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> 47bce90380STejun Heo #include <linux/nodemask.h> 48e22bee78STejun Heo 49ea138446STejun Heo #include "workqueue_internal.h" 501da177e4SLinus Torvalds 51c8e55f36STejun Heo enum { 52bc2ae0f5STejun Heo /* 5324647570STejun Heo * worker_pool flags 54bc2ae0f5STejun Heo * 5524647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 56bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 57bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 58bc2ae0f5STejun Heo * is in effect. 59bc2ae0f5STejun Heo * 60bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 61bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6224647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 63bc2ae0f5STejun Heo * 64bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 65bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6624647570STejun Heo * create_worker() is in progress. 67bc2ae0f5STejun Heo */ 6811ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 6924647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 7035b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 71db7bccf4STejun Heo 72c8e55f36STejun Heo /* worker flags */ 73c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 74c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 75c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 76e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 77fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 78f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 79a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 80e22bee78STejun Heo 81a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 82a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 83db7bccf4STejun Heo 84e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 854ce62e9eSTejun Heo 8629c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 87c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 88db7bccf4STejun Heo 89e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 90e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 91e22bee78STejun Heo 923233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 933233cdbdSTejun Heo /* call for help after 10ms 943233cdbdSTejun Heo (min two ticks) */ 95e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 96e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* 99e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 100e22bee78STejun Heo * all cpus. Give -20. 101e22bee78STejun Heo */ 102e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1033270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 104c8e55f36STejun Heo }; 105c8e55f36STejun Heo 1061da177e4SLinus Torvalds /* 1074690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1084690c4abSTejun Heo * 109e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 110e41e704bSTejun Heo * everyone else. 1114690c4abSTejun Heo * 112e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 113e22bee78STejun Heo * only be modified and accessed from the local cpu. 114e22bee78STejun Heo * 115d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1164690c4abSTejun Heo * 117d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 118d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 119d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 120d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 121e22bee78STejun Heo * 122822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 123822d8405STejun Heo * locks. Reads can happen under either lock. 124822d8405STejun Heo * 12568e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12676af4d93STejun Heo * 12768e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1285bcab335STejun Heo * 1293c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1303c25a55dSLai Jiangshan * 131b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1322e109a28STejun Heo * 1332e109a28STejun Heo * MD: wq_mayday_lock protected. 1344690c4abSTejun Heo */ 1354690c4abSTejun Heo 1362eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 137c34056a3STejun Heo 138bd7bdd43STejun Heo struct worker_pool { 139d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 140d84ff051STejun Heo int cpu; /* I: the associated cpu */ 141f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1429daf9e67STejun Heo int id; /* I: pool ID */ 14311ebea50STejun Heo unsigned int flags; /* X: flags */ 144bd7bdd43STejun Heo 145bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 146bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 147ea1abd61SLai Jiangshan 148ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 149bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 150bd7bdd43STejun Heo 151bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 152bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 153bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 154bd7bdd43STejun Heo 155c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 156c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 157c9e7cf27STejun Heo /* L: hash of busy workers */ 158c9e7cf27STejun Heo 159bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16034a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 161bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 162822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 163e19e397aSTejun Heo 1647a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16568e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 16668e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1677a4e344cSTejun Heo 168e19e397aSTejun Heo /* 169e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 170e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 171e19e397aSTejun Heo * cacheline. 172e19e397aSTejun Heo */ 173e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17429c91e99STejun Heo 17529c91e99STejun Heo /* 17629c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 17729c91e99STejun Heo * from get_work_pool(). 17829c91e99STejun Heo */ 17929c91e99STejun Heo struct rcu_head rcu; 1808b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1818b03ae3cSTejun Heo 1828b03ae3cSTejun Heo /* 183112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 184112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 185112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 186112202d9STejun Heo * number of flag bits. 1871da177e4SLinus Torvalds */ 188112202d9STejun Heo struct pool_workqueue { 189bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1904690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19173f53c4aSTejun Heo int work_color; /* L: current color */ 19273f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1938864b4e5STejun Heo int refcnt; /* L: reference count */ 19473f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19573f53c4aSTejun Heo /* L: nr of in_flight works */ 1961e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 197a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1981e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1993c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2002e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2018864b4e5STejun Heo 2028864b4e5STejun Heo /* 2038864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2048864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2058864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 206b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2078864b4e5STejun Heo */ 2088864b4e5STejun Heo struct work_struct unbound_release_work; 2098864b4e5STejun Heo struct rcu_head rcu; 210e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2111da177e4SLinus Torvalds 2121da177e4SLinus Torvalds /* 21373f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21473f53c4aSTejun Heo */ 21573f53c4aSTejun Heo struct wq_flusher { 2163c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2173c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 21873f53c4aSTejun Heo struct completion done; /* flush completion */ 21973f53c4aSTejun Heo }; 2201da177e4SLinus Torvalds 221226223abSTejun Heo struct wq_device; 222226223abSTejun Heo 22373f53c4aSTejun Heo /* 224c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 225c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2261da177e4SLinus Torvalds */ 2271da177e4SLinus Torvalds struct workqueue_struct { 22887fc741eSLai Jiangshan unsigned int flags; /* WQ: WQ_* flags */ 229420c0ddbSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ 2303c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 23168e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23273f53c4aSTejun Heo 2333c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2343c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2353c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 236112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2373c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2383c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2393c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24073f53c4aSTejun Heo 2412e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 242e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 243e22bee78STejun Heo 24487fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 245a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 246226223abSTejun Heo 247226223abSTejun Heo #ifdef CONFIG_SYSFS 248226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 249226223abSTejun Heo #endif 2504e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2514e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2524e6045f1SJohannes Berg #endif 253b196be89STejun Heo char name[]; /* I: workqueue name */ 2541da177e4SLinus Torvalds }; 2551da177e4SLinus Torvalds 256e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 257e904e6c2STejun Heo 258bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 259bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 260bce90380STejun Heo /* possible CPUs of each node */ 261bce90380STejun Heo 262bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 263bce90380STejun Heo 26468e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2652e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2665bcab335STejun Heo 26768e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 26868e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2697d19c5ceSTejun Heo 2707d19c5ceSTejun Heo /* the per-cpu worker pools */ 2717d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2727d19c5ceSTejun Heo cpu_worker_pools); 2737d19c5ceSTejun Heo 27468e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2757d19c5ceSTejun Heo 27668e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 27729c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 27829c91e99STejun Heo 279c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 28029c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 28129c91e99STejun Heo 282d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 283d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 284044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2851aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 286044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 287d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 288044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 289f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 290044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 29124d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 292d320c038STejun Heo 2937d19c5ceSTejun Heo static int worker_thread(void *__worker); 2947d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 2957d19c5ceSTejun Heo const struct workqueue_attrs *from); 2967d19c5ceSTejun Heo 29797bd2347STejun Heo #define CREATE_TRACE_POINTS 29897bd2347STejun Heo #include <trace/events/workqueue.h> 29997bd2347STejun Heo 30068e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3015bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 30268e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 30368e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3045bcab335STejun Heo 305b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 30676af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 307b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 308b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 30976af4d93STejun Heo 310822d8405STejun Heo #ifdef CONFIG_LOCKDEP 311822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 312519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 313519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 314822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 315822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 316822d8405STejun Heo #else 317822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 318822d8405STejun Heo #endif 319822d8405STejun Heo 320f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 321f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 322f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3237a62c2c8STejun Heo (pool)++) 3244ce62e9eSTejun Heo 32549e3cf44STejun Heo /** 32617116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 32717116969STejun Heo * @pool: iteration cursor 328611c92a0STejun Heo * @pi: integer used for iteration 329fa1b54e6STejun Heo * 33068e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 33168e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 33268e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 333fa1b54e6STejun Heo * 334fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 335fa1b54e6STejun Heo * ignored. 33617116969STejun Heo */ 337611c92a0STejun Heo #define for_each_pool(pool, pi) \ 338611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 33968e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 340fa1b54e6STejun Heo else 34117116969STejun Heo 34217116969STejun Heo /** 343822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 344822d8405STejun Heo * @worker: iteration cursor 345822d8405STejun Heo * @wi: integer used for iteration 346822d8405STejun Heo * @pool: worker_pool to iterate workers of 347822d8405STejun Heo * 348822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 349822d8405STejun Heo * 350822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 351822d8405STejun Heo * ignored. 352822d8405STejun Heo */ 353822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 354822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 355822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 356822d8405STejun Heo else 357822d8405STejun Heo 358822d8405STejun Heo /** 35949e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 36049e3cf44STejun Heo * @pwq: iteration cursor 36149e3cf44STejun Heo * @wq: the target workqueue 36276af4d93STejun Heo * 363b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 364794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 365794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 36676af4d93STejun Heo * 36776af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 36876af4d93STejun Heo * ignored. 36949e3cf44STejun Heo */ 37049e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 37176af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 372b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 37376af4d93STejun Heo else 374f3421797STejun Heo 375dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 376dc186ad7SThomas Gleixner 377dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 378dc186ad7SThomas Gleixner 37999777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 38099777288SStanislaw Gruszka { 38199777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 38299777288SStanislaw Gruszka } 38399777288SStanislaw Gruszka 384dc186ad7SThomas Gleixner /* 385dc186ad7SThomas Gleixner * fixup_init is called when: 386dc186ad7SThomas Gleixner * - an active object is initialized 387dc186ad7SThomas Gleixner */ 388dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 389dc186ad7SThomas Gleixner { 390dc186ad7SThomas Gleixner struct work_struct *work = addr; 391dc186ad7SThomas Gleixner 392dc186ad7SThomas Gleixner switch (state) { 393dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 394dc186ad7SThomas Gleixner cancel_work_sync(work); 395dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 396dc186ad7SThomas Gleixner return 1; 397dc186ad7SThomas Gleixner default: 398dc186ad7SThomas Gleixner return 0; 399dc186ad7SThomas Gleixner } 400dc186ad7SThomas Gleixner } 401dc186ad7SThomas Gleixner 402dc186ad7SThomas Gleixner /* 403dc186ad7SThomas Gleixner * fixup_activate is called when: 404dc186ad7SThomas Gleixner * - an active object is activated 405dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 406dc186ad7SThomas Gleixner */ 407dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 408dc186ad7SThomas Gleixner { 409dc186ad7SThomas Gleixner struct work_struct *work = addr; 410dc186ad7SThomas Gleixner 411dc186ad7SThomas Gleixner switch (state) { 412dc186ad7SThomas Gleixner 413dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 414dc186ad7SThomas Gleixner /* 415dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 416dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 417dc186ad7SThomas Gleixner * is tracked in the object tracker. 418dc186ad7SThomas Gleixner */ 41922df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 420dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 421dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 422dc186ad7SThomas Gleixner return 0; 423dc186ad7SThomas Gleixner } 424dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 425dc186ad7SThomas Gleixner return 0; 426dc186ad7SThomas Gleixner 427dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 428dc186ad7SThomas Gleixner WARN_ON(1); 429dc186ad7SThomas Gleixner 430dc186ad7SThomas Gleixner default: 431dc186ad7SThomas Gleixner return 0; 432dc186ad7SThomas Gleixner } 433dc186ad7SThomas Gleixner } 434dc186ad7SThomas Gleixner 435dc186ad7SThomas Gleixner /* 436dc186ad7SThomas Gleixner * fixup_free is called when: 437dc186ad7SThomas Gleixner * - an active object is freed 438dc186ad7SThomas Gleixner */ 439dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 440dc186ad7SThomas Gleixner { 441dc186ad7SThomas Gleixner struct work_struct *work = addr; 442dc186ad7SThomas Gleixner 443dc186ad7SThomas Gleixner switch (state) { 444dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 445dc186ad7SThomas Gleixner cancel_work_sync(work); 446dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 447dc186ad7SThomas Gleixner return 1; 448dc186ad7SThomas Gleixner default: 449dc186ad7SThomas Gleixner return 0; 450dc186ad7SThomas Gleixner } 451dc186ad7SThomas Gleixner } 452dc186ad7SThomas Gleixner 453dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 454dc186ad7SThomas Gleixner .name = "work_struct", 45599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 456dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 457dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 458dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 459dc186ad7SThomas Gleixner }; 460dc186ad7SThomas Gleixner 461dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 462dc186ad7SThomas Gleixner { 463dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 464dc186ad7SThomas Gleixner } 465dc186ad7SThomas Gleixner 466dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 467dc186ad7SThomas Gleixner { 468dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 469dc186ad7SThomas Gleixner } 470dc186ad7SThomas Gleixner 471dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 472dc186ad7SThomas Gleixner { 473dc186ad7SThomas Gleixner if (onstack) 474dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 475dc186ad7SThomas Gleixner else 476dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 477dc186ad7SThomas Gleixner } 478dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 479dc186ad7SThomas Gleixner 480dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 481dc186ad7SThomas Gleixner { 482dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 483dc186ad7SThomas Gleixner } 484dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 485dc186ad7SThomas Gleixner 486dc186ad7SThomas Gleixner #else 487dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 488dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 489dc186ad7SThomas Gleixner #endif 490dc186ad7SThomas Gleixner 4919daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4929daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4939daf9e67STejun Heo { 4949daf9e67STejun Heo int ret; 4959daf9e67STejun Heo 49668e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4975bcab335STejun Heo 498fa1b54e6STejun Heo do { 499fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 500fa1b54e6STejun Heo return -ENOMEM; 5019daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 502fa1b54e6STejun Heo } while (ret == -EAGAIN); 5039daf9e67STejun Heo 5049daf9e67STejun Heo return ret; 5059daf9e67STejun Heo } 5069daf9e67STejun Heo 50776af4d93STejun Heo /** 50876af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 50976af4d93STejun Heo * @wq: the target workqueue 51076af4d93STejun Heo * 511b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 512794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 513794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 51476af4d93STejun Heo */ 5157fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 516a848e3b6SOleg Nesterov { 517b09f4fd3SLai Jiangshan assert_rcu_or_wq_mutex(wq); 51876af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 51976af4d93STejun Heo pwqs_node); 520f3421797STejun Heo } 521a848e3b6SOleg Nesterov 52273f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 52373f53c4aSTejun Heo { 52473f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 52573f53c4aSTejun Heo } 52673f53c4aSTejun Heo 52773f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 52873f53c4aSTejun Heo { 52973f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 53073f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 53173f53c4aSTejun Heo } 53273f53c4aSTejun Heo 53373f53c4aSTejun Heo static int work_next_color(int color) 53473f53c4aSTejun Heo { 53573f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 536b1f4ec17SOleg Nesterov } 537b1f4ec17SOleg Nesterov 5384594bf15SDavid Howells /* 539112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 540112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5417c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5427a22ad75STejun Heo * 543112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 544112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 545bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 546bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5477a22ad75STejun Heo * 548112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5497c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 550112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5517c3eed5cSTejun Heo * available only while the work item is queued. 552bbb68dfaSTejun Heo * 553bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 554bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 555bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 556bbb68dfaSTejun Heo * try to steal the PENDING bit. 5574594bf15SDavid Howells */ 5587a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5597a22ad75STejun Heo unsigned long flags) 5607a22ad75STejun Heo { 5616183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5627a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5637a22ad75STejun Heo } 5647a22ad75STejun Heo 565112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5664690c4abSTejun Heo unsigned long extra_flags) 567365970a1SDavid Howells { 568112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 569112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 570365970a1SDavid Howells } 571365970a1SDavid Howells 5724468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5734468a00fSLai Jiangshan int pool_id) 5744468a00fSLai Jiangshan { 5754468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5764468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5774468a00fSLai Jiangshan } 5784468a00fSLai Jiangshan 5797c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5807c3eed5cSTejun Heo int pool_id) 5814d707b9fSOleg Nesterov { 58223657bb1STejun Heo /* 58323657bb1STejun Heo * The following wmb is paired with the implied mb in 58423657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 58523657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 58623657bb1STejun Heo * owner. 