11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4429c91e99STejun Heo #include <linux/jhash.h> 4542f8570fSSasha Levin #include <linux/hashtable.h> 4676af4d93STejun Heo #include <linux/rculist.h> 47bce90380STejun Heo #include <linux/nodemask.h> 484c16bd32STejun Heo #include <linux/moduleparam.h> 49e22bee78STejun Heo 50ea138446STejun Heo #include "workqueue_internal.h" 511da177e4SLinus Torvalds 52c8e55f36STejun Heo enum { 53bc2ae0f5STejun Heo /* 5424647570STejun Heo * worker_pool flags 55bc2ae0f5STejun Heo * 5624647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 57bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 58bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 59bc2ae0f5STejun Heo * is in effect. 60bc2ae0f5STejun Heo * 61bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 62bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6324647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 64bc2ae0f5STejun Heo * 65bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 66bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6724647570STejun Heo * create_worker() is in progress. 68bc2ae0f5STejun Heo */ 6911ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 7024647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 7135b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 72db7bccf4STejun Heo 73c8e55f36STejun Heo /* worker flags */ 74c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 75c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 76c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 77e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 78fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 79f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 80a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 81e22bee78STejun Heo 82a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 83a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 84db7bccf4STejun Heo 85e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 864ce62e9eSTejun Heo 8729c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 88c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 89db7bccf4STejun Heo 90e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 91e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 92e22bee78STejun Heo 933233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 943233cdbdSTejun Heo /* call for help after 10ms 953233cdbdSTejun Heo (min two ticks) */ 96e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 97e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 981da177e4SLinus Torvalds 991da177e4SLinus Torvalds /* 100e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 101e22bee78STejun Heo * all cpus. Give -20. 102e22bee78STejun Heo */ 103e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1043270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 105ecf6881fSTejun Heo 106ecf6881fSTejun Heo WQ_NAME_LEN = 24, 107c8e55f36STejun Heo }; 108c8e55f36STejun Heo 1091da177e4SLinus Torvalds /* 1104690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1114690c4abSTejun Heo * 112e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 113e41e704bSTejun Heo * everyone else. 1144690c4abSTejun Heo * 115e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 116e22bee78STejun Heo * only be modified and accessed from the local cpu. 117e22bee78STejun Heo * 118d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1194690c4abSTejun Heo * 120d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 121d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 122d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 123d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 124e22bee78STejun Heo * 125822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 126822d8405STejun Heo * locks. Reads can happen under either lock. 127822d8405STejun Heo * 12868e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12976af4d93STejun Heo * 13068e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1315bcab335STejun Heo * 1323c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1333c25a55dSLai Jiangshan * 134b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1352e109a28STejun Heo * 1362e109a28STejun Heo * MD: wq_mayday_lock protected. 1374690c4abSTejun Heo */ 1384690c4abSTejun Heo 1392eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 140c34056a3STejun Heo 141bd7bdd43STejun Heo struct worker_pool { 142d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 143d84ff051STejun Heo int cpu; /* I: the associated cpu */ 144f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1459daf9e67STejun Heo int id; /* I: pool ID */ 14611ebea50STejun Heo unsigned int flags; /* X: flags */ 147bd7bdd43STejun Heo 148bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 149bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 150ea1abd61SLai Jiangshan 151ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 152bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 153bd7bdd43STejun Heo 154bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 155bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 156bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 157bd7bdd43STejun Heo 158c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 159c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 160c9e7cf27STejun Heo /* L: hash of busy workers */ 161c9e7cf27STejun Heo 162bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16334a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 164bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 165822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 166e19e397aSTejun Heo 1677a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16868e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 16968e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1707a4e344cSTejun Heo 171e19e397aSTejun Heo /* 172e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 173e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 174e19e397aSTejun Heo * cacheline. 175e19e397aSTejun Heo */ 176e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17729c91e99STejun Heo 17829c91e99STejun Heo /* 17929c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18029c91e99STejun Heo * from get_work_pool(). 