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 */ 2601da177e4SLinus Torvalds }; 2611da177e4SLinus Torvalds 262e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 263e904e6c2STejun Heo 264bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 265bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 266bce90380STejun Heo /* possible CPUs of each node */ 267bce90380STejun Heo 268bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 269bce90380STejun Heo 27068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2712e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2725bcab335STejun Heo 27368e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 27468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2757d19c5ceSTejun Heo 2767d19c5ceSTejun Heo /* the per-cpu worker pools */ 2777d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2787d19c5ceSTejun Heo cpu_worker_pools); 2797d19c5ceSTejun Heo 28068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2817d19c5ceSTejun Heo 28268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 28329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 28429c91e99STejun Heo 285c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 28629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 28729c91e99STejun Heo 288d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 289d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 290044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2911aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 292044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 293d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 294044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 295f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 296044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 29724d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 298d320c038STejun Heo 2997d19c5ceSTejun Heo static int worker_thread(void *__worker); 3007d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3017d19c5ceSTejun Heo const struct workqueue_attrs *from); 3027d19c5ceSTejun Heo 30397bd2347STejun Heo #define CREATE_TRACE_POINTS 30497bd2347STejun Heo #include <trace/events/workqueue.h> 30597bd2347STejun Heo 30668e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3075bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 30868e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 30968e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3105bcab335STejun Heo 311b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 31276af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 313b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 314b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 31576af4d93STejun Heo 316822d8405STejun Heo #ifdef CONFIG_LOCKDEP 317822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 318519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 319519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 320822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 321822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 322822d8405STejun Heo #else 323822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 324822d8405STejun Heo #endif 325822d8405STejun Heo 326f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 327f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 328f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3297a62c2c8STejun Heo (pool)++) 3304ce62e9eSTejun Heo 33149e3cf44STejun Heo /** 33217116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 33317116969STejun Heo * @pool: iteration cursor 334611c92a0STejun Heo * @pi: integer used for iteration 335fa1b54e6STejun Heo * 33668e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 33768e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 33868e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 339fa1b54e6STejun Heo * 340fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 341fa1b54e6STejun Heo * ignored. 34217116969STejun Heo */ 343611c92a0STejun Heo #define for_each_pool(pool, pi) \ 344611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 34568e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 346fa1b54e6STejun Heo else 34717116969STejun Heo 34817116969STejun Heo /** 349822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 350822d8405STejun Heo * @worker: iteration cursor 351822d8405STejun Heo * @wi: integer used for iteration 352822d8405STejun Heo * @pool: worker_pool to iterate workers of 353822d8405STejun Heo * 354822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 355822d8405STejun Heo * 356822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 357822d8405STejun Heo * ignored. 358822d8405STejun Heo */ 359822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 360822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 361822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 362822d8405STejun Heo else 363822d8405STejun Heo 364822d8405STejun Heo /** 36549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 36649e3cf44STejun Heo * @pwq: iteration cursor 36749e3cf44STejun Heo * @wq: the target workqueue 36876af4d93STejun Heo * 369b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 370794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 371794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 37276af4d93STejun Heo * 37376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 37476af4d93STejun Heo * ignored. 37549e3cf44STejun Heo */ 37649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 37776af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 378b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 37976af4d93STejun Heo else 380f3421797STejun Heo 381dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 382dc186ad7SThomas Gleixner 383dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 384dc186ad7SThomas Gleixner 38599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 38699777288SStanislaw Gruszka { 38799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 38899777288SStanislaw Gruszka } 38999777288SStanislaw Gruszka 390dc186ad7SThomas Gleixner /* 391dc186ad7SThomas Gleixner * fixup_init is called when: 392dc186ad7SThomas Gleixner * - an active object is initialized 393dc186ad7SThomas Gleixner */ 394dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 395dc186ad7SThomas Gleixner { 396dc186ad7SThomas Gleixner struct work_struct *work = addr; 397dc186ad7SThomas Gleixner 398dc186ad7SThomas Gleixner switch (state) { 399dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 400dc186ad7SThomas Gleixner cancel_work_sync(work); 401dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 402dc186ad7SThomas Gleixner return 1; 403dc186ad7SThomas Gleixner default: 404dc186ad7SThomas Gleixner return 0; 405dc186ad7SThomas Gleixner } 406dc186ad7SThomas Gleixner } 407dc186ad7SThomas Gleixner 408dc186ad7SThomas Gleixner /* 409dc186ad7SThomas Gleixner * fixup_activate is called when: 410dc186ad7SThomas Gleixner * - an active object is activated 411dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 412dc186ad7SThomas Gleixner */ 413dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 414dc186ad7SThomas Gleixner { 415dc186ad7SThomas Gleixner struct work_struct *work = addr; 416dc186ad7SThomas Gleixner 417dc186ad7SThomas Gleixner switch (state) { 418dc186ad7SThomas Gleixner 419dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 420dc186ad7SThomas Gleixner /* 421dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 422dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 423dc186ad7SThomas Gleixner * is tracked in the object tracker. 424dc186ad7SThomas Gleixner */ 42522df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 426dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 427dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 428dc186ad7SThomas Gleixner return 0; 429dc186ad7SThomas Gleixner } 430dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 431dc186ad7SThomas Gleixner return 0; 432dc186ad7SThomas Gleixner 433dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 434dc186ad7SThomas Gleixner WARN_ON(1); 435dc186ad7SThomas Gleixner 436dc186ad7SThomas Gleixner default: 437dc186ad7SThomas Gleixner return 0; 438dc186ad7SThomas Gleixner } 439dc186ad7SThomas Gleixner } 440dc186ad7SThomas Gleixner 441dc186ad7SThomas Gleixner /* 442dc186ad7SThomas Gleixner * fixup_free is called when: 443dc186ad7SThomas Gleixner * - an active object is freed 444dc186ad7SThomas Gleixner */ 445dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 446dc186ad7SThomas Gleixner { 447dc186ad7SThomas Gleixner struct work_struct *work = addr; 448dc186ad7SThomas Gleixner 449dc186ad7SThomas Gleixner switch (state) { 450dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 451dc186ad7SThomas Gleixner cancel_work_sync(work); 452dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 453dc186ad7SThomas Gleixner return 1; 454dc186ad7SThomas Gleixner default: 455dc186ad7SThomas Gleixner return 0; 456dc186ad7SThomas Gleixner } 457dc186ad7SThomas Gleixner } 458dc186ad7SThomas Gleixner 459dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 460dc186ad7SThomas Gleixner .name = "work_struct", 46199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 462dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 463dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 464dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 465dc186ad7SThomas Gleixner }; 466dc186ad7SThomas Gleixner 467dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 470dc186ad7SThomas Gleixner } 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 473dc186ad7SThomas Gleixner { 474dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 475dc186ad7SThomas Gleixner } 476dc186ad7SThomas Gleixner 477dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 478dc186ad7SThomas Gleixner { 479dc186ad7SThomas Gleixner if (onstack) 480dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 481dc186ad7SThomas Gleixner else 482dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 483dc186ad7SThomas Gleixner } 484dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 485dc186ad7SThomas Gleixner 486dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 487dc186ad7SThomas Gleixner { 488dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 489dc186ad7SThomas Gleixner } 490dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 491dc186ad7SThomas Gleixner 492dc186ad7SThomas Gleixner #else 493dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 494dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 495dc186ad7SThomas Gleixner #endif 496dc186ad7SThomas Gleixner 4979daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4989daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4999daf9e67STejun Heo { 5009daf9e67STejun Heo int ret; 5019daf9e67STejun Heo 50268e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5035bcab335STejun Heo 504fa1b54e6STejun Heo do { 505fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 506fa1b54e6STejun Heo return -ENOMEM; 5079daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 508fa1b54e6STejun Heo } while (ret == -EAGAIN); 5099daf9e67STejun Heo 5109daf9e67STejun Heo return ret; 5119daf9e67STejun Heo } 5129daf9e67STejun Heo 51376af4d93STejun Heo /** 51476af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 51576af4d93STejun Heo * @wq: the target workqueue 51676af4d93STejun Heo * 517b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 518794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 519794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 52076af4d93STejun Heo */ 5217fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 522a848e3b6SOleg Nesterov { 523b09f4fd3SLai Jiangshan assert_rcu_or_wq_mutex(wq); 52476af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 52576af4d93STejun Heo pwqs_node); 526f3421797STejun Heo } 527a848e3b6SOleg Nesterov 52873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 52973f53c4aSTejun Heo { 53073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 53173f53c4aSTejun Heo } 53273f53c4aSTejun Heo 53373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 53473f53c4aSTejun Heo { 53573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 53673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 53773f53c4aSTejun Heo } 53873f53c4aSTejun Heo 53973f53c4aSTejun Heo static int work_next_color(int color) 54073f53c4aSTejun Heo { 54173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 542b1f4ec17SOleg Nesterov } 543b1f4ec17SOleg Nesterov 5444594bf15SDavid Howells /* 545112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 546112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5477c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5487a22ad75STejun Heo * 549112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 550112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 551bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 552bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5537a22ad75STejun Heo * 554112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5557c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 556112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5577c3eed5cSTejun Heo * available only while the work item is queued. 558bbb68dfaSTejun Heo * 559bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 560bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 561bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 562bbb68dfaSTejun Heo * try to steal the PENDING bit. 5634594bf15SDavid Howells */ 5647a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5657a22ad75STejun Heo unsigned long flags) 5667a22ad75STejun Heo { 5676183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5687a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5697a22ad75STejun Heo } 5707a22ad75STejun Heo 571112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5724690c4abSTejun Heo unsigned long extra_flags) 573365970a1SDavid Howells { 574112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 575112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 576365970a1SDavid Howells } 577365970a1SDavid Howells 5784468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5794468a00fSLai Jiangshan int pool_id) 5804468a00fSLai Jiangshan { 5814468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5824468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5834468a00fSLai Jiangshan } 5844468a00fSLai Jiangshan 5857c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5867c3eed5cSTejun Heo int pool_id) 5874d707b9fSOleg Nesterov { 58823657bb1STejun Heo /* 58923657bb1STejun Heo * The following wmb is paired with the implied mb in 59023657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 59123657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 59223657bb1STejun Heo * owner. 59323657bb1STejun Heo */ 59423657bb1STejun Heo smp_wmb(); 5957c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5964d707b9fSOleg Nesterov } 5974d707b9fSOleg Nesterov 5987a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 599365970a1SDavid Howells { 6007c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6017c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6027a22ad75STejun Heo } 6037a22ad75STejun Heo 604112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6057a22ad75STejun Heo { 606e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6077a22ad75STejun Heo 608112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 609e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 610e120153dSTejun Heo else 611e120153dSTejun Heo return NULL; 6127a22ad75STejun Heo } 6137a22ad75STejun Heo 6147c3eed5cSTejun Heo /** 6157c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6167c3eed5cSTejun Heo * @work: the work item of interest 6177c3eed5cSTejun Heo * 6187c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 619fa1b54e6STejun Heo * 62068e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 62168e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 62268e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 623fa1b54e6STejun Heo * 624fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 625fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 626fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 627fa1b54e6STejun Heo * returned pool is and stays online. 6287c3eed5cSTejun Heo */ 6297c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6307a22ad75STejun Heo { 631e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6327c3eed5cSTejun Heo int pool_id; 6337a22ad75STejun Heo 63468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 635fa1b54e6STejun Heo 636112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 637112202d9STejun Heo return ((struct pool_workqueue *) 6387c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6397a22ad75STejun Heo 6407c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6417c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6427a22ad75STejun Heo return NULL; 6437a22ad75STejun Heo 644fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6457c3eed5cSTejun Heo } 6467c3eed5cSTejun Heo 6477c3eed5cSTejun Heo /** 6487c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6497c3eed5cSTejun Heo * @work: the work item of interest 6507c3eed5cSTejun Heo * 6517c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6527c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6537c3eed5cSTejun Heo */ 6547c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6557c3eed5cSTejun Heo { 65654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6577c3eed5cSTejun Heo 658112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 659112202d9STejun Heo return ((struct pool_workqueue *) 66054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 66154d5b7d0SLai Jiangshan 66254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6637c3eed5cSTejun Heo } 6647c3eed5cSTejun Heo 665bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 666bbb68dfaSTejun Heo { 6677c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 668bbb68dfaSTejun Heo 6697c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6707c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 671bbb68dfaSTejun Heo } 672bbb68dfaSTejun Heo 673bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 674bbb68dfaSTejun Heo { 675bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 676bbb68dfaSTejun Heo 677112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 678bbb68dfaSTejun Heo } 679bbb68dfaSTejun Heo 680e22bee78STejun Heo /* 6813270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6823270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 683d565ed63STejun Heo * they're being called with pool->lock held. 684e22bee78STejun Heo */ 685e22bee78STejun Heo 68663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 687649027d7STejun Heo { 688e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 689649027d7STejun Heo } 690649027d7STejun Heo 691e22bee78STejun Heo /* 692e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 693e22bee78STejun Heo * running workers. 694974271c4STejun Heo * 695974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 696706026c2STejun Heo * function will always return %true for unbound pools as long as the 697974271c4STejun Heo * worklist isn't empty. 698e22bee78STejun Heo */ 69963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 700e22bee78STejun Heo { 70163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 702e22bee78STejun Heo } 703e22bee78STejun Heo 704e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 70563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 706e22bee78STejun Heo { 70763d95a91STejun Heo return pool->nr_idle; 708e22bee78STejun Heo } 709e22bee78STejun Heo 710e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 71163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 712e22bee78STejun Heo { 713e19e397aSTejun Heo return !list_empty(&pool->worklist) && 714e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 715e22bee78STejun Heo } 716e22bee78STejun Heo 717e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 71863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 719e22bee78STejun Heo { 72063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 721e22bee78STejun Heo } 722e22bee78STejun Heo 723e22bee78STejun Heo /* Do I need to be the manager? */ 72463d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 725e22bee78STejun Heo { 72663d95a91STejun Heo return need_to_create_worker(pool) || 72711ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 728e22bee78STejun Heo } 729e22bee78STejun Heo 730e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 73163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 732e22bee78STejun Heo { 73334a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 73463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 73563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 736e22bee78STejun Heo 737ea1abd61SLai Jiangshan /* 738ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 739ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 740ea1abd61SLai Jiangshan */ 741ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 742ea1abd61SLai Jiangshan return false; 743ea1abd61SLai Jiangshan 744e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 745e22bee78STejun Heo } 746e22bee78STejun Heo 747e22bee78STejun Heo /* 748e22bee78STejun Heo * Wake up functions. 749e22bee78STejun Heo */ 750e22bee78STejun Heo 7517e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 75263d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7537e11629dSTejun Heo { 75463d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7557e11629dSTejun Heo return NULL; 7567e11629dSTejun Heo 75763d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7587e11629dSTejun Heo } 7597e11629dSTejun Heo 7607e11629dSTejun Heo /** 7617e11629dSTejun Heo * wake_up_worker - wake up an idle worker 76263d95a91STejun Heo * @pool: worker pool to wake worker from 7637e11629dSTejun Heo * 76463d95a91STejun Heo * Wake up the first idle worker of @pool. 7657e11629dSTejun Heo * 7667e11629dSTejun Heo * CONTEXT: 767d565ed63STejun Heo * spin_lock_irq(pool->lock). 7687e11629dSTejun Heo */ 76963d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7707e11629dSTejun Heo { 77163d95a91STejun Heo struct worker *worker = first_worker(pool); 7727e11629dSTejun Heo 7737e11629dSTejun Heo if (likely(worker)) 7747e11629dSTejun Heo wake_up_process(worker->task); 7757e11629dSTejun Heo } 7767e11629dSTejun Heo 7774690c4abSTejun Heo /** 778e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 779e22bee78STejun Heo * @task: task waking up 780e22bee78STejun Heo * @cpu: CPU @task is waking up to 781e22bee78STejun Heo * 782e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 783e22bee78STejun Heo * being awoken. 