58723657bb1STejun Heo */ 58823657bb1STejun Heo smp_wmb(); 5897c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5904d707b9fSOleg Nesterov } 5914d707b9fSOleg Nesterov 5927a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 593365970a1SDavid Howells { 5947c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5957c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5967a22ad75STejun Heo } 5977a22ad75STejun Heo 598112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5997a22ad75STejun Heo { 600e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6017a22ad75STejun Heo 602112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 603e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 604e120153dSTejun Heo else 605e120153dSTejun Heo return NULL; 6067a22ad75STejun Heo } 6077a22ad75STejun Heo 6087c3eed5cSTejun Heo /** 6097c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6107c3eed5cSTejun Heo * @work: the work item of interest 6117c3eed5cSTejun Heo * 6127c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 613fa1b54e6STejun Heo * 61468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 61568e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 61668e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 617fa1b54e6STejun Heo * 618fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 619fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 620fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 621fa1b54e6STejun Heo * returned pool is and stays online. 6227c3eed5cSTejun Heo */ 6237c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6247a22ad75STejun Heo { 625e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6267c3eed5cSTejun Heo int pool_id; 6277a22ad75STejun Heo 62868e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 629fa1b54e6STejun Heo 630112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 631112202d9STejun Heo return ((struct pool_workqueue *) 6327c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6337a22ad75STejun Heo 6347c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6357c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6367a22ad75STejun Heo return NULL; 6377a22ad75STejun Heo 638fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6397c3eed5cSTejun Heo } 6407c3eed5cSTejun Heo 6417c3eed5cSTejun Heo /** 6427c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6437c3eed5cSTejun Heo * @work: the work item of interest 6447c3eed5cSTejun Heo * 6457c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6467c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6477c3eed5cSTejun Heo */ 6487c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6497c3eed5cSTejun Heo { 65054d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6517c3eed5cSTejun Heo 652112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 653112202d9STejun Heo return ((struct pool_workqueue *) 65454d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 65554d5b7d0SLai Jiangshan 65654d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6577c3eed5cSTejun Heo } 6587c3eed5cSTejun Heo 659bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 660bbb68dfaSTejun Heo { 6617c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 662bbb68dfaSTejun Heo 6637c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6647c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 665bbb68dfaSTejun Heo } 666bbb68dfaSTejun Heo 667bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 668bbb68dfaSTejun Heo { 669bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 670bbb68dfaSTejun Heo 671112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 672bbb68dfaSTejun Heo } 673bbb68dfaSTejun Heo 674e22bee78STejun Heo /* 6753270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6763270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 677d565ed63STejun Heo * they're being called with pool->lock held. 678e22bee78STejun Heo */ 679e22bee78STejun Heo 68063d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 681649027d7STejun Heo { 682e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 683649027d7STejun Heo } 684649027d7STejun Heo 685e22bee78STejun Heo /* 686e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 687e22bee78STejun Heo * running workers. 688974271c4STejun Heo * 689974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 690706026c2STejun Heo * function will always return %true for unbound pools as long as the 691974271c4STejun Heo * worklist isn't empty. 692e22bee78STejun Heo */ 69363d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 694e22bee78STejun Heo { 69563d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 696e22bee78STejun Heo } 697e22bee78STejun Heo 698e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 69963d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 700e22bee78STejun Heo { 70163d95a91STejun Heo return pool->nr_idle; 702e22bee78STejun Heo } 703e22bee78STejun Heo 704e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 70563d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 706e22bee78STejun Heo { 707e19e397aSTejun Heo return !list_empty(&pool->worklist) && 708e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 709e22bee78STejun Heo } 710e22bee78STejun Heo 711e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 71263d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 713e22bee78STejun Heo { 71463d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 715e22bee78STejun Heo } 716e22bee78STejun Heo 717e22bee78STejun Heo /* Do I need to be the manager? */ 71863d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 719e22bee78STejun Heo { 72063d95a91STejun Heo return need_to_create_worker(pool) || 72111ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 722e22bee78STejun Heo } 723e22bee78STejun Heo 724e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 72563d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 726e22bee78STejun Heo { 72734a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 72863d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 72963d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 730e22bee78STejun Heo 731ea1abd61SLai Jiangshan /* 732ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 733ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 734ea1abd61SLai Jiangshan */ 735ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 736ea1abd61SLai Jiangshan return false; 737ea1abd61SLai Jiangshan 738e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 739e22bee78STejun Heo } 740e22bee78STejun Heo 741e22bee78STejun Heo /* 742e22bee78STejun Heo * Wake up functions. 743e22bee78STejun Heo */ 744e22bee78STejun Heo 7457e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 74663d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7477e11629dSTejun Heo { 74863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7497e11629dSTejun Heo return NULL; 7507e11629dSTejun Heo 75163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7527e11629dSTejun Heo } 7537e11629dSTejun Heo 7547e11629dSTejun Heo /** 7557e11629dSTejun Heo * wake_up_worker - wake up an idle worker 75663d95a91STejun Heo * @pool: worker pool to wake worker from 7577e11629dSTejun Heo * 75863d95a91STejun Heo * Wake up the first idle worker of @pool. 7597e11629dSTejun Heo * 7607e11629dSTejun Heo * CONTEXT: 761d565ed63STejun Heo * spin_lock_irq(pool->lock). 7627e11629dSTejun Heo */ 76363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7647e11629dSTejun Heo { 76563d95a91STejun Heo struct worker *worker = first_worker(pool); 7667e11629dSTejun Heo 7677e11629dSTejun Heo if (likely(worker)) 7687e11629dSTejun Heo wake_up_process(worker->task); 7697e11629dSTejun Heo } 7707e11629dSTejun Heo 7714690c4abSTejun Heo /** 772e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 773e22bee78STejun Heo * @task: task waking up 774e22bee78STejun Heo * @cpu: CPU @task is waking up to 775e22bee78STejun Heo * 776e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 777e22bee78STejun Heo * being awoken. 778e22bee78STejun Heo * 779e22bee78STejun Heo * CONTEXT: 780e22bee78STejun Heo * spin_lock_irq(rq->lock) 781e22bee78STejun Heo */ 782d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 783e22bee78STejun Heo { 784e22bee78STejun Heo struct worker *worker = kthread_data(task); 785e22bee78STejun Heo 78636576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 787ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 788e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 789e22bee78STejun Heo } 79036576000SJoonsoo Kim } 791e22bee78STejun Heo 792e22bee78STejun Heo /** 793e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 794e22bee78STejun Heo * @task: task going to sleep 795e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 796e22bee78STejun Heo * 797e22bee78STejun Heo * This function is called during schedule() when a busy worker is 798e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 799e22bee78STejun Heo * returning pointer to its task. 800e22bee78STejun Heo * 801e22bee78STejun Heo * CONTEXT: 802e22bee78STejun Heo * spin_lock_irq(rq->lock) 803e22bee78STejun Heo * 804e22bee78STejun Heo * RETURNS: 805e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 806e22bee78STejun Heo */ 807d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 808e22bee78STejun Heo { 809e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 810111c225aSTejun Heo struct worker_pool *pool; 811e22bee78STejun Heo 812111c225aSTejun Heo /* 813111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 814111c225aSTejun Heo * workers, also reach here, let's not access anything before 815111c225aSTejun Heo * checking NOT_RUNNING. 816111c225aSTejun Heo */ 8172d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 818e22bee78STejun Heo return NULL; 819e22bee78STejun Heo 820111c225aSTejun Heo pool = worker->pool; 821111c225aSTejun Heo 822e22bee78STejun Heo /* this can only happen on the local cpu */ 8236183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8246183c009STejun Heo return NULL; 825e22bee78STejun Heo 826e22bee78STejun Heo /* 827e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 828e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 829e22bee78STejun Heo * Please read comment there. 830e22bee78STejun Heo * 831628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 832628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 833628c78e7STejun Heo * disabled, which in turn means that none else could be 834d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 835628c78e7STejun Heo * lock is safe. 836e22bee78STejun Heo */ 837e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 838e19e397aSTejun Heo !list_empty(&pool->worklist)) 83963d95a91STejun Heo to_wakeup = first_worker(pool); 840e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 841e22bee78STejun Heo } 842e22bee78STejun Heo 843e22bee78STejun Heo /** 844e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 845cb444766STejun Heo * @worker: self 846d302f017STejun Heo * @flags: flags to set 847d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 848d302f017STejun Heo * 849e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 850e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 851e22bee78STejun Heo * woken up. 852d302f017STejun Heo * 853cb444766STejun Heo * CONTEXT: 854d565ed63STejun Heo * spin_lock_irq(pool->lock) 855d302f017STejun Heo */ 856d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 857d302f017STejun Heo bool wakeup) 858d302f017STejun Heo { 859bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 860e22bee78STejun Heo 861cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 862cb444766STejun Heo 863e22bee78STejun Heo /* 864e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 865e22bee78STejun Heo * wake up an idle worker as necessary if requested by 866e22bee78STejun Heo * @wakeup. 867e22bee78STejun Heo */ 868e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 869e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 870e22bee78STejun Heo if (wakeup) { 871e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 872bd7bdd43STejun Heo !list_empty(&pool->worklist)) 87363d95a91STejun Heo wake_up_worker(pool); 874e22bee78STejun Heo } else 875e19e397aSTejun Heo atomic_dec(&pool->nr_running); 876e22bee78STejun Heo } 877e22bee78STejun Heo 878d302f017STejun Heo worker->flags |= flags; 879d302f017STejun Heo } 880d302f017STejun Heo 881d302f017STejun Heo /** 882e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 883cb444766STejun Heo * @worker: self 884d302f017STejun Heo * @flags: flags to clear 885d302f017STejun Heo * 886e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 887d302f017STejun Heo * 888cb444766STejun Heo * CONTEXT: 889d565ed63STejun Heo * spin_lock_irq(pool->lock) 890d302f017STejun Heo */ 891d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 892d302f017STejun Heo { 89363d95a91STejun Heo struct worker_pool *pool = worker->pool; 894e22bee78STejun Heo unsigned int oflags = worker->flags; 895e22bee78STejun Heo 896cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 897cb444766STejun Heo 898d302f017STejun Heo worker->flags &= ~flags; 899e22bee78STejun Heo 90042c025f3STejun Heo /* 90142c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 90242c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 90342c025f3STejun Heo * of multiple flags, not a single flag. 90442c025f3STejun Heo */ 905e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 906e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 907e19e397aSTejun Heo atomic_inc(&pool->nr_running); 908d302f017STejun Heo } 909d302f017STejun Heo 910d302f017STejun Heo /** 9118cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 912c9e7cf27STejun Heo * @pool: pool of interest 9138cca0eeaSTejun Heo * @work: work to find worker for 9148cca0eeaSTejun Heo * 915c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 916c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 917a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 918a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 919a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 920a2c1c57bSTejun Heo * being executed. 921a2c1c57bSTejun Heo * 922a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 923a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 924a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 925a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 926a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 927a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 928a2c1c57bSTejun Heo * 929c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 930c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 931c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 932c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 933c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 934c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9358cca0eeaSTejun Heo * 9368cca0eeaSTejun Heo * CONTEXT: 937d565ed63STejun Heo * spin_lock_irq(pool->lock). 9388cca0eeaSTejun Heo * 9398cca0eeaSTejun Heo * RETURNS: 9408cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9418cca0eeaSTejun Heo * otherwise. 9428cca0eeaSTejun Heo */ 943c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9448cca0eeaSTejun Heo struct work_struct *work) 9458cca0eeaSTejun Heo { 94642f8570fSSasha Levin struct worker *worker; 94742f8570fSSasha Levin 948b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 949a2c1c57bSTejun Heo (unsigned long)work) 950a2c1c57bSTejun Heo if (worker->current_work == work && 951a2c1c57bSTejun Heo worker->current_func == work->func) 95242f8570fSSasha Levin return worker; 95342f8570fSSasha Levin 95442f8570fSSasha Levin return NULL; 9558cca0eeaSTejun Heo } 9568cca0eeaSTejun Heo 9578cca0eeaSTejun Heo /** 958bf4ede01STejun Heo * move_linked_works - move linked works to a list 959bf4ede01STejun Heo * @work: start of series of works to be scheduled 960bf4ede01STejun Heo * @head: target list to append @work to 961bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 962bf4ede01STejun Heo * 963bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 964bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 965bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 966bf4ede01STejun Heo * 967bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 968bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 969bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 970bf4ede01STejun Heo * 971bf4ede01STejun Heo * CONTEXT: 972d565ed63STejun Heo * spin_lock_irq(pool->lock). 973bf4ede01STejun Heo */ 974bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 975bf4ede01STejun Heo struct work_struct **nextp) 976bf4ede01STejun Heo { 977bf4ede01STejun Heo struct work_struct *n; 978bf4ede01STejun Heo 979bf4ede01STejun Heo /* 980bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 981bf4ede01STejun Heo * use NULL for list head. 982bf4ede01STejun Heo */ 983bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 984bf4ede01STejun Heo list_move_tail(&work->entry, head); 985bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 986bf4ede01STejun Heo break; 987bf4ede01STejun Heo } 988bf4ede01STejun Heo 989bf4ede01STejun Heo /* 990bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 991bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 992bf4ede01STejun Heo * needs to be updated. 993bf4ede01STejun Heo */ 994bf4ede01STejun Heo if (nextp) 995bf4ede01STejun Heo *nextp = n; 996bf4ede01STejun Heo } 997bf4ede01STejun Heo 9988864b4e5STejun Heo /** 9998864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10008864b4e5STejun Heo * @pwq: pool_workqueue to get 10018864b4e5STejun Heo * 10028864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10038864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10048864b4e5STejun Heo */ 10058864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10068864b4e5STejun Heo { 10078864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10088864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10098864b4e5STejun Heo pwq->refcnt++; 10108864b4e5STejun Heo } 10118864b4e5STejun Heo 10128864b4e5STejun Heo /** 10138864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10148864b4e5STejun Heo * @pwq: pool_workqueue to put 10158864b4e5STejun Heo * 10168864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10178864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10188864b4e5STejun Heo */ 10198864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10208864b4e5STejun Heo { 10218864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10228864b4e5STejun Heo if (likely(--pwq->refcnt)) 10238864b4e5STejun Heo return; 10248864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10258864b4e5STejun Heo return; 10268864b4e5STejun Heo /* 10278864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10288864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10298864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10308864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10318864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10328864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10338864b4e5STejun Heo */ 10348864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10358864b4e5STejun Heo } 10368864b4e5STejun Heo 1037112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1038bf4ede01STejun Heo { 1039112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1040bf4ede01STejun Heo 1041bf4ede01STejun Heo trace_workqueue_activate_work(work); 1042112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1043bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1044112202d9STejun Heo pwq->nr_active++; 1045bf4ede01STejun Heo } 1046bf4ede01STejun Heo 1047112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10483aa62497SLai Jiangshan { 1049112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10503aa62497SLai Jiangshan struct work_struct, entry); 10513aa62497SLai Jiangshan 1052112202d9STejun Heo pwq_activate_delayed_work(work); 10533aa62497SLai Jiangshan } 10543aa62497SLai Jiangshan 1055bf4ede01STejun Heo /** 1056112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1057112202d9STejun Heo * @pwq: pwq of interest 1058bf4ede01STejun Heo * @color: color of work which left the queue 1059bf4ede01STejun Heo * 1060bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1061112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1062bf4ede01STejun Heo * 1063bf4ede01STejun Heo * CONTEXT: 1064d565ed63STejun Heo * spin_lock_irq(pool->lock). 1065bf4ede01STejun Heo */ 1066112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1067bf4ede01STejun Heo { 10688864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1069bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10708864b4e5STejun Heo goto out_put; 1071bf4ede01STejun Heo 1072112202d9STejun Heo pwq->nr_in_flight[color]--; 1073bf4ede01STejun Heo 1074112202d9STejun Heo pwq->nr_active--; 1075112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1076bf4ede01STejun Heo /* one down, submit a delayed one */ 1077112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1078112202d9STejun Heo pwq_activate_first_delayed(pwq); 1079bf4ede01STejun Heo } 1080bf4ede01STejun Heo 1081bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1082112202d9STejun Heo if (likely(pwq->flush_color != color)) 10838864b4e5STejun Heo goto out_put; 1084bf4ede01STejun Heo 1085bf4ede01STejun Heo /* are there still in-flight works? */ 1086112202d9STejun Heo if (pwq->nr_in_flight[color]) 10878864b4e5STejun Heo goto out_put; 1088bf4ede01STejun Heo 1089112202d9STejun Heo /* this pwq is done, clear flush_color */ 1090112202d9STejun Heo pwq->flush_color = -1; 1091bf4ede01STejun Heo 1092bf4ede01STejun Heo /* 1093112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1094bf4ede01STejun Heo * will handle the rest. 1095bf4ede01STejun Heo */ 1096112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1097112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 10988864b4e5STejun Heo out_put: 10998864b4e5STejun Heo put_pwq(pwq); 1100bf4ede01STejun Heo } 1101bf4ede01STejun Heo 110236e227d2STejun Heo /** 1103bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 110436e227d2STejun Heo * @work: work item to steal 110536e227d2STejun Heo * @is_dwork: @work is a delayed_work 1106bbb68dfaSTejun Heo * @flags: place to store irq state 110736e227d2STejun Heo * 110836e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 110936e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 111036e227d2STejun Heo * 111136e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 111236e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 111336e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1114bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1115bbb68dfaSTejun Heo * for arbitrarily long 111636e227d2STejun Heo * 1117bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1118e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1119e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1120e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1121bbb68dfaSTejun Heo * 1122bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1123bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1124bbb68dfaSTejun Heo * 1125e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1126bf4ede01STejun Heo */ 1127bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1128bbb68dfaSTejun Heo unsigned long *flags) 1129bf4ede01STejun Heo { 1130d565ed63STejun Heo struct worker_pool *pool; 1131112202d9STejun Heo struct pool_workqueue *pwq; 1132bf4ede01STejun Heo 1133bbb68dfaSTejun Heo local_irq_save(*flags); 1134bbb68dfaSTejun Heo 113536e227d2STejun Heo /* try to steal the timer if it exists */ 113636e227d2STejun Heo if (is_dwork) { 113736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 113836e227d2STejun Heo 1139e0aecdd8STejun Heo /* 1140e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1141e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1142e0aecdd8STejun Heo * running on the local CPU. 1143e0aecdd8STejun Heo */ 114436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 114536e227d2STejun Heo return 1; 114636e227d2STejun Heo } 114736e227d2STejun Heo 114836e227d2STejun Heo /* try to claim PENDING the normal way */ 1149bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1150bf4ede01STejun Heo return 0; 1151bf4ede01STejun Heo 1152bf4ede01STejun Heo /* 1153bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1154bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1155bf4ede01STejun Heo */ 1156d565ed63STejun Heo pool = get_work_pool(work); 1157d565ed63STejun Heo if (!pool) 1158bbb68dfaSTejun Heo goto fail; 1159bf4ede01STejun Heo 1160d565ed63STejun Heo spin_lock(&pool->lock); 1161bf4ede01STejun Heo /* 1162112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1163112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1164112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1165112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1166112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11670b3dae68SLai Jiangshan * item is currently queued on that pool. 1168bf4ede01STejun Heo */ 1169112202d9STejun Heo pwq = get_work_pwq(work); 1170112202d9STejun Heo if (pwq && pwq->pool == pool) { 1171bf4ede01STejun Heo debug_work_deactivate(work); 11723aa62497SLai Jiangshan 11733aa62497SLai Jiangshan /* 117416062836STejun Heo * A delayed work item cannot be grabbed directly because 117516062836STejun Heo * it might have linked NO_COLOR work items which, if left 1176112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 117716062836STejun Heo * management later on and cause stall. Make sure the work 117816062836STejun Heo * item is activated before grabbing. 