18129c91e99STejun Heo */ 18229c91e99STejun Heo struct rcu_head rcu; 1838b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1848b03ae3cSTejun Heo 1858b03ae3cSTejun Heo /* 186112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 187112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 188112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 189112202d9STejun Heo * number of flag bits. 1901da177e4SLinus Torvalds */ 191112202d9STejun Heo struct pool_workqueue { 192bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1934690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19473f53c4aSTejun Heo int work_color; /* L: current color */ 19573f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1968864b4e5STejun Heo int refcnt; /* L: reference count */ 19773f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19873f53c4aSTejun Heo /* L: nr of in_flight works */ 1991e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 200a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2011e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2023c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2032e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2048864b4e5STejun Heo 2058864b4e5STejun Heo /* 2068864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2078864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2088864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 209b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2108864b4e5STejun Heo */ 2118864b4e5STejun Heo struct work_struct unbound_release_work; 2128864b4e5STejun Heo struct rcu_head rcu; 213e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2141da177e4SLinus Torvalds 2151da177e4SLinus Torvalds /* 21673f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21773f53c4aSTejun Heo */ 21873f53c4aSTejun Heo struct wq_flusher { 2193c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2203c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22173f53c4aSTejun Heo struct completion done; /* flush completion */ 22273f53c4aSTejun Heo }; 2231da177e4SLinus Torvalds 224226223abSTejun Heo struct wq_device; 225226223abSTejun Heo 22673f53c4aSTejun Heo /* 227c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 228c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2291da177e4SLinus Torvalds */ 2301da177e4SLinus Torvalds struct workqueue_struct { 2313c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 23268e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23373f53c4aSTejun Heo 2343c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2353c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2363c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 237112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2383c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2393c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2403c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24173f53c4aSTejun Heo 2422e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 243e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 244e22bee78STejun Heo 24587fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 246a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 247226223abSTejun Heo 2486029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2494c16bd32STejun Heo struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ 2506029a918STejun Heo 251226223abSTejun Heo #ifdef CONFIG_SYSFS 252226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 253226223abSTejun Heo #endif 2544e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2554e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2564e6045f1SJohannes Berg #endif 257ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2582728fd2fSTejun Heo 2592728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2602728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2612728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 262df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2631da177e4SLinus Torvalds }; 2641da177e4SLinus Torvalds 265e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 266e904e6c2STejun Heo 267bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 268bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 269bce90380STejun Heo /* possible CPUs of each node */ 270bce90380STejun Heo 271bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 272bce90380STejun Heo 2734c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2744c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2754c16bd32STejun Heo 27668e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2772e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2785bcab335STejun Heo 27968e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 28068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2817d19c5ceSTejun Heo 2827d19c5ceSTejun Heo /* the per-cpu worker pools */ 2837d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2847d19c5ceSTejun Heo cpu_worker_pools); 2857d19c5ceSTejun Heo 28668e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2877d19c5ceSTejun Heo 28868e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 28929c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 29029c91e99STejun Heo 291c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 29229c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 29329c91e99STejun Heo 294d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 295d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 296044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 2971aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 298044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 299d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 300044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 301f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 302044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 30324d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 304d320c038STejun Heo 3057d19c5ceSTejun Heo static int worker_thread(void *__worker); 3067d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3077d19c5ceSTejun Heo const struct workqueue_attrs *from); 3087d19c5ceSTejun Heo 30997bd2347STejun Heo #define CREATE_TRACE_POINTS 31097bd2347STejun Heo #include <trace/events/workqueue.h> 31197bd2347STejun Heo 31268e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3135bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 31468e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 31568e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3165bcab335STejun Heo 317b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 31876af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 319b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 320b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 32176af4d93STejun Heo 322822d8405STejun Heo #ifdef CONFIG_LOCKDEP 323822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 324519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 325519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 326822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 327822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 328822d8405STejun Heo #else 329822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 330822d8405STejun Heo #endif 331822d8405STejun Heo 332f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 333f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 334f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3357a62c2c8STejun Heo (pool)++) 3364ce62e9eSTejun Heo 33749e3cf44STejun Heo /** 33817116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 33917116969STejun Heo * @pool: iteration cursor 340611c92a0STejun Heo * @pi: integer used for iteration 341fa1b54e6STejun Heo * 34268e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 34368e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 34468e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 345fa1b54e6STejun Heo * 346fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 347fa1b54e6STejun Heo * ignored. 