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, 104ecf6881fSTejun Heo 105ecf6881fSTejun Heo WQ_NAME_LEN = 24, 106c8e55f36STejun Heo }; 107c8e55f36STejun Heo 1081da177e4SLinus Torvalds /* 1094690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1104690c4abSTejun Heo * 111e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 112e41e704bSTejun Heo * everyone else. 1134690c4abSTejun Heo * 114e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 115e22bee78STejun Heo * only be modified and accessed from the local cpu. 116e22bee78STejun Heo * 117d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1184690c4abSTejun Heo * 119d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 120d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 121d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 122d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 123e22bee78STejun Heo * 124822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 125822d8405STejun Heo * locks. Reads can happen under either lock. 126822d8405STejun Heo * 12768e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12876af4d93STejun Heo * 12968e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1305bcab335STejun Heo * 1313c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1323c25a55dSLai Jiangshan * 133b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1342e109a28STejun Heo * 1352e109a28STejun Heo * MD: wq_mayday_lock protected. 1364690c4abSTejun Heo */ 1374690c4abSTejun Heo 1382eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 139c34056a3STejun Heo 140bd7bdd43STejun Heo struct worker_pool { 141d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 142d84ff051STejun Heo int cpu; /* I: the associated cpu */ 143f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1449daf9e67STejun Heo int id; /* I: pool ID */ 14511ebea50STejun Heo unsigned int flags; /* X: flags */ 146bd7bdd43STejun Heo 147bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 148bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 149ea1abd61SLai Jiangshan 150ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 151bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 152bd7bdd43STejun Heo 153bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 154bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 155bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 156bd7bdd43STejun Heo 157c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 158c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 159c9e7cf27STejun Heo /* L: hash of busy workers */ 160c9e7cf27STejun Heo 161bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16234a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 163bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 164822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 165e19e397aSTejun Heo 1667a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16768e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 16868e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1697a4e344cSTejun Heo 170e19e397aSTejun Heo /* 171e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 172e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 173e19e397aSTejun Heo * cacheline. 174e19e397aSTejun Heo */ 175e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17629c91e99STejun Heo 17729c91e99STejun Heo /* 17829c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 17929c91e99STejun Heo * from get_work_pool(). 18029c91e99STejun Heo */ 18129c91e99STejun Heo struct rcu_head rcu; 1828b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1838b03ae3cSTejun Heo 1848b03ae3cSTejun Heo /* 185112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 186112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 187112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 188112202d9STejun Heo * number of flag bits. 1891da177e4SLinus Torvalds */ 190112202d9STejun Heo struct pool_workqueue { 191bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1924690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19373f53c4aSTejun Heo int work_color; /* L: current color */ 19473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1958864b4e5STejun Heo int refcnt; /* L: reference count */ 19673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19773f53c4aSTejun Heo /* L: nr of in_flight works */ 1981e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 199a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2001e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2013c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2022e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2038864b4e5STejun Heo 2048864b4e5STejun Heo /* 2058864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2068864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2078864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 208b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2098864b4e5STejun Heo */ 2108864b4e5STejun Heo struct work_struct unbound_release_work; 2118864b4e5STejun Heo struct rcu_head rcu; 212e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2131da177e4SLinus Torvalds 2141da177e4SLinus Torvalds /* 21573f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21673f53c4aSTejun Heo */ 21773f53c4aSTejun Heo struct wq_flusher { 2183c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2193c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22073f53c4aSTejun Heo struct completion done; /* flush completion */ 22173f53c4aSTejun Heo }; 2221da177e4SLinus Torvalds 223226223abSTejun Heo struct wq_device; 224226223abSTejun Heo 22573f53c4aSTejun Heo /* 226c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 227c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2281da177e4SLinus Torvalds */ 2291da177e4SLinus Torvalds struct workqueue_struct { 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 2476029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2486029a918STejun Heo 249226223abSTejun Heo #ifdef CONFIG_SYSFS 250226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 251226223abSTejun Heo #endif 2524e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2534e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2544e6045f1SJohannes Berg #endif 255ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2562728fd2fSTejun Heo 2572728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2582728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2592728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 260df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2611da177e4SLinus Torvalds }; 2621da177e4SLinus Torvalds 263e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 264e904e6c2STejun Heo 265bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 266bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 267bce90380STejun Heo /* possible CPUs of each node */ 268bce90380STejun Heo 269bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 270bce90380STejun Heo 27168e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2722e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2735bcab335STejun Heo 27468e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 27568e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2767d19c5ceSTejun Heo 2777d19c5ceSTejun Heo /* the per-cpu worker pools */ 2787d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2797d19c5ceSTejun Heo cpu_worker_pools); 2807d19c5ceSTejun Heo 28168e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2827d19c5ceSTejun Heo 28368e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 28429c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 28529c91e99STejun Heo 286c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 28729c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 28829c91e99STejun Heo 289d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 290d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 291044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2921aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 293044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 294d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 295044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 296f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 297044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 29824d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 299d320c038STejun Heo 3007d19c5ceSTejun Heo static int worker_thread(void *__worker); 3017d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3027d19c5ceSTejun Heo const struct workqueue_attrs *from); 3037d19c5ceSTejun Heo 30497bd2347STejun Heo #define CREATE_TRACE_POINTS 30597bd2347STejun Heo #include <trace/events/workqueue.h> 30697bd2347STejun Heo 30768e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3085bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 30968e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 31068e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3115bcab335STejun Heo 312b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 31376af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 314b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 315b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 31676af4d93STejun Heo 317822d8405STejun Heo #ifdef CONFIG_LOCKDEP 318822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 319519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 320519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 321822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 322822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 323822d8405STejun Heo #else 324822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 325822d8405STejun Heo #endif 326822d8405STejun Heo 327f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 328f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 329f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3307a62c2c8STejun Heo (pool)++) 3314ce62e9eSTejun Heo 33249e3cf44STejun Heo /** 33317116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 33417116969STejun Heo * @pool: iteration cursor 335611c92a0STejun Heo * @pi: integer used for iteration 336fa1b54e6STejun Heo * 33768e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 33868e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 33968e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 340fa1b54e6STejun Heo * 341fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 342fa1b54e6STejun Heo * ignored. 34317116969STejun Heo */ 344611c92a0STejun Heo #define for_each_pool(pool, pi) \ 345611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 34668e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 347fa1b54e6STejun Heo else 34817116969STejun Heo 34917116969STejun Heo /** 350822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 351822d8405STejun Heo * @worker: iteration cursor 352822d8405STejun Heo * @wi: integer used for iteration 353822d8405STejun Heo * @pool: worker_pool to iterate workers of 354822d8405STejun Heo * 355822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 356822d8405STejun Heo * 357822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 358822d8405STejun Heo * ignored. 359822d8405STejun Heo */ 360822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 361822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 362822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 363822d8405STejun Heo else 364822d8405STejun Heo 365822d8405STejun Heo /** 36649e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 36749e3cf44STejun Heo * @pwq: iteration cursor 36849e3cf44STejun Heo * @wq: the target workqueue 36976af4d93STejun Heo * 370b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 371794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 372794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 37376af4d93STejun Heo * 37476af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 37576af4d93STejun Heo * ignored. 37649e3cf44STejun Heo */ 37749e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 37876af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 379b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 38076af4d93STejun Heo else 381f3421797STejun Heo 382dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 383dc186ad7SThomas Gleixner 384dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 385dc186ad7SThomas Gleixner 38699777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 38799777288SStanislaw Gruszka { 38899777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 38999777288SStanislaw Gruszka } 39099777288SStanislaw Gruszka 391dc186ad7SThomas Gleixner /* 392dc186ad7SThomas Gleixner * fixup_init is called when: 393dc186ad7SThomas Gleixner * - an active object is initialized 394dc186ad7SThomas Gleixner */ 395dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 396dc186ad7SThomas Gleixner { 397dc186ad7SThomas Gleixner struct work_struct *work = addr; 398dc186ad7SThomas Gleixner 399dc186ad7SThomas Gleixner switch (state) { 400dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 401dc186ad7SThomas Gleixner cancel_work_sync(work); 402dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 403dc186ad7SThomas Gleixner return 1; 404dc186ad7SThomas Gleixner default: 405dc186ad7SThomas Gleixner return 0; 406dc186ad7SThomas Gleixner } 407dc186ad7SThomas Gleixner } 408dc186ad7SThomas Gleixner 409dc186ad7SThomas Gleixner /* 410dc186ad7SThomas Gleixner * fixup_activate is called when: 411dc186ad7SThomas Gleixner * - an active object is activated 412dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 413dc186ad7SThomas Gleixner */ 414dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 415dc186ad7SThomas Gleixner { 416dc186ad7SThomas Gleixner struct work_struct *work = addr; 417dc186ad7SThomas Gleixner 418dc186ad7SThomas Gleixner switch (state) { 419dc186ad7SThomas Gleixner 420dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 421dc186ad7SThomas Gleixner /* 422dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 423dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 424dc186ad7SThomas Gleixner * is tracked in the object tracker. 425dc186ad7SThomas Gleixner */ 42622df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 427dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 428dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 429dc186ad7SThomas Gleixner return 0; 430dc186ad7SThomas Gleixner } 431dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 432dc186ad7SThomas Gleixner return 0; 433dc186ad7SThomas Gleixner 434dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 435dc186ad7SThomas Gleixner WARN_ON(1); 436dc186ad7SThomas Gleixner 437dc186ad7SThomas Gleixner default: 438dc186ad7SThomas Gleixner return 0; 439dc186ad7SThomas Gleixner } 440dc186ad7SThomas Gleixner } 441dc186ad7SThomas Gleixner 442dc186ad7SThomas Gleixner /* 443dc186ad7SThomas Gleixner * fixup_free is called when: 444dc186ad7SThomas Gleixner * - an active object is freed 445dc186ad7SThomas Gleixner */ 446dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 447dc186ad7SThomas Gleixner { 448dc186ad7SThomas Gleixner struct work_struct *work = addr; 449dc186ad7SThomas Gleixner 450dc186ad7SThomas Gleixner switch (state) { 451dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 452dc186ad7SThomas Gleixner cancel_work_sync(work); 453dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 454dc186ad7SThomas Gleixner return 1; 455dc186ad7SThomas Gleixner default: 456dc186ad7SThomas Gleixner return 0; 457dc186ad7SThomas Gleixner } 458dc186ad7SThomas Gleixner } 459dc186ad7SThomas Gleixner 460dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 461dc186ad7SThomas Gleixner .name = "work_struct", 46299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 463dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 464dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 465dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 466dc186ad7SThomas Gleixner }; 467dc186ad7SThomas Gleixner 468dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 469dc186ad7SThomas Gleixner { 470dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 471dc186ad7SThomas Gleixner } 472dc186ad7SThomas Gleixner 473dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 474dc186ad7SThomas Gleixner { 475dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 476dc186ad7SThomas Gleixner } 477dc186ad7SThomas Gleixner 478dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 479dc186ad7SThomas Gleixner { 480dc186ad7SThomas Gleixner if (onstack) 481dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 482dc186ad7SThomas Gleixner else 483dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 484dc186ad7SThomas Gleixner } 485dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 486dc186ad7SThomas Gleixner 487dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 488dc186ad7SThomas Gleixner { 489dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 490dc186ad7SThomas Gleixner } 491dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 492dc186ad7SThomas Gleixner 493dc186ad7SThomas Gleixner #else 494dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 495dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 496dc186ad7SThomas Gleixner #endif 497dc186ad7SThomas Gleixner 4989daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4999daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5009daf9e67STejun Heo { 5019daf9e67STejun Heo int ret; 5029daf9e67STejun Heo 50368e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5045bcab335STejun Heo 505fa1b54e6STejun Heo do { 506fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 507fa1b54e6STejun Heo return -ENOMEM; 5089daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 509fa1b54e6STejun Heo } while (ret == -EAGAIN); 5109daf9e67STejun Heo 5119daf9e67STejun Heo return ret; 5129daf9e67STejun Heo } 5139daf9e67STejun Heo 51476af4d93STejun Heo /** 51576af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 51676af4d93STejun Heo * @wq: the target workqueue 51776af4d93STejun Heo * 518b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 519794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 520794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 52176af4d93STejun Heo */ 5227fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 523a848e3b6SOleg Nesterov { 524b09f4fd3SLai Jiangshan assert_rcu_or_wq_mutex(wq); 52576af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 52676af4d93STejun Heo pwqs_node); 527f3421797STejun Heo } 528a848e3b6SOleg Nesterov 529df2d5ae4STejun Heo /** 530df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 531df2d5ae4STejun Heo * @wq: the target workqueue 532df2d5ae4STejun Heo * @node: the node ID 533df2d5ae4STejun Heo * 534df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 535df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 536df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 537df2d5ae4STejun Heo */ 538df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 539df2d5ae4STejun Heo int node) 540df2d5ae4STejun Heo { 541df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 542df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 543df2d5ae4STejun Heo } 544df2d5ae4STejun Heo 54573f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 54673f53c4aSTejun Heo { 54773f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 54873f53c4aSTejun Heo } 54973f53c4aSTejun Heo 55073f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 55173f53c4aSTejun Heo { 55273f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 55373f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 55473f53c4aSTejun Heo } 55573f53c4aSTejun Heo 55673f53c4aSTejun Heo static int work_next_color(int color) 55773f53c4aSTejun Heo { 55873f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 559b1f4ec17SOleg Nesterov } 560b1f4ec17SOleg Nesterov 5614594bf15SDavid Howells /* 562112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 563112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5647c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5657a22ad75STejun Heo * 566112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 567112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 568bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 569bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5707a22ad75STejun Heo * 571112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5727c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 573112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5747c3eed5cSTejun Heo * available only while the work item is queued. 575bbb68dfaSTejun Heo * 576bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 577bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 578bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 579bbb68dfaSTejun Heo * try to steal the PENDING bit. 5804594bf15SDavid Howells */ 5817a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5827a22ad75STejun Heo unsigned long flags) 5837a22ad75STejun Heo { 5846183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5857a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5867a22ad75STejun Heo } 5877a22ad75STejun Heo 588112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5894690c4abSTejun Heo unsigned long extra_flags) 590365970a1SDavid Howells { 591112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 592112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 593365970a1SDavid Howells } 594365970a1SDavid Howells 5954468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5964468a00fSLai Jiangshan int pool_id) 5974468a00fSLai Jiangshan { 5984468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5994468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6004468a00fSLai Jiangshan } 6014468a00fSLai Jiangshan 6027c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6037c3eed5cSTejun Heo int pool_id) 6044d707b9fSOleg Nesterov { 60523657bb1STejun Heo /* 60623657bb1STejun Heo * The following wmb is paired with the implied mb in 60723657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 60823657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 60923657bb1STejun Heo * owner. 61023657bb1STejun Heo */ 61123657bb1STejun Heo smp_wmb(); 6127c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6134d707b9fSOleg Nesterov } 6144d707b9fSOleg Nesterov 6157a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 616365970a1SDavid Howells { 6177c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6187c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6197a22ad75STejun Heo } 6207a22ad75STejun Heo 621112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6227a22ad75STejun Heo { 623e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6247a22ad75STejun Heo 625112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 626e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 627e120153dSTejun Heo else 628e120153dSTejun Heo return NULL; 6297a22ad75STejun Heo } 6307a22ad75STejun Heo 6317c3eed5cSTejun Heo /** 6327c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6337c3eed5cSTejun Heo * @work: the work item of interest 6347c3eed5cSTejun Heo * 6357c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 636fa1b54e6STejun Heo * 63768e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 63868e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 63968e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 640fa1b54e6STejun Heo * 641fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 642fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 643fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 644fa1b54e6STejun Heo * returned pool is and stays online. 