784e22bee78STejun Heo * 785e22bee78STejun Heo * CONTEXT: 786e22bee78STejun Heo * spin_lock_irq(rq->lock) 787e22bee78STejun Heo */ 788d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 789e22bee78STejun Heo { 790e22bee78STejun Heo struct worker *worker = kthread_data(task); 791e22bee78STejun Heo 79236576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 793ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 794e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 795e22bee78STejun Heo } 79636576000SJoonsoo Kim } 797e22bee78STejun Heo 798e22bee78STejun Heo /** 799e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 800e22bee78STejun Heo * @task: task going to sleep 801e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 802e22bee78STejun Heo * 803e22bee78STejun Heo * This function is called during schedule() when a busy worker is 804e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 805e22bee78STejun Heo * returning pointer to its task. 806e22bee78STejun Heo * 807e22bee78STejun Heo * CONTEXT: 808e22bee78STejun Heo * spin_lock_irq(rq->lock) 809e22bee78STejun Heo * 810e22bee78STejun Heo * RETURNS: 811e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 812e22bee78STejun Heo */ 813d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 814e22bee78STejun Heo { 815e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 816111c225aSTejun Heo struct worker_pool *pool; 817e22bee78STejun Heo 818111c225aSTejun Heo /* 819111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 820111c225aSTejun Heo * workers, also reach here, let's not access anything before 821111c225aSTejun Heo * checking NOT_RUNNING. 822111c225aSTejun Heo */ 8232d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 824e22bee78STejun Heo return NULL; 825e22bee78STejun Heo 826111c225aSTejun Heo pool = worker->pool; 827111c225aSTejun Heo 828e22bee78STejun Heo /* this can only happen on the local cpu */ 8296183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8306183c009STejun Heo return NULL; 831e22bee78STejun Heo 832e22bee78STejun Heo /* 833e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 834e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 835e22bee78STejun Heo * Please read comment there. 836e22bee78STejun Heo * 837628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 838628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 839628c78e7STejun Heo * disabled, which in turn means that none else could be 840d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 841628c78e7STejun Heo * lock is safe. 842e22bee78STejun Heo */ 843e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 844e19e397aSTejun Heo !list_empty(&pool->worklist)) 84563d95a91STejun Heo to_wakeup = first_worker(pool); 846e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 847e22bee78STejun Heo } 848e22bee78STejun Heo 849e22bee78STejun Heo /** 850e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 851cb444766STejun Heo * @worker: self 852d302f017STejun Heo * @flags: flags to set 853d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 854d302f017STejun Heo * 855e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 856e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 857e22bee78STejun Heo * woken up. 858d302f017STejun Heo * 859cb444766STejun Heo * CONTEXT: 860d565ed63STejun Heo * spin_lock_irq(pool->lock) 861d302f017STejun Heo */ 862d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 863d302f017STejun Heo bool wakeup) 864d302f017STejun Heo { 865bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 866e22bee78STejun Heo 867cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 868cb444766STejun Heo 869e22bee78STejun Heo /* 870e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 871e22bee78STejun Heo * wake up an idle worker as necessary if requested by 872e22bee78STejun Heo * @wakeup. 873e22bee78STejun Heo */ 874e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 875e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 876e22bee78STejun Heo if (wakeup) { 877e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 878bd7bdd43STejun Heo !list_empty(&pool->worklist)) 87963d95a91STejun Heo wake_up_worker(pool); 880e22bee78STejun Heo } else 881e19e397aSTejun Heo atomic_dec(&pool->nr_running); 882e22bee78STejun Heo } 883e22bee78STejun Heo 884d302f017STejun Heo worker->flags |= flags; 885d302f017STejun Heo } 886d302f017STejun Heo 887d302f017STejun Heo /** 888e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 889cb444766STejun Heo * @worker: self 890d302f017STejun Heo * @flags: flags to clear 891d302f017STejun Heo * 892e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 893d302f017STejun Heo * 894cb444766STejun Heo * CONTEXT: 895d565ed63STejun Heo * spin_lock_irq(pool->lock) 896d302f017STejun Heo */ 897d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 898d302f017STejun Heo { 89963d95a91STejun Heo struct worker_pool *pool = worker->pool; 900e22bee78STejun Heo unsigned int oflags = worker->flags; 901e22bee78STejun Heo 902cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 903cb444766STejun Heo 904d302f017STejun Heo worker->flags &= ~flags; 905e22bee78STejun Heo 90642c025f3STejun Heo /* 90742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 90842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 90942c025f3STejun Heo * of multiple flags, not a single flag. 91042c025f3STejun Heo */ 911e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 912e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 913e19e397aSTejun Heo atomic_inc(&pool->nr_running); 914d302f017STejun Heo } 915d302f017STejun Heo 916d302f017STejun Heo /** 9178cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 918c9e7cf27STejun Heo * @pool: pool of interest 9198cca0eeaSTejun Heo * @work: work to find worker for 9208cca0eeaSTejun Heo * 921c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 922c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 923a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 924a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 925a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 926a2c1c57bSTejun Heo * being executed. 927a2c1c57bSTejun Heo * 928a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 929a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 930a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 931a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 932a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 933a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 934a2c1c57bSTejun Heo * 935c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 936c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 937c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 938c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 939c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 940c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9418cca0eeaSTejun Heo * 9428cca0eeaSTejun Heo * CONTEXT: 943d565ed63STejun Heo * spin_lock_irq(pool->lock). 9448cca0eeaSTejun Heo * 9458cca0eeaSTejun Heo * RETURNS: 9468cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9478cca0eeaSTejun Heo * otherwise. 9488cca0eeaSTejun Heo */ 949c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9508cca0eeaSTejun Heo struct work_struct *work) 9518cca0eeaSTejun Heo { 95242f8570fSSasha Levin struct worker *worker; 95342f8570fSSasha Levin 954b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 955a2c1c57bSTejun Heo (unsigned long)work) 956a2c1c57bSTejun Heo if (worker->current_work == work && 957a2c1c57bSTejun Heo worker->current_func == work->func) 95842f8570fSSasha Levin return worker; 95942f8570fSSasha Levin 96042f8570fSSasha Levin return NULL; 9618cca0eeaSTejun Heo } 9628cca0eeaSTejun Heo 9638cca0eeaSTejun Heo /** 964bf4ede01STejun Heo * move_linked_works - move linked works to a list 965bf4ede01STejun Heo * @work: start of series of works to be scheduled 966bf4ede01STejun Heo * @head: target list to append @work to 967bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 968bf4ede01STejun Heo * 969bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 970bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 971bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 972bf4ede01STejun Heo * 973bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 974bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 975bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 976bf4ede01STejun Heo * 977bf4ede01STejun Heo * CONTEXT: 978d565ed63STejun Heo * spin_lock_irq(pool->lock). 979bf4ede01STejun Heo */ 980bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 981bf4ede01STejun Heo struct work_struct **nextp) 982bf4ede01STejun Heo { 983bf4ede01STejun Heo struct work_struct *n; 984bf4ede01STejun Heo 985bf4ede01STejun Heo /* 986bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 987bf4ede01STejun Heo * use NULL for list head. 988bf4ede01STejun Heo */ 989bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 990bf4ede01STejun Heo list_move_tail(&work->entry, head); 991bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 992bf4ede01STejun Heo break; 993bf4ede01STejun Heo } 994bf4ede01STejun Heo 995bf4ede01STejun Heo /* 996bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 997bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 998bf4ede01STejun Heo * needs to be updated. 999bf4ede01STejun Heo */ 1000bf4ede01STejun Heo if (nextp) 1001bf4ede01STejun Heo *nextp = n; 1002bf4ede01STejun Heo } 1003bf4ede01STejun Heo 10048864b4e5STejun Heo /** 10058864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10068864b4e5STejun Heo * @pwq: pool_workqueue to get 10078864b4e5STejun Heo * 10088864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10098864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10108864b4e5STejun Heo */ 10118864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10128864b4e5STejun Heo { 10138864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10148864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10158864b4e5STejun Heo pwq->refcnt++; 10168864b4e5STejun Heo } 10178864b4e5STejun Heo 10188864b4e5STejun Heo /** 10198864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10208864b4e5STejun Heo * @pwq: pool_workqueue to put 10218864b4e5STejun Heo * 10228864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10238864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10248864b4e5STejun Heo */ 10258864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10268864b4e5STejun Heo { 10278864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10288864b4e5STejun Heo if (likely(--pwq->refcnt)) 10298864b4e5STejun Heo return; 10308864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10318864b4e5STejun Heo return; 10328864b4e5STejun Heo /* 10338864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10348864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10358864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10368864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10378864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10388864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10398864b4e5STejun Heo */ 10408864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10418864b4e5STejun Heo } 10428864b4e5STejun Heo 1043112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1044bf4ede01STejun Heo { 1045112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1046bf4ede01STejun Heo 1047bf4ede01STejun Heo trace_workqueue_activate_work(work); 1048112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1049bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1050112202d9STejun Heo pwq->nr_active++; 1051bf4ede01STejun Heo } 1052bf4ede01STejun Heo 1053112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10543aa62497SLai Jiangshan { 1055112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10563aa62497SLai Jiangshan struct work_struct, entry); 10573aa62497SLai Jiangshan 1058112202d9STejun Heo pwq_activate_delayed_work(work); 10593aa62497SLai Jiangshan } 10603aa62497SLai Jiangshan 1061bf4ede01STejun Heo /** 1062112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1063112202d9STejun Heo * @pwq: pwq of interest 1064bf4ede01STejun Heo * @color: color of work which left the queue 1065bf4ede01STejun Heo * 1066bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1067112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1068bf4ede01STejun Heo * 1069bf4ede01STejun Heo * CONTEXT: 1070d565ed63STejun Heo * spin_lock_irq(pool->lock). 1071bf4ede01STejun Heo */ 1072112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1073bf4ede01STejun Heo { 10748864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1075bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10768864b4e5STejun Heo goto out_put; 1077bf4ede01STejun Heo 1078112202d9STejun Heo pwq->nr_in_flight[color]--; 1079bf4ede01STejun Heo 1080112202d9STejun Heo pwq->nr_active--; 1081112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1082bf4ede01STejun Heo /* one down, submit a delayed one */ 1083112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1084112202d9STejun Heo pwq_activate_first_delayed(pwq); 1085bf4ede01STejun Heo } 1086bf4ede01STejun Heo 1087bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1088112202d9STejun Heo if (likely(pwq->flush_color != color)) 10898864b4e5STejun Heo goto out_put; 1090bf4ede01STejun Heo 1091bf4ede01STejun Heo /* are there still in-flight works? */ 1092112202d9STejun Heo if (pwq->nr_in_flight[color]) 10938864b4e5STejun Heo goto out_put; 1094bf4ede01STejun Heo 1095112202d9STejun Heo /* this pwq is done, clear flush_color */ 1096112202d9STejun Heo pwq->flush_color = -1; 1097bf4ede01STejun Heo 1098bf4ede01STejun Heo /* 1099112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1100bf4ede01STejun Heo * will handle the rest. 1101bf4ede01STejun Heo */ 1102112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1103112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11048864b4e5STejun Heo out_put: 11058864b4e5STejun Heo put_pwq(pwq); 1106bf4ede01STejun Heo } 1107bf4ede01STejun Heo 110836e227d2STejun Heo /** 1109bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 111036e227d2STejun Heo * @work: work item to steal 111136e227d2STejun Heo * @is_dwork: @work is a delayed_work 1112bbb68dfaSTejun Heo * @flags: place to store irq state 111336e227d2STejun Heo * 111436e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 111536e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 111636e227d2STejun Heo * 111736e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 111836e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 111936e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1120bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1121bbb68dfaSTejun Heo * for arbitrarily long 112236e227d2STejun Heo * 1123bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1124e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1125e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1126e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1127bbb68dfaSTejun Heo * 1128bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1129bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1130bbb68dfaSTejun Heo * 1131e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1132bf4ede01STejun Heo */ 1133bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1134bbb68dfaSTejun Heo unsigned long *flags) 1135bf4ede01STejun Heo { 1136d565ed63STejun Heo struct worker_pool *pool; 1137112202d9STejun Heo struct pool_workqueue *pwq; 1138bf4ede01STejun Heo 1139bbb68dfaSTejun Heo local_irq_save(*flags); 1140bbb68dfaSTejun Heo 114136e227d2STejun Heo /* try to steal the timer if it exists */ 114236e227d2STejun Heo if (is_dwork) { 114336e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 114436e227d2STejun Heo 1145e0aecdd8STejun Heo /* 1146e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1147e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1148e0aecdd8STejun Heo * running on the local CPU. 1149e0aecdd8STejun Heo */ 115036e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 115136e227d2STejun Heo return 1; 115236e227d2STejun Heo } 115336e227d2STejun Heo 115436e227d2STejun Heo /* try to claim PENDING the normal way */ 1155bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1156bf4ede01STejun Heo return 0; 1157bf4ede01STejun Heo 1158bf4ede01STejun Heo /* 1159bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1160bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1161bf4ede01STejun Heo */ 1162d565ed63STejun Heo pool = get_work_pool(work); 1163d565ed63STejun Heo if (!pool) 1164bbb68dfaSTejun Heo goto fail; 1165bf4ede01STejun Heo 1166d565ed63STejun Heo spin_lock(&pool->lock); 1167bf4ede01STejun Heo /* 1168112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1169112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1170112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1171112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1172112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11730b3dae68SLai Jiangshan * item is currently queued on that pool. 1174bf4ede01STejun Heo */ 1175112202d9STejun Heo pwq = get_work_pwq(work); 1176112202d9STejun Heo if (pwq && pwq->pool == pool) { 1177bf4ede01STejun Heo debug_work_deactivate(work); 11783aa62497SLai Jiangshan 11793aa62497SLai Jiangshan /* 118016062836STejun Heo * A delayed work item cannot be grabbed directly because 118116062836STejun Heo * it might have linked NO_COLOR work items which, if left 1182112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 118316062836STejun Heo * management later on and cause stall. Make sure the work 118416062836STejun Heo * item is activated before grabbing. 11853aa62497SLai Jiangshan */ 11863aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1187112202d9STejun Heo pwq_activate_delayed_work(work); 11883aa62497SLai Jiangshan 1189bf4ede01STejun Heo list_del_init(&work->entry); 1190112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 119136e227d2STejun Heo 1192112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11934468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11944468a00fSLai Jiangshan 1195d565ed63STejun Heo spin_unlock(&pool->lock); 119636e227d2STejun Heo return 1; 1197bf4ede01STejun Heo } 1198d565ed63STejun Heo spin_unlock(&pool->lock); 1199bbb68dfaSTejun Heo fail: 1200bbb68dfaSTejun Heo local_irq_restore(*flags); 1201bbb68dfaSTejun Heo if (work_is_canceling(work)) 1202bbb68dfaSTejun Heo return -ENOENT; 1203bbb68dfaSTejun Heo cpu_relax(); 120436e227d2STejun Heo return -EAGAIN; 1205bf4ede01STejun Heo } 1206bf4ede01STejun Heo 1207bf4ede01STejun Heo /** 1208706026c2STejun Heo * insert_work - insert a work into a pool 1209112202d9STejun Heo * @pwq: pwq @work belongs to 12104690c4abSTejun Heo * @work: work to insert 12114690c4abSTejun Heo * @head: insertion point 12124690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12134690c4abSTejun Heo * 1214112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1215706026c2STejun Heo * work_struct flags. 12164690c4abSTejun Heo * 12174690c4abSTejun Heo * CONTEXT: 1218d565ed63STejun Heo * spin_lock_irq(pool->lock). 1219365970a1SDavid Howells */ 1220112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1221112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1222b89deed3SOleg Nesterov { 1223112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1224e1d8aa9fSFrederic Weisbecker 12254690c4abSTejun Heo /* we own @work, set data and link */ 1226112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12271a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12288864b4e5STejun Heo get_pwq(pwq); 1229e22bee78STejun Heo 1230e22bee78STejun Heo /* 1231c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1232c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1233c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1234e22bee78STejun Heo */ 1235e22bee78STejun Heo smp_mb(); 1236e22bee78STejun Heo 123763d95a91STejun Heo if (__need_more_worker(pool)) 123863d95a91STejun Heo wake_up_worker(pool); 1239b89deed3SOleg Nesterov } 1240b89deed3SOleg Nesterov 1241c8efcc25STejun Heo /* 1242c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12438d03ecfeSTejun Heo * same workqueue. 1244c8efcc25STejun Heo */ 1245c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1246c8efcc25STejun Heo { 1247c8efcc25STejun Heo struct worker *worker; 1248c8efcc25STejun Heo 12498d03ecfeSTejun Heo worker = current_wq_worker(); 1250c8efcc25STejun Heo /* 12518d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12528d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1253c8efcc25STejun Heo */ 1254112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1255c8efcc25STejun Heo } 1256c8efcc25STejun Heo 1257d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12581da177e4SLinus Torvalds struct work_struct *work) 12591da177e4SLinus Torvalds { 1260112202d9STejun Heo struct pool_workqueue *pwq; 1261c9178087STejun Heo struct worker_pool *last_pool; 12621e19ffc6STejun Heo struct list_head *worklist; 12638a2e8e5dSTejun Heo unsigned int work_flags; 1264b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12658930cabaSTejun Heo 12668930cabaSTejun Heo /* 12678930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12688930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12698930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12708930cabaSTejun Heo * happen with IRQ disabled. 12718930cabaSTejun Heo */ 12728930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12731da177e4SLinus Torvalds 1274dc186ad7SThomas Gleixner debug_work_activate(work); 12751e19ffc6STejun Heo 1276c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1277618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1278c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1279e41e704bSTejun Heo return; 12809e8cd2f5STejun Heo retry: 1281c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1282c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 128357469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1284f3421797STejun Heo cpu = raw_smp_processor_id(); 1285c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1286c9178087STejun Heo } else { 1287c9178087STejun Heo pwq = first_pwq(wq); 1288c9178087STejun Heo } 1289f3421797STejun Heo 129018aa9effSTejun Heo /* 1291c9178087STejun Heo * If @work was previously on a different pool, it might still be 1292c9178087STejun Heo * running there, in which case the work needs to be queued on that 1293c9178087STejun Heo * pool to guarantee non-reentrancy. 129418aa9effSTejun Heo */ 1295c9e7cf27STejun Heo last_pool = get_work_pool(work); 1296112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 129718aa9effSTejun Heo struct worker *worker; 129818aa9effSTejun Heo 1299d565ed63STejun Heo spin_lock(&last_pool->lock); 130018aa9effSTejun Heo 1301c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 130218aa9effSTejun Heo 1303112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1304c9178087STejun Heo pwq = worker->current_pwq; 13058594fadeSLai Jiangshan } else { 130618aa9effSTejun Heo /* meh... not running there, queue here */ 1307d565ed63STejun Heo spin_unlock(&last_pool->lock); 1308112202d9STejun Heo spin_lock(&pwq->pool->lock); 130918aa9effSTejun Heo } 13108930cabaSTejun Heo } else { 1311112202d9STejun Heo spin_lock(&pwq->pool->lock); 13128930cabaSTejun Heo } 1313502ca9d8STejun Heo 13149e8cd2f5STejun Heo /* 13159e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13169e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13179e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 13189e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 13199e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 13209e8cd2f5STejun Heo * make forward-progress. 