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 * 1265bcab335STejun Heo * WQ: wq_mutex protected. 12776af4d93STejun Heo * 1285bcab335STejun Heo * WR: wq_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 */ 1665bcab335STejun Heo struct hlist_node hash_node; /* WQ: unbound_pool_hash node */ 1675bcab335STejun Heo int refcnt; /* WQ: 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 { 2295bcab335STejun Heo unsigned int flags; /* WQ: 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 */ 2325bcab335STejun Heo struct list_head list; /* WQ: 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 2455bcab335STejun Heo int nr_drainers; /* WQ: 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 2595bcab335STejun Heo static DEFINE_MUTEX(wq_mutex); /* protects workqueues and pools */ 260794b18bcSTejun Heo static DEFINE_SPINLOCK(pwq_lock); /* protects pool_workqueues */ 2612e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2625bcab335STejun Heo 2635bcab335STejun Heo static LIST_HEAD(workqueues); /* WQ: list of all workqueues */ 2645bcab335STejun Heo static bool workqueue_freezing; /* WQ: 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 2705bcab335STejun Heo static DEFINE_IDR(worker_pool_idr); /* WR: idr of all pools */ 2717d19c5ceSTejun Heo 2725bcab335STejun Heo /* WQ: 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 2965bcab335STejun Heo #define assert_rcu_or_wq_mutex() \ 2975bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 2985bcab335STejun Heo lockdep_is_held(&wq_mutex), \ 2995bcab335STejun Heo "sched RCU or wq_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) \ 308822d8405STejun Heo WARN_ONCE(!lockdep_is_held(&(pool)->manager_mutex) && \ 309822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 310822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 311822d8405STejun Heo #else 312822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 313822d8405STejun Heo #endif 314822d8405STejun Heo 315f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 316f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 317f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3187a62c2c8STejun Heo (pool)++) 3194ce62e9eSTejun Heo 32049e3cf44STejun Heo /** 32117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 32217116969STejun Heo * @pool: iteration cursor 323611c92a0STejun Heo * @pi: integer used for iteration 324fa1b54e6STejun Heo * 3255bcab335STejun Heo * This must be called either with wq_mutex held or sched RCU read locked. 3265bcab335STejun Heo * If the pool needs to be used beyond the locking in effect, the caller is 3275bcab335STejun Heo * responsible for guaranteeing that the pool stays online. 328fa1b54e6STejun Heo * 329fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 330fa1b54e6STejun Heo * ignored. 33117116969STejun Heo */ 332611c92a0STejun Heo #define for_each_pool(pool, pi) \ 333611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 3345bcab335STejun Heo if (({ assert_rcu_or_wq_mutex(); false; })) { } \ 335fa1b54e6STejun Heo else 33617116969STejun Heo 33717116969STejun Heo /** 338822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 339822d8405STejun Heo * @worker: iteration cursor 340822d8405STejun Heo * @wi: integer used for iteration 341822d8405STejun Heo * @pool: worker_pool to iterate workers of 342822d8405STejun Heo * 343822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 344822d8405STejun Heo * 345822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 346822d8405STejun Heo * ignored. 347822d8405STejun Heo */ 348822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 349822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 350822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 351822d8405STejun Heo else 352822d8405STejun Heo 353822d8405STejun Heo /** 35449e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 35549e3cf44STejun Heo * @pwq: iteration cursor 35649e3cf44STejun Heo * @wq: the target workqueue 35776af4d93STejun Heo * 358794b18bcSTejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 359794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 360794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 36176af4d93STejun Heo * 36276af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 36376af4d93STejun Heo * ignored. 36449e3cf44STejun Heo */ 36549e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 36676af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 367794b18bcSTejun Heo if (({ assert_rcu_or_pwq_lock(); false; })) { } \ 36876af4d93STejun Heo else 369f3421797STejun Heo 370dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 371dc186ad7SThomas Gleixner 372dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 373dc186ad7SThomas Gleixner 37499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 37599777288SStanislaw Gruszka { 37699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 37799777288SStanislaw Gruszka } 37899777288SStanislaw Gruszka 379dc186ad7SThomas Gleixner /* 380dc186ad7SThomas Gleixner * fixup_init is called when: 381dc186ad7SThomas Gleixner * - an active object is initialized 382dc186ad7SThomas Gleixner */ 383dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 384dc186ad7SThomas Gleixner { 385dc186ad7SThomas Gleixner struct work_struct *work = addr; 386dc186ad7SThomas Gleixner 387dc186ad7SThomas Gleixner switch (state) { 388dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 389dc186ad7SThomas Gleixner cancel_work_sync(work); 390dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 391dc186ad7SThomas Gleixner return 1; 392dc186ad7SThomas Gleixner default: 393dc186ad7SThomas Gleixner return 0; 394dc186ad7SThomas Gleixner } 395dc186ad7SThomas Gleixner } 396dc186ad7SThomas Gleixner 397dc186ad7SThomas Gleixner /* 398dc186ad7SThomas Gleixner * fixup_activate is called when: 399dc186ad7SThomas Gleixner * - an active object is activated 400dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 401dc186ad7SThomas Gleixner */ 402dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 403dc186ad7SThomas Gleixner { 404dc186ad7SThomas Gleixner struct work_struct *work = addr; 405dc186ad7SThomas Gleixner 406dc186ad7SThomas Gleixner switch (state) { 407dc186ad7SThomas Gleixner 408dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 409dc186ad7SThomas Gleixner /* 410dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 411dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 412dc186ad7SThomas Gleixner * is tracked in the object tracker. 413dc186ad7SThomas Gleixner */ 41422df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 415dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 416dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 417dc186ad7SThomas Gleixner return 0; 418dc186ad7SThomas Gleixner } 419dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 420dc186ad7SThomas Gleixner return 0; 421dc186ad7SThomas Gleixner 422dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 423dc186ad7SThomas Gleixner WARN_ON(1); 424dc186ad7SThomas Gleixner 425dc186ad7SThomas Gleixner default: 426dc186ad7SThomas Gleixner return 0; 427dc186ad7SThomas Gleixner } 428dc186ad7SThomas Gleixner } 429dc186ad7SThomas Gleixner 430dc186ad7SThomas Gleixner /* 431dc186ad7SThomas Gleixner * fixup_free is called when: 432dc186ad7SThomas Gleixner * - an active object is freed 433dc186ad7SThomas Gleixner */ 434dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 435dc186ad7SThomas Gleixner { 436dc186ad7SThomas Gleixner struct work_struct *work = addr; 437dc186ad7SThomas Gleixner 438dc186ad7SThomas Gleixner switch (state) { 439dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 440dc186ad7SThomas Gleixner cancel_work_sync(work); 441dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 442dc186ad7SThomas Gleixner return 1; 443dc186ad7SThomas Gleixner default: 444dc186ad7SThomas Gleixner return 0; 445dc186ad7SThomas Gleixner } 446dc186ad7SThomas Gleixner } 447dc186ad7SThomas Gleixner 448dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 449dc186ad7SThomas Gleixner .name = "work_struct", 45099777288SStanislaw Gruszka .debug_hint = work_debug_hint, 451dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 452dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 453dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 454dc186ad7SThomas Gleixner }; 455dc186ad7SThomas Gleixner 456dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 457dc186ad7SThomas Gleixner { 458dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 459dc186ad7SThomas Gleixner } 460dc186ad7SThomas Gleixner 461dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 462dc186ad7SThomas Gleixner { 463dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 464dc186ad7SThomas Gleixner } 465dc186ad7SThomas Gleixner 466dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 467dc186ad7SThomas Gleixner { 468dc186ad7SThomas Gleixner if (onstack) 469dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 470dc186ad7SThomas Gleixner else 471dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 472dc186ad7SThomas Gleixner } 473dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 474dc186ad7SThomas Gleixner 475dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 476dc186ad7SThomas Gleixner { 477dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 480dc186ad7SThomas Gleixner 481dc186ad7SThomas Gleixner #else 482dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 483dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 484dc186ad7SThomas Gleixner #endif 485dc186ad7SThomas Gleixner 4869daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4879daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4889daf9e67STejun Heo { 4899daf9e67STejun Heo int ret; 4909daf9e67STejun Heo 4915bcab335STejun Heo lockdep_assert_held(&wq_mutex); 4925bcab335STejun Heo 493fa1b54e6STejun Heo do { 494fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 495fa1b54e6STejun Heo return -ENOMEM; 4969daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 497fa1b54e6STejun Heo } while (ret == -EAGAIN); 4989daf9e67STejun Heo 4999daf9e67STejun Heo return ret; 5009daf9e67STejun Heo } 5019daf9e67STejun Heo 50276af4d93STejun Heo /** 50376af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 50476af4d93STejun Heo * @wq: the target workqueue 50576af4d93STejun Heo * 506794b18bcSTejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 507794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 508794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 50976af4d93STejun Heo */ 5107fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 511a848e3b6SOleg Nesterov { 512794b18bcSTejun Heo assert_rcu_or_pwq_lock(); 51376af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 51476af4d93STejun Heo pwqs_node); 515f3421797STejun Heo } 516a848e3b6SOleg Nesterov 51773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 51873f53c4aSTejun Heo { 51973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 52073f53c4aSTejun Heo } 52173f53c4aSTejun Heo 52273f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 52373f53c4aSTejun Heo { 52473f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 52573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 52673f53c4aSTejun Heo } 52773f53c4aSTejun Heo 52873f53c4aSTejun Heo static int work_next_color(int color) 52973f53c4aSTejun Heo { 53073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 531b1f4ec17SOleg Nesterov } 532b1f4ec17SOleg Nesterov 5334594bf15SDavid Howells /* 534112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 535112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5367c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5377a22ad75STejun Heo * 538112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 539112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 540bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 541bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5427a22ad75STejun Heo * 543112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5447c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 545112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5467c3eed5cSTejun Heo * available only while the work item is queued. 547bbb68dfaSTejun Heo * 548bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 549bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 550bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 551bbb68dfaSTejun Heo * try to steal the PENDING bit. 5524594bf15SDavid Howells */ 5537a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5547a22ad75STejun Heo unsigned long flags) 5557a22ad75STejun Heo { 5566183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5577a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5587a22ad75STejun Heo } 5597a22ad75STejun Heo 560112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5614690c4abSTejun Heo unsigned long extra_flags) 562365970a1SDavid Howells { 563112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 564112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 565365970a1SDavid Howells } 566365970a1SDavid Howells 5674468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5684468a00fSLai Jiangshan int pool_id) 5694468a00fSLai Jiangshan { 5704468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5714468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5724468a00fSLai Jiangshan } 5734468a00fSLai Jiangshan 5747c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5757c3eed5cSTejun Heo int pool_id) 5764d707b9fSOleg Nesterov { 57723657bb1STejun Heo /* 57823657bb1STejun Heo * The following wmb is paired with the implied mb in 57923657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 58023657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 58123657bb1STejun Heo * owner. 58223657bb1STejun Heo */ 58323657bb1STejun Heo smp_wmb(); 5847c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5854d707b9fSOleg Nesterov } 5864d707b9fSOleg Nesterov 5877a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 588365970a1SDavid Howells { 5897c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5907c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5917a22ad75STejun Heo } 5927a22ad75STejun Heo 593112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5947a22ad75STejun Heo { 595e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5967a22ad75STejun Heo 597112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 598e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 599e120153dSTejun Heo else 600e120153dSTejun Heo return NULL; 6017a22ad75STejun Heo } 6027a22ad75STejun Heo 6037c3eed5cSTejun Heo /** 6047c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6057c3eed5cSTejun Heo * @work: the work item of interest 6067c3eed5cSTejun Heo * 6077c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 608fa1b54e6STejun Heo * 6095bcab335STejun Heo * Pools are created and destroyed under wq_mutex, and allows read access 6105bcab335STejun Heo * under sched-RCU read lock. As such, this function should be called 6115bcab335STejun Heo * under wq_mutex or with preemption disabled. 612fa1b54e6STejun Heo * 613fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 614fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 615fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 616fa1b54e6STejun Heo * returned pool is and stays online. 6177c3eed5cSTejun Heo */ 6187c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6197a22ad75STejun Heo { 620e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6217c3eed5cSTejun Heo int pool_id; 6227a22ad75STejun Heo 6235bcab335STejun Heo assert_rcu_or_wq_mutex(); 624fa1b54e6STejun Heo 625112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 626112202d9STejun Heo return ((struct pool_workqueue *) 6277c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6287a22ad75STejun Heo 6297c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6307c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6317a22ad75STejun Heo return NULL; 6327a22ad75STejun Heo 633fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6347c3eed5cSTejun Heo } 6357c3eed5cSTejun Heo 6367c3eed5cSTejun Heo /** 6377c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6387c3eed5cSTejun Heo * @work: the work item of interest 6397c3eed5cSTejun Heo * 6407c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6417c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6427c3eed5cSTejun Heo */ 6437c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6447c3eed5cSTejun Heo { 64554d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6467c3eed5cSTejun Heo 647112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 648112202d9STejun Heo return ((struct pool_workqueue *) 64954d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 65054d5b7d0SLai Jiangshan 65154d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6527c3eed5cSTejun Heo } 6537c3eed5cSTejun Heo 654bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 655bbb68dfaSTejun Heo { 6567c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 657bbb68dfaSTejun Heo 6587c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6597c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 660bbb68dfaSTejun Heo } 661bbb68dfaSTejun Heo 662bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 663bbb68dfaSTejun Heo { 664bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 665bbb68dfaSTejun Heo 666112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 667bbb68dfaSTejun Heo } 668bbb68dfaSTejun Heo 669e22bee78STejun Heo /* 6703270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6713270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 672d565ed63STejun Heo * they're being called with pool->lock held. 673e22bee78STejun Heo */ 674e22bee78STejun Heo 67563d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 676649027d7STejun Heo { 677e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 678649027d7STejun Heo } 679649027d7STejun Heo 680e22bee78STejun Heo /* 681e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 682e22bee78STejun Heo * running workers. 683974271c4STejun Heo * 684974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 685706026c2STejun Heo * function will always return %true for unbound pools as long as the 686974271c4STejun Heo * worklist isn't empty. 687e22bee78STejun Heo */ 68863d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 689e22bee78STejun Heo { 69063d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 691e22bee78STejun Heo } 692e22bee78STejun Heo 693e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 69463d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 695e22bee78STejun Heo { 69663d95a91STejun Heo return pool->nr_idle; 697e22bee78STejun Heo } 698e22bee78STejun Heo 699e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 70063d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 701e22bee78STejun Heo { 702e19e397aSTejun Heo return !list_empty(&pool->worklist) && 703e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 704e22bee78STejun Heo } 705e22bee78STejun Heo 706e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 70763d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 708e22bee78STejun Heo { 70963d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 710e22bee78STejun Heo } 711e22bee78STejun Heo 712e22bee78STejun Heo /* Do I need to be the manager? */ 71363d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 714e22bee78STejun Heo { 71563d95a91STejun Heo return need_to_create_worker(pool) || 71611ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 717e22bee78STejun Heo } 718e22bee78STejun Heo 719e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 72063d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 721e22bee78STejun Heo { 72234a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 72363d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 72463d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 725e22bee78STejun Heo 726ea1abd61SLai Jiangshan /* 727ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 728ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 729ea1abd61SLai Jiangshan */ 730ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 731ea1abd61SLai Jiangshan return false; 732ea1abd61SLai Jiangshan 733e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 734e22bee78STejun Heo } 735e22bee78STejun Heo 736e22bee78STejun Heo /* 737e22bee78STejun Heo * Wake up functions. 738e22bee78STejun Heo */ 739e22bee78STejun Heo 7407e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 74163d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7427e11629dSTejun Heo { 74363d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7447e11629dSTejun Heo return NULL; 7457e11629dSTejun Heo 74663d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7477e11629dSTejun Heo } 7487e11629dSTejun Heo 7497e11629dSTejun Heo /** 7507e11629dSTejun Heo * wake_up_worker - wake up an idle worker 75163d95a91STejun Heo * @pool: worker pool to wake worker from 7527e11629dSTejun Heo * 75363d95a91STejun Heo * Wake up the first idle worker of @pool. 7547e11629dSTejun Heo * 7557e11629dSTejun Heo * CONTEXT: 756d565ed63STejun Heo * spin_lock_irq(pool->lock). 7577e11629dSTejun Heo */ 75863d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7597e11629dSTejun Heo { 76063d95a91STejun Heo struct worker *worker = first_worker(pool); 7617e11629dSTejun Heo 7627e11629dSTejun Heo if (likely(worker)) 7637e11629dSTejun Heo wake_up_process(worker->task); 7647e11629dSTejun Heo } 7657e11629dSTejun Heo 7664690c4abSTejun Heo /** 767e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 768e22bee78STejun Heo * @task: task waking up 769e22bee78STejun Heo * @cpu: CPU @task is waking up to 770e22bee78STejun Heo * 771e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 772e22bee78STejun Heo * being awoken. 773e22bee78STejun Heo * 774e22bee78STejun Heo * CONTEXT: 775e22bee78STejun Heo * spin_lock_irq(rq->lock) 776e22bee78STejun Heo */ 777d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 778e22bee78STejun Heo { 779e22bee78STejun Heo struct worker *worker = kthread_data(task); 780e22bee78STejun Heo 78136576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 782ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 783e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 784e22bee78STejun Heo } 78536576000SJoonsoo Kim } 786e22bee78STejun Heo 787e22bee78STejun Heo /** 788e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 789e22bee78STejun Heo * @task: task going to sleep 790e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 791e22bee78STejun Heo * 792e22bee78STejun Heo * This function is called during schedule() when a busy worker is 793e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 794e22bee78STejun Heo * returning pointer to its task. 795e22bee78STejun Heo * 796e22bee78STejun Heo * CONTEXT: 797e22bee78STejun Heo * spin_lock_irq(rq->lock) 798e22bee78STejun Heo * 799e22bee78STejun Heo * RETURNS: 800e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 801e22bee78STejun Heo */ 802d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 803e22bee78STejun Heo { 804e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 805111c225aSTejun Heo struct worker_pool *pool; 806e22bee78STejun Heo 807111c225aSTejun Heo /* 808111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 809111c225aSTejun Heo * workers, also reach here, let's not access anything before 810111c225aSTejun Heo * checking NOT_RUNNING. 