6457c3eed5cSTejun Heo */ 6467c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6477a22ad75STejun Heo { 648e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6497c3eed5cSTejun Heo int pool_id; 6507a22ad75STejun Heo 65168e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 652fa1b54e6STejun Heo 653112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 654112202d9STejun Heo return ((struct pool_workqueue *) 6557c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6567a22ad75STejun Heo 6577c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6587c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6597a22ad75STejun Heo return NULL; 6607a22ad75STejun Heo 661fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6627c3eed5cSTejun Heo } 6637c3eed5cSTejun Heo 6647c3eed5cSTejun Heo /** 6657c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6667c3eed5cSTejun Heo * @work: the work item of interest 6677c3eed5cSTejun Heo * 6687c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6697c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6707c3eed5cSTejun Heo */ 6717c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6727c3eed5cSTejun Heo { 67354d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6747c3eed5cSTejun Heo 675112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 676112202d9STejun Heo return ((struct pool_workqueue *) 67754d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 67854d5b7d0SLai Jiangshan 67954d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6807c3eed5cSTejun Heo } 6817c3eed5cSTejun Heo 682bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 683bbb68dfaSTejun Heo { 6847c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 685bbb68dfaSTejun Heo 6867c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6877c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 688bbb68dfaSTejun Heo } 689bbb68dfaSTejun Heo 690bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 691bbb68dfaSTejun Heo { 692bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 693bbb68dfaSTejun Heo 694112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 695bbb68dfaSTejun Heo } 696bbb68dfaSTejun Heo 697e22bee78STejun Heo /* 6983270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6993270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 700d565ed63STejun Heo * they're being called with pool->lock held. 701e22bee78STejun Heo */ 702e22bee78STejun Heo 70363d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 704649027d7STejun Heo { 705e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 706649027d7STejun Heo } 707649027d7STejun Heo 708e22bee78STejun Heo /* 709e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 710e22bee78STejun Heo * running workers. 711974271c4STejun Heo * 712974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 713706026c2STejun Heo * function will always return %true for unbound pools as long as the 714974271c4STejun Heo * worklist isn't empty. 715e22bee78STejun Heo */ 71663d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 717e22bee78STejun Heo { 71863d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 719e22bee78STejun Heo } 720e22bee78STejun Heo 721e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 72263d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 723e22bee78STejun Heo { 72463d95a91STejun Heo return pool->nr_idle; 725e22bee78STejun Heo } 726e22bee78STejun Heo 727e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 72863d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 729e22bee78STejun Heo { 730e19e397aSTejun Heo return !list_empty(&pool->worklist) && 731e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 732e22bee78STejun Heo } 733e22bee78STejun Heo 734e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 73563d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 736e22bee78STejun Heo { 73763d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 738e22bee78STejun Heo } 739e22bee78STejun Heo 740e22bee78STejun Heo /* Do I need to be the manager? */ 74163d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 742e22bee78STejun Heo { 74363d95a91STejun Heo return need_to_create_worker(pool) || 74411ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 745e22bee78STejun Heo } 746e22bee78STejun Heo 747e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 74863d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 749e22bee78STejun Heo { 75034a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 75163d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 75263d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 753e22bee78STejun Heo 754ea1abd61SLai Jiangshan /* 755ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 756ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 757ea1abd61SLai Jiangshan */ 758ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 759ea1abd61SLai Jiangshan return false; 760ea1abd61SLai Jiangshan 761e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 762e22bee78STejun Heo } 763e22bee78STejun Heo 764e22bee78STejun Heo /* 765e22bee78STejun Heo * Wake up functions. 766e22bee78STejun Heo */ 767e22bee78STejun Heo 7687e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 76963d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7707e11629dSTejun Heo { 77163d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7727e11629dSTejun Heo return NULL; 7737e11629dSTejun Heo 77463d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7757e11629dSTejun Heo } 7767e11629dSTejun Heo 7777e11629dSTejun Heo /** 7787e11629dSTejun Heo * wake_up_worker - wake up an idle worker 77963d95a91STejun Heo * @pool: worker pool to wake worker from 7807e11629dSTejun Heo * 78163d95a91STejun Heo * Wake up the first idle worker of @pool. 7827e11629dSTejun Heo * 7837e11629dSTejun Heo * CONTEXT: 784d565ed63STejun Heo * spin_lock_irq(pool->lock). 7857e11629dSTejun Heo */ 78663d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7877e11629dSTejun Heo { 78863d95a91STejun Heo struct worker *worker = first_worker(pool); 7897e11629dSTejun Heo 7907e11629dSTejun Heo if (likely(worker)) 7917e11629dSTejun Heo wake_up_process(worker->task); 7927e11629dSTejun Heo } 7937e11629dSTejun Heo 7944690c4abSTejun Heo /** 795e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 796e22bee78STejun Heo * @task: task waking up 797e22bee78STejun Heo * @cpu: CPU @task is waking up to 798e22bee78STejun Heo * 799e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 800e22bee78STejun Heo * being awoken. 801e22bee78STejun Heo * 802e22bee78STejun Heo * CONTEXT: 803e22bee78STejun Heo * spin_lock_irq(rq->lock) 804e22bee78STejun Heo */ 805d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 806e22bee78STejun Heo { 807e22bee78STejun Heo struct worker *worker = kthread_data(task); 808e22bee78STejun Heo 80936576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 810ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 811e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 812e22bee78STejun Heo } 81336576000SJoonsoo Kim } 814e22bee78STejun Heo 815e22bee78STejun Heo /** 816e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 817e22bee78STejun Heo * @task: task going to sleep 818e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 819e22bee78STejun Heo * 820e22bee78STejun Heo * This function is called during schedule() when a busy worker is 821e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 822e22bee78STejun Heo * returning pointer to its task. 823e22bee78STejun Heo * 824e22bee78STejun Heo * CONTEXT: 825e22bee78STejun Heo * spin_lock_irq(rq->lock) 826e22bee78STejun Heo * 827e22bee78STejun Heo * RETURNS: 828e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 829e22bee78STejun Heo */ 830d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 831e22bee78STejun Heo { 832e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 833111c225aSTejun Heo struct worker_pool *pool; 834e22bee78STejun Heo 835111c225aSTejun Heo /* 836111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 837111c225aSTejun Heo * workers, also reach here, let's not access anything before 838111c225aSTejun Heo * checking NOT_RUNNING. 839111c225aSTejun Heo */ 8402d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 841e22bee78STejun Heo return NULL; 842e22bee78STejun Heo 843111c225aSTejun Heo pool = worker->pool; 844111c225aSTejun Heo 845e22bee78STejun Heo /* this can only happen on the local cpu */ 8466183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8476183c009STejun Heo return NULL; 848e22bee78STejun Heo 849e22bee78STejun Heo /* 850e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 851e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 852e22bee78STejun Heo * Please read comment there. 853e22bee78STejun Heo * 854628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 855628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 856628c78e7STejun Heo * disabled, which in turn means that none else could be 857d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 858628c78e7STejun Heo * lock is safe. 859e22bee78STejun Heo */ 860e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 861e19e397aSTejun Heo !list_empty(&pool->worklist)) 86263d95a91STejun Heo to_wakeup = first_worker(pool); 863e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 864e22bee78STejun Heo } 865e22bee78STejun Heo 866e22bee78STejun Heo /** 867e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 868cb444766STejun Heo * @worker: self 869d302f017STejun Heo * @flags: flags to set 870d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 871d302f017STejun Heo * 872e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 873e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 874e22bee78STejun Heo * woken up. 875d302f017STejun Heo * 876cb444766STejun Heo * CONTEXT: 877d565ed63STejun Heo * spin_lock_irq(pool->lock) 878d302f017STejun Heo */ 879d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 880d302f017STejun Heo bool wakeup) 881d302f017STejun Heo { 882bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 883e22bee78STejun Heo 884cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 885cb444766STejun Heo 886e22bee78STejun Heo /* 887e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 888e22bee78STejun Heo * wake up an idle worker as necessary if requested by 889e22bee78STejun Heo * @wakeup. 890e22bee78STejun Heo */ 891e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 892e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 893e22bee78STejun Heo if (wakeup) { 894e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 895bd7bdd43STejun Heo !list_empty(&pool->worklist)) 89663d95a91STejun Heo wake_up_worker(pool); 897e22bee78STejun Heo } else 898e19e397aSTejun Heo atomic_dec(&pool->nr_running); 899e22bee78STejun Heo } 900e22bee78STejun Heo 901d302f017STejun Heo worker->flags |= flags; 902d302f017STejun Heo } 903d302f017STejun Heo 904d302f017STejun Heo /** 905e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 906cb444766STejun Heo * @worker: self 907d302f017STejun Heo * @flags: flags to clear 908d302f017STejun Heo * 909e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 910d302f017STejun Heo * 911cb444766STejun Heo * CONTEXT: 912d565ed63STejun Heo * spin_lock_irq(pool->lock) 913d302f017STejun Heo */ 914d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 915d302f017STejun Heo { 91663d95a91STejun Heo struct worker_pool *pool = worker->pool; 917e22bee78STejun Heo unsigned int oflags = worker->flags; 918e22bee78STejun Heo 919cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 920cb444766STejun Heo 921d302f017STejun Heo worker->flags &= ~flags; 922e22bee78STejun Heo 92342c025f3STejun Heo /* 92442c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 92542c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 92642c025f3STejun Heo * of multiple flags, not a single flag. 92742c025f3STejun Heo */ 928e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 929e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 930e19e397aSTejun Heo atomic_inc(&pool->nr_running); 931d302f017STejun Heo } 932d302f017STejun Heo 933d302f017STejun Heo /** 9348cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 935c9e7cf27STejun Heo * @pool: pool of interest 9368cca0eeaSTejun Heo * @work: work to find worker for 9378cca0eeaSTejun Heo * 938c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 939c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 940a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 941a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 942a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 943a2c1c57bSTejun Heo * being executed. 944a2c1c57bSTejun Heo * 945a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 946a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 947a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 948a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 949a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 950a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 951a2c1c57bSTejun Heo * 952c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 953c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 954c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 955c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 956c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 957c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9588cca0eeaSTejun Heo * 9598cca0eeaSTejun Heo * CONTEXT: 960d565ed63STejun Heo * spin_lock_irq(pool->lock). 9618cca0eeaSTejun Heo * 9628cca0eeaSTejun Heo * RETURNS: 9638cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9648cca0eeaSTejun Heo * otherwise. 9658cca0eeaSTejun Heo */ 966c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9678cca0eeaSTejun Heo struct work_struct *work) 9688cca0eeaSTejun Heo { 96942f8570fSSasha Levin struct worker *worker; 97042f8570fSSasha Levin 971b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 972a2c1c57bSTejun Heo (unsigned long)work) 973a2c1c57bSTejun Heo if (worker->current_work == work && 974a2c1c57bSTejun Heo worker->current_func == work->func) 97542f8570fSSasha Levin return worker; 97642f8570fSSasha Levin 97742f8570fSSasha Levin return NULL; 9788cca0eeaSTejun Heo } 9798cca0eeaSTejun Heo 9808cca0eeaSTejun Heo /** 981bf4ede01STejun Heo * move_linked_works - move linked works to a list 982bf4ede01STejun Heo * @work: start of series of works to be scheduled 983bf4ede01STejun Heo * @head: target list to append @work to 984bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 985bf4ede01STejun Heo * 986bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 987bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 988bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 989bf4ede01STejun Heo * 990bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 991bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 992bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 993bf4ede01STejun Heo * 994bf4ede01STejun Heo * CONTEXT: 995d565ed63STejun Heo * spin_lock_irq(pool->lock). 996bf4ede01STejun Heo */ 997bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 998bf4ede01STejun Heo struct work_struct **nextp) 999bf4ede01STejun Heo { 1000bf4ede01STejun Heo struct work_struct *n; 1001bf4ede01STejun Heo 1002bf4ede01STejun Heo /* 1003bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1004bf4ede01STejun Heo * use NULL for list head. 1005bf4ede01STejun Heo */ 1006bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1007bf4ede01STejun Heo list_move_tail(&work->entry, head); 1008bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1009bf4ede01STejun Heo break; 1010bf4ede01STejun Heo } 1011bf4ede01STejun Heo 1012bf4ede01STejun Heo /* 1013bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1014bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1015bf4ede01STejun Heo * needs to be updated. 1016bf4ede01STejun Heo */ 1017bf4ede01STejun Heo if (nextp) 1018bf4ede01STejun Heo *nextp = n; 1019bf4ede01STejun Heo } 1020bf4ede01STejun Heo 10218864b4e5STejun Heo /** 10228864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10238864b4e5STejun Heo * @pwq: pool_workqueue to get 10248864b4e5STejun Heo * 10258864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10268864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10278864b4e5STejun Heo */ 10288864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10298864b4e5STejun Heo { 10308864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10318864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10328864b4e5STejun Heo pwq->refcnt++; 10338864b4e5STejun Heo } 10348864b4e5STejun Heo 10358864b4e5STejun Heo /** 10368864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10378864b4e5STejun Heo * @pwq: pool_workqueue to put 10388864b4e5STejun Heo * 10398864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10408864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10418864b4e5STejun Heo */ 10428864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10438864b4e5STejun Heo { 10448864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10458864b4e5STejun Heo if (likely(--pwq->refcnt)) 10468864b4e5STejun Heo return; 10478864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10488864b4e5STejun Heo return; 10498864b4e5STejun Heo /* 10508864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10518864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10528864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10538864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10548864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10558864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10568864b4e5STejun Heo */ 10578864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10588864b4e5STejun Heo } 10598864b4e5STejun Heo 1060112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1061bf4ede01STejun Heo { 1062112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1063bf4ede01STejun Heo 1064bf4ede01STejun Heo trace_workqueue_activate_work(work); 1065112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1066bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1067112202d9STejun Heo pwq->nr_active++; 1068bf4ede01STejun Heo } 1069bf4ede01STejun Heo 1070112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10713aa62497SLai Jiangshan { 1072112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10733aa62497SLai Jiangshan struct work_struct, entry); 10743aa62497SLai Jiangshan 1075112202d9STejun Heo pwq_activate_delayed_work(work); 10763aa62497SLai Jiangshan } 10773aa62497SLai Jiangshan 1078bf4ede01STejun Heo /** 1079112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1080112202d9STejun Heo * @pwq: pwq of interest 1081bf4ede01STejun Heo * @color: color of work which left the queue 1082bf4ede01STejun Heo * 1083bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1084112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1085bf4ede01STejun Heo * 1086bf4ede01STejun Heo * CONTEXT: 1087d565ed63STejun Heo * spin_lock_irq(pool->lock). 1088bf4ede01STejun Heo */ 1089112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1090bf4ede01STejun Heo { 10918864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1092bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10938864b4e5STejun Heo goto out_put; 1094bf4ede01STejun Heo 1095112202d9STejun Heo pwq->nr_in_flight[color]--; 1096bf4ede01STejun Heo 1097112202d9STejun Heo pwq->nr_active--; 1098112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1099bf4ede01STejun Heo /* one down, submit a delayed one */ 1100112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1101112202d9STejun Heo pwq_activate_first_delayed(pwq); 1102bf4ede01STejun Heo } 1103bf4ede01STejun Heo 1104bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1105112202d9STejun Heo if (likely(pwq->flush_color != color)) 11068864b4e5STejun Heo goto out_put; 1107bf4ede01STejun Heo 1108bf4ede01STejun Heo /* are there still in-flight works? */ 1109112202d9STejun Heo if (pwq->nr_in_flight[color]) 11108864b4e5STejun Heo goto out_put; 1111bf4ede01STejun Heo 1112112202d9STejun Heo /* this pwq is done, clear flush_color */ 1113112202d9STejun Heo pwq->flush_color = -1; 1114bf4ede01STejun Heo 1115bf4ede01STejun Heo /* 1116112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1117bf4ede01STejun Heo * will handle the rest. 1118bf4ede01STejun Heo */ 1119112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1120112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11218864b4e5STejun Heo out_put: 11228864b4e5STejun Heo put_pwq(pwq); 1123bf4ede01STejun Heo } 1124bf4ede01STejun Heo 112536e227d2STejun Heo /** 1126bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 112736e227d2STejun Heo * @work: work item to steal 112836e227d2STejun Heo * @is_dwork: @work is a delayed_work 1129bbb68dfaSTejun Heo * @flags: place to store irq state 113036e227d2STejun Heo * 113136e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 113236e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 113336e227d2STejun Heo * 113436e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 113536e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 113636e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1137bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1138bbb68dfaSTejun Heo * for arbitrarily long 113936e227d2STejun Heo * 1140bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1141e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1142e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1143e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1144bbb68dfaSTejun Heo * 1145bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1146bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1147bbb68dfaSTejun Heo * 1148e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1149bf4ede01STejun Heo */ 1150bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1151bbb68dfaSTejun Heo unsigned long *flags) 1152bf4ede01STejun Heo { 1153d565ed63STejun Heo struct worker_pool *pool; 1154112202d9STejun Heo struct pool_workqueue *pwq; 1155bf4ede01STejun Heo 1156bbb68dfaSTejun Heo local_irq_save(*flags); 1157bbb68dfaSTejun Heo 115836e227d2STejun Heo /* try to steal the timer if it exists */ 115936e227d2STejun Heo if (is_dwork) { 116036e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 116136e227d2STejun Heo 1162e0aecdd8STejun Heo /* 1163e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1164e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1165e0aecdd8STejun Heo * running on the local CPU. 1166e0aecdd8STejun Heo */ 116736e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 116836e227d2STejun Heo return 1; 116936e227d2STejun Heo } 117036e227d2STejun Heo 117136e227d2STejun Heo /* try to claim PENDING the normal way */ 1172bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1173bf4ede01STejun Heo return 0; 1174bf4ede01STejun Heo 1175bf4ede01STejun Heo /* 1176bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1177bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1178bf4ede01STejun Heo */ 1179d565ed63STejun Heo pool = get_work_pool(work); 1180d565ed63STejun Heo if (!pool) 1181bbb68dfaSTejun Heo goto fail; 1182bf4ede01STejun Heo 1183d565ed63STejun Heo spin_lock(&pool->lock); 1184bf4ede01STejun Heo /* 1185112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1186112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1187112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1188112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1189112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11900b3dae68SLai Jiangshan * item is currently queued on that pool. 1191bf4ede01STejun Heo */ 1192112202d9STejun Heo pwq = get_work_pwq(work); 1193112202d9STejun Heo if (pwq && pwq->pool == pool) { 1194bf4ede01STejun Heo debug_work_deactivate(work); 11953aa62497SLai Jiangshan 11963aa62497SLai Jiangshan /* 119716062836STejun Heo * A delayed work item cannot be grabbed directly because 119816062836STejun Heo * it might have linked NO_COLOR work items which, if left 1199112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 120016062836STejun Heo * management later on and cause stall. Make sure the work 120116062836STejun Heo * item is activated before grabbing. 12023aa62497SLai Jiangshan */ 12033aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1204112202d9STejun Heo pwq_activate_delayed_work(work); 12053aa62497SLai Jiangshan 1206bf4ede01STejun Heo list_del_init(&work->entry); 1207112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 120836e227d2STejun Heo 1209112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12104468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12114468a00fSLai Jiangshan 1212d565ed63STejun Heo spin_unlock(&pool->lock); 121336e227d2STejun Heo return 1; 1214bf4ede01STejun Heo } 1215d565ed63STejun Heo spin_unlock(&pool->lock); 1216bbb68dfaSTejun Heo fail: 1217bbb68dfaSTejun Heo local_irq_restore(*flags); 1218bbb68dfaSTejun Heo if (work_is_canceling(work)) 1219bbb68dfaSTejun Heo return -ENOENT; 1220bbb68dfaSTejun Heo cpu_relax(); 122136e227d2STejun Heo return -EAGAIN; 1222bf4ede01STejun Heo } 1223bf4ede01STejun Heo 1224bf4ede01STejun Heo /** 1225706026c2STejun Heo * insert_work - insert a work into a pool 1226112202d9STejun Heo * @pwq: pwq @work belongs to 12274690c4abSTejun Heo * @work: work to insert 12284690c4abSTejun Heo * @head: insertion point 12294690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12304690c4abSTejun Heo * 1231112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1232706026c2STejun Heo * work_struct flags. 12334690c4abSTejun Heo * 12344690c4abSTejun Heo * CONTEXT: 1235d565ed63STejun Heo * spin_lock_irq(pool->lock). 