13219e8cd2f5STejun Heo */ 13229e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13239e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13249e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13259e8cd2f5STejun Heo cpu_relax(); 13269e8cd2f5STejun Heo goto retry; 13279e8cd2f5STejun Heo } 13289e8cd2f5STejun Heo /* oops */ 13299e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13309e8cd2f5STejun Heo wq->name, cpu); 13319e8cd2f5STejun Heo } 13329e8cd2f5STejun Heo 1333112202d9STejun Heo /* pwq determined, queue */ 1334112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1335502ca9d8STejun Heo 1336f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1337112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1338f5b2552bSDan Carpenter return; 1339f5b2552bSDan Carpenter } 13401e19ffc6STejun Heo 1341112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1342112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13431e19ffc6STejun Heo 1344112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1345cdadf009STejun Heo trace_workqueue_activate_work(work); 1346112202d9STejun Heo pwq->nr_active++; 1347112202d9STejun Heo worklist = &pwq->pool->worklist; 13488a2e8e5dSTejun Heo } else { 13498a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1350112202d9STejun Heo worklist = &pwq->delayed_works; 13518a2e8e5dSTejun Heo } 13521e19ffc6STejun Heo 1353112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13541e19ffc6STejun Heo 1355112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13561da177e4SLinus Torvalds } 13571da177e4SLinus Torvalds 13580fcb78c2SRolf Eike Beer /** 1359c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1360c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1361c1a220e7SZhang Rui * @wq: workqueue to use 1362c1a220e7SZhang Rui * @work: work to queue 1363c1a220e7SZhang Rui * 1364d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1365c1a220e7SZhang Rui * 1366c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1367c1a220e7SZhang Rui * can't go away. 1368c1a220e7SZhang Rui */ 1369d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1370d4283e93STejun Heo struct work_struct *work) 1371c1a220e7SZhang Rui { 1372d4283e93STejun Heo bool ret = false; 13738930cabaSTejun Heo unsigned long flags; 13748930cabaSTejun Heo 13758930cabaSTejun Heo local_irq_save(flags); 1376c1a220e7SZhang Rui 137722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13784690c4abSTejun Heo __queue_work(cpu, wq, work); 1379d4283e93STejun Heo ret = true; 1380c1a220e7SZhang Rui } 13818930cabaSTejun Heo 13828930cabaSTejun Heo local_irq_restore(flags); 1383c1a220e7SZhang Rui return ret; 1384c1a220e7SZhang Rui } 1385c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1386c1a220e7SZhang Rui 1387d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13881da177e4SLinus Torvalds { 138952bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13901da177e4SLinus Torvalds 1391e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 139260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13931da177e4SLinus Torvalds } 13941438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13951da177e4SLinus Torvalds 13967beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 139752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13981da177e4SLinus Torvalds { 13997beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14007beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14011da177e4SLinus Torvalds 14027beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14037beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1404fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1405fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14067beb2edfSTejun Heo 14078852aac2STejun Heo /* 14088852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14098852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14108852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14118852aac2STejun Heo * on that there's no such delay when @delay is 0. 14128852aac2STejun Heo */ 14138852aac2STejun Heo if (!delay) { 14148852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14158852aac2STejun Heo return; 14168852aac2STejun Heo } 14178852aac2STejun Heo 14187beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14197beb2edfSTejun Heo 142060c057bcSLai Jiangshan dwork->wq = wq; 14211265057fSTejun Heo dwork->cpu = cpu; 14227beb2edfSTejun Heo timer->expires = jiffies + delay; 14237beb2edfSTejun Heo 14247beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14257beb2edfSTejun Heo add_timer_on(timer, cpu); 14267beb2edfSTejun Heo else 14277beb2edfSTejun Heo add_timer(timer); 14287beb2edfSTejun Heo } 14291da177e4SLinus Torvalds 14300fcb78c2SRolf Eike Beer /** 14310fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14320fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14330fcb78c2SRolf Eike Beer * @wq: workqueue to use 1434af9997e4SRandy Dunlap * @dwork: work to queue 14350fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14360fcb78c2SRolf Eike Beer * 1437715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1438715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1439715f1300STejun Heo * execution. 14400fcb78c2SRolf Eike Beer */ 1441d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 144252bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14437a6bc1cdSVenkatesh Pallipadi { 144452bad64dSDavid Howells struct work_struct *work = &dwork->work; 1445d4283e93STejun Heo bool ret = false; 14468930cabaSTejun Heo unsigned long flags; 14478930cabaSTejun Heo 14488930cabaSTejun Heo /* read the comment in __queue_work() */ 14498930cabaSTejun Heo local_irq_save(flags); 14507a6bc1cdSVenkatesh Pallipadi 145122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14527beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1453d4283e93STejun Heo ret = true; 14547a6bc1cdSVenkatesh Pallipadi } 14558930cabaSTejun Heo 14568930cabaSTejun Heo local_irq_restore(flags); 14577a6bc1cdSVenkatesh Pallipadi return ret; 14587a6bc1cdSVenkatesh Pallipadi } 1459ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14601da177e4SLinus Torvalds 1461c8e55f36STejun Heo /** 14628376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14638376fe22STejun Heo * @cpu: CPU number to execute work on 14648376fe22STejun Heo * @wq: workqueue to use 14658376fe22STejun Heo * @dwork: work to queue 14668376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14678376fe22STejun Heo * 14688376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14698376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14708376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14718376fe22STejun Heo * current state. 14728376fe22STejun Heo * 14738376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14748376fe22STejun Heo * pending and its timer was modified. 14758376fe22STejun Heo * 1476e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14778376fe22STejun Heo * See try_to_grab_pending() for details. 14788376fe22STejun Heo */ 14798376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14808376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14818376fe22STejun Heo { 14828376fe22STejun Heo unsigned long flags; 14838376fe22STejun Heo int ret; 14848376fe22STejun Heo 14858376fe22STejun Heo do { 14868376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14878376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14888376fe22STejun Heo 14898376fe22STejun Heo if (likely(ret >= 0)) { 14908376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14918376fe22STejun Heo local_irq_restore(flags); 14928376fe22STejun Heo } 14938376fe22STejun Heo 14948376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14958376fe22STejun Heo return ret; 14968376fe22STejun Heo } 14978376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14988376fe22STejun Heo 14998376fe22STejun Heo /** 1500c8e55f36STejun Heo * worker_enter_idle - enter idle state 1501c8e55f36STejun Heo * @worker: worker which is entering idle state 1502c8e55f36STejun Heo * 1503c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1504c8e55f36STejun Heo * necessary. 1505c8e55f36STejun Heo * 1506c8e55f36STejun Heo * LOCKING: 1507d565ed63STejun Heo * spin_lock_irq(pool->lock). 1508c8e55f36STejun Heo */ 1509c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15101da177e4SLinus Torvalds { 1511bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1512c8e55f36STejun Heo 15136183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15146183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15156183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15166183c009STejun Heo return; 1517c8e55f36STejun Heo 1518cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1519cb444766STejun Heo worker->flags |= WORKER_IDLE; 1520bd7bdd43STejun Heo pool->nr_idle++; 1521e22bee78STejun Heo worker->last_active = jiffies; 1522c8e55f36STejun Heo 1523c8e55f36STejun Heo /* idle_list is LIFO */ 1524bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1525db7bccf4STejun Heo 152663d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1527628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1528cb444766STejun Heo 1529544ecf31STejun Heo /* 1530706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1531d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1532628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1533628c78e7STejun Heo * unbind is not in progress. 1534544ecf31STejun Heo */ 153524647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1536bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1537e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1538c8e55f36STejun Heo } 1539c8e55f36STejun Heo 1540c8e55f36STejun Heo /** 1541c8e55f36STejun Heo * worker_leave_idle - leave idle state 1542c8e55f36STejun Heo * @worker: worker which is leaving idle state 1543c8e55f36STejun Heo * 1544c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1545c8e55f36STejun Heo * 1546c8e55f36STejun Heo * LOCKING: 1547d565ed63STejun Heo * spin_lock_irq(pool->lock). 1548c8e55f36STejun Heo */ 1549c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1550c8e55f36STejun Heo { 1551bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1552c8e55f36STejun Heo 15536183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15546183c009STejun Heo return; 1555d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1556bd7bdd43STejun Heo pool->nr_idle--; 1557c8e55f36STejun Heo list_del_init(&worker->entry); 1558c8e55f36STejun Heo } 1559c8e55f36STejun Heo 1560e22bee78STejun Heo /** 1561f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1562f36dc67bSLai Jiangshan * @pool: target worker_pool 1563f36dc67bSLai Jiangshan * 1564f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1565e22bee78STejun Heo * 1566e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1567e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1568e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1569e22bee78STejun Heo * guaranteed to execute on the cpu. 1570e22bee78STejun Heo * 1571f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1572e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1573e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1574e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1575706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1576e22bee78STejun Heo * [dis]associated in the meantime. 1577e22bee78STejun Heo * 1578706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 157924647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1580f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1581f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1582f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1583e22bee78STejun Heo * 1584e22bee78STejun Heo * CONTEXT: 1585d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1586e22bee78STejun Heo * held. 1587e22bee78STejun Heo * 1588e22bee78STejun Heo * RETURNS: 1589706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1590e22bee78STejun Heo * bound), %false if offline. 1591e22bee78STejun Heo */ 1592f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1593d565ed63STejun Heo __acquires(&pool->lock) 1594e22bee78STejun Heo { 1595e22bee78STejun Heo while (true) { 1596e22bee78STejun Heo /* 1597e22bee78STejun Heo * The following call may fail, succeed or succeed 1598e22bee78STejun Heo * without actually migrating the task to the cpu if 1599e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 160024647570STejun Heo * against POOL_DISASSOCIATED. 1601e22bee78STejun Heo */ 160224647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16037a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1604e22bee78STejun Heo 1605d565ed63STejun Heo spin_lock_irq(&pool->lock); 160624647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1607e22bee78STejun Heo return false; 1608f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16097a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1610e22bee78STejun Heo return true; 1611d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1612e22bee78STejun Heo 16135035b20fSTejun Heo /* 16145035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16155035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16165035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16175035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16185035b20fSTejun Heo */ 1619e22bee78STejun Heo cpu_relax(); 16205035b20fSTejun Heo cond_resched(); 1621e22bee78STejun Heo } 1622e22bee78STejun Heo } 1623e22bee78STejun Heo 1624c34056a3STejun Heo static struct worker *alloc_worker(void) 1625c34056a3STejun Heo { 1626c34056a3STejun Heo struct worker *worker; 1627c34056a3STejun Heo 1628c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1629c8e55f36STejun Heo if (worker) { 1630c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1631affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1632e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1633e22bee78STejun Heo worker->flags = WORKER_PREP; 1634c8e55f36STejun Heo } 1635c34056a3STejun Heo return worker; 1636c34056a3STejun Heo } 1637c34056a3STejun Heo 1638c34056a3STejun Heo /** 1639c34056a3STejun Heo * create_worker - create a new workqueue worker 164063d95a91STejun Heo * @pool: pool the new worker will belong to 1641c34056a3STejun Heo * 164263d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1643c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1644c34056a3STejun Heo * destroy_worker(). 1645c34056a3STejun Heo * 1646c34056a3STejun Heo * CONTEXT: 1647c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1648c34056a3STejun Heo * 1649c34056a3STejun Heo * RETURNS: 1650c34056a3STejun Heo * Pointer to the newly created worker. 1651c34056a3STejun Heo */ 1652bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1653c34056a3STejun Heo { 1654c34056a3STejun Heo struct worker *worker = NULL; 1655f3421797STejun Heo int id = -1; 1656e3c916a4STejun Heo char id_buf[16]; 1657c34056a3STejun Heo 1658cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1659cd549687STejun Heo 1660822d8405STejun Heo /* 1661822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1662822d8405STejun Heo * without installing the pointer. 1663822d8405STejun Heo */ 1664822d8405STejun Heo idr_preload(GFP_KERNEL); 1665d565ed63STejun Heo spin_lock_irq(&pool->lock); 1666822d8405STejun Heo 1667822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1668822d8405STejun Heo 1669d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1670822d8405STejun Heo idr_preload_end(); 1671822d8405STejun Heo if (id < 0) 1672c34056a3STejun Heo goto fail; 1673c34056a3STejun Heo 1674c34056a3STejun Heo worker = alloc_worker(); 1675c34056a3STejun Heo if (!worker) 1676c34056a3STejun Heo goto fail; 1677c34056a3STejun Heo 1678bd7bdd43STejun Heo worker->pool = pool; 1679c34056a3STejun Heo worker->id = id; 1680c34056a3STejun Heo 168129c91e99STejun Heo if (pool->cpu >= 0) 1682e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1683e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1684f3421797STejun Heo else 1685e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1686e3c916a4STejun Heo 1687f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1688e3c916a4STejun Heo "kworker/%s", id_buf); 1689c34056a3STejun Heo if (IS_ERR(worker->task)) 1690c34056a3STejun Heo goto fail; 1691c34056a3STejun Heo 1692c5aa87bbSTejun Heo /* 1693c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1694c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1695c5aa87bbSTejun Heo */ 16967a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 16977a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16983270476aSTejun Heo 169914a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 170014a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 17017a4e344cSTejun Heo 17027a4e344cSTejun Heo /* 17037a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17047a4e344cSTejun Heo * remains stable across this function. See the comments above the 17057a4e344cSTejun Heo * flag definition for details. 17067a4e344cSTejun Heo */ 17077a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1708f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1709c34056a3STejun Heo 1710822d8405STejun Heo /* successful, commit the pointer to idr */ 1711822d8405STejun Heo spin_lock_irq(&pool->lock); 1712822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1713822d8405STejun Heo spin_unlock_irq(&pool->lock); 1714822d8405STejun Heo 1715c34056a3STejun Heo return worker; 1716822d8405STejun Heo 1717c34056a3STejun Heo fail: 1718c34056a3STejun Heo if (id >= 0) { 1719d565ed63STejun Heo spin_lock_irq(&pool->lock); 1720822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1721d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1722c34056a3STejun Heo } 1723c34056a3STejun Heo kfree(worker); 1724c34056a3STejun Heo return NULL; 1725c34056a3STejun Heo } 1726c34056a3STejun Heo 1727c34056a3STejun Heo /** 1728c34056a3STejun Heo * start_worker - start a newly created worker 1729c34056a3STejun Heo * @worker: worker to start 1730c34056a3STejun Heo * 1731706026c2STejun Heo * Make the pool aware of @worker and start it. 1732c34056a3STejun Heo * 1733c34056a3STejun Heo * CONTEXT: 1734d565ed63STejun Heo * spin_lock_irq(pool->lock). 1735c34056a3STejun Heo */ 1736c34056a3STejun Heo static void start_worker(struct worker *worker) 1737c34056a3STejun Heo { 1738cb444766STejun Heo worker->flags |= WORKER_STARTED; 1739bd7bdd43STejun Heo worker->pool->nr_workers++; 1740c8e55f36STejun Heo worker_enter_idle(worker); 1741c34056a3STejun Heo wake_up_process(worker->task); 1742c34056a3STejun Heo } 1743c34056a3STejun Heo 1744c34056a3STejun Heo /** 1745ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1746ebf44d16STejun Heo * @pool: the target pool 1747ebf44d16STejun Heo * 1748cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1749ebf44d16STejun Heo */ 1750ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1751ebf44d16STejun Heo { 1752ebf44d16STejun Heo struct worker *worker; 1753ebf44d16STejun Heo 1754cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1755cd549687STejun Heo 1756ebf44d16STejun Heo worker = create_worker(pool); 1757ebf44d16STejun Heo if (worker) { 1758ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1759ebf44d16STejun Heo start_worker(worker); 1760ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1761ebf44d16STejun Heo } 1762ebf44d16STejun Heo 1763cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1764cd549687STejun Heo 1765ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1766ebf44d16STejun Heo } 1767ebf44d16STejun Heo 1768ebf44d16STejun Heo /** 1769c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1770c34056a3STejun Heo * @worker: worker to be destroyed 1771c34056a3STejun Heo * 1772706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1773c8e55f36STejun Heo * 1774c8e55f36STejun Heo * CONTEXT: 1775d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1776c34056a3STejun Heo */ 1777c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1778c34056a3STejun Heo { 1779bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1780c34056a3STejun Heo 1781cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1782cd549687STejun Heo lockdep_assert_held(&pool->lock); 1783cd549687STejun Heo 1784c34056a3STejun Heo /* sanity check frenzy */ 17856183c009STejun Heo if (WARN_ON(worker->current_work) || 17866183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 17876183c009STejun Heo return; 1788c34056a3STejun Heo 1789c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1790bd7bdd43STejun Heo pool->nr_workers--; 1791c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1792bd7bdd43STejun Heo pool->nr_idle--; 1793c8e55f36STejun Heo 1794c8e55f36STejun Heo list_del_init(&worker->entry); 1795cb444766STejun Heo worker->flags |= WORKER_DIE; 1796c8e55f36STejun Heo 1797822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1798822d8405STejun Heo 1799d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1800c8e55f36STejun Heo 1801c34056a3STejun Heo kthread_stop(worker->task); 1802c34056a3STejun Heo kfree(worker); 1803c34056a3STejun Heo 1804d565ed63STejun Heo spin_lock_irq(&pool->lock); 1805c34056a3STejun Heo } 1806c34056a3STejun Heo 180763d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1808e22bee78STejun Heo { 180963d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1810e22bee78STejun Heo 1811d565ed63STejun Heo spin_lock_irq(&pool->lock); 1812e22bee78STejun Heo 181363d95a91STejun Heo if (too_many_workers(pool)) { 1814e22bee78STejun Heo struct worker *worker; 1815e22bee78STejun Heo unsigned long expires; 1816e22bee78STejun Heo 1817e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 181863d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1819e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1820e22bee78STejun Heo 1821e22bee78STejun Heo if (time_before(jiffies, expires)) 182263d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1823e22bee78STejun Heo else { 1824e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 182511ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 182663d95a91STejun Heo wake_up_worker(pool); 1827e22bee78STejun Heo } 1828e22bee78STejun Heo } 1829e22bee78STejun Heo 1830d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1831e22bee78STejun Heo } 1832e22bee78STejun Heo 1833493a1724STejun Heo static void send_mayday(struct work_struct *work) 1834e22bee78STejun Heo { 1835112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1836112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1837493a1724STejun Heo 18382e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1839e22bee78STejun Heo 1840493008a8STejun Heo if (!wq->rescuer) 1841493a1724STejun Heo return; 1842e22bee78STejun Heo 1843e22bee78STejun Heo /* mayday mayday mayday */ 1844493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1845493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1846e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1847493a1724STejun Heo } 1848e22bee78STejun Heo } 1849e22bee78STejun Heo 1850706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1851e22bee78STejun Heo { 185263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1853e22bee78STejun Heo struct work_struct *work; 1854e22bee78STejun Heo 18552e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1856493a1724STejun Heo spin_lock(&pool->lock); 1857e22bee78STejun Heo 185863d95a91STejun Heo if (need_to_create_worker(pool)) { 1859e22bee78STejun Heo /* 1860e22bee78STejun Heo * We've been trying to create a new worker but 1861e22bee78STejun Heo * haven't been successful. We might be hitting an 1862e22bee78STejun Heo * allocation deadlock. Send distress signals to 1863e22bee78STejun Heo * rescuers. 1864e22bee78STejun Heo */ 186563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1866e22bee78STejun Heo send_mayday(work); 1867e22bee78STejun Heo } 1868e22bee78STejun Heo 1869493a1724STejun Heo spin_unlock(&pool->lock); 18702e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1871e22bee78STejun Heo 187263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1873e22bee78STejun Heo } 1874e22bee78STejun Heo 1875e22bee78STejun Heo /** 1876e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 187763d95a91STejun Heo * @pool: pool to create a new worker for 1878e22bee78STejun Heo * 187963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1880e22bee78STejun Heo * have at least one idle worker on return from this function. If 1881e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 188263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1883e22bee78STejun Heo * possible allocation deadlock. 