo 347429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 347529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 347629c91e99STejun Heo } 347729c91e99STejun Heo 347829c91e99STejun Heo /** 347929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 348029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 348129c91e99STejun Heo * 348229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 348329c91e99STejun Heo * reference count and return it. If there already is a matching 348429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 348529c91e99STejun Heo * create a new one. On failure, returns NULL. 348629c91e99STejun Heo */ 348729c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 348829c91e99STejun Heo { 348929c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 349029c91e99STejun Heo struct worker_pool *pool; 349129c91e99STejun Heo 349268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 349329c91e99STejun Heo 349429c91e99STejun Heo /* do we already have a matching pool? */ 349529c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 349629c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 349729c91e99STejun Heo pool->refcnt++; 349829c91e99STejun Heo goto out_unlock; 349929c91e99STejun Heo } 350029c91e99STejun Heo } 350129c91e99STejun Heo 350229c91e99STejun Heo /* nope, create a new one */ 350329c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 350429c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 350529c91e99STejun Heo goto fail; 350629c91e99STejun Heo 350712ee4fc6SLai Jiangshan if (workqueue_freezing) 350812ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 350912ee4fc6SLai Jiangshan 35108864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 351129c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 351229c91e99STejun Heo 351329c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 351429c91e99STejun Heo goto fail; 351529c91e99STejun Heo 351629c91e99STejun Heo /* create and start the initial worker */ 3517ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 351829c91e99STejun Heo goto fail; 351929c91e99STejun Heo 352029c91e99STejun Heo /* install */ 352129c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 352229c91e99STejun Heo out_unlock: 352368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 352429c91e99STejun Heo return pool; 352529c91e99STejun Heo fail: 352668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 352729c91e99STejun Heo if (pool) 352829c91e99STejun Heo put_unbound_pool(pool); 352929c91e99STejun Heo return NULL; 353029c91e99STejun Heo } 353129c91e99STejun Heo 35328864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35338864b4e5STejun Heo { 35348864b4e5STejun Heo kmem_cache_free(pwq_cache, 35358864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35368864b4e5STejun Heo } 35378864b4e5STejun Heo 35388864b4e5STejun Heo /* 35398864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35408864b4e5STejun Heo * and needs to be destroyed. 35418864b4e5STejun Heo */ 35428864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35438864b4e5STejun Heo { 35448864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35458864b4e5STejun Heo unbound_release_work); 35468864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35478864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 35488864b4e5STejun Heo 35498864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35508864b4e5STejun Heo return; 35518864b4e5STejun Heo 355275ccf595STejun Heo /* 355375ccf595STejun Heo * Unlink @pwq. Synchronization against flush_mutex isn't strictly 355475ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 355575ccf595STejun Heo * and consistent with the linking path. 355675ccf595STejun Heo */ 355775ccf595STejun Heo mutex_lock(&wq->flush_mutex); 3558794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 35598864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3560794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 356175ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 35628864b4e5STejun Heo 35638864b4e5STejun Heo put_unbound_pool(pool); 35648864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35658864b4e5STejun Heo 35668864b4e5STejun Heo /* 35678864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35688864b4e5STejun Heo * is gonna access it anymore. Free it. 35698864b4e5STejun Heo */ 35708864b4e5STejun Heo if (list_empty(&wq->pwqs)) 35718864b4e5STejun Heo kfree(wq); 35728864b4e5STejun Heo } 35738864b4e5STejun Heo 35740fbd95aaSTejun Heo /** 3575699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35760fbd95aaSTejun Heo * @pwq: target pool_workqueue 35770fbd95aaSTejun Heo * 3578699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3579699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3580699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35810fbd95aaSTejun Heo */ 3582699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35830fbd95aaSTejun Heo { 3584699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3585699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3586699ce097STejun Heo 3587699ce097STejun Heo /* for @wq->saved_max_active */ 3588794b18bcSTejun Heo lockdep_assert_held(&pwq_lock); 3589699ce097STejun Heo 3590699ce097STejun Heo /* fast exit for non-freezable wqs */ 3591699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3592699ce097STejun Heo return; 3593699ce097STejun Heo 3594699ce097STejun Heo spin_lock(&pwq->pool->lock); 3595699ce097STejun Heo 3596699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3597699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35980fbd95aaSTejun