1236365970a1SDavid Howells */ 1237112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1238112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1239b89deed3SOleg Nesterov { 1240112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1241e1d8aa9fSFrederic Weisbecker 12424690c4abSTejun Heo /* we own @work, set data and link */ 1243112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12441a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12458864b4e5STejun Heo get_pwq(pwq); 1246e22bee78STejun Heo 1247e22bee78STejun Heo /* 1248c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1249c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1250c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1251e22bee78STejun Heo */ 1252e22bee78STejun Heo smp_mb(); 1253e22bee78STejun Heo 125463d95a91STejun Heo if (__need_more_worker(pool)) 125563d95a91STejun Heo wake_up_worker(pool); 1256b89deed3SOleg Nesterov } 1257b89deed3SOleg Nesterov 1258c8efcc25STejun Heo /* 1259c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12608d03ecfeSTejun Heo * same workqueue. 1261c8efcc25STejun Heo */ 1262c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1263c8efcc25STejun Heo { 1264c8efcc25STejun Heo struct worker *worker; 1265c8efcc25STejun Heo 12668d03ecfeSTejun Heo worker = current_wq_worker(); 1267c8efcc25STejun Heo /* 12688d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12698d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1270c8efcc25STejun Heo */ 1271112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1272c8efcc25STejun Heo } 1273c8efcc25STejun Heo 1274d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12751da177e4SLinus Torvalds struct work_struct *work) 12761da177e4SLinus Torvalds { 1277112202d9STejun Heo struct pool_workqueue *pwq; 1278c9178087STejun Heo struct worker_pool *last_pool; 12791e19ffc6STejun Heo struct list_head *worklist; 12808a2e8e5dSTejun Heo unsigned int work_flags; 1281b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12828930cabaSTejun Heo 12838930cabaSTejun Heo /* 12848930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12858930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12868930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12878930cabaSTejun Heo * happen with IRQ disabled. 12888930cabaSTejun Heo */ 12898930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12901da177e4SLinus Torvalds 1291dc186ad7SThomas Gleixner debug_work_activate(work); 12921e19ffc6STejun Heo 1293c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1294618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1295c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1296e41e704bSTejun Heo return; 12979e8cd2f5STejun Heo retry: 1298df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1299f3421797STejun Heo cpu = raw_smp_processor_id(); 1300df2d5ae4STejun Heo 1301df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1302df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1303c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1304df2d5ae4STejun Heo else 1305df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1306f3421797STejun Heo 130718aa9effSTejun Heo /* 1308c9178087STejun Heo * If @work was previously on a different pool, it might still be 1309c9178087STejun Heo * running there, in which case the work needs to be queued on that 1310c9178087STejun Heo * pool to guarantee non-reentrancy. 131118aa9effSTejun Heo */ 1312c9e7cf27STejun Heo last_pool = get_work_pool(work); 1313112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 131418aa9effSTejun Heo struct worker *worker; 131518aa9effSTejun Heo 1316d565ed63STejun Heo spin_lock(&last_pool->lock); 131718aa9effSTejun Heo 1318c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 131918aa9effSTejun Heo 1320112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1321c9178087STejun Heo pwq = worker->current_pwq; 13228594fadeSLai Jiangshan } else { 132318aa9effSTejun Heo /* meh... not running there, queue here */ 1324d565ed63STejun Heo spin_unlock(&last_pool->lock); 1325112202d9STejun Heo spin_lock(&pwq->pool->lock); 132618aa9effSTejun Heo } 13278930cabaSTejun Heo } else { 1328112202d9STejun Heo spin_lock(&pwq->pool->lock); 13298930cabaSTejun Heo } 1330502ca9d8STejun Heo 13319e8cd2f5STejun Heo /* 13329e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13339e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13349e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1335df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1336df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13379e8cd2f5STejun Heo * make forward-progress. 13389e8cd2f5STejun Heo */ 13399e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13409e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13419e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13429e8cd2f5STejun Heo cpu_relax(); 13439e8cd2f5STejun Heo goto retry; 13449e8cd2f5STejun Heo } 13459e8cd2f5STejun Heo /* oops */ 13469e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13479e8cd2f5STejun Heo wq->name, cpu); 13489e8cd2f5STejun Heo } 13499e8cd2f5STejun Heo 1350112202d9STejun Heo /* pwq determined, queue */ 1351112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1352502ca9d8STejun Heo 1353f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1354112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1355f5b2552bSDan Carpenter return; 1356f5b2552bSDan Carpenter } 13571e19ffc6STejun Heo 1358112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1359112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13601e19ffc6STejun Heo 1361112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1362cdadf009STejun Heo trace_workqueue_activate_work(work); 1363112202d9STejun Heo pwq->nr_active++; 1364112202d9STejun Heo worklist = &pwq->pool->worklist; 13658a2e8e5dSTejun Heo } else { 13668a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1367112202d9STejun Heo worklist = &pwq->delayed_works; 13688a2e8e5dSTejun Heo } 13691e19ffc6STejun Heo 1370112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13711e19ffc6STejun Heo 1372112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13731da177e4SLinus Torvalds } 13741da177e4SLinus Torvalds 13750fcb78c2SRolf Eike Beer /** 1376c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1377c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1378c1a220e7SZhang Rui * @wq: workqueue to use 1379c1a220e7SZhang Rui * @work: work to queue 1380c1a220e7SZhang Rui * 1381d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1382c1a220e7SZhang Rui * 1383c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1384c1a220e7SZhang Rui * can't go away. 1385c1a220e7SZhang Rui */ 1386d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1387d4283e93STejun Heo struct work_struct *work) 1388c1a220e7SZhang Rui { 1389d4283e93STejun Heo bool ret = false; 13908930cabaSTejun Heo unsigned long flags; 13918930cabaSTejun Heo 13928930cabaSTejun Heo local_irq_save(flags); 1393c1a220e7SZhang Rui 139422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13954690c4abSTejun Heo __queue_work(cpu, wq, work); 1396d4283e93STejun Heo ret = true; 1397c1a220e7SZhang Rui } 13988930cabaSTejun Heo 13998930cabaSTejun Heo local_irq_restore(flags); 1400c1a220e7SZhang Rui return ret; 1401c1a220e7SZhang Rui } 1402c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1403c1a220e7SZhang Rui 1404d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14051da177e4SLinus Torvalds { 140652bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14071da177e4SLinus Torvalds 1408e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 140960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14101da177e4SLinus Torvalds } 14111438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14121da177e4SLinus Torvalds 14137beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 141452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14151da177e4SLinus Torvalds { 14167beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14177beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14181da177e4SLinus Torvalds 14197beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14207beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1421fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1422fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14237beb2edfSTejun Heo 14248852aac2STejun Heo /* 14258852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14268852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14278852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14288852aac2STejun Heo * on that there's no such delay when @delay is 0. 14298852aac2STejun Heo */ 14308852aac2STejun Heo if (!delay) { 14318852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14328852aac2STejun Heo return; 14338852aac2STejun Heo } 14348852aac2STejun Heo 14357beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14367beb2edfSTejun Heo 143760c057bcSLai Jiangshan dwork->wq = wq; 14381265057fSTejun Heo dwork->cpu = cpu; 14397beb2edfSTejun Heo timer->expires = jiffies + delay; 14407beb2edfSTejun Heo 14417beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14427beb2edfSTejun Heo add_timer_on(timer, cpu); 14437beb2edfSTejun Heo else 14447beb2edfSTejun Heo add_timer(timer); 14457beb2edfSTejun Heo } 14461da177e4SLinus Torvalds 14470fcb78c2SRolf Eike Beer /** 14480fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14490fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14500fcb78c2SRolf Eike Beer * @wq: workqueue to use 1451af9997e4SRandy Dunlap * @dwork: work to queue 14520fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14530fcb78c2SRolf Eike Beer * 1454715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1455715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1456715f1300STejun Heo * execution. 14570fcb78c2SRolf Eike Beer */ 1458d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 145952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14607a6bc1cdSVenkatesh Pallipadi { 146152bad64dSDavid Howells struct work_struct *work = &dwork->work; 1462d4283e93STejun Heo bool ret = false; 14638930cabaSTejun Heo unsigned long flags; 14648930cabaSTejun Heo 14658930cabaSTejun Heo /* read the comment in __queue_work() */ 14668930cabaSTejun Heo local_irq_save(flags); 14677a6bc1cdSVenkatesh Pallipadi 146822df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14697beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1470d4283e93STejun Heo ret = true; 14717a6bc1cdSVenkatesh Pallipadi } 14728930cabaSTejun Heo 14738930cabaSTejun Heo local_irq_restore(flags); 14747a6bc1cdSVenkatesh Pallipadi return ret; 14757a6bc1cdSVenkatesh Pallipadi } 1476ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14771da177e4SLinus Torvalds 1478c8e55f36STejun Heo /** 14798376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14808376fe22STejun Heo * @cpu: CPU number to execute work on 14818376fe22STejun Heo * @wq: workqueue to use 14828376fe22STejun Heo * @dwork: work to queue 14838376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14848376fe22STejun Heo * 14858376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14868376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14878376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14888376fe22STejun Heo * current state. 14898376fe22STejun Heo * 14908376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14918376fe22STejun Heo * pending and its timer was modified. 14928376fe22STejun Heo * 1493e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14948376fe22STejun Heo * See try_to_grab_pending() for details. 14958376fe22STejun Heo */ 14968376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14978376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14988376fe22STejun Heo { 14998376fe22STejun Heo unsigned long flags; 15008376fe22STejun Heo int ret; 15018376fe22STejun Heo 15028376fe22STejun Heo do { 15038376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15048376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15058376fe22STejun Heo 15068376fe22STejun Heo if (likely(ret >= 0)) { 15078376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15088376fe22STejun Heo local_irq_restore(flags); 15098376fe22STejun Heo } 15108376fe22STejun Heo 15118376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15128376fe22STejun Heo return ret; 15138376fe22STejun Heo } 15148376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15158376fe22STejun Heo 15168376fe22STejun Heo /** 1517c8e55f36STejun Heo * worker_enter_idle - enter idle state 1518c8e55f36STejun Heo * @worker: worker which is entering idle state 1519c8e55f36STejun Heo * 1520c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1521c8e55f36STejun Heo * necessary. 1522c8e55f36STejun Heo * 1523c8e55f36STejun Heo * LOCKING: 1524d565ed63STejun Heo * spin_lock_irq(pool->lock). 1525c8e55f36STejun Heo */ 1526c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15271da177e4SLinus Torvalds { 1528bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1529c8e55f36STejun Heo 15306183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15316183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15326183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15336183c009STejun Heo return; 1534c8e55f36STejun Heo 1535cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1536cb444766STejun Heo worker->flags |= WORKER_IDLE; 1537bd7bdd43STejun Heo pool->nr_idle++; 1538e22bee78STejun Heo worker->last_active = jiffies; 1539c8e55f36STejun Heo 1540c8e55f36STejun Heo /* idle_list is LIFO */ 1541bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1542db7bccf4STejun Heo 154363d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1544628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1545cb444766STejun Heo 1546544ecf31STejun Heo /* 1547706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1548d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1549628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1550628c78e7STejun Heo * unbind is not in progress. 1551544ecf31STejun Heo */ 155224647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1553bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1554e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1555c8e55f36STejun Heo } 1556c8e55f36STejun Heo 1557c8e55f36STejun Heo /** 1558c8e55f36STejun Heo * worker_leave_idle - leave idle state 1559c8e55f36STejun Heo * @worker: worker which is leaving idle state 1560c8e55f36STejun Heo * 1561c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1562c8e55f36STejun Heo * 1563c8e55f36STejun Heo * LOCKING: 1564d565ed63STejun Heo * spin_lock_irq(pool->lock). 1565c8e55f36STejun Heo */ 1566c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1567c8e55f36STejun Heo { 1568bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1569c8e55f36STejun Heo 15706183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15716183c009STejun Heo return; 1572d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1573bd7bdd43STejun Heo pool->nr_idle--; 1574c8e55f36STejun Heo list_del_init(&worker->entry); 1575c8e55f36STejun Heo } 1576c8e55f36STejun Heo 1577e22bee78STejun Heo /** 1578f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1579f36dc67bSLai Jiangshan * @pool: target worker_pool 1580f36dc67bSLai Jiangshan * 1581f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1582e22bee78STejun Heo * 1583e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1584e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1585e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1586e22bee78STejun Heo * guaranteed to execute on the cpu. 1587e22bee78STejun Heo * 1588f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1589e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1590e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1591e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1592706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1593e22bee78STejun Heo * [dis]associated in the meantime. 1594e22bee78STejun Heo * 1595706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 159624647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1597f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1598f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1599f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1600e22bee78STejun Heo * 1601e22bee78STejun Heo * CONTEXT: 1602d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1603e22bee78STejun Heo * held. 1604e22bee78STejun Heo * 1605e22bee78STejun Heo * RETURNS: 1606706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1607e22bee78STejun Heo * bound), %false if offline. 1608e22bee78STejun Heo */ 1609f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1610d565ed63STejun Heo __acquires(&pool->lock) 1611e22bee78STejun Heo { 1612e22bee78STejun Heo while (true) { 1613e22bee78STejun Heo /* 1614e22bee78STejun Heo * The following call may fail, succeed or succeed 1615e22bee78STejun Heo * without actually migrating the task to the cpu if 1616e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 161724647570STejun Heo * against POOL_DISASSOCIATED. 1618e22bee78STejun Heo */ 161924647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16207a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1621e22bee78STejun Heo 1622d565ed63STejun Heo spin_lock_irq(&pool->lock); 162324647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1624e22bee78STejun Heo return false; 1625f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16267a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1627e22bee78STejun Heo return true; 1628d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1629e22bee78STejun Heo 16305035b20fSTejun Heo /* 16315035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16325035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16335035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16345035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16355035b20fSTejun Heo */ 1636e22bee78STejun Heo cpu_relax(); 16375035b20fSTejun Heo cond_resched(); 1638e22bee78STejun Heo } 1639e22bee78STejun Heo } 1640e22bee78STejun Heo 1641c34056a3STejun Heo static struct worker *alloc_worker(void) 1642c34056a3STejun Heo { 1643c34056a3STejun Heo struct worker *worker; 1644c34056a3STejun Heo 1645c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1646c8e55f36STejun Heo if (worker) { 1647c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1648affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1649e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1650e22bee78STejun Heo worker->flags = WORKER_PREP; 1651c8e55f36STejun Heo } 1652c34056a3STejun Heo return worker; 1653c34056a3STejun Heo } 1654c34056a3STejun Heo 1655c34056a3STejun Heo /** 1656c34056a3STejun Heo * create_worker - create a new workqueue worker 165763d95a91STejun Heo * @pool: pool the new worker will belong to 1658c34056a3STejun Heo * 165963d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1660c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1661c34056a3STejun Heo * destroy_worker(). 1662c34056a3STejun Heo * 1663c34056a3STejun Heo * CONTEXT: 1664c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1665c34056a3STejun Heo * 1666c34056a3STejun Heo * RETURNS: 1667c34056a3STejun Heo * Pointer to the newly created worker. 1668c34056a3STejun Heo */ 1669bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1670c34056a3STejun Heo { 1671c34056a3STejun Heo struct worker *worker = NULL; 1672f3421797STejun Heo int id = -1; 1673e3c916a4STejun Heo char id_buf[16]; 1674c34056a3STejun Heo 1675cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1676cd549687STejun Heo 1677822d8405STejun Heo /* 1678822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1679822d8405STejun Heo * without installing the pointer. 1680822d8405STejun Heo */ 1681822d8405STejun Heo idr_preload(GFP_KERNEL); 1682d565ed63STejun Heo spin_lock_irq(&pool->lock); 1683822d8405STejun Heo 1684822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1685822d8405STejun Heo 1686d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1687822d8405STejun Heo idr_preload_end(); 1688822d8405STejun Heo if (id < 0) 1689c34056a3STejun Heo goto fail; 1690c34056a3STejun Heo 1691c34056a3STejun Heo worker = alloc_worker(); 1692c34056a3STejun Heo if (!worker) 1693c34056a3STejun Heo goto fail; 1694c34056a3STejun Heo 1695bd7bdd43STejun Heo worker->pool = pool; 1696c34056a3STejun Heo worker->id = id; 1697c34056a3STejun Heo 169829c91e99STejun Heo if (pool->cpu >= 0) 1699e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1700e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1701f3421797STejun Heo else 1702e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1703e3c916a4STejun Heo 1704f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1705e3c916a4STejun Heo "kworker/%s", id_buf); 1706c34056a3STejun Heo if (IS_ERR(worker->task)) 1707c34056a3STejun Heo goto fail; 1708c34056a3STejun Heo 1709c5aa87bbSTejun Heo /* 1710c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1711c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1712c5aa87bbSTejun Heo */ 17137a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 17147a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17153270476aSTejun Heo 171614a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 171714a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 17187a4e344cSTejun Heo 17197a4e344cSTejun Heo /* 17207a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17217a4e344cSTejun Heo * remains stable across this function. See the comments above the 17227a4e344cSTejun Heo * flag definition for details. 17237a4e344cSTejun Heo */ 17247a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1725f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1726c34056a3STejun Heo 1727822d8405STejun Heo /* successful, commit the pointer to idr */ 1728822d8405STejun Heo spin_lock_irq(&pool->lock); 1729822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1730822d8405STejun Heo spin_unlock_irq(&pool->lock); 1731822d8405STejun Heo 1732c34056a3STejun Heo return worker; 1733822d8405STejun Heo 1734c34056a3STejun Heo fail: 1735c34056a3STejun Heo if (id >= 0) { 1736d565ed63STejun Heo spin_lock_irq(&pool->lock); 1737822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1738d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1739c34056a3STejun Heo } 1740c34056a3STejun Heo kfree(worker); 1741c34056a3STejun Heo return NULL; 1742c34056a3STejun Heo } 1743c34056a3STejun Heo 1744c34056a3STejun Heo /** 1745c34056a3STejun Heo * start_worker - start a newly created worker 1746c34056a3STejun Heo * @worker: worker to start 1747c34056a3STejun Heo * 1748706026c2STejun Heo * Make the pool aware of @worker and start it. 1749c34056a3STejun Heo * 1750c34056a3STejun Heo * CONTEXT: 1751d565ed63STejun Heo * spin_lock_irq(pool->lock). 