811111c225aSTejun Heo */ 8122d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 813e22bee78STejun Heo return NULL; 814e22bee78STejun Heo 815111c225aSTejun Heo pool = worker->pool; 816111c225aSTejun Heo 817e22bee78STejun Heo /* this can only happen on the local cpu */ 8186183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8196183c009STejun Heo return NULL; 820e22bee78STejun Heo 821e22bee78STejun Heo /* 822e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 823e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 824e22bee78STejun Heo * Please read comment there. 825e22bee78STejun Heo * 826628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 827628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 828628c78e7STejun Heo * disabled, which in turn means that none else could be 829d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 830628c78e7STejun Heo * lock is safe. 831e22bee78STejun Heo */ 832e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 833e19e397aSTejun Heo !list_empty(&pool->worklist)) 83463d95a91STejun Heo to_wakeup = first_worker(pool); 835e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 836e22bee78STejun Heo } 837e22bee78STejun Heo 838e22bee78STejun Heo /** 839e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 840cb444766STejun Heo * @worker: self 841d302f017STejun Heo * @flags: flags to set 842d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 843d302f017STejun Heo * 844e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 845e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 846e22bee78STejun Heo * woken up. 847d302f017STejun Heo * 848cb444766STejun Heo * CONTEXT: 849d565ed63STejun Heo * spin_lock_irq(pool->lock) 850d302f017STejun Heo */ 851d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 852d302f017STejun Heo bool wakeup) 853d302f017STejun Heo { 854bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 855e22bee78STejun Heo 856cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 857cb444766STejun Heo 858e22bee78STejun Heo /* 859e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 860e22bee78STejun Heo * wake up an idle worker as necessary if requested by 861e22bee78STejun Heo * @wakeup. 862e22bee78STejun Heo */ 863e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 864e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 865e22bee78STejun Heo if (wakeup) { 866e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 867bd7bdd43STejun Heo !list_empty(&pool->worklist)) 86863d95a91STejun Heo wake_up_worker(pool); 869e22bee78STejun Heo } else 870e19e397aSTejun Heo atomic_dec(&pool->nr_running); 871e22bee78STejun Heo } 872e22bee78STejun Heo 873d302f017STejun Heo worker->flags |= flags; 874d302f017STejun Heo } 875d302f017STejun Heo 876d302f017STejun Heo /** 877e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 878cb444766STejun Heo * @worker: self 879d302f017STejun Heo * @flags: flags to clear 880d302f017STejun Heo * 881e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 882d302f017STejun Heo * 883cb444766STejun Heo * CONTEXT: 884d565ed63STejun Heo * spin_lock_irq(pool->lock) 885d302f017STejun Heo */ 886d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 887d302f017STejun Heo { 88863d95a91STejun Heo struct worker_pool *pool = worker->pool; 889e22bee78STejun Heo unsigned int oflags = worker->flags; 890e22bee78STejun Heo 891cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 892cb444766STejun Heo 893d302f017STejun Heo worker->flags &= ~flags; 894e22bee78STejun Heo 89542c025f3STejun Heo /* 89642c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 89742c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 89842c025f3STejun Heo * of multiple flags, not a single flag. 89942c025f3STejun Heo */ 900e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 901e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 902e19e397aSTejun Heo atomic_inc(&pool->nr_running); 903d302f017STejun Heo } 904d302f017STejun Heo 905d302f017STejun Heo /** 9068cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 907c9e7cf27STejun Heo * @pool: pool of interest 9088cca0eeaSTejun Heo * @work: work to find worker for 9098cca0eeaSTejun Heo * 910c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 911c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 912a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 913a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 914a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 915a2c1c57bSTejun Heo * being executed. 916a2c1c57bSTejun Heo * 917a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 918a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 919a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 920a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 921a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 922a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 923a2c1c57bSTejun Heo * 924c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 925c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 926c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 927c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 928c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 929c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9308cca0eeaSTejun Heo * 9318cca0eeaSTejun Heo * CONTEXT: 932d565ed63STejun Heo * spin_lock_irq(pool->lock). 9338cca0eeaSTejun Heo * 9348cca0eeaSTejun Heo * RETURNS: 9358cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9368cca0eeaSTejun Heo * otherwise. 9378cca0eeaSTejun Heo */ 938c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9398cca0eeaSTejun Heo struct work_struct *work) 9408cca0eeaSTejun Heo { 94142f8570fSSasha Levin struct worker *worker; 94242f8570fSSasha Levin 943b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 944a2c1c57bSTejun Heo (unsigned long)work) 945a2c1c57bSTejun Heo if (worker->current_work == work && 946a2c1c57bSTejun Heo worker->current_func == work->func) 94742f8570fSSasha Levin return worker; 94842f8570fSSasha Levin 94942f8570fSSasha Levin return NULL; 9508cca0eeaSTejun Heo } 9518cca0eeaSTejun Heo 9528cca0eeaSTejun Heo /** 953bf4ede01STejun Heo * move_linked_works - move linked works to a list 954bf4ede01STejun Heo * @work: start of series of works to be scheduled 955bf4ede01STejun Heo * @head: target list to append @work to 956bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 957bf4ede01STejun Heo * 958bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 959bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 960bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 961bf4ede01STejun Heo * 962bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 963bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 964bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 965bf4ede01STejun Heo * 966bf4ede01STejun Heo * CONTEXT: 967d565ed63STejun Heo * spin_lock_irq(pool->lock). 968bf4ede01STejun Heo */ 969bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 970bf4ede01STejun Heo struct work_struct **nextp) 971bf4ede01STejun Heo { 972bf4ede01STejun Heo struct work_struct *n; 973bf4ede01STejun Heo 974bf4ede01STejun Heo /* 975bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 976bf4ede01STejun Heo * use NULL for list head. 977bf4ede01STejun Heo */ 978bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 979bf4ede01STejun Heo list_move_tail(&work->entry, head); 980bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 981bf4ede01STejun Heo break; 982bf4ede01STejun Heo } 983bf4ede01STejun Heo 984bf4ede01STejun Heo /* 985bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 986bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 987bf4ede01STejun Heo * needs to be updated. 988bf4ede01STejun Heo */ 989bf4ede01STejun Heo if (nextp) 990bf4ede01STejun Heo *nextp = n; 991bf4ede01STejun Heo } 992bf4ede01STejun Heo 9938864b4e5STejun Heo /** 9948864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 9958864b4e5STejun Heo * @pwq: pool_workqueue to get 9968864b4e5STejun Heo * 9978864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 9988864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 9998864b4e5STejun Heo */ 10008864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10018864b4e5STejun Heo { 10028864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10038864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10048864b4e5STejun Heo pwq->refcnt++; 10058864b4e5STejun Heo } 10068864b4e5STejun Heo 10078864b4e5STejun Heo /** 10088864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10098864b4e5STejun Heo * @pwq: pool_workqueue to put 10108864b4e5STejun Heo * 10118864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10128864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10138864b4e5STejun Heo */ 10148864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10158864b4e5STejun Heo { 10168864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10178864b4e5STejun Heo if (likely(--pwq->refcnt)) 10188864b4e5STejun Heo return; 10198864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10208864b4e5STejun Heo return; 10218864b4e5STejun Heo /* 10228864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10238864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10248864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10258864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10268864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10278864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10288864b4e5STejun Heo */ 10298864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10308864b4e5STejun Heo } 10318864b4e5STejun Heo 1032112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1033bf4ede01STejun Heo { 1034112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1035bf4ede01STejun Heo 1036bf4ede01STejun Heo trace_workqueue_activate_work(work); 1037112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1038bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1039112202d9STejun Heo pwq->nr_active++; 1040bf4ede01STejun Heo } 1041bf4ede01STejun Heo 1042112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10433aa62497SLai Jiangshan { 1044112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10453aa62497SLai Jiangshan struct work_struct, entry); 10463aa62497SLai Jiangshan 1047112202d9STejun Heo pwq_activate_delayed_work(work); 10483aa62497SLai Jiangshan } 10493aa62497SLai Jiangshan 1050bf4ede01STejun Heo /** 1051112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1052112202d9STejun Heo * @pwq: pwq of interest 1053bf4ede01STejun Heo * @color: color of work which left the queue 1054bf4ede01STejun Heo * 1055bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1056112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1057bf4ede01STejun Heo * 1058bf4ede01STejun Heo * CONTEXT: 1059d565ed63STejun Heo * spin_lock_irq(pool->lock). 1060bf4ede01STejun Heo */ 1061112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1062bf4ede01STejun Heo { 10638864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1064bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10658864b4e5STejun Heo goto out_put; 1066bf4ede01STejun Heo 1067112202d9STejun Heo pwq->nr_in_flight[color]--; 1068bf4ede01STejun Heo 1069112202d9STejun Heo pwq->nr_active--; 1070112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1071bf4ede01STejun Heo /* one down, submit a delayed one */ 1072112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1073112202d9STejun Heo pwq_activate_first_delayed(pwq); 1074bf4ede01STejun Heo } 1075bf4ede01STejun Heo 1076bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1077112202d9STejun Heo if (likely(pwq->flush_color != color)) 10788864b4e5STejun Heo goto out_put; 1079bf4ede01STejun Heo 1080bf4ede01STejun Heo /* are there still in-flight works? */ 1081112202d9STejun Heo if (pwq->nr_in_flight[color]) 10828864b4e5STejun Heo goto out_put; 1083bf4ede01STejun Heo 1084112202d9STejun Heo /* this pwq is done, clear flush_color */ 1085112202d9STejun Heo pwq->flush_color = -1; 1086bf4ede01STejun Heo 1087bf4ede01STejun Heo /* 1088112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1089bf4ede01STejun Heo * will handle the rest. 1090bf4ede01STejun Heo */ 1091112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1092112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 10938864b4e5STejun Heo out_put: 10948864b4e5STejun Heo put_pwq(pwq); 1095bf4ede01STejun Heo } 1096bf4ede01STejun Heo 109736e227d2STejun Heo /** 1098bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 109936e227d2STejun Heo * @work: work item to steal 110036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1101bbb68dfaSTejun Heo * @flags: place to store irq state 110236e227d2STejun Heo * 110336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 110436e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 110536e227d2STejun Heo * 110636e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 110736e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 110836e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1109bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1110bbb68dfaSTejun Heo * for arbitrarily long 111136e227d2STejun Heo * 1112bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1113e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1114e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1115e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1116bbb68dfaSTejun Heo * 1117bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1118bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1119bbb68dfaSTejun Heo * 1120e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1121bf4ede01STejun Heo */ 1122bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1123bbb68dfaSTejun Heo unsigned long *flags) 1124bf4ede01STejun Heo { 1125d565ed63STejun Heo struct worker_pool *pool; 1126112202d9STejun Heo struct pool_workqueue *pwq; 1127bf4ede01STejun Heo 1128bbb68dfaSTejun Heo local_irq_save(*flags); 1129bbb68dfaSTejun Heo 113036e227d2STejun Heo /* try to steal the timer if it exists */ 113136e227d2STejun Heo if (is_dwork) { 113236e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 113336e227d2STejun Heo 1134e0aecdd8STejun Heo /* 1135e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1136e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1137e0aecdd8STejun Heo * running on the local CPU. 1138e0aecdd8STejun Heo */ 113936e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 114036e227d2STejun Heo return 1; 114136e227d2STejun Heo } 114236e227d2STejun Heo 114336e227d2STejun Heo /* try to claim PENDING the normal way */ 1144bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1145bf4ede01STejun Heo return 0; 1146bf4ede01STejun Heo 1147bf4ede01STejun Heo /* 1148bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1149bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1150bf4ede01STejun Heo */ 1151d565ed63STejun Heo pool = get_work_pool(work); 1152d565ed63STejun Heo if (!pool) 1153bbb68dfaSTejun Heo goto fail; 1154bf4ede01STejun Heo 1155d565ed63STejun Heo spin_lock(&pool->lock); 1156bf4ede01STejun Heo /* 1157112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1158112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1159112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1160112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1161112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11620b3dae68SLai Jiangshan * item is currently queued on that pool. 1163bf4ede01STejun Heo */ 1164112202d9STejun Heo pwq = get_work_pwq(work); 1165112202d9STejun Heo if (pwq && pwq->pool == pool) { 1166bf4ede01STejun Heo debug_work_deactivate(work); 11673aa62497SLai Jiangshan 11683aa62497SLai Jiangshan /* 116916062836STejun Heo * A delayed work item cannot be grabbed directly because 117016062836STejun Heo * it might have linked NO_COLOR work items which, if left 1171112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 117216062836STejun Heo * management later on and cause stall. Make sure the work 117316062836STejun Heo * item is activated before grabbing. 11743aa62497SLai Jiangshan */ 11753aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1176112202d9STejun Heo pwq_activate_delayed_work(work); 11773aa62497SLai Jiangshan 1178bf4ede01STejun Heo list_del_init(&work->entry); 1179112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 118036e227d2STejun Heo 1181112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11824468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11834468a00fSLai Jiangshan 1184d565ed63STejun Heo spin_unlock(&pool->lock); 118536e227d2STejun Heo return 1; 1186bf4ede01STejun Heo } 1187d565ed63STejun Heo spin_unlock(&pool->lock); 1188bbb68dfaSTejun Heo fail: 1189bbb68dfaSTejun Heo local_irq_restore(*flags); 1190bbb68dfaSTejun Heo if (work_is_canceling(work)) 1191bbb68dfaSTejun Heo return -ENOENT; 1192bbb68dfaSTejun Heo cpu_relax(); 119336e227d2STejun Heo return -EAGAIN; 1194bf4ede01STejun Heo } 1195bf4ede01STejun Heo 1196bf4ede01STejun Heo /** 1197706026c2STejun Heo * insert_work - insert a work into a pool 1198112202d9STejun Heo * @pwq: pwq @work belongs to 11994690c4abSTejun Heo * @work: work to insert 12004690c4abSTejun Heo * @head: insertion point 12014690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12024690c4abSTejun Heo * 1203112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1204706026c2STejun Heo * work_struct flags. 12054690c4abSTejun Heo * 12064690c4abSTejun Heo * CONTEXT: 1207d565ed63STejun Heo * spin_lock_irq(pool->lock). 1208365970a1SDavid Howells */ 1209112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1210112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1211b89deed3SOleg Nesterov { 1212112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1213e1d8aa9fSFrederic Weisbecker 12144690c4abSTejun Heo /* we own @work, set data and link */ 1215112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12161a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12178864b4e5STejun Heo get_pwq(pwq); 1218e22bee78STejun Heo 1219e22bee78STejun Heo /* 1220c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1221c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1222c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1223e22bee78STejun Heo */ 1224e22bee78STejun Heo smp_mb(); 1225e22bee78STejun Heo 122663d95a91STejun Heo if (__need_more_worker(pool)) 122763d95a91STejun Heo wake_up_worker(pool); 1228b89deed3SOleg Nesterov } 1229b89deed3SOleg Nesterov 1230c8efcc25STejun Heo /* 1231c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12328d03ecfeSTejun Heo * same workqueue. 1233c8efcc25STejun Heo */ 1234c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1235c8efcc25STejun Heo { 1236c8efcc25STejun Heo struct worker *worker; 1237c8efcc25STejun Heo 12388d03ecfeSTejun Heo worker = current_wq_worker(); 1239c8efcc25STejun Heo /* 12408d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12418d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1242c8efcc25STejun Heo */ 1243112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1244c8efcc25STejun Heo } 1245c8efcc25STejun Heo 1246d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12471da177e4SLinus Torvalds struct work_struct *work) 12481da177e4SLinus Torvalds { 1249112202d9STejun Heo struct pool_workqueue *pwq; 1250c9178087STejun Heo struct worker_pool *last_pool; 12511e19ffc6STejun Heo struct list_head *worklist; 12528a2e8e5dSTejun Heo unsigned int work_flags; 1253b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12548930cabaSTejun Heo 12558930cabaSTejun Heo /* 12568930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12578930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12588930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12598930cabaSTejun Heo * happen with IRQ disabled. 12608930cabaSTejun Heo */ 12618930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12621da177e4SLinus Torvalds 1263dc186ad7SThomas Gleixner debug_work_activate(work); 12641e19ffc6STejun Heo 1265c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1266618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1267c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1268e41e704bSTejun Heo return; 12699e8cd2f5STejun Heo retry: 1270c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1271c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 127257469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1273f3421797STejun Heo cpu = raw_smp_processor_id(); 1274c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1275c9178087STejun Heo } else { 1276c9178087STejun Heo pwq = first_pwq(wq); 1277c9178087STejun Heo } 1278f3421797STejun Heo 127918aa9effSTejun Heo /* 1280c9178087STejun Heo * If @work was previously on a different pool, it might still be 1281c9178087STejun Heo * running there, in which case the work needs to be queued on that 1282c9178087STejun Heo * pool to guarantee non-reentrancy. 128318aa9effSTejun Heo */ 1284c9e7cf27STejun Heo last_pool = get_work_pool(work); 1285112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 128618aa9effSTejun Heo struct worker *worker; 128718aa9effSTejun Heo 1288d565ed63STejun Heo spin_lock(&last_pool->lock); 128918aa9effSTejun Heo 1290c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 129118aa9effSTejun Heo 1292112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1293c9178087STejun Heo pwq = worker->current_pwq; 12948594fadeSLai Jiangshan } else { 129518aa9effSTejun Heo /* meh... not running there, queue here */ 1296d565ed63STejun Heo spin_unlock(&last_pool->lock); 1297112202d9STejun Heo spin_lock(&pwq->pool->lock); 129818aa9effSTejun Heo } 12998930cabaSTejun Heo } else { 1300112202d9STejun Heo spin_lock(&pwq->pool->lock); 13018930cabaSTejun Heo } 1302502ca9d8STejun Heo 13039e8cd2f5STejun Heo /* 13049e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13059e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13069e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 13079e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 13089e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 13099e8cd2f5STejun Heo * make forward-progress. 