1884e22bee78STejun Heo * 1885c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1886c5aa87bbSTejun Heo * may_start_working() %true. 1887e22bee78STejun Heo * 1888e22bee78STejun Heo * LOCKING: 1889d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1890e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1891e22bee78STejun Heo * manager. 1892e22bee78STejun Heo * 1893e22bee78STejun Heo * RETURNS: 1894c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1895e22bee78STejun Heo * otherwise. 1896e22bee78STejun Heo */ 189763d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1898d565ed63STejun Heo __releases(&pool->lock) 1899d565ed63STejun Heo __acquires(&pool->lock) 1900e22bee78STejun Heo { 190163d95a91STejun Heo if (!need_to_create_worker(pool)) 1902e22bee78STejun Heo return false; 1903e22bee78STejun Heo restart: 1904d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19059f9c2364STejun Heo 1906e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 190763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1908e22bee78STejun Heo 1909e22bee78STejun Heo while (true) { 1910e22bee78STejun Heo struct worker *worker; 1911e22bee78STejun Heo 1912bc2ae0f5STejun Heo worker = create_worker(pool); 1913e22bee78STejun Heo if (worker) { 191463d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1915d565ed63STejun Heo spin_lock_irq(&pool->lock); 1916e22bee78STejun Heo start_worker(worker); 19176183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19186183c009STejun Heo goto restart; 1919e22bee78STejun Heo return true; 1920e22bee78STejun Heo } 1921e22bee78STejun Heo 192263d95a91STejun Heo if (!need_to_create_worker(pool)) 1923e22bee78STejun Heo break; 1924e22bee78STejun Heo 1925e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1926e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19279f9c2364STejun Heo 192863d95a91STejun Heo if (!need_to_create_worker(pool)) 1929e22bee78STejun Heo break; 1930e22bee78STejun Heo } 1931e22bee78STejun Heo 193263d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1933d565ed63STejun Heo spin_lock_irq(&pool->lock); 193463d95a91STejun Heo if (need_to_create_worker(pool)) 1935e22bee78STejun Heo goto restart; 1936e22bee78STejun Heo return true; 1937e22bee78STejun Heo } 1938e22bee78STejun Heo 1939e22bee78STejun Heo /** 1940e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 194163d95a91STejun Heo * @pool: pool to destroy workers for 1942e22bee78STejun Heo * 194363d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1944e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1945e22bee78STejun Heo * 1946e22bee78STejun Heo * LOCKING: 1947d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1948e22bee78STejun Heo * multiple times. Called only from manager. 1949e22bee78STejun Heo * 1950e22bee78STejun Heo * RETURNS: 1951c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1952e22bee78STejun Heo * otherwise. 1953e22bee78STejun Heo */ 195463d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1955e22bee78STejun Heo { 1956e22bee78STejun Heo bool ret = false; 1957e22bee78STejun Heo 195863d95a91STejun Heo while (too_many_workers(pool)) { 1959e22bee78STejun Heo struct worker *worker; 1960e22bee78STejun Heo unsigned long expires; 1961e22bee78STejun Heo 196263d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1963e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1964e22bee78STejun Heo 1965e22bee78STejun Heo if (time_before(jiffies, expires)) { 196663d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1967e22bee78STejun Heo break; 1968e22bee78STejun Heo } 1969e22bee78STejun Heo 1970e22bee78STejun Heo destroy_worker(worker); 1971e22bee78STejun Heo ret = true; 1972e22bee78STejun Heo } 1973e22bee78STejun Heo 1974e22bee78STejun Heo return ret; 1975e22bee78STejun Heo } 1976e22bee78STejun Heo 1977e22bee78STejun Heo /** 1978e22bee78STejun Heo * manage_workers - manage worker pool 1979e22bee78STejun Heo * @worker: self 1980e22bee78STejun Heo * 1981706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1982e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1983706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1984e22bee78STejun Heo * 1985e22bee78STejun Heo * The caller can safely start processing works on false return. On 1986e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1987e22bee78STejun Heo * and may_start_working() is true. 1988e22bee78STejun Heo * 1989e22bee78STejun Heo * CONTEXT: 1990d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1991e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1992e22bee78STejun Heo * 1993e22bee78STejun Heo * RETURNS: 1994d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1995d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 1996e22bee78STejun Heo */ 1997e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1998e22bee78STejun Heo { 199963d95a91STejun Heo struct worker_pool *pool = worker->pool; 2000e22bee78STejun Heo bool ret = false; 2001e22bee78STejun Heo 2002bc3a1afcSTejun Heo /* 2003bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2004bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2005bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2006bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2007bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2008bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2009bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2010bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2011bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2012bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2013bc3a1afcSTejun Heo * 2014bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2015bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2016bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2017bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2018bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2019bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2020bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2021bc3a1afcSTejun Heo * manager_mutex. 2022bc3a1afcSTejun Heo */ 202334a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2024e22bee78STejun Heo return ret; 2025e22bee78STejun Heo 2026ee378aa4SLai Jiangshan /* 2027bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2028bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2029ee378aa4SLai Jiangshan */ 2030bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2031d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2032bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2033ee378aa4SLai Jiangshan ret = true; 2034ee378aa4SLai Jiangshan } 2035ee378aa4SLai Jiangshan 203611ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2037e22bee78STejun Heo 2038e22bee78STejun Heo /* 2039e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2040e22bee78STejun Heo * on return. 2041e22bee78STejun Heo */ 204263d95a91STejun Heo ret |= maybe_destroy_workers(pool); 204363d95a91STejun Heo ret |= maybe_create_worker(pool); 2044e22bee78STejun Heo 2045bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 204634a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2047e22bee78STejun Heo return ret; 2048e22bee78STejun Heo } 2049e22bee78STejun Heo 2050a62428c0STejun Heo /** 2051a62428c0STejun Heo * process_one_work - process single work 2052c34056a3STejun Heo * @worker: self 2053a62428c0STejun Heo * @work: work to process 2054a62428c0STejun Heo * 2055a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2056a62428c0STejun Heo * process a single work including synchronization against and 2057a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2058a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2059a62428c0STejun Heo * call this function to process a work. 2060a62428c0STejun Heo * 2061a62428c0STejun Heo * CONTEXT: 2062d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2063a62428c0STejun Heo */ 2064c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2065d565ed63STejun Heo __releases(&pool->lock) 2066d565ed63STejun Heo __acquires(&pool->lock) 20671da177e4SLinus Torvalds { 2068112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2069bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2070112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 207173f53c4aSTejun Heo int work_color; 20727e11629dSTejun Heo struct worker *collision; 20734e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20744e6045f1SJohannes Berg /* 2075a62428c0STejun Heo * It is permissible to free the struct work_struct from 2076a62428c0STejun Heo * inside the function that is called from it, this we need to 2077a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2078a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2079a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20804e6045f1SJohannes Berg */ 20814d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20824d82a1deSPeter Zijlstra 20834d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20844e6045f1SJohannes Berg #endif 20856fec10a1STejun Heo /* 20866fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20876fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 208824647570STejun Heo * unbound or a disassociated pool. 20896fec10a1STejun Heo */ 20905f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 209124647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2092ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 209325511a47STejun Heo 20947e11629dSTejun Heo /* 20957e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20967e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20977e11629dSTejun Heo * already processing the work. If so, defer the work to the 20987e11629dSTejun Heo * currently executing one. 20997e11629dSTejun Heo */ 2100c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21017e11629dSTejun Heo if (unlikely(collision)) { 21027e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21037e11629dSTejun Heo return; 21047e11629dSTejun Heo } 21051da177e4SLinus Torvalds 21068930cabaSTejun Heo /* claim and dequeue */ 21071da177e4SLinus Torvalds debug_work_deactivate(work); 2108c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2109c34056a3STejun Heo worker->current_work = work; 2110a2c1c57bSTejun Heo worker->current_func = work->func; 2111112202d9STejun Heo worker->current_pwq = pwq; 211273f53c4aSTejun Heo work_color = get_work_color(work); 21137a22ad75STejun Heo 2114a62428c0STejun Heo list_del_init(&work->entry); 2115a62428c0STejun Heo 2116649027d7STejun Heo /* 2117fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2118fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2119fb0e7bebSTejun Heo */ 2120fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2121fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2122fb0e7bebSTejun Heo 2123974271c4STejun Heo /* 2124d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2125974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2126974271c4STejun Heo */ 212763d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 212863d95a91STejun Heo wake_up_worker(pool); 2129974271c4STejun Heo 21308930cabaSTejun Heo /* 21317c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2132d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 213323657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 213423657bb1STejun Heo * disabled. 21358930cabaSTejun Heo */ 21367c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21371da177e4SLinus Torvalds 2138d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2139365970a1SDavid Howells 2140112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21413295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2142e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2143a2c1c57bSTejun Heo worker->current_func(work); 2144e36c886aSArjan van de Ven /* 2145e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2146e36c886aSArjan van de Ven * point will only record its address. 2147e36c886aSArjan van de Ven */ 2148e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21493295f0efSIngo Molnar lock_map_release(&lockdep_map); 2150112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21511da177e4SLinus Torvalds 2152d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2153044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2154044c782cSValentin Ilie " last function: %pf\n", 2155a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2156a2c1c57bSTejun Heo worker->current_func); 2157d5abe669SPeter Zijlstra debug_show_held_locks(current); 2158d5abe669SPeter Zijlstra dump_stack(); 2159d5abe669SPeter Zijlstra } 2160d5abe669SPeter Zijlstra 2161d565ed63STejun Heo spin_lock_irq(&pool->lock); 2162a62428c0STejun Heo 2163fb0e7bebSTejun Heo /* clear cpu intensive status */ 2164fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2165fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2166fb0e7bebSTejun Heo 2167a62428c0STejun Heo /* we're done with it, release */ 216842f8570fSSasha Levin hash_del(&worker->hentry); 2169c34056a3STejun Heo worker->current_work = NULL; 2170a2c1c57bSTejun Heo worker->current_func = NULL; 2171112202d9STejun Heo worker->current_pwq = NULL; 2172112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21731da177e4SLinus Torvalds } 21741da177e4SLinus Torvalds 2175affee4b2STejun Heo /** 2176affee4b2STejun Heo * process_scheduled_works - process scheduled works 2177affee4b2STejun Heo * @worker: self 2178affee4b2STejun Heo * 2179affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2180affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2181affee4b2STejun Heo * fetches a work from the top and executes it. 2182affee4b2STejun Heo * 2183affee4b2STejun Heo * CONTEXT: 2184d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2185affee4b2STejun Heo * multiple times. 2186affee4b2STejun Heo */ 2187affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21881da177e4SLinus Torvalds { 2189affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2190affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2191a62428c0STejun Heo struct work_struct, entry); 2192c34056a3STejun Heo process_one_work(worker, work); 2193a62428c0STejun Heo } 21941da177e4SLinus Torvalds } 21951da177e4SLinus Torvalds 21964690c4abSTejun Heo /** 21974690c4abSTejun Heo * worker_thread - the worker thread function 2198c34056a3STejun Heo * @__worker: self 21994690c4abSTejun Heo * 2200c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2201c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2202c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2203c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2204c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22054690c4abSTejun Heo */ 2206c34056a3STejun Heo static int worker_thread(void *__worker) 22071da177e4SLinus Torvalds { 2208c34056a3STejun Heo struct worker *worker = __worker; 2209bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22101da177e4SLinus Torvalds 2211e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2212e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2213c8e55f36STejun Heo woke_up: 2214d565ed63STejun Heo spin_lock_irq(&pool->lock); 2215affee4b2STejun Heo 2216a9ab775bSTejun Heo /* am I supposed to die? */ 2217a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2218d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2219a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2220e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2221c8e55f36STejun Heo return 0; 2222c8e55f36STejun Heo } 2223c8e55f36STejun Heo 2224c8e55f36STejun Heo worker_leave_idle(worker); 2225db7bccf4STejun Heo recheck: 2226e22bee78STejun Heo /* no more worker necessary? */ 222763d95a91STejun Heo if (!need_more_worker(pool)) 2228e22bee78STejun Heo goto sleep; 2229e22bee78STejun Heo 2230e22bee78STejun Heo /* do we need to manage? */ 223163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2232e22bee78STejun Heo goto recheck; 2233e22bee78STejun Heo 2234c8e55f36STejun Heo /* 2235c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2236c8e55f36STejun Heo * preparing to process a work or actually processing it. 2237c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2238c8e55f36STejun Heo */ 22396183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2240c8e55f36STejun Heo 2241e22bee78STejun Heo /* 2242a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2243a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2244a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2245a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2246a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2247e22bee78STejun Heo */ 2248a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2249e22bee78STejun Heo 2250e22bee78STejun Heo do { 2251affee4b2STejun Heo struct work_struct *work = 2252bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2253affee4b2STejun Heo struct work_struct, entry); 2254affee4b2STejun Heo 2255c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2256affee4b2STejun Heo /* optimization path, not strictly necessary */ 2257affee4b2STejun Heo process_one_work(worker, work); 2258affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2259affee4b2STejun Heo process_scheduled_works(worker); 2260affee4b2STejun Heo } else { 2261c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2262affee4b2STejun Heo process_scheduled_works(worker); 2263affee4b2STejun Heo } 226463d95a91STejun Heo } while (keep_working(pool)); 2265affee4b2STejun Heo 2266e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2267d313dd85STejun Heo sleep: 226863d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2269e22bee78STejun Heo goto recheck; 2270d313dd85STejun Heo 2271c8e55f36STejun Heo /* 2272d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2273d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2274d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2275d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2276d565ed63STejun Heo * event. 2277c8e55f36STejun Heo */ 2278c8e55f36STejun Heo worker_enter_idle(worker); 2279c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2280d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22811da177e4SLinus Torvalds schedule(); 2282c8e55f36STejun Heo goto woke_up; 22831da177e4SLinus Torvalds } 22841da177e4SLinus Torvalds 2285e22bee78STejun Heo /** 2286e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2287111c225aSTejun Heo * @__rescuer: self 2288e22bee78STejun Heo * 2289e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2290493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2291e22bee78STejun Heo * 2292706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2293e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2294e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2295e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2296e22bee78STejun Heo * the problem rescuer solves. 2297e22bee78STejun Heo * 2298706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2299706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2300e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2301e22bee78STejun Heo * 2302e22bee78STejun Heo * This should happen rarely. 2303e22bee78STejun Heo */ 2304111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2305e22bee78STejun Heo { 2306111c225aSTejun Heo struct worker *rescuer = __rescuer; 2307111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2308e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2309e22bee78STejun Heo 2310e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2311111c225aSTejun Heo 2312111c225aSTejun Heo /* 2313111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2314111c225aSTejun Heo * doesn't participate in concurrency management. 2315111c225aSTejun Heo */ 2316111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2317e22bee78STejun Heo repeat: 2318e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23191da177e4SLinus Torvalds 2320412d32e6SMike Galbraith if (kthread_should_stop()) { 2321412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2322111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2323e22bee78STejun Heo return 0; 2324412d32e6SMike Galbraith } 23251da177e4SLinus Torvalds 2326493a1724STejun Heo /* see whether any pwq is asking for help */ 23272e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2328493a1724STejun Heo 2329493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2330493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2331493a1724STejun Heo struct pool_workqueue, mayday_node); 2332112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2333e22bee78STejun Heo struct work_struct *work, *n; 2334e22bee78STejun Heo 2335e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2336493a1724STejun Heo list_del_init(&pwq->mayday_node); 2337493a1724STejun Heo 23382e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2339e22bee78STejun Heo 2340e22bee78STejun Heo /* migrate to the target cpu if possible */ 2341f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2342b3104104SLai Jiangshan rescuer->pool = pool; 2343e22bee78STejun Heo 2344e22bee78STejun Heo /* 2345e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2346e22bee78STejun Heo * process'em. 2347e22bee78STejun Heo */ 23486183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2349bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2350112202d9STejun Heo if (get_work_pwq(work) == pwq) 2351e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2352e22bee78STejun Heo 2353e22bee78STejun Heo process_scheduled_works(rescuer); 23547576958aSTejun Heo 23557576958aSTejun Heo /* 2356d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23577576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23587576958aSTejun Heo * and stalling the execution. 23597576958aSTejun Heo */ 236063d95a91STejun Heo if (keep_working(pool)) 236163d95a91STejun Heo wake_up_worker(pool); 23627576958aSTejun Heo 2363b3104104SLai Jiangshan rescuer->pool = NULL; 2364493a1724STejun Heo spin_unlock(&pool->lock); 23652e109a28STejun Heo spin_lock(&wq_mayday_lock); 23661da177e4SLinus Torvalds } 23671da177e4SLinus Torvalds 23682e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2369493a1724STejun Heo 2370111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2371111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2372e22bee78STejun Heo schedule(); 2373e22bee78STejun Heo goto repeat; 23741da177e4SLinus Torvalds } 23751da177e4SLinus Torvalds 2376fc2e4d70SOleg Nesterov struct wq_barrier { 2377fc2e4d70SOleg Nesterov struct work_struct work; 2378fc2e4d70SOleg Nesterov struct completion done; 2379fc2e4d70SOleg Nesterov }; 2380fc2e4d70SOleg Nesterov 2381fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2382fc2e4d70SOleg Nesterov { 2383fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2384fc2e4d70SOleg Nesterov complete(&barr->done); 2385fc2e4d70SOleg Nesterov } 2386fc2e4d70SOleg Nesterov 23874690c4abSTejun Heo /** 23884690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2389112202d9STejun Heo * @pwq: pwq to insert barrier into 23904690c4abSTejun Heo * @barr: wq_barrier to insert 2391affee4b2STejun Heo * @target: target work to attach @barr to 2392affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23934690c4abSTejun Heo * 2394affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2395affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2396affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2397affee4b2STejun Heo * cpu. 2398affee4b2STejun Heo * 2399affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2400affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2401affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2402affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2403affee4b2STejun Heo * after a work with LINKED flag set. 2404affee4b2STejun Heo * 2405affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2406112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24074690c4abSTejun Heo * 24084690c4abSTejun Heo * CONTEXT: 2409d565ed63STejun Heo * spin_lock_irq(pool->lock). 