1752c34056a3STejun Heo */ 1753c34056a3STejun Heo static void start_worker(struct worker *worker) 1754c34056a3STejun Heo { 1755cb444766STejun Heo worker->flags |= WORKER_STARTED; 1756bd7bdd43STejun Heo worker->pool->nr_workers++; 1757c8e55f36STejun Heo worker_enter_idle(worker); 1758c34056a3STejun Heo wake_up_process(worker->task); 1759c34056a3STejun Heo } 1760c34056a3STejun Heo 1761c34056a3STejun Heo /** 1762ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1763ebf44d16STejun Heo * @pool: the target pool 1764ebf44d16STejun Heo * 1765cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1766ebf44d16STejun Heo */ 1767ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1768ebf44d16STejun Heo { 1769ebf44d16STejun Heo struct worker *worker; 1770ebf44d16STejun Heo 1771cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1772cd549687STejun Heo 1773ebf44d16STejun Heo worker = create_worker(pool); 1774ebf44d16STejun Heo if (worker) { 1775ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1776ebf44d16STejun Heo start_worker(worker); 1777ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1778ebf44d16STejun Heo } 1779ebf44d16STejun Heo 1780cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1781cd549687STejun Heo 1782ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1783ebf44d16STejun Heo } 1784ebf44d16STejun Heo 1785ebf44d16STejun Heo /** 1786c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1787c34056a3STejun Heo * @worker: worker to be destroyed 1788c34056a3STejun Heo * 1789706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1790c8e55f36STejun Heo * 1791c8e55f36STejun Heo * CONTEXT: 1792d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1793c34056a3STejun Heo */ 1794c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1795c34056a3STejun Heo { 1796bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1797c34056a3STejun Heo 1798cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1799cd549687STejun Heo lockdep_assert_held(&pool->lock); 1800cd549687STejun Heo 1801c34056a3STejun Heo /* sanity check frenzy */ 18026183c009STejun Heo if (WARN_ON(worker->current_work) || 18036183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18046183c009STejun Heo return; 1805c34056a3STejun Heo 1806c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1807bd7bdd43STejun Heo pool->nr_workers--; 1808c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1809bd7bdd43STejun Heo pool->nr_idle--; 1810c8e55f36STejun Heo 1811c8e55f36STejun Heo list_del_init(&worker->entry); 1812cb444766STejun Heo worker->flags |= WORKER_DIE; 1813c8e55f36STejun Heo 1814822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1815822d8405STejun Heo 1816d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1817c8e55f36STejun Heo 1818c34056a3STejun Heo kthread_stop(worker->task); 1819c34056a3STejun Heo kfree(worker); 1820c34056a3STejun Heo 1821d565ed63STejun Heo spin_lock_irq(&pool->lock); 1822c34056a3STejun Heo } 1823c34056a3STejun Heo 182463d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1825e22bee78STejun Heo { 182663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1827e22bee78STejun Heo 1828d565ed63STejun Heo spin_lock_irq(&pool->lock); 1829e22bee78STejun Heo 183063d95a91STejun Heo if (too_many_workers(pool)) { 1831e22bee78STejun Heo struct worker *worker; 1832e22bee78STejun Heo unsigned long expires; 1833e22bee78STejun Heo 1834e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 183563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1836e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1837e22bee78STejun Heo 1838e22bee78STejun Heo if (time_before(jiffies, expires)) 183963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1840e22bee78STejun Heo else { 1841e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 184211ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 184363d95a91STejun Heo wake_up_worker(pool); 1844e22bee78STejun Heo } 1845e22bee78STejun Heo } 1846e22bee78STejun Heo 1847d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1848e22bee78STejun Heo } 1849e22bee78STejun Heo 1850493a1724STejun Heo static void send_mayday(struct work_struct *work) 1851e22bee78STejun Heo { 1852112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1853112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1854493a1724STejun Heo 18552e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1856e22bee78STejun Heo 1857493008a8STejun Heo if (!wq->rescuer) 1858493a1724STejun Heo return; 1859e22bee78STejun Heo 1860e22bee78STejun Heo /* mayday mayday mayday */ 1861493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1862493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1863e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1864493a1724STejun Heo } 1865e22bee78STejun Heo } 1866e22bee78STejun Heo 1867706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1868e22bee78STejun Heo { 186963d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1870e22bee78STejun Heo struct work_struct *work; 1871e22bee78STejun Heo 18722e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1873493a1724STejun Heo spin_lock(&pool->lock); 1874e22bee78STejun Heo 187563d95a91STejun Heo if (need_to_create_worker(pool)) { 1876e22bee78STejun Heo /* 1877e22bee78STejun Heo * We've been trying to create a new worker but 1878e22bee78STejun Heo * haven't been successful. We might be hitting an 1879e22bee78STejun Heo * allocation deadlock. Send distress signals to 1880e22bee78STejun Heo * rescuers. 1881e22bee78STejun Heo */ 188263d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1883e22bee78STejun Heo send_mayday(work); 1884e22bee78STejun Heo } 1885e22bee78STejun Heo 1886493a1724STejun Heo spin_unlock(&pool->lock); 18872e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1888e22bee78STejun Heo 188963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892e22bee78STejun Heo /** 1893e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 189463d95a91STejun Heo * @pool: pool to create a new worker for 1895e22bee78STejun Heo * 189663d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1897e22bee78STejun Heo * have at least one idle worker on return from this function. If 1898e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 189963d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1900e22bee78STejun Heo * possible allocation deadlock. 1901e22bee78STejun Heo * 1902c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1903c5aa87bbSTejun Heo * may_start_working() %true. 1904e22bee78STejun Heo * 1905e22bee78STejun Heo * LOCKING: 1906d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1907e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1908e22bee78STejun Heo * manager. 1909e22bee78STejun Heo * 1910e22bee78STejun Heo * RETURNS: 1911c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1912e22bee78STejun Heo * otherwise. 1913e22bee78STejun Heo */ 191463d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1915d565ed63STejun Heo __releases(&pool->lock) 1916d565ed63STejun Heo __acquires(&pool->lock) 1917e22bee78STejun Heo { 191863d95a91STejun Heo if (!need_to_create_worker(pool)) 1919e22bee78STejun Heo return false; 1920e22bee78STejun Heo restart: 1921d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19229f9c2364STejun Heo 1923e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 192463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1925e22bee78STejun Heo 1926e22bee78STejun Heo while (true) { 1927e22bee78STejun Heo struct worker *worker; 1928e22bee78STejun Heo 1929bc2ae0f5STejun Heo worker = create_worker(pool); 1930e22bee78STejun Heo if (worker) { 193163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1932d565ed63STejun Heo spin_lock_irq(&pool->lock); 1933e22bee78STejun Heo start_worker(worker); 19346183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19356183c009STejun Heo goto restart; 1936e22bee78STejun Heo return true; 1937e22bee78STejun Heo } 1938e22bee78STejun Heo 193963d95a91STejun Heo if (!need_to_create_worker(pool)) 1940e22bee78STejun Heo break; 1941e22bee78STejun Heo 1942e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1943e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19449f9c2364STejun Heo 194563d95a91STejun Heo if (!need_to_create_worker(pool)) 1946e22bee78STejun Heo break; 1947e22bee78STejun Heo } 1948e22bee78STejun Heo 194963d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1950d565ed63STejun Heo spin_lock_irq(&pool->lock); 195163d95a91STejun Heo if (need_to_create_worker(pool)) 1952e22bee78STejun Heo goto restart; 1953e22bee78STejun Heo return true; 1954e22bee78STejun Heo } 1955e22bee78STejun Heo 1956e22bee78STejun Heo /** 1957e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 195863d95a91STejun Heo * @pool: pool to destroy workers for 1959e22bee78STejun Heo * 196063d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1961e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1962e22bee78STejun Heo * 1963e22bee78STejun Heo * LOCKING: 1964d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1965e22bee78STejun Heo * multiple times. Called only from manager. 1966e22bee78STejun Heo * 1967e22bee78STejun Heo * RETURNS: 1968c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1969e22bee78STejun Heo * otherwise. 1970e22bee78STejun Heo */ 197163d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1972e22bee78STejun Heo { 1973e22bee78STejun Heo bool ret = false; 1974e22bee78STejun Heo 197563d95a91STejun Heo while (too_many_workers(pool)) { 1976e22bee78STejun Heo struct worker *worker; 1977e22bee78STejun Heo unsigned long expires; 1978e22bee78STejun Heo 197963d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1980e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1981e22bee78STejun Heo 1982e22bee78STejun Heo if (time_before(jiffies, expires)) { 198363d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1984e22bee78STejun Heo break; 1985e22bee78STejun Heo } 1986e22bee78STejun Heo 1987e22bee78STejun Heo destroy_worker(worker); 1988e22bee78STejun Heo ret = true; 1989e22bee78STejun Heo } 1990e22bee78STejun Heo 1991e22bee78STejun Heo return ret; 1992e22bee78STejun Heo } 1993e22bee78STejun Heo 1994e22bee78STejun Heo /** 1995e22bee78STejun Heo * manage_workers - manage worker pool 1996e22bee78STejun Heo * @worker: self 1997e22bee78STejun Heo * 1998706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1999e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2000706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2001e22bee78STejun Heo * 2002e22bee78STejun Heo * The caller can safely start processing works on false return. On 2003e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2004e22bee78STejun Heo * and may_start_working() is true. 2005e22bee78STejun Heo * 2006e22bee78STejun Heo * CONTEXT: 2007d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2008e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2009e22bee78STejun Heo * 2010e22bee78STejun Heo * RETURNS: 2011d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2012d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2013e22bee78STejun Heo */ 2014e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2015e22bee78STejun Heo { 201663d95a91STejun Heo struct worker_pool *pool = worker->pool; 2017e22bee78STejun Heo bool ret = false; 2018e22bee78STejun Heo 2019bc3a1afcSTejun Heo /* 2020bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2021bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2022bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2023bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2024bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2025bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2026bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2027bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2028bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2029bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2030bc3a1afcSTejun Heo * 2031bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2032bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2033bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2034bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2035bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2036bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2037bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2038bc3a1afcSTejun Heo * manager_mutex. 2039bc3a1afcSTejun Heo */ 204034a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2041e22bee78STejun Heo return ret; 2042e22bee78STejun Heo 2043ee378aa4SLai Jiangshan /* 2044bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2045bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2046ee378aa4SLai Jiangshan */ 2047bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2048d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2049bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2050ee378aa4SLai Jiangshan ret = true; 2051ee378aa4SLai Jiangshan } 2052ee378aa4SLai Jiangshan 205311ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2054e22bee78STejun Heo 2055e22bee78STejun Heo /* 2056e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2057e22bee78STejun Heo * on return. 2058e22bee78STejun Heo */ 205963d95a91STejun Heo ret |= maybe_destroy_workers(pool); 206063d95a91STejun Heo ret |= maybe_create_worker(pool); 2061e22bee78STejun Heo 2062bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 206334a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2064e22bee78STejun Heo return ret; 2065e22bee78STejun Heo } 2066e22bee78STejun Heo 2067a62428c0STejun Heo /** 2068a62428c0STejun Heo * process_one_work - process single work 2069c34056a3STejun Heo * @worker: self 2070a62428c0STejun Heo * @work: work to process 2071a62428c0STejun Heo * 2072a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2073a62428c0STejun Heo * process a single work including synchronization against and 2074a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2075a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2076a62428c0STejun Heo * call this function to process a work. 2077a62428c0STejun Heo * 2078a62428c0STejun Heo * CONTEXT: 2079d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2080a62428c0STejun Heo */ 2081c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2082d565ed63STejun Heo __releases(&pool->lock) 2083d565ed63STejun Heo __acquires(&pool->lock) 20841da177e4SLinus Torvalds { 2085112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2086bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2087112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 208873f53c4aSTejun Heo int work_color; 20897e11629dSTejun Heo struct worker *collision; 20904e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20914e6045f1SJohannes Berg /* 2092a62428c0STejun Heo * It is permissible to free the struct work_struct from 2093a62428c0STejun Heo * inside the function that is called from it, this we need to 2094a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2095a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2096a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20974e6045f1SJohannes Berg */ 20984d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20994d82a1deSPeter Zijlstra 21004d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21014e6045f1SJohannes Berg #endif 21026fec10a1STejun Heo /* 21036fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21046fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 210524647570STejun Heo * unbound or a disassociated pool. 21066fec10a1STejun Heo */ 21075f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 210824647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2109ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 211025511a47STejun Heo 21117e11629dSTejun Heo /* 21127e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21137e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21147e11629dSTejun Heo * already processing the work. If so, defer the work to the 21157e11629dSTejun Heo * currently executing one. 21167e11629dSTejun Heo */ 2117c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21187e11629dSTejun Heo if (unlikely(collision)) { 21197e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21207e11629dSTejun Heo return; 21217e11629dSTejun Heo } 21221da177e4SLinus Torvalds 21238930cabaSTejun Heo /* claim and dequeue */ 21241da177e4SLinus Torvalds debug_work_deactivate(work); 2125c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2126c34056a3STejun Heo worker->current_work = work; 2127a2c1c57bSTejun Heo worker->current_func = work->func; 2128112202d9STejun Heo worker->current_pwq = pwq; 212973f53c4aSTejun Heo work_color = get_work_color(work); 21307a22ad75STejun Heo 2131a62428c0STejun Heo list_del_init(&work->entry); 2132a62428c0STejun Heo 2133649027d7STejun Heo /* 2134fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2135fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2136fb0e7bebSTejun Heo */ 2137fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2138fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2139fb0e7bebSTejun Heo 2140974271c4STejun Heo /* 2141d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2142974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2143974271c4STejun Heo */ 214463d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 214563d95a91STejun Heo wake_up_worker(pool); 2146974271c4STejun Heo 21478930cabaSTejun Heo /* 21487c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2149d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 215023657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 215123657bb1STejun Heo * disabled. 21528930cabaSTejun Heo */ 21537c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21541da177e4SLinus Torvalds 2155d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2156365970a1SDavid Howells 2157112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21583295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2159e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2160a2c1c57bSTejun Heo worker->current_func(work); 2161e36c886aSArjan van de Ven /* 2162e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2163e36c886aSArjan van de Ven * point will only record its address. 2164e36c886aSArjan van de Ven */ 2165e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21663295f0efSIngo Molnar lock_map_release(&lockdep_map); 2167112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21681da177e4SLinus Torvalds 2169d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2170044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2171044c782cSValentin Ilie " last function: %pf\n", 2172a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2173a2c1c57bSTejun Heo worker->current_func); 2174d5abe669SPeter Zijlstra debug_show_held_locks(current); 2175d5abe669SPeter Zijlstra dump_stack(); 2176d5abe669SPeter Zijlstra } 2177d5abe669SPeter Zijlstra 2178d565ed63STejun Heo spin_lock_irq(&pool->lock); 2179a62428c0STejun Heo 2180fb0e7bebSTejun Heo /* clear cpu intensive status */ 2181fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2182fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2183fb0e7bebSTejun Heo 2184a62428c0STejun Heo /* we're done with it, release */ 218542f8570fSSasha Levin hash_del(&worker->hentry); 2186c34056a3STejun Heo worker->current_work = NULL; 2187a2c1c57bSTejun Heo worker->current_func = NULL; 2188112202d9STejun Heo worker->current_pwq = NULL; 2189112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21901da177e4SLinus Torvalds } 21911da177e4SLinus Torvalds 2192affee4b2STejun Heo /** 2193affee4b2STejun Heo * process_scheduled_works - process scheduled works 2194affee4b2STejun Heo * @worker: self 2195affee4b2STejun Heo * 2196affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2197affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2198affee4b2STejun Heo * fetches a work from the top and executes it. 2199affee4b2STejun Heo * 2200affee4b2STejun Heo * CONTEXT: 2201d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2202affee4b2STejun Heo * multiple times. 2203affee4b2STejun Heo */ 2204affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22051da177e4SLinus Torvalds { 2206affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2207affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2208a62428c0STejun Heo struct work_struct, entry); 2209c34056a3STejun Heo process_one_work(worker, work); 2210a62428c0STejun Heo } 22111da177e4SLinus Torvalds } 22121da177e4SLinus Torvalds 22134690c4abSTejun Heo /** 22144690c4abSTejun Heo * worker_thread - the worker thread function 2215c34056a3STejun Heo * @__worker: self 22164690c4abSTejun Heo * 2217c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2218c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2219c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2220c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2221c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22224690c4abSTejun Heo */ 2223c34056a3STejun Heo static int worker_thread(void *__worker) 22241da177e4SLinus Torvalds { 2225c34056a3STejun Heo struct worker *worker = __worker; 2226bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22271da177e4SLinus Torvalds 2228e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2229e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2230c8e55f36STejun Heo woke_up: 2231d565ed63STejun Heo spin_lock_irq(&pool->lock); 2232affee4b2STejun Heo 2233a9ab775bSTejun Heo /* am I supposed to die? */ 2234a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2235d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2236a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2237e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2238c8e55f36STejun Heo return 0; 2239c8e55f36STejun Heo } 2240c8e55f36STejun Heo 2241c8e55f36STejun Heo worker_leave_idle(worker); 2242db7bccf4STejun Heo recheck: 2243e22bee78STejun Heo /* no more worker necessary? */ 224463d95a91STejun Heo if (!need_more_worker(pool)) 2245e22bee78STejun Heo goto sleep; 2246e22bee78STejun Heo 2247e22bee78STejun Heo /* do we need to manage? */ 224863d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2249e22bee78STejun Heo goto recheck; 2250e22bee78STejun Heo 2251c8e55f36STejun Heo /* 2252c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2253c8e55f36STejun Heo * preparing to process a work or actually processing it. 2254c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2255c8e55f36STejun Heo */ 22566183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2257c8e55f36STejun Heo 2258e22bee78STejun Heo /* 2259a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2260a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2261a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2262a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2263a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2264e22bee78STejun Heo */ 2265a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2266e22bee78STejun Heo 2267e22bee78STejun Heo do { 2268affee4b2STejun Heo struct work_struct *work = 2269bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2270affee4b2STejun Heo struct work_struct, entry); 2271affee4b2STejun Heo 2272c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2273affee4b2STejun Heo /* optimization path, not strictly necessary */ 2274affee4b2STejun Heo process_one_work(worker, work); 2275affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2276affee4b2STejun Heo process_scheduled_works(worker); 2277affee4b2STejun Heo } else { 2278c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2279affee4b2STejun Heo process_scheduled_works(worker); 2280affee4b2STejun Heo } 228163d95a91STejun Heo } while (keep_working(pool)); 2282affee4b2STejun Heo 2283e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2284d313dd85STejun Heo sleep: 228563d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2286e22bee78STejun Heo goto recheck; 2287d313dd85STejun Heo 2288c8e55f36STejun Heo /* 2289d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2290d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2291d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2292d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2293d565ed63STejun Heo * event. 2294c8e55f36STejun Heo */ 2295c8e55f36STejun Heo worker_enter_idle(worker); 2296c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2297d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22981da177e4SLinus Torvalds schedule(); 2299c8e55f36STejun Heo goto woke_up; 23001da177e4SLinus Torvalds } 23011da177e4SLinus Torvalds 2302e22bee78STejun Heo /** 2303e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2304111c225aSTejun Heo * @__rescuer: self 2305e22bee78STejun Heo * 2306e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2307493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2308e22bee78STejun Heo * 2309706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2310e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2311e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2312e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2313e22bee78STejun Heo * the problem rescuer solves. 2314e22bee78STejun Heo * 2315706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2316706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2317e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2318e22bee78STejun Heo * 2319e22bee78STejun Heo * This should happen rarely. 2320e22bee78STejun Heo */ 2321111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2322e22bee78STejun Heo { 2323111c225aSTejun Heo struct worker *rescuer = __rescuer; 2324111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2325e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2326e22bee78STejun Heo 2327e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2328111c225aSTejun Heo 2329111c225aSTejun Heo /* 2330111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2331111c225aSTejun Heo * doesn't participate in concurrency management. 