13109e8cd2f5STejun Heo */ 13119e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13129e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13139e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13149e8cd2f5STejun Heo cpu_relax(); 13159e8cd2f5STejun Heo goto retry; 13169e8cd2f5STejun Heo } 13179e8cd2f5STejun Heo /* oops */ 13189e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13199e8cd2f5STejun Heo wq->name, cpu); 13209e8cd2f5STejun Heo } 13219e8cd2f5STejun Heo 1322112202d9STejun Heo /* pwq determined, queue */ 1323112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1324502ca9d8STejun Heo 1325f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1326112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1327f5b2552bSDan Carpenter return; 1328f5b2552bSDan Carpenter } 13291e19ffc6STejun Heo 1330112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1331112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13321e19ffc6STejun Heo 1333112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1334cdadf009STejun Heo trace_workqueue_activate_work(work); 1335112202d9STejun Heo pwq->nr_active++; 1336112202d9STejun Heo worklist = &pwq->pool->worklist; 13378a2e8e5dSTejun Heo } else { 13388a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1339112202d9STejun Heo worklist = &pwq->delayed_works; 13408a2e8e5dSTejun Heo } 13411e19ffc6STejun Heo 1342112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13431e19ffc6STejun Heo 1344112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13451da177e4SLinus Torvalds } 13461da177e4SLinus Torvalds 13470fcb78c2SRolf Eike Beer /** 1348c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1349c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1350c1a220e7SZhang Rui * @wq: workqueue to use 1351c1a220e7SZhang Rui * @work: work to queue 1352c1a220e7SZhang Rui * 1353d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1354c1a220e7SZhang Rui * 1355c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1356c1a220e7SZhang Rui * can't go away. 1357c1a220e7SZhang Rui */ 1358d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1359d4283e93STejun Heo struct work_struct *work) 1360c1a220e7SZhang Rui { 1361d4283e93STejun Heo bool ret = false; 13628930cabaSTejun Heo unsigned long flags; 13638930cabaSTejun Heo 13648930cabaSTejun Heo local_irq_save(flags); 1365c1a220e7SZhang Rui 136622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13674690c4abSTejun Heo __queue_work(cpu, wq, work); 1368d4283e93STejun Heo ret = true; 1369c1a220e7SZhang Rui } 13708930cabaSTejun Heo 13718930cabaSTejun Heo local_irq_restore(flags); 1372c1a220e7SZhang Rui return ret; 1373c1a220e7SZhang Rui } 1374c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1375c1a220e7SZhang Rui 1376d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13771da177e4SLinus Torvalds { 137852bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13791da177e4SLinus Torvalds 1380e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 138160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13821da177e4SLinus Torvalds } 13831438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13841da177e4SLinus Torvalds 13857beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 138652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13871da177e4SLinus Torvalds { 13887beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13897beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13901da177e4SLinus Torvalds 13917beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13927beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1393fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1394fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13957beb2edfSTejun Heo 13968852aac2STejun Heo /* 13978852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13988852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13998852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14008852aac2STejun Heo * on that there's no such delay when @delay is 0. 14018852aac2STejun Heo */ 14028852aac2STejun Heo if (!delay) { 14038852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14048852aac2STejun Heo return; 14058852aac2STejun Heo } 14068852aac2STejun Heo 14077beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14087beb2edfSTejun Heo 140960c057bcSLai Jiangshan dwork->wq = wq; 14101265057fSTejun Heo dwork->cpu = cpu; 14117beb2edfSTejun Heo timer->expires = jiffies + delay; 14127beb2edfSTejun Heo 14137beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14147beb2edfSTejun Heo add_timer_on(timer, cpu); 14157beb2edfSTejun Heo else 14167beb2edfSTejun Heo add_timer(timer); 14177beb2edfSTejun Heo } 14181da177e4SLinus Torvalds 14190fcb78c2SRolf Eike Beer /** 14200fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14210fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14220fcb78c2SRolf Eike Beer * @wq: workqueue to use 1423af9997e4SRandy Dunlap * @dwork: work to queue 14240fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14250fcb78c2SRolf Eike Beer * 1426715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1427715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1428715f1300STejun Heo * execution. 14290fcb78c2SRolf Eike Beer */ 1430d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 143152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14327a6bc1cdSVenkatesh Pallipadi { 143352bad64dSDavid Howells struct work_struct *work = &dwork->work; 1434d4283e93STejun Heo bool ret = false; 14358930cabaSTejun Heo unsigned long flags; 14368930cabaSTejun Heo 14378930cabaSTejun Heo /* read the comment in __queue_work() */ 14388930cabaSTejun Heo local_irq_save(flags); 14397a6bc1cdSVenkatesh Pallipadi 144022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14417beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1442d4283e93STejun Heo ret = true; 14437a6bc1cdSVenkatesh Pallipadi } 14448930cabaSTejun Heo 14458930cabaSTejun Heo local_irq_restore(flags); 14467a6bc1cdSVenkatesh Pallipadi return ret; 14477a6bc1cdSVenkatesh Pallipadi } 1448ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14491da177e4SLinus Torvalds 1450c8e55f36STejun Heo /** 14518376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14528376fe22STejun Heo * @cpu: CPU number to execute work on 14538376fe22STejun Heo * @wq: workqueue to use 14548376fe22STejun Heo * @dwork: work to queue 14558376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14568376fe22STejun Heo * 14578376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14588376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14598376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14608376fe22STejun Heo * current state. 14618376fe22STejun Heo * 14628376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14638376fe22STejun Heo * pending and its timer was modified. 14648376fe22STejun Heo * 1465e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14668376fe22STejun Heo * See try_to_grab_pending() for details. 14678376fe22STejun Heo */ 14688376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14698376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14708376fe22STejun Heo { 14718376fe22STejun Heo unsigned long flags; 14728376fe22STejun Heo int ret; 14738376fe22STejun Heo 14748376fe22STejun Heo do { 14758376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14768376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14778376fe22STejun Heo 14788376fe22STejun Heo if (likely(ret >= 0)) { 14798376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14808376fe22STejun Heo local_irq_restore(flags); 14818376fe22STejun Heo } 14828376fe22STejun Heo 14838376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14848376fe22STejun Heo return ret; 14858376fe22STejun Heo } 14868376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14878376fe22STejun Heo 14888376fe22STejun Heo /** 1489c8e55f36STejun Heo * worker_enter_idle - enter idle state 1490c8e55f36STejun Heo * @worker: worker which is entering idle state 1491c8e55f36STejun Heo * 1492c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1493c8e55f36STejun Heo * necessary. 1494c8e55f36STejun Heo * 1495c8e55f36STejun Heo * LOCKING: 1496d565ed63STejun Heo * spin_lock_irq(pool->lock). 1497c8e55f36STejun Heo */ 1498c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14991da177e4SLinus Torvalds { 1500bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1501c8e55f36STejun Heo 15026183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15036183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15046183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15056183c009STejun Heo return; 1506c8e55f36STejun Heo 1507cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1508cb444766STejun Heo worker->flags |= WORKER_IDLE; 1509bd7bdd43STejun Heo pool->nr_idle++; 1510e22bee78STejun Heo worker->last_active = jiffies; 1511c8e55f36STejun Heo 1512c8e55f36STejun Heo /* idle_list is LIFO */ 1513bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1514db7bccf4STejun Heo 151563d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1516628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1517cb444766STejun Heo 1518544ecf31STejun Heo /* 1519706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1520d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1521628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1522628c78e7STejun Heo * unbind is not in progress. 1523544ecf31STejun Heo */ 152424647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1525bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1526e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1527c8e55f36STejun Heo } 1528c8e55f36STejun Heo 1529c8e55f36STejun Heo /** 1530c8e55f36STejun Heo * worker_leave_idle - leave idle state 1531c8e55f36STejun Heo * @worker: worker which is leaving idle state 1532c8e55f36STejun Heo * 1533c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1534c8e55f36STejun Heo * 1535c8e55f36STejun Heo * LOCKING: 1536d565ed63STejun Heo * spin_lock_irq(pool->lock). 1537c8e55f36STejun Heo */ 1538c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1539c8e55f36STejun Heo { 1540bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1541c8e55f36STejun Heo 15426183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15436183c009STejun Heo return; 1544d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1545bd7bdd43STejun Heo pool->nr_idle--; 1546c8e55f36STejun Heo list_del_init(&worker->entry); 1547c8e55f36STejun Heo } 1548c8e55f36STejun Heo 1549e22bee78STejun Heo /** 1550f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1551f36dc67bSLai Jiangshan * @pool: target worker_pool 1552f36dc67bSLai Jiangshan * 1553f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1554e22bee78STejun Heo * 1555e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1556e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1557e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1558e22bee78STejun Heo * guaranteed to execute on the cpu. 1559e22bee78STejun Heo * 1560f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1561e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1562e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1563e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1564706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1565e22bee78STejun Heo * [dis]associated in the meantime. 1566e22bee78STejun Heo * 1567706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 156824647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1569f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1570f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1571f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1572e22bee78STejun Heo * 1573e22bee78STejun Heo * CONTEXT: 1574d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1575e22bee78STejun Heo * held. 1576e22bee78STejun Heo * 1577e22bee78STejun Heo * RETURNS: 1578706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1579e22bee78STejun Heo * bound), %false if offline. 1580e22bee78STejun Heo */ 1581f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1582d565ed63STejun Heo __acquires(&pool->lock) 1583e22bee78STejun Heo { 1584e22bee78STejun Heo while (true) { 1585e22bee78STejun Heo /* 1586e22bee78STejun Heo * The following call may fail, succeed or succeed 1587e22bee78STejun Heo * without actually migrating the task to the cpu if 1588e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 158924647570STejun Heo * against POOL_DISASSOCIATED. 1590e22bee78STejun Heo */ 159124647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 15927a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1593e22bee78STejun Heo 1594d565ed63STejun Heo spin_lock_irq(&pool->lock); 159524647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1596e22bee78STejun Heo return false; 1597f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 15987a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1599e22bee78STejun Heo return true; 1600d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1601e22bee78STejun Heo 16025035b20fSTejun Heo /* 16035035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16045035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16055035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16065035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16075035b20fSTejun Heo */ 1608e22bee78STejun Heo cpu_relax(); 16095035b20fSTejun Heo cond_resched(); 1610e22bee78STejun Heo } 1611e22bee78STejun Heo } 1612e22bee78STejun Heo 1613c34056a3STejun Heo static struct worker *alloc_worker(void) 1614c34056a3STejun Heo { 1615c34056a3STejun Heo struct worker *worker; 1616c34056a3STejun Heo 1617c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1618c8e55f36STejun Heo if (worker) { 1619c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1620affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1621e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1622e22bee78STejun Heo worker->flags = WORKER_PREP; 1623c8e55f36STejun Heo } 1624c34056a3STejun Heo return worker; 1625c34056a3STejun Heo } 1626c34056a3STejun Heo 1627c34056a3STejun Heo /** 1628c34056a3STejun Heo * create_worker - create a new workqueue worker 162963d95a91STejun Heo * @pool: pool the new worker will belong to 1630c34056a3STejun Heo * 163163d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1632c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1633c34056a3STejun Heo * destroy_worker(). 1634c34056a3STejun Heo * 1635c34056a3STejun Heo * CONTEXT: 1636c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1637c34056a3STejun Heo * 1638c34056a3STejun Heo * RETURNS: 1639c34056a3STejun Heo * Pointer to the newly created worker. 1640c34056a3STejun Heo */ 1641bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1642c34056a3STejun Heo { 16437a4e344cSTejun Heo const char *pri = pool->attrs->nice < 0 ? "H" : ""; 1644c34056a3STejun Heo struct worker *worker = NULL; 1645f3421797STejun Heo int id = -1; 1646c34056a3STejun Heo 1647cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1648cd549687STejun Heo 1649822d8405STejun Heo /* 1650822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1651822d8405STejun Heo * without installing the pointer. 1652822d8405STejun Heo */ 1653822d8405STejun Heo idr_preload(GFP_KERNEL); 1654d565ed63STejun Heo spin_lock_irq(&pool->lock); 1655822d8405STejun Heo 1656822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1657822d8405STejun Heo 1658d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1659822d8405STejun Heo idr_preload_end(); 1660822d8405STejun Heo if (id < 0) 1661c34056a3STejun Heo goto fail; 1662c34056a3STejun Heo 1663c34056a3STejun Heo worker = alloc_worker(); 1664c34056a3STejun Heo if (!worker) 1665c34056a3STejun Heo goto fail; 1666c34056a3STejun Heo 1667bd7bdd43STejun Heo worker->pool = pool; 1668c34056a3STejun Heo worker->id = id; 1669c34056a3STejun Heo 167029c91e99STejun Heo if (pool->cpu >= 0) 167194dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1672ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1673d84ff051STejun Heo "kworker/%d:%d%s", pool->cpu, id, pri); 1674f3421797STejun Heo else 1675f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1676ac6104cdSTejun Heo "kworker/u%d:%d%s", 1677ac6104cdSTejun Heo pool->id, id, pri); 1678c34056a3STejun Heo if (IS_ERR(worker->task)) 1679c34056a3STejun Heo goto fail; 1680c34056a3STejun Heo 1681c5aa87bbSTejun Heo /* 1682c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1683c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1684c5aa87bbSTejun Heo */ 16857a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 16867a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16873270476aSTejun Heo 168814a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 168914a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 16907a4e344cSTejun Heo 16917a4e344cSTejun Heo /* 16927a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 16937a4e344cSTejun Heo * remains stable across this function. See the comments above the 16947a4e344cSTejun Heo * flag definition for details. 16957a4e344cSTejun Heo */ 16967a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1697f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1698c34056a3STejun Heo 1699822d8405STejun Heo /* successful, commit the pointer to idr */ 1700822d8405STejun Heo spin_lock_irq(&pool->lock); 1701822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1702822d8405STejun Heo spin_unlock_irq(&pool->lock); 1703822d8405STejun Heo 1704c34056a3STejun Heo return worker; 1705822d8405STejun Heo 1706c34056a3STejun Heo fail: 1707c34056a3STejun Heo if (id >= 0) { 1708d565ed63STejun Heo spin_lock_irq(&pool->lock); 1709822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1710d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1711c34056a3STejun Heo } 1712c34056a3STejun Heo kfree(worker); 1713c34056a3STejun Heo return NULL; 1714c34056a3STejun Heo } 1715c34056a3STejun Heo 1716c34056a3STejun Heo /** 1717c34056a3STejun Heo * start_worker - start a newly created worker 1718c34056a3STejun Heo * @worker: worker to start 1719c34056a3STejun Heo * 1720706026c2STejun Heo * Make the pool aware of @worker and start it. 1721c34056a3STejun Heo * 1722c34056a3STejun Heo * CONTEXT: 1723d565ed63STejun Heo * spin_lock_irq(pool->lock). 1724c34056a3STejun Heo */ 1725c34056a3STejun Heo static void start_worker(struct worker *worker) 1726c34056a3STejun Heo { 1727cb444766STejun Heo worker->flags |= WORKER_STARTED; 1728bd7bdd43STejun Heo worker->pool->nr_workers++; 1729c8e55f36STejun Heo worker_enter_idle(worker); 1730c34056a3STejun Heo wake_up_process(worker->task); 1731c34056a3STejun Heo } 1732c34056a3STejun Heo 1733c34056a3STejun Heo /** 1734ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1735ebf44d16STejun Heo * @pool: the target pool 1736ebf44d16STejun Heo * 1737cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1738ebf44d16STejun Heo */ 1739ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1740ebf44d16STejun Heo { 1741ebf44d16STejun Heo struct worker *worker; 1742ebf44d16STejun Heo 1743cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1744cd549687STejun Heo 1745ebf44d16STejun Heo worker = create_worker(pool); 1746ebf44d16STejun Heo if (worker) { 1747ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1748ebf44d16STejun Heo start_worker(worker); 1749ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1750ebf44d16STejun Heo } 1751ebf44d16STejun Heo 1752cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1753cd549687STejun Heo 1754ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1755ebf44d16STejun Heo } 1756ebf44d16STejun Heo 1757ebf44d16STejun Heo /** 1758c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1759c34056a3STejun Heo * @worker: worker to be destroyed 1760c34056a3STejun Heo * 1761706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1762c8e55f36STejun Heo * 1763c8e55f36STejun Heo * CONTEXT: 1764d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1765c34056a3STejun Heo */ 1766c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1767c34056a3STejun Heo { 1768bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1769c34056a3STejun Heo 1770cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1771cd549687STejun Heo lockdep_assert_held(&pool->lock); 1772cd549687STejun Heo 1773c34056a3STejun Heo /* sanity check frenzy */ 17746183c009STejun Heo if (WARN_ON(worker->current_work) || 17756183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 17766183c009STejun Heo return; 1777c34056a3STejun Heo 1778c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1779bd7bdd43STejun Heo pool->nr_workers--; 1780c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1781bd7bdd43STejun Heo pool->nr_idle--; 1782c8e55f36STejun Heo 1783c8e55f36STejun Heo list_del_init(&worker->entry); 1784cb444766STejun Heo worker->flags |= WORKER_DIE; 1785c8e55f36STejun Heo 1786822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1787822d8405STejun Heo 1788d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1789c8e55f36STejun Heo 1790c34056a3STejun Heo kthread_stop(worker->task); 1791c34056a3STejun Heo kfree(worker); 1792c34056a3STejun Heo 1793d565ed63STejun Heo spin_lock_irq(&pool->lock); 1794c34056a3STejun Heo } 1795c34056a3STejun Heo 179663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1797e22bee78STejun Heo { 179863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1799e22bee78STejun Heo 1800d565ed63STejun Heo spin_lock_irq(&pool->lock); 1801e22bee78STejun Heo 180263d95a91STejun Heo if (too_many_workers(pool)) { 1803e22bee78STejun Heo struct worker *worker; 1804e22bee78STejun Heo unsigned long expires; 1805e22bee78STejun Heo 1806e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 180763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1808e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1809e22bee78STejun Heo 1810e22bee78STejun Heo if (time_before(jiffies, expires)) 181163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1812e22bee78STejun Heo else { 1813e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 181411ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 181563d95a91STejun Heo wake_up_worker(pool); 1816e22bee78STejun Heo } 1817e22bee78STejun Heo } 1818e22bee78STejun Heo 1819d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1820e22bee78STejun Heo } 1821e22bee78STejun Heo 1822493a1724STejun Heo static void send_mayday(struct work_struct *work) 1823e22bee78STejun Heo { 1824112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1825112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1826493a1724STejun Heo 18272e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1828e22bee78STejun Heo 1829493008a8STejun Heo if (!wq->rescuer) 1830493a1724STejun Heo return; 1831e22bee78STejun Heo 1832e22bee78STejun Heo /* mayday mayday mayday */ 1833493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1834493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1835e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1836493a1724STejun Heo } 1837e22bee78STejun Heo } 1838e22bee78STejun Heo 1839706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1840e22bee78STejun Heo { 184163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1842e22bee78STejun Heo struct work_struct *work; 1843e22bee78STejun Heo 18442e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1845493a1724STejun Heo spin_lock(&pool->lock); 1846e22bee78STejun Heo 184763d95a91STejun Heo if (need_to_create_worker(pool)) { 1848e22bee78STejun Heo /* 1849e22bee78STejun Heo * We've been trying to create a new worker but 1850e22bee78STejun Heo * haven't been successful. We might be hitting an 1851e22bee78STejun Heo * allocation deadlock. Send distress signals to 1852e22bee78STejun Heo * rescuers. 