24104690c4abSTejun Heo */ 2411112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2412affee4b2STejun Heo struct wq_barrier *barr, 2413affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2414fc2e4d70SOleg Nesterov { 2415affee4b2STejun Heo struct list_head *head; 2416affee4b2STejun Heo unsigned int linked = 0; 2417affee4b2STejun Heo 2418dc186ad7SThomas Gleixner /* 2419d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2420dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2421dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2422dc186ad7SThomas Gleixner * might deadlock. 2423dc186ad7SThomas Gleixner */ 2424ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 242522df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2426fc2e4d70SOleg Nesterov init_completion(&barr->done); 242783c22520SOleg Nesterov 2428affee4b2STejun Heo /* 2429affee4b2STejun Heo * If @target is currently being executed, schedule the 2430affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2431affee4b2STejun Heo */ 2432affee4b2STejun Heo if (worker) 2433affee4b2STejun Heo head = worker->scheduled.next; 2434affee4b2STejun Heo else { 2435affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2436affee4b2STejun Heo 2437affee4b2STejun Heo head = target->entry.next; 2438affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2439affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2440affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2441affee4b2STejun Heo } 2442affee4b2STejun Heo 2443dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2444112202d9STejun Heo insert_work(pwq, &barr->work, head, 2445affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2446fc2e4d70SOleg Nesterov } 2447fc2e4d70SOleg Nesterov 244873f53c4aSTejun Heo /** 2449112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 245073f53c4aSTejun Heo * @wq: workqueue being flushed 245173f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 245273f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 245373f53c4aSTejun Heo * 2454112202d9STejun Heo * Prepare pwqs for workqueue flushing. 245573f53c4aSTejun Heo * 2456112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2457112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2458112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2459112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2460112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 246173f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 246273f53c4aSTejun Heo * 246373f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 246473f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 246573f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 246673f53c4aSTejun Heo * is returned. 246773f53c4aSTejun Heo * 2468112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 246973f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 247073f53c4aSTejun Heo * advanced to @work_color. 247173f53c4aSTejun Heo * 247273f53c4aSTejun Heo * CONTEXT: 24733c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 247473f53c4aSTejun Heo * 247573f53c4aSTejun Heo * RETURNS: 247673f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 247773f53c4aSTejun Heo * otherwise. 247873f53c4aSTejun Heo */ 2479112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 248073f53c4aSTejun Heo int flush_color, int work_color) 24811da177e4SLinus Torvalds { 248273f53c4aSTejun Heo bool wait = false; 248349e3cf44STejun Heo struct pool_workqueue *pwq; 24841da177e4SLinus Torvalds 248573f53c4aSTejun Heo if (flush_color >= 0) { 24866183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2487112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2488dc186ad7SThomas Gleixner } 248914441960SOleg Nesterov 249049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2491112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24921da177e4SLinus Torvalds 2493b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 249473f53c4aSTejun Heo 249573f53c4aSTejun Heo if (flush_color >= 0) { 24966183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 249773f53c4aSTejun Heo 2498112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2499112202d9STejun Heo pwq->flush_color = flush_color; 2500112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 250173f53c4aSTejun Heo wait = true; 25021da177e4SLinus Torvalds } 250373f53c4aSTejun Heo } 250473f53c4aSTejun Heo 250573f53c4aSTejun Heo if (work_color >= 0) { 25066183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2507112202d9STejun Heo pwq->work_color = work_color; 250873f53c4aSTejun Heo } 250973f53c4aSTejun Heo 2510b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25111da177e4SLinus Torvalds } 25121da177e4SLinus Torvalds 2513112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 251473f53c4aSTejun Heo complete(&wq->first_flusher->done); 251573f53c4aSTejun Heo 251673f53c4aSTejun Heo return wait; 251783c22520SOleg Nesterov } 25181da177e4SLinus Torvalds 25190fcb78c2SRolf Eike Beer /** 25201da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25210fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25221da177e4SLinus Torvalds * 2523c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2524c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25251da177e4SLinus Torvalds */ 25267ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25271da177e4SLinus Torvalds { 252873f53c4aSTejun Heo struct wq_flusher this_flusher = { 252973f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 253073f53c4aSTejun Heo .flush_color = -1, 253173f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 253273f53c4aSTejun Heo }; 253373f53c4aSTejun Heo int next_color; 2534b1f4ec17SOleg Nesterov 25353295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25363295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 253773f53c4aSTejun Heo 25383c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 253973f53c4aSTejun Heo 254073f53c4aSTejun Heo /* 254173f53c4aSTejun Heo * Start-to-wait phase 254273f53c4aSTejun Heo */ 254373f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 254473f53c4aSTejun Heo 254573f53c4aSTejun Heo if (next_color != wq->flush_color) { 254673f53c4aSTejun Heo /* 254773f53c4aSTejun Heo * Color space is not full. The current work_color 254873f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 254973f53c4aSTejun Heo * by one. 255073f53c4aSTejun Heo */ 25516183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 255273f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 255373f53c4aSTejun Heo wq->work_color = next_color; 255473f53c4aSTejun Heo 255573f53c4aSTejun Heo if (!wq->first_flusher) { 255673f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25576183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 255873f53c4aSTejun Heo 255973f53c4aSTejun Heo wq->first_flusher = &this_flusher; 256073f53c4aSTejun Heo 2561112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 256273f53c4aSTejun Heo wq->work_color)) { 256373f53c4aSTejun Heo /* nothing to flush, done */ 256473f53c4aSTejun Heo wq->flush_color = next_color; 256573f53c4aSTejun Heo wq->first_flusher = NULL; 256673f53c4aSTejun Heo goto out_unlock; 256773f53c4aSTejun Heo } 256873f53c4aSTejun Heo } else { 256973f53c4aSTejun Heo /* wait in queue */ 25706183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 257173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2572112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 257373f53c4aSTejun Heo } 257473f53c4aSTejun Heo } else { 257573f53c4aSTejun Heo /* 257673f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 257773f53c4aSTejun Heo * The next flush completion will assign us 257873f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 257973f53c4aSTejun Heo */ 258073f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 258173f53c4aSTejun Heo } 258273f53c4aSTejun Heo 25833c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 258473f53c4aSTejun Heo 258573f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 258673f53c4aSTejun Heo 258773f53c4aSTejun Heo /* 258873f53c4aSTejun Heo * Wake-up-and-cascade phase 258973f53c4aSTejun Heo * 259073f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 259173f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 259273f53c4aSTejun Heo */ 259373f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 259473f53c4aSTejun Heo return; 259573f53c4aSTejun Heo 25963c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 259773f53c4aSTejun Heo 25984ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25994ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26004ce48b37STejun Heo goto out_unlock; 26014ce48b37STejun Heo 260273f53c4aSTejun Heo wq->first_flusher = NULL; 260373f53c4aSTejun Heo 26046183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26056183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 260673f53c4aSTejun Heo 260773f53c4aSTejun Heo while (true) { 260873f53c4aSTejun Heo struct wq_flusher *next, *tmp; 260973f53c4aSTejun Heo 261073f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 261173f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 261273f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 261373f53c4aSTejun Heo break; 261473f53c4aSTejun Heo list_del_init(&next->list); 261573f53c4aSTejun Heo complete(&next->done); 261673f53c4aSTejun Heo } 261773f53c4aSTejun Heo 26186183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 261973f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 262073f53c4aSTejun Heo 262173f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 262273f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 262373f53c4aSTejun Heo 262473f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 262573f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 262673f53c4aSTejun Heo /* 262773f53c4aSTejun Heo * Assign the same color to all overflowed 262873f53c4aSTejun Heo * flushers, advance work_color and append to 262973f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 263073f53c4aSTejun Heo * phase for these overflowed flushers. 263173f53c4aSTejun Heo */ 263273f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 263373f53c4aSTejun Heo tmp->flush_color = wq->work_color; 263473f53c4aSTejun Heo 263573f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 263673f53c4aSTejun Heo 263773f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 263873f53c4aSTejun Heo &wq->flusher_queue); 2639112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 264073f53c4aSTejun Heo } 264173f53c4aSTejun Heo 264273f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26436183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 264473f53c4aSTejun Heo break; 264573f53c4aSTejun Heo } 264673f53c4aSTejun Heo 264773f53c4aSTejun Heo /* 264873f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2649112202d9STejun Heo * the new first flusher and arm pwqs. 265073f53c4aSTejun Heo */ 26516183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26526183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 265373f53c4aSTejun Heo 265473f53c4aSTejun Heo list_del_init(&next->list); 265573f53c4aSTejun Heo wq->first_flusher = next; 265673f53c4aSTejun Heo 2657112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 265873f53c4aSTejun Heo break; 265973f53c4aSTejun Heo 266073f53c4aSTejun Heo /* 266173f53c4aSTejun Heo * Meh... this color is already done, clear first 266273f53c4aSTejun Heo * flusher and repeat cascading. 266373f53c4aSTejun Heo */ 266473f53c4aSTejun Heo wq->first_flusher = NULL; 266573f53c4aSTejun Heo } 266673f53c4aSTejun Heo 266773f53c4aSTejun Heo out_unlock: 26683c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26691da177e4SLinus Torvalds } 2670ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26711da177e4SLinus Torvalds 26729c5a2ba7STejun Heo /** 26739c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26749c5a2ba7STejun Heo * @wq: workqueue to drain 26759c5a2ba7STejun Heo * 26769c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26779c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26789c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26799c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26809c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26819c5a2ba7STejun Heo * takes too long. 26829c5a2ba7STejun Heo */ 26839c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26849c5a2ba7STejun Heo { 26859c5a2ba7STejun Heo unsigned int flush_cnt = 0; 268649e3cf44STejun Heo struct pool_workqueue *pwq; 26879c5a2ba7STejun Heo 26889c5a2ba7STejun Heo /* 26899c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26909c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2691618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26929c5a2ba7STejun Heo */ 269387fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26949c5a2ba7STejun Heo if (!wq->nr_drainers++) 2695618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 269687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26979c5a2ba7STejun Heo reflush: 26989c5a2ba7STejun Heo flush_workqueue(wq); 26999c5a2ba7STejun Heo 2700b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 270176af4d93STejun Heo 270249e3cf44STejun Heo for_each_pwq(pwq, wq) { 2703fa2563e4SThomas Tuttle bool drained; 27049c5a2ba7STejun Heo 2705b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2706112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2707b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2708fa2563e4SThomas Tuttle 2709fa2563e4SThomas Tuttle if (drained) 27109c5a2ba7STejun Heo continue; 27119c5a2ba7STejun Heo 27129c5a2ba7STejun Heo if (++flush_cnt == 10 || 27139c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2714c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27159c5a2ba7STejun Heo wq->name, flush_cnt); 271676af4d93STejun Heo 2717b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27189c5a2ba7STejun Heo goto reflush; 27199c5a2ba7STejun Heo } 27209c5a2ba7STejun Heo 27219c5a2ba7STejun Heo if (!--wq->nr_drainers) 2722618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 272387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27249c5a2ba7STejun Heo } 27259c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27269c5a2ba7STejun Heo 2727606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2728baf59022STejun Heo { 2729baf59022STejun Heo struct worker *worker = NULL; 2730c9e7cf27STejun Heo struct worker_pool *pool; 2731112202d9STejun Heo struct pool_workqueue *pwq; 2732baf59022STejun Heo 2733baf59022STejun Heo might_sleep(); 2734baf59022STejun Heo 2735fa1b54e6STejun Heo local_irq_disable(); 2736fa1b54e6STejun Heo pool = get_work_pool(work); 2737fa1b54e6STejun Heo if (!pool) { 2738fa1b54e6STejun Heo local_irq_enable(); 2739fa1b54e6STejun Heo return false; 2740fa1b54e6STejun Heo } 2741fa1b54e6STejun Heo 2742fa1b54e6STejun Heo spin_lock(&pool->lock); 27430b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2744112202d9STejun Heo pwq = get_work_pwq(work); 2745112202d9STejun Heo if (pwq) { 2746112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2747baf59022STejun Heo goto already_gone; 2748606a5020STejun Heo } else { 2749c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2750baf59022STejun Heo if (!worker) 2751baf59022STejun Heo goto already_gone; 2752112202d9STejun Heo pwq = worker->current_pwq; 2753606a5020STejun Heo } 2754baf59022STejun Heo 2755112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2756d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2757baf59022STejun Heo 2758e159489bSTejun Heo /* 2759e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2760e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2761e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2762e159489bSTejun Heo * access. 2763e159489bSTejun Heo */ 2764493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2765112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2766e159489bSTejun Heo else 2767112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2768112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2769e159489bSTejun Heo 2770baf59022STejun Heo return true; 2771baf59022STejun Heo already_gone: 2772d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2773baf59022STejun Heo return false; 2774baf59022STejun Heo } 2775baf59022STejun Heo 2776db700897SOleg Nesterov /** 2777401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2778401a8d04STejun Heo * @work: the work to flush 2779db700897SOleg Nesterov * 2780606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2781606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2782401a8d04STejun Heo * 2783401a8d04STejun Heo * RETURNS: 2784401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2785401a8d04STejun Heo * %false if it was already idle. 2786db700897SOleg Nesterov */ 2787401a8d04STejun Heo bool flush_work(struct work_struct *work) 2788db700897SOleg Nesterov { 2789db700897SOleg Nesterov struct wq_barrier barr; 2790db700897SOleg Nesterov 27910976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27920976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27930976dfc1SStephen Boyd 2794606a5020STejun Heo if (start_flush_work(work, &barr)) { 2795db700897SOleg Nesterov wait_for_completion(&barr.done); 2796dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2797401a8d04STejun Heo return true; 2798606a5020STejun Heo } else { 2799401a8d04STejun Heo return false; 2800db700897SOleg Nesterov } 2801606a5020STejun Heo } 2802db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2803db700897SOleg Nesterov 280436e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2805401a8d04STejun Heo { 2806bbb68dfaSTejun Heo unsigned long flags; 28071f1f642eSOleg Nesterov int ret; 28081f1f642eSOleg Nesterov 28091f1f642eSOleg Nesterov do { 2810bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2811bbb68dfaSTejun Heo /* 2812bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2813bbb68dfaSTejun Heo * would be waiting for before retrying. 2814bbb68dfaSTejun Heo */ 2815bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2816606a5020STejun Heo flush_work(work); 28171f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28181f1f642eSOleg Nesterov 2819bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2820bbb68dfaSTejun Heo mark_work_canceling(work); 2821bbb68dfaSTejun Heo local_irq_restore(flags); 2822bbb68dfaSTejun Heo 2823606a5020STejun Heo flush_work(work); 28247a22ad75STejun Heo clear_work_data(work); 28251f1f642eSOleg Nesterov return ret; 28261f1f642eSOleg Nesterov } 28271f1f642eSOleg Nesterov 28286e84d644SOleg Nesterov /** 2829401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2830401a8d04STejun Heo * @work: the work to cancel 28316e84d644SOleg Nesterov * 2832401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2833401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2834401a8d04STejun Heo * another workqueue. On return from this function, @work is 2835401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28361f1f642eSOleg Nesterov * 2837401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2838401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28396e84d644SOleg Nesterov * 2840401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28416e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2842401a8d04STejun Heo * 2843401a8d04STejun Heo * RETURNS: 2844401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28456e84d644SOleg Nesterov */ 2846401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28476e84d644SOleg Nesterov { 284836e227d2STejun Heo return __cancel_work_timer(work, false); 2849b89deed3SOleg Nesterov } 285028e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2851b89deed3SOleg Nesterov 28526e84d644SOleg Nesterov /** 2853401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2854401a8d04STejun Heo * @dwork: the delayed work to flush 28556e84d644SOleg Nesterov * 2856401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2857401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2858401a8d04STejun Heo * considers the last queueing instance of @dwork. 28591f1f642eSOleg Nesterov * 2860401a8d04STejun Heo * RETURNS: 2861401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2862401a8d04STejun Heo * %false if it was already idle. 28636e84d644SOleg Nesterov */ 2864401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2865401a8d04STejun Heo { 28668930cabaSTejun Heo local_irq_disable(); 2867401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 286860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28698930cabaSTejun Heo local_irq_enable(); 2870401a8d04STejun Heo return flush_work(&dwork->work); 2871401a8d04STejun Heo } 2872401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2873401a8d04STejun Heo 2874401a8d04STejun Heo /** 287557b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 287657b30ae7STejun Heo * @dwork: delayed_work to cancel 287709383498STejun Heo * 287857b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 287957b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 288057b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 288157b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 288257b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 288309383498STejun Heo * 288457b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 288509383498STejun Heo */ 288657b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 288709383498STejun Heo { 288857b30ae7STejun Heo unsigned long flags; 288957b30ae7STejun Heo int ret; 289057b30ae7STejun Heo 289157b30ae7STejun Heo do { 289257b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 289357b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 289457b30ae7STejun Heo 289557b30ae7STejun Heo if (unlikely(ret < 0)) 289657b30ae7STejun Heo return false; 289757b30ae7STejun Heo 28987c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28997c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 290057b30ae7STejun Heo local_irq_restore(flags); 2901c0158ca6SDan Magenheimer return ret; 290209383498STejun Heo } 290357b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 290409383498STejun Heo 290509383498STejun Heo /** 2906401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2907401a8d04STejun Heo * @dwork: the delayed work cancel 2908401a8d04STejun Heo * 2909401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2910401a8d04STejun Heo * 2911401a8d04STejun Heo * RETURNS: 2912401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2913401a8d04STejun Heo */ 2914401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29156e84d644SOleg Nesterov { 291636e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29176e84d644SOleg Nesterov } 2918f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29191da177e4SLinus Torvalds 29200fcb78c2SRolf Eike Beer /** 292131ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2922b6136773SAndrew Morton * @func: the function to call 2923b6136773SAndrew Morton * 292431ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 292531ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2926b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 292731ddd871STejun Heo * 292831ddd871STejun Heo * RETURNS: 292931ddd871STejun Heo * 0 on success, -errno on failure. 