2332111c225aSTejun Heo */ 2333111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2334e22bee78STejun Heo repeat: 2335e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23361da177e4SLinus Torvalds 2337412d32e6SMike Galbraith if (kthread_should_stop()) { 2338412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2339111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2340e22bee78STejun Heo return 0; 2341412d32e6SMike Galbraith } 23421da177e4SLinus Torvalds 2343493a1724STejun Heo /* see whether any pwq is asking for help */ 23442e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2345493a1724STejun Heo 2346493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2347493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2348493a1724STejun Heo struct pool_workqueue, mayday_node); 2349112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2350e22bee78STejun Heo struct work_struct *work, *n; 2351e22bee78STejun Heo 2352e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2353493a1724STejun Heo list_del_init(&pwq->mayday_node); 2354493a1724STejun Heo 23552e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2356e22bee78STejun Heo 2357e22bee78STejun Heo /* migrate to the target cpu if possible */ 2358f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2359b3104104SLai Jiangshan rescuer->pool = pool; 2360e22bee78STejun Heo 2361e22bee78STejun Heo /* 2362e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2363e22bee78STejun Heo * process'em. 2364e22bee78STejun Heo */ 23656183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2366bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2367112202d9STejun Heo if (get_work_pwq(work) == pwq) 2368e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2369e22bee78STejun Heo 2370e22bee78STejun Heo process_scheduled_works(rescuer); 23717576958aSTejun Heo 23727576958aSTejun Heo /* 2373d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23747576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23757576958aSTejun Heo * and stalling the execution. 23767576958aSTejun Heo */ 237763d95a91STejun Heo if (keep_working(pool)) 237863d95a91STejun Heo wake_up_worker(pool); 23797576958aSTejun Heo 2380b3104104SLai Jiangshan rescuer->pool = NULL; 2381493a1724STejun Heo spin_unlock(&pool->lock); 23822e109a28STejun Heo spin_lock(&wq_mayday_lock); 23831da177e4SLinus Torvalds } 23841da177e4SLinus Torvalds 23852e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2386493a1724STejun Heo 2387111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2388111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2389e22bee78STejun Heo schedule(); 2390e22bee78STejun Heo goto repeat; 23911da177e4SLinus Torvalds } 23921da177e4SLinus Torvalds 2393fc2e4d70SOleg Nesterov struct wq_barrier { 2394fc2e4d70SOleg Nesterov struct work_struct work; 2395fc2e4d70SOleg Nesterov struct completion done; 2396fc2e4d70SOleg Nesterov }; 2397fc2e4d70SOleg Nesterov 2398fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2399fc2e4d70SOleg Nesterov { 2400fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2401fc2e4d70SOleg Nesterov complete(&barr->done); 2402fc2e4d70SOleg Nesterov } 2403fc2e4d70SOleg Nesterov 24044690c4abSTejun Heo /** 24054690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2406112202d9STejun Heo * @pwq: pwq to insert barrier into 24074690c4abSTejun Heo * @barr: wq_barrier to insert 2408affee4b2STejun Heo * @target: target work to attach @barr to 2409affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24104690c4abSTejun Heo * 2411affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2412affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2413affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2414affee4b2STejun Heo * cpu. 2415affee4b2STejun Heo * 2416affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2417affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2418affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2419affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2420affee4b2STejun Heo * after a work with LINKED flag set. 2421affee4b2STejun Heo * 2422affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2423112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24244690c4abSTejun Heo * 24254690c4abSTejun Heo * CONTEXT: 2426d565ed63STejun Heo * spin_lock_irq(pool->lock). 24274690c4abSTejun Heo */ 2428112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2429affee4b2STejun Heo struct wq_barrier *barr, 2430affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2431fc2e4d70SOleg Nesterov { 2432affee4b2STejun Heo struct list_head *head; 2433affee4b2STejun Heo unsigned int linked = 0; 2434affee4b2STejun Heo 2435dc186ad7SThomas Gleixner /* 2436d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2437dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2438dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2439dc186ad7SThomas Gleixner * might deadlock. 2440dc186ad7SThomas Gleixner */ 2441ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 244222df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2443fc2e4d70SOleg Nesterov init_completion(&barr->done); 244483c22520SOleg Nesterov 2445affee4b2STejun Heo /* 2446affee4b2STejun Heo * If @target is currently being executed, schedule the 2447affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2448affee4b2STejun Heo */ 2449affee4b2STejun Heo if (worker) 2450affee4b2STejun Heo head = worker->scheduled.next; 2451affee4b2STejun Heo else { 2452affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2453affee4b2STejun Heo 2454affee4b2STejun Heo head = target->entry.next; 2455affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2456affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2457affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2458affee4b2STejun Heo } 2459affee4b2STejun Heo 2460dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2461112202d9STejun Heo insert_work(pwq, &barr->work, head, 2462affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2463fc2e4d70SOleg Nesterov } 2464fc2e4d70SOleg Nesterov 246573f53c4aSTejun Heo /** 2466112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 246773f53c4aSTejun Heo * @wq: workqueue being flushed 246873f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 246973f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 247073f53c4aSTejun Heo * 2471112202d9STejun Heo * Prepare pwqs for workqueue flushing. 247273f53c4aSTejun Heo * 2473112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2474112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2475112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2476112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2477112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 247873f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 247973f53c4aSTejun Heo * 248073f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 248173f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 248273f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 248373f53c4aSTejun Heo * is returned. 248473f53c4aSTejun Heo * 2485112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 248673f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 248773f53c4aSTejun Heo * advanced to @work_color. 248873f53c4aSTejun Heo * 248973f53c4aSTejun Heo * CONTEXT: 24903c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 249173f53c4aSTejun Heo * 249273f53c4aSTejun Heo * RETURNS: 249373f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 249473f53c4aSTejun Heo * otherwise. 249573f53c4aSTejun Heo */ 2496112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 249773f53c4aSTejun Heo int flush_color, int work_color) 24981da177e4SLinus Torvalds { 249973f53c4aSTejun Heo bool wait = false; 250049e3cf44STejun Heo struct pool_workqueue *pwq; 25011da177e4SLinus Torvalds 250273f53c4aSTejun Heo if (flush_color >= 0) { 25036183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2504112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2505dc186ad7SThomas Gleixner } 250614441960SOleg Nesterov 250749e3cf44STejun Heo for_each_pwq(pwq, wq) { 2508112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25091da177e4SLinus Torvalds 2510b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 251173f53c4aSTejun Heo 251273f53c4aSTejun Heo if (flush_color >= 0) { 25136183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 251473f53c4aSTejun Heo 2515112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2516112202d9STejun Heo pwq->flush_color = flush_color; 2517112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 251873f53c4aSTejun Heo wait = true; 25191da177e4SLinus Torvalds } 252073f53c4aSTejun Heo } 252173f53c4aSTejun Heo 252273f53c4aSTejun Heo if (work_color >= 0) { 25236183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2524112202d9STejun Heo pwq->work_color = work_color; 252573f53c4aSTejun Heo } 252673f53c4aSTejun Heo 2527b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25281da177e4SLinus Torvalds } 25291da177e4SLinus Torvalds 2530112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 253173f53c4aSTejun Heo complete(&wq->first_flusher->done); 253273f53c4aSTejun Heo 253373f53c4aSTejun Heo return wait; 253483c22520SOleg Nesterov } 25351da177e4SLinus Torvalds 25360fcb78c2SRolf Eike Beer /** 25371da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25380fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25391da177e4SLinus Torvalds * 2540c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2541c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25421da177e4SLinus Torvalds */ 25437ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25441da177e4SLinus Torvalds { 254573f53c4aSTejun Heo struct wq_flusher this_flusher = { 254673f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 254773f53c4aSTejun Heo .flush_color = -1, 254873f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 254973f53c4aSTejun Heo }; 255073f53c4aSTejun Heo int next_color; 2551b1f4ec17SOleg Nesterov 25523295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25533295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 255473f53c4aSTejun Heo 25553c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 255673f53c4aSTejun Heo 255773f53c4aSTejun Heo /* 255873f53c4aSTejun Heo * Start-to-wait phase 255973f53c4aSTejun Heo */ 256073f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 256173f53c4aSTejun Heo 256273f53c4aSTejun Heo if (next_color != wq->flush_color) { 256373f53c4aSTejun Heo /* 256473f53c4aSTejun Heo * Color space is not full. The current work_color 256573f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 256673f53c4aSTejun Heo * by one. 256773f53c4aSTejun Heo */ 25686183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 256973f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 257073f53c4aSTejun Heo wq->work_color = next_color; 257173f53c4aSTejun Heo 257273f53c4aSTejun Heo if (!wq->first_flusher) { 257373f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25746183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 257573f53c4aSTejun Heo 257673f53c4aSTejun Heo wq->first_flusher = &this_flusher; 257773f53c4aSTejun Heo 2578112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 257973f53c4aSTejun Heo wq->work_color)) { 258073f53c4aSTejun Heo /* nothing to flush, done */ 258173f53c4aSTejun Heo wq->flush_color = next_color; 258273f53c4aSTejun Heo wq->first_flusher = NULL; 258373f53c4aSTejun Heo goto out_unlock; 258473f53c4aSTejun Heo } 258573f53c4aSTejun Heo } else { 258673f53c4aSTejun Heo /* wait in queue */ 25876183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 258873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2589112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 259073f53c4aSTejun Heo } 259173f53c4aSTejun Heo } else { 259273f53c4aSTejun Heo /* 259373f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 259473f53c4aSTejun Heo * The next flush completion will assign us 259573f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 259673f53c4aSTejun Heo */ 259773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 259873f53c4aSTejun Heo } 259973f53c4aSTejun Heo 26003c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 260173f53c4aSTejun Heo 260273f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 260373f53c4aSTejun Heo 260473f53c4aSTejun Heo /* 260573f53c4aSTejun Heo * Wake-up-and-cascade phase 260673f53c4aSTejun Heo * 260773f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 260873f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 260973f53c4aSTejun Heo */ 261073f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 261173f53c4aSTejun Heo return; 261273f53c4aSTejun Heo 26133c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 261473f53c4aSTejun Heo 26154ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26164ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26174ce48b37STejun Heo goto out_unlock; 26184ce48b37STejun Heo 261973f53c4aSTejun Heo wq->first_flusher = NULL; 262073f53c4aSTejun Heo 26216183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26226183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 262373f53c4aSTejun Heo 262473f53c4aSTejun Heo while (true) { 262573f53c4aSTejun Heo struct wq_flusher *next, *tmp; 262673f53c4aSTejun Heo 262773f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 262873f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 262973f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 263073f53c4aSTejun Heo break; 263173f53c4aSTejun Heo list_del_init(&next->list); 263273f53c4aSTejun Heo complete(&next->done); 263373f53c4aSTejun Heo } 263473f53c4aSTejun Heo 26356183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 263673f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 263973f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 264073f53c4aSTejun Heo 264173f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 264273f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 264373f53c4aSTejun Heo /* 264473f53c4aSTejun Heo * Assign the same color to all overflowed 264573f53c4aSTejun Heo * flushers, advance work_color and append to 264673f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 264773f53c4aSTejun Heo * phase for these overflowed flushers. 264873f53c4aSTejun Heo */ 264973f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 265073f53c4aSTejun Heo tmp->flush_color = wq->work_color; 265173f53c4aSTejun Heo 265273f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 265373f53c4aSTejun Heo 265473f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 265573f53c4aSTejun Heo &wq->flusher_queue); 2656112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 265773f53c4aSTejun Heo } 265873f53c4aSTejun Heo 265973f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26606183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 266173f53c4aSTejun Heo break; 266273f53c4aSTejun Heo } 266373f53c4aSTejun Heo 266473f53c4aSTejun Heo /* 266573f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2666112202d9STejun Heo * the new first flusher and arm pwqs. 266773f53c4aSTejun Heo */ 26686183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26696183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 267073f53c4aSTejun Heo 267173f53c4aSTejun Heo list_del_init(&next->list); 267273f53c4aSTejun Heo wq->first_flusher = next; 267373f53c4aSTejun Heo 2674112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 267573f53c4aSTejun Heo break; 267673f53c4aSTejun Heo 267773f53c4aSTejun Heo /* 267873f53c4aSTejun Heo * Meh... this color is already done, clear first 267973f53c4aSTejun Heo * flusher and repeat cascading. 268073f53c4aSTejun Heo */ 268173f53c4aSTejun Heo wq->first_flusher = NULL; 268273f53c4aSTejun Heo } 268373f53c4aSTejun Heo 268473f53c4aSTejun Heo out_unlock: 26853c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26861da177e4SLinus Torvalds } 2687ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26881da177e4SLinus Torvalds 26899c5a2ba7STejun Heo /** 26909c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26919c5a2ba7STejun Heo * @wq: workqueue to drain 26929c5a2ba7STejun Heo * 26939c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26949c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26959c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26969c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26979c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26989c5a2ba7STejun Heo * takes too long. 26999c5a2ba7STejun Heo */ 27009c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27019c5a2ba7STejun Heo { 27029c5a2ba7STejun Heo unsigned int flush_cnt = 0; 270349e3cf44STejun Heo struct pool_workqueue *pwq; 27049c5a2ba7STejun Heo 27059c5a2ba7STejun Heo /* 27069c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27079c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2708618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27099c5a2ba7STejun Heo */ 271087fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27119c5a2ba7STejun Heo if (!wq->nr_drainers++) 2712618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 271387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27149c5a2ba7STejun Heo reflush: 27159c5a2ba7STejun Heo flush_workqueue(wq); 27169c5a2ba7STejun Heo 2717b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 271876af4d93STejun Heo 271949e3cf44STejun Heo for_each_pwq(pwq, wq) { 2720fa2563e4SThomas Tuttle bool drained; 27219c5a2ba7STejun Heo 2722b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2723112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2724b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2725fa2563e4SThomas Tuttle 2726fa2563e4SThomas Tuttle if (drained) 27279c5a2ba7STejun Heo continue; 27289c5a2ba7STejun Heo 27299c5a2ba7STejun Heo if (++flush_cnt == 10 || 27309c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2731c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27329c5a2ba7STejun Heo wq->name, flush_cnt); 273376af4d93STejun Heo 2734b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27359c5a2ba7STejun Heo goto reflush; 27369c5a2ba7STejun Heo } 27379c5a2ba7STejun Heo 27389c5a2ba7STejun Heo if (!--wq->nr_drainers) 2739618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 274087fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27419c5a2ba7STejun Heo } 27429c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27439c5a2ba7STejun Heo 2744606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2745baf59022STejun Heo { 2746baf59022STejun Heo struct worker *worker = NULL; 2747c9e7cf27STejun Heo struct worker_pool *pool; 2748112202d9STejun Heo struct pool_workqueue *pwq; 2749baf59022STejun Heo 2750baf59022STejun Heo might_sleep(); 2751baf59022STejun Heo 2752fa1b54e6STejun Heo local_irq_disable(); 2753fa1b54e6STejun Heo pool = get_work_pool(work); 2754fa1b54e6STejun Heo if (!pool) { 2755fa1b54e6STejun Heo local_irq_enable(); 2756fa1b54e6STejun Heo return false; 2757fa1b54e6STejun Heo } 2758fa1b54e6STejun Heo 2759fa1b54e6STejun Heo spin_lock(&pool->lock); 27600b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2761112202d9STejun Heo pwq = get_work_pwq(work); 2762112202d9STejun Heo if (pwq) { 2763112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2764baf59022STejun Heo goto already_gone; 2765606a5020STejun Heo } else { 2766c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2767baf59022STejun Heo if (!worker) 2768baf59022STejun Heo goto already_gone; 2769112202d9STejun Heo pwq = worker->current_pwq; 2770606a5020STejun Heo } 2771baf59022STejun Heo 2772112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2773d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2774baf59022STejun Heo 2775e159489bSTejun Heo /* 2776e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2777e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2778e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2779e159489bSTejun Heo * access. 2780e159489bSTejun Heo */ 2781493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2782112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2783e159489bSTejun Heo else 2784112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2785112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2786e159489bSTejun Heo 2787baf59022STejun Heo return true; 2788baf59022STejun Heo already_gone: 2789d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2790baf59022STejun Heo return false; 2791baf59022STejun Heo } 2792baf59022STejun Heo 2793db700897SOleg Nesterov /** 2794401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2795401a8d04STejun Heo * @work: the work to flush 2796db700897SOleg Nesterov * 2797606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2798606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2799401a8d04STejun Heo * 2800401a8d04STejun Heo * RETURNS: 2801401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2802401a8d04STejun Heo * %false if it was already idle. 2803db700897SOleg Nesterov */ 2804401a8d04STejun Heo bool flush_work(struct work_struct *work) 2805db700897SOleg Nesterov { 2806db700897SOleg Nesterov struct wq_barrier barr; 2807db700897SOleg Nesterov 28080976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28090976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28100976dfc1SStephen Boyd 2811606a5020STejun Heo if (start_flush_work(work, &barr)) { 2812db700897SOleg Nesterov wait_for_completion(&barr.done); 2813dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2814401a8d04STejun Heo return true; 2815606a5020STejun Heo } else { 2816401a8d04STejun Heo return false; 2817db700897SOleg Nesterov } 2818606a5020STejun Heo } 2819db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2820db700897SOleg Nesterov 282136e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2822401a8d04STejun Heo { 2823bbb68dfaSTejun Heo unsigned long flags; 28241f1f642eSOleg Nesterov int ret; 28251f1f642eSOleg Nesterov 28261f1f642eSOleg Nesterov do { 2827bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2828bbb68dfaSTejun Heo /* 2829bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2830bbb68dfaSTejun Heo * would be waiting for before retrying. 2831bbb68dfaSTejun Heo */ 2832bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2833606a5020STejun Heo flush_work(work); 28341f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28351f1f642eSOleg Nesterov 2836bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2837bbb68dfaSTejun Heo mark_work_canceling(work); 2838bbb68dfaSTejun Heo local_irq_restore(flags); 2839bbb68dfaSTejun Heo 2840606a5020STejun Heo flush_work(work); 28417a22ad75STejun Heo clear_work_data(work); 28421f1f642eSOleg Nesterov return ret; 28431f1f642eSOleg Nesterov } 28441f1f642eSOleg Nesterov 28456e84d644SOleg Nesterov /** 2846401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2847401a8d04STejun Heo * @work: the work to cancel 28486e84d644SOleg Nesterov * 2849401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2850401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2851401a8d04STejun Heo * another workqueue. On return from this function, @work is 2852401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28531f1f642eSOleg Nesterov * 2854401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2855401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28566e84d644SOleg Nesterov * 2857401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28586e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2859401a8d04STejun Heo * 2860401a8d04STejun Heo * RETURNS: 2861401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28626e84d644SOleg Nesterov */ 2863401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28646e84d644SOleg Nesterov { 286536e227d2STejun Heo return __cancel_work_timer(work, false); 2866b89deed3SOleg Nesterov } 286728e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2868b89deed3SOleg Nesterov 28696e84d644SOleg Nesterov /** 2870401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2871401a8d04STejun Heo * @dwork: the delayed work to flush 28726e84d644SOleg Nesterov * 2873401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2874401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2875401a8d04STejun Heo * considers the last queueing instance of @dwork. 