Heo 35990fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36000fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36010fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3602951a078aSLai Jiangshan 3603951a078aSLai Jiangshan /* 3604951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3605951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3606951a078aSLai Jiangshan */ 3607951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3608699ce097STejun Heo } else { 3609699ce097STejun Heo pwq->max_active = 0; 3610699ce097STejun Heo } 3611699ce097STejun Heo 3612699ce097STejun Heo spin_unlock(&pwq->pool->lock); 36130fbd95aaSTejun Heo } 36140fbd95aaSTejun Heo 3615d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3616d2c1d404STejun Heo struct workqueue_struct *wq, 36179e8cd2f5STejun Heo struct worker_pool *pool, 36189e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3619d2c1d404STejun Heo { 3620d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3621d2c1d404STejun Heo 3622d2c1d404STejun Heo pwq->pool = pool; 3623d2c1d404STejun Heo pwq->wq = wq; 3624d2c1d404STejun Heo pwq->flush_color = -1; 36258864b4e5STejun Heo pwq->refcnt = 1; 3626d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3627d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36288864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3629d2c1d404STejun Heo 363075ccf595STejun Heo mutex_lock(&wq->flush_mutex); 3631794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 363275ccf595STejun Heo 3633983ca25eSTejun Heo /* 3634983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 3635983ca25eSTejun Heo * flush_mutex to avoid confusing flush_workqueue(). 3636983ca25eSTejun Heo */ 36379e8cd2f5STejun Heo if (p_last_pwq) 36389e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 363975ccf595STejun Heo pwq->work_color = wq->work_color; 3640983ca25eSTejun Heo 3641983ca25eSTejun Heo /* sync max_active to the current setting */ 3642983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3643983ca25eSTejun Heo 3644983ca25eSTejun Heo /* link in @pwq */ 36459e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 364675ccf595STejun Heo 3647794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 364875ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 3649d2c1d404STejun Heo } 3650d2c1d404STejun Heo 36519e8cd2f5STejun Heo /** 36529e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36539e8cd2f5STejun Heo * @wq: the target workqueue 36549e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36559e8cd2f5STejun Heo * 36569e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 36579e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 36589e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 36599e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 36609e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 36619e8cd2f5STejun Heo * 36629e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 36639e8cd2f5STejun Heo * failure. 36649e8cd2f5STejun Heo */ 36659e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36669e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36679e8cd2f5STejun Heo { 36689e8cd2f5STejun Heo struct pool_workqueue *pwq, *last_pwq; 36699e8cd2f5STejun Heo struct worker_pool *pool; 36709e8cd2f5STejun Heo 36718719dceaSTejun Heo /* only unbound workqueues can change attributes */ 36729e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 36739e8cd2f5STejun Heo return -EINVAL; 36749e8cd2f5STejun Heo 36758719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 36768719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 36778719dceaSTejun Heo return -EINVAL; 36788719dceaSTejun Heo 36799e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 36809e8cd2f5STejun Heo if (!pwq) 36819e8cd2f5STejun Heo return -ENOMEM; 36829e8cd2f5STejun Heo 36839e8cd2f5STejun Heo pool = get_unbound_pool(attrs); 36849e8cd2f5STejun Heo if (!pool) { 36859e8cd2f5STejun Heo kmem_cache_free(pwq_cache, pwq); 36869e8cd2f5STejun Heo return -ENOMEM; 36879e8cd2f5STejun Heo } 36889e8cd2f5STejun Heo 36899e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 36909e8cd2f5STejun Heo if (last_pwq) { 36919e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 36929e8cd2f5STejun Heo put_pwq(last_pwq); 36939e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 36949e8cd2f5STejun Heo } 36959e8cd2f5STejun Heo 36969e8cd2f5STejun Heo return 0; 36979e8cd2f5STejun Heo } 36989e8cd2f5STejun Heo 369930cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37001da177e4SLinus Torvalds { 370149e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 370230cdf249STejun Heo int cpu; 3703e1d8aa9fSFrederic Weisbecker 370430cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3705420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3706420c0ddbSTejun Heo if (!wq->cpu_pwqs) 370730cdf249STejun Heo return -ENOMEM; 370830cdf249STejun Heo 370930cdf249STejun Heo for_each_possible_cpu(cpu) { 37107fb98ea7STejun Heo struct pool_workqueue *pwq = 37117fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37127a62c2c8STejun Heo struct worker_pool *cpu_pools = 3713f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 371430cdf249STejun Heo 37159e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 371630cdf249STejun Heo } 371730cdf249STejun Heo return 0; 37189e8cd2f5STejun Heo } else { 37199e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37209e8cd2f5STejun Heo } 37210f900049STejun Heo } 37220f900049STejun Heo 3723f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3724f3421797STejun Heo const char *name) 3725b71ab8c2STejun Heo { 3726f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3727f3421797STejun Heo 3728f3421797STejun Heo if (max_active < 1 || max_active > lim) 3729044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3730f3421797STejun Heo max_active, name, 1, lim); 3731b71ab8c2STejun Heo 3732f3421797STejun Heo return clamp_val(max_active, 1, lim); 3733b71ab8c2STejun Heo } 3734b71ab8c2STejun Heo 3735b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 373697e37d7bSTejun Heo unsigned int flags, 37371e19ffc6STejun Heo int max_active, 3738eb13ba87SJohannes Berg struct lock_class_key *key, 3739b196be89STejun Heo const char *lock_name, ...) 