11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4429c91e99STejun Heo #include <linux/jhash.h> 4542f8570fSSasha Levin #include <linux/hashtable.h> 4676af4d93STejun Heo #include <linux/rculist.h> 47e22bee78STejun Heo 48ea138446STejun Heo #include "workqueue_internal.h" 491da177e4SLinus Torvalds 50c8e55f36STejun Heo enum { 51bc2ae0f5STejun Heo /* 5224647570STejun Heo * worker_pool flags 53bc2ae0f5STejun Heo * 5424647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 55bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 56bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 57bc2ae0f5STejun Heo * is in effect. 58bc2ae0f5STejun Heo * 59bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 60bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6124647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 62bc2ae0f5STejun Heo * 63bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 64bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6524647570STejun Heo * create_worker() is in progress. 66bc2ae0f5STejun Heo */ 6711ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 6824647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 6935b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 70db7bccf4STejun Heo 71c8e55f36STejun Heo /* worker flags */ 72c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 73c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 74c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 75e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 76fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 77f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 78a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 79e22bee78STejun Heo 80a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 81a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 82db7bccf4STejun Heo 83e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 844ce62e9eSTejun Heo 8529c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 86c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 87db7bccf4STejun Heo 88e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 89e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 90e22bee78STejun Heo 913233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 923233cdbdSTejun Heo /* call for help after 10ms 933233cdbdSTejun Heo (min two ticks) */ 94e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 95e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 961da177e4SLinus Torvalds 971da177e4SLinus Torvalds /* 98e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 99e22bee78STejun Heo * all cpus. Give -20. 100e22bee78STejun Heo */ 101e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1023270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 103c8e55f36STejun Heo }; 104c8e55f36STejun Heo 1051da177e4SLinus Torvalds /* 1064690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1074690c4abSTejun Heo * 108e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 109e41e704bSTejun Heo * everyone else. 1104690c4abSTejun Heo * 111e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 112e22bee78STejun Heo * only be modified and accessed from the local cpu. 113e22bee78STejun Heo * 114d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1154690c4abSTejun Heo * 116d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 117d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 118d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 119d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 120e22bee78STejun Heo * 121822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 122822d8405STejun Heo * locks. Reads can happen under either lock. 123822d8405STejun Heo * 12468e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12576af4d93STejun Heo * 12668e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1275bcab335STejun Heo * 1283c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1293c25a55dSLai Jiangshan * 130b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1312e109a28STejun Heo * 1322e109a28STejun Heo * MD: wq_mayday_lock protected. 1334690c4abSTejun Heo */ 1344690c4abSTejun Heo 1352eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 136c34056a3STejun Heo 137bd7bdd43STejun Heo struct worker_pool { 138d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 139d84ff051STejun Heo int cpu; /* I: the associated cpu */ 1409daf9e67STejun Heo int id; /* I: pool ID */ 14111ebea50STejun Heo unsigned int flags; /* X: flags */ 142bd7bdd43STejun Heo 143bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 144bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 145ea1abd61SLai Jiangshan 146ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 147bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 148bd7bdd43STejun Heo 149bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 150bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 151bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 152bd7bdd43STejun Heo 153c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 154c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 155c9e7cf27STejun Heo /* L: hash of busy workers */ 156c9e7cf27STejun Heo 157bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 15834a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 159bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 160822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 161e19e397aSTejun Heo 1627a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16368e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 16468e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1657a4e344cSTejun Heo 166e19e397aSTejun Heo /* 167e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 168e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 169e19e397aSTejun Heo * cacheline. 170e19e397aSTejun Heo */ 171e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17229c91e99STejun Heo 17329c91e99STejun Heo /* 17429c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 17529c91e99STejun Heo * from get_work_pool(). 17629c91e99STejun Heo */ 17729c91e99STejun Heo struct rcu_head rcu; 1788b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1798b03ae3cSTejun Heo 1808b03ae3cSTejun Heo /* 181112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 182112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 183112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 184112202d9STejun Heo * number of flag bits. 1851da177e4SLinus Torvalds */ 186112202d9STejun Heo struct pool_workqueue { 187bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1884690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 18973f53c4aSTejun Heo int work_color; /* L: current color */ 19073f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1918864b4e5STejun Heo int refcnt; /* L: reference count */ 19273f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19373f53c4aSTejun Heo /* L: nr of in_flight works */ 1941e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 195a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 1961e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 1973c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 1982e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 1998864b4e5STejun Heo 2008864b4e5STejun Heo /* 2018864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2028864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2038864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 204b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2058864b4e5STejun Heo */ 2068864b4e5STejun Heo struct work_struct unbound_release_work; 2078864b4e5STejun Heo struct rcu_head rcu; 208e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2091da177e4SLinus Torvalds 2101da177e4SLinus Torvalds /* 21173f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21273f53c4aSTejun Heo */ 21373f53c4aSTejun Heo struct wq_flusher { 2143c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2153c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 21673f53c4aSTejun Heo struct completion done; /* flush completion */ 21773f53c4aSTejun Heo }; 2181da177e4SLinus Torvalds 219226223abSTejun Heo struct wq_device; 220226223abSTejun Heo 22173f53c4aSTejun Heo /* 222c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 223c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2241da177e4SLinus Torvalds */ 2251da177e4SLinus Torvalds struct workqueue_struct { 22687fc741eSLai Jiangshan unsigned int flags; /* WQ: WQ_* flags */ 227420c0ddbSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */ 2283c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 22968e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23073f53c4aSTejun Heo 2313c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2323c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2333c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 234112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2353c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2363c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2373c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 23873f53c4aSTejun Heo 2392e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 240e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 241e22bee78STejun Heo 24287fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 243a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 244226223abSTejun Heo 245226223abSTejun Heo #ifdef CONFIG_SYSFS 246226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 247226223abSTejun Heo #endif 2484e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2494e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2504e6045f1SJohannes Berg #endif 251b196be89STejun Heo char name[]; /* I: workqueue name */ 2521da177e4SLinus Torvalds }; 2531da177e4SLinus Torvalds 254e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 255e904e6c2STejun Heo 25668e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2572e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2585bcab335STejun Heo 25968e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 26068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2617d19c5ceSTejun Heo 2627d19c5ceSTejun Heo /* the per-cpu worker pools */ 2637d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2647d19c5ceSTejun Heo cpu_worker_pools); 2657d19c5ceSTejun Heo 26668e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2677d19c5ceSTejun Heo 26868e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 26929c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 27029c91e99STejun Heo 271c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 27229c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 27329c91e99STejun Heo 274d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 275d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 276044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2771aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 278044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 279d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 280044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 281f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 282044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 28324d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 284d320c038STejun Heo 2857d19c5ceSTejun Heo static int worker_thread(void *__worker); 2867d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 2877d19c5ceSTejun Heo const struct workqueue_attrs *from); 2887d19c5ceSTejun Heo 28997bd2347STejun Heo #define CREATE_TRACE_POINTS 29097bd2347STejun Heo #include <trace/events/workqueue.h> 29197bd2347STejun Heo 29268e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 2935bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 29468e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 29568e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 2965bcab335STejun Heo 297b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 29876af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 299b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 300b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 30176af4d93STejun Heo 302822d8405STejun Heo #ifdef CONFIG_LOCKDEP 303822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 304519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 305519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 306822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 307822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 308822d8405STejun Heo #else 309822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 310822d8405STejun Heo #endif 311822d8405STejun Heo 312f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 313f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 314f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3157a62c2c8STejun Heo (pool)++) 3164ce62e9eSTejun Heo 31749e3cf44STejun Heo /** 31817116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 31917116969STejun Heo * @pool: iteration cursor 320611c92a0STejun Heo * @pi: integer used for iteration 321fa1b54e6STejun Heo * 32268e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 32368e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 32468e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 325fa1b54e6STejun Heo * 326fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 327fa1b54e6STejun Heo * ignored. 32817116969STejun Heo */ 329611c92a0STejun Heo #define for_each_pool(pool, pi) \ 330611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 33168e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 332fa1b54e6STejun Heo else 33317116969STejun Heo 33417116969STejun Heo /** 335822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 336822d8405STejun Heo * @worker: iteration cursor 337822d8405STejun Heo * @wi: integer used for iteration 338822d8405STejun Heo * @pool: worker_pool to iterate workers of 339822d8405STejun Heo * 340822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 341822d8405STejun Heo * 342822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 343822d8405STejun Heo * ignored. 344822d8405STejun Heo */ 345822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 346822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 347822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 348822d8405STejun Heo else 349822d8405STejun Heo 350822d8405STejun Heo /** 35149e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 35249e3cf44STejun Heo * @pwq: iteration cursor 35349e3cf44STejun Heo * @wq: the target workqueue 35476af4d93STejun Heo * 355b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 356794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 357794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 35876af4d93STejun Heo * 35976af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 36076af4d93STejun Heo * ignored. 36149e3cf44STejun Heo */ 36249e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 36376af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 364b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 36576af4d93STejun Heo else 366f3421797STejun Heo 367dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 368dc186ad7SThomas Gleixner 369dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 370dc186ad7SThomas Gleixner 37199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 37299777288SStanislaw Gruszka { 37399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 37499777288SStanislaw Gruszka } 37599777288SStanislaw Gruszka 376dc186ad7SThomas Gleixner /* 377dc186ad7SThomas Gleixner * fixup_init is called when: 378dc186ad7SThomas Gleixner * - an active object is initialized 379dc186ad7SThomas Gleixner */ 380dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 381dc186ad7SThomas Gleixner { 382dc186ad7SThomas Gleixner struct work_struct *work = addr; 383dc186ad7SThomas Gleixner 384dc186ad7SThomas Gleixner switch (state) { 385dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 386dc186ad7SThomas Gleixner cancel_work_sync(work); 387dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 388dc186ad7SThomas Gleixner return 1; 389dc186ad7SThomas Gleixner default: 390dc186ad7SThomas Gleixner return 0; 391dc186ad7SThomas Gleixner } 392dc186ad7SThomas Gleixner } 393dc186ad7SThomas Gleixner 394dc186ad7SThomas Gleixner /* 395dc186ad7SThomas Gleixner * fixup_activate is called when: 396dc186ad7SThomas Gleixner * - an active object is activated 397dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 398dc186ad7SThomas Gleixner */ 399dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 400dc186ad7SThomas Gleixner { 401dc186ad7SThomas Gleixner struct work_struct *work = addr; 402dc186ad7SThomas Gleixner 403dc186ad7SThomas Gleixner switch (state) { 404dc186ad7SThomas Gleixner 405dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 406dc186ad7SThomas Gleixner /* 407dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 408dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 409dc186ad7SThomas Gleixner * is tracked in the object tracker. 410dc186ad7SThomas Gleixner */ 41122df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 412dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 413dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 414dc186ad7SThomas Gleixner return 0; 415dc186ad7SThomas Gleixner } 416dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 417dc186ad7SThomas Gleixner return 0; 418dc186ad7SThomas Gleixner 419dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 420dc186ad7SThomas Gleixner WARN_ON(1); 421dc186ad7SThomas Gleixner 422dc186ad7SThomas Gleixner default: 423dc186ad7SThomas Gleixner return 0; 424dc186ad7SThomas Gleixner } 425dc186ad7SThomas Gleixner } 426dc186ad7SThomas Gleixner 427dc186ad7SThomas Gleixner /* 428dc186ad7SThomas Gleixner * fixup_free is called when: 429dc186ad7SThomas Gleixner * - an active object is freed 430dc186ad7SThomas Gleixner */ 431dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 432dc186ad7SThomas Gleixner { 433dc186ad7SThomas Gleixner struct work_struct *work = addr; 434dc186ad7SThomas Gleixner 435dc186ad7SThomas Gleixner switch (state) { 436dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 437dc186ad7SThomas Gleixner cancel_work_sync(work); 438dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 439dc186ad7SThomas Gleixner return 1; 440dc186ad7SThomas Gleixner default: 441dc186ad7SThomas Gleixner return 0; 442dc186ad7SThomas Gleixner } 443dc186ad7SThomas Gleixner } 444dc186ad7SThomas Gleixner 445dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 446dc186ad7SThomas Gleixner .name = "work_struct", 44799777288SStanislaw Gruszka .debug_hint = work_debug_hint, 448dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 449dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 450dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 451dc186ad7SThomas Gleixner }; 452dc186ad7SThomas Gleixner 453dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 454dc186ad7SThomas Gleixner { 455dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 456dc186ad7SThomas Gleixner } 457dc186ad7SThomas Gleixner 458dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 459dc186ad7SThomas Gleixner { 460dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 464dc186ad7SThomas Gleixner { 465dc186ad7SThomas Gleixner if (onstack) 466dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 467dc186ad7SThomas Gleixner else 468dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 469dc186ad7SThomas Gleixner } 470dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 471dc186ad7SThomas Gleixner 472dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 473dc186ad7SThomas Gleixner { 474dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 475dc186ad7SThomas Gleixner } 476dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 477dc186ad7SThomas Gleixner 478dc186ad7SThomas Gleixner #else 479dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 480dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 481dc186ad7SThomas Gleixner #endif 482dc186ad7SThomas Gleixner 4839daf9e67STejun Heo /* allocate ID and assign it to @pool */ 4849daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 4859daf9e67STejun Heo { 4869daf9e67STejun Heo int ret; 4879daf9e67STejun Heo 48868e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4895bcab335STejun Heo 490fa1b54e6STejun Heo do { 491fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 492fa1b54e6STejun Heo return -ENOMEM; 4939daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 494fa1b54e6STejun Heo } while (ret == -EAGAIN); 4959daf9e67STejun Heo 4969daf9e67STejun Heo return ret; 4979daf9e67STejun Heo } 4989daf9e67STejun Heo 49976af4d93STejun Heo /** 50076af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 50176af4d93STejun Heo * @wq: the target workqueue 50276af4d93STejun Heo * 503b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 504794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 505794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 50676af4d93STejun Heo */ 5077fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 508a848e3b6SOleg Nesterov { 509b09f4fd3SLai Jiangshan assert_rcu_or_wq_mutex(wq); 51076af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 51176af4d93STejun Heo pwqs_node); 512f3421797STejun Heo } 513a848e3b6SOleg Nesterov 51473f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 51573f53c4aSTejun Heo { 51673f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 51773f53c4aSTejun Heo } 51873f53c4aSTejun Heo 51973f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 52073f53c4aSTejun Heo { 52173f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 52273f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 52373f53c4aSTejun Heo } 52473f53c4aSTejun Heo 52573f53c4aSTejun Heo static int work_next_color(int color) 52673f53c4aSTejun Heo { 52773f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 528b1f4ec17SOleg Nesterov } 529b1f4ec17SOleg Nesterov 5304594bf15SDavid Howells /* 531112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 532112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5337c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5347a22ad75STejun Heo * 535112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 536112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 537bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 538bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5397a22ad75STejun Heo * 540112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5417c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 542112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5437c3eed5cSTejun Heo * available only while the work item is queued. 544bbb68dfaSTejun Heo * 545bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 546bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 547bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 548bbb68dfaSTejun Heo * try to steal the PENDING bit. 5494594bf15SDavid Howells */ 5507a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5517a22ad75STejun Heo unsigned long flags) 5527a22ad75STejun Heo { 5536183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5547a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5557a22ad75STejun Heo } 5567a22ad75STejun Heo 557112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5584690c4abSTejun Heo unsigned long extra_flags) 559365970a1SDavid Howells { 560112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 561112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 562365970a1SDavid Howells } 563365970a1SDavid Howells 5644468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5654468a00fSLai Jiangshan int pool_id) 5664468a00fSLai Jiangshan { 5674468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5684468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5694468a00fSLai Jiangshan } 5704468a00fSLai Jiangshan 5717c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5727c3eed5cSTejun Heo int pool_id) 5734d707b9fSOleg Nesterov { 57423657bb1STejun Heo /* 57523657bb1STejun Heo * The following wmb is paired with the implied mb in 57623657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 57723657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 57823657bb1STejun Heo * owner. 