11793aa62497SLai Jiangshan */ 11803aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1181112202d9STejun Heo pwq_activate_delayed_work(work); 11823aa62497SLai Jiangshan 1183bf4ede01STejun Heo list_del_init(&work->entry); 1184112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 118536e227d2STejun Heo 1186112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11874468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11884468a00fSLai Jiangshan 1189d565ed63STejun Heo spin_unlock(&pool->lock); 119036e227d2STejun Heo return 1; 1191bf4ede01STejun Heo } 1192d565ed63STejun Heo spin_unlock(&pool->lock); 1193bbb68dfaSTejun Heo fail: 1194bbb68dfaSTejun Heo local_irq_restore(*flags); 1195bbb68dfaSTejun Heo if (work_is_canceling(work)) 1196bbb68dfaSTejun Heo return -ENOENT; 1197bbb68dfaSTejun Heo cpu_relax(); 119836e227d2STejun Heo return -EAGAIN; 1199bf4ede01STejun Heo } 1200bf4ede01STejun Heo 1201bf4ede01STejun Heo /** 1202706026c2STejun Heo * insert_work - insert a work into a pool 1203112202d9STejun Heo * @pwq: pwq @work belongs to 12044690c4abSTejun Heo * @work: work to insert 12054690c4abSTejun Heo * @head: insertion point 12064690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12074690c4abSTejun Heo * 1208112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1209706026c2STejun Heo * work_struct flags. 12104690c4abSTejun Heo * 12114690c4abSTejun Heo * CONTEXT: 1212d565ed63STejun Heo * spin_lock_irq(pool->lock). 1213365970a1SDavid Howells */ 1214112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1215112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1216b89deed3SOleg Nesterov { 1217112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1218e1d8aa9fSFrederic Weisbecker 12194690c4abSTejun Heo /* we own @work, set data and link */ 1220112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12211a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12228864b4e5STejun Heo get_pwq(pwq); 1223e22bee78STejun Heo 1224e22bee78STejun Heo /* 1225c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1226c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1227c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1228e22bee78STejun Heo */ 1229e22bee78STejun Heo smp_mb(); 1230e22bee78STejun Heo 123163d95a91STejun Heo if (__need_more_worker(pool)) 123263d95a91STejun Heo wake_up_worker(pool); 1233b89deed3SOleg Nesterov } 1234b89deed3SOleg Nesterov 1235c8efcc25STejun Heo /* 1236c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12378d03ecfeSTejun Heo * same workqueue. 1238c8efcc25STejun Heo */ 1239c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1240c8efcc25STejun Heo { 1241c8efcc25STejun Heo struct worker *worker; 1242c8efcc25STejun Heo 12438d03ecfeSTejun Heo worker = current_wq_worker(); 1244c8efcc25STejun Heo /* 12458d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12468d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1247c8efcc25STejun Heo */ 1248112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1249c8efcc25STejun Heo } 1250c8efcc25STejun Heo 1251d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12521da177e4SLinus Torvalds struct work_struct *work) 12531da177e4SLinus Torvalds { 1254112202d9STejun Heo struct pool_workqueue *pwq; 1255c9178087STejun Heo struct worker_pool *last_pool; 12561e19ffc6STejun Heo struct list_head *worklist; 12578a2e8e5dSTejun Heo unsigned int work_flags; 1258b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12598930cabaSTejun Heo 12608930cabaSTejun Heo /* 12618930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12628930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12638930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12648930cabaSTejun Heo * happen with IRQ disabled. 12658930cabaSTejun Heo */ 12668930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12671da177e4SLinus Torvalds 1268dc186ad7SThomas Gleixner debug_work_activate(work); 12691e19ffc6STejun Heo 1270c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1271618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1272c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1273e41e704bSTejun Heo return; 12749e8cd2f5STejun Heo retry: 1275c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1276c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 127757469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1278f3421797STejun Heo cpu = raw_smp_processor_id(); 1279c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1280c9178087STejun Heo } else { 1281c9178087STejun Heo pwq = first_pwq(wq); 1282c9178087STejun Heo } 1283f3421797STejun Heo 128418aa9effSTejun Heo /* 1285c9178087STejun Heo * If @work was previously on a different pool, it might still be 1286c9178087STejun Heo * running there, in which case the work needs to be queued on that 1287c9178087STejun Heo * pool to guarantee non-reentrancy. 128818aa9effSTejun Heo */ 1289c9e7cf27STejun Heo last_pool = get_work_pool(work); 1290112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 129118aa9effSTejun Heo struct worker *worker; 129218aa9effSTejun Heo 1293d565ed63STejun Heo spin_lock(&last_pool->lock); 129418aa9effSTejun Heo 1295c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 129618aa9effSTejun Heo 1297112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1298c9178087STejun Heo pwq = worker->current_pwq; 12998594fadeSLai Jiangshan } else { 130018aa9effSTejun Heo /* meh... not running there, queue here */ 1301d565ed63STejun Heo spin_unlock(&last_pool->lock); 1302112202d9STejun Heo spin_lock(&pwq->pool->lock); 130318aa9effSTejun Heo } 13048930cabaSTejun Heo } else { 1305112202d9STejun Heo spin_lock(&pwq->pool->lock); 13068930cabaSTejun Heo } 1307502ca9d8STejun Heo 13089e8cd2f5STejun Heo /* 13099e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13109e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13119e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 13129e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 13139e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 13149e8cd2f5STejun Heo * make forward-progress. 13159e8cd2f5STejun Heo */ 13169e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13179e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13189e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13199e8cd2f5STejun Heo cpu_relax(); 13209e8cd2f5STejun Heo goto retry; 13219e8cd2f5STejun Heo } 13229e8cd2f5STejun Heo /* oops */ 13239e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13249e8cd2f5STejun Heo wq->name, cpu); 13259e8cd2f5STejun Heo } 13269e8cd2f5STejun Heo 1327112202d9STejun Heo /* pwq determined, queue */ 1328112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1329502ca9d8STejun Heo 1330f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1331112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1332f5b2552bSDan Carpenter return; 1333f5b2552bSDan Carpenter } 13341e19ffc6STejun Heo 1335112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1336112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13371e19ffc6STejun Heo 1338112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1339cdadf009STejun Heo trace_workqueue_activate_work(work); 1340112202d9STejun Heo pwq->nr_active++; 1341112202d9STejun Heo worklist = &pwq->pool->worklist; 13428a2e8e5dSTejun Heo } else { 13438a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1344112202d9STejun Heo worklist = &pwq->delayed_works; 13458a2e8e5dSTejun Heo } 13461e19ffc6STejun Heo 1347112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13481e19ffc6STejun Heo 1349112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13501da177e4SLinus Torvalds } 13511da177e4SLinus Torvalds 13520fcb78c2SRolf Eike Beer /** 1353c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1354c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1355c1a220e7SZhang Rui * @wq: workqueue to use 1356c1a220e7SZhang Rui * @work: work to queue 1357c1a220e7SZhang Rui * 1358d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1359c1a220e7SZhang Rui * 1360c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1361c1a220e7SZhang Rui * can't go away. 1362c1a220e7SZhang Rui */ 1363d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1364d4283e93STejun Heo struct work_struct *work) 1365c1a220e7SZhang Rui { 1366d4283e93STejun Heo bool ret = false; 13678930cabaSTejun Heo unsigned long flags; 13688930cabaSTejun Heo 13698930cabaSTejun Heo local_irq_save(flags); 1370c1a220e7SZhang Rui 137122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13724690c4abSTejun Heo __queue_work(cpu, wq, work); 1373d4283e93STejun Heo ret = true; 1374c1a220e7SZhang Rui } 13758930cabaSTejun Heo 13768930cabaSTejun Heo local_irq_restore(flags); 1377c1a220e7SZhang Rui return ret; 1378c1a220e7SZhang Rui } 1379c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1380c1a220e7SZhang Rui 1381d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13821da177e4SLinus Torvalds { 138352bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13841da177e4SLinus Torvalds 1385e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 138660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13871da177e4SLinus Torvalds } 13881438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13891da177e4SLinus Torvalds 13907beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 139152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13921da177e4SLinus Torvalds { 13937beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13947beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13951da177e4SLinus Torvalds 13967beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13977beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1398fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1399fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14007beb2edfSTejun Heo 14018852aac2STejun Heo /* 14028852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14038852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14048852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14058852aac2STejun Heo * on that there's no such delay when @delay is 0. 14068852aac2STejun Heo */ 14078852aac2STejun Heo if (!delay) { 14088852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14098852aac2STejun Heo return; 14108852aac2STejun Heo } 14118852aac2STejun Heo 14127beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14137beb2edfSTejun Heo 141460c057bcSLai Jiangshan dwork->wq = wq; 14151265057fSTejun Heo dwork->cpu = cpu; 14167beb2edfSTejun Heo timer->expires = jiffies + delay; 14177beb2edfSTejun Heo 14187beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14197beb2edfSTejun Heo add_timer_on(timer, cpu); 14207beb2edfSTejun Heo else 14217beb2edfSTejun Heo add_timer(timer); 14227beb2edfSTejun Heo } 14231da177e4SLinus Torvalds 14240fcb78c2SRolf Eike Beer /** 14250fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14260fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14270fcb78c2SRolf Eike Beer * @wq: workqueue to use 1428af9997e4SRandy Dunlap * @dwork: work to queue 14290fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14300fcb78c2SRolf Eike Beer * 1431715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1432715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1433715f1300STejun Heo * execution. 14340fcb78c2SRolf Eike Beer */ 1435d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 143652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14377a6bc1cdSVenkatesh Pallipadi { 143852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1439d4283e93STejun Heo bool ret = false; 14408930cabaSTejun Heo unsigned long flags; 14418930cabaSTejun Heo 14428930cabaSTejun Heo /* read the comment in __queue_work() */ 14438930cabaSTejun Heo local_irq_save(flags); 14447a6bc1cdSVenkatesh Pallipadi 144522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14467beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1447d4283e93STejun Heo ret = true; 14487a6bc1cdSVenkatesh Pallipadi } 14498930cabaSTejun Heo 14508930cabaSTejun Heo local_irq_restore(flags); 14517a6bc1cdSVenkatesh Pallipadi return ret; 14527a6bc1cdSVenkatesh Pallipadi } 1453ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14541da177e4SLinus Torvalds 1455c8e55f36STejun Heo /** 14568376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14578376fe22STejun Heo * @cpu: CPU number to execute work on 14588376fe22STejun Heo * @wq: workqueue to use 14598376fe22STejun Heo * @dwork: work to queue 14608376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14618376fe22STejun Heo * 14628376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14638376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14648376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14658376fe22STejun Heo * current state. 14668376fe22STejun Heo * 14678376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14688376fe22STejun Heo * pending and its timer was modified. 14698376fe22STejun Heo * 1470e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14718376fe22STejun Heo * See try_to_grab_pending() for details. 14728376fe22STejun Heo */ 14738376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14748376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14758376fe22STejun Heo { 14768376fe22STejun Heo unsigned long flags; 14778376fe22STejun Heo int ret; 14788376fe22STejun Heo 14798376fe22STejun Heo do { 14808376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14818376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14828376fe22STejun Heo 14838376fe22STejun Heo if (likely(ret >= 0)) { 14848376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14858376fe22STejun Heo local_irq_restore(flags); 14868376fe22STejun Heo } 14878376fe22STejun Heo 14888376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14898376fe22STejun Heo return ret; 14908376fe22STejun Heo } 14918376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14928376fe22STejun Heo 14938376fe22STejun Heo /** 1494c8e55f36STejun Heo * worker_enter_idle - enter idle state 1495c8e55f36STejun Heo * @worker: worker which is entering idle state 1496c8e55f36STejun Heo * 1497c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1498c8e55f36STejun Heo * necessary. 1499c8e55f36STejun Heo * 1500c8e55f36STejun Heo * LOCKING: 1501d565ed63STejun Heo * spin_lock_irq(pool->lock). 1502c8e55f36STejun Heo */ 1503c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15041da177e4SLinus Torvalds { 1505bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1506c8e55f36STejun Heo 15076183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15086183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15096183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15106183c009STejun Heo return; 1511c8e55f36STejun Heo 1512cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1513cb444766STejun Heo worker->flags |= WORKER_IDLE; 1514bd7bdd43STejun Heo pool->nr_idle++; 1515e22bee78STejun Heo worker->last_active = jiffies; 1516c8e55f36STejun Heo 1517c8e55f36STejun Heo /* idle_list is LIFO */ 1518bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1519db7bccf4STejun Heo 152063d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1521628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1522cb444766STejun Heo 1523544ecf31STejun Heo /* 1524706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1525d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1526628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1527628c78e7STejun Heo * unbind is not in progress. 1528544ecf31STejun Heo */ 152924647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1530bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1531e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1532c8e55f36STejun Heo } 1533c8e55f36STejun Heo 1534c8e55f36STejun Heo /** 1535c8e55f36STejun Heo * worker_leave_idle - leave idle state 1536c8e55f36STejun Heo * @worker: worker which is leaving idle state 1537c8e55f36STejun Heo * 1538c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1539c8e55f36STejun Heo * 1540c8e55f36STejun Heo * LOCKING: 1541d565ed63STejun Heo * spin_lock_irq(pool->lock). 1542c8e55f36STejun Heo */ 1543c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1544c8e55f36STejun Heo { 1545bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1546c8e55f36STejun Heo 15476183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15486183c009STejun Heo return; 1549d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1550bd7bdd43STejun Heo pool->nr_idle--; 1551c8e55f36STejun Heo list_del_init(&worker->entry); 1552c8e55f36STejun Heo } 1553c8e55f36STejun Heo 1554e22bee78STejun Heo /** 1555f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1556f36dc67bSLai Jiangshan * @pool: target worker_pool 1557f36dc67bSLai Jiangshan * 1558f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1559e22bee78STejun Heo * 1560e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1561e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1562e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1563e22bee78STejun Heo * guaranteed to execute on the cpu. 1564e22bee78STejun Heo * 1565f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1566e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1567e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1568e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1569706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1570e22bee78STejun Heo * [dis]associated in the meantime. 1571e22bee78STejun Heo * 1572706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 157324647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1574f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1575f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1576f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1577e22bee78STejun Heo * 1578e22bee78STejun Heo * CONTEXT: 1579d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1580e22bee78STejun Heo * held. 1581e22bee78STejun Heo * 1582e22bee78STejun Heo * RETURNS: 1583706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1584e22bee78STejun Heo * bound), %false if offline. 1585e22bee78STejun Heo */ 1586f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1587d565ed63STejun Heo __acquires(&pool->lock) 1588e22bee78STejun Heo { 1589e22bee78STejun Heo while (true) { 1590e22bee78STejun Heo /* 1591e22bee78STejun Heo * The following call may fail, succeed or succeed 1592e22bee78STejun Heo * without actually migrating the task to the cpu if 1593e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 159424647570STejun Heo * against POOL_DISASSOCIATED. 1595e22bee78STejun Heo */ 159624647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 15977a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1598e22bee78STejun Heo 1599d565ed63STejun Heo spin_lock_irq(&pool->lock); 160024647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1601e22bee78STejun Heo return false; 1602f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16037a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1604e22bee78STejun Heo return true; 1605d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1606e22bee78STejun Heo 16075035b20fSTejun Heo /* 16085035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16095035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16105035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16115035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16125035b20fSTejun Heo */ 1613e22bee78STejun Heo cpu_relax(); 16145035b20fSTejun Heo cond_resched(); 1615e22bee78STejun Heo } 1616e22bee78STejun Heo } 1617e22bee78STejun Heo 1618c34056a3STejun Heo static struct worker *alloc_worker(void) 1619c34056a3STejun Heo { 1620c34056a3STejun Heo struct worker *worker; 1621c34056a3STejun Heo 1622c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1623c8e55f36STejun Heo if (worker) { 1624c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1625affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1626e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1627e22bee78STejun Heo worker->flags = WORKER_PREP; 1628c8e55f36STejun Heo } 1629c34056a3STejun Heo return worker; 1630c34056a3STejun Heo } 1631c34056a3STejun Heo 1632c34056a3STejun Heo /** 1633c34056a3STejun Heo * create_worker - create a new workqueue worker 163463d95a91STejun Heo * @pool: pool the new worker will belong to 1635c34056a3STejun Heo * 163663d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1637c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1638c34056a3STejun Heo * destroy_worker(). 1639c34056a3STejun Heo * 1640c34056a3STejun Heo * CONTEXT: 1641c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1642c34056a3STejun Heo * 1643c34056a3STejun Heo * RETURNS: 1644c34056a3STejun Heo * Pointer to the newly created worker. 1645c34056a3STejun Heo */ 1646bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1647c34056a3STejun Heo { 1648c34056a3STejun Heo struct worker *worker = NULL; 1649f3421797STejun Heo int id = -1; 1650e3c916a4STejun Heo char id_buf[16]; 1651c34056a3STejun Heo 1652cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1653cd549687STejun Heo 1654822d8405STejun Heo /* 1655822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1656822d8405STejun Heo * without installing the pointer. 1657822d8405STejun Heo */ 1658822d8405STejun Heo idr_preload(GFP_KERNEL); 1659d565ed63STejun Heo spin_lock_irq(&pool->lock); 1660822d8405STejun Heo 1661822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1662822d8405STejun Heo 1663d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1664822d8405STejun Heo idr_preload_end(); 1665822d8405STejun Heo if (id < 0) 1666c34056a3STejun Heo goto fail; 1667c34056a3STejun Heo 1668c34056a3STejun Heo worker = alloc_worker(); 1669c34056a3STejun Heo if (!worker) 1670c34056a3STejun Heo goto fail; 1671c34056a3STejun Heo 1672bd7bdd43STejun Heo worker->pool = pool; 1673c34056a3STejun Heo worker->id = id; 1674c34056a3STejun Heo 167529c91e99STejun Heo if (pool->cpu >= 0) 1676e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1677e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1678f3421797STejun Heo else 1679e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1680e3c916a4STejun Heo 1681f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1682e3c916a4STejun Heo "kworker/%s", id_buf); 1683c34056a3STejun Heo if (IS_ERR(worker->task)) 1684c34056a3STejun Heo goto fail; 1685c34056a3STejun Heo 1686c5aa87bbSTejun Heo /* 1687c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1688c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1689c5aa87bbSTejun Heo */ 16907a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 16917a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16923270476aSTejun Heo 169314a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 169414a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 16957a4e344cSTejun Heo 16967a4e344cSTejun Heo /* 16977a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 16987a4e344cSTejun Heo * remains stable across this function. See the comments above the 16997a4e344cSTejun Heo * flag definition for details. 17007a4e344cSTejun Heo */ 17017a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1702f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1703c34056a3STejun Heo 1704822d8405STejun Heo /* successful, commit the pointer to idr */ 1705822d8405STejun Heo spin_lock_irq(&pool->lock); 1706822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1707822d8405STejun Heo spin_unlock_irq(&pool->lock); 1708822d8405STejun Heo 1709c34056a3STejun Heo return worker; 1710822d8405STejun Heo 1711c34056a3STejun Heo fail: 1712c34056a3STejun Heo if (id >= 0) { 1713d565ed63STejun Heo spin_lock_irq(&pool->lock); 1714822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1715d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1716c34056a3STejun Heo } 1717c34056a3STejun Heo kfree(worker); 1718c34056a3STejun Heo return NULL; 1719c34056a3STejun Heo } 1720c34056a3STejun Heo 1721c34056a3STejun Heo /** 1722c34056a3STejun Heo * start_worker - start a newly created worker 1723c34056a3STejun Heo * @worker: worker to start 1724c34056a3STejun Heo * 1725706026c2STejun Heo * Make the pool aware of @worker and start it. 1726c34056a3STejun Heo * 1727c34056a3STejun Heo * CONTEXT: 1728d565ed63STejun Heo * spin_lock_irq(pool->lock). 