28761f1f642eSOleg Nesterov * 2877401a8d04STejun Heo * RETURNS: 2878401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2879401a8d04STejun Heo * %false if it was already idle. 28806e84d644SOleg Nesterov */ 2881401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2882401a8d04STejun Heo { 28838930cabaSTejun Heo local_irq_disable(); 2884401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 288560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28868930cabaSTejun Heo local_irq_enable(); 2887401a8d04STejun Heo return flush_work(&dwork->work); 2888401a8d04STejun Heo } 2889401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2890401a8d04STejun Heo 2891401a8d04STejun Heo /** 289257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 289357b30ae7STejun Heo * @dwork: delayed_work to cancel 289409383498STejun Heo * 289557b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 289657b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 289757b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 289857b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 289957b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 290009383498STejun Heo * 290157b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 290209383498STejun Heo */ 290357b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 290409383498STejun Heo { 290557b30ae7STejun Heo unsigned long flags; 290657b30ae7STejun Heo int ret; 290757b30ae7STejun Heo 290857b30ae7STejun Heo do { 290957b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 291057b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 291157b30ae7STejun Heo 291257b30ae7STejun Heo if (unlikely(ret < 0)) 291357b30ae7STejun Heo return false; 291457b30ae7STejun Heo 29157c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29167c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 291757b30ae7STejun Heo local_irq_restore(flags); 2918c0158ca6SDan Magenheimer return ret; 291909383498STejun Heo } 292057b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 292109383498STejun Heo 292209383498STejun Heo /** 2923401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2924401a8d04STejun Heo * @dwork: the delayed work cancel 2925401a8d04STejun Heo * 2926401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2927401a8d04STejun Heo * 2928401a8d04STejun Heo * RETURNS: 2929401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2930401a8d04STejun Heo */ 2931401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29326e84d644SOleg Nesterov { 293336e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29346e84d644SOleg Nesterov } 2935f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29361da177e4SLinus Torvalds 29370fcb78c2SRolf Eike Beer /** 293831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2939b6136773SAndrew Morton * @func: the function to call 2940b6136773SAndrew Morton * 294131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 294231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2943b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 294431ddd871STejun Heo * 294531ddd871STejun Heo * RETURNS: 294631ddd871STejun Heo * 0 on success, -errno on failure. 2947b6136773SAndrew Morton */ 294865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 294915316ba8SChristoph Lameter { 295015316ba8SChristoph Lameter int cpu; 295138f51568SNamhyung Kim struct work_struct __percpu *works; 295215316ba8SChristoph Lameter 2953b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2954b6136773SAndrew Morton if (!works) 295515316ba8SChristoph Lameter return -ENOMEM; 2956b6136773SAndrew Morton 295795402b38SGautham R Shenoy get_online_cpus(); 295893981800STejun Heo 295915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29609bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29619bfb1839SIngo Molnar 29629bfb1839SIngo Molnar INIT_WORK(work, func); 29638de6d308SOleg Nesterov schedule_work_on(cpu, work); 296415316ba8SChristoph Lameter } 296593981800STejun Heo 296693981800STejun Heo for_each_online_cpu(cpu) 29678616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 296893981800STejun Heo 296995402b38SGautham R Shenoy put_online_cpus(); 2970b6136773SAndrew Morton free_percpu(works); 297115316ba8SChristoph Lameter return 0; 297215316ba8SChristoph Lameter } 297315316ba8SChristoph Lameter 2974eef6a7d5SAlan Stern /** 2975eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2976eef6a7d5SAlan Stern * 2977eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2978eef6a7d5SAlan Stern * completion. 2979eef6a7d5SAlan Stern * 2980eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2981eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2982eef6a7d5SAlan Stern * will lead to deadlock: 2983eef6a7d5SAlan Stern * 2984eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2985eef6a7d5SAlan Stern * a lock held by your code or its caller. 2986eef6a7d5SAlan Stern * 2987eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2988eef6a7d5SAlan Stern * 2989eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2990eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2991eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2992eef6a7d5SAlan Stern * 2993eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2994eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2995eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2996eef6a7d5SAlan Stern * cancel_work_sync() instead. 2997eef6a7d5SAlan Stern */ 29981da177e4SLinus Torvalds void flush_scheduled_work(void) 29991da177e4SLinus Torvalds { 3000d320c038STejun Heo flush_workqueue(system_wq); 30011da177e4SLinus Torvalds } 3002ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30031da177e4SLinus Torvalds 30041da177e4SLinus Torvalds /** 30051fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30061fa44ecaSJames Bottomley * @fn: the function to execute 30071fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30081fa44ecaSJames Bottomley * be available when the work executes) 30091fa44ecaSJames Bottomley * 30101fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30111fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30121fa44ecaSJames Bottomley * 30131fa44ecaSJames Bottomley * Returns: 0 - function was executed 30141fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30151fa44ecaSJames Bottomley */ 301665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30171fa44ecaSJames Bottomley { 30181fa44ecaSJames Bottomley if (!in_interrupt()) { 301965f27f38SDavid Howells fn(&ew->work); 30201fa44ecaSJames Bottomley return 0; 30211fa44ecaSJames Bottomley } 30221fa44ecaSJames Bottomley 302365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30241fa44ecaSJames Bottomley schedule_work(&ew->work); 30251fa44ecaSJames Bottomley 30261fa44ecaSJames Bottomley return 1; 30271fa44ecaSJames Bottomley } 30281fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30291fa44ecaSJames Bottomley 3030226223abSTejun Heo #ifdef CONFIG_SYSFS 3031226223abSTejun Heo /* 3032226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3033226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3034226223abSTejun Heo * following attributes. 3035226223abSTejun Heo * 3036226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3037226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3038226223abSTejun Heo * 3039226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3040226223abSTejun Heo * 3041226223abSTejun Heo * id RO int : the associated pool ID 3042226223abSTejun Heo * nice RW int : nice value of the workers 3043226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3044226223abSTejun Heo */ 3045226223abSTejun Heo struct wq_device { 3046226223abSTejun Heo struct workqueue_struct *wq; 3047226223abSTejun Heo struct device dev; 3048226223abSTejun Heo }; 3049226223abSTejun Heo 3050226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3051226223abSTejun Heo { 3052226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3053226223abSTejun Heo 3054226223abSTejun Heo return wq_dev->wq; 3055226223abSTejun Heo } 3056226223abSTejun Heo 3057226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3058226223abSTejun Heo struct device_attribute *attr, char *buf) 3059226223abSTejun Heo { 3060226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3061226223abSTejun Heo 3062226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3063226223abSTejun Heo } 3064226223abSTejun Heo 3065226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3066226223abSTejun Heo struct device_attribute *attr, char *buf) 3067226223abSTejun Heo { 3068226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3069226223abSTejun Heo 3070226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3071226223abSTejun Heo } 3072226223abSTejun Heo 3073226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3074226223abSTejun Heo struct device_attribute *attr, 3075226223abSTejun Heo const char *buf, size_t count) 3076226223abSTejun Heo { 3077226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3078226223abSTejun Heo int val; 3079226223abSTejun Heo 3080226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3081226223abSTejun Heo return -EINVAL; 3082226223abSTejun Heo 3083226223abSTejun Heo workqueue_set_max_active(wq, val); 3084226223abSTejun Heo return count; 3085226223abSTejun Heo } 3086226223abSTejun Heo 3087226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3088226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3089226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3090226223abSTejun Heo __ATTR_NULL, 3091226223abSTejun Heo }; 3092226223abSTejun Heo 3093226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3094226223abSTejun Heo struct device_attribute *attr, char *buf) 3095226223abSTejun Heo { 3096226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3097226223abSTejun Heo struct worker_pool *pool; 3098226223abSTejun Heo int written; 3099226223abSTejun Heo 3100226223abSTejun Heo rcu_read_lock_sched(); 3101226223abSTejun Heo pool = first_pwq(wq)->pool; 3102226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3103226223abSTejun Heo rcu_read_unlock_sched(); 3104226223abSTejun Heo 3105226223abSTejun Heo return written; 3106226223abSTejun Heo } 3107226223abSTejun Heo 3108226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3109226223abSTejun Heo char *buf) 3110226223abSTejun Heo { 3111226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3112226223abSTejun Heo int written; 3113226223abSTejun Heo 31146029a918STejun Heo mutex_lock(&wq->mutex); 31156029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 31166029a918STejun Heo mutex_unlock(&wq->mutex); 3117226223abSTejun Heo 3118226223abSTejun Heo return written; 3119226223abSTejun Heo } 3120226223abSTejun Heo 3121226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3122226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3123226223abSTejun Heo { 3124226223abSTejun Heo struct workqueue_attrs *attrs; 3125226223abSTejun Heo 3126226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3127226223abSTejun Heo if (!attrs) 3128226223abSTejun Heo return NULL; 3129226223abSTejun Heo 31306029a918STejun Heo mutex_lock(&wq->mutex); 31316029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31326029a918STejun Heo mutex_unlock(&wq->mutex); 3133226223abSTejun Heo return attrs; 3134226223abSTejun Heo } 3135226223abSTejun Heo 3136226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3137226223abSTejun Heo const char *buf, size_t count) 3138226223abSTejun Heo { 3139226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3140226223abSTejun Heo struct workqueue_attrs *attrs; 3141226223abSTejun Heo int ret; 3142226223abSTejun Heo 3143226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3144226223abSTejun Heo if (!attrs) 3145226223abSTejun Heo return -ENOMEM; 3146226223abSTejun Heo 3147226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3148226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3149226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3150226223abSTejun Heo else 3151226223abSTejun Heo ret = -EINVAL; 3152226223abSTejun Heo 3153226223abSTejun Heo free_workqueue_attrs(attrs); 3154226223abSTejun Heo return ret ?: count; 3155226223abSTejun Heo } 3156226223abSTejun Heo 3157226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3158226223abSTejun Heo struct device_attribute *attr, char *buf) 3159226223abSTejun Heo { 3160226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3161226223abSTejun Heo int written; 3162226223abSTejun Heo 31636029a918STejun Heo mutex_lock(&wq->mutex); 31646029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 31656029a918STejun Heo mutex_unlock(&wq->mutex); 3166226223abSTejun Heo 3167226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3168226223abSTejun Heo return written; 3169226223abSTejun Heo } 3170226223abSTejun Heo 3171226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3172226223abSTejun Heo struct device_attribute *attr, 3173226223abSTejun Heo const char *buf, size_t count) 3174226223abSTejun Heo { 3175226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3176226223abSTejun Heo struct workqueue_attrs *attrs; 3177226223abSTejun Heo int ret; 3178226223abSTejun Heo 3179226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3180226223abSTejun Heo if (!attrs) 3181226223abSTejun Heo return -ENOMEM; 3182226223abSTejun Heo 3183226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3184226223abSTejun Heo if (!ret) 3185226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3186226223abSTejun Heo 3187226223abSTejun Heo free_workqueue_attrs(attrs); 3188226223abSTejun Heo return ret ?: count; 3189226223abSTejun Heo } 3190226223abSTejun Heo 3191226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3192226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3193226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3194226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3195226223abSTejun Heo __ATTR_NULL, 3196226223abSTejun Heo }; 3197226223abSTejun Heo 3198226223abSTejun Heo static struct bus_type wq_subsys = { 3199226223abSTejun Heo .name = "workqueue", 3200226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3201226223abSTejun Heo }; 3202226223abSTejun Heo 3203226223abSTejun Heo static int __init wq_sysfs_init(void) 3204226223abSTejun Heo { 3205226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3206226223abSTejun Heo } 3207226223abSTejun Heo core_initcall(wq_sysfs_init); 3208226223abSTejun Heo 3209226223abSTejun Heo static void wq_device_release(struct device *dev) 3210226223abSTejun Heo { 3211226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3212226223abSTejun Heo 3213226223abSTejun Heo kfree(wq_dev); 3214226223abSTejun Heo } 3215226223abSTejun Heo 3216226223abSTejun Heo /** 3217226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3218226223abSTejun Heo * @wq: the workqueue to register 3219226223abSTejun Heo * 3220226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3221226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3222226223abSTejun Heo * which is the preferred method. 3223226223abSTejun Heo * 3224226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3225226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3226226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3227226223abSTejun Heo * attributes. 3228226223abSTejun Heo * 3229226223abSTejun Heo * Returns 0 on success, -errno on failure. 3230226223abSTejun Heo */ 3231226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3232226223abSTejun Heo { 3233226223abSTejun Heo struct wq_device *wq_dev; 3234226223abSTejun Heo int ret; 3235226223abSTejun Heo 3236226223abSTejun Heo /* 3237226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3238226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3239226223abSTejun Heo * workqueues. 3240226223abSTejun Heo */ 3241226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3242226223abSTejun Heo return -EINVAL; 3243226223abSTejun Heo 3244226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3245226223abSTejun Heo if (!wq_dev) 3246226223abSTejun Heo return -ENOMEM; 3247226223abSTejun Heo 3248226223abSTejun Heo wq_dev->wq = wq; 3249226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3250226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3251226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3252226223abSTejun Heo 3253226223abSTejun Heo /* 3254226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3255226223abSTejun Heo * everything is ready. 3256226223abSTejun Heo */ 3257226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3258226223abSTejun Heo 3259226223abSTejun Heo ret = device_register(&wq_dev->dev); 3260226223abSTejun Heo if (ret) { 3261226223abSTejun Heo kfree(wq_dev); 3262226223abSTejun Heo wq->wq_dev = NULL; 3263226223abSTejun Heo return ret; 3264226223abSTejun Heo } 3265226223abSTejun Heo 3266226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3267226223abSTejun Heo struct device_attribute *attr; 3268226223abSTejun Heo 3269226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3270226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3271226223abSTejun Heo if (ret) { 3272226223abSTejun Heo device_unregister(&wq_dev->dev); 3273226223abSTejun Heo wq->wq_dev = NULL; 3274226223abSTejun Heo return ret; 3275226223abSTejun Heo } 3276226223abSTejun Heo } 3277226223abSTejun Heo } 3278226223abSTejun Heo 3279226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3280226223abSTejun Heo return 0; 3281226223abSTejun Heo } 3282226223abSTejun Heo 3283226223abSTejun Heo /** 3284226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3285226223abSTejun Heo * @wq: the workqueue to unregister 3286226223abSTejun Heo * 3287226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3288226223abSTejun Heo */ 3289226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3290226223abSTejun Heo { 3291226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3292226223abSTejun Heo 3293226223abSTejun Heo if (!wq->wq_dev) 3294226223abSTejun Heo return; 3295226223abSTejun Heo 3296226223abSTejun Heo wq->wq_dev = NULL; 3297226223abSTejun Heo device_unregister(&wq_dev->dev); 3298226223abSTejun Heo } 3299226223abSTejun Heo #else /* CONFIG_SYSFS */ 3300226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3301226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3302226223abSTejun Heo 33037a4e344cSTejun Heo /** 33047a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33057a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33067a4e344cSTejun Heo * 33077a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33087a4e344cSTejun Heo */ 33097a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33107a4e344cSTejun Heo { 33117a4e344cSTejun Heo if (attrs) { 33127a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33137a4e344cSTejun Heo kfree(attrs); 33147a4e344cSTejun Heo } 33157a4e344cSTejun Heo } 33167a4e344cSTejun Heo 33177a4e344cSTejun Heo /** 33187a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33197a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33207a4e344cSTejun Heo * 33217a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33227a4e344cSTejun Heo * return it. Returns NULL on failure. 33237a4e344cSTejun Heo */ 33247a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33257a4e344cSTejun Heo { 33267a4e344cSTejun Heo struct workqueue_attrs *attrs; 33277a4e344cSTejun Heo 33287a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33297a4e344cSTejun Heo if (!attrs) 33307a4e344cSTejun Heo goto fail; 33317a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33327a4e344cSTejun Heo goto fail; 33337a4e344cSTejun Heo 333413e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33357a4e344cSTejun Heo return attrs; 33367a4e344cSTejun Heo fail: 33377a4e344cSTejun Heo free_workqueue_attrs(attrs); 33387a4e344cSTejun Heo return NULL; 33397a4e344cSTejun Heo } 33407a4e344cSTejun Heo 334129c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 334229c91e99STejun Heo const struct workqueue_attrs *from) 334329c91e99STejun Heo { 334429c91e99STejun Heo to->nice = from->nice; 334529c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 334629c91e99STejun Heo } 334729c91e99STejun Heo 334829c91e99STejun Heo /* hash value of the content of @attr */ 334929c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 335029c91e99STejun Heo { 335129c91e99STejun Heo u32 hash = 0; 335229c91e99STejun Heo 335329c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 335413e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 335513e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 335629c91e99STejun Heo return hash; 335729c91e99STejun Heo } 335829c91e99STejun Heo 335929c91e99STejun Heo /* content equality test */ 336029c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 336129c91e99STejun Heo const struct workqueue_attrs *b) 336229c91e99STejun Heo { 336329c91e99STejun Heo if (a->nice != b->nice) 336429c91e99STejun Heo return false; 336529c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 336629c91e99STejun Heo return false; 336729c91e99STejun Heo return true; 336829c91e99STejun Heo } 336929c91e99STejun Heo 33707a4e344cSTejun Heo /** 33717a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33727a4e344cSTejun Heo * @pool: worker_pool to initialize 33737a4e344cSTejun Heo * 33747a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 337529c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 337629c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 337729c91e99STejun Heo * on @pool safely to release it. 33787a4e344cSTejun Heo */ 33797a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33804e1a1f9aSTejun Heo { 33814e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 338229c91e99STejun Heo pool->id = -1; 338329c91e99STejun Heo pool->cpu = -1; 3384f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 33854e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33864e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33874e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33884e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33894e1a1f9aSTejun Heo 33904e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33914e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33924e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33934e1a1f9aSTejun Heo 33944e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33954e1a1f9aSTejun Heo (unsigned long)pool); 33964e1a1f9aSTejun Heo 33974e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3398bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3399822d8405STejun Heo idr_init(&pool->worker_idr); 34007a4e344cSTejun Heo 340129c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 340229c91e99STejun Heo pool->refcnt = 1; 340329c91e99STejun Heo 340429c91e99STejun Heo /* shouldn't fail above this point */ 34057a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34067a4e344cSTejun Heo if (!pool->attrs) 34077a4e344cSTejun Heo return -ENOMEM; 34087a4e344cSTejun Heo return 0; 34094e1a1f9aSTejun Heo } 34104e1a1f9aSTejun Heo 341129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 341229c91e99STejun Heo { 341329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 341429c91e99STejun Heo 3415822d8405STejun Heo idr_destroy(&pool->worker_idr); 341629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 341729c91e99STejun Heo kfree(pool); 341829c91e99STejun Heo } 341929c91e99STejun Heo 342029c91e99STejun Heo /** 342129c91e99STejun Heo * put_unbound_pool - put a worker_pool 342229c91e99STejun Heo * @pool: worker_pool to put 342329c91e99STejun Heo * 342429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3425c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3426c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3427c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3428a892caccSTejun Heo * 3429a892caccSTejun Heo * Should be called with wq_pool_mutex held. 343029c91e99STejun Heo */ 343129c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 343229c91e99STejun Heo { 343329c91e99STejun Heo struct worker *worker; 343429c91e99STejun Heo 3435a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3436a892caccSTejun Heo 3437a892caccSTejun Heo if (--pool->refcnt) 343829c91e99STejun Heo return; 343929c91e99STejun Heo 344029c91e99STejun Heo /* sanity checks */ 344129c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3442a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 344329c91e99STejun Heo return; 344429c91e99STejun Heo 344529c91e99STejun Heo /* release id and unhash */ 344629c91e99STejun Heo if (pool->id >= 0) 344729c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 344829c91e99STejun Heo hash_del(&pool->hash_node); 344929c91e99STejun Heo 3450c5aa87bbSTejun Heo /* 3451c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3452c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3453c5aa87bbSTejun Heo * manager_mutex. 