37403af24433SOleg Nesterov { 3741b196be89STejun Heo va_list args, args1; 37423af24433SOleg Nesterov struct workqueue_struct *wq; 374349e3cf44STejun Heo struct pool_workqueue *pwq; 3744b196be89STejun Heo size_t namelen; 3745b196be89STejun Heo 3746b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3747b196be89STejun Heo va_start(args, lock_name); 3748b196be89STejun Heo va_copy(args1, args); 3749b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3750b196be89STejun Heo 3751b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3752b196be89STejun Heo if (!wq) 3753d2c1d404STejun Heo return NULL; 3754b196be89STejun Heo 3755b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3756b196be89STejun Heo va_end(args); 3757b196be89STejun Heo va_end(args1); 37583af24433SOleg Nesterov 3759d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3760b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 37613af24433SOleg Nesterov 3762b196be89STejun Heo /* init wq */ 376397e37d7bSTejun Heo wq->flags = flags; 3764a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 376573f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3766112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 376730cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 376873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 376973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3770493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 37713af24433SOleg Nesterov 3772eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3773cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 37743af24433SOleg Nesterov 377530cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3776d2c1d404STejun Heo goto err_free_wq; 37771537663fSTejun Heo 3778493008a8STejun Heo /* 3779493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3780493008a8STejun Heo * have a rescuer to guarantee forward progress. 3781493008a8STejun Heo */ 3782493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3783e22bee78STejun Heo struct worker *rescuer; 3784e22bee78STejun Heo 3785d2c1d404STejun Heo rescuer = alloc_worker(); 3786e22bee78STejun Heo if (!rescuer) 3787d2c1d404STejun Heo goto err_destroy; 3788e22bee78STejun Heo 3789111c225aSTejun Heo rescuer->rescue_wq = wq; 3790111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3791b196be89STejun Heo wq->name); 3792d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3793d2c1d404STejun Heo kfree(rescuer); 3794d2c1d404STejun Heo goto err_destroy; 3795d2c1d404STejun Heo } 3796e22bee78STejun Heo 3797d2c1d404STejun Heo wq->rescuer = rescuer; 379814a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3799e22bee78STejun Heo wake_up_process(rescuer->task); 38003af24433SOleg Nesterov } 38011537663fSTejun Heo 3802226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3803226223abSTejun Heo goto err_destroy; 3804226223abSTejun Heo 38053af24433SOleg Nesterov /* 380668e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 380768e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 380868e13a67SLai Jiangshan * list. 38093af24433SOleg Nesterov */ 381068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3811a0a1a5fdSTejun Heo 3812794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 381349e3cf44STejun Heo for_each_pwq(pwq, wq) 3814699ce097STejun Heo pwq_adjust_max_active(pwq); 3815794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 3816a0a1a5fdSTejun Heo 38173af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3818a0a1a5fdSTejun Heo 381968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38203af24433SOleg Nesterov 38213af24433SOleg Nesterov return wq; 3822d2c1d404STejun Heo 3823d2c1d404STejun Heo err_free_wq: 38244690c4abSTejun Heo kfree(wq); 3825d2c1d404STejun Heo return NULL; 3826d2c1d404STejun Heo err_destroy: 3827d2c1d404STejun Heo destroy_workqueue(wq); 38284690c4abSTejun Heo return NULL; 38291da177e4SLinus Torvalds } 3830d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 38311da177e4SLinus Torvalds 38323af24433SOleg Nesterov /** 38333af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 38343af24433SOleg Nesterov * @wq: target workqueue 38353af24433SOleg Nesterov * 38363af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 38373af24433SOleg Nesterov */ 38383af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 38393af24433SOleg Nesterov { 384049e3cf44STejun Heo struct pool_workqueue *pwq; 38413af24433SOleg Nesterov 38429c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 38439c5a2ba7STejun Heo drain_workqueue(wq); 3844c8efcc25STejun Heo 38456183c009STejun Heo /* sanity checks */ 3846794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 384749e3cf44STejun Heo for_each_pwq(pwq, wq) { 38486183c009STejun Heo int i; 38496183c009STejun Heo 385076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 385176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3852794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 38536183c009STejun Heo return; 385476af4d93STejun