1729c34056a3STejun Heo */ 1730c34056a3STejun Heo static void start_worker(struct worker *worker) 1731c34056a3STejun Heo { 1732cb444766STejun Heo worker->flags |= WORKER_STARTED; 1733bd7bdd43STejun Heo worker->pool->nr_workers++; 1734c8e55f36STejun Heo worker_enter_idle(worker); 1735c34056a3STejun Heo wake_up_process(worker->task); 1736c34056a3STejun Heo } 1737c34056a3STejun Heo 1738c34056a3STejun Heo /** 1739ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1740ebf44d16STejun Heo * @pool: the target pool 1741ebf44d16STejun Heo * 1742cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1743ebf44d16STejun Heo */ 1744ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1745ebf44d16STejun Heo { 1746ebf44d16STejun Heo struct worker *worker; 1747ebf44d16STejun Heo 1748cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1749cd549687STejun Heo 1750ebf44d16STejun Heo worker = create_worker(pool); 1751ebf44d16STejun Heo if (worker) { 1752ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1753ebf44d16STejun Heo start_worker(worker); 1754ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1755ebf44d16STejun Heo } 1756ebf44d16STejun Heo 1757cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1758cd549687STejun Heo 1759ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1760ebf44d16STejun Heo } 1761ebf44d16STejun Heo 1762ebf44d16STejun Heo /** 1763c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1764c34056a3STejun Heo * @worker: worker to be destroyed 1765c34056a3STejun Heo * 1766706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1767c8e55f36STejun Heo * 1768c8e55f36STejun Heo * CONTEXT: 1769d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1770c34056a3STejun Heo */ 1771c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1772c34056a3STejun Heo { 1773bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1774c34056a3STejun Heo 1775cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1776cd549687STejun Heo lockdep_assert_held(&pool->lock); 1777cd549687STejun Heo 1778c34056a3STejun Heo /* sanity check frenzy */ 17796183c009STejun Heo if (WARN_ON(worker->current_work) || 17806183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 17816183c009STejun Heo return; 1782c34056a3STejun Heo 1783c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1784bd7bdd43STejun Heo pool->nr_workers--; 1785c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1786bd7bdd43STejun Heo pool->nr_idle--; 1787c8e55f36STejun Heo 1788c8e55f36STejun Heo list_del_init(&worker->entry); 1789cb444766STejun Heo worker->flags |= WORKER_DIE; 1790c8e55f36STejun Heo 1791822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1792822d8405STejun Heo 1793d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1794c8e55f36STejun Heo 1795c34056a3STejun Heo kthread_stop(worker->task); 1796c34056a3STejun Heo kfree(worker); 1797c34056a3STejun Heo 1798d565ed63STejun Heo spin_lock_irq(&pool->lock); 1799c34056a3STejun Heo } 1800c34056a3STejun Heo 180163d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1802e22bee78STejun Heo { 180363d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1804e22bee78STejun Heo 1805d565ed63STejun Heo spin_lock_irq(&pool->lock); 1806e22bee78STejun Heo 180763d95a91STejun Heo if (too_many_workers(pool)) { 1808e22bee78STejun Heo struct worker *worker; 1809e22bee78STejun Heo unsigned long expires; 1810e22bee78STejun Heo 1811e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 181263d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1813e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1814e22bee78STejun Heo 1815e22bee78STejun Heo if (time_before(jiffies, expires)) 181663d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1817e22bee78STejun Heo else { 1818e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 181911ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 182063d95a91STejun Heo wake_up_worker(pool); 1821e22bee78STejun Heo } 1822e22bee78STejun Heo } 1823e22bee78STejun Heo 1824d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1825e22bee78STejun Heo } 1826e22bee78STejun Heo 1827493a1724STejun Heo static void send_mayday(struct work_struct *work) 1828e22bee78STejun Heo { 1829112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1830112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1831493a1724STejun Heo 18322e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1833e22bee78STejun Heo 1834493008a8STejun Heo if (!wq->rescuer) 1835493a1724STejun Heo return; 1836e22bee78STejun Heo 1837e22bee78STejun Heo /* mayday mayday mayday */ 1838493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1839493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1840e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1841493a1724STejun Heo } 1842e22bee78STejun Heo } 1843e22bee78STejun Heo 1844706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1845e22bee78STejun Heo { 184663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1847e22bee78STejun Heo struct work_struct *work; 1848e22bee78STejun Heo 18492e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1850493a1724STejun Heo spin_lock(&pool->lock); 1851e22bee78STejun Heo 185263d95a91STejun Heo if (need_to_create_worker(pool)) { 1853e22bee78STejun Heo /* 1854e22bee78STejun Heo * We've been trying to create a new worker but 1855e22bee78STejun Heo * haven't been successful. We might be hitting an 1856e22bee78STejun Heo * allocation deadlock. Send distress signals to 1857e22bee78STejun Heo * rescuers. 1858e22bee78STejun Heo */ 185963d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1860e22bee78STejun Heo send_mayday(work); 1861e22bee78STejun Heo } 1862e22bee78STejun Heo 1863493a1724STejun Heo spin_unlock(&pool->lock); 18642e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1865e22bee78STejun Heo 186663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1867e22bee78STejun Heo } 1868e22bee78STejun Heo 1869e22bee78STejun Heo /** 1870e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 187163d95a91STejun Heo * @pool: pool to create a new worker for 1872e22bee78STejun Heo * 187363d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1874e22bee78STejun Heo * have at least one idle worker on return from this function. If 1875e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 187663d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1877e22bee78STejun Heo * possible allocation deadlock. 1878e22bee78STejun Heo * 1879c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1880c5aa87bbSTejun Heo * may_start_working() %true. 1881e22bee78STejun Heo * 1882e22bee78STejun Heo * LOCKING: 1883d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1884e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1885e22bee78STejun Heo * manager. 1886e22bee78STejun Heo * 1887e22bee78STejun Heo * RETURNS: 1888c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1889e22bee78STejun Heo * otherwise. 1890e22bee78STejun Heo */ 189163d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1892d565ed63STejun Heo __releases(&pool->lock) 1893d565ed63STejun Heo __acquires(&pool->lock) 1894e22bee78STejun Heo { 189563d95a91STejun Heo if (!need_to_create_worker(pool)) 1896e22bee78STejun Heo return false; 1897e22bee78STejun Heo restart: 1898d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18999f9c2364STejun Heo 1900e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 190163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1902e22bee78STejun Heo 1903e22bee78STejun Heo while (true) { 1904e22bee78STejun Heo struct worker *worker; 1905e22bee78STejun Heo 1906bc2ae0f5STejun Heo worker = create_worker(pool); 1907e22bee78STejun Heo if (worker) { 190863d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1909d565ed63STejun Heo spin_lock_irq(&pool->lock); 1910e22bee78STejun Heo start_worker(worker); 19116183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19126183c009STejun Heo goto restart; 1913e22bee78STejun Heo return true; 1914e22bee78STejun Heo } 1915e22bee78STejun Heo 191663d95a91STejun Heo if (!need_to_create_worker(pool)) 1917e22bee78STejun Heo break; 1918e22bee78STejun Heo 1919e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1920e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19219f9c2364STejun Heo 192263d95a91STejun Heo if (!need_to_create_worker(pool)) 1923e22bee78STejun Heo break; 1924e22bee78STejun Heo } 1925e22bee78STejun Heo 192663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1927d565ed63STejun Heo spin_lock_irq(&pool->lock); 192863d95a91STejun Heo if (need_to_create_worker(pool)) 1929e22bee78STejun Heo goto restart; 1930e22bee78STejun Heo return true; 1931e22bee78STejun Heo } 1932e22bee78STejun Heo 1933e22bee78STejun Heo /** 1934e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 193563d95a91STejun Heo * @pool: pool to destroy workers for 1936e22bee78STejun Heo * 193763d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1938e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1939e22bee78STejun Heo * 1940e22bee78STejun Heo * LOCKING: 1941d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1942e22bee78STejun Heo * multiple times. Called only from manager. 1943e22bee78STejun Heo * 1944e22bee78STejun Heo * RETURNS: 1945c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1946e22bee78STejun Heo * otherwise. 1947e22bee78STejun Heo */ 194863d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1949e22bee78STejun Heo { 1950e22bee78STejun Heo bool ret = false; 1951e22bee78STejun Heo 195263d95a91STejun Heo while (too_many_workers(pool)) { 1953e22bee78STejun Heo struct worker *worker; 1954e22bee78STejun Heo unsigned long expires; 1955e22bee78STejun Heo 195663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1957e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1958e22bee78STejun Heo 1959e22bee78STejun Heo if (time_before(jiffies, expires)) { 196063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1961e22bee78STejun Heo break; 1962e22bee78STejun Heo } 1963e22bee78STejun Heo 1964e22bee78STejun Heo destroy_worker(worker); 1965e22bee78STejun Heo ret = true; 1966e22bee78STejun Heo } 1967e22bee78STejun Heo 1968e22bee78STejun Heo return ret; 1969e22bee78STejun Heo } 1970e22bee78STejun Heo 1971e22bee78STejun Heo /** 1972e22bee78STejun Heo * manage_workers - manage worker pool 1973e22bee78STejun Heo * @worker: self 1974e22bee78STejun Heo * 1975706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1976e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1977706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1978e22bee78STejun Heo * 1979e22bee78STejun Heo * The caller can safely start processing works on false return. On 1980e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1981e22bee78STejun Heo * and may_start_working() is true. 1982e22bee78STejun Heo * 1983e22bee78STejun Heo * CONTEXT: 1984d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1985e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1986e22bee78STejun Heo * 1987e22bee78STejun Heo * RETURNS: 1988d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1989d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 1990e22bee78STejun Heo */ 1991e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1992e22bee78STejun Heo { 199363d95a91STejun Heo struct worker_pool *pool = worker->pool; 1994e22bee78STejun Heo bool ret = false; 1995e22bee78STejun Heo 1996bc3a1afcSTejun Heo /* 1997bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 1998bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 1999bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2000bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2001bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2002bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2003bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2004bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2005bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2006bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2007bc3a1afcSTejun Heo * 2008bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2009bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2010bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2011bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2012bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2013bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2014bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2015bc3a1afcSTejun Heo * manager_mutex. 2016bc3a1afcSTejun Heo */ 201734a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2018e22bee78STejun Heo return ret; 2019e22bee78STejun Heo 2020ee378aa4SLai Jiangshan /* 2021bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2022bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2023ee378aa4SLai Jiangshan */ 2024bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2025d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2026bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2027ee378aa4SLai Jiangshan ret = true; 2028ee378aa4SLai Jiangshan } 2029ee378aa4SLai Jiangshan 203011ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2031e22bee78STejun Heo 2032e22bee78STejun Heo /* 2033e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2034e22bee78STejun Heo * on return. 2035e22bee78STejun Heo */ 203663d95a91STejun Heo ret |= maybe_destroy_workers(pool); 203763d95a91STejun Heo ret |= maybe_create_worker(pool); 2038e22bee78STejun Heo 2039bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 204034a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2041e22bee78STejun Heo return ret; 2042e22bee78STejun Heo } 2043e22bee78STejun Heo 2044a62428c0STejun Heo /** 2045a62428c0STejun Heo * process_one_work - process single work 2046c34056a3STejun Heo * @worker: self 2047a62428c0STejun Heo * @work: work to process 2048a62428c0STejun Heo * 2049a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2050a62428c0STejun Heo * process a single work including synchronization against and 2051a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2052a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2053a62428c0STejun Heo * call this function to process a work. 2054a62428c0STejun Heo * 2055a62428c0STejun Heo * CONTEXT: 2056d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2057a62428c0STejun Heo */ 2058c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2059d565ed63STejun Heo __releases(&pool->lock) 2060d565ed63STejun Heo __acquires(&pool->lock) 20611da177e4SLinus Torvalds { 2062112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2063bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2064112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 206573f53c4aSTejun Heo int work_color; 20667e11629dSTejun Heo struct worker *collision; 20674e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20684e6045f1SJohannes Berg /* 2069a62428c0STejun Heo * It is permissible to free the struct work_struct from 2070a62428c0STejun Heo * inside the function that is called from it, this we need to 2071a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2072a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2073a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20744e6045f1SJohannes Berg */ 20754d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20764d82a1deSPeter Zijlstra 20774d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20784e6045f1SJohannes Berg #endif 20796fec10a1STejun Heo /* 20806fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20816fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 208224647570STejun Heo * unbound or a disassociated pool. 20836fec10a1STejun Heo */ 20845f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 208524647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2086ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 208725511a47STejun Heo 20887e11629dSTejun Heo /* 20897e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20907e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20917e11629dSTejun Heo * already processing the work. If so, defer the work to the 20927e11629dSTejun Heo * currently executing one. 20937e11629dSTejun Heo */ 2094c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20957e11629dSTejun Heo if (unlikely(collision)) { 20967e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20977e11629dSTejun Heo return; 20987e11629dSTejun Heo } 20991da177e4SLinus Torvalds 21008930cabaSTejun Heo /* claim and dequeue */ 21011da177e4SLinus Torvalds debug_work_deactivate(work); 2102c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2103c34056a3STejun Heo worker->current_work = work; 2104a2c1c57bSTejun Heo worker->current_func = work->func; 2105112202d9STejun Heo worker->current_pwq = pwq; 210673f53c4aSTejun Heo work_color = get_work_color(work); 21077a22ad75STejun Heo 2108a62428c0STejun Heo list_del_init(&work->entry); 2109a62428c0STejun Heo 2110649027d7STejun Heo /* 2111fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2112fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2113fb0e7bebSTejun Heo */ 2114fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2115fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2116fb0e7bebSTejun Heo 2117974271c4STejun Heo /* 2118d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2119974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2120974271c4STejun Heo */ 212163d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 212263d95a91STejun Heo wake_up_worker(pool); 2123974271c4STejun Heo 21248930cabaSTejun Heo /* 21257c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2126d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 212723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 212823657bb1STejun Heo * disabled. 21298930cabaSTejun Heo */ 21307c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21311da177e4SLinus Torvalds 2132d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2133365970a1SDavid Howells 2134112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21353295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2136e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2137a2c1c57bSTejun Heo worker->current_func(work); 2138e36c886aSArjan van de Ven /* 2139e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2140e36c886aSArjan van de Ven * point will only record its address. 2141e36c886aSArjan van de Ven */ 2142e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21433295f0efSIngo Molnar lock_map_release(&lockdep_map); 2144112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21451da177e4SLinus Torvalds 2146d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2147044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2148044c782cSValentin Ilie " last function: %pf\n", 2149a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2150a2c1c57bSTejun Heo worker->current_func); 2151d5abe669SPeter Zijlstra debug_show_held_locks(current); 2152d5abe669SPeter Zijlstra dump_stack(); 2153d5abe669SPeter Zijlstra } 2154d5abe669SPeter Zijlstra 2155d565ed63STejun Heo spin_lock_irq(&pool->lock); 2156a62428c0STejun Heo 2157fb0e7bebSTejun Heo /* clear cpu intensive status */ 2158fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2159fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2160fb0e7bebSTejun Heo 2161a62428c0STejun Heo /* we're done with it, release */ 216242f8570fSSasha Levin hash_del(&worker->hentry); 2163c34056a3STejun Heo worker->current_work = NULL; 2164a2c1c57bSTejun Heo worker->current_func = NULL; 2165112202d9STejun Heo worker->current_pwq = NULL; 2166112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21671da177e4SLinus Torvalds } 21681da177e4SLinus Torvalds 2169affee4b2STejun Heo /** 2170affee4b2STejun Heo * process_scheduled_works - process scheduled works 2171affee4b2STejun Heo * @worker: self 2172affee4b2STejun Heo * 2173affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2174affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2175affee4b2STejun Heo * fetches a work from the top and executes it. 2176affee4b2STejun Heo * 2177affee4b2STejun Heo * CONTEXT: 2178d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2179affee4b2STejun Heo * multiple times. 2180affee4b2STejun Heo */ 2181affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21821da177e4SLinus Torvalds { 2183affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2184affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2185a62428c0STejun Heo struct work_struct, entry); 2186c34056a3STejun Heo process_one_work(worker, work); 2187a62428c0STejun Heo } 21881da177e4SLinus Torvalds } 21891da177e4SLinus Torvalds 21904690c4abSTejun Heo /** 21914690c4abSTejun Heo * worker_thread - the worker thread function 2192c34056a3STejun Heo * @__worker: self 21934690c4abSTejun Heo * 2194c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2195c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2196c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2197c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2198c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 21994690c4abSTejun Heo */ 2200c34056a3STejun Heo static int worker_thread(void *__worker) 22011da177e4SLinus Torvalds { 2202c34056a3STejun Heo struct worker *worker = __worker; 2203bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22041da177e4SLinus Torvalds 2205e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2206e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2207c8e55f36STejun Heo woke_up: 2208d565ed63STejun Heo spin_lock_irq(&pool->lock); 2209affee4b2STejun Heo 2210a9ab775bSTejun Heo /* am I supposed to die? */ 2211a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2212d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2213a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2214e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2215c8e55f36STejun Heo return 0; 2216c8e55f36STejun Heo } 2217c8e55f36STejun Heo 2218c8e55f36STejun Heo worker_leave_idle(worker); 2219db7bccf4STejun Heo recheck: 2220e22bee78STejun Heo /* no more worker necessary? */ 222163d95a91STejun Heo if (!need_more_worker(pool)) 2222e22bee78STejun Heo goto sleep; 2223e22bee78STejun Heo 2224e22bee78STejun Heo /* do we need to manage? */ 222563d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2226e22bee78STejun Heo goto recheck; 2227e22bee78STejun Heo 2228c8e55f36STejun Heo /* 2229c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2230c8e55f36STejun Heo * preparing to process a work or actually processing it. 2231c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2232c8e55f36STejun Heo */ 22336183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2234c8e55f36STejun Heo 2235e22bee78STejun Heo /* 2236a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2237a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2238a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2239a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2240a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2241e22bee78STejun Heo */ 2242a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2243e22bee78STejun Heo 2244e22bee78STejun Heo do { 2245affee4b2STejun Heo struct work_struct *work = 2246bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2247affee4b2STejun Heo struct work_struct, entry); 2248affee4b2STejun Heo 2249c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2250affee4b2STejun Heo /* optimization path, not strictly necessary */ 2251affee4b2STejun Heo process_one_work(worker, work); 2252affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2253affee4b2STejun Heo process_scheduled_works(worker); 2254affee4b2STejun Heo } else { 2255c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2256affee4b2STejun Heo process_scheduled_works(worker); 2257affee4b2STejun Heo } 225863d95a91STejun Heo } while (keep_working(pool)); 2259affee4b2STejun Heo 2260e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2261d313dd85STejun Heo sleep: 226263d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2263e22bee78STejun Heo goto recheck; 2264d313dd85STejun Heo 2265c8e55f36STejun Heo /* 2266d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2267d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2268d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2269d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2270d565ed63STejun Heo * event. 