3454c5aa87bbSTejun Heo */ 345529c91e99STejun Heo mutex_lock(&pool->manager_arb); 3456cd549687STejun Heo mutex_lock(&pool->manager_mutex); 345729c91e99STejun Heo spin_lock_irq(&pool->lock); 345829c91e99STejun Heo 345929c91e99STejun Heo while ((worker = first_worker(pool))) 346029c91e99STejun Heo destroy_worker(worker); 346129c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 346229c91e99STejun Heo 346329c91e99STejun Heo spin_unlock_irq(&pool->lock); 3464cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 346529c91e99STejun Heo mutex_unlock(&pool->manager_arb); 346629c91e99STejun Heo 346729c91e99STejun Heo /* shut down the timers */ 346829c91e99STejun Heo del_timer_sync(&pool->idle_timer); 346929c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 347029c91e99STejun Heo 347129c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 347229c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 347329c91e99STejun Heo } 347429c91e99STejun Heo 347529c91e99STejun Heo /** 347629c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 347729c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 347829c91e99STejun Heo * 347929c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 348029c91e99STejun Heo * reference count and return it. If there already is a matching 348129c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 348229c91e99STejun Heo * create a new one. On failure, returns NULL. 3483a892caccSTejun Heo * 3484a892caccSTejun Heo * Should be called with wq_pool_mutex held. 348529c91e99STejun Heo */ 348629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 348729c91e99STejun Heo { 348829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 348929c91e99STejun Heo struct worker_pool *pool; 3490f3f90ad4STejun Heo int node; 349129c91e99STejun Heo 3492a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 349329c91e99STejun Heo 349429c91e99STejun Heo /* do we already have a matching pool? */ 349529c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 349629c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 349729c91e99STejun Heo pool->refcnt++; 349829c91e99STejun Heo goto out_unlock; 349929c91e99STejun Heo } 350029c91e99STejun Heo } 350129c91e99STejun Heo 350229c91e99STejun Heo /* nope, create a new one */ 350329c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 350429c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 350529c91e99STejun Heo goto fail; 350629c91e99STejun Heo 350712ee4fc6SLai Jiangshan if (workqueue_freezing) 350812ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 350912ee4fc6SLai Jiangshan 35108864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 351129c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 351229c91e99STejun Heo 3513f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3514f3f90ad4STejun Heo if (wq_numa_enabled) { 3515f3f90ad4STejun Heo for_each_node(node) { 3516f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3517f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3518f3f90ad4STejun Heo pool->node = node; 3519f3f90ad4STejun Heo break; 3520f3f90ad4STejun Heo } 3521f3f90ad4STejun Heo } 3522f3f90ad4STejun Heo } 3523f3f90ad4STejun Heo 352429c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 352529c91e99STejun Heo goto fail; 352629c91e99STejun Heo 352729c91e99STejun Heo /* create and start the initial worker */ 3528ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 352929c91e99STejun Heo goto fail; 353029c91e99STejun Heo 353129c91e99STejun Heo /* install */ 353229c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 353329c91e99STejun Heo out_unlock: 353429c91e99STejun Heo return pool; 353529c91e99STejun Heo fail: 353629c91e99STejun Heo if (pool) 353729c91e99STejun Heo put_unbound_pool(pool); 353829c91e99STejun Heo return NULL; 353929c91e99STejun Heo } 354029c91e99STejun Heo 35418864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35428864b4e5STejun Heo { 35438864b4e5STejun Heo kmem_cache_free(pwq_cache, 35448864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35458864b4e5STejun Heo } 35468864b4e5STejun Heo 35478864b4e5STejun Heo /* 35488864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35498864b4e5STejun Heo * and needs to be destroyed. 35508864b4e5STejun Heo */ 35518864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35528864b4e5STejun Heo { 35538864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35548864b4e5STejun Heo unbound_release_work); 35558864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35568864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3557bc0caf09STejun Heo bool is_last; 35588864b4e5STejun Heo 35598864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35608864b4e5STejun Heo return; 35618864b4e5STejun Heo 356275ccf595STejun Heo /* 35633c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 356475ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 356575ccf595STejun Heo * and consistent with the linking path. 356675ccf595STejun Heo */ 35673c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35688864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3569bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35703c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35718864b4e5STejun Heo 3572a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35738864b4e5STejun Heo put_unbound_pool(pool); 3574a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3575a892caccSTejun Heo 35768864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35778864b4e5STejun Heo 35788864b4e5STejun Heo /* 35798864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35808864b4e5STejun Heo * is gonna access it anymore. Free it. 35818864b4e5STejun Heo */ 35826029a918STejun Heo if (is_last) { 35836029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 35848864b4e5STejun Heo kfree(wq); 35858864b4e5STejun Heo } 35866029a918STejun Heo } 35878864b4e5STejun Heo 35880fbd95aaSTejun Heo /** 3589699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35900fbd95aaSTejun Heo * @pwq: target pool_workqueue 35910fbd95aaSTejun Heo * 3592699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3593699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3594699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35950fbd95aaSTejun Heo */ 3596699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35970fbd95aaSTejun Heo { 3598699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3599699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3600699ce097STejun Heo 3601699ce097STejun Heo /* for @wq->saved_max_active */ 3602a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3603699ce097STejun Heo 3604699ce097STejun Heo /* fast exit for non-freezable wqs */ 3605699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3606699ce097STejun Heo return; 3607699ce097STejun Heo 3608a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3609699ce097STejun Heo 3610699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3611699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36120fbd95aaSTejun Heo 36130fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36140fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36150fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3616951a078aSLai Jiangshan 3617951a078aSLai Jiangshan /* 3618951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3619951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3620951a078aSLai Jiangshan */ 3621951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3622699ce097STejun Heo } else { 3623699ce097STejun Heo pwq->max_active = 0; 3624699ce097STejun Heo } 3625699ce097STejun Heo 3626a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36270fbd95aaSTejun Heo } 36280fbd95aaSTejun Heo 3629e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3630f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3631f147f29eSTejun Heo struct worker_pool *pool) 3632d2c1d404STejun Heo { 3633d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3634d2c1d404STejun Heo 3635e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3636e50aba9aSTejun Heo 3637d2c1d404STejun Heo pwq->pool = pool; 3638d2c1d404STejun Heo pwq->wq = wq; 3639d2c1d404STejun Heo pwq->flush_color = -1; 36408864b4e5STejun Heo pwq->refcnt = 1; 3641d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36421befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3643d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36448864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3645f147f29eSTejun Heo } 3646d2c1d404STejun Heo 3647f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 36481befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3649f147f29eSTejun Heo { 3650f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3651f147f29eSTejun Heo 3652f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 365375ccf595STejun Heo 36541befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 36551befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 36561befcf30STejun Heo return; 36571befcf30STejun Heo 3658983ca25eSTejun Heo /* 3659983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36603c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3661983ca25eSTejun Heo */ 366275ccf595STejun Heo pwq->work_color = wq->work_color; 3663983ca25eSTejun Heo 3664983ca25eSTejun Heo /* sync max_active to the current setting */ 3665983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3666983ca25eSTejun Heo 3667983ca25eSTejun Heo /* link in @pwq */ 36689e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3669df2d5ae4STejun Heo } 36706029a918STejun Heo 3671f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3672f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3673f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3674f147f29eSTejun Heo { 3675f147f29eSTejun Heo struct worker_pool *pool; 3676f147f29eSTejun Heo struct pool_workqueue *pwq; 3677f147f29eSTejun Heo 3678f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3679f147f29eSTejun Heo 3680f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3681f147f29eSTejun Heo if (!pool) 3682f147f29eSTejun Heo return NULL; 3683f147f29eSTejun Heo 3684e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3685f147f29eSTejun Heo if (!pwq) { 3686f147f29eSTejun Heo put_unbound_pool(pool); 3687f147f29eSTejun Heo return NULL; 3688f147f29eSTejun Heo } 3689f147f29eSTejun Heo 3690f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3691f147f29eSTejun Heo return pwq; 3692d2c1d404STejun Heo } 3693d2c1d404STejun Heo 36941befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 36951befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 36961befcf30STejun Heo int node, 36971befcf30STejun Heo struct pool_workqueue *pwq) 36981befcf30STejun Heo { 36991befcf30STejun Heo struct pool_workqueue *old_pwq; 37001befcf30STejun Heo 37011befcf30STejun Heo lockdep_assert_held(&wq->mutex); 37021befcf30STejun Heo 37031befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 37041befcf30STejun Heo link_pwq(pwq); 37051befcf30STejun Heo 37061befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 37071befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 37081befcf30STejun Heo return old_pwq; 37091befcf30STejun Heo } 37101befcf30STejun Heo 37119e8cd2f5STejun Heo /** 37129e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37139e8cd2f5STejun Heo * @wq: the target workqueue 37149e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37159e8cd2f5STejun Heo * 37169e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 37179e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 37189e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 37199e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 37209e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 37219e8cd2f5STejun Heo * 37229e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 37239e8cd2f5STejun Heo * failure. 37249e8cd2f5STejun Heo */ 37259e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 37269e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 37279e8cd2f5STejun Heo { 372813e2e556STejun Heo struct workqueue_attrs *new_attrs; 37291befcf30STejun Heo struct pool_workqueue *pwq, *last_pwq = NULL; 3730f147f29eSTejun Heo int node, ret; 37319e8cd2f5STejun Heo 37328719dceaSTejun Heo /* only unbound workqueues can change attributes */ 37339e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 37349e8cd2f5STejun Heo return -EINVAL; 37359e8cd2f5STejun Heo 37368719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 37378719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 37388719dceaSTejun Heo return -EINVAL; 37398719dceaSTejun Heo 374013e2e556STejun Heo /* make a copy of @attrs and sanitize it */ 374113e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 374213e2e556STejun Heo if (!new_attrs) 374313e2e556STejun Heo goto enomem; 374413e2e556STejun Heo 374513e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 374613e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 374713e2e556STejun Heo 3748a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 3749f147f29eSTejun Heo pwq = alloc_unbound_pwq(wq, new_attrs); 3750a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3751f147f29eSTejun Heo if (!pwq) 375213e2e556STejun Heo goto enomem; 37539e8cd2f5STejun Heo 3754f147f29eSTejun Heo mutex_lock(&wq->mutex); 3755a892caccSTejun Heo 3756f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 3757f147f29eSTejun Heo for_each_node(node) 37581befcf30STejun Heo last_pwq = numa_pwq_tbl_install(wq, node, pwq); 3759f147f29eSTejun Heo 3760f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3761f147f29eSTejun Heo 37629e8cd2f5STejun Heo if (last_pwq) { 37639e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 37649e8cd2f5STejun Heo put_pwq(last_pwq); 37659e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 37669e8cd2f5STejun Heo } 37679e8cd2f5STejun Heo 37684862125bSTejun Heo ret = 0; 37694862125bSTejun Heo /* fall through */ 37704862125bSTejun Heo out_free: 37714862125bSTejun Heo free_workqueue_attrs(new_attrs); 37724862125bSTejun Heo return ret; 377313e2e556STejun Heo 377413e2e556STejun Heo enomem: 37754862125bSTejun Heo ret = -ENOMEM; 37764862125bSTejun Heo goto out_free; 37779e8cd2f5STejun Heo } 37789e8cd2f5STejun Heo 377930cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37801da177e4SLinus Torvalds { 378149e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 378230cdf249STejun Heo int cpu; 3783e1d8aa9fSFrederic Weisbecker 378430cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3785420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3786420c0ddbSTejun Heo if (!wq->cpu_pwqs) 378730cdf249STejun Heo return -ENOMEM; 378830cdf249STejun Heo 378930cdf249STejun Heo for_each_possible_cpu(cpu) { 37907fb98ea7STejun Heo struct pool_workqueue *pwq = 37917fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37927a62c2c8STejun Heo struct worker_pool *cpu_pools = 3793f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 379430cdf249STejun Heo 3795f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3796f147f29eSTejun Heo 3797f147f29eSTejun Heo mutex_lock(&wq->mutex); 37981befcf30STejun Heo link_pwq(pwq); 3799f147f29eSTejun Heo mutex_unlock(&wq->mutex); 380030cdf249STejun Heo } 380130cdf249STejun Heo return 0; 38029e8cd2f5STejun Heo } else { 38039e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 38049e8cd2f5STejun Heo } 38050f900049STejun Heo } 38060f900049STejun Heo 3807f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3808f3421797STejun Heo const char *name) 3809b71ab8c2STejun Heo { 3810f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3811f3421797STejun Heo 3812f3421797STejun Heo if (max_active < 1 || max_active > lim) 3813044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3814f3421797STejun Heo max_active, name, 1, lim); 3815b71ab8c2STejun Heo 3816f3421797STejun Heo return clamp_val(max_active, 1, lim); 3817b71ab8c2STejun Heo } 3818b71ab8c2STejun Heo 3819b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 382097e37d7bSTejun Heo unsigned int flags, 38211e19ffc6STejun Heo int max_active, 3822eb13ba87SJohannes Berg struct lock_class_key *key, 3823b196be89STejun Heo const char *lock_name, ...) 38243af24433SOleg Nesterov { 3825df2d5ae4STejun Heo size_t tbl_size = 0; 3826ecf6881fSTejun Heo va_list args; 38273af24433SOleg Nesterov struct workqueue_struct *wq; 382849e3cf44STejun Heo struct pool_workqueue *pwq; 3829b196be89STejun Heo 3830ecf6881fSTejun Heo /* allocate wq and format name */ 3831df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 3832df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 3833df2d5ae4STejun Heo 3834df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 3835b196be89STejun Heo if (!wq) 3836d2c1d404STejun Heo return NULL; 3837b196be89STejun Heo 38386029a918STejun Heo if (flags & WQ_UNBOUND) { 38396029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38406029a918STejun Heo if (!wq->unbound_attrs) 38416029a918STejun Heo goto err_free_wq; 38426029a918STejun Heo } 38436029a918STejun Heo 3844ecf6881fSTejun Heo va_start(args, lock_name); 3845ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3846b196be89STejun Heo va_end(args); 38473af24433SOleg Nesterov 3848d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3849b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 38503af24433SOleg Nesterov 3851b196be89STejun Heo /* init wq */ 385297e37d7bSTejun Heo wq->flags = flags; 3853a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 38543c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3855112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 385630cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 385773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 385873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3859493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 38603af24433SOleg Nesterov 3861eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3862cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 38633af24433SOleg Nesterov 386430cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3865d2c1d404STejun Heo goto err_free_wq; 38661537663fSTejun Heo 3867493008a8STejun Heo /* 3868493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3869493008a8STejun Heo * have a rescuer to guarantee forward progress. 3870493008a8STejun Heo */ 3871493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3872e22bee78STejun Heo struct worker *rescuer; 3873e22bee78STejun Heo 3874d2c1d404STejun Heo rescuer = alloc_worker(); 3875e22bee78STejun Heo if (!rescuer) 3876d2c1d404STejun Heo goto err_destroy; 3877e22bee78STejun Heo 3878111c225aSTejun Heo rescuer->rescue_wq = wq; 3879111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3880b196be89STejun Heo wq->name); 3881d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3882d2c1d404STejun Heo kfree(rescuer); 3883d2c1d404STejun Heo goto err_destroy; 3884d2c1d404STejun Heo } 3885e22bee78STejun Heo 3886d2c1d404STejun Heo wq->rescuer = rescuer; 388714a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3888e22bee78STejun Heo wake_up_process(rescuer->task); 38893af24433SOleg Nesterov } 38901537663fSTejun Heo 3891226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3892226223abSTejun Heo goto err_destroy; 3893226223abSTejun Heo 38943af24433SOleg Nesterov /* 389568e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 389668e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 389768e13a67SLai Jiangshan * list. 38983af24433SOleg Nesterov */ 389968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3900a0a1a5fdSTejun Heo 3901a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 390249e3cf44STejun Heo for_each_pwq(pwq, wq) 3903699ce097STejun Heo pwq_adjust_max_active(pwq); 3904a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3905a0a1a5fdSTejun Heo 39063af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3907a0a1a5fdSTejun Heo 390868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39093af24433SOleg Nesterov 39103af24433SOleg Nesterov return wq; 3911d2c1d404STejun Heo 3912d2c1d404STejun Heo err_free_wq: 39136029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 39144690c4abSTejun Heo kfree(wq); 3915d2c1d404STejun Heo return NULL; 3916d2c1d404STejun Heo err_destroy: 3917d2c1d404STejun Heo destroy_workqueue(wq); 39184690c4abSTejun Heo return NULL; 39191da177e4SLinus Torvalds } 3920d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 39211da177e4SLinus Torvalds 39223af24433SOleg Nesterov /** 39233af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 39243af24433SOleg Nesterov * @wq: target workqueue 39253af24433SOleg Nesterov * 39263af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 39273af24433SOleg Nesterov */ 39283af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 39293af24433SOleg Nesterov { 393049e3cf44STejun Heo struct pool_workqueue *pwq; 39313af24433SOleg Nesterov 39329c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 39339c5a2ba7STejun Heo drain_workqueue(wq); 3934c8efcc25STejun Heo 39356183c009STejun Heo /* sanity checks */ 3936b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 393749e3cf44STejun Heo for_each_pwq(pwq, wq) { 39386183c009STejun Heo int i; 39396183c009STejun Heo 394076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 394176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3942b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39436183c009STejun Heo return; 394476af4d93STejun Heo } 394576af4d93STejun Heo } 394676af4d93STejun Heo 39478864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 39488864b4e5STejun Heo WARN_ON(pwq->nr_active) || 394976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3950b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39516183c009STejun Heo return; 39526183c009STejun Heo } 395376af4d93STejun Heo } 3954b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 39556183c009STejun Heo 3956a0a1a5fdSTejun Heo /* 3957a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3958a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3959a0a1a5fdSTejun Heo */ 396068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3961d2c1d404STejun Heo list_del_init(&wq->list); 396268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 39633af24433SOleg Nesterov 3964226223abSTejun Heo workqueue_sysfs_unregister(wq); 3965226223abSTejun Heo 3966493008a8STejun Heo if (wq->rescuer) { 3967e22bee78STejun Heo kthread_stop(wq->rescuer->task); 39688d9df9f0SXiaotian Feng kfree(wq->rescuer); 3969493008a8STejun Heo wq->rescuer = NULL; 3970e22bee78STejun Heo } 3971e22bee78STejun Heo 39728864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 397329c91e99STejun Heo /* 39748864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 39758864b4e5STejun Heo * free the pwqs and wq. 397629c91e99STejun Heo */ 39778864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 39788864b4e5STejun Heo kfree(wq); 39798864b4e5STejun Heo } else { 39808864b4e5STejun Heo /* 39818864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39828864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 39838864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 39848864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 39858864b4e5STejun Heo * released. 39868864b4e5STejun Heo */ 398729c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 398829c91e99STejun Heo pwqs_node); 39898864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 39908864b4e5STejun Heo put_pwq(pwq); 39918864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 399229c91e99STejun Heo } 39933af24433SOleg Nesterov } 39943af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39953af24433SOleg Nesterov 3996dcd989cbSTejun Heo /** 3997dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3998dcd989cbSTejun Heo * @wq: target workqueue 3999dcd989cbSTejun Heo * @max_active: new max_active value. 4000dcd989cbSTejun Heo * 4001dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4002dcd989cbSTejun Heo * 4003dcd989cbSTejun Heo * CONTEXT: 4004dcd989cbSTejun Heo * Don't call from IRQ context. 4005dcd989cbSTejun Heo */ 4006dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4007dcd989cbSTejun Heo { 400849e3cf44STejun Heo struct pool_workqueue *pwq; 4009dcd989cbSTejun Heo 40108719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 40118719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 40128719dceaSTejun Heo return; 40138719dceaSTejun Heo 4014f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4015dcd989cbSTejun Heo 4016a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4017dcd989cbSTejun Heo 4018dcd989cbSTejun Heo wq->saved_max_active = max_active; 4019dcd989cbSTejun Heo 4020699ce097STejun Heo for_each_pwq(pwq, wq) 4021699ce097STejun Heo pwq_adjust_max_active(pwq); 4022dcd989cbSTejun Heo 4023a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4024dcd989cbSTejun Heo } 4025dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4026dcd989cbSTejun Heo 4027dcd989cbSTejun Heo /** 4028e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4029e6267616STejun Heo * 4030e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4031e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4032e6267616STejun Heo */ 4033e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4034e6267616STejun Heo { 4035e6267616STejun Heo struct worker *worker = current_wq_worker(); 4036e6267616STejun Heo 40376a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4038e6267616STejun Heo } 4039e6267616STejun Heo 4040e6267616STejun Heo /** 4041dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4042dcd989cbSTejun Heo * @cpu: CPU in question 4043dcd989cbSTejun Heo * @wq: target workqueue 4044dcd989cbSTejun Heo * 4045dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4046dcd989cbSTejun Heo * no synchronization around this function and the test result is 4047dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4048dcd989cbSTejun Heo * 4049dcd989cbSTejun Heo * RETURNS: 4050dcd989cbSTejun Heo * %true if congested, %false otherwise. 