Heo } 385576af4d93STejun Heo } 385676af4d93STejun Heo 38578864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 38588864b4e5STejun Heo WARN_ON(pwq->nr_active) || 385976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3860794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 38616183c009STejun Heo return; 38626183c009STejun Heo } 386376af4d93STejun Heo } 3864794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 38656183c009STejun Heo 3866a0a1a5fdSTejun Heo /* 3867a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3868a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3869a0a1a5fdSTejun Heo */ 387068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3871d2c1d404STejun Heo list_del_init(&wq->list); 387268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38733af24433SOleg Nesterov 3874226223abSTejun Heo workqueue_sysfs_unregister(wq); 3875226223abSTejun Heo 3876493008a8STejun Heo if (wq->rescuer) { 3877e22bee78STejun Heo kthread_stop(wq->rescuer->task); 38788d9df9f0SXiaotian Feng kfree(wq->rescuer); 3879493008a8STejun Heo wq->rescuer = NULL; 3880e22bee78STejun Heo } 3881e22bee78STejun Heo 38828864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 388329c91e99STejun Heo /* 38848864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 38858864b4e5STejun Heo * free the pwqs and wq. 388629c91e99STejun Heo */ 38878864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 38888864b4e5STejun Heo kfree(wq); 38898864b4e5STejun Heo } else { 38908864b4e5STejun Heo /* 38918864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 38928864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 38938864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 38948864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 38958864b4e5STejun Heo * released. 38968864b4e5STejun Heo */ 389729c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 389829c91e99STejun Heo pwqs_node); 38998864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 39008864b4e5STejun Heo put_pwq(pwq); 39018864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 390229c91e99STejun Heo } 39033af24433SOleg Nesterov } 39043af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39053af24433SOleg Nesterov 3906dcd989cbSTejun Heo /** 3907dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3908dcd989cbSTejun Heo * @wq: target workqueue 3909dcd989cbSTejun Heo * @max_active: new max_active value. 3910dcd989cbSTejun Heo * 3911dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3912dcd989cbSTejun Heo * 3913dcd989cbSTejun Heo * CONTEXT: 3914dcd989cbSTejun Heo * Don't call from IRQ context. 3915dcd989cbSTejun Heo */ 3916dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3917dcd989cbSTejun Heo { 391849e3cf44STejun Heo struct pool_workqueue *pwq; 3919dcd989cbSTejun Heo 39208719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39218719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39228719dceaSTejun Heo return; 39238719dceaSTejun Heo 3924f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3925dcd989cbSTejun Heo 3926794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 3927dcd989cbSTejun Heo 3928dcd989cbSTejun Heo wq->saved_max_active = max_active; 3929dcd989cbSTejun Heo 3930699ce097STejun Heo for_each_pwq(pwq, wq) 3931699ce097STejun Heo pwq_adjust_max_active(pwq); 3932dcd989cbSTejun Heo 3933794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 3934dcd989cbSTejun Heo } 3935dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3936dcd989cbSTejun Heo 3937dcd989cbSTejun Heo /** 3938e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 3939e6267616STejun Heo * 3940e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 3941e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 3942e6267616STejun Heo */ 3943e6267616STejun Heo bool current_is_workqueue_rescuer(void) 3944e6267616STejun Heo { 3945e6267616STejun Heo struct worker *worker = current_wq_worker(); 3946e6267616STejun Heo 39476a092dfdSLai Jiangshan return worker && worker->rescue_wq; 3948e6267616STejun Heo } 3949e6267616STejun Heo 3950e6267616STejun Heo /** 3951dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3952dcd989cbSTejun Heo * @cpu: CPU in question 3953dcd989cbSTejun Heo * @wq: target workqueue 3954dcd989cbSTejun Heo * 3955dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3956dcd989cbSTejun Heo * no synchronization around this function and the test result is 3957dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3958dcd989cbSTejun Heo * 3959dcd989cbSTejun Heo * RETURNS: 3960dcd989cbSTejun Heo * %true if congested, %false otherwise. 3961dcd989cbSTejun Heo */ 3962d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 3963dcd989cbSTejun Heo { 39647fb98ea7STejun Heo struct pool_workqueue *pwq; 396576af4d93STejun Heo bool ret; 396676af4d93STejun Heo 396788109453SLai Jiangshan rcu_read_lock_sched(); 39687fb98ea7STejun Heo 39697fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 39707fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 39717fb98ea7STejun Heo else 39727fb98ea7STejun Heo pwq = first_pwq(wq); 3973dcd989cbSTejun Heo 397476af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 397588109453SLai Jiangshan rcu_read_unlock_sched(); 397676af4d93STejun Heo 397776af4d93STejun Heo return ret; 3978dcd989cbSTejun Heo } 3979dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3980dcd989cbSTejun Heo 3981dcd989cbSTejun Heo /** 3982dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3983dcd989cbSTejun Heo * @work: the work to be tested 3984dcd989cbSTejun Heo * 3985dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3986dcd989cbSTejun