1853e22bee78STejun Heo */ 185463d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1855e22bee78STejun Heo send_mayday(work); 1856e22bee78STejun Heo } 1857e22bee78STejun Heo 1858493a1724STejun Heo spin_unlock(&pool->lock); 18592e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1860e22bee78STejun Heo 186163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1862e22bee78STejun Heo } 1863e22bee78STejun Heo 1864e22bee78STejun Heo /** 1865e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 186663d95a91STejun Heo * @pool: pool to create a new worker for 1867e22bee78STejun Heo * 186863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1869e22bee78STejun Heo * have at least one idle worker on return from this function. If 1870e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 187163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1872e22bee78STejun Heo * possible allocation deadlock. 1873e22bee78STejun Heo * 1874c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1875c5aa87bbSTejun Heo * may_start_working() %true. 1876e22bee78STejun Heo * 1877e22bee78STejun Heo * LOCKING: 1878d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1879e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1880e22bee78STejun Heo * manager. 1881e22bee78STejun Heo * 1882e22bee78STejun Heo * RETURNS: 1883c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1884e22bee78STejun Heo * otherwise. 1885e22bee78STejun Heo */ 188663d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1887d565ed63STejun Heo __releases(&pool->lock) 1888d565ed63STejun Heo __acquires(&pool->lock) 1889e22bee78STejun Heo { 189063d95a91STejun Heo if (!need_to_create_worker(pool)) 1891e22bee78STejun Heo return false; 1892e22bee78STejun Heo restart: 1893d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18949f9c2364STejun Heo 1895e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 189663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1897e22bee78STejun Heo 1898e22bee78STejun Heo while (true) { 1899e22bee78STejun Heo struct worker *worker; 1900e22bee78STejun Heo 1901bc2ae0f5STejun Heo worker = create_worker(pool); 1902e22bee78STejun Heo if (worker) { 190363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1904d565ed63STejun Heo spin_lock_irq(&pool->lock); 1905e22bee78STejun Heo start_worker(worker); 19066183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19076183c009STejun Heo goto restart; 1908e22bee78STejun Heo return true; 1909e22bee78STejun Heo } 1910e22bee78STejun Heo 191163d95a91STejun Heo if (!need_to_create_worker(pool)) 1912e22bee78STejun Heo break; 1913e22bee78STejun Heo 1914e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1915e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19169f9c2364STejun Heo 191763d95a91STejun Heo if (!need_to_create_worker(pool)) 1918e22bee78STejun Heo break; 1919e22bee78STejun Heo } 1920e22bee78STejun Heo 192163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1922d565ed63STejun Heo spin_lock_irq(&pool->lock); 192363d95a91STejun Heo if (need_to_create_worker(pool)) 1924e22bee78STejun Heo goto restart; 1925e22bee78STejun Heo return true; 1926e22bee78STejun Heo } 1927e22bee78STejun Heo 1928e22bee78STejun Heo /** 1929e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 193063d95a91STejun Heo * @pool: pool to destroy workers for 1931e22bee78STejun Heo * 193263d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1933e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1934e22bee78STejun Heo * 1935e22bee78STejun Heo * LOCKING: 1936d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1937e22bee78STejun Heo * multiple times. Called only from manager. 1938e22bee78STejun Heo * 1939e22bee78STejun Heo * RETURNS: 1940c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1941e22bee78STejun Heo * otherwise. 1942e22bee78STejun Heo */ 194363d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1944e22bee78STejun Heo { 1945e22bee78STejun Heo bool ret = false; 1946e22bee78STejun Heo 194763d95a91STejun Heo while (too_many_workers(pool)) { 1948e22bee78STejun Heo struct worker *worker; 1949e22bee78STejun Heo unsigned long expires; 1950e22bee78STejun Heo 195163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1952e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1953e22bee78STejun Heo 1954e22bee78STejun Heo if (time_before(jiffies, expires)) { 195563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1956e22bee78STejun Heo break; 1957e22bee78STejun Heo } 1958e22bee78STejun Heo 1959e22bee78STejun Heo destroy_worker(worker); 1960e22bee78STejun Heo ret = true; 1961e22bee78STejun Heo } 1962e22bee78STejun Heo 1963e22bee78STejun Heo return ret; 1964e22bee78STejun Heo } 1965e22bee78STejun Heo 1966e22bee78STejun Heo /** 1967e22bee78STejun Heo * manage_workers - manage worker pool 1968e22bee78STejun Heo * @worker: self 1969e22bee78STejun Heo * 1970706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1971e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1972706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1973e22bee78STejun Heo * 1974e22bee78STejun Heo * The caller can safely start processing works on false return. On 1975e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1976e22bee78STejun Heo * and may_start_working() is true. 1977e22bee78STejun Heo * 1978e22bee78STejun Heo * CONTEXT: 1979d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1980e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1981e22bee78STejun Heo * 1982e22bee78STejun Heo * RETURNS: 1983d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1984d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 1985e22bee78STejun Heo */ 1986e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1987e22bee78STejun Heo { 198863d95a91STejun Heo struct worker_pool *pool = worker->pool; 1989e22bee78STejun Heo bool ret = false; 1990e22bee78STejun Heo 1991bc3a1afcSTejun Heo /* 1992bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 1993bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 1994bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1995bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1996bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1997bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1998bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1999bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2000bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2001bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2002bc3a1afcSTejun Heo * 2003bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2004bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2005bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2006bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2007bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2008bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2009bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2010bc3a1afcSTejun Heo * manager_mutex. 2011bc3a1afcSTejun Heo */ 201234a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2013e22bee78STejun Heo return ret; 2014e22bee78STejun Heo 2015ee378aa4SLai Jiangshan /* 2016bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2017bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2018ee378aa4SLai Jiangshan */ 2019bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2020d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2021bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2022ee378aa4SLai Jiangshan ret = true; 2023ee378aa4SLai Jiangshan } 2024ee378aa4SLai Jiangshan 202511ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2026e22bee78STejun Heo 2027e22bee78STejun Heo /* 2028e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2029e22bee78STejun Heo * on return. 2030e22bee78STejun Heo */ 203163d95a91STejun Heo ret |= maybe_destroy_workers(pool); 203263d95a91STejun Heo ret |= maybe_create_worker(pool); 2033e22bee78STejun Heo 2034bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 203534a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2036e22bee78STejun Heo return ret; 2037e22bee78STejun Heo } 2038e22bee78STejun Heo 2039a62428c0STejun Heo /** 2040a62428c0STejun Heo * process_one_work - process single work 2041c34056a3STejun Heo * @worker: self 2042a62428c0STejun Heo * @work: work to process 2043a62428c0STejun Heo * 2044a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2045a62428c0STejun Heo * process a single work including synchronization against and 2046a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2047a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2048a62428c0STejun Heo * call this function to process a work. 2049a62428c0STejun Heo * 2050a62428c0STejun Heo * CONTEXT: 2051d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2052a62428c0STejun Heo */ 2053c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2054d565ed63STejun Heo __releases(&pool->lock) 2055d565ed63STejun Heo __acquires(&pool->lock) 20561da177e4SLinus Torvalds { 2057112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2058bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2059112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 206073f53c4aSTejun Heo int work_color; 20617e11629dSTejun Heo struct worker *collision; 20624e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20634e6045f1SJohannes Berg /* 2064a62428c0STejun Heo * It is permissible to free the struct work_struct from 2065a62428c0STejun Heo * inside the function that is called from it, this we need to 2066a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2067a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2068a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20694e6045f1SJohannes Berg */ 20704d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20714d82a1deSPeter Zijlstra 20724d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20734e6045f1SJohannes Berg #endif 20746fec10a1STejun Heo /* 20756fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20766fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 207724647570STejun Heo * unbound or a disassociated pool. 20786fec10a1STejun Heo */ 20795f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 208024647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2081ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 208225511a47STejun Heo 20837e11629dSTejun Heo /* 20847e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20857e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20867e11629dSTejun Heo * already processing the work. If so, defer the work to the 20877e11629dSTejun Heo * currently executing one. 20887e11629dSTejun Heo */ 2089c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20907e11629dSTejun Heo if (unlikely(collision)) { 20917e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20927e11629dSTejun Heo return; 20937e11629dSTejun Heo } 20941da177e4SLinus Torvalds 20958930cabaSTejun Heo /* claim and dequeue */ 20961da177e4SLinus Torvalds debug_work_deactivate(work); 2097c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2098c34056a3STejun Heo worker->current_work = work; 2099a2c1c57bSTejun Heo worker->current_func = work->func; 2100112202d9STejun Heo worker->current_pwq = pwq; 210173f53c4aSTejun Heo work_color = get_work_color(work); 21027a22ad75STejun Heo 2103a62428c0STejun Heo list_del_init(&work->entry); 2104a62428c0STejun Heo 2105649027d7STejun Heo /* 2106fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2107fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2108fb0e7bebSTejun Heo */ 2109fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2110fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2111fb0e7bebSTejun Heo 2112974271c4STejun Heo /* 2113d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2114974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2115974271c4STejun Heo */ 211663d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 211763d95a91STejun Heo wake_up_worker(pool); 2118974271c4STejun Heo 21198930cabaSTejun Heo /* 21207c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2121d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 212223657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 212323657bb1STejun Heo * disabled. 21248930cabaSTejun Heo */ 21257c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21261da177e4SLinus Torvalds 2127d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2128365970a1SDavid Howells 2129112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21303295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2131e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2132a2c1c57bSTejun Heo worker->current_func(work); 2133e36c886aSArjan van de Ven /* 2134e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2135e36c886aSArjan van de Ven * point will only record its address. 2136e36c886aSArjan van de Ven */ 2137e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21383295f0efSIngo Molnar lock_map_release(&lockdep_map); 2139112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21401da177e4SLinus Torvalds 2141d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2142044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2143044c782cSValentin Ilie " last function: %pf\n", 2144a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2145a2c1c57bSTejun Heo worker->current_func); 2146d5abe669SPeter Zijlstra debug_show_held_locks(current); 2147d5abe669SPeter Zijlstra dump_stack(); 2148d5abe669SPeter Zijlstra } 2149d5abe669SPeter Zijlstra 2150d565ed63STejun Heo spin_lock_irq(&pool->lock); 2151a62428c0STejun Heo 2152fb0e7bebSTejun Heo /* clear cpu intensive status */ 2153fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2154fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2155fb0e7bebSTejun Heo 2156a62428c0STejun Heo /* we're done with it, release */ 215742f8570fSSasha Levin hash_del(&worker->hentry); 2158c34056a3STejun Heo worker->current_work = NULL; 2159a2c1c57bSTejun Heo worker->current_func = NULL; 2160112202d9STejun Heo worker->current_pwq = NULL; 2161112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21621da177e4SLinus Torvalds } 21631da177e4SLinus Torvalds 2164affee4b2STejun Heo /** 2165affee4b2STejun Heo * process_scheduled_works - process scheduled works 2166affee4b2STejun Heo * @worker: self 2167affee4b2STejun Heo * 2168affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2169affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2170affee4b2STejun Heo * fetches a work from the top and executes it. 2171affee4b2STejun Heo * 2172affee4b2STejun Heo * CONTEXT: 2173d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2174affee4b2STejun Heo * multiple times. 2175affee4b2STejun Heo */ 2176affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21771da177e4SLinus Torvalds { 2178affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2179affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2180a62428c0STejun Heo struct work_struct, entry); 2181c34056a3STejun Heo process_one_work(worker, work); 2182a62428c0STejun Heo } 21831da177e4SLinus Torvalds } 21841da177e4SLinus Torvalds 21854690c4abSTejun Heo /** 21864690c4abSTejun Heo * worker_thread - the worker thread function 2187c34056a3STejun Heo * @__worker: self 21884690c4abSTejun Heo * 2189c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2190c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2191c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2192c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2193c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 21944690c4abSTejun Heo */ 2195c34056a3STejun Heo static int worker_thread(void *__worker) 21961da177e4SLinus Torvalds { 2197c34056a3STejun Heo struct worker *worker = __worker; 2198bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21991da177e4SLinus Torvalds 2200e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2201e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2202c8e55f36STejun Heo woke_up: 2203d565ed63STejun Heo spin_lock_irq(&pool->lock); 2204affee4b2STejun Heo 2205a9ab775bSTejun Heo /* am I supposed to die? */ 2206a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2207d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2208a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2209e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2210c8e55f36STejun Heo return 0; 2211c8e55f36STejun Heo } 2212c8e55f36STejun Heo 2213c8e55f36STejun Heo worker_leave_idle(worker); 2214db7bccf4STejun Heo recheck: 2215e22bee78STejun Heo /* no more worker necessary? */ 221663d95a91STejun Heo if (!need_more_worker(pool)) 2217e22bee78STejun Heo goto sleep; 2218e22bee78STejun Heo 2219e22bee78STejun Heo /* do we need to manage? */ 222063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2221e22bee78STejun Heo goto recheck; 2222e22bee78STejun Heo 2223c8e55f36STejun Heo /* 2224c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2225c8e55f36STejun Heo * preparing to process a work or actually processing it. 2226c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2227c8e55f36STejun Heo */ 22286183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2229c8e55f36STejun Heo 2230e22bee78STejun Heo /* 2231a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2232a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2233a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2234a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2235a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2236e22bee78STejun Heo */ 2237a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2238e22bee78STejun Heo 2239e22bee78STejun Heo do { 2240affee4b2STejun Heo struct work_struct *work = 2241bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2242affee4b2STejun Heo struct work_struct, entry); 2243affee4b2STejun Heo 2244c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2245affee4b2STejun Heo /* optimization path, not strictly necessary */ 2246affee4b2STejun Heo process_one_work(worker, work); 2247affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2248affee4b2STejun Heo process_scheduled_works(worker); 2249affee4b2STejun Heo } else { 2250c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2251affee4b2STejun Heo process_scheduled_works(worker); 2252affee4b2STejun Heo } 225363d95a91STejun Heo } while (keep_working(pool)); 2254affee4b2STejun Heo 2255e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2256d313dd85STejun Heo sleep: 225763d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2258e22bee78STejun Heo goto recheck; 2259d313dd85STejun Heo 2260c8e55f36STejun Heo /* 2261d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2262d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2263d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2264d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2265d565ed63STejun Heo * event. 2266c8e55f36STejun Heo */ 2267c8e55f36STejun Heo worker_enter_idle(worker); 2268c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2269d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22701da177e4SLinus Torvalds schedule(); 2271c8e55f36STejun Heo goto woke_up; 22721da177e4SLinus Torvalds } 22731da177e4SLinus Torvalds 2274e22bee78STejun Heo /** 2275e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2276111c225aSTejun Heo * @__rescuer: self 2277e22bee78STejun Heo * 2278e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2279493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2280e22bee78STejun Heo * 2281706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2282e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2283e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2284e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2285e22bee78STejun Heo * the problem rescuer solves. 2286e22bee78STejun Heo * 2287706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2288706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2289e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2290e22bee78STejun Heo * 2291e22bee78STejun Heo * This should happen rarely. 2292e22bee78STejun Heo */ 2293111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2294e22bee78STejun Heo { 2295111c225aSTejun Heo struct worker *rescuer = __rescuer; 2296111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2297e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2298e22bee78STejun Heo 2299e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2300111c225aSTejun Heo 2301111c225aSTejun Heo /* 2302111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2303111c225aSTejun Heo * doesn't participate in concurrency management. 2304111c225aSTejun Heo */ 2305111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2306e22bee78STejun Heo repeat: 2307e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23081da177e4SLinus Torvalds 2309412d32e6SMike Galbraith if (kthread_should_stop()) { 2310412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2311111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2312e22bee78STejun Heo return 0; 2313412d32e6SMike Galbraith } 23141da177e4SLinus Torvalds 2315493a1724STejun Heo /* see whether any pwq is asking for help */ 23162e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2317493a1724STejun Heo 2318493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2319493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2320493a1724STejun Heo struct pool_workqueue, mayday_node); 2321112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2322e22bee78STejun Heo struct work_struct *work, *n; 2323e22bee78STejun Heo 2324e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2325493a1724STejun Heo list_del_init(&pwq->mayday_node); 2326493a1724STejun Heo 23272e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2328e22bee78STejun Heo 2329e22bee78STejun Heo /* migrate to the target cpu if possible */ 2330f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2331b3104104SLai Jiangshan rescuer->pool = pool; 2332e22bee78STejun Heo 2333e22bee78STejun Heo /* 2334e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2335e22bee78STejun Heo * process'em. 2336e22bee78STejun Heo */ 23376183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2338bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2339112202d9STejun Heo if (get_work_pwq(work) == pwq) 2340e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2341e22bee78STejun Heo 2342e22bee78STejun Heo process_scheduled_works(rescuer); 23437576958aSTejun Heo 23447576958aSTejun Heo /* 2345d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23467576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23477576958aSTejun Heo * and stalling the execution. 