57923657bb1STejun Heo */ 58023657bb1STejun Heo smp_wmb(); 5817c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 5824d707b9fSOleg Nesterov } 5834d707b9fSOleg Nesterov 5847a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 585365970a1SDavid Howells { 5867c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 5877c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 5887a22ad75STejun Heo } 5897a22ad75STejun Heo 590112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 5917a22ad75STejun Heo { 592e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5937a22ad75STejun Heo 594112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 595e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 596e120153dSTejun Heo else 597e120153dSTejun Heo return NULL; 5987a22ad75STejun Heo } 5997a22ad75STejun Heo 6007c3eed5cSTejun Heo /** 6017c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6027c3eed5cSTejun Heo * @work: the work item of interest 6037c3eed5cSTejun Heo * 6047c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 605fa1b54e6STejun Heo * 60668e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 60768e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 60868e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 609fa1b54e6STejun Heo * 610fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 611fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 612fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 613fa1b54e6STejun Heo * returned pool is and stays online. 6147c3eed5cSTejun Heo */ 6157c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6167a22ad75STejun Heo { 617e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6187c3eed5cSTejun Heo int pool_id; 6197a22ad75STejun Heo 62068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 621fa1b54e6STejun Heo 622112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 623112202d9STejun Heo return ((struct pool_workqueue *) 6247c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6257a22ad75STejun Heo 6267c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6277c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6287a22ad75STejun Heo return NULL; 6297a22ad75STejun Heo 630fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6317c3eed5cSTejun Heo } 6327c3eed5cSTejun Heo 6337c3eed5cSTejun Heo /** 6347c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6357c3eed5cSTejun Heo * @work: the work item of interest 6367c3eed5cSTejun Heo * 6377c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6387c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6397c3eed5cSTejun Heo */ 6407c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6417c3eed5cSTejun Heo { 64254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6437c3eed5cSTejun Heo 644112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 645112202d9STejun Heo return ((struct pool_workqueue *) 64654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 64754d5b7d0SLai Jiangshan 64854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6497c3eed5cSTejun Heo } 6507c3eed5cSTejun Heo 651bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 652bbb68dfaSTejun Heo { 6537c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 654bbb68dfaSTejun Heo 6557c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6567c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 657bbb68dfaSTejun Heo } 658bbb68dfaSTejun Heo 659bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 660bbb68dfaSTejun Heo { 661bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 662bbb68dfaSTejun Heo 663112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 664bbb68dfaSTejun Heo } 665bbb68dfaSTejun Heo 666e22bee78STejun Heo /* 6673270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6683270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 669d565ed63STejun Heo * they're being called with pool->lock held. 670e22bee78STejun Heo */ 671e22bee78STejun Heo 67263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 673649027d7STejun Heo { 674e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 675649027d7STejun Heo } 676649027d7STejun Heo 677e22bee78STejun Heo /* 678e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 679e22bee78STejun Heo * running workers. 680974271c4STejun Heo * 681974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 682706026c2STejun Heo * function will always return %true for unbound pools as long as the 683974271c4STejun Heo * worklist isn't empty. 684e22bee78STejun Heo */ 68563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 686e22bee78STejun Heo { 68763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 688e22bee78STejun Heo } 689e22bee78STejun Heo 690e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 69163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 692e22bee78STejun Heo { 69363d95a91STejun Heo return pool->nr_idle; 694e22bee78STejun Heo } 695e22bee78STejun Heo 696e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 69763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 698e22bee78STejun Heo { 699e19e397aSTejun Heo return !list_empty(&pool->worklist) && 700e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 701e22bee78STejun Heo } 702e22bee78STejun Heo 703e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 70463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 705e22bee78STejun Heo { 70663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 707e22bee78STejun Heo } 708e22bee78STejun Heo 709e22bee78STejun Heo /* Do I need to be the manager? */ 71063d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 711e22bee78STejun Heo { 71263d95a91STejun Heo return need_to_create_worker(pool) || 71311ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 714e22bee78STejun Heo } 715e22bee78STejun Heo 716e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 71763d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 718e22bee78STejun Heo { 71934a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 72063d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 72163d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 722e22bee78STejun Heo 723ea1abd61SLai Jiangshan /* 724ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 725ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 726ea1abd61SLai Jiangshan */ 727ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 728ea1abd61SLai Jiangshan return false; 729ea1abd61SLai Jiangshan 730e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 731e22bee78STejun Heo } 732e22bee78STejun Heo 733e22bee78STejun Heo /* 734e22bee78STejun Heo * Wake up functions. 735e22bee78STejun Heo */ 736e22bee78STejun Heo 7377e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 73863d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7397e11629dSTejun Heo { 74063d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7417e11629dSTejun Heo return NULL; 7427e11629dSTejun Heo 74363d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7447e11629dSTejun Heo } 7457e11629dSTejun Heo 7467e11629dSTejun Heo /** 7477e11629dSTejun Heo * wake_up_worker - wake up an idle worker 74863d95a91STejun Heo * @pool: worker pool to wake worker from 7497e11629dSTejun Heo * 75063d95a91STejun Heo * Wake up the first idle worker of @pool. 7517e11629dSTejun Heo * 7527e11629dSTejun Heo * CONTEXT: 753d565ed63STejun Heo * spin_lock_irq(pool->lock). 7547e11629dSTejun Heo */ 75563d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7567e11629dSTejun Heo { 75763d95a91STejun Heo struct worker *worker = first_worker(pool); 7587e11629dSTejun Heo 7597e11629dSTejun Heo if (likely(worker)) 7607e11629dSTejun Heo wake_up_process(worker->task); 7617e11629dSTejun Heo } 7627e11629dSTejun Heo 7634690c4abSTejun Heo /** 764e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 765e22bee78STejun Heo * @task: task waking up 766e22bee78STejun Heo * @cpu: CPU @task is waking up to 767e22bee78STejun Heo * 768e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 769e22bee78STejun Heo * being awoken. 770e22bee78STejun Heo * 771e22bee78STejun Heo * CONTEXT: 772e22bee78STejun Heo * spin_lock_irq(rq->lock) 773e22bee78STejun Heo */ 774d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 775e22bee78STejun Heo { 776e22bee78STejun Heo struct worker *worker = kthread_data(task); 777e22bee78STejun Heo 77836576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 779ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 780e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 781e22bee78STejun Heo } 78236576000SJoonsoo Kim } 783e22bee78STejun Heo 784e22bee78STejun Heo /** 785e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 786e22bee78STejun Heo * @task: task going to sleep 787e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 788e22bee78STejun Heo * 789e22bee78STejun Heo * This function is called during schedule() when a busy worker is 790e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 791e22bee78STejun Heo * returning pointer to its task. 792e22bee78STejun Heo * 793e22bee78STejun Heo * CONTEXT: 794e22bee78STejun Heo * spin_lock_irq(rq->lock) 795e22bee78STejun Heo * 796e22bee78STejun Heo * RETURNS: 797e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 798e22bee78STejun Heo */ 799d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 800e22bee78STejun Heo { 801e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 802111c225aSTejun Heo struct worker_pool *pool; 803e22bee78STejun Heo 804111c225aSTejun Heo /* 805111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 806111c225aSTejun Heo * workers, also reach here, let's not access anything before 807111c225aSTejun Heo * checking NOT_RUNNING. 808111c225aSTejun Heo */ 8092d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 810e22bee78STejun Heo return NULL; 811e22bee78STejun Heo 812111c225aSTejun Heo pool = worker->pool; 813111c225aSTejun Heo 814e22bee78STejun Heo /* this can only happen on the local cpu */ 8156183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8166183c009STejun Heo return NULL; 817e22bee78STejun Heo 818e22bee78STejun Heo /* 819e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 820e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 821e22bee78STejun Heo * Please read comment there. 822e22bee78STejun Heo * 823628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 824628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 825628c78e7STejun Heo * disabled, which in turn means that none else could be 826d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 827628c78e7STejun Heo * lock is safe. 828e22bee78STejun Heo */ 829e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 830e19e397aSTejun Heo !list_empty(&pool->worklist)) 83163d95a91STejun Heo to_wakeup = first_worker(pool); 832e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 833e22bee78STejun Heo } 834e22bee78STejun Heo 835e22bee78STejun Heo /** 836e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 837cb444766STejun Heo * @worker: self 838d302f017STejun Heo * @flags: flags to set 839d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 840d302f017STejun Heo * 841e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 842e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 843e22bee78STejun Heo * woken up. 844d302f017STejun Heo * 845cb444766STejun Heo * CONTEXT: 846d565ed63STejun Heo * spin_lock_irq(pool->lock) 847d302f017STejun Heo */ 848d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 849d302f017STejun Heo bool wakeup) 850d302f017STejun Heo { 851bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 852e22bee78STejun Heo 853cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 854cb444766STejun Heo 855e22bee78STejun Heo /* 856e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 857e22bee78STejun Heo * wake up an idle worker as necessary if requested by 858e22bee78STejun Heo * @wakeup. 859e22bee78STejun Heo */ 860e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 861e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 862e22bee78STejun Heo if (wakeup) { 863e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 864bd7bdd43STejun Heo !list_empty(&pool->worklist)) 86563d95a91STejun Heo wake_up_worker(pool); 866e22bee78STejun Heo } else 867e19e397aSTejun Heo atomic_dec(&pool->nr_running); 868e22bee78STejun Heo } 869e22bee78STejun Heo 870d302f017STejun Heo worker->flags |= flags; 871d302f017STejun Heo } 872d302f017STejun Heo 873d302f017STejun Heo /** 874e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 875cb444766STejun Heo * @worker: self 876d302f017STejun Heo * @flags: flags to clear 877d302f017STejun Heo * 878e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 879d302f017STejun Heo * 880cb444766STejun Heo * CONTEXT: 881d565ed63STejun Heo * spin_lock_irq(pool->lock) 882d302f017STejun Heo */ 883d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 884d302f017STejun Heo { 88563d95a91STejun Heo struct worker_pool *pool = worker->pool; 886e22bee78STejun Heo unsigned int oflags = worker->flags; 887e22bee78STejun Heo 888cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 889cb444766STejun Heo 890d302f017STejun Heo worker->flags &= ~flags; 891e22bee78STejun Heo 89242c025f3STejun Heo /* 89342c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 89442c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 89542c025f3STejun Heo * of multiple flags, not a single flag. 89642c025f3STejun Heo */ 897e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 898e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 899e19e397aSTejun Heo atomic_inc(&pool->nr_running); 900d302f017STejun Heo } 901d302f017STejun Heo 902d302f017STejun Heo /** 9038cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 904c9e7cf27STejun Heo * @pool: pool of interest 9058cca0eeaSTejun Heo * @work: work to find worker for 9068cca0eeaSTejun Heo * 907c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 908c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 909a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 910a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 911a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 912a2c1c57bSTejun Heo * being executed. 913a2c1c57bSTejun Heo * 914a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 915a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 916a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 917a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 918a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 919a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 920a2c1c57bSTejun Heo * 921c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 922c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 923c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 924c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 925c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 926c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9278cca0eeaSTejun Heo * 9288cca0eeaSTejun Heo * CONTEXT: 929d565ed63STejun Heo * spin_lock_irq(pool->lock). 9308cca0eeaSTejun Heo * 9318cca0eeaSTejun Heo * RETURNS: 9328cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9338cca0eeaSTejun Heo * otherwise. 9348cca0eeaSTejun Heo */ 935c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9368cca0eeaSTejun Heo struct work_struct *work) 9378cca0eeaSTejun Heo { 93842f8570fSSasha Levin struct worker *worker; 93942f8570fSSasha Levin 940b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 941a2c1c57bSTejun Heo (unsigned long)work) 942a2c1c57bSTejun Heo if (worker->current_work == work && 943a2c1c57bSTejun Heo worker->current_func == work->func) 94442f8570fSSasha Levin return worker; 94542f8570fSSasha Levin 94642f8570fSSasha Levin return NULL; 9478cca0eeaSTejun Heo } 9488cca0eeaSTejun Heo 9498cca0eeaSTejun Heo /** 950bf4ede01STejun Heo * move_linked_works - move linked works to a list 951bf4ede01STejun Heo * @work: start of series of works to be scheduled 952bf4ede01STejun Heo * @head: target list to append @work to 953bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 954bf4ede01STejun Heo * 955bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 956bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 957bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 958bf4ede01STejun Heo * 959bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 960bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 961bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 962bf4ede01STejun Heo * 963bf4ede01STejun Heo * CONTEXT: 964d565ed63STejun Heo * spin_lock_irq(pool->lock). 965bf4ede01STejun Heo */ 966bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 967bf4ede01STejun Heo struct work_struct **nextp) 968bf4ede01STejun Heo { 969bf4ede01STejun Heo struct work_struct *n; 970bf4ede01STejun Heo 971bf4ede01STejun Heo /* 972bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 973bf4ede01STejun Heo * use NULL for list head. 974bf4ede01STejun Heo */ 975bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 976bf4ede01STejun Heo list_move_tail(&work->entry, head); 977bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 978bf4ede01STejun Heo break; 979bf4ede01STejun Heo } 980bf4ede01STejun Heo 981bf4ede01STejun Heo /* 982bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 983bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 984bf4ede01STejun Heo * needs to be updated. 985bf4ede01STejun Heo */ 986bf4ede01STejun Heo if (nextp) 987bf4ede01STejun Heo *nextp = n; 988bf4ede01STejun Heo } 989bf4ede01STejun Heo 9908864b4e5STejun Heo /** 9918864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 9928864b4e5STejun Heo * @pwq: pool_workqueue to get 9938864b4e5STejun Heo * 9948864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 9958864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 9968864b4e5STejun Heo */ 9978864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 9988864b4e5STejun Heo { 9998864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10008864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10018864b4e5STejun Heo pwq->refcnt++; 10028864b4e5STejun Heo } 10038864b4e5STejun Heo 10048864b4e5STejun Heo /** 10058864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10068864b4e5STejun Heo * @pwq: pool_workqueue to put 10078864b4e5STejun Heo * 10088864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10098864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10108864b4e5STejun Heo */ 10118864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10128864b4e5STejun Heo { 10138864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10148864b4e5STejun Heo if (likely(--pwq->refcnt)) 10158864b4e5STejun Heo return; 10168864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10178864b4e5STejun Heo return; 10188864b4e5STejun Heo /* 10198864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10208864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10218864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10228864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10238864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10248864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10258864b4e5STejun Heo */ 10268864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10278864b4e5STejun Heo } 10288864b4e5STejun Heo 1029112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1030bf4ede01STejun Heo { 1031112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1032bf4ede01STejun Heo 1033bf4ede01STejun Heo trace_workqueue_activate_work(work); 1034112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1035bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1036112202d9STejun Heo pwq->nr_active++; 1037bf4ede01STejun Heo } 1038bf4ede01STejun Heo 1039112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10403aa62497SLai Jiangshan { 1041112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10423aa62497SLai Jiangshan struct work_struct, entry); 10433aa62497SLai Jiangshan 1044112202d9STejun Heo pwq_activate_delayed_work(work); 10453aa62497SLai Jiangshan } 10463aa62497SLai Jiangshan 1047bf4ede01STejun Heo /** 1048112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1049112202d9STejun Heo * @pwq: pwq of interest 1050bf4ede01STejun Heo * @color: color of work which left the queue 1051bf4ede01STejun Heo * 1052bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1053112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1054bf4ede01STejun Heo * 1055bf4ede01STejun Heo * CONTEXT: 1056d565ed63STejun Heo * spin_lock_irq(pool->lock). 1057bf4ede01STejun Heo */ 1058112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1059bf4ede01STejun Heo { 10608864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1061bf4ede01STejun Heo if (color == WORK_NO_COLOR) 10628864b4e5STejun Heo goto out_put; 1063bf4ede01STejun Heo 1064112202d9STejun Heo pwq->nr_in_flight[color]--; 1065bf4ede01STejun Heo 1066112202d9STejun Heo pwq->nr_active--; 1067112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1068bf4ede01STejun Heo /* one down, submit a delayed one */ 1069112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1070112202d9STejun Heo pwq_activate_first_delayed(pwq); 1071bf4ede01STejun Heo } 1072bf4ede01STejun Heo 1073bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1074112202d9STejun Heo if (likely(pwq->flush_color != color)) 10758864b4e5STejun Heo goto out_put; 1076bf4ede01STejun Heo 1077bf4ede01STejun Heo /* are there still in-flight works? */ 1078112202d9STejun Heo if (pwq->nr_in_flight[color]) 10798864b4e5STejun Heo goto out_put; 1080bf4ede01STejun Heo 1081112202d9STejun Heo /* this pwq is done, clear flush_color */ 1082112202d9STejun Heo pwq->flush_color = -1; 1083bf4ede01STejun Heo 1084bf4ede01STejun Heo /* 1085112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1086bf4ede01STejun Heo * will handle the rest. 1087bf4ede01STejun Heo */ 1088112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1089112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 10908864b4e5STejun Heo out_put: 10918864b4e5STejun Heo put_pwq(pwq); 1092bf4ede01STejun Heo } 1093bf4ede01STejun Heo 109436e227d2STejun Heo /** 1095bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 109636e227d2STejun Heo * @work: work item to steal 109736e227d2STejun Heo * @is_dwork: @work is a delayed_work 1098bbb68dfaSTejun Heo * @flags: place to store irq state 109936e227d2STejun Heo * 110036e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 110136e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 110236e227d2STejun Heo * 110336e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 110436e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 110536e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1106bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1107bbb68dfaSTejun Heo * for arbitrarily long 110836e227d2STejun Heo * 1109bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1110e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1111e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1112e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1113bbb68dfaSTejun Heo * 1114bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1115bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1116bbb68dfaSTejun Heo * 1117e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1118bf4ede01STejun Heo */ 1119bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1120bbb68dfaSTejun Heo unsigned long *flags) 1121bf4ede01STejun Heo { 1122d565ed63STejun Heo struct worker_pool *pool; 1123112202d9STejun Heo struct pool_workqueue *pwq; 1124bf4ede01STejun Heo 1125bbb68dfaSTejun Heo local_irq_save(*flags); 1126bbb68dfaSTejun Heo 112736e227d2STejun Heo /* try to steal the timer if it exists */ 112836e227d2STejun Heo if (is_dwork) { 112936e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 113036e227d2STejun Heo 1131e0aecdd8STejun Heo /* 1132e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1133e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1134e0aecdd8STejun Heo * running on the local CPU. 1135e0aecdd8STejun Heo */ 113636e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 113736e227d2STejun Heo return 1; 113836e227d2STejun Heo } 113936e227d2STejun Heo 114036e227d2STejun Heo /* try to claim PENDING the normal way */ 1141bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1142bf4ede01STejun Heo return 0; 1143bf4ede01STejun Heo 1144bf4ede01STejun Heo /* 1145bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1146bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1147bf4ede01STejun Heo */ 1148d565ed63STejun Heo pool = get_work_pool(work); 1149d565ed63STejun Heo if (!pool) 1150bbb68dfaSTejun Heo goto fail; 1151bf4ede01STejun Heo 1152d565ed63STejun Heo spin_lock(&pool->lock); 1153bf4ede01STejun Heo /* 1154112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1155112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1156112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1157112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1158112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 11590b3dae68SLai Jiangshan * item is currently queued on that pool. 1160bf4ede01STejun Heo */ 1161112202d9STejun Heo pwq = get_work_pwq(work); 1162112202d9STejun Heo if (pwq && pwq->pool == pool) { 1163bf4ede01STejun Heo debug_work_deactivate(work); 11643aa62497SLai Jiangshan 11653aa62497SLai Jiangshan /* 116616062836STejun Heo * A delayed work item cannot be grabbed directly because 116716062836STejun Heo * it might have linked NO_COLOR work items which, if left 1168112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 116916062836STejun Heo * management later on and cause stall. Make sure the work 117016062836STejun Heo * item is activated before grabbing. 