2271c8e55f36STejun Heo */ 2272c8e55f36STejun Heo worker_enter_idle(worker); 2273c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2274d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22751da177e4SLinus Torvalds schedule(); 2276c8e55f36STejun Heo goto woke_up; 22771da177e4SLinus Torvalds } 22781da177e4SLinus Torvalds 2279e22bee78STejun Heo /** 2280e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2281111c225aSTejun Heo * @__rescuer: self 2282e22bee78STejun Heo * 2283e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2284493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2285e22bee78STejun Heo * 2286706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2287e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2288e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2289e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2290e22bee78STejun Heo * the problem rescuer solves. 2291e22bee78STejun Heo * 2292706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2293706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2294e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2295e22bee78STejun Heo * 2296e22bee78STejun Heo * This should happen rarely. 2297e22bee78STejun Heo */ 2298111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2299e22bee78STejun Heo { 2300111c225aSTejun Heo struct worker *rescuer = __rescuer; 2301111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2302e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2303e22bee78STejun Heo 2304e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2305111c225aSTejun Heo 2306111c225aSTejun Heo /* 2307111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2308111c225aSTejun Heo * doesn't participate in concurrency management. 2309111c225aSTejun Heo */ 2310111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2311e22bee78STejun Heo repeat: 2312e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23131da177e4SLinus Torvalds 2314412d32e6SMike Galbraith if (kthread_should_stop()) { 2315412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2316111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2317e22bee78STejun Heo return 0; 2318412d32e6SMike Galbraith } 23191da177e4SLinus Torvalds 2320493a1724STejun Heo /* see whether any pwq is asking for help */ 23212e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2322493a1724STejun Heo 2323493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2324493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2325493a1724STejun Heo struct pool_workqueue, mayday_node); 2326112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2327e22bee78STejun Heo struct work_struct *work, *n; 2328e22bee78STejun Heo 2329e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2330493a1724STejun Heo list_del_init(&pwq->mayday_node); 2331493a1724STejun Heo 23322e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2333e22bee78STejun Heo 2334e22bee78STejun Heo /* migrate to the target cpu if possible */ 2335f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2336b3104104SLai Jiangshan rescuer->pool = pool; 2337e22bee78STejun Heo 2338e22bee78STejun Heo /* 2339e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2340e22bee78STejun Heo * process'em. 2341e22bee78STejun Heo */ 23426183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2343bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2344112202d9STejun Heo if (get_work_pwq(work) == pwq) 2345e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2346e22bee78STejun Heo 2347e22bee78STejun Heo process_scheduled_works(rescuer); 23487576958aSTejun Heo 23497576958aSTejun Heo /* 2350d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23517576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23527576958aSTejun Heo * and stalling the execution. 23537576958aSTejun Heo */ 235463d95a91STejun Heo if (keep_working(pool)) 235563d95a91STejun Heo wake_up_worker(pool); 23567576958aSTejun Heo 2357b3104104SLai Jiangshan rescuer->pool = NULL; 2358493a1724STejun Heo spin_unlock(&pool->lock); 23592e109a28STejun Heo spin_lock(&wq_mayday_lock); 23601da177e4SLinus Torvalds } 23611da177e4SLinus Torvalds 23622e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2363493a1724STejun Heo 2364111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2365111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2366e22bee78STejun Heo schedule(); 2367e22bee78STejun Heo goto repeat; 23681da177e4SLinus Torvalds } 23691da177e4SLinus Torvalds 2370fc2e4d70SOleg Nesterov struct wq_barrier { 2371fc2e4d70SOleg Nesterov struct work_struct work; 2372fc2e4d70SOleg Nesterov struct completion done; 2373fc2e4d70SOleg Nesterov }; 2374fc2e4d70SOleg Nesterov 2375fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2376fc2e4d70SOleg Nesterov { 2377fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2378fc2e4d70SOleg Nesterov complete(&barr->done); 2379fc2e4d70SOleg Nesterov } 2380fc2e4d70SOleg Nesterov 23814690c4abSTejun Heo /** 23824690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2383112202d9STejun Heo * @pwq: pwq to insert barrier into 23844690c4abSTejun Heo * @barr: wq_barrier to insert 2385affee4b2STejun Heo * @target: target work to attach @barr to 2386affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23874690c4abSTejun Heo * 2388affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2389affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2390affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2391affee4b2STejun Heo * cpu. 2392affee4b2STejun Heo * 2393affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2394affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2395affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2396affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2397affee4b2STejun Heo * after a work with LINKED flag set. 2398affee4b2STejun Heo * 2399affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2400112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24014690c4abSTejun Heo * 24024690c4abSTejun Heo * CONTEXT: 2403d565ed63STejun Heo * spin_lock_irq(pool->lock). 24044690c4abSTejun Heo */ 2405112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2406affee4b2STejun Heo struct wq_barrier *barr, 2407affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2408fc2e4d70SOleg Nesterov { 2409affee4b2STejun Heo struct list_head *head; 2410affee4b2STejun Heo unsigned int linked = 0; 2411affee4b2STejun Heo 2412dc186ad7SThomas Gleixner /* 2413d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2414dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2415dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2416dc186ad7SThomas Gleixner * might deadlock. 2417dc186ad7SThomas Gleixner */ 2418ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 241922df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2420fc2e4d70SOleg Nesterov init_completion(&barr->done); 242183c22520SOleg Nesterov 2422affee4b2STejun Heo /* 2423affee4b2STejun Heo * If @target is currently being executed, schedule the 2424affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2425affee4b2STejun Heo */ 2426affee4b2STejun Heo if (worker) 2427affee4b2STejun Heo head = worker->scheduled.next; 2428affee4b2STejun Heo else { 2429affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2430affee4b2STejun Heo 2431affee4b2STejun Heo head = target->entry.next; 2432affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2433affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2434affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2435affee4b2STejun Heo } 2436affee4b2STejun Heo 2437dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2438112202d9STejun Heo insert_work(pwq, &barr->work, head, 2439affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2440fc2e4d70SOleg Nesterov } 2441fc2e4d70SOleg Nesterov 244273f53c4aSTejun Heo /** 2443112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 244473f53c4aSTejun Heo * @wq: workqueue being flushed 244573f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 244673f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 244773f53c4aSTejun Heo * 2448112202d9STejun Heo * Prepare pwqs for workqueue flushing. 244973f53c4aSTejun Heo * 2450112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2451112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2452112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2453112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2454112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 245573f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 245673f53c4aSTejun Heo * 245773f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 245873f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 245973f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 246073f53c4aSTejun Heo * is returned. 246173f53c4aSTejun Heo * 2462112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 246373f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 246473f53c4aSTejun Heo * advanced to @work_color. 246573f53c4aSTejun Heo * 246673f53c4aSTejun Heo * CONTEXT: 24673c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 246873f53c4aSTejun Heo * 246973f53c4aSTejun Heo * RETURNS: 247073f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 247173f53c4aSTejun Heo * otherwise. 247273f53c4aSTejun Heo */ 2473112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 247473f53c4aSTejun Heo int flush_color, int work_color) 24751da177e4SLinus Torvalds { 247673f53c4aSTejun Heo bool wait = false; 247749e3cf44STejun Heo struct pool_workqueue *pwq; 24781da177e4SLinus Torvalds 247973f53c4aSTejun Heo if (flush_color >= 0) { 24806183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2481112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2482dc186ad7SThomas Gleixner } 248314441960SOleg Nesterov 248449e3cf44STejun Heo for_each_pwq(pwq, wq) { 2485112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24861da177e4SLinus Torvalds 2487b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 248873f53c4aSTejun Heo 248973f53c4aSTejun Heo if (flush_color >= 0) { 24906183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 249173f53c4aSTejun Heo 2492112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2493112202d9STejun Heo pwq->flush_color = flush_color; 2494112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 249573f53c4aSTejun Heo wait = true; 24961da177e4SLinus Torvalds } 249773f53c4aSTejun Heo } 249873f53c4aSTejun Heo 249973f53c4aSTejun Heo if (work_color >= 0) { 25006183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2501112202d9STejun Heo pwq->work_color = work_color; 250273f53c4aSTejun Heo } 250373f53c4aSTejun Heo 2504b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25051da177e4SLinus Torvalds } 25061da177e4SLinus Torvalds 2507112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 250873f53c4aSTejun Heo complete(&wq->first_flusher->done); 250973f53c4aSTejun Heo 251073f53c4aSTejun Heo return wait; 251183c22520SOleg Nesterov } 25121da177e4SLinus Torvalds 25130fcb78c2SRolf Eike Beer /** 25141da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25150fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25161da177e4SLinus Torvalds * 2517c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2518c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25191da177e4SLinus Torvalds */ 25207ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25211da177e4SLinus Torvalds { 252273f53c4aSTejun Heo struct wq_flusher this_flusher = { 252373f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 252473f53c4aSTejun Heo .flush_color = -1, 252573f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 252673f53c4aSTejun Heo }; 252773f53c4aSTejun Heo int next_color; 2528b1f4ec17SOleg Nesterov 25293295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25303295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 253173f53c4aSTejun Heo 25323c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 253373f53c4aSTejun Heo 253473f53c4aSTejun Heo /* 253573f53c4aSTejun Heo * Start-to-wait phase 253673f53c4aSTejun Heo */ 253773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 253873f53c4aSTejun Heo 253973f53c4aSTejun Heo if (next_color != wq->flush_color) { 254073f53c4aSTejun Heo /* 254173f53c4aSTejun Heo * Color space is not full. The current work_color 254273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 254373f53c4aSTejun Heo * by one. 254473f53c4aSTejun Heo */ 25456183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 254673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 254773f53c4aSTejun Heo wq->work_color = next_color; 254873f53c4aSTejun Heo 254973f53c4aSTejun Heo if (!wq->first_flusher) { 255073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25516183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 255273f53c4aSTejun Heo 255373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 255473f53c4aSTejun Heo 2555112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 255673f53c4aSTejun Heo wq->work_color)) { 255773f53c4aSTejun Heo /* nothing to flush, done */ 255873f53c4aSTejun Heo wq->flush_color = next_color; 255973f53c4aSTejun Heo wq->first_flusher = NULL; 256073f53c4aSTejun Heo goto out_unlock; 256173f53c4aSTejun Heo } 256273f53c4aSTejun Heo } else { 256373f53c4aSTejun Heo /* wait in queue */ 25646183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 256573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2566112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 256773f53c4aSTejun Heo } 256873f53c4aSTejun Heo } else { 256973f53c4aSTejun Heo /* 257073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 257173f53c4aSTejun Heo * The next flush completion will assign us 257273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 257373f53c4aSTejun Heo */ 257473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 257573f53c4aSTejun Heo } 257673f53c4aSTejun Heo 25773c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 257873f53c4aSTejun Heo 257973f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 258073f53c4aSTejun Heo 258173f53c4aSTejun Heo /* 258273f53c4aSTejun Heo * Wake-up-and-cascade phase 258373f53c4aSTejun Heo * 258473f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 258573f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 258673f53c4aSTejun Heo */ 258773f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 258873f53c4aSTejun Heo return; 258973f53c4aSTejun Heo 25903c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 259173f53c4aSTejun Heo 25924ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25934ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25944ce48b37STejun Heo goto out_unlock; 25954ce48b37STejun Heo 259673f53c4aSTejun Heo wq->first_flusher = NULL; 259773f53c4aSTejun Heo 25986183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25996183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 260073f53c4aSTejun Heo 260173f53c4aSTejun Heo while (true) { 260273f53c4aSTejun Heo struct wq_flusher *next, *tmp; 260373f53c4aSTejun Heo 260473f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 260573f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 260673f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 260773f53c4aSTejun Heo break; 260873f53c4aSTejun Heo list_del_init(&next->list); 260973f53c4aSTejun Heo complete(&next->done); 261073f53c4aSTejun Heo } 261173f53c4aSTejun Heo 26126183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 261373f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 261473f53c4aSTejun Heo 261573f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 261673f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 261773f53c4aSTejun Heo 261873f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 261973f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 262073f53c4aSTejun Heo /* 262173f53c4aSTejun Heo * Assign the same color to all overflowed 262273f53c4aSTejun Heo * flushers, advance work_color and append to 262373f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 262473f53c4aSTejun Heo * phase for these overflowed flushers. 262573f53c4aSTejun Heo */ 262673f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 262773f53c4aSTejun Heo tmp->flush_color = wq->work_color; 262873f53c4aSTejun Heo 262973f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 263073f53c4aSTejun Heo 263173f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 263273f53c4aSTejun Heo &wq->flusher_queue); 2633112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 263473f53c4aSTejun Heo } 263573f53c4aSTejun Heo 263673f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26376183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 263873f53c4aSTejun Heo break; 263973f53c4aSTejun Heo } 264073f53c4aSTejun Heo 264173f53c4aSTejun Heo /* 264273f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2643112202d9STejun Heo * the new first flusher and arm pwqs. 264473f53c4aSTejun Heo */ 26456183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26466183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 264773f53c4aSTejun Heo 264873f53c4aSTejun Heo list_del_init(&next->list); 264973f53c4aSTejun Heo wq->first_flusher = next; 265073f53c4aSTejun Heo 2651112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 265273f53c4aSTejun Heo break; 265373f53c4aSTejun Heo 265473f53c4aSTejun Heo /* 265573f53c4aSTejun Heo * Meh... this color is already done, clear first 265673f53c4aSTejun Heo * flusher and repeat cascading. 265773f53c4aSTejun Heo */ 265873f53c4aSTejun Heo wq->first_flusher = NULL; 265973f53c4aSTejun Heo } 266073f53c4aSTejun Heo 266173f53c4aSTejun Heo out_unlock: 26623c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26631da177e4SLinus Torvalds } 2664ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26651da177e4SLinus Torvalds 26669c5a2ba7STejun Heo /** 26679c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26689c5a2ba7STejun Heo * @wq: workqueue to drain 26699c5a2ba7STejun Heo * 26709c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26719c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26729c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26739c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26749c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26759c5a2ba7STejun Heo * takes too long. 26769c5a2ba7STejun Heo */ 26779c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26789c5a2ba7STejun Heo { 26799c5a2ba7STejun Heo unsigned int flush_cnt = 0; 268049e3cf44STejun Heo struct pool_workqueue *pwq; 26819c5a2ba7STejun Heo 26829c5a2ba7STejun Heo /* 26839c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26849c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2685618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26869c5a2ba7STejun Heo */ 268787fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26889c5a2ba7STejun Heo if (!wq->nr_drainers++) 2689618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 269087fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26919c5a2ba7STejun Heo reflush: 26929c5a2ba7STejun Heo flush_workqueue(wq); 26939c5a2ba7STejun Heo 2694b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 269576af4d93STejun Heo 269649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2697fa2563e4SThomas Tuttle bool drained; 26989c5a2ba7STejun Heo 2699b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2700112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2701b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2702fa2563e4SThomas Tuttle 2703fa2563e4SThomas Tuttle if (drained) 27049c5a2ba7STejun Heo continue; 27059c5a2ba7STejun Heo 27069c5a2ba7STejun Heo if (++flush_cnt == 10 || 27079c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2708c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27099c5a2ba7STejun Heo wq->name, flush_cnt); 271076af4d93STejun Heo 2711b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27129c5a2ba7STejun Heo goto reflush; 27139c5a2ba7STejun Heo } 27149c5a2ba7STejun Heo 27159c5a2ba7STejun Heo if (!--wq->nr_drainers) 2716618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 271787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27189c5a2ba7STejun Heo } 27199c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27209c5a2ba7STejun Heo 2721606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2722baf59022STejun Heo { 2723baf59022STejun Heo struct worker *worker = NULL; 2724c9e7cf27STejun Heo struct worker_pool *pool; 2725112202d9STejun Heo struct pool_workqueue *pwq; 2726baf59022STejun Heo 2727baf59022STejun Heo might_sleep(); 2728baf59022STejun Heo 2729fa1b54e6STejun Heo local_irq_disable(); 2730fa1b54e6STejun Heo pool = get_work_pool(work); 2731fa1b54e6STejun Heo if (!pool) { 2732fa1b54e6STejun Heo local_irq_enable(); 2733fa1b54e6STejun Heo return false; 2734fa1b54e6STejun Heo } 2735fa1b54e6STejun Heo 2736fa1b54e6STejun Heo spin_lock(&pool->lock); 27370b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2738112202d9STejun Heo pwq = get_work_pwq(work); 2739112202d9STejun Heo if (pwq) { 2740112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2741baf59022STejun Heo goto already_gone; 2742606a5020STejun Heo } else { 2743c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2744baf59022STejun Heo if (!worker) 2745baf59022STejun Heo goto already_gone; 2746112202d9STejun Heo pwq = worker->current_pwq; 2747606a5020STejun Heo } 2748baf59022STejun Heo 2749112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2750d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2751baf59022STejun Heo 2752e159489bSTejun Heo /* 2753e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2754e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2755e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2756e159489bSTejun Heo * access. 2757e159489bSTejun Heo */ 2758493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2759112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2760e159489bSTejun Heo else 2761112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2762112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2763e159489bSTejun Heo 2764baf59022STejun Heo return true; 2765baf59022STejun Heo already_gone: 2766d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2767baf59022STejun Heo return false; 2768baf59022STejun Heo } 2769baf59022STejun Heo 2770db700897SOleg Nesterov /** 2771401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2772401a8d04STejun Heo * @work: the work to flush 2773db700897SOleg Nesterov * 2774606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2775606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2776401a8d04STejun Heo * 2777401a8d04STejun Heo * RETURNS: 2778401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2779401a8d04STejun Heo * %false if it was already idle. 2780db700897SOleg Nesterov */ 2781401a8d04STejun Heo bool flush_work(struct work_struct *work) 2782db700897SOleg Nesterov { 2783db700897SOleg Nesterov struct wq_barrier barr; 2784db700897SOleg Nesterov 27850976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27860976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27870976dfc1SStephen Boyd 2788606a5020STejun Heo if (start_flush_work(work, &barr)) { 2789db700897SOleg Nesterov wait_for_completion(&barr.done); 2790dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2791401a8d04STejun Heo return true; 2792606a5020STejun Heo } else { 2793401a8d04STejun Heo return false; 2794db700897SOleg Nesterov } 2795606a5020STejun Heo } 2796db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2797db700897SOleg Nesterov 279836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2799401a8d04STejun Heo { 2800bbb68dfaSTejun Heo unsigned long flags; 28011f1f642eSOleg Nesterov int ret; 28021f1f642eSOleg Nesterov 28031f1f642eSOleg Nesterov do { 2804bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2805bbb68dfaSTejun Heo /* 2806bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2807bbb68dfaSTejun Heo * would be waiting for before retrying. 2808bbb68dfaSTejun Heo */ 2809bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2810606a5020STejun Heo flush_work(work); 28111f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28121f1f642eSOleg Nesterov 2813bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2814bbb68dfaSTejun Heo mark_work_canceling(work); 2815bbb68dfaSTejun Heo local_irq_restore(flags); 2816bbb68dfaSTejun Heo 2817606a5020STejun Heo flush_work(work); 28187a22ad75STejun Heo clear_work_data(work); 28191f1f642eSOleg Nesterov return ret; 28201f1f642eSOleg Nesterov } 28211f1f642eSOleg Nesterov 28226e84d644SOleg Nesterov /** 2823401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2824401a8d04STejun Heo * @work: the work to cancel 28256e84d644SOleg Nesterov * 2826401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2827401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2828401a8d04STejun Heo * another workqueue. On return from this function, @work is 2829401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28301f1f642eSOleg Nesterov * 2831401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2832401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28336e84d644SOleg Nesterov * 2834401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28356e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2836401a8d04STejun Heo * 2837401a8d04STejun Heo * RETURNS: 2838401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28396e84d644SOleg Nesterov */ 2840401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28416e84d644SOleg Nesterov { 284236e227d2STejun Heo return __cancel_work_timer(work, false); 2843b89deed3SOleg Nesterov } 284428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2845b89deed3SOleg Nesterov 28466e84d644SOleg Nesterov /** 2847401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2848401a8d04STejun Heo * @dwork: the delayed work to flush 28496e84d644SOleg Nesterov * 2850401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2851401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2852401a8d04STejun Heo * considers the last queueing instance of @dwork. 