2930b6136773SAndrew Morton */ 293165f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 293215316ba8SChristoph Lameter { 293315316ba8SChristoph Lameter int cpu; 293438f51568SNamhyung Kim struct work_struct __percpu *works; 293515316ba8SChristoph Lameter 2936b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2937b6136773SAndrew Morton if (!works) 293815316ba8SChristoph Lameter return -ENOMEM; 2939b6136773SAndrew Morton 294095402b38SGautham R Shenoy get_online_cpus(); 294193981800STejun Heo 294215316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29439bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29449bfb1839SIngo Molnar 29459bfb1839SIngo Molnar INIT_WORK(work, func); 29468de6d308SOleg Nesterov schedule_work_on(cpu, work); 294715316ba8SChristoph Lameter } 294893981800STejun Heo 294993981800STejun Heo for_each_online_cpu(cpu) 29508616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 295193981800STejun Heo 295295402b38SGautham R Shenoy put_online_cpus(); 2953b6136773SAndrew Morton free_percpu(works); 295415316ba8SChristoph Lameter return 0; 295515316ba8SChristoph Lameter } 295615316ba8SChristoph Lameter 2957eef6a7d5SAlan Stern /** 2958eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2959eef6a7d5SAlan Stern * 2960eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2961eef6a7d5SAlan Stern * completion. 2962eef6a7d5SAlan Stern * 2963eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2964eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2965eef6a7d5SAlan Stern * will lead to deadlock: 2966eef6a7d5SAlan Stern * 2967eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2968eef6a7d5SAlan Stern * a lock held by your code or its caller. 2969eef6a7d5SAlan Stern * 2970eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2971eef6a7d5SAlan Stern * 2972eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2973eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2974eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2975eef6a7d5SAlan Stern * 2976eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2977eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2978eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2979eef6a7d5SAlan Stern * cancel_work_sync() instead. 2980eef6a7d5SAlan Stern */ 29811da177e4SLinus Torvalds void flush_scheduled_work(void) 29821da177e4SLinus Torvalds { 2983d320c038STejun Heo flush_workqueue(system_wq); 29841da177e4SLinus Torvalds } 2985ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29861da177e4SLinus Torvalds 29871da177e4SLinus Torvalds /** 29881fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29891fa44ecaSJames Bottomley * @fn: the function to execute 29901fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29911fa44ecaSJames Bottomley * be available when the work executes) 29921fa44ecaSJames Bottomley * 29931fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29941fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29951fa44ecaSJames Bottomley * 29961fa44ecaSJames Bottomley * Returns: 0 - function was executed 29971fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29981fa44ecaSJames Bottomley */ 299965f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30001fa44ecaSJames Bottomley { 30011fa44ecaSJames Bottomley if (!in_interrupt()) { 300265f27f38SDavid Howells fn(&ew->work); 30031fa44ecaSJames Bottomley return 0; 30041fa44ecaSJames Bottomley } 30051fa44ecaSJames Bottomley 300665f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30071fa44ecaSJames Bottomley schedule_work(&ew->work); 30081fa44ecaSJames Bottomley 30091fa44ecaSJames Bottomley return 1; 30101fa44ecaSJames Bottomley } 30111fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30121fa44ecaSJames Bottomley 3013226223abSTejun Heo #ifdef CONFIG_SYSFS 3014226223abSTejun Heo /* 3015226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3016226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3017226223abSTejun Heo * following attributes. 3018226223abSTejun Heo * 3019226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3020226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3021226223abSTejun Heo * 3022226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3023226223abSTejun Heo * 3024226223abSTejun Heo * id RO int : the associated pool ID 3025226223abSTejun Heo * nice RW int : nice value of the workers 3026226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3027226223abSTejun Heo */ 3028226223abSTejun Heo struct wq_device { 3029226223abSTejun Heo struct workqueue_struct *wq; 3030226223abSTejun Heo struct device dev; 3031226223abSTejun Heo }; 3032226223abSTejun Heo 3033226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3034226223abSTejun Heo { 3035226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3036226223abSTejun Heo 3037226223abSTejun Heo return wq_dev->wq; 3038226223abSTejun Heo } 3039226223abSTejun Heo 3040226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3041226223abSTejun Heo struct device_attribute *attr, char *buf) 3042226223abSTejun Heo { 3043226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3044226223abSTejun Heo 3045226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3046226223abSTejun Heo } 3047226223abSTejun Heo 3048226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3049226223abSTejun Heo struct device_attribute *attr, char *buf) 3050226223abSTejun Heo { 3051226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3052226223abSTejun Heo 3053226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3054226223abSTejun Heo } 3055226223abSTejun Heo 3056226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3057226223abSTejun Heo struct device_attribute *attr, 3058226223abSTejun Heo const char *buf, size_t count) 3059226223abSTejun Heo { 3060226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3061226223abSTejun Heo int val; 3062226223abSTejun Heo 3063226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3064226223abSTejun Heo return -EINVAL; 3065226223abSTejun Heo 3066226223abSTejun Heo workqueue_set_max_active(wq, val); 3067226223abSTejun Heo return count; 3068226223abSTejun Heo } 3069226223abSTejun Heo 3070226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3071226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3072226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3073226223abSTejun Heo __ATTR_NULL, 3074226223abSTejun Heo }; 3075226223abSTejun Heo 3076226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3077226223abSTejun Heo struct device_attribute *attr, char *buf) 3078226223abSTejun Heo { 3079226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3080226223abSTejun Heo struct worker_pool *pool; 3081226223abSTejun Heo int written; 3082226223abSTejun Heo 3083226223abSTejun Heo rcu_read_lock_sched(); 3084226223abSTejun Heo pool = first_pwq(wq)->pool; 3085226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3086226223abSTejun Heo rcu_read_unlock_sched(); 3087226223abSTejun Heo 3088226223abSTejun Heo return written; 3089226223abSTejun Heo } 3090226223abSTejun Heo 3091226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3092226223abSTejun Heo char *buf) 3093226223abSTejun Heo { 3094226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3095226223abSTejun Heo int written; 3096226223abSTejun Heo 30976029a918STejun Heo mutex_lock(&wq->mutex); 30986029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 30996029a918STejun Heo mutex_unlock(&wq->mutex); 3100226223abSTejun Heo 3101226223abSTejun Heo return written; 3102226223abSTejun Heo } 3103226223abSTejun Heo 3104226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3105226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3106226223abSTejun Heo { 3107226223abSTejun Heo struct workqueue_attrs *attrs; 3108226223abSTejun Heo 3109226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3110226223abSTejun Heo if (!attrs) 3111226223abSTejun Heo return NULL; 3112226223abSTejun Heo 31136029a918STejun Heo mutex_lock(&wq->mutex); 31146029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31156029a918STejun Heo mutex_unlock(&wq->mutex); 3116226223abSTejun Heo return attrs; 3117226223abSTejun Heo } 3118226223abSTejun Heo 3119226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3120226223abSTejun Heo const char *buf, size_t count) 3121226223abSTejun Heo { 3122226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3123226223abSTejun Heo struct workqueue_attrs *attrs; 3124226223abSTejun Heo int ret; 3125226223abSTejun Heo 3126226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3127226223abSTejun Heo if (!attrs) 3128226223abSTejun Heo return -ENOMEM; 3129226223abSTejun Heo 3130226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3131226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3132226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3133226223abSTejun Heo else 3134226223abSTejun Heo ret = -EINVAL; 3135226223abSTejun Heo 3136226223abSTejun Heo free_workqueue_attrs(attrs); 3137226223abSTejun Heo return ret ?: count; 3138226223abSTejun Heo } 3139226223abSTejun Heo 3140226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3141226223abSTejun Heo struct device_attribute *attr, char *buf) 3142226223abSTejun Heo { 3143226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3144226223abSTejun Heo int written; 3145226223abSTejun Heo 31466029a918STejun Heo mutex_lock(&wq->mutex); 31476029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 31486029a918STejun Heo mutex_unlock(&wq->mutex); 3149226223abSTejun Heo 3150226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3151226223abSTejun Heo return written; 3152226223abSTejun Heo } 3153226223abSTejun Heo 3154226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3155226223abSTejun Heo struct device_attribute *attr, 3156226223abSTejun Heo const char *buf, size_t count) 3157226223abSTejun Heo { 3158226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3159226223abSTejun Heo struct workqueue_attrs *attrs; 3160226223abSTejun Heo int ret; 3161226223abSTejun Heo 3162226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3163226223abSTejun Heo if (!attrs) 3164226223abSTejun Heo return -ENOMEM; 3165226223abSTejun Heo 3166226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3167226223abSTejun Heo if (!ret) 3168226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3169226223abSTejun Heo 3170226223abSTejun Heo free_workqueue_attrs(attrs); 3171226223abSTejun Heo return ret ?: count; 3172226223abSTejun Heo } 3173226223abSTejun Heo 3174226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3175226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3176226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3177226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3178226223abSTejun Heo __ATTR_NULL, 3179226223abSTejun Heo }; 3180226223abSTejun Heo 3181226223abSTejun Heo static struct bus_type wq_subsys = { 3182226223abSTejun Heo .name = "workqueue", 3183226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3184226223abSTejun Heo }; 3185226223abSTejun Heo 3186226223abSTejun Heo static int __init wq_sysfs_init(void) 3187226223abSTejun Heo { 3188226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3189226223abSTejun Heo } 3190226223abSTejun Heo core_initcall(wq_sysfs_init); 3191226223abSTejun Heo 3192226223abSTejun Heo static void wq_device_release(struct device *dev) 3193226223abSTejun Heo { 3194226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3195226223abSTejun Heo 3196226223abSTejun Heo kfree(wq_dev); 3197226223abSTejun Heo } 3198226223abSTejun Heo 3199226223abSTejun Heo /** 3200226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3201226223abSTejun Heo * @wq: the workqueue to register 3202226223abSTejun Heo * 3203226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3204226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3205226223abSTejun Heo * which is the preferred method. 3206226223abSTejun Heo * 3207226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3208226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3209226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3210226223abSTejun Heo * attributes. 3211226223abSTejun Heo * 3212226223abSTejun Heo * Returns 0 on success, -errno on failure. 3213226223abSTejun Heo */ 3214226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3215226223abSTejun Heo { 3216226223abSTejun Heo struct wq_device *wq_dev; 3217226223abSTejun Heo int ret; 3218226223abSTejun Heo 3219226223abSTejun Heo /* 3220226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3221226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3222226223abSTejun Heo * workqueues. 3223226223abSTejun Heo */ 3224226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3225226223abSTejun Heo return -EINVAL; 3226226223abSTejun Heo 3227226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3228226223abSTejun Heo if (!wq_dev) 3229226223abSTejun Heo return -ENOMEM; 3230226223abSTejun Heo 3231226223abSTejun Heo wq_dev->wq = wq; 3232226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3233226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3234226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3235226223abSTejun Heo 3236226223abSTejun Heo /* 3237226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3238226223abSTejun Heo * everything is ready. 3239226223abSTejun Heo */ 3240226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3241226223abSTejun Heo 3242226223abSTejun Heo ret = device_register(&wq_dev->dev); 3243226223abSTejun Heo if (ret) { 3244226223abSTejun Heo kfree(wq_dev); 3245226223abSTejun Heo wq->wq_dev = NULL; 3246226223abSTejun Heo return ret; 3247226223abSTejun Heo } 3248226223abSTejun Heo 3249226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3250226223abSTejun Heo struct device_attribute *attr; 3251226223abSTejun Heo 3252226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3253226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3254226223abSTejun Heo if (ret) { 3255226223abSTejun Heo device_unregister(&wq_dev->dev); 3256226223abSTejun Heo wq->wq_dev = NULL; 3257226223abSTejun Heo return ret; 3258226223abSTejun Heo } 3259226223abSTejun Heo } 3260226223abSTejun Heo } 3261226223abSTejun Heo 3262226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3263226223abSTejun Heo return 0; 3264226223abSTejun Heo } 3265226223abSTejun Heo 3266226223abSTejun Heo /** 3267226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3268226223abSTejun Heo * @wq: the workqueue to unregister 3269226223abSTejun Heo * 3270226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3271226223abSTejun Heo */ 3272226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3273226223abSTejun Heo { 3274226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3275226223abSTejun Heo 3276226223abSTejun Heo if (!wq->wq_dev) 3277226223abSTejun Heo return; 3278226223abSTejun Heo 3279226223abSTejun Heo wq->wq_dev = NULL; 3280226223abSTejun Heo device_unregister(&wq_dev->dev); 3281226223abSTejun Heo } 3282226223abSTejun Heo #else /* CONFIG_SYSFS */ 3283226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3284226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3285226223abSTejun Heo 32867a4e344cSTejun Heo /** 32877a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32887a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32897a4e344cSTejun Heo * 32907a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32917a4e344cSTejun Heo */ 32927a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32937a4e344cSTejun Heo { 32947a4e344cSTejun Heo if (attrs) { 32957a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32967a4e344cSTejun Heo kfree(attrs); 32977a4e344cSTejun Heo } 32987a4e344cSTejun Heo } 32997a4e344cSTejun Heo 33007a4e344cSTejun Heo /** 33017a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33027a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33037a4e344cSTejun Heo * 33047a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33057a4e344cSTejun Heo * return it. Returns NULL on failure. 33067a4e344cSTejun Heo */ 33077a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33087a4e344cSTejun Heo { 33097a4e344cSTejun Heo struct workqueue_attrs *attrs; 33107a4e344cSTejun Heo 33117a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33127a4e344cSTejun Heo if (!attrs) 33137a4e344cSTejun Heo goto fail; 33147a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33157a4e344cSTejun Heo goto fail; 33167a4e344cSTejun Heo 331713e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33187a4e344cSTejun Heo return attrs; 33197a4e344cSTejun Heo fail: 33207a4e344cSTejun Heo free_workqueue_attrs(attrs); 33217a4e344cSTejun Heo return NULL; 33227a4e344cSTejun Heo } 33237a4e344cSTejun Heo 332429c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 332529c91e99STejun Heo const struct workqueue_attrs *from) 332629c91e99STejun Heo { 332729c91e99STejun Heo to->nice = from->nice; 332829c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 332929c91e99STejun Heo } 333029c91e99STejun Heo 333129c91e99STejun Heo /* hash value of the content of @attr */ 333229c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 333329c91e99STejun Heo { 333429c91e99STejun Heo u32 hash = 0; 333529c91e99STejun Heo 333629c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 333713e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 333813e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 333929c91e99STejun Heo return hash; 334029c91e99STejun Heo } 334129c91e99STejun Heo 334229c91e99STejun Heo /* content equality test */ 334329c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 334429c91e99STejun Heo const struct workqueue_attrs *b) 334529c91e99STejun Heo { 334629c91e99STejun Heo if (a->nice != b->nice) 334729c91e99STejun Heo return false; 334829c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 334929c91e99STejun Heo return false; 335029c91e99STejun Heo return true; 335129c91e99STejun Heo } 335229c91e99STejun Heo 33537a4e344cSTejun Heo /** 33547a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33557a4e344cSTejun Heo * @pool: worker_pool to initialize 33567a4e344cSTejun Heo * 33577a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 335829c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 335929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 336029c91e99STejun Heo * on @pool safely to release it. 33617a4e344cSTejun Heo */ 33627a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33634e1a1f9aSTejun Heo { 33644e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 336529c91e99STejun Heo pool->id = -1; 336629c91e99STejun Heo pool->cpu = -1; 3367f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 33684e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33694e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33704e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33714e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33724e1a1f9aSTejun Heo 33734e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33744e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33754e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33764e1a1f9aSTejun Heo 33774e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33784e1a1f9aSTejun Heo (unsigned long)pool); 33794e1a1f9aSTejun Heo 33804e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3381bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3382822d8405STejun Heo idr_init(&pool->worker_idr); 33837a4e344cSTejun Heo 338429c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 338529c91e99STejun Heo pool->refcnt = 1; 338629c91e99STejun Heo 338729c91e99STejun Heo /* shouldn't fail above this point */ 33887a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 33897a4e344cSTejun Heo if (!pool->attrs) 33907a4e344cSTejun Heo return -ENOMEM; 33917a4e344cSTejun Heo return 0; 33924e1a1f9aSTejun Heo } 33934e1a1f9aSTejun Heo 339429c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 339529c91e99STejun Heo { 339629c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 339729c91e99STejun Heo 3398822d8405STejun Heo idr_destroy(&pool->worker_idr); 339929c91e99STejun Heo free_workqueue_attrs(pool->attrs); 340029c91e99STejun Heo kfree(pool); 340129c91e99STejun Heo } 340229c91e99STejun Heo 340329c91e99STejun Heo /** 340429c91e99STejun Heo * put_unbound_pool - put a worker_pool 340529c91e99STejun Heo * @pool: worker_pool to put 340629c91e99STejun Heo * 340729c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3408c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3409c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3410c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3411a892caccSTejun Heo * 3412a892caccSTejun Heo * Should be called with wq_pool_mutex held. 341329c91e99STejun Heo */ 341429c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 341529c91e99STejun Heo { 341629c91e99STejun Heo struct worker *worker; 341729c91e99STejun Heo 3418a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3419a892caccSTejun Heo 3420a892caccSTejun Heo if (--pool->refcnt) 342129c91e99STejun Heo return; 342229c91e99STejun Heo 342329c91e99STejun Heo /* sanity checks */ 342429c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3425a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 342629c91e99STejun Heo return; 342729c91e99STejun Heo 342829c91e99STejun Heo /* release id and unhash */ 342929c91e99STejun Heo if (pool->id >= 0) 343029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 343129c91e99STejun Heo hash_del(&pool->hash_node); 343229c91e99STejun Heo 3433c5aa87bbSTejun Heo /* 3434c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3435c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3436c5aa87bbSTejun Heo * manager_mutex. 