11713aa62497SLai Jiangshan */ 11723aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1173112202d9STejun Heo pwq_activate_delayed_work(work); 11743aa62497SLai Jiangshan 1175bf4ede01STejun Heo list_del_init(&work->entry); 1176112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 117736e227d2STejun Heo 1178112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 11794468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 11804468a00fSLai Jiangshan 1181d565ed63STejun Heo spin_unlock(&pool->lock); 118236e227d2STejun Heo return 1; 1183bf4ede01STejun Heo } 1184d565ed63STejun Heo spin_unlock(&pool->lock); 1185bbb68dfaSTejun Heo fail: 1186bbb68dfaSTejun Heo local_irq_restore(*flags); 1187bbb68dfaSTejun Heo if (work_is_canceling(work)) 1188bbb68dfaSTejun Heo return -ENOENT; 1189bbb68dfaSTejun Heo cpu_relax(); 119036e227d2STejun Heo return -EAGAIN; 1191bf4ede01STejun Heo } 1192bf4ede01STejun Heo 1193bf4ede01STejun Heo /** 1194706026c2STejun Heo * insert_work - insert a work into a pool 1195112202d9STejun Heo * @pwq: pwq @work belongs to 11964690c4abSTejun Heo * @work: work to insert 11974690c4abSTejun Heo * @head: insertion point 11984690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 11994690c4abSTejun Heo * 1200112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1201706026c2STejun Heo * work_struct flags. 12024690c4abSTejun Heo * 12034690c4abSTejun Heo * CONTEXT: 1204d565ed63STejun Heo * spin_lock_irq(pool->lock). 1205365970a1SDavid Howells */ 1206112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1207112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1208b89deed3SOleg Nesterov { 1209112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1210e1d8aa9fSFrederic Weisbecker 12114690c4abSTejun Heo /* we own @work, set data and link */ 1212112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12131a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12148864b4e5STejun Heo get_pwq(pwq); 1215e22bee78STejun Heo 1216e22bee78STejun Heo /* 1217c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1218c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1219c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1220e22bee78STejun Heo */ 1221e22bee78STejun Heo smp_mb(); 1222e22bee78STejun Heo 122363d95a91STejun Heo if (__need_more_worker(pool)) 122463d95a91STejun Heo wake_up_worker(pool); 1225b89deed3SOleg Nesterov } 1226b89deed3SOleg Nesterov 1227c8efcc25STejun Heo /* 1228c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12298d03ecfeSTejun Heo * same workqueue. 1230c8efcc25STejun Heo */ 1231c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1232c8efcc25STejun Heo { 1233c8efcc25STejun Heo struct worker *worker; 1234c8efcc25STejun Heo 12358d03ecfeSTejun Heo worker = current_wq_worker(); 1236c8efcc25STejun Heo /* 12378d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12388d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1239c8efcc25STejun Heo */ 1240112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1241c8efcc25STejun Heo } 1242c8efcc25STejun Heo 1243d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12441da177e4SLinus Torvalds struct work_struct *work) 12451da177e4SLinus Torvalds { 1246112202d9STejun Heo struct pool_workqueue *pwq; 1247c9178087STejun Heo struct worker_pool *last_pool; 12481e19ffc6STejun Heo struct list_head *worklist; 12498a2e8e5dSTejun Heo unsigned int work_flags; 1250b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12518930cabaSTejun Heo 12528930cabaSTejun Heo /* 12538930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12548930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12558930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12568930cabaSTejun Heo * happen with IRQ disabled. 12578930cabaSTejun Heo */ 12588930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 12591da177e4SLinus Torvalds 1260dc186ad7SThomas Gleixner debug_work_activate(work); 12611e19ffc6STejun Heo 1262c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1263618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1264c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1265e41e704bSTejun Heo return; 12669e8cd2f5STejun Heo retry: 1267c9178087STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1268c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 126957469821STejun Heo if (cpu == WORK_CPU_UNBOUND) 1270f3421797STejun Heo cpu = raw_smp_processor_id(); 1271c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1272c9178087STejun Heo } else { 1273c9178087STejun Heo pwq = first_pwq(wq); 1274c9178087STejun Heo } 1275f3421797STejun Heo 127618aa9effSTejun Heo /* 1277c9178087STejun Heo * If @work was previously on a different pool, it might still be 1278c9178087STejun Heo * running there, in which case the work needs to be queued on that 1279c9178087STejun Heo * pool to guarantee non-reentrancy. 128018aa9effSTejun Heo */ 1281c9e7cf27STejun Heo last_pool = get_work_pool(work); 1282112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 128318aa9effSTejun Heo struct worker *worker; 128418aa9effSTejun Heo 1285d565ed63STejun Heo spin_lock(&last_pool->lock); 128618aa9effSTejun Heo 1287c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 128818aa9effSTejun Heo 1289112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1290c9178087STejun Heo pwq = worker->current_pwq; 12918594fadeSLai Jiangshan } else { 129218aa9effSTejun Heo /* meh... not running there, queue here */ 1293d565ed63STejun Heo spin_unlock(&last_pool->lock); 1294112202d9STejun Heo spin_lock(&pwq->pool->lock); 129518aa9effSTejun Heo } 12968930cabaSTejun Heo } else { 1297112202d9STejun Heo spin_lock(&pwq->pool->lock); 12988930cabaSTejun Heo } 1299502ca9d8STejun Heo 13009e8cd2f5STejun Heo /* 13019e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13029e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13039e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 13049e8cd2f5STejun Heo * without another pwq replacing it as the first pwq or while a 13059e8cd2f5STejun Heo * work item is executing on it, so the retying is guaranteed to 13069e8cd2f5STejun Heo * make forward-progress. 13079e8cd2f5STejun Heo */ 13089e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13099e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13109e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13119e8cd2f5STejun Heo cpu_relax(); 13129e8cd2f5STejun Heo goto retry; 13139e8cd2f5STejun Heo } 13149e8cd2f5STejun Heo /* oops */ 13159e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13169e8cd2f5STejun Heo wq->name, cpu); 13179e8cd2f5STejun Heo } 13189e8cd2f5STejun Heo 1319112202d9STejun Heo /* pwq determined, queue */ 1320112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1321502ca9d8STejun Heo 1322f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1323112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1324f5b2552bSDan Carpenter return; 1325f5b2552bSDan Carpenter } 13261e19ffc6STejun Heo 1327112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1328112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13291e19ffc6STejun Heo 1330112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1331cdadf009STejun Heo trace_workqueue_activate_work(work); 1332112202d9STejun Heo pwq->nr_active++; 1333112202d9STejun Heo worklist = &pwq->pool->worklist; 13348a2e8e5dSTejun Heo } else { 13358a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1336112202d9STejun Heo worklist = &pwq->delayed_works; 13378a2e8e5dSTejun Heo } 13381e19ffc6STejun Heo 1339112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13401e19ffc6STejun Heo 1341112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13421da177e4SLinus Torvalds } 13431da177e4SLinus Torvalds 13440fcb78c2SRolf Eike Beer /** 1345c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1346c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1347c1a220e7SZhang Rui * @wq: workqueue to use 1348c1a220e7SZhang Rui * @work: work to queue 1349c1a220e7SZhang Rui * 1350d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1351c1a220e7SZhang Rui * 1352c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1353c1a220e7SZhang Rui * can't go away. 1354c1a220e7SZhang Rui */ 1355d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1356d4283e93STejun Heo struct work_struct *work) 1357c1a220e7SZhang Rui { 1358d4283e93STejun Heo bool ret = false; 13598930cabaSTejun Heo unsigned long flags; 13608930cabaSTejun Heo 13618930cabaSTejun Heo local_irq_save(flags); 1362c1a220e7SZhang Rui 136322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 13644690c4abSTejun Heo __queue_work(cpu, wq, work); 1365d4283e93STejun Heo ret = true; 1366c1a220e7SZhang Rui } 13678930cabaSTejun Heo 13688930cabaSTejun Heo local_irq_restore(flags); 1369c1a220e7SZhang Rui return ret; 1370c1a220e7SZhang Rui } 1371c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1372c1a220e7SZhang Rui 1373d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 13741da177e4SLinus Torvalds { 137552bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 13761da177e4SLinus Torvalds 1377e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 137860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 13791da177e4SLinus Torvalds } 13801438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 13811da177e4SLinus Torvalds 13827beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 138352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 13841da177e4SLinus Torvalds { 13857beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 13867beb2edfSTejun Heo struct work_struct *work = &dwork->work; 13871da177e4SLinus Torvalds 13887beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 13897beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1390fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1391fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 13927beb2edfSTejun Heo 13938852aac2STejun Heo /* 13948852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 13958852aac2STejun Heo * both optimization and correctness. The earliest @timer can 13968852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 13978852aac2STejun Heo * on that there's no such delay when @delay is 0. 13988852aac2STejun Heo */ 13998852aac2STejun Heo if (!delay) { 14008852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14018852aac2STejun Heo return; 14028852aac2STejun Heo } 14038852aac2STejun Heo 14047beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14057beb2edfSTejun Heo 140660c057bcSLai Jiangshan dwork->wq = wq; 14071265057fSTejun Heo dwork->cpu = cpu; 14087beb2edfSTejun Heo timer->expires = jiffies + delay; 14097beb2edfSTejun Heo 14107beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14117beb2edfSTejun Heo add_timer_on(timer, cpu); 14127beb2edfSTejun Heo else 14137beb2edfSTejun Heo add_timer(timer); 14147beb2edfSTejun Heo } 14151da177e4SLinus Torvalds 14160fcb78c2SRolf Eike Beer /** 14170fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14180fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14190fcb78c2SRolf Eike Beer * @wq: workqueue to use 1420af9997e4SRandy Dunlap * @dwork: work to queue 14210fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14220fcb78c2SRolf Eike Beer * 1423715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1424715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1425715f1300STejun Heo * execution. 14260fcb78c2SRolf Eike Beer */ 1427d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 142852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14297a6bc1cdSVenkatesh Pallipadi { 143052bad64dSDavid Howells struct work_struct *work = &dwork->work; 1431d4283e93STejun Heo bool ret = false; 14328930cabaSTejun Heo unsigned long flags; 14338930cabaSTejun Heo 14348930cabaSTejun Heo /* read the comment in __queue_work() */ 14358930cabaSTejun Heo local_irq_save(flags); 14367a6bc1cdSVenkatesh Pallipadi 143722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14387beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1439d4283e93STejun Heo ret = true; 14407a6bc1cdSVenkatesh Pallipadi } 14418930cabaSTejun Heo 14428930cabaSTejun Heo local_irq_restore(flags); 14437a6bc1cdSVenkatesh Pallipadi return ret; 14447a6bc1cdSVenkatesh Pallipadi } 1445ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 14461da177e4SLinus Torvalds 1447c8e55f36STejun Heo /** 14488376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14498376fe22STejun Heo * @cpu: CPU number to execute work on 14508376fe22STejun Heo * @wq: workqueue to use 14518376fe22STejun Heo * @dwork: work to queue 14528376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14538376fe22STejun Heo * 14548376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14558376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14568376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 14578376fe22STejun Heo * current state. 14588376fe22STejun Heo * 14598376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 14608376fe22STejun Heo * pending and its timer was modified. 14618376fe22STejun Heo * 1462e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 14638376fe22STejun Heo * See try_to_grab_pending() for details. 14648376fe22STejun Heo */ 14658376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 14668376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 14678376fe22STejun Heo { 14688376fe22STejun Heo unsigned long flags; 14698376fe22STejun Heo int ret; 14708376fe22STejun Heo 14718376fe22STejun Heo do { 14728376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 14738376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 14748376fe22STejun Heo 14758376fe22STejun Heo if (likely(ret >= 0)) { 14768376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 14778376fe22STejun Heo local_irq_restore(flags); 14788376fe22STejun Heo } 14798376fe22STejun Heo 14808376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 14818376fe22STejun Heo return ret; 14828376fe22STejun Heo } 14838376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 14848376fe22STejun Heo 14858376fe22STejun Heo /** 1486c8e55f36STejun Heo * worker_enter_idle - enter idle state 1487c8e55f36STejun Heo * @worker: worker which is entering idle state 1488c8e55f36STejun Heo * 1489c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1490c8e55f36STejun Heo * necessary. 1491c8e55f36STejun Heo * 1492c8e55f36STejun Heo * LOCKING: 1493d565ed63STejun Heo * spin_lock_irq(pool->lock). 1494c8e55f36STejun Heo */ 1495c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 14961da177e4SLinus Torvalds { 1497bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1498c8e55f36STejun Heo 14996183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15006183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15016183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15026183c009STejun Heo return; 1503c8e55f36STejun Heo 1504cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1505cb444766STejun Heo worker->flags |= WORKER_IDLE; 1506bd7bdd43STejun Heo pool->nr_idle++; 1507e22bee78STejun Heo worker->last_active = jiffies; 1508c8e55f36STejun Heo 1509c8e55f36STejun Heo /* idle_list is LIFO */ 1510bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1511db7bccf4STejun Heo 151263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1513628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1514cb444766STejun Heo 1515544ecf31STejun Heo /* 1516706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1517d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1518628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1519628c78e7STejun Heo * unbind is not in progress. 1520544ecf31STejun Heo */ 152124647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1522bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1523e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1524c8e55f36STejun Heo } 1525c8e55f36STejun Heo 1526c8e55f36STejun Heo /** 1527c8e55f36STejun Heo * worker_leave_idle - leave idle state 1528c8e55f36STejun Heo * @worker: worker which is leaving idle state 1529c8e55f36STejun Heo * 1530c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1531c8e55f36STejun Heo * 1532c8e55f36STejun Heo * LOCKING: 1533d565ed63STejun Heo * spin_lock_irq(pool->lock). 1534c8e55f36STejun Heo */ 1535c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1536c8e55f36STejun Heo { 1537bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1538c8e55f36STejun Heo 15396183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15406183c009STejun Heo return; 1541d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1542bd7bdd43STejun Heo pool->nr_idle--; 1543c8e55f36STejun Heo list_del_init(&worker->entry); 1544c8e55f36STejun Heo } 1545c8e55f36STejun Heo 1546e22bee78STejun Heo /** 1547f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1548f36dc67bSLai Jiangshan * @pool: target worker_pool 1549f36dc67bSLai Jiangshan * 1550f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1551e22bee78STejun Heo * 1552e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1553e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1554e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1555e22bee78STejun Heo * guaranteed to execute on the cpu. 1556e22bee78STejun Heo * 1557f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1558e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1559e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1560e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1561706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1562e22bee78STejun Heo * [dis]associated in the meantime. 1563e22bee78STejun Heo * 1564706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 156524647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1566f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1567f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1568f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1569e22bee78STejun Heo * 1570e22bee78STejun Heo * CONTEXT: 1571d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1572e22bee78STejun Heo * held. 1573e22bee78STejun Heo * 1574e22bee78STejun Heo * RETURNS: 1575706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1576e22bee78STejun Heo * bound), %false if offline. 1577e22bee78STejun Heo */ 1578f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1579d565ed63STejun Heo __acquires(&pool->lock) 1580e22bee78STejun Heo { 1581e22bee78STejun Heo while (true) { 1582e22bee78STejun Heo /* 1583e22bee78STejun Heo * The following call may fail, succeed or succeed 1584e22bee78STejun Heo * without actually migrating the task to the cpu if 1585e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 158624647570STejun Heo * against POOL_DISASSOCIATED. 1587e22bee78STejun Heo */ 158824647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 15897a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1590e22bee78STejun Heo 1591d565ed63STejun Heo spin_lock_irq(&pool->lock); 159224647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1593e22bee78STejun Heo return false; 1594f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 15957a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1596e22bee78STejun Heo return true; 1597d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1598e22bee78STejun Heo 15995035b20fSTejun Heo /* 16005035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16015035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16025035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16035035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16045035b20fSTejun Heo */ 1605e22bee78STejun Heo cpu_relax(); 16065035b20fSTejun Heo cond_resched(); 1607e22bee78STejun Heo } 1608e22bee78STejun Heo } 1609e22bee78STejun Heo 1610c34056a3STejun Heo static struct worker *alloc_worker(void) 1611c34056a3STejun Heo { 1612c34056a3STejun Heo struct worker *worker; 1613c34056a3STejun Heo 1614c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1615c8e55f36STejun Heo if (worker) { 1616c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1617affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1618e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1619e22bee78STejun Heo worker->flags = WORKER_PREP; 1620c8e55f36STejun Heo } 1621c34056a3STejun Heo return worker; 1622c34056a3STejun Heo } 1623c34056a3STejun Heo 1624c34056a3STejun Heo /** 1625c34056a3STejun Heo * create_worker - create a new workqueue worker 162663d95a91STejun Heo * @pool: pool the new worker will belong to 1627c34056a3STejun Heo * 162863d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1629c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1630c34056a3STejun Heo * destroy_worker(). 1631c34056a3STejun Heo * 1632c34056a3STejun Heo * CONTEXT: 1633c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1634c34056a3STejun Heo * 1635c34056a3STejun Heo * RETURNS: 1636c34056a3STejun Heo * Pointer to the newly created worker. 1637c34056a3STejun Heo */ 1638bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1639c34056a3STejun Heo { 16407a4e344cSTejun Heo const char *pri = pool->attrs->nice < 0 ? "H" : ""; 1641c34056a3STejun Heo struct worker *worker = NULL; 1642f3421797STejun Heo int id = -1; 1643c34056a3STejun Heo 1644cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1645cd549687STejun Heo 1646822d8405STejun Heo /* 1647822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1648822d8405STejun Heo * without installing the pointer. 1649822d8405STejun Heo */ 1650822d8405STejun Heo idr_preload(GFP_KERNEL); 1651d565ed63STejun Heo spin_lock_irq(&pool->lock); 1652822d8405STejun Heo 1653822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1654822d8405STejun Heo 1655d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1656822d8405STejun Heo idr_preload_end(); 1657822d8405STejun Heo if (id < 0) 1658c34056a3STejun Heo goto fail; 1659c34056a3STejun Heo 1660c34056a3STejun Heo worker = alloc_worker(); 1661c34056a3STejun Heo if (!worker) 1662c34056a3STejun Heo goto fail; 1663c34056a3STejun Heo 1664bd7bdd43STejun Heo worker->pool = pool; 1665c34056a3STejun Heo worker->id = id; 1666c34056a3STejun Heo 166729c91e99STejun Heo if (pool->cpu >= 0) 166894dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 1669ec22ca5eSTejun Heo worker, cpu_to_node(pool->cpu), 1670d84ff051STejun Heo "kworker/%d:%d%s", pool->cpu, id, pri); 1671f3421797STejun Heo else 1672f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 1673ac6104cdSTejun Heo "kworker/u%d:%d%s", 1674ac6104cdSTejun Heo pool->id, id, pri); 1675c34056a3STejun Heo if (IS_ERR(worker->task)) 1676c34056a3STejun Heo goto fail; 1677c34056a3STejun Heo 1678c5aa87bbSTejun Heo /* 1679c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1680c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1681c5aa87bbSTejun Heo */ 16827a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 16837a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 16843270476aSTejun Heo 168514a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 168614a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 16877a4e344cSTejun Heo 16887a4e344cSTejun Heo /* 16897a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 16907a4e344cSTejun Heo * remains stable across this function. See the comments above the 16917a4e344cSTejun Heo * flag definition for details. 