23487576958aSTejun Heo */ 234963d95a91STejun Heo if (keep_working(pool)) 235063d95a91STejun Heo wake_up_worker(pool); 23517576958aSTejun Heo 2352b3104104SLai Jiangshan rescuer->pool = NULL; 2353493a1724STejun Heo spin_unlock(&pool->lock); 23542e109a28STejun Heo spin_lock(&wq_mayday_lock); 23551da177e4SLinus Torvalds } 23561da177e4SLinus Torvalds 23572e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2358493a1724STejun Heo 2359111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2360111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2361e22bee78STejun Heo schedule(); 2362e22bee78STejun Heo goto repeat; 23631da177e4SLinus Torvalds } 23641da177e4SLinus Torvalds 2365fc2e4d70SOleg Nesterov struct wq_barrier { 2366fc2e4d70SOleg Nesterov struct work_struct work; 2367fc2e4d70SOleg Nesterov struct completion done; 2368fc2e4d70SOleg Nesterov }; 2369fc2e4d70SOleg Nesterov 2370fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2371fc2e4d70SOleg Nesterov { 2372fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2373fc2e4d70SOleg Nesterov complete(&barr->done); 2374fc2e4d70SOleg Nesterov } 2375fc2e4d70SOleg Nesterov 23764690c4abSTejun Heo /** 23774690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2378112202d9STejun Heo * @pwq: pwq to insert barrier into 23794690c4abSTejun Heo * @barr: wq_barrier to insert 2380affee4b2STejun Heo * @target: target work to attach @barr to 2381affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23824690c4abSTejun Heo * 2383affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2384affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2385affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2386affee4b2STejun Heo * cpu. 2387affee4b2STejun Heo * 2388affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2389affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2390affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2391affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2392affee4b2STejun Heo * after a work with LINKED flag set. 2393affee4b2STejun Heo * 2394affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2395112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23964690c4abSTejun Heo * 23974690c4abSTejun Heo * CONTEXT: 2398d565ed63STejun Heo * spin_lock_irq(pool->lock). 23994690c4abSTejun Heo */ 2400112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2401affee4b2STejun Heo struct wq_barrier *barr, 2402affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2403fc2e4d70SOleg Nesterov { 2404affee4b2STejun Heo struct list_head *head; 2405affee4b2STejun Heo unsigned int linked = 0; 2406affee4b2STejun Heo 2407dc186ad7SThomas Gleixner /* 2408d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2409dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2410dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2411dc186ad7SThomas Gleixner * might deadlock. 2412dc186ad7SThomas Gleixner */ 2413ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 241422df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2415fc2e4d70SOleg Nesterov init_completion(&barr->done); 241683c22520SOleg Nesterov 2417affee4b2STejun Heo /* 2418affee4b2STejun Heo * If @target is currently being executed, schedule the 2419affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2420affee4b2STejun Heo */ 2421affee4b2STejun Heo if (worker) 2422affee4b2STejun Heo head = worker->scheduled.next; 2423affee4b2STejun Heo else { 2424affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2425affee4b2STejun Heo 2426affee4b2STejun Heo head = target->entry.next; 2427affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2428affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2429affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2430affee4b2STejun Heo } 2431affee4b2STejun Heo 2432dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2433112202d9STejun Heo insert_work(pwq, &barr->work, head, 2434affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2435fc2e4d70SOleg Nesterov } 2436fc2e4d70SOleg Nesterov 243773f53c4aSTejun Heo /** 2438112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 243973f53c4aSTejun Heo * @wq: workqueue being flushed 244073f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 244173f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 244273f53c4aSTejun Heo * 2443112202d9STejun Heo * Prepare pwqs for workqueue flushing. 244473f53c4aSTejun Heo * 2445112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2446112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2447112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2448112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2449112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 245073f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 245173f53c4aSTejun Heo * 245273f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 245373f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 245473f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 245573f53c4aSTejun Heo * is returned. 245673f53c4aSTejun Heo * 2457112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 245873f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 245973f53c4aSTejun Heo * advanced to @work_color. 246073f53c4aSTejun Heo * 246173f53c4aSTejun Heo * CONTEXT: 246273f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 246373f53c4aSTejun Heo * 246473f53c4aSTejun Heo * RETURNS: 246573f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 246673f53c4aSTejun Heo * otherwise. 246773f53c4aSTejun Heo */ 2468112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 246973f53c4aSTejun Heo int flush_color, int work_color) 24701da177e4SLinus Torvalds { 247173f53c4aSTejun Heo bool wait = false; 247249e3cf44STejun Heo struct pool_workqueue *pwq; 24731da177e4SLinus Torvalds 247473f53c4aSTejun Heo if (flush_color >= 0) { 24756183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2476112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2477dc186ad7SThomas Gleixner } 247814441960SOleg Nesterov 247976af4d93STejun Heo local_irq_disable(); 248076af4d93STejun Heo 248149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2482112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24831da177e4SLinus Torvalds 248476af4d93STejun Heo spin_lock(&pool->lock); 248573f53c4aSTejun Heo 248673f53c4aSTejun Heo if (flush_color >= 0) { 24876183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 248873f53c4aSTejun Heo 2489112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2490112202d9STejun Heo pwq->flush_color = flush_color; 2491112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 249273f53c4aSTejun Heo wait = true; 24931da177e4SLinus Torvalds } 249473f53c4aSTejun Heo } 249573f53c4aSTejun Heo 249673f53c4aSTejun Heo if (work_color >= 0) { 24976183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2498112202d9STejun Heo pwq->work_color = work_color; 249973f53c4aSTejun Heo } 250073f53c4aSTejun Heo 250176af4d93STejun Heo spin_unlock(&pool->lock); 25021da177e4SLinus Torvalds } 25031da177e4SLinus Torvalds 250476af4d93STejun Heo local_irq_enable(); 250576af4d93STejun Heo 2506112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 250773f53c4aSTejun Heo complete(&wq->first_flusher->done); 250873f53c4aSTejun Heo 250973f53c4aSTejun Heo return wait; 251083c22520SOleg Nesterov } 25111da177e4SLinus Torvalds 25120fcb78c2SRolf Eike Beer /** 25131da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25140fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25151da177e4SLinus Torvalds * 2516c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2517c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25181da177e4SLinus Torvalds */ 25197ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25201da177e4SLinus Torvalds { 252173f53c4aSTejun Heo struct wq_flusher this_flusher = { 252273f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 252373f53c4aSTejun Heo .flush_color = -1, 252473f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 252573f53c4aSTejun Heo }; 252673f53c4aSTejun Heo int next_color; 2527b1f4ec17SOleg Nesterov 25283295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25293295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 253073f53c4aSTejun Heo 253173f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 253273f53c4aSTejun Heo 253373f53c4aSTejun Heo /* 253473f53c4aSTejun Heo * Start-to-wait phase 253573f53c4aSTejun Heo */ 253673f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 253773f53c4aSTejun Heo 253873f53c4aSTejun Heo if (next_color != wq->flush_color) { 253973f53c4aSTejun Heo /* 254073f53c4aSTejun Heo * Color space is not full. The current work_color 254173f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 254273f53c4aSTejun Heo * by one. 254373f53c4aSTejun Heo */ 25446183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 254573f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 254673f53c4aSTejun Heo wq->work_color = next_color; 254773f53c4aSTejun Heo 254873f53c4aSTejun Heo if (!wq->first_flusher) { 254973f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25506183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 255173f53c4aSTejun Heo 255273f53c4aSTejun Heo wq->first_flusher = &this_flusher; 255373f53c4aSTejun Heo 2554112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 255573f53c4aSTejun Heo wq->work_color)) { 255673f53c4aSTejun Heo /* nothing to flush, done */ 255773f53c4aSTejun Heo wq->flush_color = next_color; 255873f53c4aSTejun Heo wq->first_flusher = NULL; 255973f53c4aSTejun Heo goto out_unlock; 256073f53c4aSTejun Heo } 256173f53c4aSTejun Heo } else { 256273f53c4aSTejun Heo /* wait in queue */ 25636183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 256473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2565112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 256673f53c4aSTejun Heo } 256773f53c4aSTejun Heo } else { 256873f53c4aSTejun Heo /* 256973f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 257073f53c4aSTejun Heo * The next flush completion will assign us 257173f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 257273f53c4aSTejun Heo */ 257373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 257473f53c4aSTejun Heo } 257573f53c4aSTejun Heo 257673f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 257773f53c4aSTejun Heo 257873f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 257973f53c4aSTejun Heo 258073f53c4aSTejun Heo /* 258173f53c4aSTejun Heo * Wake-up-and-cascade phase 258273f53c4aSTejun Heo * 258373f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 258473f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 258573f53c4aSTejun Heo */ 258673f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 258773f53c4aSTejun Heo return; 258873f53c4aSTejun Heo 258973f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 259073f53c4aSTejun Heo 25914ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25924ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25934ce48b37STejun Heo goto out_unlock; 25944ce48b37STejun Heo 259573f53c4aSTejun Heo wq->first_flusher = NULL; 259673f53c4aSTejun Heo 25976183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25986183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 259973f53c4aSTejun Heo 260073f53c4aSTejun Heo while (true) { 260173f53c4aSTejun Heo struct wq_flusher *next, *tmp; 260273f53c4aSTejun Heo 260373f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 260473f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 260573f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 260673f53c4aSTejun Heo break; 260773f53c4aSTejun Heo list_del_init(&next->list); 260873f53c4aSTejun Heo complete(&next->done); 260973f53c4aSTejun Heo } 261073f53c4aSTejun Heo 26116183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 261273f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 261373f53c4aSTejun Heo 261473f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 261573f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 261673f53c4aSTejun Heo 261773f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 261873f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 261973f53c4aSTejun Heo /* 262073f53c4aSTejun Heo * Assign the same color to all overflowed 262173f53c4aSTejun Heo * flushers, advance work_color and append to 262273f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 262373f53c4aSTejun Heo * phase for these overflowed flushers. 262473f53c4aSTejun Heo */ 262573f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 262673f53c4aSTejun Heo tmp->flush_color = wq->work_color; 262773f53c4aSTejun Heo 262873f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 262973f53c4aSTejun Heo 263073f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 263173f53c4aSTejun Heo &wq->flusher_queue); 2632112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 263373f53c4aSTejun Heo } 263473f53c4aSTejun Heo 263573f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26366183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 263773f53c4aSTejun Heo break; 263873f53c4aSTejun Heo } 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo /* 264173f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2642112202d9STejun Heo * the new first flusher and arm pwqs. 264373f53c4aSTejun Heo */ 26446183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26456183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 264673f53c4aSTejun Heo 264773f53c4aSTejun Heo list_del_init(&next->list); 264873f53c4aSTejun Heo wq->first_flusher = next; 264973f53c4aSTejun Heo 2650112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 265173f53c4aSTejun Heo break; 265273f53c4aSTejun Heo 265373f53c4aSTejun Heo /* 265473f53c4aSTejun Heo * Meh... this color is already done, clear first 265573f53c4aSTejun Heo * flusher and repeat cascading. 265673f53c4aSTejun Heo */ 265773f53c4aSTejun Heo wq->first_flusher = NULL; 265873f53c4aSTejun Heo } 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo out_unlock: 266173f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 26621da177e4SLinus Torvalds } 2663ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26641da177e4SLinus Torvalds 26659c5a2ba7STejun Heo /** 26669c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26679c5a2ba7STejun Heo * @wq: workqueue to drain 26689c5a2ba7STejun Heo * 26699c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26709c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26719c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26729c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26739c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26749c5a2ba7STejun Heo * takes too long. 26759c5a2ba7STejun Heo */ 26769c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26779c5a2ba7STejun Heo { 26789c5a2ba7STejun Heo unsigned int flush_cnt = 0; 267949e3cf44STejun Heo struct pool_workqueue *pwq; 26809c5a2ba7STejun Heo 26819c5a2ba7STejun Heo /* 26829c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26839c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2684618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26859c5a2ba7STejun Heo */ 26865bcab335STejun Heo mutex_lock(&wq_mutex); 26879c5a2ba7STejun Heo if (!wq->nr_drainers++) 2688618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 26895bcab335STejun Heo mutex_unlock(&wq_mutex); 26909c5a2ba7STejun Heo reflush: 26919c5a2ba7STejun Heo flush_workqueue(wq); 26929c5a2ba7STejun Heo 269376af4d93STejun Heo local_irq_disable(); 269476af4d93STejun Heo 269549e3cf44STejun Heo for_each_pwq(pwq, wq) { 2696fa2563e4SThomas Tuttle bool drained; 26979c5a2ba7STejun Heo 269876af4d93STejun Heo spin_lock(&pwq->pool->lock); 2699112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 270076af4d93STejun Heo spin_unlock(&pwq->pool->lock); 2701fa2563e4SThomas Tuttle 2702fa2563e4SThomas Tuttle if (drained) 27039c5a2ba7STejun Heo continue; 27049c5a2ba7STejun Heo 27059c5a2ba7STejun Heo if (++flush_cnt == 10 || 27069c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2707c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27089c5a2ba7STejun Heo wq->name, flush_cnt); 270976af4d93STejun Heo 271076af4d93STejun Heo local_irq_enable(); 27119c5a2ba7STejun Heo goto reflush; 27129c5a2ba7STejun Heo } 27139c5a2ba7STejun Heo 27145bcab335STejun Heo local_irq_enable(); 27155bcab335STejun Heo 27165bcab335STejun Heo mutex_lock(&wq_mutex); 27179c5a2ba7STejun Heo if (!--wq->nr_drainers) 2718618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 27195bcab335STejun Heo mutex_unlock(&wq_mutex); 27209c5a2ba7STejun Heo } 27219c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27229c5a2ba7STejun Heo 2723606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2724baf59022STejun Heo { 2725baf59022STejun Heo struct worker *worker = NULL; 2726c9e7cf27STejun Heo struct worker_pool *pool; 2727112202d9STejun Heo struct pool_workqueue *pwq; 2728baf59022STejun Heo 2729baf59022STejun Heo might_sleep(); 2730baf59022STejun Heo 2731fa1b54e6STejun Heo local_irq_disable(); 2732fa1b54e6STejun Heo pool = get_work_pool(work); 2733fa1b54e6STejun Heo if (!pool) { 2734fa1b54e6STejun Heo local_irq_enable(); 2735fa1b54e6STejun Heo return false; 2736fa1b54e6STejun Heo } 2737fa1b54e6STejun Heo 2738fa1b54e6STejun Heo spin_lock(&pool->lock); 27390b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2740112202d9STejun Heo pwq = get_work_pwq(work); 2741112202d9STejun Heo if (pwq) { 2742112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2743baf59022STejun Heo goto already_gone; 2744606a5020STejun Heo } else { 2745c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2746baf59022STejun Heo if (!worker) 2747baf59022STejun Heo goto already_gone; 2748112202d9STejun Heo pwq = worker->current_pwq; 2749606a5020STejun Heo } 2750baf59022STejun Heo 2751112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2752d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2753baf59022STejun Heo 2754e159489bSTejun Heo /* 2755e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2756e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2757e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2758e159489bSTejun Heo * access. 2759e159489bSTejun Heo */ 2760493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2761112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2762e159489bSTejun Heo else 2763112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2764112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2765e159489bSTejun Heo 2766baf59022STejun Heo return true; 2767baf59022STejun Heo already_gone: 2768d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2769baf59022STejun Heo return false; 2770baf59022STejun Heo } 2771baf59022STejun Heo 2772db700897SOleg Nesterov /** 2773401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2774401a8d04STejun Heo * @work: the work to flush 2775db700897SOleg Nesterov * 2776606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2777606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2778401a8d04STejun Heo * 2779401a8d04STejun Heo * RETURNS: 2780401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2781401a8d04STejun Heo * %false if it was already idle. 2782db700897SOleg Nesterov */ 2783401a8d04STejun Heo bool flush_work(struct work_struct *work) 2784db700897SOleg Nesterov { 2785db700897SOleg Nesterov struct wq_barrier barr; 2786db700897SOleg Nesterov 27870976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27880976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27890976dfc1SStephen Boyd 2790606a5020STejun Heo if (start_flush_work(work, &barr)) { 2791db700897SOleg Nesterov wait_for_completion(&barr.done); 2792dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2793401a8d04STejun Heo return true; 2794606a5020STejun Heo } else { 2795401a8d04STejun Heo return false; 2796db700897SOleg Nesterov } 2797606a5020STejun Heo } 2798db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2799db700897SOleg Nesterov 280036e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2801401a8d04STejun Heo { 2802bbb68dfaSTejun Heo unsigned long flags; 28031f1f642eSOleg Nesterov int ret; 28041f1f642eSOleg Nesterov 28051f1f642eSOleg Nesterov do { 2806bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2807bbb68dfaSTejun Heo /* 2808bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2809bbb68dfaSTejun Heo * would be waiting for before retrying. 2810bbb68dfaSTejun Heo */ 2811bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2812606a5020STejun Heo flush_work(work); 28131f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28141f1f642eSOleg Nesterov 2815bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2816bbb68dfaSTejun Heo mark_work_canceling(work); 2817bbb68dfaSTejun Heo local_irq_restore(flags); 2818bbb68dfaSTejun Heo 2819606a5020STejun Heo flush_work(work); 28207a22ad75STejun Heo clear_work_data(work); 28211f1f642eSOleg Nesterov return ret; 28221f1f642eSOleg Nesterov } 28231f1f642eSOleg Nesterov 28246e84d644SOleg Nesterov /** 2825401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2826401a8d04STejun Heo * @work: the work to cancel 28276e84d644SOleg Nesterov * 2828401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2829401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2830401a8d04STejun Heo * another workqueue. On return from this function, @work is 2831401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28321f1f642eSOleg Nesterov * 2833401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2834401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28356e84d644SOleg Nesterov * 2836401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28376e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2838401a8d04STejun Heo * 2839401a8d04STejun Heo * RETURNS: 2840401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28416e84d644SOleg Nesterov */ 2842401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28436e84d644SOleg Nesterov { 284436e227d2STejun Heo return __cancel_work_timer(work, false); 2845b89deed3SOleg Nesterov } 284628e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2847b89deed3SOleg Nesterov 28486e84d644SOleg Nesterov /** 2849401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2850401a8d04STejun Heo * @dwork: the delayed work to flush 28516e84d644SOleg Nesterov * 2852401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2853401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2854401a8d04STejun Heo * considers the last queueing instance of @dwork. 28551f1f642eSOleg Nesterov * 2856401a8d04STejun Heo * RETURNS: 2857401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2858401a8d04STejun Heo * %false if it was already idle. 28596e84d644SOleg Nesterov */ 2860401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2861401a8d04STejun Heo { 28628930cabaSTejun Heo local_irq_disable(); 2863401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 286460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28658930cabaSTejun Heo local_irq_enable(); 2866401a8d04STejun Heo return flush_work(&dwork->work); 2867401a8d04STejun Heo } 2868401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2869401a8d04STejun Heo 2870401a8d04STejun Heo /** 287157b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 287257b30ae7STejun Heo * @dwork: delayed_work to cancel 287309383498STejun Heo * 287457b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 287557b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 287657b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 287757b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 287857b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 287909383498STejun Heo * 288057b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 288109383498STejun Heo */ 288257b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 288309383498STejun Heo { 288457b30ae7STejun Heo unsigned long flags; 288557b30ae7STejun Heo int ret; 288657b30ae7STejun Heo 288757b30ae7STejun Heo do { 288857b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 288957b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 289057b30ae7STejun Heo 289157b30ae7STejun Heo if (unlikely(ret < 0)) 289257b30ae7STejun Heo return false; 289357b30ae7STejun Heo 28947c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28957c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289657b30ae7STejun Heo local_irq_restore(flags); 2897c0158ca6SDan Magenheimer return ret; 289809383498STejun Heo } 289957b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 290009383498STejun Heo 290109383498STejun Heo /** 2902401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2903401a8d04STejun Heo * @dwork: the delayed work cancel 2904401a8d04STejun Heo * 2905401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2906401a8d04STejun Heo * 2907401a8d04STejun Heo * RETURNS: 2908401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2909401a8d04STejun Heo */ 2910401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29116e84d644SOleg Nesterov { 291236e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29136e84d644SOleg Nesterov } 2914f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29151da177e4SLinus Torvalds 29160fcb78c2SRolf Eike Beer /** 291731ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2918b6136773SAndrew Morton * @func: the function to call 2919b6136773SAndrew Morton * 292031ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 292131ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2922b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 292331ddd871STejun Heo * 292431ddd871STejun Heo * RETURNS: 292531ddd871STejun Heo * 0 on success, -errno on failure. 