3437c5aa87bbSTejun Heo */ 343829c91e99STejun Heo mutex_lock(&pool->manager_arb); 3439cd549687STejun Heo mutex_lock(&pool->manager_mutex); 344029c91e99STejun Heo spin_lock_irq(&pool->lock); 344129c91e99STejun Heo 344229c91e99STejun Heo while ((worker = first_worker(pool))) 344329c91e99STejun Heo destroy_worker(worker); 344429c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 344529c91e99STejun Heo 344629c91e99STejun Heo spin_unlock_irq(&pool->lock); 3447cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 344829c91e99STejun Heo mutex_unlock(&pool->manager_arb); 344929c91e99STejun Heo 345029c91e99STejun Heo /* shut down the timers */ 345129c91e99STejun Heo del_timer_sync(&pool->idle_timer); 345229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 345329c91e99STejun Heo 345429c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 345529c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 345629c91e99STejun Heo } 345729c91e99STejun Heo 345829c91e99STejun Heo /** 345929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 346029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 346129c91e99STejun Heo * 346229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 346329c91e99STejun Heo * reference count and return it. If there already is a matching 346429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 346529c91e99STejun Heo * create a new one. On failure, returns NULL. 3466a892caccSTejun Heo * 3467a892caccSTejun Heo * Should be called with wq_pool_mutex held. 346829c91e99STejun Heo */ 346929c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 347029c91e99STejun Heo { 347129c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 347229c91e99STejun Heo struct worker_pool *pool; 3473f3f90ad4STejun Heo int node; 347429c91e99STejun Heo 3475a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 347629c91e99STejun Heo 347729c91e99STejun Heo /* do we already have a matching pool? */ 347829c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 347929c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 348029c91e99STejun Heo pool->refcnt++; 348129c91e99STejun Heo goto out_unlock; 348229c91e99STejun Heo } 348329c91e99STejun Heo } 348429c91e99STejun Heo 348529c91e99STejun Heo /* nope, create a new one */ 348629c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 348729c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 348829c91e99STejun Heo goto fail; 348929c91e99STejun Heo 349012ee4fc6SLai Jiangshan if (workqueue_freezing) 349112ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 349212ee4fc6SLai Jiangshan 34938864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 349429c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 349529c91e99STejun Heo 3496f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3497f3f90ad4STejun Heo if (wq_numa_enabled) { 3498f3f90ad4STejun Heo for_each_node(node) { 3499f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3500f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3501f3f90ad4STejun Heo pool->node = node; 3502f3f90ad4STejun Heo break; 3503f3f90ad4STejun Heo } 3504f3f90ad4STejun Heo } 3505f3f90ad4STejun Heo } 3506f3f90ad4STejun Heo 350729c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 350829c91e99STejun Heo goto fail; 350929c91e99STejun Heo 351029c91e99STejun Heo /* create and start the initial worker */ 3511ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 351229c91e99STejun Heo goto fail; 351329c91e99STejun Heo 351429c91e99STejun Heo /* install */ 351529c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 351629c91e99STejun Heo out_unlock: 351729c91e99STejun Heo return pool; 351829c91e99STejun Heo fail: 351929c91e99STejun Heo if (pool) 352029c91e99STejun Heo put_unbound_pool(pool); 352129c91e99STejun Heo return NULL; 352229c91e99STejun Heo } 352329c91e99STejun Heo 35248864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35258864b4e5STejun Heo { 35268864b4e5STejun Heo kmem_cache_free(pwq_cache, 35278864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35288864b4e5STejun Heo } 35298864b4e5STejun Heo 35308864b4e5STejun Heo /* 35318864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35328864b4e5STejun Heo * and needs to be destroyed. 35338864b4e5STejun Heo */ 35348864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35358864b4e5STejun Heo { 35368864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35378864b4e5STejun Heo unbound_release_work); 35388864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35398864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3540bc0caf09STejun Heo bool is_last; 35418864b4e5STejun Heo 35428864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35438864b4e5STejun Heo return; 35448864b4e5STejun Heo 354575ccf595STejun Heo /* 35463c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 354775ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 354875ccf595STejun Heo * and consistent with the linking path. 354975ccf595STejun Heo */ 35503c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35518864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3552bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35533c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35548864b4e5STejun Heo 3555a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35568864b4e5STejun Heo put_unbound_pool(pool); 3557a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3558a892caccSTejun Heo 35598864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35608864b4e5STejun Heo 35618864b4e5STejun Heo /* 35628864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35638864b4e5STejun Heo * is gonna access it anymore. Free it. 35648864b4e5STejun Heo */ 35656029a918STejun Heo if (is_last) { 35666029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 35678864b4e5STejun Heo kfree(wq); 35688864b4e5STejun Heo } 35696029a918STejun Heo } 35708864b4e5STejun Heo 35710fbd95aaSTejun Heo /** 3572699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35730fbd95aaSTejun Heo * @pwq: target pool_workqueue 35740fbd95aaSTejun Heo * 3575699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3576699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3577699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35780fbd95aaSTejun Heo */ 3579699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35800fbd95aaSTejun Heo { 3581699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3582699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3583699ce097STejun Heo 3584699ce097STejun Heo /* for @wq->saved_max_active */ 3585a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3586699ce097STejun Heo 3587699ce097STejun Heo /* fast exit for non-freezable wqs */ 3588699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3589699ce097STejun Heo return; 3590699ce097STejun Heo 3591a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3592699ce097STejun Heo 3593699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3594699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35950fbd95aaSTejun Heo 35960fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35970fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35980fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3599951a078aSLai Jiangshan 3600951a078aSLai Jiangshan /* 3601951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3602951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3603951a078aSLai Jiangshan */ 3604951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3605699ce097STejun Heo } else { 3606699ce097STejun Heo pwq->max_active = 0; 3607699ce097STejun Heo } 3608699ce097STejun Heo 3609a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36100fbd95aaSTejun Heo } 36110fbd95aaSTejun Heo 3612d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3613d2c1d404STejun Heo struct workqueue_struct *wq, 36149e8cd2f5STejun Heo struct worker_pool *pool, 36159e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3616d2c1d404STejun Heo { 3617d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3618d2c1d404STejun Heo 3619d2c1d404STejun Heo pwq->pool = pool; 3620d2c1d404STejun Heo pwq->wq = wq; 3621d2c1d404STejun Heo pwq->flush_color = -1; 36228864b4e5STejun Heo pwq->refcnt = 1; 3623d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3624d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36258864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3626d2c1d404STejun Heo 36273c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 362875ccf595STejun Heo 3629983ca25eSTejun Heo /* 3630983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36313c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3632983ca25eSTejun Heo */ 36339e8cd2f5STejun Heo if (p_last_pwq) 36349e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 363575ccf595STejun Heo pwq->work_color = wq->work_color; 3636983ca25eSTejun Heo 3637983ca25eSTejun Heo /* sync max_active to the current setting */ 3638983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3639983ca25eSTejun Heo 3640983ca25eSTejun Heo /* link in @pwq */ 36419e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3642a357fc03SLai Jiangshan 36436029a918STejun Heo if (wq->flags & WQ_UNBOUND) 36446029a918STejun Heo copy_workqueue_attrs(wq->unbound_attrs, pool->attrs); 36456029a918STejun Heo 36463c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3647d2c1d404STejun Heo } 3648d2c1d404STejun Heo 36499e8cd2f5STejun Heo /** 36509e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36519e8cd2f5STejun Heo * @wq: the target workqueue 36529e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36539e8cd2f5STejun Heo * 36549e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 36559e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 36569e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 36579e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 36589e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 36599e8cd2f5STejun Heo * 36609e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 36619e8cd2f5STejun Heo * failure. 36629e8cd2f5STejun Heo */ 36639e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36649e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36659e8cd2f5STejun Heo { 366613e2e556STejun Heo struct workqueue_attrs *new_attrs; 366713e2e556STejun Heo struct pool_workqueue *pwq = NULL, *last_pwq; 36689e8cd2f5STejun Heo struct worker_pool *pool; 36694862125bSTejun Heo int ret; 36709e8cd2f5STejun Heo 36718719dceaSTejun Heo /* only unbound workqueues can change attributes */ 36729e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 36739e8cd2f5STejun Heo return -EINVAL; 36749e8cd2f5STejun Heo 36758719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 36768719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 36778719dceaSTejun Heo return -EINVAL; 36788719dceaSTejun Heo 367913e2e556STejun Heo /* make a copy of @attrs and sanitize it */ 368013e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 368113e2e556STejun Heo if (!new_attrs) 368213e2e556STejun Heo goto enomem; 368313e2e556STejun Heo 368413e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 368513e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 368613e2e556STejun Heo 3687a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 3688a892caccSTejun Heo 36899e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 3690a892caccSTejun Heo if (!pwq) { 3691a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 369213e2e556STejun Heo goto enomem; 3693a892caccSTejun Heo } 36949e8cd2f5STejun Heo 369513e2e556STejun Heo pool = get_unbound_pool(new_attrs); 3696a892caccSTejun Heo if (!pool) { 3697a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 369813e2e556STejun Heo goto enomem; 3699a892caccSTejun Heo } 3700a892caccSTejun Heo 3701a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 37029e8cd2f5STejun Heo 37039e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 37049e8cd2f5STejun Heo if (last_pwq) { 37059e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 37069e8cd2f5STejun Heo put_pwq(last_pwq); 37079e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 37089e8cd2f5STejun Heo } 37099e8cd2f5STejun Heo 37104862125bSTejun Heo ret = 0; 37114862125bSTejun Heo /* fall through */ 37124862125bSTejun Heo out_free: 37134862125bSTejun Heo free_workqueue_attrs(new_attrs); 37144862125bSTejun Heo return ret; 371513e2e556STejun Heo 371613e2e556STejun Heo enomem: 371713e2e556STejun Heo kmem_cache_free(pwq_cache, pwq); 37184862125bSTejun Heo ret = -ENOMEM; 37194862125bSTejun Heo goto out_free; 37209e8cd2f5STejun Heo } 37219e8cd2f5STejun Heo 372230cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 37231da177e4SLinus Torvalds { 372449e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 372530cdf249STejun Heo int cpu; 3726e1d8aa9fSFrederic Weisbecker 372730cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3728420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3729420c0ddbSTejun Heo if (!wq->cpu_pwqs) 373030cdf249STejun Heo return -ENOMEM; 373130cdf249STejun Heo 373230cdf249STejun Heo for_each_possible_cpu(cpu) { 37337fb98ea7STejun Heo struct pool_workqueue *pwq = 37347fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37357a62c2c8STejun Heo struct worker_pool *cpu_pools = 3736f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 373730cdf249STejun Heo 37389e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 373930cdf249STejun Heo } 374030cdf249STejun Heo return 0; 37419e8cd2f5STejun Heo } else { 37429e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37439e8cd2f5STejun Heo } 37440f900049STejun Heo } 37450f900049STejun Heo 3746f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3747f3421797STejun Heo const char *name) 3748b71ab8c2STejun Heo { 3749f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3750f3421797STejun Heo 3751f3421797STejun Heo if (max_active < 1 || max_active > lim) 3752044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3753f3421797STejun Heo max_active, name, 1, lim); 3754b71ab8c2STejun Heo 3755f3421797STejun Heo return clamp_val(max_active, 1, lim); 3756b71ab8c2STejun Heo } 3757b71ab8c2STejun Heo 3758b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 375997e37d7bSTejun Heo unsigned int flags, 37601e19ffc6STejun Heo int max_active, 3761eb13ba87SJohannes Berg struct lock_class_key *key, 3762b196be89STejun Heo const char *lock_name, ...) 37633af24433SOleg Nesterov { 3764ecf6881fSTejun Heo va_list args; 37653af24433SOleg Nesterov struct workqueue_struct *wq; 376649e3cf44STejun Heo struct pool_workqueue *pwq; 3767b196be89STejun Heo 3768ecf6881fSTejun Heo /* allocate wq and format name */ 3769ecf6881fSTejun Heo wq = kzalloc(sizeof(*wq), GFP_KERNEL); 3770b196be89STejun Heo if (!wq) 3771d2c1d404STejun Heo return NULL; 3772b196be89STejun Heo 37736029a918STejun Heo if (flags & WQ_UNBOUND) { 37746029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37756029a918STejun Heo if (!wq->unbound_attrs) 37766029a918STejun Heo goto err_free_wq; 37776029a918STejun Heo } 37786029a918STejun Heo 3779ecf6881fSTejun Heo va_start(args, lock_name); 3780ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 3781b196be89STejun Heo va_end(args); 37823af24433SOleg Nesterov 3783d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3784b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 37853af24433SOleg Nesterov 3786b196be89STejun Heo /* init wq */ 378797e37d7bSTejun Heo wq->flags = flags; 3788a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 37893c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3790112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 379130cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 379273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 379373f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3794493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 37953af24433SOleg Nesterov 3796eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3797cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 37983af24433SOleg Nesterov 379930cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3800d2c1d404STejun Heo goto err_free_wq; 38011537663fSTejun Heo 3802493008a8STejun Heo /* 3803493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3804493008a8STejun Heo * have a rescuer to guarantee forward progress. 3805493008a8STejun Heo */ 3806493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3807e22bee78STejun Heo struct worker *rescuer; 3808e22bee78STejun Heo 3809d2c1d404STejun Heo rescuer = alloc_worker(); 3810e22bee78STejun Heo if (!rescuer) 3811d2c1d404STejun Heo goto err_destroy; 3812e22bee78STejun Heo 3813111c225aSTejun Heo rescuer->rescue_wq = wq; 3814111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3815b196be89STejun Heo wq->name); 3816d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3817d2c1d404STejun Heo kfree(rescuer); 3818d2c1d404STejun Heo goto err_destroy; 3819d2c1d404STejun Heo } 3820e22bee78STejun Heo 3821d2c1d404STejun Heo wq->rescuer = rescuer; 382214a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3823e22bee78STejun Heo wake_up_process(rescuer->task); 38243af24433SOleg Nesterov } 38251537663fSTejun Heo 3826226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3827226223abSTejun Heo goto err_destroy; 3828226223abSTejun Heo 38293af24433SOleg Nesterov /* 383068e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 383168e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 383268e13a67SLai Jiangshan * list. 38333af24433SOleg Nesterov */ 383468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3835a0a1a5fdSTejun Heo 3836a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 383749e3cf44STejun Heo for_each_pwq(pwq, wq) 3838699ce097STejun Heo pwq_adjust_max_active(pwq); 3839a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3840a0a1a5fdSTejun Heo 38413af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3842a0a1a5fdSTejun Heo 384368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38443af24433SOleg Nesterov 38453af24433SOleg Nesterov return wq; 3846d2c1d404STejun Heo 3847d2c1d404STejun Heo err_free_wq: 38486029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 38494690c4abSTejun Heo kfree(wq); 3850d2c1d404STejun Heo return NULL; 3851d2c1d404STejun Heo err_destroy: 3852d2c1d404STejun Heo destroy_workqueue(wq); 38534690c4abSTejun Heo return NULL; 38541da177e4SLinus Torvalds } 3855d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 38561da177e4SLinus Torvalds 38573af24433SOleg Nesterov /** 38583af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 38593af24433SOleg Nesterov * @wq: target workqueue 38603af24433SOleg Nesterov * 38613af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 38623af24433SOleg Nesterov */ 38633af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 38643af24433SOleg Nesterov { 386549e3cf44STejun Heo struct pool_workqueue *pwq; 38663af24433SOleg Nesterov 38679c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 38689c5a2ba7STejun Heo drain_workqueue(wq); 3869c8efcc25STejun Heo 38706183c009STejun Heo /* sanity checks */ 3871b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 387249e3cf44STejun Heo for_each_pwq(pwq, wq) { 38736183c009STejun Heo int i; 38746183c009STejun Heo 387576af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 387676af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3877b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38786183c009STejun Heo return; 387976af4d93STejun Heo } 388076af4d93STejun Heo } 388176af4d93STejun Heo 38828864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 38838864b4e5STejun Heo WARN_ON(pwq->nr_active) || 388476af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3885b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38866183c009STejun Heo return; 38876183c009STejun Heo } 388876af4d93STejun Heo } 3889b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38906183c009STejun Heo 3891a0a1a5fdSTejun Heo /* 3892a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3893a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3894a0a1a5fdSTejun Heo */ 389568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3896d2c1d404STejun Heo list_del_init(&wq->list); 389768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38983af24433SOleg Nesterov 3899226223abSTejun Heo workqueue_sysfs_unregister(wq); 3900226223abSTejun Heo 3901493008a8STejun Heo if (wq->rescuer) { 3902e22bee78STejun Heo kthread_stop(wq->rescuer->task); 39038d9df9f0SXiaotian Feng kfree(wq->rescuer); 3904493008a8STejun Heo wq->rescuer = NULL; 3905e22bee78STejun Heo } 3906e22bee78STejun Heo 39078864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 390829c91e99STejun Heo /* 39098864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 39108864b4e5STejun Heo * free the pwqs and wq. 391129c91e99STejun Heo */ 39128864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 39138864b4e5STejun Heo kfree(wq); 39148864b4e5STejun Heo } else { 39158864b4e5STejun Heo /* 39168864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 39178864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 39188864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 39198864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 39208864b4e5STejun Heo * released. 39218864b4e5STejun Heo */ 392229c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 392329c91e99STejun Heo pwqs_node); 39248864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 39258864b4e5STejun Heo put_pwq(pwq); 39268864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 392729c91e99STejun Heo } 39283af24433SOleg Nesterov } 39293af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 39303af24433SOleg Nesterov 3931dcd989cbSTejun Heo /** 3932dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3933dcd989cbSTejun Heo * @wq: target workqueue 3934dcd989cbSTejun Heo * @max_active: new max_active value. 3935dcd989cbSTejun Heo * 3936dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3937dcd989cbSTejun Heo * 3938dcd989cbSTejun Heo * CONTEXT: 3939dcd989cbSTejun Heo * Don't call from IRQ context. 3940dcd989cbSTejun Heo */ 3941dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3942dcd989cbSTejun Heo { 394349e3cf44STejun Heo struct pool_workqueue *pwq; 3944dcd989cbSTejun Heo 39458719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39468719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39478719dceaSTejun Heo return; 39488719dceaSTejun Heo 3949f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3950dcd989cbSTejun Heo 3951a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 3952dcd989cbSTejun Heo 3953dcd989cbSTejun Heo wq->saved_max_active = max_active; 3954dcd989cbSTejun Heo 3955699ce097STejun Heo for_each_pwq(pwq, wq) 3956699ce097STejun Heo pwq_adjust_max_active(pwq); 3957dcd989cbSTejun Heo 3958a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3959dcd989cbSTejun Heo } 3960dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3961dcd989cbSTejun Heo 3962dcd989cbSTejun Heo /** 3963e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 3964e6267616STejun Heo * 3965e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 3966e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 3967e6267616STejun Heo */ 3968e6267616STejun Heo bool current_is_workqueue_rescuer(void) 3969e6267616STejun Heo { 3970e6267616STejun Heo struct worker *worker = current_wq_worker(); 3971e6267616STejun Heo 39726a092dfdSLai Jiangshan return worker && worker->rescue_wq; 3973e6267616STejun Heo } 3974e6267616STejun Heo 3975e6267616STejun Heo /** 3976dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3977dcd989cbSTejun Heo * @cpu: CPU in question 3978dcd989cbSTejun Heo * @wq: target workqueue 3979dcd989cbSTejun Heo * 3980dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3981dcd989cbSTejun Heo * no synchronization around this function and the test result is 3982dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3983dcd989cbSTejun Heo * 3984dcd989cbSTejun Heo * RETURNS: 3985dcd989cbSTejun Heo * %true if congested, %false otherwise. 3986dcd989cbSTejun Heo */ 3987d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 3988dcd989cbSTejun Heo { 39897fb98ea7STejun Heo struct pool_workqueue *pwq; 399076af4d93STejun Heo bool ret; 399176af4d93STejun Heo 399288109453SLai Jiangshan rcu_read_lock_sched(); 39937fb98ea7STejun Heo 39947fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 39957fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 39967fb98ea7STejun Heo else 39977fb98ea7STejun Heo pwq = first_pwq(wq); 3998dcd989cbSTejun Heo 399976af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 400088109453SLai Jiangshan rcu_read_unlock_sched(); 400176af4d93STejun Heo 400276af4d93STejun Heo return ret; 4003dcd989cbSTejun Heo } 4004dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4005dcd989cbSTejun Heo 4006dcd989cbSTejun Heo /** 4007dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4008dcd989cbSTejun Heo * @work: the work to be tested 4009dcd989cbSTejun Heo * 4010dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4011dcd989cbSTejun Heo * synchronization around this function and the test result is 4012dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4013dcd989cbSTejun Heo * 4014dcd989cbSTejun Heo * RETURNS: 4015dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4016dcd989cbSTejun Heo */ 4017dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4018dcd989cbSTejun Heo { 4019fa1b54e6STejun Heo struct worker_pool *pool; 4020dcd989cbSTejun Heo unsigned long flags; 4021dcd989cbSTejun Heo unsigned int ret = 0; 4022dcd989cbSTejun Heo 4023dcd989cbSTejun Heo if (work_pending(work)) 4024dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4025038366c5SLai Jiangshan 4026fa1b54e6STejun Heo local_irq_save(flags); 4027fa1b54e6STejun Heo pool = get_work_pool(work); 4028038366c5SLai Jiangshan if (pool) { 4029fa1b54e6STejun Heo spin_lock(&pool->lock); 4030c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4031dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4032fa1b54e6STejun Heo spin_unlock(&pool->lock); 4033038366c5SLai Jiangshan } 4034fa1b54e6STejun Heo local_irq_restore(flags); 4035dcd989cbSTejun Heo 4036dcd989cbSTejun Heo return ret; 4037dcd989cbSTejun Heo } 4038dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4039dcd989cbSTejun Heo 4040db7bccf4STejun Heo /* 4041db7bccf4STejun Heo * CPU hotplug. 