16927a4e344cSTejun Heo */ 16937a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1694f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1695c34056a3STejun Heo 1696822d8405STejun Heo /* successful, commit the pointer to idr */ 1697822d8405STejun Heo spin_lock_irq(&pool->lock); 1698822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1699822d8405STejun Heo spin_unlock_irq(&pool->lock); 1700822d8405STejun Heo 1701c34056a3STejun Heo return worker; 1702822d8405STejun Heo 1703c34056a3STejun Heo fail: 1704c34056a3STejun Heo if (id >= 0) { 1705d565ed63STejun Heo spin_lock_irq(&pool->lock); 1706822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1707d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1708c34056a3STejun Heo } 1709c34056a3STejun Heo kfree(worker); 1710c34056a3STejun Heo return NULL; 1711c34056a3STejun Heo } 1712c34056a3STejun Heo 1713c34056a3STejun Heo /** 1714c34056a3STejun Heo * start_worker - start a newly created worker 1715c34056a3STejun Heo * @worker: worker to start 1716c34056a3STejun Heo * 1717706026c2STejun Heo * Make the pool aware of @worker and start it. 1718c34056a3STejun Heo * 1719c34056a3STejun Heo * CONTEXT: 1720d565ed63STejun Heo * spin_lock_irq(pool->lock). 1721c34056a3STejun Heo */ 1722c34056a3STejun Heo static void start_worker(struct worker *worker) 1723c34056a3STejun Heo { 1724cb444766STejun Heo worker->flags |= WORKER_STARTED; 1725bd7bdd43STejun Heo worker->pool->nr_workers++; 1726c8e55f36STejun Heo worker_enter_idle(worker); 1727c34056a3STejun Heo wake_up_process(worker->task); 1728c34056a3STejun Heo } 1729c34056a3STejun Heo 1730c34056a3STejun Heo /** 1731ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1732ebf44d16STejun Heo * @pool: the target pool 1733ebf44d16STejun Heo * 1734cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1735ebf44d16STejun Heo */ 1736ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1737ebf44d16STejun Heo { 1738ebf44d16STejun Heo struct worker *worker; 1739ebf44d16STejun Heo 1740cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1741cd549687STejun Heo 1742ebf44d16STejun Heo worker = create_worker(pool); 1743ebf44d16STejun Heo if (worker) { 1744ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1745ebf44d16STejun Heo start_worker(worker); 1746ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1747ebf44d16STejun Heo } 1748ebf44d16STejun Heo 1749cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1750cd549687STejun Heo 1751ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1752ebf44d16STejun Heo } 1753ebf44d16STejun Heo 1754ebf44d16STejun Heo /** 1755c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1756c34056a3STejun Heo * @worker: worker to be destroyed 1757c34056a3STejun Heo * 1758706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1759c8e55f36STejun Heo * 1760c8e55f36STejun Heo * CONTEXT: 1761d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1762c34056a3STejun Heo */ 1763c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1764c34056a3STejun Heo { 1765bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1766c34056a3STejun Heo 1767cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1768cd549687STejun Heo lockdep_assert_held(&pool->lock); 1769cd549687STejun Heo 1770c34056a3STejun Heo /* sanity check frenzy */ 17716183c009STejun Heo if (WARN_ON(worker->current_work) || 17726183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 17736183c009STejun Heo return; 1774c34056a3STejun Heo 1775c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1776bd7bdd43STejun Heo pool->nr_workers--; 1777c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1778bd7bdd43STejun Heo pool->nr_idle--; 1779c8e55f36STejun Heo 1780c8e55f36STejun Heo list_del_init(&worker->entry); 1781cb444766STejun Heo worker->flags |= WORKER_DIE; 1782c8e55f36STejun Heo 1783822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1784822d8405STejun Heo 1785d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1786c8e55f36STejun Heo 1787c34056a3STejun Heo kthread_stop(worker->task); 1788c34056a3STejun Heo kfree(worker); 1789c34056a3STejun Heo 1790d565ed63STejun Heo spin_lock_irq(&pool->lock); 1791c34056a3STejun Heo } 1792c34056a3STejun Heo 179363d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1794e22bee78STejun Heo { 179563d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1796e22bee78STejun Heo 1797d565ed63STejun Heo spin_lock_irq(&pool->lock); 1798e22bee78STejun Heo 179963d95a91STejun Heo if (too_many_workers(pool)) { 1800e22bee78STejun Heo struct worker *worker; 1801e22bee78STejun Heo unsigned long expires; 1802e22bee78STejun Heo 1803e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 180463d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1805e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1806e22bee78STejun Heo 1807e22bee78STejun Heo if (time_before(jiffies, expires)) 180863d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1809e22bee78STejun Heo else { 1810e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 181111ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 181263d95a91STejun Heo wake_up_worker(pool); 1813e22bee78STejun Heo } 1814e22bee78STejun Heo } 1815e22bee78STejun Heo 1816d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1817e22bee78STejun Heo } 1818e22bee78STejun Heo 1819493a1724STejun Heo static void send_mayday(struct work_struct *work) 1820e22bee78STejun Heo { 1821112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1822112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1823493a1724STejun Heo 18242e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1825e22bee78STejun Heo 1826493008a8STejun Heo if (!wq->rescuer) 1827493a1724STejun Heo return; 1828e22bee78STejun Heo 1829e22bee78STejun Heo /* mayday mayday mayday */ 1830493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1831493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1832e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1833493a1724STejun Heo } 1834e22bee78STejun Heo } 1835e22bee78STejun Heo 1836706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1837e22bee78STejun Heo { 183863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1839e22bee78STejun Heo struct work_struct *work; 1840e22bee78STejun Heo 18412e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1842493a1724STejun Heo spin_lock(&pool->lock); 1843e22bee78STejun Heo 184463d95a91STejun Heo if (need_to_create_worker(pool)) { 1845e22bee78STejun Heo /* 1846e22bee78STejun Heo * We've been trying to create a new worker but 1847e22bee78STejun Heo * haven't been successful. We might be hitting an 1848e22bee78STejun Heo * allocation deadlock. Send distress signals to 1849e22bee78STejun Heo * rescuers. 1850e22bee78STejun Heo */ 185163d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1852e22bee78STejun Heo send_mayday(work); 1853e22bee78STejun Heo } 1854e22bee78STejun Heo 1855493a1724STejun Heo spin_unlock(&pool->lock); 18562e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1857e22bee78STejun Heo 185863d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1859e22bee78STejun Heo } 1860e22bee78STejun Heo 1861e22bee78STejun Heo /** 1862e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 186363d95a91STejun Heo * @pool: pool to create a new worker for 1864e22bee78STejun Heo * 186563d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1866e22bee78STejun Heo * have at least one idle worker on return from this function. If 1867e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 186863d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1869e22bee78STejun Heo * possible allocation deadlock. 1870e22bee78STejun Heo * 1871c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1872c5aa87bbSTejun Heo * may_start_working() %true. 1873e22bee78STejun Heo * 1874e22bee78STejun Heo * LOCKING: 1875d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1876e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1877e22bee78STejun Heo * manager. 1878e22bee78STejun Heo * 1879e22bee78STejun Heo * RETURNS: 1880c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1881e22bee78STejun Heo * otherwise. 1882e22bee78STejun Heo */ 188363d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1884d565ed63STejun Heo __releases(&pool->lock) 1885d565ed63STejun Heo __acquires(&pool->lock) 1886e22bee78STejun Heo { 188763d95a91STejun Heo if (!need_to_create_worker(pool)) 1888e22bee78STejun Heo return false; 1889e22bee78STejun Heo restart: 1890d565ed63STejun Heo spin_unlock_irq(&pool->lock); 18919f9c2364STejun Heo 1892e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 189363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1894e22bee78STejun Heo 1895e22bee78STejun Heo while (true) { 1896e22bee78STejun Heo struct worker *worker; 1897e22bee78STejun Heo 1898bc2ae0f5STejun Heo worker = create_worker(pool); 1899e22bee78STejun Heo if (worker) { 190063d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1901d565ed63STejun Heo spin_lock_irq(&pool->lock); 1902e22bee78STejun Heo start_worker(worker); 19036183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19046183c009STejun Heo goto restart; 1905e22bee78STejun Heo return true; 1906e22bee78STejun Heo } 1907e22bee78STejun Heo 190863d95a91STejun Heo if (!need_to_create_worker(pool)) 1909e22bee78STejun Heo break; 1910e22bee78STejun Heo 1911e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1912e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19139f9c2364STejun Heo 191463d95a91STejun Heo if (!need_to_create_worker(pool)) 1915e22bee78STejun Heo break; 1916e22bee78STejun Heo } 1917e22bee78STejun Heo 191863d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1919d565ed63STejun Heo spin_lock_irq(&pool->lock); 192063d95a91STejun Heo if (need_to_create_worker(pool)) 1921e22bee78STejun Heo goto restart; 1922e22bee78STejun Heo return true; 1923e22bee78STejun Heo } 1924e22bee78STejun Heo 1925e22bee78STejun Heo /** 1926e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 192763d95a91STejun Heo * @pool: pool to destroy workers for 1928e22bee78STejun Heo * 192963d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1930e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1931e22bee78STejun Heo * 1932e22bee78STejun Heo * LOCKING: 1933d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1934e22bee78STejun Heo * multiple times. Called only from manager. 1935e22bee78STejun Heo * 1936e22bee78STejun Heo * RETURNS: 1937c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1938e22bee78STejun Heo * otherwise. 1939e22bee78STejun Heo */ 194063d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1941e22bee78STejun Heo { 1942e22bee78STejun Heo bool ret = false; 1943e22bee78STejun Heo 194463d95a91STejun Heo while (too_many_workers(pool)) { 1945e22bee78STejun Heo struct worker *worker; 1946e22bee78STejun Heo unsigned long expires; 1947e22bee78STejun Heo 194863d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1949e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1950e22bee78STejun Heo 1951e22bee78STejun Heo if (time_before(jiffies, expires)) { 195263d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1953e22bee78STejun Heo break; 1954e22bee78STejun Heo } 1955e22bee78STejun Heo 1956e22bee78STejun Heo destroy_worker(worker); 1957e22bee78STejun Heo ret = true; 1958e22bee78STejun Heo } 1959e22bee78STejun Heo 1960e22bee78STejun Heo return ret; 1961e22bee78STejun Heo } 1962e22bee78STejun Heo 1963e22bee78STejun Heo /** 1964e22bee78STejun Heo * manage_workers - manage worker pool 1965e22bee78STejun Heo * @worker: self 1966e22bee78STejun Heo * 1967706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1968e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1969706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1970e22bee78STejun Heo * 1971e22bee78STejun Heo * The caller can safely start processing works on false return. On 1972e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1973e22bee78STejun Heo * and may_start_working() is true. 1974e22bee78STejun Heo * 1975e22bee78STejun Heo * CONTEXT: 1976d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1977e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1978e22bee78STejun Heo * 1979e22bee78STejun Heo * RETURNS: 1980d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1981d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 1982e22bee78STejun Heo */ 1983e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1984e22bee78STejun Heo { 198563d95a91STejun Heo struct worker_pool *pool = worker->pool; 1986e22bee78STejun Heo bool ret = false; 1987e22bee78STejun Heo 1988bc3a1afcSTejun Heo /* 1989bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 1990bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 1991bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 1992bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 1993bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 1994bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 1995bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 1996bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 1997bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 1998bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 1999bc3a1afcSTejun Heo * 2000bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2001bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2002bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2003bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2004bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2005bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2006bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2007bc3a1afcSTejun Heo * manager_mutex. 2008bc3a1afcSTejun Heo */ 200934a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2010e22bee78STejun Heo return ret; 2011e22bee78STejun Heo 2012ee378aa4SLai Jiangshan /* 2013bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2014bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2015ee378aa4SLai Jiangshan */ 2016bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2017d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2018bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2019ee378aa4SLai Jiangshan ret = true; 2020ee378aa4SLai Jiangshan } 2021ee378aa4SLai Jiangshan 202211ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2023e22bee78STejun Heo 2024e22bee78STejun Heo /* 2025e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2026e22bee78STejun Heo * on return. 2027e22bee78STejun Heo */ 202863d95a91STejun Heo ret |= maybe_destroy_workers(pool); 202963d95a91STejun Heo ret |= maybe_create_worker(pool); 2030e22bee78STejun Heo 2031bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 203234a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2033e22bee78STejun Heo return ret; 2034e22bee78STejun Heo } 2035e22bee78STejun Heo 2036a62428c0STejun Heo /** 2037a62428c0STejun Heo * process_one_work - process single work 2038c34056a3STejun Heo * @worker: self 2039a62428c0STejun Heo * @work: work to process 2040a62428c0STejun Heo * 2041a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2042a62428c0STejun Heo * process a single work including synchronization against and 2043a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2044a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2045a62428c0STejun Heo * call this function to process a work. 2046a62428c0STejun Heo * 2047a62428c0STejun Heo * CONTEXT: 2048d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2049a62428c0STejun Heo */ 2050c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2051d565ed63STejun Heo __releases(&pool->lock) 2052d565ed63STejun Heo __acquires(&pool->lock) 20531da177e4SLinus Torvalds { 2054112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2055bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2056112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 205773f53c4aSTejun Heo int work_color; 20587e11629dSTejun Heo struct worker *collision; 20594e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20604e6045f1SJohannes Berg /* 2061a62428c0STejun Heo * It is permissible to free the struct work_struct from 2062a62428c0STejun Heo * inside the function that is called from it, this we need to 2063a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2064a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2065a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20664e6045f1SJohannes Berg */ 20674d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20684d82a1deSPeter Zijlstra 20694d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20704e6045f1SJohannes Berg #endif 20716fec10a1STejun Heo /* 20726fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 20736fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 207424647570STejun Heo * unbound or a disassociated pool. 20756fec10a1STejun Heo */ 20765f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 207724647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2078ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 207925511a47STejun Heo 20807e11629dSTejun Heo /* 20817e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20827e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20837e11629dSTejun Heo * already processing the work. If so, defer the work to the 20847e11629dSTejun Heo * currently executing one. 20857e11629dSTejun Heo */ 2086c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20877e11629dSTejun Heo if (unlikely(collision)) { 20887e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20897e11629dSTejun Heo return; 20907e11629dSTejun Heo } 20911da177e4SLinus Torvalds 20928930cabaSTejun Heo /* claim and dequeue */ 20931da177e4SLinus Torvalds debug_work_deactivate(work); 2094c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2095c34056a3STejun Heo worker->current_work = work; 2096a2c1c57bSTejun Heo worker->current_func = work->func; 2097112202d9STejun Heo worker->current_pwq = pwq; 209873f53c4aSTejun Heo work_color = get_work_color(work); 20997a22ad75STejun Heo 2100a62428c0STejun Heo list_del_init(&work->entry); 2101a62428c0STejun Heo 2102649027d7STejun Heo /* 2103fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2104fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2105fb0e7bebSTejun Heo */ 2106fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2107fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2108fb0e7bebSTejun Heo 2109974271c4STejun Heo /* 2110d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2111974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2112974271c4STejun Heo */ 211363d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 211463d95a91STejun Heo wake_up_worker(pool); 2115974271c4STejun Heo 21168930cabaSTejun Heo /* 21177c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2118d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 211923657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 212023657bb1STejun Heo * disabled. 21218930cabaSTejun Heo */ 21227c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21231da177e4SLinus Torvalds 2124d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2125365970a1SDavid Howells 2126112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21273295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2128e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2129a2c1c57bSTejun Heo worker->current_func(work); 2130e36c886aSArjan van de Ven /* 2131e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2132e36c886aSArjan van de Ven * point will only record its address. 2133e36c886aSArjan van de Ven */ 2134e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21353295f0efSIngo Molnar lock_map_release(&lockdep_map); 2136112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21371da177e4SLinus Torvalds 2138d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2139044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2140044c782cSValentin Ilie " last function: %pf\n", 2141a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2142a2c1c57bSTejun Heo worker->current_func); 2143d5abe669SPeter Zijlstra debug_show_held_locks(current); 2144d5abe669SPeter Zijlstra dump_stack(); 2145d5abe669SPeter Zijlstra } 2146d5abe669SPeter Zijlstra 2147d565ed63STejun Heo spin_lock_irq(&pool->lock); 2148a62428c0STejun Heo 2149fb0e7bebSTejun Heo /* clear cpu intensive status */ 2150fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2151fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2152fb0e7bebSTejun Heo 2153a62428c0STejun Heo /* we're done with it, release */ 215442f8570fSSasha Levin hash_del(&worker->hentry); 2155c34056a3STejun Heo worker->current_work = NULL; 2156a2c1c57bSTejun Heo worker->current_func = NULL; 2157112202d9STejun Heo worker->current_pwq = NULL; 2158112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21591da177e4SLinus Torvalds } 21601da177e4SLinus Torvalds 2161affee4b2STejun Heo /** 2162affee4b2STejun Heo * process_scheduled_works - process scheduled works 2163affee4b2STejun Heo * @worker: self 2164affee4b2STejun Heo * 2165affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2166affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2167affee4b2STejun Heo * fetches a work from the top and executes it. 2168affee4b2STejun Heo * 2169affee4b2STejun Heo * CONTEXT: 2170d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2171affee4b2STejun Heo * multiple times. 2172affee4b2STejun Heo */ 2173affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 21741da177e4SLinus Torvalds { 2175affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2176affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2177a62428c0STejun Heo struct work_struct, entry); 2178c34056a3STejun Heo process_one_work(worker, work); 2179a62428c0STejun Heo } 21801da177e4SLinus Torvalds } 21811da177e4SLinus Torvalds 21824690c4abSTejun Heo /** 21834690c4abSTejun Heo * worker_thread - the worker thread function 2184c34056a3STejun Heo * @__worker: self 21854690c4abSTejun Heo * 2186c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2187c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2188c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2189c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2190c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 21914690c4abSTejun Heo */ 2192c34056a3STejun Heo static int worker_thread(void *__worker) 21931da177e4SLinus Torvalds { 2194c34056a3STejun Heo struct worker *worker = __worker; 2195bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 21961da177e4SLinus Torvalds 2197e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2198e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2199c8e55f36STejun Heo woke_up: 2200d565ed63STejun Heo spin_lock_irq(&pool->lock); 2201affee4b2STejun Heo 2202a9ab775bSTejun Heo /* am I supposed to die? */ 2203a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2204d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2205a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2206e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2207c8e55f36STejun Heo return 0; 2208c8e55f36STejun Heo } 2209c8e55f36STejun Heo 2210c8e55f36STejun Heo worker_leave_idle(worker); 2211db7bccf4STejun Heo recheck: 2212e22bee78STejun Heo /* no more worker necessary? */ 221363d95a91STejun Heo if (!need_more_worker(pool)) 2214e22bee78STejun Heo goto sleep; 2215e22bee78STejun Heo 2216e22bee78STejun Heo /* do we need to manage? */ 221763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2218e22bee78STejun Heo goto recheck; 2219e22bee78STejun Heo 2220c8e55f36STejun Heo /* 2221c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2222c8e55f36STejun Heo * preparing to process a work or actually processing it. 2223c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2224c8e55f36STejun Heo */ 22256183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2226c8e55f36STejun Heo 2227e22bee78STejun Heo /* 2228a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2229a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2230a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2231a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2232a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2233e22bee78STejun Heo */ 2234a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2235e22bee78STejun Heo 2236e22bee78STejun Heo do { 2237affee4b2STejun Heo struct work_struct *work = 2238bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2239affee4b2STejun Heo struct work_struct, entry); 2240affee4b2STejun Heo 2241c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2242affee4b2STejun Heo /* optimization path, not strictly necessary */ 2243affee4b2STejun Heo process_one_work(worker, work); 2244affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2245affee4b2STejun Heo process_scheduled_works(worker); 2246affee4b2STejun Heo } else { 2247c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2248affee4b2STejun Heo process_scheduled_works(worker); 2249affee4b2STejun Heo } 225063d95a91STejun Heo } while (keep_working(pool)); 2251affee4b2STejun Heo 2252e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2253d313dd85STejun Heo sleep: 225463d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2255e22bee78STejun Heo goto recheck; 2256d313dd85STejun Heo 2257c8e55f36STejun Heo /* 2258d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2259d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2260d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2261d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2262d565ed63STejun Heo * event. 