4051dcd989cbSTejun Heo */ 4052d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4053dcd989cbSTejun Heo { 40547fb98ea7STejun Heo struct pool_workqueue *pwq; 405576af4d93STejun Heo bool ret; 405676af4d93STejun Heo 405788109453SLai Jiangshan rcu_read_lock_sched(); 40587fb98ea7STejun Heo 40597fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 40607fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 40617fb98ea7STejun Heo else 4062df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4063dcd989cbSTejun Heo 406476af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 406588109453SLai Jiangshan rcu_read_unlock_sched(); 406676af4d93STejun Heo 406776af4d93STejun Heo return ret; 4068dcd989cbSTejun Heo } 4069dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4070dcd989cbSTejun Heo 4071dcd989cbSTejun Heo /** 4072dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4073dcd989cbSTejun Heo * @work: the work to be tested 4074dcd989cbSTejun Heo * 4075dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4076dcd989cbSTejun Heo * synchronization around this function and the test result is 4077dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4078dcd989cbSTejun Heo * 4079dcd989cbSTejun Heo * RETURNS: 4080dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4081dcd989cbSTejun Heo */ 4082dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4083dcd989cbSTejun Heo { 4084fa1b54e6STejun Heo struct worker_pool *pool; 4085dcd989cbSTejun Heo unsigned long flags; 4086dcd989cbSTejun Heo unsigned int ret = 0; 4087dcd989cbSTejun Heo 4088dcd989cbSTejun Heo if (work_pending(work)) 4089dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4090038366c5SLai Jiangshan 4091fa1b54e6STejun Heo local_irq_save(flags); 4092fa1b54e6STejun Heo pool = get_work_pool(work); 4093038366c5SLai Jiangshan if (pool) { 4094fa1b54e6STejun Heo spin_lock(&pool->lock); 4095c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4096dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4097fa1b54e6STejun Heo spin_unlock(&pool->lock); 4098038366c5SLai Jiangshan } 4099fa1b54e6STejun Heo local_irq_restore(flags); 4100dcd989cbSTejun Heo 4101dcd989cbSTejun Heo return ret; 4102dcd989cbSTejun Heo } 4103dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4104dcd989cbSTejun Heo 4105db7bccf4STejun Heo /* 4106db7bccf4STejun Heo * CPU hotplug. 4107db7bccf4STejun Heo * 4108e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4109112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4110706026c2STejun Heo * pool which make migrating pending and scheduled works very 4111e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 411294cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4113e22bee78STejun Heo * blocked draining impractical. 4114e22bee78STejun Heo * 411524647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4116628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4117628c78e7STejun Heo * cpu comes back online. 4118db7bccf4STejun Heo */ 4119db7bccf4STejun Heo 4120706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4121db7bccf4STejun Heo { 412238db41d9STejun Heo int cpu = smp_processor_id(); 41234ce62e9eSTejun Heo struct worker_pool *pool; 4124db7bccf4STejun Heo struct worker *worker; 4125a9ab775bSTejun Heo int wi; 4126db7bccf4STejun Heo 4127f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41286183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4129db7bccf4STejun Heo 4130bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 413194cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4132e22bee78STejun Heo 4133f2d5a0eeSTejun Heo /* 4134bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 413594cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 413694cf58bbSTejun Heo * except for the ones which are still executing works from 413794cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 413894cf58bbSTejun Heo * this, they may become diasporas. 4139f2d5a0eeSTejun Heo */ 4140a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4141403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4142db7bccf4STejun Heo 414324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4144f2d5a0eeSTejun Heo 414594cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4146bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 414794cf58bbSTejun Heo } 4148e22bee78STejun Heo 4149e22bee78STejun Heo /* 4150628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4151628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4152628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4153628c78e7STejun Heo */ 4154628c78e7STejun Heo schedule(); 4155628c78e7STejun Heo 4156628c78e7STejun Heo /* 4157628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4158628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 415938db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 416038db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 416138db41d9STejun Heo * pools which are served by workers tied to the CPU. 4162628c78e7STejun Heo * 4163628c78e7STejun Heo * On return from this function, the current worker would trigger 4164628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4165628c78e7STejun Heo * didn't already. 4166e22bee78STejun Heo */ 4167f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4168e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4169db7bccf4STejun Heo } 4170db7bccf4STejun Heo 4171bd7c089eSTejun Heo /** 4172bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4173bd7c089eSTejun Heo * @pool: pool of interest 4174bd7c089eSTejun Heo * 4175a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4176bd7c089eSTejun Heo */ 4177bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4178bd7c089eSTejun Heo { 4179a9ab775bSTejun Heo struct worker *worker; 4180a9ab775bSTejun Heo int wi; 4181bd7c089eSTejun Heo 4182bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4183bd7c089eSTejun Heo 4184bd7c089eSTejun Heo /* 4185a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4186a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4187a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4188a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4189a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4190bd7c089eSTejun Heo */ 4191a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4192a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4193a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4194a9ab775bSTejun Heo 4195a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4196a9ab775bSTejun Heo 4197a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4198a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4199a9ab775bSTejun Heo 4200a9ab775bSTejun Heo /* 4201a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4202a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4203a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4204a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4205a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4206a9ab775bSTejun Heo * be bound before @pool->lock is released. 4207a9ab775bSTejun Heo */ 4208a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4209bd7c089eSTejun Heo wake_up_process(worker->task); 4210bd7c089eSTejun Heo 4211bd7c089eSTejun Heo /* 4212a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4213a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4214a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4215a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4216a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4217a9ab775bSTejun Heo * concurrency management. Note that when or whether 4218a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4219a9ab775bSTejun Heo * 4220a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4221a9ab775bSTejun Heo * tested without holding any lock in 4222a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4223a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4224a9ab775bSTejun Heo * management operations. 4225bd7c089eSTejun Heo */ 4226a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4227a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4228a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4229a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4230bd7c089eSTejun Heo } 4231a9ab775bSTejun Heo 4232a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4233bd7c089eSTejun Heo } 4234bd7c089eSTejun Heo 42357dbc725eSTejun Heo /** 42367dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 42377dbc725eSTejun Heo * @pool: unbound pool of interest 42387dbc725eSTejun Heo * @cpu: the CPU which is coming up 42397dbc725eSTejun Heo * 42407dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 42417dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 42427dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 42437dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 42447dbc725eSTejun Heo */ 42457dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 42467dbc725eSTejun Heo { 42477dbc725eSTejun Heo static cpumask_t cpumask; 42487dbc725eSTejun Heo struct worker *worker; 42497dbc725eSTejun Heo int wi; 42507dbc725eSTejun Heo 42517dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 42527dbc725eSTejun Heo 42537dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 42547dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 42557dbc725eSTejun Heo return; 42567dbc725eSTejun Heo 42577dbc725eSTejun Heo /* is @cpu the only online CPU? */ 42587dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 42597dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 42607dbc725eSTejun Heo return; 42617dbc725eSTejun Heo 42627dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 42637dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 42647dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 42657dbc725eSTejun Heo pool->attrs->cpumask) < 0); 42667dbc725eSTejun Heo } 42677dbc725eSTejun Heo 42688db25e78STejun Heo /* 42698db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 42708db25e78STejun Heo * This will be registered high priority CPU notifier. 42718db25e78STejun Heo */ 42729fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 42731da177e4SLinus Torvalds unsigned long action, 42741da177e4SLinus Torvalds void *hcpu) 42751da177e4SLinus Torvalds { 4276d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42774ce62e9eSTejun Heo struct worker_pool *pool; 42787dbc725eSTejun Heo int pi; 42791da177e4SLinus Torvalds 42808db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 42813af24433SOleg Nesterov case CPU_UP_PREPARE: 4282f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 42833ce63377STejun Heo if (pool->nr_workers) 42843ce63377STejun Heo continue; 4285ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 42863ce63377STejun Heo return NOTIFY_BAD; 42873af24433SOleg Nesterov } 42881da177e4SLinus Torvalds break; 42891da177e4SLinus Torvalds 429065758202STejun Heo case CPU_DOWN_FAILED: 429165758202STejun Heo case CPU_ONLINE: 429268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 42937dbc725eSTejun Heo 42947dbc725eSTejun Heo for_each_pool(pool, pi) { 4295bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 429694cf58bbSTejun Heo 42977dbc725eSTejun Heo if (pool->cpu == cpu) { 4298a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 429924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4300a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4301a9ab775bSTejun Heo 430294cf58bbSTejun Heo rebind_workers(pool); 43037dbc725eSTejun Heo } else if (pool->cpu < 0) { 43047dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 43057dbc725eSTejun Heo } 430694cf58bbSTejun Heo 4307bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 430894cf58bbSTejun Heo } 43097dbc725eSTejun Heo 431068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43118db25e78STejun Heo break; 431265758202STejun Heo } 431365758202STejun Heo return NOTIFY_OK; 431465758202STejun Heo } 431565758202STejun Heo 431665758202STejun Heo /* 431765758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 431865758202STejun Heo * This will be registered as low priority CPU notifier. 431965758202STejun Heo */ 43209fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 432165758202STejun Heo unsigned long action, 432265758202STejun Heo void *hcpu) 432365758202STejun Heo { 4324d84ff051STejun Heo int cpu = (unsigned long)hcpu; 43258db25e78STejun Heo struct work_struct unbind_work; 43268db25e78STejun Heo 432765758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 432865758202STejun Heo case CPU_DOWN_PREPARE: 43298db25e78STejun Heo /* unbinding should happen on the local CPU */ 4330706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 43317635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 43328db25e78STejun Heo flush_work(&unbind_work); 43338db25e78STejun Heo break; 433465758202STejun Heo } 433565758202STejun Heo return NOTIFY_OK; 433665758202STejun Heo } 433765758202STejun Heo 43382d3854a3SRusty Russell #ifdef CONFIG_SMP 43398ccad40dSRusty Russell 43402d3854a3SRusty Russell struct work_for_cpu { 4341ed48ece2STejun Heo struct work_struct work; 43422d3854a3SRusty Russell long (*fn)(void *); 43432d3854a3SRusty Russell void *arg; 43442d3854a3SRusty Russell long ret; 43452d3854a3SRusty Russell }; 43462d3854a3SRusty Russell 4347ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 43482d3854a3SRusty Russell { 4349ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4350ed48ece2STejun Heo 43512d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 43522d3854a3SRusty Russell } 43532d3854a3SRusty Russell 43542d3854a3SRusty Russell /** 43552d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 43562d3854a3SRusty Russell * @cpu: the cpu to run on 43572d3854a3SRusty Russell * @fn: the function to run 43582d3854a3SRusty Russell * @arg: the function arg 43592d3854a3SRusty Russell * 436031ad9081SRusty Russell * This will return the value @fn returns. 436131ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 43626b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 43632d3854a3SRusty Russell */ 4364d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 43652d3854a3SRusty Russell { 4366ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 43672d3854a3SRusty Russell 4368ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4369ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4370ed48ece2STejun Heo flush_work(&wfc.work); 43712d3854a3SRusty Russell return wfc.ret; 43722d3854a3SRusty Russell } 43732d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 43742d3854a3SRusty Russell #endif /* CONFIG_SMP */ 43752d3854a3SRusty Russell 4376a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4377e7577c50SRusty Russell 4378a0a1a5fdSTejun Heo /** 4379a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4380a0a1a5fdSTejun Heo * 438158a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4382c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4383706026c2STejun Heo * pool->worklist. 4384a0a1a5fdSTejun Heo * 4385a0a1a5fdSTejun Heo * CONTEXT: 4386a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4387a0a1a5fdSTejun Heo */ 4388a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4389a0a1a5fdSTejun Heo { 439017116969STejun Heo struct worker_pool *pool; 439124b8a847STejun Heo struct workqueue_struct *wq; 439224b8a847STejun Heo struct pool_workqueue *pwq; 4393611c92a0STejun Heo int pi; 4394a0a1a5fdSTejun Heo 439568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4396a0a1a5fdSTejun Heo 43976183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4398a0a1a5fdSTejun Heo workqueue_freezing = true; 4399a0a1a5fdSTejun Heo 440024b8a847STejun Heo /* set FREEZING */ 4401611c92a0STejun Heo for_each_pool(pool, pi) { 44025bcab335STejun Heo spin_lock_irq(&pool->lock); 440335b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 440435b6bb63STejun Heo pool->flags |= POOL_FREEZING; 44055bcab335STejun Heo spin_unlock_irq(&pool->lock); 44061da177e4SLinus Torvalds } 44078b03ae3cSTejun Heo 440824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4409a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4410699ce097STejun Heo for_each_pwq(pwq, wq) 4411699ce097STejun Heo pwq_adjust_max_active(pwq); 4412a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4413a1056305STejun Heo } 44145bcab335STejun Heo 441568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4416a0a1a5fdSTejun Heo } 4417a0a1a5fdSTejun Heo 4418a0a1a5fdSTejun Heo /** 441958a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4420a0a1a5fdSTejun Heo * 4421a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4422a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4423a0a1a5fdSTejun Heo * 4424a0a1a5fdSTejun Heo * CONTEXT: 442568e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4426a0a1a5fdSTejun Heo * 4427a0a1a5fdSTejun Heo * RETURNS: 442858a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 442958a69cb4STejun Heo * is complete. 4430a0a1a5fdSTejun Heo */ 4431a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4432a0a1a5fdSTejun Heo { 4433a0a1a5fdSTejun Heo bool busy = false; 443424b8a847STejun Heo struct workqueue_struct *wq; 443524b8a847STejun Heo struct pool_workqueue *pwq; 4436a0a1a5fdSTejun Heo 443768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4438a0a1a5fdSTejun Heo 44396183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4440a0a1a5fdSTejun Heo 444124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 444224b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 444324b8a847STejun Heo continue; 4444a0a1a5fdSTejun Heo /* 4445a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4446a0a1a5fdSTejun Heo * to peek without lock. 4447a0a1a5fdSTejun Heo */ 444888109453SLai Jiangshan rcu_read_lock_sched(); 444924b8a847STejun Heo for_each_pwq(pwq, wq) { 44506183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4451112202d9STejun Heo if (pwq->nr_active) { 4452a0a1a5fdSTejun Heo busy = true; 445388109453SLai Jiangshan rcu_read_unlock_sched(); 4454a0a1a5fdSTejun Heo goto out_unlock; 4455a0a1a5fdSTejun Heo } 4456a0a1a5fdSTejun Heo } 445788109453SLai Jiangshan rcu_read_unlock_sched(); 4458a0a1a5fdSTejun Heo } 4459a0a1a5fdSTejun Heo out_unlock: 446068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4461a0a1a5fdSTejun Heo return busy; 4462a0a1a5fdSTejun Heo } 4463a0a1a5fdSTejun Heo 4464a0a1a5fdSTejun Heo /** 4465a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4466a0a1a5fdSTejun Heo * 4467a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4468706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4469a0a1a5fdSTejun Heo * 4470a0a1a5fdSTejun Heo * CONTEXT: 4471a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4472a0a1a5fdSTejun Heo */ 4473a0a1a5fdSTejun Heo void thaw_workqueues(void) 4474a0a1a5fdSTejun Heo { 447524b8a847STejun Heo struct workqueue_struct *wq; 447624b8a847STejun Heo struct pool_workqueue *pwq; 447724b8a847STejun Heo struct worker_pool *pool; 4478611c92a0STejun Heo int pi; 4479a0a1a5fdSTejun Heo 448068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4481a0a1a5fdSTejun Heo 4482a0a1a5fdSTejun Heo if (!workqueue_freezing) 4483a0a1a5fdSTejun Heo goto out_unlock; 4484a0a1a5fdSTejun Heo 448524b8a847STejun Heo /* clear FREEZING */ 4486611c92a0STejun Heo for_each_pool(pool, pi) { 44875bcab335STejun Heo spin_lock_irq(&pool->lock); 448835b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 448935b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 44905bcab335STejun Heo spin_unlock_irq(&pool->lock); 4491d565ed63STejun Heo } 449224b8a847STejun Heo 449324b8a847STejun Heo /* restore max_active and repopulate worklist */ 449424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4495a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4496699ce097STejun Heo for_each_pwq(pwq, wq) 4497699ce097STejun Heo pwq_adjust_max_active(pwq); 4498a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 449924b8a847STejun Heo } 450024b8a847STejun Heo 4501a0a1a5fdSTejun Heo workqueue_freezing = false; 4502a0a1a5fdSTejun Heo out_unlock: 450368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4504a0a1a5fdSTejun Heo } 4505a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4506a0a1a5fdSTejun Heo 4507bce90380STejun Heo static void __init wq_numa_init(void) 4508bce90380STejun Heo { 4509bce90380STejun Heo cpumask_var_t *tbl; 4510bce90380STejun Heo int node, cpu; 4511bce90380STejun Heo 4512bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4513bce90380STejun Heo for_each_node(node) 4514bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4515bce90380STejun Heo 4516bce90380STejun Heo if (num_possible_nodes() <= 1) 4517bce90380STejun Heo return; 4518bce90380STejun Heo 4519bce90380STejun Heo /* 4520bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4521bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4522bce90380STejun Heo * fully initialized by now. 4523bce90380STejun Heo */ 4524bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4525bce90380STejun Heo BUG_ON(!tbl); 4526bce90380STejun Heo 4527bce90380STejun Heo for_each_node(node) 4528bce90380STejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node)); 4529bce90380STejun Heo 4530bce90380STejun Heo for_each_possible_cpu(cpu) { 4531bce90380STejun Heo node = cpu_to_node(cpu); 4532bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4533bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4534bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4535bce90380STejun Heo return; 4536bce90380STejun Heo } 4537bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4538bce90380STejun Heo } 4539bce90380STejun Heo 4540bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4541bce90380STejun Heo wq_numa_enabled = true; 4542bce90380STejun Heo } 4543bce90380STejun Heo 45446ee0578bSSuresh Siddha static int __init init_workqueues(void) 45451da177e4SLinus Torvalds { 45467a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 45477a4e344cSTejun Heo int i, cpu; 4548c34056a3STejun Heo 45497c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 45507c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 45516be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4552b5490077STejun Heo 4553e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4554e904e6c2STejun Heo 4555e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4556e904e6c2STejun Heo 455765758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4558a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 45598b03ae3cSTejun Heo 4560bce90380STejun Heo wq_numa_init(); 4561bce90380STejun Heo 4562706026c2STejun Heo /* initialize CPU pools */ 456329c91e99STejun Heo for_each_possible_cpu(cpu) { 45644ce62e9eSTejun Heo struct worker_pool *pool; 45658b03ae3cSTejun Heo 45667a4e344cSTejun Heo i = 0; 4567f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45687a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4569ec22ca5eSTejun Heo pool->cpu = cpu; 45707a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 45717a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4572f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 45737a4e344cSTejun Heo 45749daf9e67STejun Heo /* alloc pool ID */ 457568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45769daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 457768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45784ce62e9eSTejun Heo } 45798b03ae3cSTejun Heo } 45808b03ae3cSTejun Heo 4581e22bee78STejun Heo /* create the initial worker */ 458229c91e99STejun Heo for_each_online_cpu(cpu) { 45834ce62e9eSTejun Heo struct worker_pool *pool; 4584e22bee78STejun Heo 4585f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 458624647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4587ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4588e22bee78STejun Heo } 45894ce62e9eSTejun Heo } 4590e22bee78STejun Heo 459129c91e99STejun Heo /* create default unbound wq attrs */ 459229c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 459329c91e99STejun Heo struct workqueue_attrs *attrs; 459429c91e99STejun Heo 459529c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 459629c91e99STejun Heo attrs->nice = std_nice[i]; 459729c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 459829c91e99STejun Heo } 459929c91e99STejun Heo 4600d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 46011aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4602d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4603f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4604f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 460524d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 460624d51addSTejun Heo WQ_FREEZABLE, 0); 46071aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4608ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 46096ee0578bSSuresh Siddha return 0; 46101da177e4SLinus Torvalds } 46116ee0578bSSuresh Siddha early_initcall(init_workqueues); 4612