2926b6136773SAndrew Morton */ 292765f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 292815316ba8SChristoph Lameter { 292915316ba8SChristoph Lameter int cpu; 293038f51568SNamhyung Kim struct work_struct __percpu *works; 293115316ba8SChristoph Lameter 2932b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2933b6136773SAndrew Morton if (!works) 293415316ba8SChristoph Lameter return -ENOMEM; 2935b6136773SAndrew Morton 293695402b38SGautham R Shenoy get_online_cpus(); 293793981800STejun Heo 293815316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29399bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29409bfb1839SIngo Molnar 29419bfb1839SIngo Molnar INIT_WORK(work, func); 29428de6d308SOleg Nesterov schedule_work_on(cpu, work); 294315316ba8SChristoph Lameter } 294493981800STejun Heo 294593981800STejun Heo for_each_online_cpu(cpu) 29468616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 294793981800STejun Heo 294895402b38SGautham R Shenoy put_online_cpus(); 2949b6136773SAndrew Morton free_percpu(works); 295015316ba8SChristoph Lameter return 0; 295115316ba8SChristoph Lameter } 295215316ba8SChristoph Lameter 2953eef6a7d5SAlan Stern /** 2954eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2955eef6a7d5SAlan Stern * 2956eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2957eef6a7d5SAlan Stern * completion. 2958eef6a7d5SAlan Stern * 2959eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2960eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2961eef6a7d5SAlan Stern * will lead to deadlock: 2962eef6a7d5SAlan Stern * 2963eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2964eef6a7d5SAlan Stern * a lock held by your code or its caller. 2965eef6a7d5SAlan Stern * 2966eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2967eef6a7d5SAlan Stern * 2968eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2969eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2970eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2971eef6a7d5SAlan Stern * 2972eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2973eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2974eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2975eef6a7d5SAlan Stern * cancel_work_sync() instead. 2976eef6a7d5SAlan Stern */ 29771da177e4SLinus Torvalds void flush_scheduled_work(void) 29781da177e4SLinus Torvalds { 2979d320c038STejun Heo flush_workqueue(system_wq); 29801da177e4SLinus Torvalds } 2981ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29821da177e4SLinus Torvalds 29831da177e4SLinus Torvalds /** 29841fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29851fa44ecaSJames Bottomley * @fn: the function to execute 29861fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29871fa44ecaSJames Bottomley * be available when the work executes) 29881fa44ecaSJames Bottomley * 29891fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29901fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29911fa44ecaSJames Bottomley * 29921fa44ecaSJames Bottomley * Returns: 0 - function was executed 29931fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29941fa44ecaSJames Bottomley */ 299565f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29961fa44ecaSJames Bottomley { 29971fa44ecaSJames Bottomley if (!in_interrupt()) { 299865f27f38SDavid Howells fn(&ew->work); 29991fa44ecaSJames Bottomley return 0; 30001fa44ecaSJames Bottomley } 30011fa44ecaSJames Bottomley 300265f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30031fa44ecaSJames Bottomley schedule_work(&ew->work); 30041fa44ecaSJames Bottomley 30051fa44ecaSJames Bottomley return 1; 30061fa44ecaSJames Bottomley } 30071fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30081fa44ecaSJames Bottomley 3009226223abSTejun Heo #ifdef CONFIG_SYSFS 3010226223abSTejun Heo /* 3011226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3012226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3013226223abSTejun Heo * following attributes. 3014226223abSTejun Heo * 3015226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3016226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3017226223abSTejun Heo * 3018226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3019226223abSTejun Heo * 3020226223abSTejun Heo * id RO int : the associated pool ID 3021226223abSTejun Heo * nice RW int : nice value of the workers 3022226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3023226223abSTejun Heo */ 3024226223abSTejun Heo struct wq_device { 3025226223abSTejun Heo struct workqueue_struct *wq; 3026226223abSTejun Heo struct device dev; 3027226223abSTejun Heo }; 3028226223abSTejun Heo 3029226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3030226223abSTejun Heo { 3031226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3032226223abSTejun Heo 3033226223abSTejun Heo return wq_dev->wq; 3034226223abSTejun Heo } 3035226223abSTejun Heo 3036226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3037226223abSTejun Heo struct device_attribute *attr, char *buf) 3038226223abSTejun Heo { 3039226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3040226223abSTejun Heo 3041226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3042226223abSTejun Heo } 3043226223abSTejun Heo 3044226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3045226223abSTejun Heo struct device_attribute *attr, char *buf) 3046226223abSTejun Heo { 3047226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3048226223abSTejun Heo 3049226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3050226223abSTejun Heo } 3051226223abSTejun Heo 3052226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3053226223abSTejun Heo struct device_attribute *attr, 3054226223abSTejun Heo const char *buf, size_t count) 3055226223abSTejun Heo { 3056226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3057226223abSTejun Heo int val; 3058226223abSTejun Heo 3059226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3060226223abSTejun Heo return -EINVAL; 3061226223abSTejun Heo 3062226223abSTejun Heo workqueue_set_max_active(wq, val); 3063226223abSTejun Heo return count; 3064226223abSTejun Heo } 3065226223abSTejun Heo 3066226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3067226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3068226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3069226223abSTejun Heo __ATTR_NULL, 3070226223abSTejun Heo }; 3071226223abSTejun Heo 3072226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3073226223abSTejun Heo struct device_attribute *attr, char *buf) 3074226223abSTejun Heo { 3075226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3076226223abSTejun Heo struct worker_pool *pool; 3077226223abSTejun Heo int written; 3078226223abSTejun Heo 3079226223abSTejun Heo rcu_read_lock_sched(); 3080226223abSTejun Heo pool = first_pwq(wq)->pool; 3081226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3082226223abSTejun Heo rcu_read_unlock_sched(); 3083226223abSTejun Heo 3084226223abSTejun Heo return written; 3085226223abSTejun Heo } 3086226223abSTejun Heo 3087226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3088226223abSTejun Heo char *buf) 3089226223abSTejun Heo { 3090226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3091226223abSTejun Heo int written; 3092226223abSTejun Heo 3093226223abSTejun Heo rcu_read_lock_sched(); 3094226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3095226223abSTejun Heo first_pwq(wq)->pool->attrs->nice); 3096226223abSTejun Heo rcu_read_unlock_sched(); 3097226223abSTejun Heo 3098226223abSTejun Heo return written; 3099226223abSTejun Heo } 3100226223abSTejun Heo 3101226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3102226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3103226223abSTejun Heo { 3104226223abSTejun Heo struct workqueue_attrs *attrs; 3105226223abSTejun Heo 3106226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3107226223abSTejun Heo if (!attrs) 3108226223abSTejun Heo return NULL; 3109226223abSTejun Heo 3110226223abSTejun Heo rcu_read_lock_sched(); 3111226223abSTejun Heo copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs); 3112226223abSTejun Heo rcu_read_unlock_sched(); 3113226223abSTejun Heo return attrs; 3114226223abSTejun Heo } 3115226223abSTejun Heo 3116226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3117226223abSTejun Heo const char *buf, size_t count) 3118226223abSTejun Heo { 3119226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3120226223abSTejun Heo struct workqueue_attrs *attrs; 3121226223abSTejun Heo int ret; 3122226223abSTejun Heo 3123226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3124226223abSTejun Heo if (!attrs) 3125226223abSTejun Heo return -ENOMEM; 3126226223abSTejun Heo 3127226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3128226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3129226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3130226223abSTejun Heo else 3131226223abSTejun Heo ret = -EINVAL; 3132226223abSTejun Heo 3133226223abSTejun Heo free_workqueue_attrs(attrs); 3134226223abSTejun Heo return ret ?: count; 3135226223abSTejun Heo } 3136226223abSTejun Heo 3137226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3138226223abSTejun Heo struct device_attribute *attr, char *buf) 3139226223abSTejun Heo { 3140226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3141226223abSTejun Heo int written; 3142226223abSTejun Heo 3143226223abSTejun Heo rcu_read_lock_sched(); 3144226223abSTejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, 3145226223abSTejun Heo first_pwq(wq)->pool->attrs->cpumask); 3146226223abSTejun Heo rcu_read_unlock_sched(); 3147226223abSTejun Heo 3148226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3149226223abSTejun Heo return written; 3150226223abSTejun Heo } 3151226223abSTejun Heo 3152226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3153226223abSTejun Heo struct device_attribute *attr, 3154226223abSTejun Heo const char *buf, size_t count) 3155226223abSTejun Heo { 3156226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3157226223abSTejun Heo struct workqueue_attrs *attrs; 3158226223abSTejun Heo int ret; 3159226223abSTejun Heo 3160226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3161226223abSTejun Heo if (!attrs) 3162226223abSTejun Heo return -ENOMEM; 3163226223abSTejun Heo 3164226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3165226223abSTejun Heo if (!ret) 3166226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3167226223abSTejun Heo 3168226223abSTejun Heo free_workqueue_attrs(attrs); 3169226223abSTejun Heo return ret ?: count; 3170226223abSTejun Heo } 3171226223abSTejun Heo 3172226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3173226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3174226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3175226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3176226223abSTejun Heo __ATTR_NULL, 3177226223abSTejun Heo }; 3178226223abSTejun Heo 3179226223abSTejun Heo static struct bus_type wq_subsys = { 3180226223abSTejun Heo .name = "workqueue", 3181226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3182226223abSTejun Heo }; 3183226223abSTejun Heo 3184226223abSTejun Heo static int __init wq_sysfs_init(void) 3185226223abSTejun Heo { 3186226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3187226223abSTejun Heo } 3188226223abSTejun Heo core_initcall(wq_sysfs_init); 3189226223abSTejun Heo 3190226223abSTejun Heo static void wq_device_release(struct device *dev) 3191226223abSTejun Heo { 3192226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3193226223abSTejun Heo 3194226223abSTejun Heo kfree(wq_dev); 3195226223abSTejun Heo } 3196226223abSTejun Heo 3197226223abSTejun Heo /** 3198226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3199226223abSTejun Heo * @wq: the workqueue to register 3200226223abSTejun Heo * 3201226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3202226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3203226223abSTejun Heo * which is the preferred method. 3204226223abSTejun Heo * 3205226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3206226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3207226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3208226223abSTejun Heo * attributes. 3209226223abSTejun Heo * 3210226223abSTejun Heo * Returns 0 on success, -errno on failure. 3211226223abSTejun Heo */ 3212226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3213226223abSTejun Heo { 3214226223abSTejun Heo struct wq_device *wq_dev; 3215226223abSTejun Heo int ret; 3216226223abSTejun Heo 3217226223abSTejun Heo /* 3218226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3219226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3220226223abSTejun Heo * workqueues. 3221226223abSTejun Heo */ 3222226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3223226223abSTejun Heo return -EINVAL; 3224226223abSTejun Heo 3225226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3226226223abSTejun Heo if (!wq_dev) 3227226223abSTejun Heo return -ENOMEM; 3228226223abSTejun Heo 3229226223abSTejun Heo wq_dev->wq = wq; 3230226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3231226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3232226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3233226223abSTejun Heo 3234226223abSTejun Heo /* 3235226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3236226223abSTejun Heo * everything is ready. 3237226223abSTejun Heo */ 3238226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3239226223abSTejun Heo 3240226223abSTejun Heo ret = device_register(&wq_dev->dev); 3241226223abSTejun Heo if (ret) { 3242226223abSTejun Heo kfree(wq_dev); 3243226223abSTejun Heo wq->wq_dev = NULL; 3244226223abSTejun Heo return ret; 3245226223abSTejun Heo } 3246226223abSTejun Heo 3247226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3248226223abSTejun Heo struct device_attribute *attr; 3249226223abSTejun Heo 3250226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3251226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3252226223abSTejun Heo if (ret) { 3253226223abSTejun Heo device_unregister(&wq_dev->dev); 3254226223abSTejun Heo wq->wq_dev = NULL; 3255226223abSTejun Heo return ret; 3256226223abSTejun Heo } 3257226223abSTejun Heo } 3258226223abSTejun Heo } 3259226223abSTejun Heo 3260226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3261226223abSTejun Heo return 0; 3262226223abSTejun Heo } 3263226223abSTejun Heo 3264226223abSTejun Heo /** 3265226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3266226223abSTejun Heo * @wq: the workqueue to unregister 3267226223abSTejun Heo * 3268226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3269226223abSTejun Heo */ 3270226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3271226223abSTejun Heo { 3272226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3273226223abSTejun Heo 3274226223abSTejun Heo if (!wq->wq_dev) 3275226223abSTejun Heo return; 3276226223abSTejun Heo 3277226223abSTejun Heo wq->wq_dev = NULL; 3278226223abSTejun Heo device_unregister(&wq_dev->dev); 3279226223abSTejun Heo } 3280226223abSTejun Heo #else /* CONFIG_SYSFS */ 3281226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3282226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3283226223abSTejun Heo 32847a4e344cSTejun Heo /** 32857a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32867a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32877a4e344cSTejun Heo * 32887a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32897a4e344cSTejun Heo */ 32907a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32917a4e344cSTejun Heo { 32927a4e344cSTejun Heo if (attrs) { 32937a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32947a4e344cSTejun Heo kfree(attrs); 32957a4e344cSTejun Heo } 32967a4e344cSTejun Heo } 32977a4e344cSTejun Heo 32987a4e344cSTejun Heo /** 32997a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33007a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33017a4e344cSTejun Heo * 33027a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33037a4e344cSTejun Heo * return it. Returns NULL on failure. 33047a4e344cSTejun Heo */ 33057a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33067a4e344cSTejun Heo { 33077a4e344cSTejun Heo struct workqueue_attrs *attrs; 33087a4e344cSTejun Heo 33097a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33107a4e344cSTejun Heo if (!attrs) 33117a4e344cSTejun Heo goto fail; 33127a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33137a4e344cSTejun Heo goto fail; 33147a4e344cSTejun Heo 33157a4e344cSTejun Heo cpumask_setall(attrs->cpumask); 33167a4e344cSTejun Heo return attrs; 33177a4e344cSTejun Heo fail: 33187a4e344cSTejun Heo free_workqueue_attrs(attrs); 33197a4e344cSTejun Heo return NULL; 33207a4e344cSTejun Heo } 33217a4e344cSTejun Heo 332229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 332329c91e99STejun Heo const struct workqueue_attrs *from) 332429c91e99STejun Heo { 332529c91e99STejun Heo to->nice = from->nice; 332629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 332729c91e99STejun Heo } 332829c91e99STejun Heo 332929c91e99STejun Heo /* 333029c91e99STejun Heo * Hacky implementation of jhash of bitmaps which only considers the 333129c91e99STejun Heo * specified number of bits. We probably want a proper implementation in 333229c91e99STejun Heo * include/linux/jhash.h. 333329c91e99STejun Heo */ 333429c91e99STejun Heo static u32 jhash_bitmap(const unsigned long *bitmap, int bits, u32 hash) 333529c91e99STejun Heo { 333629c91e99STejun Heo int nr_longs = bits / BITS_PER_LONG; 333729c91e99STejun Heo int nr_leftover = bits % BITS_PER_LONG; 333829c91e99STejun Heo unsigned long leftover = 0; 333929c91e99STejun Heo 334029c91e99STejun Heo if (nr_longs) 334129c91e99STejun Heo hash = jhash(bitmap, nr_longs * sizeof(long), hash); 334229c91e99STejun Heo if (nr_leftover) { 334329c91e99STejun Heo bitmap_copy(&leftover, bitmap + nr_longs, nr_leftover); 334429c91e99STejun Heo hash = jhash(&leftover, sizeof(long), hash); 334529c91e99STejun Heo } 334629c91e99STejun Heo return hash; 334729c91e99STejun Heo } 334829c91e99STejun Heo 334929c91e99STejun Heo /* hash value of the content of @attr */ 335029c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 335129c91e99STejun Heo { 335229c91e99STejun Heo u32 hash = 0; 335329c91e99STejun Heo 335429c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 335529c91e99STejun Heo hash = jhash_bitmap(cpumask_bits(attrs->cpumask), nr_cpu_ids, hash); 335629c91e99STejun Heo return hash; 335729c91e99STejun Heo } 335829c91e99STejun Heo 335929c91e99STejun Heo /* content equality test */ 336029c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 336129c91e99STejun Heo const struct workqueue_attrs *b) 336229c91e99STejun Heo { 336329c91e99STejun Heo if (a->nice != b->nice) 336429c91e99STejun Heo return false; 336529c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 336629c91e99STejun Heo return false; 336729c91e99STejun Heo return true; 336829c91e99STejun Heo } 336929c91e99STejun Heo 33707a4e344cSTejun Heo /** 33717a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33727a4e344cSTejun Heo * @pool: worker_pool to initialize 33737a4e344cSTejun Heo * 33747a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 337529c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 337629c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 337729c91e99STejun Heo * on @pool safely to release it. 33787a4e344cSTejun Heo */ 33797a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33804e1a1f9aSTejun Heo { 33814e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 338229c91e99STejun Heo pool->id = -1; 338329c91e99STejun Heo pool->cpu = -1; 33844e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33854e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33864e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33874e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33884e1a1f9aSTejun Heo 33894e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33904e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33914e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33924e1a1f9aSTejun Heo 33934e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33944e1a1f9aSTejun Heo (unsigned long)pool); 33954e1a1f9aSTejun Heo 33964e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3397bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3398822d8405STejun Heo idr_init(&pool->worker_idr); 33997a4e344cSTejun Heo 340029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 340129c91e99STejun Heo pool->refcnt = 1; 340229c91e99STejun Heo 340329c91e99STejun Heo /* shouldn't fail above this point */ 34047a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34057a4e344cSTejun Heo if (!pool->attrs) 34067a4e344cSTejun Heo return -ENOMEM; 34077a4e344cSTejun Heo return 0; 34084e1a1f9aSTejun Heo } 34094e1a1f9aSTejun Heo 341029c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 341129c91e99STejun Heo { 341229c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 341329c91e99STejun Heo 3414822d8405STejun Heo idr_destroy(&pool->worker_idr); 341529c91e99STejun Heo free_workqueue_attrs(pool->attrs); 341629c91e99STejun Heo kfree(pool); 341729c91e99STejun Heo } 341829c91e99STejun Heo 341929c91e99STejun Heo /** 342029c91e99STejun Heo * put_unbound_pool - put a worker_pool 342129c91e99STejun Heo * @pool: worker_pool to put 342229c91e99STejun Heo * 342329c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3424c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3425c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3426c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 342729c91e99STejun Heo */ 342829c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 342929c91e99STejun Heo { 343029c91e99STejun Heo struct worker *worker; 343129c91e99STejun Heo 34325bcab335STejun Heo mutex_lock(&wq_mutex); 343329c91e99STejun Heo if (--pool->refcnt) { 34345bcab335STejun Heo mutex_unlock(&wq_mutex); 343529c91e99STejun Heo return; 343629c91e99STejun Heo } 343729c91e99STejun Heo 343829c91e99STejun Heo /* sanity checks */ 343929c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 344029c91e99STejun Heo WARN_ON(!list_empty(&pool->worklist))) { 34415bcab335STejun Heo mutex_unlock(&wq_mutex); 344229c91e99STejun Heo return; 344329c91e99STejun Heo } 344429c91e99STejun Heo 344529c91e99STejun Heo /* release id and unhash */ 344629c91e99STejun Heo if (pool->id >= 0) 344729c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 344829c91e99STejun Heo hash_del(&pool->hash_node); 344929c91e99STejun Heo 34505bcab335STejun Heo mutex_unlock(&wq_mutex); 345129c91e99STejun Heo 3452c5aa87bbSTejun Heo /* 3453c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3454c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3455c5aa87bbSTejun Heo * manager_mutex. 