4042db7bccf4STejun Heo * 4043e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4044112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4045706026c2STejun Heo * pool which make migrating pending and scheduled works very 4046e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 404794cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4048e22bee78STejun Heo * blocked draining impractical. 4049e22bee78STejun Heo * 405024647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4051628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4052628c78e7STejun Heo * cpu comes back online. 4053db7bccf4STejun Heo */ 4054db7bccf4STejun Heo 4055706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4056db7bccf4STejun Heo { 405738db41d9STejun Heo int cpu = smp_processor_id(); 40584ce62e9eSTejun Heo struct worker_pool *pool; 4059db7bccf4STejun Heo struct worker *worker; 4060a9ab775bSTejun Heo int wi; 4061db7bccf4STejun Heo 4062f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 40636183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4064db7bccf4STejun Heo 4065bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 406694cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4067e22bee78STejun Heo 4068f2d5a0eeSTejun Heo /* 4069bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 407094cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 407194cf58bbSTejun Heo * except for the ones which are still executing works from 407294cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 407394cf58bbSTejun Heo * this, they may become diasporas. 4074f2d5a0eeSTejun Heo */ 4075a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4076403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4077db7bccf4STejun Heo 407824647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4079f2d5a0eeSTejun Heo 408094cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4081bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 408294cf58bbSTejun Heo } 4083e22bee78STejun Heo 4084e22bee78STejun Heo /* 4085628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4086628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4087628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4088628c78e7STejun Heo */ 4089628c78e7STejun Heo schedule(); 4090628c78e7STejun Heo 4091628c78e7STejun Heo /* 4092628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4093628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 409438db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 409538db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 409638db41d9STejun Heo * pools which are served by workers tied to the CPU. 4097628c78e7STejun Heo * 4098628c78e7STejun Heo * On return from this function, the current worker would trigger 4099628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4100628c78e7STejun Heo * didn't already. 4101e22bee78STejun Heo */ 4102f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4103e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4104db7bccf4STejun Heo } 4105db7bccf4STejun Heo 4106bd7c089eSTejun Heo /** 4107bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4108bd7c089eSTejun Heo * @pool: pool of interest 4109bd7c089eSTejun Heo * 4110a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4111bd7c089eSTejun Heo */ 4112bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4113bd7c089eSTejun Heo { 4114a9ab775bSTejun Heo struct worker *worker; 4115a9ab775bSTejun Heo int wi; 4116bd7c089eSTejun Heo 4117bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4118bd7c089eSTejun Heo 4119bd7c089eSTejun Heo /* 4120a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4121a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4122a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4123a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4124a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4125bd7c089eSTejun Heo */ 4126a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4127a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4128a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4129a9ab775bSTejun Heo 4130a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4131a9ab775bSTejun Heo 4132a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4133a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4134a9ab775bSTejun Heo 4135a9ab775bSTejun Heo /* 4136a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4137a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4138a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4139a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4140a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4141a9ab775bSTejun Heo * be bound before @pool->lock is released. 4142a9ab775bSTejun Heo */ 4143a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4144bd7c089eSTejun Heo wake_up_process(worker->task); 4145bd7c089eSTejun Heo 4146bd7c089eSTejun Heo /* 4147a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4148a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4149a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4150a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4151a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4152a9ab775bSTejun Heo * concurrency management. Note that when or whether 4153a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4154a9ab775bSTejun Heo * 4155a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4156a9ab775bSTejun Heo * tested without holding any lock in 4157a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4158a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4159a9ab775bSTejun Heo * management operations. 4160bd7c089eSTejun Heo */ 4161a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4162a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4163a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4164a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4165bd7c089eSTejun Heo } 4166a9ab775bSTejun Heo 4167a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4168bd7c089eSTejun Heo } 4169bd7c089eSTejun Heo 41707dbc725eSTejun Heo /** 41717dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 41727dbc725eSTejun Heo * @pool: unbound pool of interest 41737dbc725eSTejun Heo * @cpu: the CPU which is coming up 41747dbc725eSTejun Heo * 41757dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 41767dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 41777dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 41787dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 41797dbc725eSTejun Heo */ 41807dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 41817dbc725eSTejun Heo { 41827dbc725eSTejun Heo static cpumask_t cpumask; 41837dbc725eSTejun Heo struct worker *worker; 41847dbc725eSTejun Heo int wi; 41857dbc725eSTejun Heo 41867dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 41877dbc725eSTejun Heo 41887dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 41897dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 41907dbc725eSTejun Heo return; 41917dbc725eSTejun Heo 41927dbc725eSTejun Heo /* is @cpu the only online CPU? */ 41937dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 41947dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 41957dbc725eSTejun Heo return; 41967dbc725eSTejun Heo 41977dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 41987dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 41997dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 42007dbc725eSTejun Heo pool->attrs->cpumask) < 0); 42017dbc725eSTejun Heo } 42027dbc725eSTejun Heo 42038db25e78STejun Heo /* 42048db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 42058db25e78STejun Heo * This will be registered high priority CPU notifier. 42068db25e78STejun Heo */ 42079fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 42081da177e4SLinus Torvalds unsigned long action, 42091da177e4SLinus Torvalds void *hcpu) 42101da177e4SLinus Torvalds { 4211d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42124ce62e9eSTejun Heo struct worker_pool *pool; 42137dbc725eSTejun Heo int pi; 42141da177e4SLinus Torvalds 42158db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 42163af24433SOleg Nesterov case CPU_UP_PREPARE: 4217f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 42183ce63377STejun Heo if (pool->nr_workers) 42193ce63377STejun Heo continue; 4220ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 42213ce63377STejun Heo return NOTIFY_BAD; 42223af24433SOleg Nesterov } 42231da177e4SLinus Torvalds break; 42241da177e4SLinus Torvalds 422565758202STejun Heo case CPU_DOWN_FAILED: 422665758202STejun Heo case CPU_ONLINE: 422768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 42287dbc725eSTejun Heo 42297dbc725eSTejun Heo for_each_pool(pool, pi) { 4230bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 423194cf58bbSTejun Heo 42327dbc725eSTejun Heo if (pool->cpu == cpu) { 4233a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 423424647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4235a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4236a9ab775bSTejun Heo 423794cf58bbSTejun Heo rebind_workers(pool); 42387dbc725eSTejun Heo } else if (pool->cpu < 0) { 42397dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 42407dbc725eSTejun Heo } 424194cf58bbSTejun Heo 4242bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 424394cf58bbSTejun Heo } 42447dbc725eSTejun Heo 424568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42468db25e78STejun Heo break; 424765758202STejun Heo } 424865758202STejun Heo return NOTIFY_OK; 424965758202STejun Heo } 425065758202STejun Heo 425165758202STejun Heo /* 425265758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 425365758202STejun Heo * This will be registered as low priority CPU notifier. 425465758202STejun Heo */ 42559fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 425665758202STejun Heo unsigned long action, 425765758202STejun Heo void *hcpu) 425865758202STejun Heo { 4259d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42608db25e78STejun Heo struct work_struct unbind_work; 42618db25e78STejun Heo 426265758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 426365758202STejun Heo case CPU_DOWN_PREPARE: 42648db25e78STejun Heo /* unbinding should happen on the local CPU */ 4265706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42667635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42678db25e78STejun Heo flush_work(&unbind_work); 42688db25e78STejun Heo break; 426965758202STejun Heo } 427065758202STejun Heo return NOTIFY_OK; 427165758202STejun Heo } 427265758202STejun Heo 42732d3854a3SRusty Russell #ifdef CONFIG_SMP 42748ccad40dSRusty Russell 42752d3854a3SRusty Russell struct work_for_cpu { 4276ed48ece2STejun Heo struct work_struct work; 42772d3854a3SRusty Russell long (*fn)(void *); 42782d3854a3SRusty Russell void *arg; 42792d3854a3SRusty Russell long ret; 42802d3854a3SRusty Russell }; 42812d3854a3SRusty Russell 4282ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 42832d3854a3SRusty Russell { 4284ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4285ed48ece2STejun Heo 42862d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 42872d3854a3SRusty Russell } 42882d3854a3SRusty Russell 42892d3854a3SRusty Russell /** 42902d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 42912d3854a3SRusty Russell * @cpu: the cpu to run on 42922d3854a3SRusty Russell * @fn: the function to run 42932d3854a3SRusty Russell * @arg: the function arg 42942d3854a3SRusty Russell * 429531ad9081SRusty Russell * This will return the value @fn returns. 429631ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 42976b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 42982d3854a3SRusty Russell */ 4299d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 43002d3854a3SRusty Russell { 4301ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 43022d3854a3SRusty Russell 4303ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4304ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4305ed48ece2STejun Heo flush_work(&wfc.work); 43062d3854a3SRusty Russell return wfc.ret; 43072d3854a3SRusty Russell } 43082d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 43092d3854a3SRusty Russell #endif /* CONFIG_SMP */ 43102d3854a3SRusty Russell 4311a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4312e7577c50SRusty Russell 4313a0a1a5fdSTejun Heo /** 4314a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4315a0a1a5fdSTejun Heo * 431658a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4317c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4318706026c2STejun Heo * pool->worklist. 4319a0a1a5fdSTejun Heo * 4320a0a1a5fdSTejun Heo * CONTEXT: 4321a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4322a0a1a5fdSTejun Heo */ 4323a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4324a0a1a5fdSTejun Heo { 432517116969STejun Heo struct worker_pool *pool; 432624b8a847STejun Heo struct workqueue_struct *wq; 432724b8a847STejun Heo struct pool_workqueue *pwq; 4328611c92a0STejun Heo int pi; 4329a0a1a5fdSTejun Heo 433068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4331a0a1a5fdSTejun Heo 43326183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4333a0a1a5fdSTejun Heo workqueue_freezing = true; 4334a0a1a5fdSTejun Heo 433524b8a847STejun Heo /* set FREEZING */ 4336611c92a0STejun Heo for_each_pool(pool, pi) { 43375bcab335STejun Heo spin_lock_irq(&pool->lock); 433835b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 433935b6bb63STejun Heo pool->flags |= POOL_FREEZING; 43405bcab335STejun Heo spin_unlock_irq(&pool->lock); 43411da177e4SLinus Torvalds } 43428b03ae3cSTejun Heo 434324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4344a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4345699ce097STejun Heo for_each_pwq(pwq, wq) 4346699ce097STejun Heo pwq_adjust_max_active(pwq); 4347a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4348a1056305STejun Heo } 43495bcab335STejun Heo 435068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4351a0a1a5fdSTejun Heo } 4352a0a1a5fdSTejun Heo 4353a0a1a5fdSTejun Heo /** 435458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4355a0a1a5fdSTejun Heo * 4356a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4357a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4358a0a1a5fdSTejun Heo * 4359a0a1a5fdSTejun Heo * CONTEXT: 436068e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4361a0a1a5fdSTejun Heo * 4362a0a1a5fdSTejun Heo * RETURNS: 436358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 436458a69cb4STejun Heo * is complete. 4365a0a1a5fdSTejun Heo */ 4366a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4367a0a1a5fdSTejun Heo { 4368a0a1a5fdSTejun Heo bool busy = false; 436924b8a847STejun Heo struct workqueue_struct *wq; 437024b8a847STejun Heo struct pool_workqueue *pwq; 4371a0a1a5fdSTejun Heo 437268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4373a0a1a5fdSTejun Heo 43746183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4375a0a1a5fdSTejun Heo 437624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 437724b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 437824b8a847STejun Heo continue; 4379a0a1a5fdSTejun Heo /* 4380a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4381a0a1a5fdSTejun Heo * to peek without lock. 4382a0a1a5fdSTejun Heo */ 438388109453SLai Jiangshan rcu_read_lock_sched(); 438424b8a847STejun Heo for_each_pwq(pwq, wq) { 43856183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4386112202d9STejun Heo if (pwq->nr_active) { 4387a0a1a5fdSTejun Heo busy = true; 438888109453SLai Jiangshan rcu_read_unlock_sched(); 4389a0a1a5fdSTejun Heo goto out_unlock; 4390a0a1a5fdSTejun Heo } 4391a0a1a5fdSTejun Heo } 439288109453SLai Jiangshan rcu_read_unlock_sched(); 4393a0a1a5fdSTejun Heo } 4394a0a1a5fdSTejun Heo out_unlock: 439568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4396a0a1a5fdSTejun Heo return busy; 4397a0a1a5fdSTejun Heo } 4398a0a1a5fdSTejun Heo 4399a0a1a5fdSTejun Heo /** 4400a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4401a0a1a5fdSTejun Heo * 4402a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4403706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4404a0a1a5fdSTejun Heo * 4405a0a1a5fdSTejun Heo * CONTEXT: 4406a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4407a0a1a5fdSTejun Heo */ 4408a0a1a5fdSTejun Heo void thaw_workqueues(void) 4409a0a1a5fdSTejun Heo { 441024b8a847STejun Heo struct workqueue_struct *wq; 441124b8a847STejun Heo struct pool_workqueue *pwq; 441224b8a847STejun Heo struct worker_pool *pool; 4413611c92a0STejun Heo int pi; 4414a0a1a5fdSTejun Heo 441568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4416a0a1a5fdSTejun Heo 4417a0a1a5fdSTejun Heo if (!workqueue_freezing) 4418a0a1a5fdSTejun Heo goto out_unlock; 4419a0a1a5fdSTejun Heo 442024b8a847STejun Heo /* clear FREEZING */ 4421611c92a0STejun Heo for_each_pool(pool, pi) { 44225bcab335STejun Heo spin_lock_irq(&pool->lock); 442335b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 442435b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 44255bcab335STejun Heo spin_unlock_irq(&pool->lock); 4426d565ed63STejun Heo } 442724b8a847STejun Heo 442824b8a847STejun Heo /* restore max_active and repopulate worklist */ 442924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4430a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4431699ce097STejun Heo for_each_pwq(pwq, wq) 4432699ce097STejun Heo pwq_adjust_max_active(pwq); 4433a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 443424b8a847STejun Heo } 443524b8a847STejun Heo 4436a0a1a5fdSTejun Heo workqueue_freezing = false; 4437a0a1a5fdSTejun Heo out_unlock: 443868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4439a0a1a5fdSTejun Heo } 4440a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4441a0a1a5fdSTejun Heo 4442bce90380STejun Heo static void __init wq_numa_init(void) 4443bce90380STejun Heo { 4444bce90380STejun Heo cpumask_var_t *tbl; 4445bce90380STejun Heo int node, cpu; 4446bce90380STejun Heo 4447bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4448bce90380STejun Heo for_each_node(node) 4449bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4450bce90380STejun Heo 4451bce90380STejun Heo if (num_possible_nodes() <= 1) 4452bce90380STejun Heo return; 4453bce90380STejun Heo 4454bce90380STejun Heo /* 4455bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4456bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4457bce90380STejun Heo * fully initialized by now. 4458bce90380STejun Heo */ 4459bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4460bce90380STejun Heo BUG_ON(!tbl); 4461bce90380STejun Heo 4462bce90380STejun Heo for_each_node(node) 4463bce90380STejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node)); 4464bce90380STejun Heo 4465bce90380STejun Heo for_each_possible_cpu(cpu) { 4466bce90380STejun Heo node = cpu_to_node(cpu); 4467bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4468bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4469bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4470bce90380STejun Heo return; 4471bce90380STejun Heo } 4472bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4473bce90380STejun Heo } 4474bce90380STejun Heo 4475bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4476bce90380STejun Heo wq_numa_enabled = true; 4477bce90380STejun Heo } 4478bce90380STejun Heo 44796ee0578bSSuresh Siddha static int __init init_workqueues(void) 44801da177e4SLinus Torvalds { 44817a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44827a4e344cSTejun Heo int i, cpu; 4483c34056a3STejun Heo 44847c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44857c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44866be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4487b5490077STejun Heo 4488e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4489e904e6c2STejun Heo 4490e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4491e904e6c2STejun Heo 449265758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4493a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44948b03ae3cSTejun Heo 4495bce90380STejun Heo wq_numa_init(); 4496bce90380STejun Heo 4497706026c2STejun Heo /* initialize CPU pools */ 449829c91e99STejun Heo for_each_possible_cpu(cpu) { 44994ce62e9eSTejun Heo struct worker_pool *pool; 45008b03ae3cSTejun Heo 45017a4e344cSTejun Heo i = 0; 4502f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45037a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4504ec22ca5eSTejun Heo pool->cpu = cpu; 45057a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 45067a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4507f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 45087a4e344cSTejun Heo 45099daf9e67STejun Heo /* alloc pool ID */ 451068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45119daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 451268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45134ce62e9eSTejun Heo } 45148b03ae3cSTejun Heo } 45158b03ae3cSTejun Heo 4516e22bee78STejun Heo /* create the initial worker */ 451729c91e99STejun Heo for_each_online_cpu(cpu) { 45184ce62e9eSTejun Heo struct worker_pool *pool; 4519e22bee78STejun Heo 4520f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 452124647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4522ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4523e22bee78STejun Heo } 45244ce62e9eSTejun Heo } 4525e22bee78STejun Heo 452629c91e99STejun Heo /* create default unbound wq attrs */ 452729c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 452829c91e99STejun Heo struct workqueue_attrs *attrs; 452929c91e99STejun Heo 453029c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 453129c91e99STejun Heo attrs->nice = std_nice[i]; 453229c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 453329c91e99STejun Heo } 453429c91e99STejun Heo 4535d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 45361aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4537d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4538f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4539f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 454024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 454124d51addSTejun Heo WQ_FREEZABLE, 0); 45421aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4543ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 45446ee0578bSSuresh Siddha return 0; 45451da177e4SLinus Torvalds } 45466ee0578bSSuresh Siddha early_initcall(init_workqueues); 4547