2263c8e55f36STejun Heo */ 2264c8e55f36STejun Heo worker_enter_idle(worker); 2265c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2266d565ed63STejun Heo spin_unlock_irq(&pool->lock); 22671da177e4SLinus Torvalds schedule(); 2268c8e55f36STejun Heo goto woke_up; 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds 2271e22bee78STejun Heo /** 2272e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2273111c225aSTejun Heo * @__rescuer: self 2274e22bee78STejun Heo * 2275e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2276493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2277e22bee78STejun Heo * 2278706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2279e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2280e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2281e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2282e22bee78STejun Heo * the problem rescuer solves. 2283e22bee78STejun Heo * 2284706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2285706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2286e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2287e22bee78STejun Heo * 2288e22bee78STejun Heo * This should happen rarely. 2289e22bee78STejun Heo */ 2290111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2291e22bee78STejun Heo { 2292111c225aSTejun Heo struct worker *rescuer = __rescuer; 2293111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2294e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2295e22bee78STejun Heo 2296e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2297111c225aSTejun Heo 2298111c225aSTejun Heo /* 2299111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2300111c225aSTejun Heo * doesn't participate in concurrency management. 2301111c225aSTejun Heo */ 2302111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2303e22bee78STejun Heo repeat: 2304e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23051da177e4SLinus Torvalds 2306412d32e6SMike Galbraith if (kthread_should_stop()) { 2307412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2308111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2309e22bee78STejun Heo return 0; 2310412d32e6SMike Galbraith } 23111da177e4SLinus Torvalds 2312493a1724STejun Heo /* see whether any pwq is asking for help */ 23132e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2314493a1724STejun Heo 2315493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2316493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2317493a1724STejun Heo struct pool_workqueue, mayday_node); 2318112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2319e22bee78STejun Heo struct work_struct *work, *n; 2320e22bee78STejun Heo 2321e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2322493a1724STejun Heo list_del_init(&pwq->mayday_node); 2323493a1724STejun Heo 23242e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2325e22bee78STejun Heo 2326e22bee78STejun Heo /* migrate to the target cpu if possible */ 2327f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2328b3104104SLai Jiangshan rescuer->pool = pool; 2329e22bee78STejun Heo 2330e22bee78STejun Heo /* 2331e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2332e22bee78STejun Heo * process'em. 2333e22bee78STejun Heo */ 23346183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2335bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2336112202d9STejun Heo if (get_work_pwq(work) == pwq) 2337e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2338e22bee78STejun Heo 2339e22bee78STejun Heo process_scheduled_works(rescuer); 23407576958aSTejun Heo 23417576958aSTejun Heo /* 2342d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23437576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23447576958aSTejun Heo * and stalling the execution. 23457576958aSTejun Heo */ 234663d95a91STejun Heo if (keep_working(pool)) 234763d95a91STejun Heo wake_up_worker(pool); 23487576958aSTejun Heo 2349b3104104SLai Jiangshan rescuer->pool = NULL; 2350493a1724STejun Heo spin_unlock(&pool->lock); 23512e109a28STejun Heo spin_lock(&wq_mayday_lock); 23521da177e4SLinus Torvalds } 23531da177e4SLinus Torvalds 23542e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2355493a1724STejun Heo 2356111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2357111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2358e22bee78STejun Heo schedule(); 2359e22bee78STejun Heo goto repeat; 23601da177e4SLinus Torvalds } 23611da177e4SLinus Torvalds 2362fc2e4d70SOleg Nesterov struct wq_barrier { 2363fc2e4d70SOleg Nesterov struct work_struct work; 2364fc2e4d70SOleg Nesterov struct completion done; 2365fc2e4d70SOleg Nesterov }; 2366fc2e4d70SOleg Nesterov 2367fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2368fc2e4d70SOleg Nesterov { 2369fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2370fc2e4d70SOleg Nesterov complete(&barr->done); 2371fc2e4d70SOleg Nesterov } 2372fc2e4d70SOleg Nesterov 23734690c4abSTejun Heo /** 23744690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2375112202d9STejun Heo * @pwq: pwq to insert barrier into 23764690c4abSTejun Heo * @barr: wq_barrier to insert 2377affee4b2STejun Heo * @target: target work to attach @barr to 2378affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 23794690c4abSTejun Heo * 2380affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2381affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2382affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2383affee4b2STejun Heo * cpu. 2384affee4b2STejun Heo * 2385affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2386affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2387affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2388affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2389affee4b2STejun Heo * after a work with LINKED flag set. 2390affee4b2STejun Heo * 2391affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2392112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 23934690c4abSTejun Heo * 23944690c4abSTejun Heo * CONTEXT: 2395d565ed63STejun Heo * spin_lock_irq(pool->lock). 23964690c4abSTejun Heo */ 2397112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2398affee4b2STejun Heo struct wq_barrier *barr, 2399affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2400fc2e4d70SOleg Nesterov { 2401affee4b2STejun Heo struct list_head *head; 2402affee4b2STejun Heo unsigned int linked = 0; 2403affee4b2STejun Heo 2404dc186ad7SThomas Gleixner /* 2405d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2406dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2407dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2408dc186ad7SThomas Gleixner * might deadlock. 2409dc186ad7SThomas Gleixner */ 2410ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 241122df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2412fc2e4d70SOleg Nesterov init_completion(&barr->done); 241383c22520SOleg Nesterov 2414affee4b2STejun Heo /* 2415affee4b2STejun Heo * If @target is currently being executed, schedule the 2416affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2417affee4b2STejun Heo */ 2418affee4b2STejun Heo if (worker) 2419affee4b2STejun Heo head = worker->scheduled.next; 2420affee4b2STejun Heo else { 2421affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2422affee4b2STejun Heo 2423affee4b2STejun Heo head = target->entry.next; 2424affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2425affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2426affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2427affee4b2STejun Heo } 2428affee4b2STejun Heo 2429dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2430112202d9STejun Heo insert_work(pwq, &barr->work, head, 2431affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2432fc2e4d70SOleg Nesterov } 2433fc2e4d70SOleg Nesterov 243473f53c4aSTejun Heo /** 2435112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 243673f53c4aSTejun Heo * @wq: workqueue being flushed 243773f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 243873f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 243973f53c4aSTejun Heo * 2440112202d9STejun Heo * Prepare pwqs for workqueue flushing. 244173f53c4aSTejun Heo * 2442112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2443112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2444112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2445112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2446112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 244773f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 244873f53c4aSTejun Heo * 244973f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 245073f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 245173f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 245273f53c4aSTejun Heo * is returned. 245373f53c4aSTejun Heo * 2454112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 245573f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 245673f53c4aSTejun Heo * advanced to @work_color. 245773f53c4aSTejun Heo * 245873f53c4aSTejun Heo * CONTEXT: 24593c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 246073f53c4aSTejun Heo * 246173f53c4aSTejun Heo * RETURNS: 246273f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 246373f53c4aSTejun Heo * otherwise. 246473f53c4aSTejun Heo */ 2465112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 246673f53c4aSTejun Heo int flush_color, int work_color) 24671da177e4SLinus Torvalds { 246873f53c4aSTejun Heo bool wait = false; 246949e3cf44STejun Heo struct pool_workqueue *pwq; 24701da177e4SLinus Torvalds 247173f53c4aSTejun Heo if (flush_color >= 0) { 24726183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2473112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2474dc186ad7SThomas Gleixner } 247514441960SOleg Nesterov 247649e3cf44STejun Heo for_each_pwq(pwq, wq) { 2477112202d9STejun Heo struct worker_pool *pool = pwq->pool; 24781da177e4SLinus Torvalds 2479b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 248073f53c4aSTejun Heo 248173f53c4aSTejun Heo if (flush_color >= 0) { 24826183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 248373f53c4aSTejun Heo 2484112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2485112202d9STejun Heo pwq->flush_color = flush_color; 2486112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 248773f53c4aSTejun Heo wait = true; 24881da177e4SLinus Torvalds } 248973f53c4aSTejun Heo } 249073f53c4aSTejun Heo 249173f53c4aSTejun Heo if (work_color >= 0) { 24926183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2493112202d9STejun Heo pwq->work_color = work_color; 249473f53c4aSTejun Heo } 249573f53c4aSTejun Heo 2496b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 24971da177e4SLinus Torvalds } 24981da177e4SLinus Torvalds 2499112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 250073f53c4aSTejun Heo complete(&wq->first_flusher->done); 250173f53c4aSTejun Heo 250273f53c4aSTejun Heo return wait; 250383c22520SOleg Nesterov } 25041da177e4SLinus Torvalds 25050fcb78c2SRolf Eike Beer /** 25061da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25070fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25081da177e4SLinus Torvalds * 2509c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2510c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25111da177e4SLinus Torvalds */ 25127ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25131da177e4SLinus Torvalds { 251473f53c4aSTejun Heo struct wq_flusher this_flusher = { 251573f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 251673f53c4aSTejun Heo .flush_color = -1, 251773f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 251873f53c4aSTejun Heo }; 251973f53c4aSTejun Heo int next_color; 2520b1f4ec17SOleg Nesterov 25213295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25223295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 252373f53c4aSTejun Heo 25243c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 252573f53c4aSTejun Heo 252673f53c4aSTejun Heo /* 252773f53c4aSTejun Heo * Start-to-wait phase 252873f53c4aSTejun Heo */ 252973f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 253073f53c4aSTejun Heo 253173f53c4aSTejun Heo if (next_color != wq->flush_color) { 253273f53c4aSTejun Heo /* 253373f53c4aSTejun Heo * Color space is not full. The current work_color 253473f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 253573f53c4aSTejun Heo * by one. 253673f53c4aSTejun Heo */ 25376183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 253873f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 253973f53c4aSTejun Heo wq->work_color = next_color; 254073f53c4aSTejun Heo 254173f53c4aSTejun Heo if (!wq->first_flusher) { 254273f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25436183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 254473f53c4aSTejun Heo 254573f53c4aSTejun Heo wq->first_flusher = &this_flusher; 254673f53c4aSTejun Heo 2547112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 254873f53c4aSTejun Heo wq->work_color)) { 254973f53c4aSTejun Heo /* nothing to flush, done */ 255073f53c4aSTejun Heo wq->flush_color = next_color; 255173f53c4aSTejun Heo wq->first_flusher = NULL; 255273f53c4aSTejun Heo goto out_unlock; 255373f53c4aSTejun Heo } 255473f53c4aSTejun Heo } else { 255573f53c4aSTejun Heo /* wait in queue */ 25566183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 255773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2558112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 255973f53c4aSTejun Heo } 256073f53c4aSTejun Heo } else { 256173f53c4aSTejun Heo /* 256273f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 256373f53c4aSTejun Heo * The next flush completion will assign us 256473f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 256573f53c4aSTejun Heo */ 256673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 256773f53c4aSTejun Heo } 256873f53c4aSTejun Heo 25693c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 257073f53c4aSTejun Heo 257173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 257273f53c4aSTejun Heo 257373f53c4aSTejun Heo /* 257473f53c4aSTejun Heo * Wake-up-and-cascade phase 257573f53c4aSTejun Heo * 257673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 257773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 257873f53c4aSTejun Heo */ 257973f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 258073f53c4aSTejun Heo return; 258173f53c4aSTejun Heo 25823c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 258373f53c4aSTejun Heo 25844ce48b37STejun Heo /* we might have raced, check again with mutex held */ 25854ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 25864ce48b37STejun Heo goto out_unlock; 25874ce48b37STejun Heo 258873f53c4aSTejun Heo wq->first_flusher = NULL; 258973f53c4aSTejun Heo 25906183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 25916183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 259273f53c4aSTejun Heo 259373f53c4aSTejun Heo while (true) { 259473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 259573f53c4aSTejun Heo 259673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 259773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 259873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 259973f53c4aSTejun Heo break; 260073f53c4aSTejun Heo list_del_init(&next->list); 260173f53c4aSTejun Heo complete(&next->done); 260273f53c4aSTejun Heo } 260373f53c4aSTejun Heo 26046183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 260573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 260673f53c4aSTejun Heo 260773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 260873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 260973f53c4aSTejun Heo 261073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 261173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 261273f53c4aSTejun Heo /* 261373f53c4aSTejun Heo * Assign the same color to all overflowed 261473f53c4aSTejun Heo * flushers, advance work_color and append to 261573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 261673f53c4aSTejun Heo * phase for these overflowed flushers. 261773f53c4aSTejun Heo */ 261873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 261973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 262073f53c4aSTejun Heo 262173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 262273f53c4aSTejun Heo 262373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 262473f53c4aSTejun Heo &wq->flusher_queue); 2625112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 262673f53c4aSTejun Heo } 262773f53c4aSTejun Heo 262873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26296183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 263073f53c4aSTejun Heo break; 263173f53c4aSTejun Heo } 263273f53c4aSTejun Heo 263373f53c4aSTejun Heo /* 263473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2635112202d9STejun Heo * the new first flusher and arm pwqs. 263673f53c4aSTejun Heo */ 26376183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26386183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 263973f53c4aSTejun Heo 264073f53c4aSTejun Heo list_del_init(&next->list); 264173f53c4aSTejun Heo wq->first_flusher = next; 264273f53c4aSTejun Heo 2643112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 264473f53c4aSTejun Heo break; 264573f53c4aSTejun Heo 264673f53c4aSTejun Heo /* 264773f53c4aSTejun Heo * Meh... this color is already done, clear first 264873f53c4aSTejun Heo * flusher and repeat cascading. 264973f53c4aSTejun Heo */ 265073f53c4aSTejun Heo wq->first_flusher = NULL; 265173f53c4aSTejun Heo } 265273f53c4aSTejun Heo 265373f53c4aSTejun Heo out_unlock: 26543c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 26551da177e4SLinus Torvalds } 2656ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 26571da177e4SLinus Torvalds 26589c5a2ba7STejun Heo /** 26599c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 26609c5a2ba7STejun Heo * @wq: workqueue to drain 26619c5a2ba7STejun Heo * 26629c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 26639c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 26649c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 26659c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 26669c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 26679c5a2ba7STejun Heo * takes too long. 26689c5a2ba7STejun Heo */ 26699c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 26709c5a2ba7STejun Heo { 26719c5a2ba7STejun Heo unsigned int flush_cnt = 0; 267249e3cf44STejun Heo struct pool_workqueue *pwq; 26739c5a2ba7STejun Heo 26749c5a2ba7STejun Heo /* 26759c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 26769c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2677618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 26789c5a2ba7STejun Heo */ 267987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 26809c5a2ba7STejun Heo if (!wq->nr_drainers++) 2681618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 268287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 26839c5a2ba7STejun Heo reflush: 26849c5a2ba7STejun Heo flush_workqueue(wq); 26859c5a2ba7STejun Heo 2686b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 268776af4d93STejun Heo 268849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2689fa2563e4SThomas Tuttle bool drained; 26909c5a2ba7STejun Heo 2691b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2692112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2693b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2694fa2563e4SThomas Tuttle 2695fa2563e4SThomas Tuttle if (drained) 26969c5a2ba7STejun Heo continue; 26979c5a2ba7STejun Heo 26989c5a2ba7STejun Heo if (++flush_cnt == 10 || 26999c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2700c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27019c5a2ba7STejun Heo wq->name, flush_cnt); 270276af4d93STejun Heo 2703b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27049c5a2ba7STejun Heo goto reflush; 27059c5a2ba7STejun Heo } 27069c5a2ba7STejun Heo 27079c5a2ba7STejun Heo if (!--wq->nr_drainers) 2708618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 270987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27109c5a2ba7STejun Heo } 27119c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27129c5a2ba7STejun Heo 2713606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2714baf59022STejun Heo { 2715baf59022STejun Heo struct worker *worker = NULL; 2716c9e7cf27STejun Heo struct worker_pool *pool; 2717112202d9STejun Heo struct pool_workqueue *pwq; 2718baf59022STejun Heo 2719baf59022STejun Heo might_sleep(); 2720baf59022STejun Heo 2721fa1b54e6STejun Heo local_irq_disable(); 2722fa1b54e6STejun Heo pool = get_work_pool(work); 2723fa1b54e6STejun Heo if (!pool) { 2724fa1b54e6STejun Heo local_irq_enable(); 2725fa1b54e6STejun Heo return false; 2726fa1b54e6STejun Heo } 2727fa1b54e6STejun Heo 2728fa1b54e6STejun Heo spin_lock(&pool->lock); 27290b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2730112202d9STejun Heo pwq = get_work_pwq(work); 2731112202d9STejun Heo if (pwq) { 2732112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2733baf59022STejun Heo goto already_gone; 2734606a5020STejun Heo } else { 2735c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2736baf59022STejun Heo if (!worker) 2737baf59022STejun Heo goto already_gone; 2738112202d9STejun Heo pwq = worker->current_pwq; 2739606a5020STejun Heo } 2740baf59022STejun Heo 2741112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2742d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2743baf59022STejun Heo 2744e159489bSTejun Heo /* 2745e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2746e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2747e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2748e159489bSTejun Heo * access. 2749e159489bSTejun Heo */ 2750493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2751112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2752e159489bSTejun Heo else 2753112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2754112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2755e159489bSTejun Heo 2756baf59022STejun Heo return true; 2757baf59022STejun Heo already_gone: 2758d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2759baf59022STejun Heo return false; 2760baf59022STejun Heo } 2761baf59022STejun Heo 2762db700897SOleg Nesterov /** 2763401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2764401a8d04STejun Heo * @work: the work to flush 2765db700897SOleg Nesterov * 2766606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2767606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2768401a8d04STejun Heo * 2769401a8d04STejun Heo * RETURNS: 2770401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2771401a8d04STejun Heo * %false if it was already idle. 2772db700897SOleg Nesterov */ 2773401a8d04STejun Heo bool flush_work(struct work_struct *work) 2774db700897SOleg Nesterov { 2775db700897SOleg Nesterov struct wq_barrier barr; 2776db700897SOleg Nesterov 27770976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 27780976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 27790976dfc1SStephen Boyd 2780606a5020STejun Heo if (start_flush_work(work, &barr)) { 2781db700897SOleg Nesterov wait_for_completion(&barr.done); 2782dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2783401a8d04STejun Heo return true; 2784606a5020STejun Heo } else { 2785401a8d04STejun Heo return false; 2786db700897SOleg Nesterov } 2787606a5020STejun Heo } 2788db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2789db700897SOleg Nesterov 279036e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2791401a8d04STejun Heo { 2792bbb68dfaSTejun Heo unsigned long flags; 27931f1f642eSOleg Nesterov int ret; 27941f1f642eSOleg Nesterov 27951f1f642eSOleg Nesterov do { 2796bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2797bbb68dfaSTejun Heo /* 2798bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2799bbb68dfaSTejun Heo * would be waiting for before retrying. 