28531f1f642eSOleg Nesterov * 2854401a8d04STejun Heo * RETURNS: 2855401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2856401a8d04STejun Heo * %false if it was already idle. 28576e84d644SOleg Nesterov */ 2858401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2859401a8d04STejun Heo { 28608930cabaSTejun Heo local_irq_disable(); 2861401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 286260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28638930cabaSTejun Heo local_irq_enable(); 2864401a8d04STejun Heo return flush_work(&dwork->work); 2865401a8d04STejun Heo } 2866401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2867401a8d04STejun Heo 2868401a8d04STejun Heo /** 286957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 287057b30ae7STejun Heo * @dwork: delayed_work to cancel 287109383498STejun Heo * 287257b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 287357b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 287457b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 287557b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 287657b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 287709383498STejun Heo * 287857b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 287909383498STejun Heo */ 288057b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 288109383498STejun Heo { 288257b30ae7STejun Heo unsigned long flags; 288357b30ae7STejun Heo int ret; 288457b30ae7STejun Heo 288557b30ae7STejun Heo do { 288657b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 288757b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 288857b30ae7STejun Heo 288957b30ae7STejun Heo if (unlikely(ret < 0)) 289057b30ae7STejun Heo return false; 289157b30ae7STejun Heo 28927c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28937c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 289457b30ae7STejun Heo local_irq_restore(flags); 2895c0158ca6SDan Magenheimer return ret; 289609383498STejun Heo } 289757b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 289809383498STejun Heo 289909383498STejun Heo /** 2900401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2901401a8d04STejun Heo * @dwork: the delayed work cancel 2902401a8d04STejun Heo * 2903401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2904401a8d04STejun Heo * 2905401a8d04STejun Heo * RETURNS: 2906401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2907401a8d04STejun Heo */ 2908401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29096e84d644SOleg Nesterov { 291036e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29116e84d644SOleg Nesterov } 2912f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29131da177e4SLinus Torvalds 29140fcb78c2SRolf Eike Beer /** 291531ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2916b6136773SAndrew Morton * @func: the function to call 2917b6136773SAndrew Morton * 291831ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 291931ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2920b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 292131ddd871STejun Heo * 292231ddd871STejun Heo * RETURNS: 292331ddd871STejun Heo * 0 on success, -errno on failure. 2924b6136773SAndrew Morton */ 292565f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 292615316ba8SChristoph Lameter { 292715316ba8SChristoph Lameter int cpu; 292838f51568SNamhyung Kim struct work_struct __percpu *works; 292915316ba8SChristoph Lameter 2930b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2931b6136773SAndrew Morton if (!works) 293215316ba8SChristoph Lameter return -ENOMEM; 2933b6136773SAndrew Morton 293495402b38SGautham R Shenoy get_online_cpus(); 293593981800STejun Heo 293615316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29379bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29389bfb1839SIngo Molnar 29399bfb1839SIngo Molnar INIT_WORK(work, func); 29408de6d308SOleg Nesterov schedule_work_on(cpu, work); 294115316ba8SChristoph Lameter } 294293981800STejun Heo 294393981800STejun Heo for_each_online_cpu(cpu) 29448616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 294593981800STejun Heo 294695402b38SGautham R Shenoy put_online_cpus(); 2947b6136773SAndrew Morton free_percpu(works); 294815316ba8SChristoph Lameter return 0; 294915316ba8SChristoph Lameter } 295015316ba8SChristoph Lameter 2951eef6a7d5SAlan Stern /** 2952eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2953eef6a7d5SAlan Stern * 2954eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2955eef6a7d5SAlan Stern * completion. 2956eef6a7d5SAlan Stern * 2957eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2958eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2959eef6a7d5SAlan Stern * will lead to deadlock: 2960eef6a7d5SAlan Stern * 2961eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2962eef6a7d5SAlan Stern * a lock held by your code or its caller. 2963eef6a7d5SAlan Stern * 2964eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2965eef6a7d5SAlan Stern * 2966eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2967eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2968eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2969eef6a7d5SAlan Stern * 2970eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2971eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2972eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2973eef6a7d5SAlan Stern * cancel_work_sync() instead. 2974eef6a7d5SAlan Stern */ 29751da177e4SLinus Torvalds void flush_scheduled_work(void) 29761da177e4SLinus Torvalds { 2977d320c038STejun Heo flush_workqueue(system_wq); 29781da177e4SLinus Torvalds } 2979ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29801da177e4SLinus Torvalds 29811da177e4SLinus Torvalds /** 29821fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29831fa44ecaSJames Bottomley * @fn: the function to execute 29841fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29851fa44ecaSJames Bottomley * be available when the work executes) 29861fa44ecaSJames Bottomley * 29871fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29881fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29891fa44ecaSJames Bottomley * 29901fa44ecaSJames Bottomley * Returns: 0 - function was executed 29911fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29921fa44ecaSJames Bottomley */ 299365f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29941fa44ecaSJames Bottomley { 29951fa44ecaSJames Bottomley if (!in_interrupt()) { 299665f27f38SDavid Howells fn(&ew->work); 29971fa44ecaSJames Bottomley return 0; 29981fa44ecaSJames Bottomley } 29991fa44ecaSJames Bottomley 300065f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30011fa44ecaSJames Bottomley schedule_work(&ew->work); 30021fa44ecaSJames Bottomley 30031fa44ecaSJames Bottomley return 1; 30041fa44ecaSJames Bottomley } 30051fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30061fa44ecaSJames Bottomley 3007226223abSTejun Heo #ifdef CONFIG_SYSFS 3008226223abSTejun Heo /* 3009226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3010226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3011226223abSTejun Heo * following attributes. 3012226223abSTejun Heo * 3013226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3014226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3015226223abSTejun Heo * 3016226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3017226223abSTejun Heo * 3018226223abSTejun Heo * id RO int : the associated pool ID 3019226223abSTejun Heo * nice RW int : nice value of the workers 3020226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3021226223abSTejun Heo */ 3022226223abSTejun Heo struct wq_device { 3023226223abSTejun Heo struct workqueue_struct *wq; 3024226223abSTejun Heo struct device dev; 3025226223abSTejun Heo }; 3026226223abSTejun Heo 3027226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3028226223abSTejun Heo { 3029226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3030226223abSTejun Heo 3031226223abSTejun Heo return wq_dev->wq; 3032226223abSTejun Heo } 3033226223abSTejun Heo 3034226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3035226223abSTejun Heo struct device_attribute *attr, char *buf) 3036226223abSTejun Heo { 3037226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3038226223abSTejun Heo 3039226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3040226223abSTejun Heo } 3041226223abSTejun Heo 3042226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3043226223abSTejun Heo struct device_attribute *attr, char *buf) 3044226223abSTejun Heo { 3045226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3046226223abSTejun Heo 3047226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3048226223abSTejun Heo } 3049226223abSTejun Heo 3050226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3051226223abSTejun Heo struct device_attribute *attr, 3052226223abSTejun Heo const char *buf, size_t count) 3053226223abSTejun Heo { 3054226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3055226223abSTejun Heo int val; 3056226223abSTejun Heo 3057226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3058226223abSTejun Heo return -EINVAL; 3059226223abSTejun Heo 3060226223abSTejun Heo workqueue_set_max_active(wq, val); 3061226223abSTejun Heo return count; 3062226223abSTejun Heo } 3063226223abSTejun Heo 3064226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3065226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3066226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3067226223abSTejun Heo __ATTR_NULL, 3068226223abSTejun Heo }; 3069226223abSTejun Heo 3070226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3071226223abSTejun Heo struct device_attribute *attr, char *buf) 3072226223abSTejun Heo { 3073226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3074226223abSTejun Heo struct worker_pool *pool; 3075226223abSTejun Heo int written; 3076226223abSTejun Heo 3077226223abSTejun Heo rcu_read_lock_sched(); 3078226223abSTejun Heo pool = first_pwq(wq)->pool; 3079226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3080226223abSTejun Heo rcu_read_unlock_sched(); 3081226223abSTejun Heo 3082226223abSTejun Heo return written; 3083226223abSTejun Heo } 3084226223abSTejun Heo 3085226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3086226223abSTejun Heo char *buf) 3087226223abSTejun Heo { 3088226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3089226223abSTejun Heo int written; 3090226223abSTejun Heo 3091226223abSTejun Heo rcu_read_lock_sched(); 3092226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3093226223abSTejun Heo first_pwq(wq)->pool->attrs->nice); 3094226223abSTejun Heo rcu_read_unlock_sched(); 3095226223abSTejun Heo 3096226223abSTejun Heo return written; 3097226223abSTejun Heo } 3098226223abSTejun Heo 3099226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3100226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3101226223abSTejun Heo { 3102226223abSTejun Heo struct workqueue_attrs *attrs; 3103226223abSTejun Heo 3104226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3105226223abSTejun Heo if (!attrs) 3106226223abSTejun Heo return NULL; 3107226223abSTejun Heo 3108226223abSTejun Heo rcu_read_lock_sched(); 3109226223abSTejun Heo copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs); 3110226223abSTejun Heo rcu_read_unlock_sched(); 3111226223abSTejun Heo return attrs; 3112226223abSTejun Heo } 3113226223abSTejun Heo 3114226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3115226223abSTejun Heo const char *buf, size_t count) 3116226223abSTejun Heo { 3117226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3118226223abSTejun Heo struct workqueue_attrs *attrs; 3119226223abSTejun Heo int ret; 3120226223abSTejun Heo 3121226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3122226223abSTejun Heo if (!attrs) 3123226223abSTejun Heo return -ENOMEM; 3124226223abSTejun Heo 3125226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3126226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3127226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3128226223abSTejun Heo else 3129226223abSTejun Heo ret = -EINVAL; 3130226223abSTejun Heo 3131226223abSTejun Heo free_workqueue_attrs(attrs); 3132226223abSTejun Heo return ret ?: count; 3133226223abSTejun Heo } 3134226223abSTejun Heo 3135226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3136226223abSTejun Heo struct device_attribute *attr, char *buf) 3137226223abSTejun Heo { 3138226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3139226223abSTejun Heo int written; 3140226223abSTejun Heo 3141226223abSTejun Heo rcu_read_lock_sched(); 3142226223abSTejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, 3143226223abSTejun Heo first_pwq(wq)->pool->attrs->cpumask); 3144226223abSTejun Heo rcu_read_unlock_sched(); 3145226223abSTejun Heo 3146226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3147226223abSTejun Heo return written; 3148226223abSTejun Heo } 3149226223abSTejun Heo 3150226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3151226223abSTejun Heo struct device_attribute *attr, 3152226223abSTejun Heo const char *buf, size_t count) 3153226223abSTejun Heo { 3154226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3155226223abSTejun Heo struct workqueue_attrs *attrs; 3156226223abSTejun Heo int ret; 3157226223abSTejun Heo 3158226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3159226223abSTejun Heo if (!attrs) 3160226223abSTejun Heo return -ENOMEM; 3161226223abSTejun Heo 3162226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3163226223abSTejun Heo if (!ret) 3164226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3165226223abSTejun Heo 3166226223abSTejun Heo free_workqueue_attrs(attrs); 3167226223abSTejun Heo return ret ?: count; 3168226223abSTejun Heo } 3169226223abSTejun Heo 3170226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3171226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3172226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3173226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3174226223abSTejun Heo __ATTR_NULL, 3175226223abSTejun Heo }; 3176226223abSTejun Heo 3177226223abSTejun Heo static struct bus_type wq_subsys = { 3178226223abSTejun Heo .name = "workqueue", 3179226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3180226223abSTejun Heo }; 3181226223abSTejun Heo 3182226223abSTejun Heo static int __init wq_sysfs_init(void) 3183226223abSTejun Heo { 3184226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3185226223abSTejun Heo } 3186226223abSTejun Heo core_initcall(wq_sysfs_init); 3187226223abSTejun Heo 3188226223abSTejun Heo static void wq_device_release(struct device *dev) 3189226223abSTejun Heo { 3190226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3191226223abSTejun Heo 3192226223abSTejun Heo kfree(wq_dev); 3193226223abSTejun Heo } 3194226223abSTejun Heo 3195226223abSTejun Heo /** 3196226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3197226223abSTejun Heo * @wq: the workqueue to register 3198226223abSTejun Heo * 3199226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3200226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3201226223abSTejun Heo * which is the preferred method. 3202226223abSTejun Heo * 3203226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3204226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3205226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3206226223abSTejun Heo * attributes. 3207226223abSTejun Heo * 3208226223abSTejun Heo * Returns 0 on success, -errno on failure. 3209226223abSTejun Heo */ 3210226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3211226223abSTejun Heo { 3212226223abSTejun Heo struct wq_device *wq_dev; 3213226223abSTejun Heo int ret; 3214226223abSTejun Heo 3215226223abSTejun Heo /* 3216226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3217226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3218226223abSTejun Heo * workqueues. 3219226223abSTejun Heo */ 3220226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3221226223abSTejun Heo return -EINVAL; 3222226223abSTejun Heo 3223226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3224226223abSTejun Heo if (!wq_dev) 3225226223abSTejun Heo return -ENOMEM; 3226226223abSTejun Heo 3227226223abSTejun Heo wq_dev->wq = wq; 3228226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3229226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3230226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3231226223abSTejun Heo 3232226223abSTejun Heo /* 3233226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3234226223abSTejun Heo * everything is ready. 3235226223abSTejun Heo */ 3236226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3237226223abSTejun Heo 3238226223abSTejun Heo ret = device_register(&wq_dev->dev); 3239226223abSTejun Heo if (ret) { 3240226223abSTejun Heo kfree(wq_dev); 3241226223abSTejun Heo wq->wq_dev = NULL; 3242226223abSTejun Heo return ret; 3243226223abSTejun Heo } 3244226223abSTejun Heo 3245226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3246226223abSTejun Heo struct device_attribute *attr; 3247226223abSTejun Heo 3248226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3249226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3250226223abSTejun Heo if (ret) { 3251226223abSTejun Heo device_unregister(&wq_dev->dev); 3252226223abSTejun Heo wq->wq_dev = NULL; 3253226223abSTejun Heo return ret; 3254226223abSTejun Heo } 3255226223abSTejun Heo } 3256226223abSTejun Heo } 3257226223abSTejun Heo 3258226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3259226223abSTejun Heo return 0; 3260226223abSTejun Heo } 3261226223abSTejun Heo 3262226223abSTejun Heo /** 3263226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3264226223abSTejun Heo * @wq: the workqueue to unregister 3265226223abSTejun Heo * 3266226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3267226223abSTejun Heo */ 3268226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3269226223abSTejun Heo { 3270226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3271226223abSTejun Heo 3272226223abSTejun Heo if (!wq->wq_dev) 3273226223abSTejun Heo return; 3274226223abSTejun Heo 3275226223abSTejun Heo wq->wq_dev = NULL; 3276226223abSTejun Heo device_unregister(&wq_dev->dev); 3277226223abSTejun Heo } 3278226223abSTejun Heo #else /* CONFIG_SYSFS */ 3279226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3280226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3281226223abSTejun Heo 32827a4e344cSTejun Heo /** 32837a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32847a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32857a4e344cSTejun Heo * 32867a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32877a4e344cSTejun Heo */ 32887a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32897a4e344cSTejun Heo { 32907a4e344cSTejun Heo if (attrs) { 32917a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32927a4e344cSTejun Heo kfree(attrs); 32937a4e344cSTejun Heo } 32947a4e344cSTejun Heo } 32957a4e344cSTejun Heo 32967a4e344cSTejun Heo /** 32977a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 32987a4e344cSTejun Heo * @gfp_mask: allocation mask to use 32997a4e344cSTejun Heo * 33007a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33017a4e344cSTejun Heo * return it. Returns NULL on failure. 33027a4e344cSTejun Heo */ 33037a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33047a4e344cSTejun Heo { 33057a4e344cSTejun Heo struct workqueue_attrs *attrs; 33067a4e344cSTejun Heo 33077a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33087a4e344cSTejun Heo if (!attrs) 33097a4e344cSTejun Heo goto fail; 33107a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33117a4e344cSTejun Heo goto fail; 33127a4e344cSTejun Heo 331313e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33147a4e344cSTejun Heo return attrs; 33157a4e344cSTejun Heo fail: 33167a4e344cSTejun Heo free_workqueue_attrs(attrs); 33177a4e344cSTejun Heo return NULL; 33187a4e344cSTejun Heo } 33197a4e344cSTejun Heo 332029c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 332129c91e99STejun Heo const struct workqueue_attrs *from) 332229c91e99STejun Heo { 332329c91e99STejun Heo to->nice = from->nice; 332429c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 332529c91e99STejun Heo } 332629c91e99STejun Heo 332729c91e99STejun Heo /* hash value of the content of @attr */ 332829c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 332929c91e99STejun Heo { 333029c91e99STejun Heo u32 hash = 0; 333129c91e99STejun Heo 333229c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 333313e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 333413e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 333529c91e99STejun Heo return hash; 333629c91e99STejun Heo } 333729c91e99STejun Heo 333829c91e99STejun Heo /* content equality test */ 333929c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 334029c91e99STejun Heo const struct workqueue_attrs *b) 334129c91e99STejun Heo { 334229c91e99STejun Heo if (a->nice != b->nice) 334329c91e99STejun Heo return false; 334429c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 334529c91e99STejun Heo return false; 334629c91e99STejun Heo return true; 334729c91e99STejun Heo } 334829c91e99STejun Heo 33497a4e344cSTejun Heo /** 33507a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33517a4e344cSTejun Heo * @pool: worker_pool to initialize 33527a4e344cSTejun Heo * 33537a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 335429c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 335529c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 335629c91e99STejun Heo * on @pool safely to release it. 33577a4e344cSTejun Heo */ 33587a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33594e1a1f9aSTejun Heo { 33604e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 336129c91e99STejun Heo pool->id = -1; 336229c91e99STejun Heo pool->cpu = -1; 3363f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 33644e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33654e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33664e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33674e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33684e1a1f9aSTejun Heo 33694e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33704e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33714e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33724e1a1f9aSTejun Heo 33734e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33744e1a1f9aSTejun Heo (unsigned long)pool); 33754e1a1f9aSTejun Heo 33764e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3377bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3378822d8405STejun Heo idr_init(&pool->worker_idr); 33797a4e344cSTejun Heo 338029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 338129c91e99STejun Heo pool->refcnt = 1; 338229c91e99STejun Heo 338329c91e99STejun Heo /* shouldn't fail above this point */ 33847a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 33857a4e344cSTejun Heo if (!pool->attrs) 33867a4e344cSTejun Heo return -ENOMEM; 33877a4e344cSTejun Heo return 0; 33884e1a1f9aSTejun Heo } 33894e1a1f9aSTejun Heo 339029c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 339129c91e99STejun Heo { 339229c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 339329c91e99STejun Heo 3394822d8405STejun Heo idr_destroy(&pool->worker_idr); 339529c91e99STejun Heo free_workqueue_attrs(pool->attrs); 339629c91e99STejun Heo kfree(pool); 339729c91e99STejun Heo } 339829c91e99STejun Heo 339929c91e99STejun Heo /** 340029c91e99STejun Heo * put_unbound_pool - put a worker_pool 340129c91e99STejun Heo * @pool: worker_pool to put 340229c91e99STejun Heo * 340329c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3404c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3405c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3406c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3407a892caccSTejun Heo * 3408a892caccSTejun Heo * Should be called with wq_pool_mutex held. 340929c91e99STejun Heo */ 341029c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 341129c91e99STejun Heo { 341229c91e99STejun Heo struct worker *worker; 341329c91e99STejun Heo 3414a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3415a892caccSTejun Heo 3416a892caccSTejun Heo if (--pool->refcnt) 341729c91e99STejun Heo return; 341829c91e99STejun Heo 341929c91e99STejun Heo /* sanity checks */ 342029c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3421a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 342229c91e99STejun Heo return; 342329c91e99STejun Heo 342429c91e99STejun Heo /* release id and unhash */ 342529c91e99STejun Heo if (pool->id >= 0) 342629c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 342729c91e99STejun Heo hash_del(&pool->hash_node); 342829c91e99STejun Heo 3429c5aa87bbSTejun Heo /* 3430c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3431c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3432c5aa87bbSTejun Heo * manager_mutex. 