Heo * synchronization around this function and the test result is 3987dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3988dcd989cbSTejun Heo * 3989dcd989cbSTejun Heo * RETURNS: 3990dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3991dcd989cbSTejun Heo */ 3992dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3993dcd989cbSTejun Heo { 3994fa1b54e6STejun Heo struct worker_pool *pool; 3995dcd989cbSTejun Heo unsigned long flags; 3996dcd989cbSTejun Heo unsigned int ret = 0; 3997dcd989cbSTejun Heo 3998dcd989cbSTejun Heo if (work_pending(work)) 3999dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4000038366c5SLai Jiangshan 4001fa1b54e6STejun Heo local_irq_save(flags); 4002fa1b54e6STejun Heo pool = get_work_pool(work); 4003038366c5SLai Jiangshan if (pool) { 4004fa1b54e6STejun Heo spin_lock(&pool->lock); 4005c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4006dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4007fa1b54e6STejun Heo spin_unlock(&pool->lock); 4008038366c5SLai Jiangshan } 4009fa1b54e6STejun Heo local_irq_restore(flags); 4010dcd989cbSTejun Heo 4011dcd989cbSTejun Heo return ret; 4012dcd989cbSTejun Heo } 4013dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4014dcd989cbSTejun Heo 4015db7bccf4STejun Heo /* 4016db7bccf4STejun Heo * CPU hotplug. 4017db7bccf4STejun Heo * 4018e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4019112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4020706026c2STejun Heo * pool which make migrating pending and scheduled works very 4021e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 402294cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4023e22bee78STejun Heo * blocked draining impractical. 4024e22bee78STejun Heo * 402524647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4026628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4027628c78e7STejun Heo * cpu comes back online. 4028db7bccf4STejun Heo */ 4029db7bccf4STejun Heo 4030706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4031db7bccf4STejun Heo { 403238db41d9STejun Heo int cpu = smp_processor_id(); 40334ce62e9eSTejun Heo struct worker_pool *pool; 4034db7bccf4STejun Heo struct worker *worker; 4035a9ab775bSTejun Heo int wi; 4036db7bccf4STejun Heo 4037f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 40386183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4039db7bccf4STejun Heo 4040bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 404194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4042e22bee78STejun Heo 4043f2d5a0eeSTejun Heo /* 4044bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 404594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 404694cf58bbSTejun Heo * except for the ones which are still executing works from 404794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 404894cf58bbSTejun Heo * this, they may become diasporas. 4049f2d5a0eeSTejun Heo */ 4050a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4051403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4052db7bccf4STejun Heo 405324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4054f2d5a0eeSTejun Heo 405594cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4056bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 405794cf58bbSTejun Heo } 4058e22bee78STejun Heo 4059e22bee78STejun Heo /* 4060628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4061628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4062628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4063628c78e7STejun Heo */ 4064628c78e7STejun Heo schedule(); 4065628c78e7STejun Heo 4066628c78e7STejun Heo /* 4067628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4068628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 406938db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 407038db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 407138db41d9STejun Heo * pools which are served by workers tied to the CPU. 4072628c78e7STejun Heo * 4073628c78e7STejun Heo * On return from this function, the current worker would trigger 4074628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4075628c78e7STejun Heo * didn't already. 4076e22bee78STejun Heo */ 4077f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4078e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4079db7bccf4STejun Heo } 4080db7bccf4STejun Heo 4081bd7c089eSTejun Heo /** 4082bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4083bd7c089eSTejun Heo * @pool: pool of interest 4084bd7c089eSTejun Heo * 4085a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4086bd7c089eSTejun Heo */ 4087bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4088bd7c089eSTejun Heo { 4089a9ab775bSTejun Heo struct worker *worker; 4090a9ab775bSTejun Heo int wi; 4091bd7c089eSTejun Heo 4092bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4093bd7c089eSTejun Heo 4094bd7c089eSTejun Heo /* 4095a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4096a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4097a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4098a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4099a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4100bd7c089eSTejun Heo */ 4101a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4102a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4103a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4104a9ab775bSTejun Heo 4105a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4106a9ab775bSTejun Heo 4107a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4108a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4109a9ab775bSTejun Heo 4110a9ab775bSTejun Heo /* 4111a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4112a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4113a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4114a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4115a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4116a9ab775bSTejun Heo * be bound before @pool->lock is released. 