2800bbb68dfaSTejun Heo */ 2801bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2802606a5020STejun Heo flush_work(work); 28031f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28041f1f642eSOleg Nesterov 2805bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2806bbb68dfaSTejun Heo mark_work_canceling(work); 2807bbb68dfaSTejun Heo local_irq_restore(flags); 2808bbb68dfaSTejun Heo 2809606a5020STejun Heo flush_work(work); 28107a22ad75STejun Heo clear_work_data(work); 28111f1f642eSOleg Nesterov return ret; 28121f1f642eSOleg Nesterov } 28131f1f642eSOleg Nesterov 28146e84d644SOleg Nesterov /** 2815401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2816401a8d04STejun Heo * @work: the work to cancel 28176e84d644SOleg Nesterov * 2818401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2819401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2820401a8d04STejun Heo * another workqueue. On return from this function, @work is 2821401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28221f1f642eSOleg Nesterov * 2823401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2824401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28256e84d644SOleg Nesterov * 2826401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28276e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2828401a8d04STejun Heo * 2829401a8d04STejun Heo * RETURNS: 2830401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28316e84d644SOleg Nesterov */ 2832401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28336e84d644SOleg Nesterov { 283436e227d2STejun Heo return __cancel_work_timer(work, false); 2835b89deed3SOleg Nesterov } 283628e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2837b89deed3SOleg Nesterov 28386e84d644SOleg Nesterov /** 2839401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2840401a8d04STejun Heo * @dwork: the delayed work to flush 28416e84d644SOleg Nesterov * 2842401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2843401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2844401a8d04STejun Heo * considers the last queueing instance of @dwork. 28451f1f642eSOleg Nesterov * 2846401a8d04STejun Heo * RETURNS: 2847401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2848401a8d04STejun Heo * %false if it was already idle. 28496e84d644SOleg Nesterov */ 2850401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2851401a8d04STejun Heo { 28528930cabaSTejun Heo local_irq_disable(); 2853401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 285460c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 28558930cabaSTejun Heo local_irq_enable(); 2856401a8d04STejun Heo return flush_work(&dwork->work); 2857401a8d04STejun Heo } 2858401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2859401a8d04STejun Heo 2860401a8d04STejun Heo /** 286157b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 286257b30ae7STejun Heo * @dwork: delayed_work to cancel 286309383498STejun Heo * 286457b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 286557b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 286657b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 286757b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 286857b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 286909383498STejun Heo * 287057b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 287109383498STejun Heo */ 287257b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 287309383498STejun Heo { 287457b30ae7STejun Heo unsigned long flags; 287557b30ae7STejun Heo int ret; 287657b30ae7STejun Heo 287757b30ae7STejun Heo do { 287857b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 287957b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 288057b30ae7STejun Heo 288157b30ae7STejun Heo if (unlikely(ret < 0)) 288257b30ae7STejun Heo return false; 288357b30ae7STejun Heo 28847c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 28857c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 288657b30ae7STejun Heo local_irq_restore(flags); 2887c0158ca6SDan Magenheimer return ret; 288809383498STejun Heo } 288957b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 289009383498STejun Heo 289109383498STejun Heo /** 2892401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2893401a8d04STejun Heo * @dwork: the delayed work cancel 2894401a8d04STejun Heo * 2895401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2896401a8d04STejun Heo * 2897401a8d04STejun Heo * RETURNS: 2898401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2899401a8d04STejun Heo */ 2900401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29016e84d644SOleg Nesterov { 290236e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29036e84d644SOleg Nesterov } 2904f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29051da177e4SLinus Torvalds 29060fcb78c2SRolf Eike Beer /** 290731ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2908b6136773SAndrew Morton * @func: the function to call 2909b6136773SAndrew Morton * 291031ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 291131ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2912b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 291331ddd871STejun Heo * 291431ddd871STejun Heo * RETURNS: 291531ddd871STejun Heo * 0 on success, -errno on failure. 2916b6136773SAndrew Morton */ 291765f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 291815316ba8SChristoph Lameter { 291915316ba8SChristoph Lameter int cpu; 292038f51568SNamhyung Kim struct work_struct __percpu *works; 292115316ba8SChristoph Lameter 2922b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2923b6136773SAndrew Morton if (!works) 292415316ba8SChristoph Lameter return -ENOMEM; 2925b6136773SAndrew Morton 292695402b38SGautham R Shenoy get_online_cpus(); 292793981800STejun Heo 292815316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29299bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29309bfb1839SIngo Molnar 29319bfb1839SIngo Molnar INIT_WORK(work, func); 29328de6d308SOleg Nesterov schedule_work_on(cpu, work); 293315316ba8SChristoph Lameter } 293493981800STejun Heo 293593981800STejun Heo for_each_online_cpu(cpu) 29368616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 293793981800STejun Heo 293895402b38SGautham R Shenoy put_online_cpus(); 2939b6136773SAndrew Morton free_percpu(works); 294015316ba8SChristoph Lameter return 0; 294115316ba8SChristoph Lameter } 294215316ba8SChristoph Lameter 2943eef6a7d5SAlan Stern /** 2944eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2945eef6a7d5SAlan Stern * 2946eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2947eef6a7d5SAlan Stern * completion. 2948eef6a7d5SAlan Stern * 2949eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2950eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2951eef6a7d5SAlan Stern * will lead to deadlock: 2952eef6a7d5SAlan Stern * 2953eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2954eef6a7d5SAlan Stern * a lock held by your code or its caller. 2955eef6a7d5SAlan Stern * 2956eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 2957eef6a7d5SAlan Stern * 2958eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 2959eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 2960eef6a7d5SAlan Stern * what locks they need, which you have no control over. 2961eef6a7d5SAlan Stern * 2962eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 2963eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 2964eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 2965eef6a7d5SAlan Stern * cancel_work_sync() instead. 2966eef6a7d5SAlan Stern */ 29671da177e4SLinus Torvalds void flush_scheduled_work(void) 29681da177e4SLinus Torvalds { 2969d320c038STejun Heo flush_workqueue(system_wq); 29701da177e4SLinus Torvalds } 2971ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 29721da177e4SLinus Torvalds 29731da177e4SLinus Torvalds /** 29741fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 29751fa44ecaSJames Bottomley * @fn: the function to execute 29761fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 29771fa44ecaSJames Bottomley * be available when the work executes) 29781fa44ecaSJames Bottomley * 29791fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 29801fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 29811fa44ecaSJames Bottomley * 29821fa44ecaSJames Bottomley * Returns: 0 - function was executed 29831fa44ecaSJames Bottomley * 1 - function was scheduled for execution 29841fa44ecaSJames Bottomley */ 298565f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 29861fa44ecaSJames Bottomley { 29871fa44ecaSJames Bottomley if (!in_interrupt()) { 298865f27f38SDavid Howells fn(&ew->work); 29891fa44ecaSJames Bottomley return 0; 29901fa44ecaSJames Bottomley } 29911fa44ecaSJames Bottomley 299265f27f38SDavid Howells INIT_WORK(&ew->work, fn); 29931fa44ecaSJames Bottomley schedule_work(&ew->work); 29941fa44ecaSJames Bottomley 29951fa44ecaSJames Bottomley return 1; 29961fa44ecaSJames Bottomley } 29971fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 29981fa44ecaSJames Bottomley 2999226223abSTejun Heo #ifdef CONFIG_SYSFS 3000226223abSTejun Heo /* 3001226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3002226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3003226223abSTejun Heo * following attributes. 3004226223abSTejun Heo * 3005226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3006226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3007226223abSTejun Heo * 3008226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3009226223abSTejun Heo * 3010226223abSTejun Heo * id RO int : the associated pool ID 3011226223abSTejun Heo * nice RW int : nice value of the workers 3012226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3013226223abSTejun Heo */ 3014226223abSTejun Heo struct wq_device { 3015226223abSTejun Heo struct workqueue_struct *wq; 3016226223abSTejun Heo struct device dev; 3017226223abSTejun Heo }; 3018226223abSTejun Heo 3019226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3020226223abSTejun Heo { 3021226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3022226223abSTejun Heo 3023226223abSTejun Heo return wq_dev->wq; 3024226223abSTejun Heo } 3025226223abSTejun Heo 3026226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3027226223abSTejun Heo struct device_attribute *attr, char *buf) 3028226223abSTejun Heo { 3029226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3030226223abSTejun Heo 3031226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3032226223abSTejun Heo } 3033226223abSTejun Heo 3034226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3035226223abSTejun Heo struct device_attribute *attr, char *buf) 3036226223abSTejun Heo { 3037226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3038226223abSTejun Heo 3039226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3040226223abSTejun Heo } 3041226223abSTejun Heo 3042226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3043226223abSTejun Heo struct device_attribute *attr, 3044226223abSTejun Heo const char *buf, size_t count) 3045226223abSTejun Heo { 3046226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3047226223abSTejun Heo int val; 3048226223abSTejun Heo 3049226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3050226223abSTejun Heo return -EINVAL; 3051226223abSTejun Heo 3052226223abSTejun Heo workqueue_set_max_active(wq, val); 3053226223abSTejun Heo return count; 3054226223abSTejun Heo } 3055226223abSTejun Heo 3056226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3057226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3058226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3059226223abSTejun Heo __ATTR_NULL, 3060226223abSTejun Heo }; 3061226223abSTejun Heo 3062226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3063226223abSTejun Heo struct device_attribute *attr, char *buf) 3064226223abSTejun Heo { 3065226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3066226223abSTejun Heo struct worker_pool *pool; 3067226223abSTejun Heo int written; 3068226223abSTejun Heo 3069226223abSTejun Heo rcu_read_lock_sched(); 3070226223abSTejun Heo pool = first_pwq(wq)->pool; 3071226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3072226223abSTejun Heo rcu_read_unlock_sched(); 3073226223abSTejun Heo 3074226223abSTejun Heo return written; 3075226223abSTejun Heo } 3076226223abSTejun Heo 3077226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3078226223abSTejun Heo char *buf) 3079226223abSTejun Heo { 3080226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3081226223abSTejun Heo int written; 3082226223abSTejun Heo 3083226223abSTejun Heo rcu_read_lock_sched(); 3084226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3085226223abSTejun Heo first_pwq(wq)->pool->attrs->nice); 3086226223abSTejun Heo rcu_read_unlock_sched(); 3087226223abSTejun Heo 3088226223abSTejun Heo return written; 3089226223abSTejun Heo } 3090226223abSTejun Heo 3091226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3092226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3093226223abSTejun Heo { 3094226223abSTejun Heo struct workqueue_attrs *attrs; 3095226223abSTejun Heo 3096226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3097226223abSTejun Heo if (!attrs) 3098226223abSTejun Heo return NULL; 3099226223abSTejun Heo 3100226223abSTejun Heo rcu_read_lock_sched(); 3101226223abSTejun Heo copy_workqueue_attrs(attrs, first_pwq(wq)->pool->attrs); 3102226223abSTejun Heo rcu_read_unlock_sched(); 3103226223abSTejun Heo return attrs; 3104226223abSTejun Heo } 3105226223abSTejun Heo 3106226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3107226223abSTejun Heo const char *buf, size_t count) 3108226223abSTejun Heo { 3109226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3110226223abSTejun Heo struct workqueue_attrs *attrs; 3111226223abSTejun Heo int ret; 3112226223abSTejun Heo 3113226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3114226223abSTejun Heo if (!attrs) 3115226223abSTejun Heo return -ENOMEM; 3116226223abSTejun Heo 3117226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3118226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3119226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3120226223abSTejun Heo else 3121226223abSTejun Heo ret = -EINVAL; 3122226223abSTejun Heo 3123226223abSTejun Heo free_workqueue_attrs(attrs); 3124226223abSTejun Heo return ret ?: count; 3125226223abSTejun Heo } 3126226223abSTejun Heo 3127226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3128226223abSTejun Heo struct device_attribute *attr, char *buf) 3129226223abSTejun Heo { 3130226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3131226223abSTejun Heo int written; 3132226223abSTejun Heo 3133226223abSTejun Heo rcu_read_lock_sched(); 3134226223abSTejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, 3135226223abSTejun Heo first_pwq(wq)->pool->attrs->cpumask); 3136226223abSTejun Heo rcu_read_unlock_sched(); 3137226223abSTejun Heo 3138226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3139226223abSTejun Heo return written; 3140226223abSTejun Heo } 3141226223abSTejun Heo 3142226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3143226223abSTejun Heo struct device_attribute *attr, 3144226223abSTejun Heo const char *buf, size_t count) 3145226223abSTejun Heo { 3146226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3147226223abSTejun Heo struct workqueue_attrs *attrs; 3148226223abSTejun Heo int ret; 3149226223abSTejun Heo 3150226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3151226223abSTejun Heo if (!attrs) 3152226223abSTejun Heo return -ENOMEM; 3153226223abSTejun Heo 3154226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3155226223abSTejun Heo if (!ret) 3156226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3157226223abSTejun Heo 3158226223abSTejun Heo free_workqueue_attrs(attrs); 3159226223abSTejun Heo return ret ?: count; 3160226223abSTejun Heo } 3161226223abSTejun Heo 3162226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3163226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3164226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3165226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3166226223abSTejun Heo __ATTR_NULL, 3167226223abSTejun Heo }; 3168226223abSTejun Heo 3169226223abSTejun Heo static struct bus_type wq_subsys = { 3170226223abSTejun Heo .name = "workqueue", 3171226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3172226223abSTejun Heo }; 3173226223abSTejun Heo 3174226223abSTejun Heo static int __init wq_sysfs_init(void) 3175226223abSTejun Heo { 3176226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3177226223abSTejun Heo } 3178226223abSTejun Heo core_initcall(wq_sysfs_init); 3179226223abSTejun Heo 3180226223abSTejun Heo static void wq_device_release(struct device *dev) 3181226223abSTejun Heo { 3182226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3183226223abSTejun Heo 3184226223abSTejun Heo kfree(wq_dev); 3185226223abSTejun Heo } 3186226223abSTejun Heo 3187226223abSTejun Heo /** 3188226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3189226223abSTejun Heo * @wq: the workqueue to register 3190226223abSTejun Heo * 3191226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3192226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3193226223abSTejun Heo * which is the preferred method. 3194226223abSTejun Heo * 3195226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3196226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3197226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3198226223abSTejun Heo * attributes. 3199226223abSTejun Heo * 3200226223abSTejun Heo * Returns 0 on success, -errno on failure. 3201226223abSTejun Heo */ 3202226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3203226223abSTejun Heo { 3204226223abSTejun Heo struct wq_device *wq_dev; 3205226223abSTejun Heo int ret; 3206226223abSTejun Heo 3207226223abSTejun Heo /* 3208226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3209226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3210226223abSTejun Heo * workqueues. 3211226223abSTejun Heo */ 3212226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3213226223abSTejun Heo return -EINVAL; 3214226223abSTejun Heo 3215226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3216226223abSTejun Heo if (!wq_dev) 3217226223abSTejun Heo return -ENOMEM; 3218226223abSTejun Heo 3219226223abSTejun Heo wq_dev->wq = wq; 3220226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3221226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3222226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3223226223abSTejun Heo 3224226223abSTejun Heo /* 3225226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3226226223abSTejun Heo * everything is ready. 3227226223abSTejun Heo */ 3228226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3229226223abSTejun Heo 3230226223abSTejun Heo ret = device_register(&wq_dev->dev); 3231226223abSTejun Heo if (ret) { 3232226223abSTejun Heo kfree(wq_dev); 3233226223abSTejun Heo wq->wq_dev = NULL; 3234226223abSTejun Heo return ret; 3235226223abSTejun Heo } 3236226223abSTejun Heo 3237226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3238226223abSTejun Heo struct device_attribute *attr; 3239226223abSTejun Heo 3240226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3241226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3242226223abSTejun Heo if (ret) { 3243226223abSTejun Heo device_unregister(&wq_dev->dev); 3244226223abSTejun Heo wq->wq_dev = NULL; 3245226223abSTejun Heo return ret; 3246226223abSTejun Heo } 3247226223abSTejun Heo } 3248226223abSTejun Heo } 3249226223abSTejun Heo 3250226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3251226223abSTejun Heo return 0; 3252226223abSTejun Heo } 3253226223abSTejun Heo 3254226223abSTejun Heo /** 3255226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3256226223abSTejun Heo * @wq: the workqueue to unregister 3257226223abSTejun Heo * 3258226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3259226223abSTejun Heo */ 3260226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3261226223abSTejun Heo { 3262226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3263226223abSTejun Heo 3264226223abSTejun Heo if (!wq->wq_dev) 3265226223abSTejun Heo return; 3266226223abSTejun Heo 3267226223abSTejun Heo wq->wq_dev = NULL; 3268226223abSTejun Heo device_unregister(&wq_dev->dev); 3269226223abSTejun Heo } 3270226223abSTejun Heo #else /* CONFIG_SYSFS */ 3271226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3272226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3273226223abSTejun Heo 32747a4e344cSTejun Heo /** 32757a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 32767a4e344cSTejun Heo * @attrs: workqueue_attrs to free 32777a4e344cSTejun Heo * 32787a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 32797a4e344cSTejun Heo */ 32807a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 32817a4e344cSTejun Heo { 32827a4e344cSTejun Heo if (attrs) { 32837a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 32847a4e344cSTejun Heo kfree(attrs); 32857a4e344cSTejun Heo } 32867a4e344cSTejun Heo } 32877a4e344cSTejun Heo 32887a4e344cSTejun Heo /** 32897a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 32907a4e344cSTejun Heo * @gfp_mask: allocation mask to use 32917a4e344cSTejun Heo * 32927a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 32937a4e344cSTejun Heo * return it. Returns NULL on failure. 32947a4e344cSTejun Heo */ 32957a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 32967a4e344cSTejun Heo { 32977a4e344cSTejun Heo struct workqueue_attrs *attrs; 32987a4e344cSTejun Heo 32997a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33007a4e344cSTejun Heo if (!attrs) 33017a4e344cSTejun Heo goto fail; 33027a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33037a4e344cSTejun Heo goto fail; 33047a4e344cSTejun Heo 330513e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33067a4e344cSTejun Heo return attrs; 33077a4e344cSTejun Heo fail: 33087a4e344cSTejun Heo free_workqueue_attrs(attrs); 33097a4e344cSTejun Heo return NULL; 33107a4e344cSTejun Heo } 33117a4e344cSTejun Heo 331229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 331329c91e99STejun Heo const struct workqueue_attrs *from) 331429c91e99STejun Heo { 331529c91e99STejun Heo to->nice = from->nice; 331629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 331729c91e99STejun Heo } 331829c91e99STejun Heo 331929c91e99STejun Heo /* hash value of the content of @attr */ 332029c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 332129c91e99STejun Heo { 332229c91e99STejun Heo u32 hash = 0; 332329c91e99STejun Heo 332429c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 332513e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 332613e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 332729c91e99STejun Heo return hash; 332829c91e99STejun Heo } 332929c91e99STejun Heo 333029c91e99STejun Heo /* content equality test */ 333129c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 333229c91e99STejun Heo const struct workqueue_attrs *b) 333329c91e99STejun Heo { 333429c91e99STejun Heo if (a->nice != b->nice) 333529c91e99STejun Heo return false; 333629c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 333729c91e99STejun Heo return false; 333829c91e99STejun Heo return true; 333929c91e99STejun Heo } 334029c91e99STejun Heo 33417a4e344cSTejun Heo /** 33427a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33437a4e344cSTejun Heo * @pool: worker_pool to initialize 33447a4e344cSTejun Heo * 33457a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 334629c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 334729c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 334829c91e99STejun Heo * on @pool safely to release it. 33497a4e344cSTejun Heo */ 33507a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 33514e1a1f9aSTejun Heo { 33524e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 335329c91e99STejun Heo pool->id = -1; 335429c91e99STejun Heo pool->cpu = -1; 33554e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 33564e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 33574e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 33584e1a1f9aSTejun Heo hash_init(pool->busy_hash); 33594e1a1f9aSTejun Heo 33604e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 33614e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 33624e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 33634e1a1f9aSTejun Heo 33644e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 33654e1a1f9aSTejun Heo (unsigned long)pool); 33664e1a1f9aSTejun Heo 33674e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3368bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3369822d8405STejun Heo idr_init(&pool->worker_idr); 33707a4e344cSTejun Heo 337129c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 337229c91e99STejun Heo pool->refcnt = 1; 337329c91e99STejun Heo 337429c91e99STejun Heo /* shouldn't fail above this point */ 33757a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 33767a4e344cSTejun Heo if (!pool->attrs) 33777a4e344cSTejun Heo return -ENOMEM; 33787a4e344cSTejun Heo return 0; 33794e1a1f9aSTejun Heo } 33804e1a1f9aSTejun Heo 338129c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 338229c91e99STejun Heo { 338329c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 338429c91e99STejun Heo 3385822d8405STejun Heo idr_destroy(&pool->worker_idr); 338629c91e99STejun Heo free_workqueue_attrs(pool->attrs); 338729c91e99STejun Heo kfree(pool); 338829c91e99STejun Heo } 338929c91e99STejun Heo 339029c91e99STejun Heo /** 339129c91e99STejun Heo * put_unbound_pool - put a worker_pool 339229c91e99STejun Heo * @pool: worker_pool to put 339329c91e99STejun Heo * 339429c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3395c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3396c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3397c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3398a892caccSTejun Heo * 3399a892caccSTejun Heo * Should be called with wq_pool_mutex held. 340029c91e99STejun Heo */ 340129c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 340229c91e99STejun Heo { 340329c91e99STejun Heo struct worker *worker; 340429c91e99STejun Heo 3405a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3406a892caccSTejun Heo 3407a892caccSTejun Heo if (--pool->refcnt) 340829c91e99STejun Heo return; 340929c91e99STejun Heo 341029c91e99STejun Heo /* sanity checks */ 341129c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3412a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 341329c91e99STejun Heo return; 341429c91e99STejun Heo 341529c91e99STejun Heo /* release id and unhash */ 341629c91e99STejun Heo if (pool->id >= 0) 341729c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 341829c91e99STejun Heo hash_del(&pool->hash_node); 341929c91e99STejun Heo 3420c5aa87bbSTejun Heo /* 3421c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3422c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3423c5aa87bbSTejun Heo * manager_mutex. 