3433c5aa87bbSTejun Heo */ 343429c91e99STejun Heo mutex_lock(&pool->manager_arb); 3435cd549687STejun Heo mutex_lock(&pool->manager_mutex); 343629c91e99STejun Heo spin_lock_irq(&pool->lock); 343729c91e99STejun Heo 343829c91e99STejun Heo while ((worker = first_worker(pool))) 343929c91e99STejun Heo destroy_worker(worker); 344029c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 344129c91e99STejun Heo 344229c91e99STejun Heo spin_unlock_irq(&pool->lock); 3443cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 344429c91e99STejun Heo mutex_unlock(&pool->manager_arb); 344529c91e99STejun Heo 344629c91e99STejun Heo /* shut down the timers */ 344729c91e99STejun Heo del_timer_sync(&pool->idle_timer); 344829c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 344929c91e99STejun Heo 345029c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 345129c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 345229c91e99STejun Heo } 345329c91e99STejun Heo 345429c91e99STejun Heo /** 345529c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 345629c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 345729c91e99STejun Heo * 345829c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 345929c91e99STejun Heo * reference count and return it. If there already is a matching 346029c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 346129c91e99STejun Heo * create a new one. On failure, returns NULL. 3462a892caccSTejun Heo * 3463a892caccSTejun Heo * Should be called with wq_pool_mutex held. 346429c91e99STejun Heo */ 346529c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 346629c91e99STejun Heo { 346729c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 346829c91e99STejun Heo struct worker_pool *pool; 3469f3f90ad4STejun Heo int node; 347029c91e99STejun Heo 3471a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 347229c91e99STejun Heo 347329c91e99STejun Heo /* do we already have a matching pool? */ 347429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 347529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 347629c91e99STejun Heo pool->refcnt++; 347729c91e99STejun Heo goto out_unlock; 347829c91e99STejun Heo } 347929c91e99STejun Heo } 348029c91e99STejun Heo 348129c91e99STejun Heo /* nope, create a new one */ 348229c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 348329c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 348429c91e99STejun Heo goto fail; 348529c91e99STejun Heo 348612ee4fc6SLai Jiangshan if (workqueue_freezing) 348712ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 348812ee4fc6SLai Jiangshan 34898864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 349029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 349129c91e99STejun Heo 3492f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3493f3f90ad4STejun Heo if (wq_numa_enabled) { 3494f3f90ad4STejun Heo for_each_node(node) { 3495f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3496f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3497f3f90ad4STejun Heo pool->node = node; 3498f3f90ad4STejun Heo break; 3499f3f90ad4STejun Heo } 3500f3f90ad4STejun Heo } 3501f3f90ad4STejun Heo } 3502f3f90ad4STejun Heo 350329c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 350429c91e99STejun Heo goto fail; 350529c91e99STejun Heo 350629c91e99STejun Heo /* create and start the initial worker */ 3507ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 350829c91e99STejun Heo goto fail; 350929c91e99STejun Heo 351029c91e99STejun Heo /* install */ 351129c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 351229c91e99STejun Heo out_unlock: 351329c91e99STejun Heo return pool; 351429c91e99STejun Heo fail: 351529c91e99STejun Heo if (pool) 351629c91e99STejun Heo put_unbound_pool(pool); 351729c91e99STejun Heo return NULL; 351829c91e99STejun Heo } 351929c91e99STejun Heo 35208864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35218864b4e5STejun Heo { 35228864b4e5STejun Heo kmem_cache_free(pwq_cache, 35238864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35248864b4e5STejun Heo } 35258864b4e5STejun Heo 35268864b4e5STejun Heo /* 35278864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35288864b4e5STejun Heo * and needs to be destroyed. 35298864b4e5STejun Heo */ 35308864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35318864b4e5STejun Heo { 35328864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35338864b4e5STejun Heo unbound_release_work); 35348864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35358864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3536bc0caf09STejun Heo bool is_last; 35378864b4e5STejun Heo 35388864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35398864b4e5STejun Heo return; 35408864b4e5STejun Heo 354175ccf595STejun Heo /* 35423c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 354375ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 354475ccf595STejun Heo * and consistent with the linking path. 354575ccf595STejun Heo */ 35463c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35478864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3548bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35493c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35508864b4e5STejun Heo 3551a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35528864b4e5STejun Heo put_unbound_pool(pool); 3553a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3554a892caccSTejun Heo 35558864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35568864b4e5STejun Heo 35578864b4e5STejun Heo /* 35588864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35598864b4e5STejun Heo * is gonna access it anymore. Free it. 35608864b4e5STejun Heo */ 3561bc0caf09STejun Heo if (is_last) 35628864b4e5STejun Heo kfree(wq); 35638864b4e5STejun Heo } 35648864b4e5STejun Heo 35650fbd95aaSTejun Heo /** 3566699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35670fbd95aaSTejun Heo * @pwq: target pool_workqueue 35680fbd95aaSTejun Heo * 3569699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3570699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3571699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35720fbd95aaSTejun Heo */ 3573699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35740fbd95aaSTejun Heo { 3575699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3576699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3577699ce097STejun Heo 3578699ce097STejun Heo /* for @wq->saved_max_active */ 3579a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3580699ce097STejun Heo 3581699ce097STejun Heo /* fast exit for non-freezable wqs */ 3582699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3583699ce097STejun Heo return; 3584699ce097STejun Heo 3585a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3586699ce097STejun Heo 3587699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3588699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35890fbd95aaSTejun Heo 35900fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35910fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35920fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3593951a078aSLai Jiangshan 3594951a078aSLai Jiangshan /* 3595951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3596951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3597951a078aSLai Jiangshan */ 3598951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3599699ce097STejun Heo } else { 3600699ce097STejun Heo pwq->max_active = 0; 3601699ce097STejun Heo } 3602699ce097STejun Heo 3603a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36040fbd95aaSTejun Heo } 36050fbd95aaSTejun Heo 3606d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3607d2c1d404STejun Heo struct workqueue_struct *wq, 36089e8cd2f5STejun Heo struct worker_pool *pool, 36099e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3610d2c1d404STejun Heo { 3611d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3612d2c1d404STejun Heo 3613d2c1d404STejun Heo pwq->pool = pool; 3614d2c1d404STejun Heo pwq->wq = wq; 3615d2c1d404STejun Heo pwq->flush_color = -1; 36168864b4e5STejun Heo pwq->refcnt = 1; 3617d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3618d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36198864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3620d2c1d404STejun Heo 36213c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 362275ccf595STejun Heo 3623983ca25eSTejun Heo /* 3624983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36253c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3626983ca25eSTejun Heo */ 36279e8cd2f5STejun Heo if (p_last_pwq) 36289e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 362975ccf595STejun Heo pwq->work_color = wq->work_color; 3630983ca25eSTejun Heo 3631983ca25eSTejun Heo /* sync max_active to the current setting */ 3632983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3633983ca25eSTejun Heo 3634983ca25eSTejun Heo /* link in @pwq */ 36359e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3636a357fc03SLai Jiangshan 36373c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3638d2c1d404STejun Heo } 3639d2c1d404STejun Heo 36409e8cd2f5STejun Heo /** 36419e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36429e8cd2f5STejun Heo * @wq: the target workqueue 36439e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36449e8cd2f5STejun Heo * 36459e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 36469e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 36479e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 36489e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 36499e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 36509e8cd2f5STejun Heo * 36519e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 36529e8cd2f5STejun Heo * failure. 36539e8cd2f5STejun Heo */ 36549e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36559e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36569e8cd2f5STejun Heo { 365713e2e556STejun Heo struct workqueue_attrs *new_attrs; 365813e2e556STejun Heo struct pool_workqueue *pwq = NULL, *last_pwq; 36599e8cd2f5STejun Heo struct worker_pool *pool; 36604862125bSTejun Heo int ret; 36619e8cd2f5STejun Heo 36628719dceaSTejun Heo /* only unbound workqueues can change attributes */ 36639e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 36649e8cd2f5STejun Heo return -EINVAL; 36659e8cd2f5STejun Heo 36668719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 36678719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 36688719dceaSTejun Heo return -EINVAL; 36698719dceaSTejun Heo 367013e2e556STejun Heo /* make a copy of @attrs and sanitize it */ 367113e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 367213e2e556STejun Heo if (!new_attrs) 367313e2e556STejun Heo goto enomem; 367413e2e556STejun Heo 367513e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 367613e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 367713e2e556STejun Heo 3678a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 3679a892caccSTejun Heo 36809e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 3681a892caccSTejun Heo if (!pwq) { 3682a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 368313e2e556STejun Heo goto enomem; 3684a892caccSTejun Heo } 36859e8cd2f5STejun Heo 368613e2e556STejun Heo pool = get_unbound_pool(new_attrs); 3687a892caccSTejun Heo if (!pool) { 3688a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 368913e2e556STejun Heo goto enomem; 3690a892caccSTejun Heo } 3691a892caccSTejun Heo 3692a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 36939e8cd2f5STejun Heo 36949e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 36959e8cd2f5STejun Heo if (last_pwq) { 36969e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 36979e8cd2f5STejun Heo put_pwq(last_pwq); 36989e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 36999e8cd2f5STejun Heo } 37009e8cd2f5STejun Heo 37014862125bSTejun Heo ret = 0; 37024862125bSTejun Heo /* fall through */ 37034862125bSTejun Heo out_free: 37044862125bSTejun Heo free_workqueue_attrs(new_attrs); 37054862125bSTejun Heo return ret; 370613e2e556STejun Heo 370713e2e556STejun Heo enomem: 370813e2e556STejun Heo kmem_cache_free(pwq_cache, pwq); 37094862125bSTejun Heo ret = -ENOMEM; 37104862125bSTejun Heo goto out_free; 37119e8cd2f5STejun Heo } 37129e8cd2f5STejun Heo 371330cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37141da177e4SLinus Torvalds { 371549e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 371630cdf249STejun Heo int cpu; 3717e1d8aa9fSFrederic Weisbecker 371830cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3719420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3720420c0ddbSTejun Heo if (!wq->cpu_pwqs) 372130cdf249STejun Heo return -ENOMEM; 372230cdf249STejun Heo 372330cdf249STejun Heo for_each_possible_cpu(cpu) { 37247fb98ea7STejun Heo struct pool_workqueue *pwq = 37257fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37267a62c2c8STejun Heo struct worker_pool *cpu_pools = 3727f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 372830cdf249STejun Heo 37299e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 373030cdf249STejun Heo } 373130cdf249STejun Heo return 0; 37329e8cd2f5STejun Heo } else { 37339e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37349e8cd2f5STejun Heo } 37350f900049STejun Heo } 37360f900049STejun Heo 3737f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3738f3421797STejun Heo const char *name) 3739b71ab8c2STejun Heo { 3740f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3741f3421797STejun Heo 3742f3421797STejun Heo if (max_active < 1 || max_active > lim) 3743044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3744f3421797STejun Heo max_active, name, 1, lim); 3745b71ab8c2STejun Heo 3746f3421797STejun Heo return clamp_val(max_active, 1, lim); 3747b71ab8c2STejun Heo } 3748b71ab8c2STejun Heo 3749b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 375097e37d7bSTejun Heo unsigned int flags, 37511e19ffc6STejun Heo int max_active, 3752eb13ba87SJohannes Berg struct lock_class_key *key, 3753b196be89STejun Heo const char *lock_name, ...) 37543af24433SOleg Nesterov { 3755b196be89STejun Heo va_list args, args1; 37563af24433SOleg Nesterov struct workqueue_struct *wq; 375749e3cf44STejun Heo struct pool_workqueue *pwq; 3758b196be89STejun Heo size_t namelen; 3759b196be89STejun Heo 3760b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3761b196be89STejun Heo va_start(args, lock_name); 3762b196be89STejun Heo va_copy(args1, args); 3763b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3764b196be89STejun Heo 3765b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3766b196be89STejun Heo if (!wq) 3767d2c1d404STejun Heo return NULL; 3768b196be89STejun Heo 3769b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3770b196be89STejun Heo va_end(args); 3771b196be89STejun Heo va_end(args1); 37723af24433SOleg Nesterov 3773d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3774b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 37753af24433SOleg Nesterov 3776b196be89STejun Heo /* init wq */ 377797e37d7bSTejun Heo wq->flags = flags; 3778a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 37793c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3780112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 378130cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 378273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 378373f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3784493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 37853af24433SOleg Nesterov 3786eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3787cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 37883af24433SOleg Nesterov 378930cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3790d2c1d404STejun Heo goto err_free_wq; 37911537663fSTejun Heo 3792493008a8STejun Heo /* 3793493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3794493008a8STejun Heo * have a rescuer to guarantee forward progress. 3795493008a8STejun Heo */ 3796493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3797e22bee78STejun Heo struct worker *rescuer; 3798e22bee78STejun Heo 3799d2c1d404STejun Heo rescuer = alloc_worker(); 3800e22bee78STejun Heo if (!rescuer) 3801d2c1d404STejun Heo goto err_destroy; 3802e22bee78STejun Heo 3803111c225aSTejun Heo rescuer->rescue_wq = wq; 3804111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3805b196be89STejun Heo wq->name); 3806d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3807d2c1d404STejun Heo kfree(rescuer); 3808d2c1d404STejun Heo goto err_destroy; 3809d2c1d404STejun Heo } 3810e22bee78STejun Heo 3811d2c1d404STejun Heo wq->rescuer = rescuer; 381214a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3813e22bee78STejun Heo wake_up_process(rescuer->task); 38143af24433SOleg Nesterov } 38151537663fSTejun Heo 3816226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3817226223abSTejun Heo goto err_destroy; 3818226223abSTejun Heo 38193af24433SOleg Nesterov /* 382068e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 382168e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 382268e13a67SLai Jiangshan * list. 38233af24433SOleg Nesterov */ 382468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3825a0a1a5fdSTejun Heo 3826a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 382749e3cf44STejun Heo for_each_pwq(pwq, wq) 3828699ce097STejun Heo pwq_adjust_max_active(pwq); 3829a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3830a0a1a5fdSTejun Heo 38313af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3832a0a1a5fdSTejun Heo 383368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38343af24433SOleg Nesterov 38353af24433SOleg Nesterov return wq; 3836d2c1d404STejun Heo 3837d2c1d404STejun Heo err_free_wq: 38384690c4abSTejun Heo kfree(wq); 3839d2c1d404STejun Heo return NULL; 3840d2c1d404STejun Heo err_destroy: 3841d2c1d404STejun Heo destroy_workqueue(wq); 38424690c4abSTejun Heo return NULL; 38431da177e4SLinus Torvalds } 3844d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 38451da177e4SLinus Torvalds 38463af24433SOleg Nesterov /** 38473af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 38483af24433SOleg Nesterov * @wq: target workqueue 38493af24433SOleg Nesterov * 38503af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 38513af24433SOleg Nesterov */ 38523af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 38533af24433SOleg Nesterov { 385449e3cf44STejun Heo struct pool_workqueue *pwq; 38553af24433SOleg Nesterov 38569c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 38579c5a2ba7STejun Heo drain_workqueue(wq); 3858c8efcc25STejun Heo 38596183c009STejun Heo /* sanity checks */ 3860b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 386149e3cf44STejun Heo for_each_pwq(pwq, wq) { 38626183c009STejun Heo int i; 38636183c009STejun Heo 386476af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 386576af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3866b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38676183c009STejun Heo return; 386876af4d93STejun Heo } 386976af4d93STejun Heo } 387076af4d93STejun Heo 38718864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 38728864b4e5STejun Heo WARN_ON(pwq->nr_active) || 387376af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3874b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38756183c009STejun Heo return; 38766183c009STejun Heo } 387776af4d93STejun Heo } 3878b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38796183c009STejun Heo 3880a0a1a5fdSTejun Heo /* 3881a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3882a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3883a0a1a5fdSTejun Heo */ 388468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3885d2c1d404STejun Heo list_del_init(&wq->list); 388668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38873af24433SOleg Nesterov 3888226223abSTejun Heo workqueue_sysfs_unregister(wq); 3889226223abSTejun Heo 3890493008a8STejun Heo if (wq->rescuer) { 3891e22bee78STejun Heo kthread_stop(wq->rescuer->task); 38928d9df9f0SXiaotian Feng kfree(wq->rescuer); 3893493008a8STejun Heo wq->rescuer = NULL; 3894e22bee78STejun Heo } 3895e22bee78STejun Heo 38968864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 389729c91e99STejun Heo /* 38988864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 38998864b4e5STejun Heo * free the pwqs and wq. 390029c91e99STejun Heo */ 39018864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 39028864b4e5STejun Heo kfree(wq); 39038864b4e5STejun Heo } else { 39048864b4e5STejun Heo /* 39058864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39068864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 39078864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 39088864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 39098864b4e5STejun Heo * released. 39108864b4e5STejun Heo */ 391129c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 391229c91e99STejun Heo pwqs_node); 39138864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 39148864b4e5STejun Heo put_pwq(pwq); 39158864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 391629c91e99STejun Heo } 39173af24433SOleg Nesterov } 39183af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39193af24433SOleg Nesterov 3920dcd989cbSTejun Heo /** 3921dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3922dcd989cbSTejun Heo * @wq: target workqueue 3923dcd989cbSTejun Heo * @max_active: new max_active value. 3924dcd989cbSTejun Heo * 3925dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3926dcd989cbSTejun Heo * 3927dcd989cbSTejun Heo * CONTEXT: 3928dcd989cbSTejun Heo * Don't call from IRQ context. 3929dcd989cbSTejun Heo */ 3930dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3931dcd989cbSTejun Heo { 393249e3cf44STejun Heo struct pool_workqueue *pwq; 3933dcd989cbSTejun Heo 39348719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39358719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39368719dceaSTejun Heo return; 39378719dceaSTejun Heo 3938f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3939dcd989cbSTejun Heo 3940a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 3941dcd989cbSTejun Heo 3942dcd989cbSTejun Heo wq->saved_max_active = max_active; 3943dcd989cbSTejun Heo 3944699ce097STejun Heo for_each_pwq(pwq, wq) 3945699ce097STejun Heo pwq_adjust_max_active(pwq); 3946dcd989cbSTejun Heo 3947a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3948dcd989cbSTejun Heo } 3949dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3950dcd989cbSTejun Heo 3951dcd989cbSTejun Heo /** 3952e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 3953e6267616STejun Heo * 3954e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 3955e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 3956e6267616STejun Heo */ 3957e6267616STejun Heo bool current_is_workqueue_rescuer(void) 3958e6267616STejun Heo { 3959e6267616STejun Heo struct worker *worker = current_wq_worker(); 3960e6267616STejun Heo 39616a092dfdSLai Jiangshan return worker && worker->rescue_wq; 3962e6267616STejun Heo } 3963e6267616STejun Heo 3964e6267616STejun Heo /** 3965dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3966dcd989cbSTejun Heo * @cpu: CPU in question 3967dcd989cbSTejun Heo * @wq: target workqueue 3968dcd989cbSTejun Heo * 3969dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3970dcd989cbSTejun Heo * no synchronization around this function and the test result is 3971dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3972dcd989cbSTejun Heo * 3973dcd989cbSTejun Heo * RETURNS: 3974dcd989cbSTejun Heo * %true if congested, %false otherwise. 