3456c5aa87bbSTejun Heo */ 345729c91e99STejun Heo mutex_lock(&pool->manager_arb); 3458cd549687STejun Heo mutex_lock(&pool->manager_mutex); 345929c91e99STejun Heo spin_lock_irq(&pool->lock); 346029c91e99STejun Heo 346129c91e99STejun Heo while ((worker = first_worker(pool))) 346229c91e99STejun Heo destroy_worker(worker); 346329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 346429c91e99STejun Heo 346529c91e99STejun Heo spin_unlock_irq(&pool->lock); 3466cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 346729c91e99STejun Heo mutex_unlock(&pool->manager_arb); 346829c91e99STejun Heo 346929c91e99STejun Heo /* shut down the timers */ 347029c91e99STejun Heo del_timer_sync(&pool->idle_timer); 347129c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 347229c91e99STejun Heo 347329c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 347429c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 347529c91e99STejun Heo } 347629c91e99STejun Heo 347729c91e99STejun Heo /** 347829c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 347929c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 348029c91e99STejun Heo * 348129c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 348229c91e99STejun Heo * reference count and return it. If there already is a matching 348329c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 348429c91e99STejun Heo * create a new one. On failure, returns NULL. 348529c91e99STejun Heo */ 348629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 348729c91e99STejun Heo { 348829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 348929c91e99STejun Heo struct worker_pool *pool; 349029c91e99STejun Heo 34915bcab335STejun Heo mutex_lock(&wq_mutex); 349229c91e99STejun Heo 349329c91e99STejun Heo /* do we already have a matching pool? */ 349429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 349529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 349629c91e99STejun Heo pool->refcnt++; 349729c91e99STejun Heo goto out_unlock; 349829c91e99STejun Heo } 349929c91e99STejun Heo } 350029c91e99STejun Heo 350129c91e99STejun Heo /* nope, create a new one */ 350229c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 350329c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 350429c91e99STejun Heo goto fail; 350529c91e99STejun Heo 350612ee4fc6SLai Jiangshan if (workqueue_freezing) 350712ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 350812ee4fc6SLai Jiangshan 35098864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 351029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 351129c91e99STejun Heo 351229c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 351329c91e99STejun Heo goto fail; 351429c91e99STejun Heo 351529c91e99STejun Heo /* create and start the initial worker */ 3516ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 351729c91e99STejun Heo goto fail; 351829c91e99STejun Heo 351929c91e99STejun Heo /* install */ 352029c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 352129c91e99STejun Heo out_unlock: 35225bcab335STejun Heo mutex_unlock(&wq_mutex); 352329c91e99STejun Heo return pool; 352429c91e99STejun Heo fail: 35255bcab335STejun Heo mutex_unlock(&wq_mutex); 352629c91e99STejun Heo if (pool) 352729c91e99STejun Heo put_unbound_pool(pool); 352829c91e99STejun Heo return NULL; 352929c91e99STejun Heo } 353029c91e99STejun Heo 35318864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35328864b4e5STejun Heo { 35338864b4e5STejun Heo kmem_cache_free(pwq_cache, 35348864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35358864b4e5STejun Heo } 35368864b4e5STejun Heo 35378864b4e5STejun Heo /* 35388864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35398864b4e5STejun Heo * and needs to be destroyed. 35408864b4e5STejun Heo */ 35418864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35428864b4e5STejun Heo { 35438864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35448864b4e5STejun Heo unbound_release_work); 35458864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35468864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 35478864b4e5STejun Heo 35488864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35498864b4e5STejun Heo return; 35508864b4e5STejun Heo 355175ccf595STejun Heo /* 355275ccf595STejun Heo * Unlink @pwq. Synchronization against flush_mutex isn't strictly 355375ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 355475ccf595STejun Heo * and consistent with the linking path. 355575ccf595STejun Heo */ 355675ccf595STejun Heo mutex_lock(&wq->flush_mutex); 3557794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 35588864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3559794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 356075ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 35618864b4e5STejun Heo 35628864b4e5STejun Heo put_unbound_pool(pool); 35638864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35648864b4e5STejun Heo 35658864b4e5STejun Heo /* 35668864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35678864b4e5STejun Heo * is gonna access it anymore. Free it. 35688864b4e5STejun Heo */ 35698864b4e5STejun Heo if (list_empty(&wq->pwqs)) 35708864b4e5STejun Heo kfree(wq); 35718864b4e5STejun Heo } 35728864b4e5STejun Heo 35730fbd95aaSTejun Heo /** 3574699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35750fbd95aaSTejun Heo * @pwq: target pool_workqueue 35760fbd95aaSTejun Heo * 3577699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3578699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3579699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35800fbd95aaSTejun Heo */ 3581699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35820fbd95aaSTejun Heo { 3583699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3584699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3585699ce097STejun Heo 3586699ce097STejun Heo /* for @wq->saved_max_active */ 3587794b18bcSTejun Heo lockdep_assert_held(&pwq_lock); 3588699ce097STejun Heo 3589699ce097STejun Heo /* fast exit for non-freezable wqs */ 3590699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3591699ce097STejun Heo return; 3592699ce097STejun Heo 3593699ce097STejun Heo spin_lock(&pwq->pool->lock); 3594699ce097STejun Heo 3595699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3596699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35970fbd95aaSTejun Heo 35980fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35990fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36000fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3601699ce097STejun Heo } else { 3602699ce097STejun Heo pwq->max_active = 0; 3603699ce097STejun Heo } 3604699ce097STejun Heo 3605699ce097STejun Heo spin_unlock(&pwq->pool->lock); 36060fbd95aaSTejun Heo } 36070fbd95aaSTejun Heo 3608d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3609d2c1d404STejun Heo struct workqueue_struct *wq, 36109e8cd2f5STejun Heo struct worker_pool *pool, 36119e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3612d2c1d404STejun Heo { 3613d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3614d2c1d404STejun Heo 3615d2c1d404STejun Heo pwq->pool = pool; 3616d2c1d404STejun Heo pwq->wq = wq; 3617d2c1d404STejun Heo pwq->flush_color = -1; 36188864b4e5STejun Heo pwq->refcnt = 1; 3619d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3620d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36218864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3622d2c1d404STejun Heo 362375ccf595STejun Heo mutex_lock(&wq->flush_mutex); 3624794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 362575ccf595STejun Heo 3626983ca25eSTejun Heo /* 3627983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 3628983ca25eSTejun Heo * flush_mutex to avoid confusing flush_workqueue(). 3629983ca25eSTejun Heo */ 36309e8cd2f5STejun Heo if (p_last_pwq) 36319e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 363275ccf595STejun Heo pwq->work_color = wq->work_color; 3633983ca25eSTejun Heo 3634983ca25eSTejun Heo /* sync max_active to the current setting */ 3635983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3636983ca25eSTejun Heo 3637983ca25eSTejun Heo /* link in @pwq */ 36389e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 363975ccf595STejun Heo 3640794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 364175ccf595STejun Heo mutex_unlock(&wq->flush_mutex); 3642d2c1d404STejun Heo } 3643d2c1d404STejun Heo 36449e8cd2f5STejun Heo /** 36459e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36469e8cd2f5STejun Heo * @wq: the target workqueue 36479e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36489e8cd2f5STejun Heo * 36499e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 36509e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 36519e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 36529e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 36539e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 36549e8cd2f5STejun Heo * 36559e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 36569e8cd2f5STejun Heo * failure. 36579e8cd2f5STejun Heo */ 36589e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36599e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36609e8cd2f5STejun Heo { 36619e8cd2f5STejun Heo struct pool_workqueue *pwq, *last_pwq; 36629e8cd2f5STejun Heo struct worker_pool *pool; 36639e8cd2f5STejun Heo 36648719dceaSTejun Heo /* only unbound workqueues can change attributes */ 36659e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 36669e8cd2f5STejun Heo return -EINVAL; 36679e8cd2f5STejun Heo 36688719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 36698719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 36708719dceaSTejun Heo return -EINVAL; 36718719dceaSTejun Heo 36729e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 36739e8cd2f5STejun Heo if (!pwq) 36749e8cd2f5STejun Heo return -ENOMEM; 36759e8cd2f5STejun Heo 36769e8cd2f5STejun Heo pool = get_unbound_pool(attrs); 36779e8cd2f5STejun Heo if (!pool) { 36789e8cd2f5STejun Heo kmem_cache_free(pwq_cache, pwq); 36799e8cd2f5STejun Heo return -ENOMEM; 36809e8cd2f5STejun Heo } 36819e8cd2f5STejun Heo 36829e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 36839e8cd2f5STejun Heo if (last_pwq) { 36849e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 36859e8cd2f5STejun Heo put_pwq(last_pwq); 36869e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 36879e8cd2f5STejun Heo } 36889e8cd2f5STejun Heo 36899e8cd2f5STejun Heo return 0; 36909e8cd2f5STejun Heo } 36919e8cd2f5STejun Heo 369230cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 36931da177e4SLinus Torvalds { 369449e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 369530cdf249STejun Heo int cpu; 3696e1d8aa9fSFrederic Weisbecker 369730cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3698420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3699420c0ddbSTejun Heo if (!wq->cpu_pwqs) 370030cdf249STejun Heo return -ENOMEM; 370130cdf249STejun Heo 370230cdf249STejun Heo for_each_possible_cpu(cpu) { 37037fb98ea7STejun Heo struct pool_workqueue *pwq = 37047fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37057a62c2c8STejun Heo struct worker_pool *cpu_pools = 3706f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 370730cdf249STejun Heo 37089e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 370930cdf249STejun Heo } 371030cdf249STejun Heo return 0; 37119e8cd2f5STejun Heo } else { 37129e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37139e8cd2f5STejun Heo } 37140f900049STejun Heo } 37150f900049STejun Heo 3716f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3717f3421797STejun Heo const char *name) 3718b71ab8c2STejun Heo { 3719f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3720f3421797STejun Heo 3721f3421797STejun Heo if (max_active < 1 || max_active > lim) 3722044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3723f3421797STejun Heo max_active, name, 1, lim); 3724b71ab8c2STejun Heo 3725f3421797STejun Heo return clamp_val(max_active, 1, lim); 3726b71ab8c2STejun Heo } 3727b71ab8c2STejun Heo 3728b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 372997e37d7bSTejun Heo unsigned int flags, 37301e19ffc6STejun Heo int max_active, 3731eb13ba87SJohannes Berg struct lock_class_key *key, 3732b196be89STejun Heo const char *lock_name, ...) 37333af24433SOleg Nesterov { 3734b196be89STejun Heo va_list args, args1; 37353af24433SOleg Nesterov struct workqueue_struct *wq; 373649e3cf44STejun Heo struct pool_workqueue *pwq; 3737b196be89STejun Heo size_t namelen; 3738b196be89STejun Heo 3739b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3740b196be89STejun Heo va_start(args, lock_name); 3741b196be89STejun Heo va_copy(args1, args); 3742b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3743b196be89STejun Heo 3744b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3745b196be89STejun Heo if (!wq) 3746d2c1d404STejun Heo return NULL; 3747b196be89STejun Heo 3748b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3749b196be89STejun Heo va_end(args); 3750b196be89STejun Heo va_end(args1); 37513af24433SOleg Nesterov 3752d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3753b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 37543af24433SOleg Nesterov 3755b196be89STejun Heo /* init wq */ 375697e37d7bSTejun Heo wq->flags = flags; 3757a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 375873f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 3759112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 376030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 376173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 376273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3763493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 37643af24433SOleg Nesterov 3765eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3766cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 37673af24433SOleg Nesterov 376830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3769d2c1d404STejun Heo goto err_free_wq; 37701537663fSTejun Heo 3771493008a8STejun Heo /* 3772493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3773493008a8STejun Heo * have a rescuer to guarantee forward progress. 3774493008a8STejun Heo */ 3775493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3776e22bee78STejun Heo struct worker *rescuer; 3777e22bee78STejun Heo 3778d2c1d404STejun Heo rescuer = alloc_worker(); 3779e22bee78STejun Heo if (!rescuer) 3780d2c1d404STejun Heo goto err_destroy; 3781e22bee78STejun Heo 3782111c225aSTejun Heo rescuer->rescue_wq = wq; 3783111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3784b196be89STejun Heo wq->name); 3785d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3786d2c1d404STejun Heo kfree(rescuer); 3787d2c1d404STejun Heo goto err_destroy; 3788d2c1d404STejun Heo } 3789e22bee78STejun Heo 3790d2c1d404STejun Heo wq->rescuer = rescuer; 379114a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3792e22bee78STejun Heo wake_up_process(rescuer->task); 37933af24433SOleg Nesterov } 37941537663fSTejun Heo 3795226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3796226223abSTejun Heo goto err_destroy; 3797226223abSTejun Heo 37983af24433SOleg Nesterov /* 37995bcab335STejun Heo * wq_mutex protects global freeze state and workqueues list. Grab 38005bcab335STejun Heo * it, adjust max_active and add the new @wq to workqueues list. 38013af24433SOleg Nesterov */ 38025bcab335STejun Heo mutex_lock(&wq_mutex); 3803a0a1a5fdSTejun Heo 3804794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 380549e3cf44STejun Heo for_each_pwq(pwq, wq) 3806699ce097STejun Heo pwq_adjust_max_active(pwq); 3807794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 3808a0a1a5fdSTejun Heo 38093af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3810a0a1a5fdSTejun Heo 38115bcab335STejun Heo mutex_unlock(&wq_mutex); 38123af24433SOleg Nesterov 38133af24433SOleg Nesterov return wq; 3814d2c1d404STejun Heo 3815d2c1d404STejun Heo err_free_wq: 38164690c4abSTejun Heo kfree(wq); 3817d2c1d404STejun Heo return NULL; 3818d2c1d404STejun Heo err_destroy: 3819d2c1d404STejun Heo destroy_workqueue(wq); 38204690c4abSTejun Heo return NULL; 38211da177e4SLinus Torvalds } 3822d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 38231da177e4SLinus Torvalds 38243af24433SOleg Nesterov /** 38253af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 38263af24433SOleg Nesterov * @wq: target workqueue 38273af24433SOleg Nesterov * 38283af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 38293af24433SOleg Nesterov */ 38303af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 38313af24433SOleg Nesterov { 383249e3cf44STejun Heo struct pool_workqueue *pwq; 38333af24433SOleg Nesterov 38349c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 38359c5a2ba7STejun Heo drain_workqueue(wq); 3836c8efcc25STejun Heo 38376183c009STejun Heo /* sanity checks */ 3838794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 383949e3cf44STejun Heo for_each_pwq(pwq, wq) { 38406183c009STejun Heo int i; 38416183c009STejun Heo 384276af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 384376af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3844794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 38456183c009STejun Heo return; 384676af4d93STejun Heo } 384776af4d93STejun Heo } 384876af4d93STejun Heo 38498864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 38508864b4e5STejun Heo WARN_ON(pwq->nr_active) || 385176af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3852794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 38536183c009STejun Heo return; 38546183c009STejun Heo } 385576af4d93STejun Heo } 3856794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 38576183c009STejun Heo 3858a0a1a5fdSTejun Heo /* 3859a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3860a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3861a0a1a5fdSTejun Heo */ 38625bcab335STejun Heo mutex_lock(&wq_mutex); 3863d2c1d404STejun Heo list_del_init(&wq->list); 38645bcab335STejun Heo mutex_unlock(&wq_mutex); 38653af24433SOleg Nesterov 3866226223abSTejun Heo workqueue_sysfs_unregister(wq); 3867226223abSTejun Heo 3868493008a8STejun Heo if (wq->rescuer) { 3869e22bee78STejun Heo kthread_stop(wq->rescuer->task); 38708d9df9f0SXiaotian Feng kfree(wq->rescuer); 3871493008a8STejun Heo wq->rescuer = NULL; 3872e22bee78STejun Heo } 3873e22bee78STejun Heo 38748864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 387529c91e99STejun Heo /* 38768864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 38778864b4e5STejun Heo * free the pwqs and wq. 387829c91e99STejun Heo */ 38798864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 38808864b4e5STejun Heo kfree(wq); 38818864b4e5STejun Heo } else { 38828864b4e5STejun Heo /* 38838864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 38848864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 38858864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 38868864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 38878864b4e5STejun Heo * released. 38888864b4e5STejun Heo */ 388929c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 389029c91e99STejun Heo pwqs_node); 38918864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 38928864b4e5STejun Heo put_pwq(pwq); 38938864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 389429c91e99STejun Heo } 38953af24433SOleg Nesterov } 38963af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 38973af24433SOleg Nesterov 3898dcd989cbSTejun Heo /** 3899dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3900dcd989cbSTejun Heo * @wq: target workqueue 3901dcd989cbSTejun Heo * @max_active: new max_active value. 3902dcd989cbSTejun Heo * 3903dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3904dcd989cbSTejun Heo * 3905dcd989cbSTejun Heo * CONTEXT: 3906dcd989cbSTejun Heo * Don't call from IRQ context. 3907dcd989cbSTejun Heo */ 3908dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3909dcd989cbSTejun Heo { 391049e3cf44STejun Heo struct pool_workqueue *pwq; 3911dcd989cbSTejun Heo 39128719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39138719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39148719dceaSTejun Heo return; 39158719dceaSTejun Heo 3916f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3917dcd989cbSTejun Heo 3918794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 3919dcd989cbSTejun Heo 3920dcd989cbSTejun Heo wq->saved_max_active = max_active; 3921dcd989cbSTejun Heo 3922699ce097STejun Heo for_each_pwq(pwq, wq) 3923699ce097STejun Heo pwq_adjust_max_active(pwq); 3924dcd989cbSTejun Heo 3925794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 3926dcd989cbSTejun Heo } 3927dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3928dcd989cbSTejun Heo 3929dcd989cbSTejun Heo /** 3930e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 3931e6267616STejun Heo * 3932e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 3933e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 3934e6267616STejun Heo */ 3935e6267616STejun Heo bool current_is_workqueue_rescuer(void) 3936e6267616STejun Heo { 3937e6267616STejun Heo struct worker *worker = current_wq_worker(); 3938e6267616STejun Heo 39396a092dfdSLai Jiangshan return worker && worker->rescue_wq; 3940e6267616STejun Heo } 3941e6267616STejun Heo 3942e6267616STejun Heo /** 3943dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3944dcd989cbSTejun Heo * @cpu: CPU in question 3945dcd989cbSTejun Heo * @wq: target workqueue 3946dcd989cbSTejun Heo * 3947dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3948dcd989cbSTejun Heo * no synchronization around this function and the test result is 3949dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3950dcd989cbSTejun Heo * 3951dcd989cbSTejun Heo * RETURNS: 3952dcd989cbSTejun Heo * %true if congested, %false otherwise. 