4117a9ab775bSTejun Heo */ 4118a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4119bd7c089eSTejun Heo wake_up_process(worker->task); 4120bd7c089eSTejun Heo 4121bd7c089eSTejun Heo /* 4122a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4123a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4124a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4125a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4126a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4127a9ab775bSTejun Heo * concurrency management. Note that when or whether 4128a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4129a9ab775bSTejun Heo * 4130a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4131a9ab775bSTejun Heo * tested without holding any lock in 4132a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4133a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4134a9ab775bSTejun Heo * management operations. 4135bd7c089eSTejun Heo */ 4136a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4137a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4138a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4139a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4140bd7c089eSTejun Heo } 4141a9ab775bSTejun Heo 4142a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4143bd7c089eSTejun Heo } 4144bd7c089eSTejun Heo 41457dbc725eSTejun Heo /** 41467dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 41477dbc725eSTejun Heo * @pool: unbound pool of interest 41487dbc725eSTejun Heo * @cpu: the CPU which is coming up 41497dbc725eSTejun Heo * 41507dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 41517dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 41527dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 41537dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 41547dbc725eSTejun Heo */ 41557dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 41567dbc725eSTejun Heo { 41577dbc725eSTejun Heo static cpumask_t cpumask; 41587dbc725eSTejun Heo struct worker *worker; 41597dbc725eSTejun Heo int wi; 41607dbc725eSTejun Heo 41617dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 41627dbc725eSTejun Heo 41637dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 41647dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 41657dbc725eSTejun Heo return; 41667dbc725eSTejun Heo 41677dbc725eSTejun Heo /* is @cpu the only online CPU? */ 41687dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 41697dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 41707dbc725eSTejun Heo return; 41717dbc725eSTejun Heo 41727dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 41737dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 41747dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 41757dbc725eSTejun Heo pool->attrs->cpumask) < 0); 41767dbc725eSTejun Heo } 41777dbc725eSTejun Heo 41788db25e78STejun Heo /* 41798db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 41808db25e78STejun Heo * This will be registered high priority CPU notifier. 41818db25e78STejun Heo */ 41829fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 41831da177e4SLinus Torvalds unsigned long action, 41841da177e4SLinus Torvalds void *hcpu) 41851da177e4SLinus Torvalds { 4186d84ff051STejun Heo int cpu = (unsigned long)hcpu; 41874ce62e9eSTejun Heo struct worker_pool *pool; 41887dbc725eSTejun Heo int pi; 41891da177e4SLinus Torvalds 41908db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 41913af24433SOleg Nesterov case CPU_UP_PREPARE: 4192f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41933ce63377STejun Heo if (pool->nr_workers) 41943ce63377STejun Heo continue; 4195ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 41963ce63377STejun Heo return NOTIFY_BAD; 41973af24433SOleg Nesterov } 41981da177e4SLinus Torvalds break; 41991da177e4SLinus Torvalds 420065758202STejun Heo case CPU_DOWN_FAILED: 420165758202STejun Heo case CPU_ONLINE: 420268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 42037dbc725eSTejun Heo 42047dbc725eSTejun Heo for_each_pool(pool, pi) { 4205bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 420694cf58bbSTejun Heo 42077dbc725eSTejun Heo if (pool->cpu == cpu) { 4208a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 420924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4210a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4211a9ab775bSTejun Heo 421294cf58bbSTejun Heo rebind_workers(pool); 42137dbc725eSTejun Heo } else if (pool->cpu < 0) { 42147dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 42157dbc725eSTejun Heo } 421694cf58bbSTejun Heo 4217bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 421894cf58bbSTejun Heo } 42197dbc725eSTejun Heo 422068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42218db25e78STejun Heo break; 422265758202STejun Heo } 422365758202STejun Heo return NOTIFY_OK; 422465758202STejun Heo } 422565758202STejun Heo 422665758202STejun Heo /* 422765758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 422865758202STejun Heo * This will be registered as low priority CPU notifier. 