3424c5aa87bbSTejun Heo */ 342529c91e99STejun Heo mutex_lock(&pool->manager_arb); 3426cd549687STejun Heo mutex_lock(&pool->manager_mutex); 342729c91e99STejun Heo spin_lock_irq(&pool->lock); 342829c91e99STejun Heo 342929c91e99STejun Heo while ((worker = first_worker(pool))) 343029c91e99STejun Heo destroy_worker(worker); 343129c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 343229c91e99STejun Heo 343329c91e99STejun Heo spin_unlock_irq(&pool->lock); 3434cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 343529c91e99STejun Heo mutex_unlock(&pool->manager_arb); 343629c91e99STejun Heo 343729c91e99STejun Heo /* shut down the timers */ 343829c91e99STejun Heo del_timer_sync(&pool->idle_timer); 343929c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 344029c91e99STejun Heo 344129c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 344229c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 344329c91e99STejun Heo } 344429c91e99STejun Heo 344529c91e99STejun Heo /** 344629c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 344729c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 344829c91e99STejun Heo * 344929c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 345029c91e99STejun Heo * reference count and return it. If there already is a matching 345129c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 345229c91e99STejun Heo * create a new one. On failure, returns NULL. 3453a892caccSTejun Heo * 3454a892caccSTejun Heo * Should be called with wq_pool_mutex held. 345529c91e99STejun Heo */ 345629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 345729c91e99STejun Heo { 345829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 345929c91e99STejun Heo struct worker_pool *pool; 346029c91e99STejun Heo 3461a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 346229c91e99STejun Heo 346329c91e99STejun Heo /* do we already have a matching pool? */ 346429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 346529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 346629c91e99STejun Heo pool->refcnt++; 346729c91e99STejun Heo goto out_unlock; 346829c91e99STejun Heo } 346929c91e99STejun Heo } 347029c91e99STejun Heo 347129c91e99STejun Heo /* nope, create a new one */ 347229c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 347329c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 347429c91e99STejun Heo goto fail; 347529c91e99STejun Heo 347612ee4fc6SLai Jiangshan if (workqueue_freezing) 347712ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 347812ee4fc6SLai Jiangshan 34798864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 348029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 348129c91e99STejun Heo 348229c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 348329c91e99STejun Heo goto fail; 348429c91e99STejun Heo 348529c91e99STejun Heo /* create and start the initial worker */ 3486ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 348729c91e99STejun Heo goto fail; 348829c91e99STejun Heo 348929c91e99STejun Heo /* install */ 349029c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 349129c91e99STejun Heo out_unlock: 349229c91e99STejun Heo return pool; 349329c91e99STejun Heo fail: 349429c91e99STejun Heo if (pool) 349529c91e99STejun Heo put_unbound_pool(pool); 349629c91e99STejun Heo return NULL; 349729c91e99STejun Heo } 349829c91e99STejun Heo 34998864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35008864b4e5STejun Heo { 35018864b4e5STejun Heo kmem_cache_free(pwq_cache, 35028864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35038864b4e5STejun Heo } 35048864b4e5STejun Heo 35058864b4e5STejun Heo /* 35068864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35078864b4e5STejun Heo * and needs to be destroyed. 35088864b4e5STejun Heo */ 35098864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35108864b4e5STejun Heo { 35118864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35128864b4e5STejun Heo unbound_release_work); 35138864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35148864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3515bc0caf09STejun Heo bool is_last; 35168864b4e5STejun Heo 35178864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35188864b4e5STejun Heo return; 35198864b4e5STejun Heo 352075ccf595STejun Heo /* 35213c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 352275ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 352375ccf595STejun Heo * and consistent with the linking path. 352475ccf595STejun Heo */ 35253c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35268864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3527bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35283c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35298864b4e5STejun Heo 3530a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35318864b4e5STejun Heo put_unbound_pool(pool); 3532a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3533a892caccSTejun Heo 35348864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 35358864b4e5STejun Heo 35368864b4e5STejun Heo /* 35378864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 35388864b4e5STejun Heo * is gonna access it anymore. Free it. 35398864b4e5STejun Heo */ 3540bc0caf09STejun Heo if (is_last) 35418864b4e5STejun Heo kfree(wq); 35428864b4e5STejun Heo } 35438864b4e5STejun Heo 35440fbd95aaSTejun Heo /** 3545699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 35460fbd95aaSTejun Heo * @pwq: target pool_workqueue 35470fbd95aaSTejun Heo * 3548699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3549699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3550699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35510fbd95aaSTejun Heo */ 3552699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35530fbd95aaSTejun Heo { 3554699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3555699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3556699ce097STejun Heo 3557699ce097STejun Heo /* for @wq->saved_max_active */ 3558a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3559699ce097STejun Heo 3560699ce097STejun Heo /* fast exit for non-freezable wqs */ 3561699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3562699ce097STejun Heo return; 3563699ce097STejun Heo 3564a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3565699ce097STejun Heo 3566699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3567699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35680fbd95aaSTejun Heo 35690fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35700fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35710fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3572951a078aSLai Jiangshan 3573951a078aSLai Jiangshan /* 3574951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3575951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3576951a078aSLai Jiangshan */ 3577951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3578699ce097STejun Heo } else { 3579699ce097STejun Heo pwq->max_active = 0; 3580699ce097STejun Heo } 3581699ce097STejun Heo 3582a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 35830fbd95aaSTejun Heo } 35840fbd95aaSTejun Heo 3585d2c1d404STejun Heo static void init_and_link_pwq(struct pool_workqueue *pwq, 3586d2c1d404STejun Heo struct workqueue_struct *wq, 35879e8cd2f5STejun Heo struct worker_pool *pool, 35889e8cd2f5STejun Heo struct pool_workqueue **p_last_pwq) 3589d2c1d404STejun Heo { 3590d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3591d2c1d404STejun Heo 3592d2c1d404STejun Heo pwq->pool = pool; 3593d2c1d404STejun Heo pwq->wq = wq; 3594d2c1d404STejun Heo pwq->flush_color = -1; 35958864b4e5STejun Heo pwq->refcnt = 1; 3596d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 3597d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 35988864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3599d2c1d404STejun Heo 36003c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 360175ccf595STejun Heo 3602983ca25eSTejun Heo /* 3603983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36043c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3605983ca25eSTejun Heo */ 36069e8cd2f5STejun Heo if (p_last_pwq) 36079e8cd2f5STejun Heo *p_last_pwq = first_pwq(wq); 360875ccf595STejun Heo pwq->work_color = wq->work_color; 3609983ca25eSTejun Heo 3610983ca25eSTejun Heo /* sync max_active to the current setting */ 3611983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3612983ca25eSTejun Heo 3613983ca25eSTejun Heo /* link in @pwq */ 36149e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3615a357fc03SLai Jiangshan 36163c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3617d2c1d404STejun Heo } 3618d2c1d404STejun Heo 36199e8cd2f5STejun Heo /** 36209e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 36219e8cd2f5STejun Heo * @wq: the target workqueue 36229e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 36239e8cd2f5STejun Heo * 36249e8cd2f5STejun Heo * Apply @attrs to an unbound workqueue @wq. If @attrs doesn't match the 36259e8cd2f5STejun Heo * current attributes, a new pwq is created and made the first pwq which 36269e8cd2f5STejun Heo * will serve all new work items. Older pwqs are released as in-flight 36279e8cd2f5STejun Heo * work items finish. Note that a work item which repeatedly requeues 36289e8cd2f5STejun Heo * itself back-to-back will stay on its current pwq. 36299e8cd2f5STejun Heo * 36309e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 36319e8cd2f5STejun Heo * failure. 36329e8cd2f5STejun Heo */ 36339e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 36349e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 36359e8cd2f5STejun Heo { 363613e2e556STejun Heo struct workqueue_attrs *new_attrs; 363713e2e556STejun Heo struct pool_workqueue *pwq = NULL, *last_pwq; 36389e8cd2f5STejun Heo struct worker_pool *pool; 36394862125bSTejun Heo int ret; 36409e8cd2f5STejun Heo 36418719dceaSTejun Heo /* only unbound workqueues can change attributes */ 36429e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 36439e8cd2f5STejun Heo return -EINVAL; 36449e8cd2f5STejun Heo 36458719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 36468719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 36478719dceaSTejun Heo return -EINVAL; 36488719dceaSTejun Heo 364913e2e556STejun Heo /* make a copy of @attrs and sanitize it */ 365013e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 365113e2e556STejun Heo if (!new_attrs) 365213e2e556STejun Heo goto enomem; 365313e2e556STejun Heo 365413e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 365513e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 365613e2e556STejun Heo 3657a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 3658a892caccSTejun Heo 36599e8cd2f5STejun Heo pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL); 3660a892caccSTejun Heo if (!pwq) { 3661a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 366213e2e556STejun Heo goto enomem; 3663a892caccSTejun Heo } 36649e8cd2f5STejun Heo 366513e2e556STejun Heo pool = get_unbound_pool(new_attrs); 3666a892caccSTejun Heo if (!pool) { 3667a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 366813e2e556STejun Heo goto enomem; 3669a892caccSTejun Heo } 3670a892caccSTejun Heo 3671a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 36729e8cd2f5STejun Heo 36739e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, pool, &last_pwq); 36749e8cd2f5STejun Heo if (last_pwq) { 36759e8cd2f5STejun Heo spin_lock_irq(&last_pwq->pool->lock); 36769e8cd2f5STejun Heo put_pwq(last_pwq); 36779e8cd2f5STejun Heo spin_unlock_irq(&last_pwq->pool->lock); 36789e8cd2f5STejun Heo } 36799e8cd2f5STejun Heo 36804862125bSTejun Heo ret = 0; 36814862125bSTejun Heo /* fall through */ 36824862125bSTejun Heo out_free: 36834862125bSTejun Heo free_workqueue_attrs(new_attrs); 36844862125bSTejun Heo return ret; 368513e2e556STejun Heo 368613e2e556STejun Heo enomem: 368713e2e556STejun Heo kmem_cache_free(pwq_cache, pwq); 36884862125bSTejun Heo ret = -ENOMEM; 36894862125bSTejun Heo goto out_free; 36909e8cd2f5STejun Heo } 36919e8cd2f5STejun Heo 369230cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 36931da177e4SLinus Torvalds { 369449e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 369530cdf249STejun Heo int cpu; 3696e1d8aa9fSFrederic Weisbecker 369730cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3698420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3699420c0ddbSTejun Heo if (!wq->cpu_pwqs) 370030cdf249STejun Heo return -ENOMEM; 370130cdf249STejun Heo 370230cdf249STejun Heo for_each_possible_cpu(cpu) { 37037fb98ea7STejun Heo struct pool_workqueue *pwq = 37047fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 37057a62c2c8STejun Heo struct worker_pool *cpu_pools = 3706f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 370730cdf249STejun Heo 37089e8cd2f5STejun Heo init_and_link_pwq(pwq, wq, &cpu_pools[highpri], NULL); 370930cdf249STejun Heo } 371030cdf249STejun Heo return 0; 37119e8cd2f5STejun Heo } else { 37129e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 37139e8cd2f5STejun Heo } 37140f900049STejun Heo } 37150f900049STejun Heo 3716f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3717f3421797STejun Heo const char *name) 3718b71ab8c2STejun Heo { 3719f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3720f3421797STejun Heo 3721f3421797STejun Heo if (max_active < 1 || max_active > lim) 3722044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3723f3421797STejun Heo max_active, name, 1, lim); 3724b71ab8c2STejun Heo 3725f3421797STejun Heo return clamp_val(max_active, 1, lim); 3726b71ab8c2STejun Heo } 3727b71ab8c2STejun Heo 3728b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 372997e37d7bSTejun Heo unsigned int flags, 37301e19ffc6STejun Heo int max_active, 3731eb13ba87SJohannes Berg struct lock_class_key *key, 3732b196be89STejun Heo const char *lock_name, ...) 37333af24433SOleg Nesterov { 3734b196be89STejun Heo va_list args, args1; 37353af24433SOleg Nesterov struct workqueue_struct *wq; 373649e3cf44STejun Heo struct pool_workqueue *pwq; 3737b196be89STejun Heo size_t namelen; 3738b196be89STejun Heo 3739b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3740b196be89STejun Heo va_start(args, lock_name); 3741b196be89STejun Heo va_copy(args1, args); 3742b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3743b196be89STejun Heo 3744b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3745b196be89STejun Heo if (!wq) 3746d2c1d404STejun Heo return NULL; 3747b196be89STejun Heo 3748b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3749b196be89STejun Heo va_end(args); 3750b196be89STejun Heo va_end(args1); 37513af24433SOleg Nesterov 3752d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3753b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 37543af24433SOleg Nesterov 3755b196be89STejun Heo /* init wq */ 375697e37d7bSTejun Heo wq->flags = flags; 3757a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 37583c25a55dSLai Jiangshan mutex_init(&wq->mutex); 3759112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 376030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 376173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 376273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 3763493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 37643af24433SOleg Nesterov 3765eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3766cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 37673af24433SOleg Nesterov 376830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 3769d2c1d404STejun Heo goto err_free_wq; 37701537663fSTejun Heo 3771493008a8STejun Heo /* 3772493008a8STejun Heo * Workqueues which may be used during memory reclaim should 3773493008a8STejun Heo * have a rescuer to guarantee forward progress. 3774493008a8STejun Heo */ 3775493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 3776e22bee78STejun Heo struct worker *rescuer; 3777e22bee78STejun Heo 3778d2c1d404STejun Heo rescuer = alloc_worker(); 3779e22bee78STejun Heo if (!rescuer) 3780d2c1d404STejun Heo goto err_destroy; 3781e22bee78STejun Heo 3782111c225aSTejun Heo rescuer->rescue_wq = wq; 3783111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 3784b196be89STejun Heo wq->name); 3785d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 3786d2c1d404STejun Heo kfree(rescuer); 3787d2c1d404STejun Heo goto err_destroy; 3788d2c1d404STejun Heo } 3789e22bee78STejun Heo 3790d2c1d404STejun Heo wq->rescuer = rescuer; 379114a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 3792e22bee78STejun Heo wake_up_process(rescuer->task); 37933af24433SOleg Nesterov } 37941537663fSTejun Heo 3795226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 3796226223abSTejun Heo goto err_destroy; 3797226223abSTejun Heo 37983af24433SOleg Nesterov /* 379968e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 380068e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 380168e13a67SLai Jiangshan * list. 38023af24433SOleg Nesterov */ 380368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3804a0a1a5fdSTejun Heo 3805a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 380649e3cf44STejun Heo for_each_pwq(pwq, wq) 3807699ce097STejun Heo pwq_adjust_max_active(pwq); 3808a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3809a0a1a5fdSTejun Heo 38103af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3811a0a1a5fdSTejun Heo 381268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38133af24433SOleg Nesterov 38143af24433SOleg Nesterov return wq; 3815d2c1d404STejun Heo 3816d2c1d404STejun Heo err_free_wq: 38174690c4abSTejun Heo kfree(wq); 3818d2c1d404STejun Heo return NULL; 3819d2c1d404STejun Heo err_destroy: 3820d2c1d404STejun Heo destroy_workqueue(wq); 38214690c4abSTejun Heo return NULL; 38221da177e4SLinus Torvalds } 3823d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 38241da177e4SLinus Torvalds 38253af24433SOleg Nesterov /** 38263af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 38273af24433SOleg Nesterov * @wq: target workqueue 38283af24433SOleg Nesterov * 38293af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 38303af24433SOleg Nesterov */ 38313af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 38323af24433SOleg Nesterov { 383349e3cf44STejun Heo struct pool_workqueue *pwq; 38343af24433SOleg Nesterov 38359c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 38369c5a2ba7STejun Heo drain_workqueue(wq); 3837c8efcc25STejun Heo 38386183c009STejun Heo /* sanity checks */ 3839b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 384049e3cf44STejun Heo for_each_pwq(pwq, wq) { 38416183c009STejun Heo int i; 38426183c009STejun Heo 384376af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 384476af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 3845b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38466183c009STejun Heo return; 384776af4d93STejun Heo } 384876af4d93STejun Heo } 384976af4d93STejun Heo 38508864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 38518864b4e5STejun Heo WARN_ON(pwq->nr_active) || 385276af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 3853b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38546183c009STejun Heo return; 38556183c009STejun Heo } 385676af4d93STejun Heo } 3857b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 38586183c009STejun Heo 3859a0a1a5fdSTejun Heo /* 3860a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3861a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3862a0a1a5fdSTejun Heo */ 386368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 3864d2c1d404STejun Heo list_del_init(&wq->list); 386568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 38663af24433SOleg Nesterov 3867226223abSTejun Heo workqueue_sysfs_unregister(wq); 3868226223abSTejun Heo 3869493008a8STejun Heo if (wq->rescuer) { 3870e22bee78STejun Heo kthread_stop(wq->rescuer->task); 38718d9df9f0SXiaotian Feng kfree(wq->rescuer); 3872493008a8STejun Heo wq->rescuer = NULL; 3873e22bee78STejun Heo } 3874e22bee78STejun Heo 38758864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 387629c91e99STejun Heo /* 38778864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 38788864b4e5STejun Heo * free the pwqs and wq. 387929c91e99STejun Heo */ 38808864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 38818864b4e5STejun Heo kfree(wq); 38828864b4e5STejun Heo } else { 38838864b4e5STejun Heo /* 38848864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 38858864b4e5STejun Heo * access the first pwq and put the base ref. As both pwqs 38868864b4e5STejun Heo * and pools are sched-RCU protected, the lock operations 38878864b4e5STejun Heo * are safe. @wq will be freed when the last pwq is 38888864b4e5STejun Heo * released. 38898864b4e5STejun Heo */ 389029c91e99STejun Heo pwq = list_first_entry(&wq->pwqs, struct pool_workqueue, 389129c91e99STejun Heo pwqs_node); 38928864b4e5STejun Heo spin_lock_irq(&pwq->pool->lock); 38938864b4e5STejun Heo put_pwq(pwq); 38948864b4e5STejun Heo spin_unlock_irq(&pwq->pool->lock); 389529c91e99STejun Heo } 38963af24433SOleg Nesterov } 38973af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 38983af24433SOleg Nesterov 3899dcd989cbSTejun Heo /** 3900dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3901dcd989cbSTejun Heo * @wq: target workqueue 3902dcd989cbSTejun Heo * @max_active: new max_active value. 3903dcd989cbSTejun Heo * 3904dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3905dcd989cbSTejun Heo * 3906dcd989cbSTejun Heo * CONTEXT: 3907dcd989cbSTejun Heo * Don't call from IRQ context. 