3953dcd989cbSTejun Heo */ 3954d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 3955dcd989cbSTejun Heo { 39567fb98ea7STejun Heo struct pool_workqueue *pwq; 395776af4d93STejun Heo bool ret; 395876af4d93STejun Heo 395976af4d93STejun Heo preempt_disable(); 39607fb98ea7STejun Heo 39617fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 39627fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 39637fb98ea7STejun Heo else 39647fb98ea7STejun Heo pwq = first_pwq(wq); 3965dcd989cbSTejun Heo 396676af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 396776af4d93STejun Heo preempt_enable(); 396876af4d93STejun Heo 396976af4d93STejun Heo return ret; 3970dcd989cbSTejun Heo } 3971dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3972dcd989cbSTejun Heo 3973dcd989cbSTejun Heo /** 3974dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3975dcd989cbSTejun Heo * @work: the work to be tested 3976dcd989cbSTejun Heo * 3977dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3978dcd989cbSTejun Heo * synchronization around this function and the test result is 3979dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3980dcd989cbSTejun Heo * 3981dcd989cbSTejun Heo * RETURNS: 3982dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3983dcd989cbSTejun Heo */ 3984dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3985dcd989cbSTejun Heo { 3986fa1b54e6STejun Heo struct worker_pool *pool; 3987dcd989cbSTejun Heo unsigned long flags; 3988dcd989cbSTejun Heo unsigned int ret = 0; 3989dcd989cbSTejun Heo 3990dcd989cbSTejun Heo if (work_pending(work)) 3991dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3992038366c5SLai Jiangshan 3993fa1b54e6STejun Heo local_irq_save(flags); 3994fa1b54e6STejun Heo pool = get_work_pool(work); 3995038366c5SLai Jiangshan if (pool) { 3996fa1b54e6STejun Heo spin_lock(&pool->lock); 3997c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3998dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3999fa1b54e6STejun Heo spin_unlock(&pool->lock); 4000038366c5SLai Jiangshan } 4001fa1b54e6STejun Heo local_irq_restore(flags); 4002dcd989cbSTejun Heo 4003dcd989cbSTejun Heo return ret; 4004dcd989cbSTejun Heo } 4005dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4006dcd989cbSTejun Heo 4007db7bccf4STejun Heo /* 4008db7bccf4STejun Heo * CPU hotplug. 4009db7bccf4STejun Heo * 4010e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4011112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4012706026c2STejun Heo * pool which make migrating pending and scheduled works very 4013e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 401494cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4015e22bee78STejun Heo * blocked draining impractical. 4016e22bee78STejun Heo * 401724647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4018628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4019628c78e7STejun Heo * cpu comes back online. 4020db7bccf4STejun Heo */ 4021db7bccf4STejun Heo 4022706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4023db7bccf4STejun Heo { 402438db41d9STejun Heo int cpu = smp_processor_id(); 40254ce62e9eSTejun Heo struct worker_pool *pool; 4026db7bccf4STejun Heo struct worker *worker; 4027a9ab775bSTejun Heo int wi; 4028db7bccf4STejun Heo 4029f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 40306183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4031db7bccf4STejun Heo 4032bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 403394cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4034e22bee78STejun Heo 4035f2d5a0eeSTejun Heo /* 4036bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 403794cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 403894cf58bbSTejun Heo * except for the ones which are still executing works from 403994cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 404094cf58bbSTejun Heo * this, they may become diasporas. 4041f2d5a0eeSTejun Heo */ 4042a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4043403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4044db7bccf4STejun Heo 404524647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4046f2d5a0eeSTejun Heo 404794cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4048bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 404994cf58bbSTejun Heo } 4050e22bee78STejun Heo 4051e22bee78STejun Heo /* 4052628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4053628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4054628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4055628c78e7STejun Heo */ 4056628c78e7STejun Heo schedule(); 4057628c78e7STejun Heo 4058628c78e7STejun Heo /* 4059628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4060628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 406138db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 406238db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 406338db41d9STejun Heo * pools which are served by workers tied to the CPU. 4064628c78e7STejun Heo * 4065628c78e7STejun Heo * On return from this function, the current worker would trigger 4066628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4067628c78e7STejun Heo * didn't already. 4068e22bee78STejun Heo */ 4069f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4070e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4071db7bccf4STejun Heo } 4072db7bccf4STejun Heo 4073bd7c089eSTejun Heo /** 4074bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4075bd7c089eSTejun Heo * @pool: pool of interest 4076bd7c089eSTejun Heo * 4077a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4078bd7c089eSTejun Heo */ 4079bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4080bd7c089eSTejun Heo { 4081a9ab775bSTejun Heo struct worker *worker; 4082a9ab775bSTejun Heo int wi; 4083bd7c089eSTejun Heo 4084bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4085bd7c089eSTejun Heo 4086bd7c089eSTejun Heo /* 4087a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4088a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4089a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4090a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4091a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4092bd7c089eSTejun Heo */ 4093a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4094a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4095a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4096a9ab775bSTejun Heo 4097a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4098a9ab775bSTejun Heo 4099a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4100a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4101a9ab775bSTejun Heo 4102a9ab775bSTejun Heo /* 4103a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4104a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4105a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4106a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4107a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4108a9ab775bSTejun Heo * be bound before @pool->lock is released. 4109a9ab775bSTejun Heo */ 4110a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4111bd7c089eSTejun Heo wake_up_process(worker->task); 4112bd7c089eSTejun Heo 4113bd7c089eSTejun Heo /* 4114a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4115a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4116a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4117a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4118a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4119a9ab775bSTejun Heo * concurrency management. Note that when or whether 4120a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4121a9ab775bSTejun Heo * 4122a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4123a9ab775bSTejun Heo * tested without holding any lock in 4124a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4125a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4126a9ab775bSTejun Heo * management operations. 4127bd7c089eSTejun Heo */ 4128a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4129a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4130a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4131a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4132bd7c089eSTejun Heo } 4133a9ab775bSTejun Heo 4134a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4135bd7c089eSTejun Heo } 4136bd7c089eSTejun Heo 41377dbc725eSTejun Heo /** 41387dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 41397dbc725eSTejun Heo * @pool: unbound pool of interest 41407dbc725eSTejun Heo * @cpu: the CPU which is coming up 41417dbc725eSTejun Heo * 41427dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 41437dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 41447dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 41457dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 41467dbc725eSTejun Heo */ 41477dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 41487dbc725eSTejun Heo { 41497dbc725eSTejun Heo static cpumask_t cpumask; 41507dbc725eSTejun Heo struct worker *worker; 41517dbc725eSTejun Heo int wi; 41527dbc725eSTejun Heo 41537dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 41547dbc725eSTejun Heo 41557dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 41567dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 41577dbc725eSTejun Heo return; 41587dbc725eSTejun Heo 41597dbc725eSTejun Heo /* is @cpu the only online CPU? */ 41607dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 41617dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 41627dbc725eSTejun Heo return; 41637dbc725eSTejun Heo 41647dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 41657dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 41667dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 41677dbc725eSTejun Heo pool->attrs->cpumask) < 0); 41687dbc725eSTejun Heo } 41697dbc725eSTejun Heo 41708db25e78STejun Heo /* 41718db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 41728db25e78STejun Heo * This will be registered high priority CPU notifier. 41738db25e78STejun Heo */ 41749fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 41751da177e4SLinus Torvalds unsigned long action, 41761da177e4SLinus Torvalds void *hcpu) 41771da177e4SLinus Torvalds { 4178d84ff051STejun Heo int cpu = (unsigned long)hcpu; 41794ce62e9eSTejun Heo struct worker_pool *pool; 41807dbc725eSTejun Heo int pi; 41811da177e4SLinus Torvalds 41828db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 41833af24433SOleg Nesterov case CPU_UP_PREPARE: 4184f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41853ce63377STejun Heo if (pool->nr_workers) 41863ce63377STejun Heo continue; 4187ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 41883ce63377STejun Heo return NOTIFY_BAD; 41893af24433SOleg Nesterov } 41901da177e4SLinus Torvalds break; 41911da177e4SLinus Torvalds 419265758202STejun Heo case CPU_DOWN_FAILED: 419365758202STejun Heo case CPU_ONLINE: 41947dbc725eSTejun Heo mutex_lock(&wq_mutex); 41957dbc725eSTejun Heo 41967dbc725eSTejun Heo for_each_pool(pool, pi) { 4197bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 419894cf58bbSTejun Heo 41997dbc725eSTejun Heo if (pool->cpu == cpu) { 4200a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 420124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4202a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4203a9ab775bSTejun Heo 420494cf58bbSTejun Heo rebind_workers(pool); 42057dbc725eSTejun Heo } else if (pool->cpu < 0) { 42067dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 42077dbc725eSTejun Heo } 420894cf58bbSTejun Heo 4209bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 421094cf58bbSTejun Heo } 42117dbc725eSTejun Heo 42127dbc725eSTejun Heo mutex_unlock(&wq_mutex); 42138db25e78STejun Heo break; 421465758202STejun Heo } 421565758202STejun Heo return NOTIFY_OK; 421665758202STejun Heo } 421765758202STejun Heo 421865758202STejun Heo /* 421965758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 422065758202STejun Heo * This will be registered as low priority CPU notifier. 422165758202STejun Heo */ 42229fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 422365758202STejun Heo unsigned long action, 422465758202STejun Heo void *hcpu) 422565758202STejun Heo { 4226d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42278db25e78STejun Heo struct work_struct unbind_work; 42288db25e78STejun Heo 422965758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 423065758202STejun Heo case CPU_DOWN_PREPARE: 42318db25e78STejun Heo /* unbinding should happen on the local CPU */ 4232706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42337635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42348db25e78STejun Heo flush_work(&unbind_work); 42358db25e78STejun Heo break; 423665758202STejun Heo } 423765758202STejun Heo return NOTIFY_OK; 423865758202STejun Heo } 423965758202STejun Heo 42402d3854a3SRusty Russell #ifdef CONFIG_SMP 42418ccad40dSRusty Russell 42422d3854a3SRusty Russell struct work_for_cpu { 4243ed48ece2STejun Heo struct work_struct work; 42442d3854a3SRusty Russell long (*fn)(void *); 42452d3854a3SRusty Russell void *arg; 42462d3854a3SRusty Russell long ret; 42472d3854a3SRusty Russell }; 42482d3854a3SRusty Russell 4249ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 42502d3854a3SRusty Russell { 4251ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4252ed48ece2STejun Heo 42532d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 42542d3854a3SRusty Russell } 42552d3854a3SRusty Russell 42562d3854a3SRusty Russell /** 42572d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 42582d3854a3SRusty Russell * @cpu: the cpu to run on 42592d3854a3SRusty Russell * @fn: the function to run 42602d3854a3SRusty Russell * @arg: the function arg 42612d3854a3SRusty Russell * 426231ad9081SRusty Russell * This will return the value @fn returns. 426331ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 42646b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 42652d3854a3SRusty Russell */ 4266d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 42672d3854a3SRusty Russell { 4268ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 42692d3854a3SRusty Russell 4270ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4271ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4272ed48ece2STejun Heo flush_work(&wfc.work); 42732d3854a3SRusty Russell return wfc.ret; 42742d3854a3SRusty Russell } 42752d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 42762d3854a3SRusty Russell #endif /* CONFIG_SMP */ 42772d3854a3SRusty Russell 4278a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4279e7577c50SRusty Russell 4280a0a1a5fdSTejun Heo /** 4281a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4282a0a1a5fdSTejun Heo * 428358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4284c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4285706026c2STejun Heo * pool->worklist. 4286a0a1a5fdSTejun Heo * 4287a0a1a5fdSTejun Heo * CONTEXT: 4288794b18bcSTejun Heo * Grabs and releases wq_mutex, pwq_lock and pool->lock's. 4289a0a1a5fdSTejun Heo */ 4290a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4291a0a1a5fdSTejun Heo { 429217116969STejun Heo struct worker_pool *pool; 429324b8a847STejun Heo struct workqueue_struct *wq; 429424b8a847STejun Heo struct pool_workqueue *pwq; 4295611c92a0STejun Heo int pi; 4296a0a1a5fdSTejun Heo 42975bcab335STejun Heo mutex_lock(&wq_mutex); 4298a0a1a5fdSTejun Heo 42996183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4300a0a1a5fdSTejun Heo workqueue_freezing = true; 4301a0a1a5fdSTejun Heo 430224b8a847STejun Heo /* set FREEZING */ 4303611c92a0STejun Heo for_each_pool(pool, pi) { 43045bcab335STejun Heo spin_lock_irq(&pool->lock); 430535b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 430635b6bb63STejun Heo pool->flags |= POOL_FREEZING; 43075bcab335STejun Heo spin_unlock_irq(&pool->lock); 43081da177e4SLinus Torvalds } 43098b03ae3cSTejun Heo 431024b8a847STejun Heo /* suppress further executions by setting max_active to zero */ 4311794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 431224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4313699ce097STejun Heo for_each_pwq(pwq, wq) 4314699ce097STejun Heo pwq_adjust_max_active(pwq); 4315a1056305STejun Heo } 4316794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 43175bcab335STejun Heo 43185bcab335STejun Heo mutex_unlock(&wq_mutex); 4319a0a1a5fdSTejun Heo } 4320a0a1a5fdSTejun Heo 4321a0a1a5fdSTejun Heo /** 432258a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4323a0a1a5fdSTejun Heo * 4324a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4325a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4326a0a1a5fdSTejun Heo * 4327a0a1a5fdSTejun Heo * CONTEXT: 43285bcab335STejun Heo * Grabs and releases wq_mutex. 4329a0a1a5fdSTejun Heo * 4330a0a1a5fdSTejun Heo * RETURNS: 433158a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 433258a69cb4STejun Heo * is complete. 4333a0a1a5fdSTejun Heo */ 4334a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4335a0a1a5fdSTejun Heo { 4336a0a1a5fdSTejun Heo bool busy = false; 433724b8a847STejun Heo struct workqueue_struct *wq; 433824b8a847STejun Heo struct pool_workqueue *pwq; 4339a0a1a5fdSTejun Heo 43405bcab335STejun Heo mutex_lock(&wq_mutex); 4341a0a1a5fdSTejun Heo 43426183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4343a0a1a5fdSTejun Heo 434424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 434524b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 434624b8a847STejun Heo continue; 4347a0a1a5fdSTejun Heo /* 4348a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4349a0a1a5fdSTejun Heo * to peek without lock. 4350a0a1a5fdSTejun Heo */ 43515bcab335STejun Heo preempt_disable(); 435224b8a847STejun Heo for_each_pwq(pwq, wq) { 43536183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4354112202d9STejun Heo if (pwq->nr_active) { 4355a0a1a5fdSTejun Heo busy = true; 43565bcab335STejun Heo preempt_enable(); 4357a0a1a5fdSTejun Heo goto out_unlock; 4358a0a1a5fdSTejun Heo } 4359a0a1a5fdSTejun Heo } 43605bcab335STejun Heo preempt_enable(); 4361a0a1a5fdSTejun Heo } 4362a0a1a5fdSTejun Heo out_unlock: 43635bcab335STejun Heo mutex_unlock(&wq_mutex); 4364a0a1a5fdSTejun Heo return busy; 4365a0a1a5fdSTejun Heo } 4366a0a1a5fdSTejun Heo 4367a0a1a5fdSTejun Heo /** 4368a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4369a0a1a5fdSTejun Heo * 4370a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4371706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4372a0a1a5fdSTejun Heo * 4373a0a1a5fdSTejun Heo * CONTEXT: 4374794b18bcSTejun Heo * Grabs and releases wq_mutex, pwq_lock and pool->lock's. 4375a0a1a5fdSTejun Heo */ 4376a0a1a5fdSTejun Heo void thaw_workqueues(void) 4377a0a1a5fdSTejun Heo { 437824b8a847STejun Heo struct workqueue_struct *wq; 437924b8a847STejun Heo struct pool_workqueue *pwq; 438024b8a847STejun Heo struct worker_pool *pool; 4381611c92a0STejun Heo int pi; 4382a0a1a5fdSTejun Heo 43835bcab335STejun Heo mutex_lock(&wq_mutex); 4384a0a1a5fdSTejun Heo 4385a0a1a5fdSTejun Heo if (!workqueue_freezing) 4386a0a1a5fdSTejun Heo goto out_unlock; 4387a0a1a5fdSTejun Heo 438824b8a847STejun Heo /* clear FREEZING */ 4389611c92a0STejun Heo for_each_pool(pool, pi) { 43905bcab335STejun Heo spin_lock_irq(&pool->lock); 439135b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 439235b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 43935bcab335STejun Heo spin_unlock_irq(&pool->lock); 4394d565ed63STejun Heo } 439524b8a847STejun Heo 439624b8a847STejun Heo /* restore max_active and repopulate worklist */ 4397794b18bcSTejun Heo spin_lock_irq(&pwq_lock); 439824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4399699ce097STejun Heo for_each_pwq(pwq, wq) 4400699ce097STejun Heo pwq_adjust_max_active(pwq); 440124b8a847STejun Heo } 4402794b18bcSTejun Heo spin_unlock_irq(&pwq_lock); 440324b8a847STejun Heo 440424b8a847STejun Heo /* kick workers */ 4405611c92a0STejun Heo for_each_pool(pool, pi) { 44065bcab335STejun Heo spin_lock_irq(&pool->lock); 440724b8a847STejun Heo wake_up_worker(pool); 44085bcab335STejun Heo spin_unlock_irq(&pool->lock); 4409a0a1a5fdSTejun Heo } 4410a0a1a5fdSTejun Heo 4411a0a1a5fdSTejun Heo workqueue_freezing = false; 4412a0a1a5fdSTejun Heo out_unlock: 44135bcab335STejun Heo mutex_unlock(&wq_mutex); 4414a0a1a5fdSTejun Heo } 4415a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4416a0a1a5fdSTejun Heo 44176ee0578bSSuresh Siddha static int __init init_workqueues(void) 44181da177e4SLinus Torvalds { 44197a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44207a4e344cSTejun Heo int i, cpu; 4421c34056a3STejun Heo 44227c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44237c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44246be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4425b5490077STejun Heo 4426e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4427e904e6c2STejun Heo 4428e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4429e904e6c2STejun Heo 443065758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4431a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44328b03ae3cSTejun Heo 4433706026c2STejun Heo /* initialize CPU pools */ 443429c91e99STejun Heo for_each_possible_cpu(cpu) { 44354ce62e9eSTejun Heo struct worker_pool *pool; 44368b03ae3cSTejun Heo 44377a4e344cSTejun Heo i = 0; 4438f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44397a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4440ec22ca5eSTejun Heo pool->cpu = cpu; 44417a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 44427a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 44437a4e344cSTejun Heo 44449daf9e67STejun Heo /* alloc pool ID */ 44455bcab335STejun Heo mutex_lock(&wq_mutex); 44469daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 44475bcab335STejun Heo mutex_unlock(&wq_mutex); 44484ce62e9eSTejun Heo } 44498b03ae3cSTejun Heo } 44508b03ae3cSTejun Heo 4451e22bee78STejun Heo /* create the initial worker */ 445229c91e99STejun Heo for_each_online_cpu(cpu) { 44534ce62e9eSTejun Heo struct worker_pool *pool; 4454e22bee78STejun Heo 4455f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 445624647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4457ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4458e22bee78STejun Heo } 44594ce62e9eSTejun Heo } 4460e22bee78STejun Heo 446129c91e99STejun Heo /* create default unbound wq attrs */ 446229c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 446329c91e99STejun Heo struct workqueue_attrs *attrs; 446429c91e99STejun Heo 446529c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 446629c91e99STejun Heo 446729c91e99STejun Heo attrs->nice = std_nice[i]; 446829c91e99STejun Heo cpumask_setall(attrs->cpumask); 446929c91e99STejun Heo 447029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 447129c91e99STejun Heo } 447229c91e99STejun Heo 4473d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 44741aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4475d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4476f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4477f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 447824d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 447924d51addSTejun Heo WQ_FREEZABLE, 0); 44801aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4481ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 44826ee0578bSSuresh Siddha return 0; 44831da177e4SLinus Torvalds } 44846ee0578bSSuresh Siddha early_initcall(init_workqueues); 4485