422965758202STejun Heo */ 42309fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 423165758202STejun Heo unsigned long action, 423265758202STejun Heo void *hcpu) 423365758202STejun Heo { 4234d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42358db25e78STejun Heo struct work_struct unbind_work; 42368db25e78STejun Heo 423765758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 423865758202STejun Heo case CPU_DOWN_PREPARE: 42398db25e78STejun Heo /* unbinding should happen on the local CPU */ 4240706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42417635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42428db25e78STejun Heo flush_work(&unbind_work); 42438db25e78STejun Heo break; 424465758202STejun Heo } 424565758202STejun Heo return NOTIFY_OK; 424665758202STejun Heo } 424765758202STejun Heo 42482d3854a3SRusty Russell #ifdef CONFIG_SMP 42498ccad40dSRusty Russell 42502d3854a3SRusty Russell struct work_for_cpu { 4251ed48ece2STejun Heo struct work_struct work; 42522d3854a3SRusty Russell long (*fn)(void *); 42532d3854a3SRusty Russell void *arg; 42542d3854a3SRusty Russell long ret; 42552d3854a3SRusty Russell }; 42562d3854a3SRusty Russell 4257ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 42582d3854a3SRusty Russell { 4259ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4260ed48ece2STejun Heo 42612d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 42622d3854a3SRusty Russell } 42632d3854a3SRusty Russell 42642d3854a3SRusty Russell /** 42652d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 42662d3854a3SRusty Russell * @cpu: the cpu to run on 42672d3854a3SRusty Russell * @fn: the function to run 42682d3854a3SRusty Russell * @arg: the function arg 42692d3854a3SRusty Russell * 427031ad9081SRusty Russell * This will return the value @fn returns. 427131ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 42726b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 42732d3854a3SRusty Russell */ 4274d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 42752d3854a3SRusty Russell { 4276ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 42772d3854a3SRusty Russell 4278ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4279ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4280ed48ece2STejun Heo flush_work(&wfc.work); 42812d3854a3SRusty Russell return wfc.ret; 42822d3854a3SRusty Russell } 42832d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 42842d3854a3SRusty Russell #endif /* CONFIG_SMP */ 42852d3854a3SRusty Russell 4286a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4287e7577c50SRusty Russell 4288a0a1a5fdSTejun Heo /** 4289a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4290a0a1a5fdSTejun Heo * 429158a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4292c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4293706026c2STejun Heo * pool->worklist. 4294a0a1a5fdSTejun Heo * 4295a0a1a5fdSTejun Heo * CONTEXT: 429668e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex, pwq_lock and pool->lock's. 4297a0a1a5fdSTejun Heo */ 4298a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4299a0a1a5fdSTejun Heo { 430017116969STejun Heo struct worker_pool *pool; 430124b8a847STejun Heo struct workqueue_struct *wq; 430224b8a847STejun Heo struct pool_workqueue *pwq; 4303611c92a0STejun Heo int pi; 4304a0a1a5fdSTejun Heo 430568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4306a0a1a5fdSTejun Heo 43076183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4308a0a1a5fdSTejun Heo workqueue_freezing = true; 4309a0a1a5fdSTejun Heo 431024b8a847STejun Heo /* set FREEZING */ 4311611c92a0STejun Heo for_each_pool(pool, pi) { 43125bcab335STejun Heo spin_lock_irq(&pool->lock); 431335b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 431435b6bb63STejun Heo pool->flags |= POOL_FREEZING; 43155bcab335STejun Heo spin_unlock_irq(&pool->lock); 43161da177e4SLinus Torvalds } 43178b03ae3cSTejun Heo 431824b8a847STejun Heo /* suppress further executions by setting max_active to zero */ 4319794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 432024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4321699ce097STejun Heo for_each_pwq(pwq, wq) 4322699ce097STejun Heo pwq_adjust_max_active(pwq); 4323a1056305STejun Heo } 4324794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 43255bcab335STejun Heo 432668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4327a0a1a5fdSTejun Heo } 4328a0a1a5fdSTejun Heo 4329a0a1a5fdSTejun Heo /** 433058a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4331a0a1a5fdSTejun Heo * 4332a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4333a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4334a0a1a5fdSTejun Heo * 4335a0a1a5fdSTejun Heo * CONTEXT: 433668e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4337a0a1a5fdSTejun Heo * 4338a0a1a5fdSTejun Heo * RETURNS: 433958a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 434058a69cb4STejun Heo * is complete. 