3908dcd989cbSTejun Heo */ 3909dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3910dcd989cbSTejun Heo { 391149e3cf44STejun Heo struct pool_workqueue *pwq; 3912dcd989cbSTejun Heo 39138719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 39148719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 39158719dceaSTejun Heo return; 39168719dceaSTejun Heo 3917f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3918dcd989cbSTejun Heo 3919a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 3920dcd989cbSTejun Heo 3921dcd989cbSTejun Heo wq->saved_max_active = max_active; 3922dcd989cbSTejun Heo 3923699ce097STejun Heo for_each_pwq(pwq, wq) 3924699ce097STejun Heo pwq_adjust_max_active(pwq); 3925dcd989cbSTejun Heo 3926a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 3927dcd989cbSTejun Heo } 3928dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3929dcd989cbSTejun Heo 3930dcd989cbSTejun Heo /** 3931e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 3932e6267616STejun Heo * 3933e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 3934e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 3935e6267616STejun Heo */ 3936e6267616STejun Heo bool current_is_workqueue_rescuer(void) 3937e6267616STejun Heo { 3938e6267616STejun Heo struct worker *worker = current_wq_worker(); 3939e6267616STejun Heo 39406a092dfdSLai Jiangshan return worker && worker->rescue_wq; 3941e6267616STejun Heo } 3942e6267616STejun Heo 3943e6267616STejun Heo /** 3944dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3945dcd989cbSTejun Heo * @cpu: CPU in question 3946dcd989cbSTejun Heo * @wq: target workqueue 3947dcd989cbSTejun Heo * 3948dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3949dcd989cbSTejun Heo * no synchronization around this function and the test result is 3950dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3951dcd989cbSTejun Heo * 3952dcd989cbSTejun Heo * RETURNS: 3953dcd989cbSTejun Heo * %true if congested, %false otherwise. 3954dcd989cbSTejun Heo */ 3955d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 3956dcd989cbSTejun Heo { 39577fb98ea7STejun Heo struct pool_workqueue *pwq; 395876af4d93STejun Heo bool ret; 395976af4d93STejun Heo 396088109453SLai Jiangshan rcu_read_lock_sched(); 39617fb98ea7STejun Heo 39627fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 39637fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 39647fb98ea7STejun Heo else 39657fb98ea7STejun Heo pwq = first_pwq(wq); 3966dcd989cbSTejun Heo 396776af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 396888109453SLai Jiangshan rcu_read_unlock_sched(); 396976af4d93STejun Heo 397076af4d93STejun Heo return ret; 3971dcd989cbSTejun Heo } 3972dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3973dcd989cbSTejun Heo 3974dcd989cbSTejun Heo /** 3975dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3976dcd989cbSTejun Heo * @work: the work to be tested 3977dcd989cbSTejun Heo * 3978dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3979dcd989cbSTejun Heo * synchronization around this function and the test result is 3980dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3981dcd989cbSTejun Heo * 3982dcd989cbSTejun Heo * RETURNS: 3983dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3984dcd989cbSTejun Heo */ 3985dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3986dcd989cbSTejun Heo { 3987fa1b54e6STejun Heo struct worker_pool *pool; 3988dcd989cbSTejun Heo unsigned long flags; 3989dcd989cbSTejun Heo unsigned int ret = 0; 3990dcd989cbSTejun Heo 3991dcd989cbSTejun Heo if (work_pending(work)) 3992dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3993038366c5SLai Jiangshan 3994fa1b54e6STejun Heo local_irq_save(flags); 3995fa1b54e6STejun Heo pool = get_work_pool(work); 3996038366c5SLai Jiangshan if (pool) { 3997fa1b54e6STejun Heo spin_lock(&pool->lock); 3998c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 3999dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4000fa1b54e6STejun Heo spin_unlock(&pool->lock); 4001038366c5SLai Jiangshan } 4002fa1b54e6STejun Heo local_irq_restore(flags); 4003dcd989cbSTejun Heo 4004dcd989cbSTejun Heo return ret; 4005dcd989cbSTejun Heo } 4006dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4007dcd989cbSTejun Heo 4008db7bccf4STejun Heo /* 4009db7bccf4STejun Heo * CPU hotplug. 4010db7bccf4STejun Heo * 4011e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4012112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4013706026c2STejun Heo * pool which make migrating pending and scheduled works very 4014e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 401594cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4016e22bee78STejun Heo * blocked draining impractical. 4017e22bee78STejun Heo * 401824647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4019628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4020628c78e7STejun Heo * cpu comes back online. 4021db7bccf4STejun Heo */ 4022db7bccf4STejun Heo 4023706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4024db7bccf4STejun Heo { 402538db41d9STejun Heo int cpu = smp_processor_id(); 40264ce62e9eSTejun Heo struct worker_pool *pool; 4027db7bccf4STejun Heo struct worker *worker; 4028a9ab775bSTejun Heo int wi; 4029db7bccf4STejun Heo 4030f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 40316183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4032db7bccf4STejun Heo 4033bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 403494cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4035e22bee78STejun Heo 4036f2d5a0eeSTejun Heo /* 4037bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 403894cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 403994cf58bbSTejun Heo * except for the ones which are still executing works from 404094cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 404194cf58bbSTejun Heo * this, they may become diasporas. 4042f2d5a0eeSTejun Heo */ 4043a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4044403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4045db7bccf4STejun Heo 404624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4047f2d5a0eeSTejun Heo 404894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4049bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 405094cf58bbSTejun Heo } 4051e22bee78STejun Heo 4052e22bee78STejun Heo /* 4053628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4054628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4055628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4056628c78e7STejun Heo */ 4057628c78e7STejun Heo schedule(); 4058628c78e7STejun Heo 4059628c78e7STejun Heo /* 4060628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4061628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 406238db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 406338db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 406438db41d9STejun Heo * pools which are served by workers tied to the CPU. 4065628c78e7STejun Heo * 4066628c78e7STejun Heo * On return from this function, the current worker would trigger 4067628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4068628c78e7STejun Heo * didn't already. 4069e22bee78STejun Heo */ 4070f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4071e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4072db7bccf4STejun Heo } 4073db7bccf4STejun Heo 4074bd7c089eSTejun Heo /** 4075bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4076bd7c089eSTejun Heo * @pool: pool of interest 4077bd7c089eSTejun Heo * 4078a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4079bd7c089eSTejun Heo */ 4080bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4081bd7c089eSTejun Heo { 4082a9ab775bSTejun Heo struct worker *worker; 4083a9ab775bSTejun Heo int wi; 4084bd7c089eSTejun Heo 4085bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4086bd7c089eSTejun Heo 4087bd7c089eSTejun Heo /* 4088a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4089a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4090a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4091a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4092a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4093bd7c089eSTejun Heo */ 4094a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4095a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4096a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4097a9ab775bSTejun Heo 4098a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4099a9ab775bSTejun Heo 4100a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4101a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4102a9ab775bSTejun Heo 4103a9ab775bSTejun Heo /* 4104a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4105a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4106a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4107a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4108a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4109a9ab775bSTejun Heo * be bound before @pool->lock is released. 4110a9ab775bSTejun Heo */ 4111a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4112bd7c089eSTejun Heo wake_up_process(worker->task); 4113bd7c089eSTejun Heo 4114bd7c089eSTejun Heo /* 4115a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4116a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4117a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4118a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4119a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4120a9ab775bSTejun Heo * concurrency management. Note that when or whether 4121a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4122a9ab775bSTejun Heo * 4123a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4124a9ab775bSTejun Heo * tested without holding any lock in 4125a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4126a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4127a9ab775bSTejun Heo * management operations. 4128bd7c089eSTejun Heo */ 4129a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4130a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4131a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4132a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4133bd7c089eSTejun Heo } 4134a9ab775bSTejun Heo 4135a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4136bd7c089eSTejun Heo } 4137bd7c089eSTejun Heo 41387dbc725eSTejun Heo /** 41397dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 41407dbc725eSTejun Heo * @pool: unbound pool of interest 41417dbc725eSTejun Heo * @cpu: the CPU which is coming up 41427dbc725eSTejun Heo * 41437dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 41447dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 41457dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 41467dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 41477dbc725eSTejun Heo */ 41487dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 41497dbc725eSTejun Heo { 41507dbc725eSTejun Heo static cpumask_t cpumask; 41517dbc725eSTejun Heo struct worker *worker; 41527dbc725eSTejun Heo int wi; 41537dbc725eSTejun Heo 41547dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 41557dbc725eSTejun Heo 41567dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 41577dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 41587dbc725eSTejun Heo return; 41597dbc725eSTejun Heo 41607dbc725eSTejun Heo /* is @cpu the only online CPU? */ 41617dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 41627dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 41637dbc725eSTejun Heo return; 41647dbc725eSTejun Heo 41657dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 41667dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 41677dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 41687dbc725eSTejun Heo pool->attrs->cpumask) < 0); 41697dbc725eSTejun Heo } 41707dbc725eSTejun Heo 41718db25e78STejun Heo /* 41728db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 41738db25e78STejun Heo * This will be registered high priority CPU notifier. 41748db25e78STejun Heo */ 41759fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 41761da177e4SLinus Torvalds unsigned long action, 41771da177e4SLinus Torvalds void *hcpu) 41781da177e4SLinus Torvalds { 4179d84ff051STejun Heo int cpu = (unsigned long)hcpu; 41804ce62e9eSTejun Heo struct worker_pool *pool; 41817dbc725eSTejun Heo int pi; 41821da177e4SLinus Torvalds 41838db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 41843af24433SOleg Nesterov case CPU_UP_PREPARE: 4185f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 41863ce63377STejun Heo if (pool->nr_workers) 41873ce63377STejun Heo continue; 4188ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 41893ce63377STejun Heo return NOTIFY_BAD; 41903af24433SOleg Nesterov } 41911da177e4SLinus Torvalds break; 41921da177e4SLinus Torvalds 419365758202STejun Heo case CPU_DOWN_FAILED: 419465758202STejun Heo case CPU_ONLINE: 419568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 41967dbc725eSTejun Heo 41977dbc725eSTejun Heo for_each_pool(pool, pi) { 4198bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 419994cf58bbSTejun Heo 42007dbc725eSTejun Heo if (pool->cpu == cpu) { 4201a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 420224647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4203a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4204a9ab775bSTejun Heo 420594cf58bbSTejun Heo rebind_workers(pool); 42067dbc725eSTejun Heo } else if (pool->cpu < 0) { 42077dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 42087dbc725eSTejun Heo } 420994cf58bbSTejun Heo 4210bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 421194cf58bbSTejun Heo } 42127dbc725eSTejun Heo 421368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42148db25e78STejun Heo break; 421565758202STejun Heo } 421665758202STejun Heo return NOTIFY_OK; 421765758202STejun Heo } 421865758202STejun Heo 421965758202STejun Heo /* 422065758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 422165758202STejun Heo * This will be registered as low priority CPU notifier. 422265758202STejun Heo */ 42239fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 422465758202STejun Heo unsigned long action, 422565758202STejun Heo void *hcpu) 422665758202STejun Heo { 4227d84ff051STejun Heo int cpu = (unsigned long)hcpu; 42288db25e78STejun Heo struct work_struct unbind_work; 42298db25e78STejun Heo 423065758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 423165758202STejun Heo case CPU_DOWN_PREPARE: 42328db25e78STejun Heo /* unbinding should happen on the local CPU */ 4233706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 42347635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 42358db25e78STejun Heo flush_work(&unbind_work); 42368db25e78STejun Heo break; 423765758202STejun Heo } 423865758202STejun Heo return NOTIFY_OK; 423965758202STejun Heo } 424065758202STejun Heo 42412d3854a3SRusty Russell #ifdef CONFIG_SMP 42428ccad40dSRusty Russell 42432d3854a3SRusty Russell struct work_for_cpu { 4244ed48ece2STejun Heo struct work_struct work; 42452d3854a3SRusty Russell long (*fn)(void *); 42462d3854a3SRusty Russell void *arg; 42472d3854a3SRusty Russell long ret; 42482d3854a3SRusty Russell }; 42492d3854a3SRusty Russell 4250ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 42512d3854a3SRusty Russell { 4252ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4253ed48ece2STejun Heo 42542d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 42552d3854a3SRusty Russell } 42562d3854a3SRusty Russell 42572d3854a3SRusty Russell /** 42582d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 42592d3854a3SRusty Russell * @cpu: the cpu to run on 42602d3854a3SRusty Russell * @fn: the function to run 42612d3854a3SRusty Russell * @arg: the function arg 42622d3854a3SRusty Russell * 426331ad9081SRusty Russell * This will return the value @fn returns. 426431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 42656b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 42662d3854a3SRusty Russell */ 4267d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 42682d3854a3SRusty Russell { 4269ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 42702d3854a3SRusty Russell 4271ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4272ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4273ed48ece2STejun Heo flush_work(&wfc.work); 42742d3854a3SRusty Russell return wfc.ret; 42752d3854a3SRusty Russell } 42762d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 42772d3854a3SRusty Russell #endif /* CONFIG_SMP */ 42782d3854a3SRusty Russell 4279a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4280e7577c50SRusty Russell 4281a0a1a5fdSTejun Heo /** 4282a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4283a0a1a5fdSTejun Heo * 428458a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4285c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4286706026c2STejun Heo * pool->worklist. 4287a0a1a5fdSTejun Heo * 4288a0a1a5fdSTejun Heo * CONTEXT: 4289a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4290a0a1a5fdSTejun Heo */ 4291a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4292a0a1a5fdSTejun Heo { 429317116969STejun Heo struct worker_pool *pool; 429424b8a847STejun Heo struct workqueue_struct *wq; 429524b8a847STejun Heo struct pool_workqueue *pwq; 4296611c92a0STejun Heo int pi; 4297a0a1a5fdSTejun Heo 429868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4299a0a1a5fdSTejun Heo 43006183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4301a0a1a5fdSTejun Heo workqueue_freezing = true; 4302a0a1a5fdSTejun Heo 430324b8a847STejun Heo /* set FREEZING */ 4304611c92a0STejun Heo for_each_pool(pool, pi) { 43055bcab335STejun Heo spin_lock_irq(&pool->lock); 430635b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 430735b6bb63STejun Heo pool->flags |= POOL_FREEZING; 43085bcab335STejun Heo spin_unlock_irq(&pool->lock); 43091da177e4SLinus Torvalds } 43108b03ae3cSTejun Heo 431124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4312a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4313699ce097STejun Heo for_each_pwq(pwq, wq) 4314699ce097STejun Heo pwq_adjust_max_active(pwq); 4315a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4316a1056305STejun Heo } 43175bcab335STejun Heo 431868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4319a0a1a5fdSTejun Heo } 4320a0a1a5fdSTejun Heo 4321a0a1a5fdSTejun Heo /** 432258a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4323a0a1a5fdSTejun Heo * 4324a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4325a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4326a0a1a5fdSTejun Heo * 4327a0a1a5fdSTejun Heo * CONTEXT: 432868e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4329a0a1a5fdSTejun Heo * 4330a0a1a5fdSTejun Heo * RETURNS: 433158a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 433258a69cb4STejun Heo * is complete. 4333a0a1a5fdSTejun Heo */ 4334a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4335a0a1a5fdSTejun Heo { 4336a0a1a5fdSTejun Heo bool busy = false; 433724b8a847STejun Heo struct workqueue_struct *wq; 433824b8a847STejun Heo struct pool_workqueue *pwq; 4339a0a1a5fdSTejun Heo 434068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4341a0a1a5fdSTejun Heo 43426183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4343a0a1a5fdSTejun Heo 434424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 434524b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 434624b8a847STejun Heo continue; 4347a0a1a5fdSTejun Heo /* 4348a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4349a0a1a5fdSTejun Heo * to peek without lock. 4350a0a1a5fdSTejun Heo */ 435188109453SLai Jiangshan rcu_read_lock_sched(); 435224b8a847STejun Heo for_each_pwq(pwq, wq) { 43536183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4354112202d9STejun Heo if (pwq->nr_active) { 4355a0a1a5fdSTejun Heo busy = true; 435688109453SLai Jiangshan rcu_read_unlock_sched(); 4357a0a1a5fdSTejun Heo goto out_unlock; 4358a0a1a5fdSTejun Heo } 4359a0a1a5fdSTejun Heo } 436088109453SLai Jiangshan rcu_read_unlock_sched(); 4361a0a1a5fdSTejun Heo } 4362a0a1a5fdSTejun Heo out_unlock: 436368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4364a0a1a5fdSTejun Heo return busy; 4365a0a1a5fdSTejun Heo } 4366a0a1a5fdSTejun Heo 4367a0a1a5fdSTejun Heo /** 4368a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4369a0a1a5fdSTejun Heo * 4370a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4371706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4372a0a1a5fdSTejun Heo * 4373a0a1a5fdSTejun Heo * CONTEXT: 4374a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4375a0a1a5fdSTejun Heo */ 4376a0a1a5fdSTejun Heo void thaw_workqueues(void) 4377a0a1a5fdSTejun Heo { 437824b8a847STejun Heo struct workqueue_struct *wq; 437924b8a847STejun Heo struct pool_workqueue *pwq; 438024b8a847STejun Heo struct worker_pool *pool; 4381611c92a0STejun Heo int pi; 4382a0a1a5fdSTejun Heo 438368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4384a0a1a5fdSTejun Heo 4385a0a1a5fdSTejun Heo if (!workqueue_freezing) 4386a0a1a5fdSTejun Heo goto out_unlock; 4387a0a1a5fdSTejun Heo 438824b8a847STejun Heo /* clear FREEZING */ 4389611c92a0STejun Heo for_each_pool(pool, pi) { 43905bcab335STejun Heo spin_lock_irq(&pool->lock); 439135b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 439235b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 43935bcab335STejun Heo spin_unlock_irq(&pool->lock); 4394d565ed63STejun Heo } 439524b8a847STejun Heo 439624b8a847STejun Heo /* restore max_active and repopulate worklist */ 439724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4398a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4399699ce097STejun Heo for_each_pwq(pwq, wq) 4400699ce097STejun Heo pwq_adjust_max_active(pwq); 4401a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 440224b8a847STejun Heo } 440324b8a847STejun Heo 4404a0a1a5fdSTejun Heo workqueue_freezing = false; 4405a0a1a5fdSTejun Heo out_unlock: 440668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4407a0a1a5fdSTejun Heo } 4408a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4409a0a1a5fdSTejun Heo 44106ee0578bSSuresh Siddha static int __init init_workqueues(void) 44111da177e4SLinus Torvalds { 44127a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 44137a4e344cSTejun Heo int i, cpu; 4414c34056a3STejun Heo 44157c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 44167c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 44176be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4418b5490077STejun Heo 4419e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4420e904e6c2STejun Heo 4421e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4422e904e6c2STejun Heo 442365758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4424a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 44258b03ae3cSTejun Heo 4426706026c2STejun Heo /* initialize CPU pools */ 442729c91e99STejun Heo for_each_possible_cpu(cpu) { 44284ce62e9eSTejun Heo struct worker_pool *pool; 44298b03ae3cSTejun Heo 44307a4e344cSTejun Heo i = 0; 4431f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44327a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4433ec22ca5eSTejun Heo pool->cpu = cpu; 44347a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 44357a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 44367a4e344cSTejun Heo 44379daf9e67STejun Heo /* alloc pool ID */ 443868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 44399daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 444068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44414ce62e9eSTejun Heo } 44428b03ae3cSTejun Heo } 44438b03ae3cSTejun Heo 4444e22bee78STejun Heo /* create the initial worker */ 444529c91e99STejun Heo for_each_online_cpu(cpu) { 44464ce62e9eSTejun Heo struct worker_pool *pool; 4447e22bee78STejun Heo 4448f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 444924647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4450ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4451e22bee78STejun Heo } 44524ce62e9eSTejun Heo } 4453e22bee78STejun Heo 445429c91e99STejun Heo /* create default unbound wq attrs */ 445529c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 445629c91e99STejun Heo struct workqueue_attrs *attrs; 445729c91e99STejun Heo 445829c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 445929c91e99STejun Heo attrs->nice = std_nice[i]; 446029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 446129c91e99STejun Heo } 446229c91e99STejun Heo 4463d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 44641aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4465d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4466f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4467f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 446824d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 446924d51addSTejun Heo WQ_FREEZABLE, 0); 44701aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4471ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 44726ee0578bSSuresh Siddha return 0; 44731da177e4SLinus Torvalds } 44746ee0578bSSuresh Siddha early_initcall(init_workqueues); 4475