3975dcd989cbSTejun Heo */ 3976d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 3977dcd989cbSTejun Heo { 39787fb98ea7STejun Heo struct pool_workqueue *pwq; 397976af4d93STejun Heo bool ret; 398076af4d93STejun Heo 398188109453SLai Jiangshan rcu_read_lock_sched(); 39827fb98ea7STejun Heo 39837fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 39847fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 39857fb98ea7STejun Heo else 39867fb98ea7STejun Heo pwq = first_pwq(wq); 3987dcd989cbSTejun Heo 398876af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 398988109453SLai Jiangshan rcu_read_unlock_sched(); 399076af4d93STejun Heo 399176af4d93STejun Heo return ret; 3992dcd989cbSTejun Heo } 3993dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3994dcd989cbSTejun Heo 3995dcd989cbSTejun Heo /** 3996dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3997dcd989cbSTejun Heo * @work: the work to be tested 3998dcd989cbSTejun Heo * 3999dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4000dcd989cbSTejun Heo * synchronization around this function and the test result is 4001dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4002dcd989cbSTejun Heo * 4003dcd989cbSTejun Heo * RETURNS: 4004dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4005dcd989cbSTejun Heo */ 4006dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4007dcd989cbSTejun Heo { 4008fa1b54e6STejun Heo struct worker_pool *pool; 4009dcd989cbSTejun Heo unsigned long flags; 4010dcd989cbSTejun Heo unsigned int ret = 0; 4011dcd989cbSTejun Heo 4012dcd989cbSTejun Heo if (work_pending(work)) 4013dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4014038366c5SLai Jiangshan 4015fa1b54e6STejun Heo local_irq_save(flags); 4016fa1b54e6STejun Heo pool = get_work_pool(work); 4017038366c5SLai Jiangshan if (pool) { 4018fa1b54e6STejun Heo spin_lock(&pool->lock); 4019c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4020dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4021fa1b54e6STejun Heo spin_unlock(&pool->lock); 4022038366c5SLai Jiangshan } 4023fa1b54e6STejun Heo local_irq_restore(flags); 4024dcd989cbSTejun Heo 4025dcd989cbSTejun Heo return ret; 4026dcd989cbSTejun Heo } 4027dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4028dcd989cbSTejun Heo 4029db7bccf4STejun Heo /* 4030db7bccf4STejun Heo * CPU hotplug. 4031db7bccf4STejun Heo * 4032e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4033112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4034706026c2STejun Heo * pool which make migrating pending and scheduled works very 4035e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 403694cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4037e22bee78STejun Heo * blocked draining impractical. 4038e22bee78STejun Heo * 403924647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4040628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4041628c78e7STejun Heo * cpu comes back online. 4042db7bccf4STejun Heo */ 4043db7bccf4STejun Heo 4044706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4045db7bccf4STejun Heo { 404638db41d9STejun Heo int cpu = smp_processor_id(); 40474ce62e9eSTejun Heo struct worker_pool *pool; 4048db7bccf4STejun Heo struct worker *worker; 4049a9ab775bSTejun Heo int wi; 4050db7bccf4STejun Heo 4051f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 40526183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4053db7bccf4STejun Heo 4054bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 405594cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4056e22bee78STejun Heo 4057f2d5a0eeSTejun Heo /* 4058bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 405994cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 406094cf58bbSTejun Heo * except for the ones which are still executing works from 406194cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 406294cf58bbSTejun Heo * this, they may become diasporas. 4063f2d5a0eeSTejun Heo */ 4064a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4065403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4066db7bccf4STejun Heo 406724647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4068f2d5a0eeSTejun Heo 406994cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4070bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 407194cf58bbSTejun Heo } 4072e22bee78STejun Heo 4073e22bee78STejun Heo /* 4074628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4075628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4076628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4077628c78e7STejun Heo */ 4078628c78e7STejun Heo schedule(); 4079628c78e7STejun Heo 4080628c78e7STejun Heo /* 4081628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4082628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 408338db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 408438db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 408538db41d9STejun Heo * pools which are served by workers tied to the CPU. 4086628c78e7STejun Heo * 4087628c78e7STejun Heo * On return from this function, the current worker would trigger 4088628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4089628c78e7STejun Heo * didn't already. 4090e22bee78STejun Heo */ 4091f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4092e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4093db7bccf4STejun Heo } 4094db7bccf4STejun Heo 4095bd7c089eSTejun Heo /** 4096bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4097bd7c089eSTejun Heo * @pool: pool of interest 4098bd7c089eSTejun Heo * 4099a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4100bd7c089eSTejun Heo */ 4101bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4102bd7c089eSTejun Heo { 4103a9ab775bSTejun Heo struct worker *worker; 4104a9ab775bSTejun Heo int wi; 4105bd7c089eSTejun Heo 4106bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4107bd7c089eSTejun Heo 4108bd7c089eSTejun Heo /* 4109a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4110a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4111a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4112a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4113a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4114bd7c089eSTejun Heo */ 4115a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4116a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4117a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4118a9ab775bSTejun Heo 4119a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4120a9ab775bSTejun Heo 4121a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4122a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4123a9ab775bSTejun Heo 4124a9ab775bSTejun Heo /* 4125a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4126a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4127a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4128a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4129a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4130a9ab775bSTejun Heo * be bound before @pool->lock is released. 4131a9ab775bSTejun Heo */ 4132a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4133bd7c089eSTejun Heo wake_up_process(worker->task); 4134bd7c089eSTejun Heo 4135bd7c089eSTejun Heo /* 4136a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4137a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4138a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4139a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4140a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4141a9ab775bSTejun Heo * concurrency management. Note that when or whether 4142a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4143a9ab775bSTejun Heo * 4144a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4145a9ab775bSTejun Heo * tested without holding any lock in 4146a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4147a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4148a9ab775bSTejun Heo * management operations. 4149bd7c089eSTejun Heo */ 4150a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4151a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4152a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4153a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4154bd7c089eSTejun Heo } 4155a9ab775bSTejun Heo 4156a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4157bd7c089eSTejun Heo } 4158bd7c089eSTejun Heo 41597dbc725eSTejun Heo /** 41607dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 41617dbc725eSTejun Heo * @pool: unbound pool of interest 41627dbc725eSTejun Heo * @cpu: the CPU which is coming up 41637dbc725eSTejun Heo * 41647dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 41657dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 41667dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 41677dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 41687dbc725eSTejun Heo */ 41697dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 41707dbc725eSTejun Heo { 41717dbc725eSTejun Heo static cpumask_t cpumask; 41727dbc725eSTejun Heo struct worker *worker; 41737dbc725eSTejun Heo int wi; 41747dbc725eSTejun Heo 41757dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 41767dbc725eSTejun Heo 41777dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 41787dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 41797dbc725eSTejun Heo return; 41807dbc725eSTejun Heo 41817dbc725eSTejun Heo /* is @cpu the only online CPU? */ 41827dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 41837dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 41847dbc725eSTejun Heo return; 41857dbc725eSTejun Heo 41867dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 41877dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 41887dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 41897dbc725eSTejun Heo pool->attrs->cpumask) < 0); 41907dbc725eSTejun Heo } 41917dbc725eSTejun Heo 41928db25e78STejun Heo /* 41938db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 41948db25e78STejun Heo * This will be registered high priority CPU notifier. 41958db25e78STejun Heo */ 41969fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 41971da177e4SLinus Torvalds unsigned long action, 41981da177e4SLinus Torvalds void *hcpu) 41991da177e4SLinus Torvalds { 4200d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42014ce62e9eSTejun Heo struct worker_pool *pool; 42027dbc725eSTejun Heo int pi; 42031da177e4SLinus Torvalds 42048db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 42053af24433SOleg Nesterov case CPU_UP_PREPARE: 4206f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 42073ce63377STejun Heo if (pool->nr_workers) 42083ce63377STejun Heo continue; 4209ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 42103ce63377STejun Heo return NOTIFY_BAD; 42113af24433SOleg Nesterov } 42121da177e4SLinus Torvalds break; 42131da177e4SLinus Torvalds 421465758202STejun Heo case CPU_DOWN_FAILED: 421565758202STejun Heo case CPU_ONLINE: 421668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 42177dbc725eSTejun Heo 42187dbc725eSTejun Heo for_each_pool(pool, pi) { 4219bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 422094cf58bbSTejun Heo 42217dbc725eSTejun Heo if (pool->cpu == cpu) { 4222a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 422324647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4224a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4225a9ab775bSTejun Heo 422694cf58bbSTejun Heo rebind_workers(pool); 42277dbc725eSTejun Heo } else if (pool->cpu < 0) { 42287dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 42297dbc725eSTejun Heo } 423094cf58bbSTejun Heo 4231bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 423294cf58bbSTejun Heo } 42337dbc725eSTejun Heo 423468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42358db25e78STejun Heo break; 423665758202STejun Heo } 423765758202STejun Heo return NOTIFY_OK; 423865758202STejun Heo } 423965758202STejun Heo 424065758202STejun Heo /* 424165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 424265758202STejun Heo * This will be registered as low priority CPU notifier. 424365758202STejun Heo */ 42449fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 424565758202STejun Heo unsigned long action, 424665758202STejun Heo void *hcpu) 424765758202STejun Heo { 4248d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42498db25e78STejun Heo struct work_struct unbind_work; 42508db25e78STejun Heo 425165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 425265758202STejun Heo case CPU_DOWN_PREPARE: 42538db25e78STejun Heo /* unbinding should happen on the local CPU */ 4254706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42557635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42568db25e78STejun Heo flush_work(&unbind_work); 42578db25e78STejun Heo break; 425865758202STejun Heo } 425965758202STejun Heo return NOTIFY_OK; 426065758202STejun Heo } 426165758202STejun Heo 42622d3854a3SRusty Russell #ifdef CONFIG_SMP 42638ccad40dSRusty Russell 42642d3854a3SRusty Russell struct work_for_cpu { 4265ed48ece2STejun Heo struct work_struct work; 42662d3854a3SRusty Russell long (*fn)(void *); 42672d3854a3SRusty Russell void *arg; 42682d3854a3SRusty Russell long ret; 42692d3854a3SRusty Russell }; 42702d3854a3SRusty Russell 4271ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 42722d3854a3SRusty Russell { 4273ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4274ed48ece2STejun Heo 42752d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 42762d3854a3SRusty Russell } 42772d3854a3SRusty Russell 42782d3854a3SRusty Russell /** 42792d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 42802d3854a3SRusty Russell * @cpu: the cpu to run on 42812d3854a3SRusty Russell * @fn: the function to run 42822d3854a3SRusty Russell * @arg: the function arg 42832d3854a3SRusty Russell * 428431ad9081SRusty Russell * This will return the value @fn returns. 428531ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 42866b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 42872d3854a3SRusty Russell */ 4288d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 42892d3854a3SRusty Russell { 4290ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 42912d3854a3SRusty Russell 4292ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4293ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4294ed48ece2STejun Heo flush_work(&wfc.work); 42952d3854a3SRusty Russell return wfc.ret; 42962d3854a3SRusty Russell } 42972d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 42982d3854a3SRusty Russell #endif /* CONFIG_SMP */ 42992d3854a3SRusty Russell 4300a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4301e7577c50SRusty Russell 4302a0a1a5fdSTejun Heo /** 4303a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4304a0a1a5fdSTejun Heo * 430558a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4306c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4307706026c2STejun Heo * pool->worklist. 4308a0a1a5fdSTejun Heo * 4309a0a1a5fdSTejun Heo * CONTEXT: 4310a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4311a0a1a5fdSTejun Heo */ 4312a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4313a0a1a5fdSTejun Heo { 431417116969STejun Heo struct worker_pool *pool; 431524b8a847STejun Heo struct workqueue_struct *wq; 431624b8a847STejun Heo struct pool_workqueue *pwq; 4317611c92a0STejun Heo int pi; 4318a0a1a5fdSTejun Heo 431968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4320a0a1a5fdSTejun Heo 43216183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4322a0a1a5fdSTejun Heo workqueue_freezing = true; 4323a0a1a5fdSTejun Heo 432424b8a847STejun Heo /* set FREEZING */ 4325611c92a0STejun Heo for_each_pool(pool, pi) { 43265bcab335STejun Heo spin_lock_irq(&pool->lock); 432735b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 432835b6bb63STejun Heo pool->flags |= POOL_FREEZING; 43295bcab335STejun Heo spin_unlock_irq(&pool->lock); 43301da177e4SLinus Torvalds } 43318b03ae3cSTejun Heo 433224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4333a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4334699ce097STejun Heo for_each_pwq(pwq, wq) 4335699ce097STejun Heo pwq_adjust_max_active(pwq); 4336a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4337a1056305STejun Heo } 43385bcab335STejun Heo 433968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4340a0a1a5fdSTejun Heo } 4341a0a1a5fdSTejun Heo 4342a0a1a5fdSTejun Heo /** 434358a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4344a0a1a5fdSTejun Heo * 4345a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4346a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4347a0a1a5fdSTejun Heo * 4348a0a1a5fdSTejun Heo * CONTEXT: 434968e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4350a0a1a5fdSTejun Heo * 4351a0a1a5fdSTejun Heo * RETURNS: 435258a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 435358a69cb4STejun Heo * is complete. 4354a0a1a5fdSTejun Heo */ 4355a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4356a0a1a5fdSTejun Heo { 4357a0a1a5fdSTejun Heo bool busy = false; 435824b8a847STejun Heo struct workqueue_struct *wq; 435924b8a847STejun Heo struct pool_workqueue *pwq; 4360a0a1a5fdSTejun Heo 436168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4362a0a1a5fdSTejun Heo 43636183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4364a0a1a5fdSTejun Heo 436524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 436624b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 436724b8a847STejun Heo continue; 4368a0a1a5fdSTejun Heo /* 4369a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4370a0a1a5fdSTejun Heo * to peek without lock. 4371a0a1a5fdSTejun Heo */ 437288109453SLai Jiangshan rcu_read_lock_sched(); 437324b8a847STejun Heo for_each_pwq(pwq, wq) { 43746183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4375112202d9STejun Heo if (pwq->nr_active) { 4376a0a1a5fdSTejun Heo busy = true; 437788109453SLai Jiangshan rcu_read_unlock_sched(); 4378a0a1a5fdSTejun Heo goto out_unlock; 4379a0a1a5fdSTejun Heo } 4380a0a1a5fdSTejun Heo } 438188109453SLai Jiangshan rcu_read_unlock_sched(); 4382a0a1a5fdSTejun Heo } 4383a0a1a5fdSTejun Heo out_unlock: 438468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4385a0a1a5fdSTejun Heo return busy; 4386a0a1a5fdSTejun Heo } 4387a0a1a5fdSTejun Heo 4388a0a1a5fdSTejun Heo /** 4389a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4390a0a1a5fdSTejun Heo * 4391a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4392706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4393a0a1a5fdSTejun Heo * 4394a0a1a5fdSTejun Heo * CONTEXT: 4395a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4396a0a1a5fdSTejun Heo */ 4397a0a1a5fdSTejun Heo void thaw_workqueues(void) 4398a0a1a5fdSTejun Heo { 439924b8a847STejun Heo struct workqueue_struct *wq; 440024b8a847STejun Heo struct pool_workqueue *pwq; 440124b8a847STejun Heo struct worker_pool *pool; 4402611c92a0STejun Heo int pi; 4403a0a1a5fdSTejun Heo 440468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4405a0a1a5fdSTejun Heo 4406a0a1a5fdSTejun Heo if (!workqueue_freezing) 4407a0a1a5fdSTejun Heo goto out_unlock; 4408a0a1a5fdSTejun Heo 440924b8a847STejun Heo /* clear FREEZING */ 4410611c92a0STejun Heo for_each_pool(pool, pi) { 44115bcab335STejun Heo spin_lock_irq(&pool->lock); 441235b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 441335b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 44145bcab335STejun Heo spin_unlock_irq(&pool->lock); 4415d565ed63STejun Heo } 441624b8a847STejun Heo 441724b8a847STejun Heo /* restore max_active and repopulate worklist */ 441824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4419a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4420699ce097STejun Heo for_each_pwq(pwq, wq) 4421699ce097STejun Heo pwq_adjust_max_active(pwq); 4422a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 442324b8a847STejun Heo } 442424b8a847STejun Heo 4425a0a1a5fdSTejun Heo workqueue_freezing = false; 4426a0a1a5fdSTejun Heo out_unlock: 442768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4428a0a1a5fdSTejun Heo } 4429a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4430a0a1a5fdSTejun Heo 4431bce90380STejun Heo static void __init wq_numa_init(void) 4432bce90380STejun Heo { 4433bce90380STejun Heo cpumask_var_t *tbl; 4434bce90380STejun Heo int node, cpu; 4435bce90380STejun Heo 4436bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4437bce90380STejun Heo for_each_node(node) 4438bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4439bce90380STejun Heo 4440bce90380STejun Heo if (num_possible_nodes() <= 1) 4441bce90380STejun Heo return; 4442bce90380STejun Heo 4443bce90380STejun Heo /* 4444bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4445bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4446bce90380STejun Heo * fully initialized by now. 4447bce90380STejun Heo */ 4448bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4449bce90380STejun Heo BUG_ON(!tbl); 4450bce90380STejun Heo 4451bce90380STejun Heo for_each_node(node) 4452bce90380STejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node)); 4453bce90380STejun Heo 4454bce90380STejun Heo for_each_possible_cpu(cpu) { 4455bce90380STejun Heo node = cpu_to_node(cpu); 4456bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4457bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4458bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4459bce90380STejun Heo return; 4460bce90380STejun Heo } 4461bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4462bce90380STejun Heo } 4463bce90380STejun Heo 4464bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4465bce90380STejun Heo wq_numa_enabled = true; 4466bce90380STejun Heo } 4467bce90380STejun Heo 44686ee0578bSSuresh Siddha static int __init init_workqueues(void) 44691da177e4SLinus Torvalds { 44707a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44717a4e344cSTejun Heo int i, cpu; 4472c34056a3STejun Heo 44737c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44747c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44756be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4476b5490077STejun Heo 4477e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4478e904e6c2STejun Heo 4479e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4480e904e6c2STejun Heo 448165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4482a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44838b03ae3cSTejun Heo 4484bce90380STejun Heo wq_numa_init(); 4485bce90380STejun Heo 4486706026c2STejun Heo /* initialize CPU pools */ 448729c91e99STejun Heo for_each_possible_cpu(cpu) { 44884ce62e9eSTejun Heo struct worker_pool *pool; 44898b03ae3cSTejun Heo 44907a4e344cSTejun Heo i = 0; 4491f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44927a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4493ec22ca5eSTejun Heo pool->cpu = cpu; 44947a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 44957a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4496f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 44977a4e344cSTejun Heo 44989daf9e67STejun Heo /* alloc pool ID */ 449968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45009daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 450168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45024ce62e9eSTejun Heo } 45038b03ae3cSTejun Heo } 45048b03ae3cSTejun Heo 4505e22bee78STejun Heo /* create the initial worker */ 450629c91e99STejun Heo for_each_online_cpu(cpu) { 45074ce62e9eSTejun Heo struct worker_pool *pool; 4508e22bee78STejun Heo 4509f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 451024647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4511ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4512e22bee78STejun Heo } 45134ce62e9eSTejun Heo } 4514e22bee78STejun Heo 451529c91e99STejun Heo /* create default unbound wq attrs */ 451629c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 451729c91e99STejun Heo struct workqueue_attrs *attrs; 451829c91e99STejun Heo 451929c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 452029c91e99STejun Heo attrs->nice = std_nice[i]; 452129c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 452229c91e99STejun Heo } 452329c91e99STejun Heo 4524d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 45251aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4526d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4527f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4528f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 452924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 453024d51addSTejun Heo WQ_FREEZABLE, 0); 45311aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4532ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 45336ee0578bSSuresh Siddha return 0; 45341da177e4SLinus Torvalds } 45356ee0578bSSuresh Siddha early_initcall(init_workqueues); 4536