4341a0a1a5fdSTejun Heo */ 4342a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4343a0a1a5fdSTejun Heo { 4344a0a1a5fdSTejun Heo bool busy = false; 434524b8a847STejun Heo struct workqueue_struct *wq; 434624b8a847STejun Heo struct pool_workqueue *pwq; 4347a0a1a5fdSTejun Heo 434868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4349a0a1a5fdSTejun Heo 43506183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4351a0a1a5fdSTejun Heo 435224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 435324b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 435424b8a847STejun Heo continue; 4355a0a1a5fdSTejun Heo /* 4356a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4357a0a1a5fdSTejun Heo * to peek without lock. 4358a0a1a5fdSTejun Heo */ 435988109453SLai Jiangshan rcu_read_lock_sched(); 436024b8a847STejun Heo for_each_pwq(pwq, wq) { 43616183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4362112202d9STejun Heo if (pwq->nr_active) { 4363a0a1a5fdSTejun Heo busy = true; 436488109453SLai Jiangshan rcu_read_unlock_sched(); 4365a0a1a5fdSTejun Heo goto out_unlock; 4366a0a1a5fdSTejun Heo } 4367a0a1a5fdSTejun Heo } 436888109453SLai Jiangshan rcu_read_unlock_sched(); 4369a0a1a5fdSTejun Heo } 4370a0a1a5fdSTejun Heo out_unlock: 437168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4372a0a1a5fdSTejun Heo return busy; 4373a0a1a5fdSTejun Heo } 4374a0a1a5fdSTejun Heo 4375a0a1a5fdSTejun Heo /** 4376a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4377a0a1a5fdSTejun Heo * 4378a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4379706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4380a0a1a5fdSTejun Heo * 4381a0a1a5fdSTejun Heo * CONTEXT: 438268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex, pwq_lock and pool->lock's. 4383a0a1a5fdSTejun Heo */ 4384a0a1a5fdSTejun Heo void thaw_workqueues(void) 4385a0a1a5fdSTejun Heo { 438624b8a847STejun Heo struct workqueue_struct *wq; 438724b8a847STejun Heo struct pool_workqueue *pwq; 438824b8a847STejun Heo struct worker_pool *pool; 4389611c92a0STejun Heo int pi; 4390a0a1a5fdSTejun Heo 439168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4392a0a1a5fdSTejun Heo 4393a0a1a5fdSTejun Heo if (!workqueue_freezing) 4394a0a1a5fdSTejun Heo goto out_unlock; 4395a0a1a5fdSTejun Heo 439624b8a847STejun Heo /* clear FREEZING */ 4397611c92a0STejun Heo for_each_pool(pool, pi) { 43985bcab335STejun Heo spin_lock_irq(&pool->lock); 439935b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 440035b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 44015bcab335STejun Heo spin_unlock_irq(&pool->lock); 4402d565ed63STejun Heo } 440324b8a847STejun Heo 440424b8a847STejun Heo /* restore max_active and repopulate worklist */ 4405794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 440624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4407699ce097STejun Heo for_each_pwq(pwq, wq) 4408699ce097STejun Heo pwq_adjust_max_active(pwq); 440924b8a847STejun Heo } 4410794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 441124b8a847STejun Heo 4412a0a1a5fdSTejun Heo workqueue_freezing = false; 4413a0a1a5fdSTejun Heo out_unlock: 441468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4415a0a1a5fdSTejun Heo } 4416a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4417a0a1a5fdSTejun Heo 44186ee0578bSSuresh Siddha static int __init init_workqueues(void) 44191da177e4SLinus Torvalds { 44207a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44217a4e344cSTejun Heo int i, cpu; 4422c34056a3STejun Heo 44237c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44247c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44256be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4426b5490077STejun Heo 4427e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4428e904e6c2STejun Heo 4429e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4430e904e6c2STejun Heo 443165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4432a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44338b03ae3cSTejun Heo 4434706026c2STejun Heo /* initialize CPU pools */ 443529c91e99STejun Heo for_each_possible_cpu(cpu) { 44364ce62e9eSTejun Heo struct worker_pool *pool; 44378b03ae3cSTejun Heo 44387a4e344cSTejun Heo i = 0; 4439f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44407a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4441ec22ca5eSTejun Heo pool->cpu = cpu; 44427a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 44437a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 44447a4e344cSTejun Heo 44459daf9e67STejun Heo /* alloc pool ID */ 444668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 44479daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 444868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44494ce62e9eSTejun Heo } 44508b03ae3cSTejun Heo } 44518b03ae3cSTejun Heo 4452e22bee78STejun Heo /* create the initial worker */ 445329c91e99STejun Heo for_each_online_cpu(cpu) { 44544ce62e9eSTejun Heo struct worker_pool *pool; 4455e22bee78STejun Heo 4456f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 445724647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4458ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4459e22bee78STejun Heo } 44604ce62e9eSTejun Heo } 4461e22bee78STejun Heo 446229c91e99STejun Heo /* create default unbound wq attrs */ 446329c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 446429c91e99STejun Heo struct workqueue_attrs *attrs; 446529c91e99STejun Heo 446629c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 446729c91e99STejun Heo 446829c91e99STejun Heo attrs->nice = std_nice[i]; 446929c91e99STejun Heo cpumask_setall(attrs->cpumask); 447029c91e99STejun Heo 447129c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 447229c91e99STejun Heo } 447329c91e99STejun Heo 4474d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 44751aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4476d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4477f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4478f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 447924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 448024d51addSTejun Heo WQ_FREEZABLE, 0); 44811aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4482ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 44836ee0578bSSuresh Siddha return 0; 44841da177e4SLinus Torvalds } 44856ee0578bSSuresh Siddha early_initcall(init_workqueues); 4486