34817116969STejun Heo */ 349611c92a0STejun Heo #define for_each_pool(pool, pi) \ 350611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 35168e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 352fa1b54e6STejun Heo else 35317116969STejun Heo 35417116969STejun Heo /** 355822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 356822d8405STejun Heo * @worker: iteration cursor 357822d8405STejun Heo * @wi: integer used for iteration 358822d8405STejun Heo * @pool: worker_pool to iterate workers of 359822d8405STejun Heo * 360822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 361822d8405STejun Heo * 362822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 363822d8405STejun Heo * ignored. 364822d8405STejun Heo */ 365822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 366822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 367822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 368822d8405STejun Heo else 369822d8405STejun Heo 370822d8405STejun Heo /** 37149e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 37249e3cf44STejun Heo * @pwq: iteration cursor 37349e3cf44STejun Heo * @wq: the target workqueue 37476af4d93STejun Heo * 375b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 376794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 377794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 37876af4d93STejun Heo * 37976af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 38076af4d93STejun Heo * ignored. 38149e3cf44STejun Heo */ 38249e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 38376af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 384b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 38576af4d93STejun Heo else 386f3421797STejun Heo 387dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 388dc186ad7SThomas Gleixner 389dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 390dc186ad7SThomas Gleixner 39199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 39299777288SStanislaw Gruszka { 39399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 39499777288SStanislaw Gruszka } 39599777288SStanislaw Gruszka 396dc186ad7SThomas Gleixner /* 397dc186ad7SThomas Gleixner * fixup_init is called when: 398dc186ad7SThomas Gleixner * - an active object is initialized 399dc186ad7SThomas Gleixner */ 400dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 401dc186ad7SThomas Gleixner { 402dc186ad7SThomas Gleixner struct work_struct *work = addr; 403dc186ad7SThomas Gleixner 404dc186ad7SThomas Gleixner switch (state) { 405dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 406dc186ad7SThomas Gleixner cancel_work_sync(work); 407dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 408dc186ad7SThomas Gleixner return 1; 409dc186ad7SThomas Gleixner default: 410dc186ad7SThomas Gleixner return 0; 411dc186ad7SThomas Gleixner } 412dc186ad7SThomas Gleixner } 413dc186ad7SThomas Gleixner 414dc186ad7SThomas Gleixner /* 415dc186ad7SThomas Gleixner * fixup_activate is called when: 416dc186ad7SThomas Gleixner * - an active object is activated 417dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 418dc186ad7SThomas Gleixner */ 419dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 420dc186ad7SThomas Gleixner { 421dc186ad7SThomas Gleixner struct work_struct *work = addr; 422dc186ad7SThomas Gleixner 423dc186ad7SThomas Gleixner switch (state) { 424dc186ad7SThomas Gleixner 425dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 426dc186ad7SThomas Gleixner /* 427dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 428dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 429dc186ad7SThomas Gleixner * is tracked in the object tracker. 430dc186ad7SThomas Gleixner */ 43122df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 432dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 433dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 434dc186ad7SThomas Gleixner return 0; 435dc186ad7SThomas Gleixner } 436dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 437dc186ad7SThomas Gleixner return 0; 438dc186ad7SThomas Gleixner 439dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 440dc186ad7SThomas Gleixner WARN_ON(1); 441dc186ad7SThomas Gleixner 442dc186ad7SThomas Gleixner default: 443dc186ad7SThomas Gleixner return 0; 444dc186ad7SThomas Gleixner } 445dc186ad7SThomas Gleixner } 446dc186ad7SThomas Gleixner 447dc186ad7SThomas Gleixner /* 448dc186ad7SThomas Gleixner * fixup_free is called when: 449dc186ad7SThomas Gleixner * - an active object is freed 450dc186ad7SThomas Gleixner */ 451dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 452dc186ad7SThomas Gleixner { 453dc186ad7SThomas Gleixner struct work_struct *work = addr; 454dc186ad7SThomas Gleixner 455dc186ad7SThomas Gleixner switch (state) { 456dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 457dc186ad7SThomas Gleixner cancel_work_sync(work); 458dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 459dc186ad7SThomas Gleixner return 1; 460dc186ad7SThomas Gleixner default: 461dc186ad7SThomas Gleixner return 0; 462dc186ad7SThomas Gleixner } 463dc186ad7SThomas Gleixner } 464dc186ad7SThomas Gleixner 465dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 466dc186ad7SThomas Gleixner .name = "work_struct", 46799777288SStanislaw Gruszka .debug_hint = work_debug_hint, 468dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 469dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 470dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 471dc186ad7SThomas Gleixner }; 472dc186ad7SThomas Gleixner 473dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 474dc186ad7SThomas Gleixner { 475dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 476dc186ad7SThomas Gleixner } 477dc186ad7SThomas Gleixner 478dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 479dc186ad7SThomas Gleixner { 480dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 481dc186ad7SThomas Gleixner } 482dc186ad7SThomas Gleixner 483dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 484dc186ad7SThomas Gleixner { 485dc186ad7SThomas Gleixner if (onstack) 486dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 487dc186ad7SThomas Gleixner else 488dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 489dc186ad7SThomas Gleixner } 490dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 491dc186ad7SThomas Gleixner 492dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 493dc186ad7SThomas Gleixner { 494dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 495dc186ad7SThomas Gleixner } 496dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 497dc186ad7SThomas Gleixner 498dc186ad7SThomas Gleixner #else 499dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 500dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 501dc186ad7SThomas Gleixner #endif 502dc186ad7SThomas Gleixner 5039daf9e67STejun Heo /* allocate ID and assign it to @pool */ 5049daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5059daf9e67STejun Heo { 5069daf9e67STejun Heo int ret; 5079daf9e67STejun Heo 50868e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5095bcab335STejun Heo 510fa1b54e6STejun Heo do { 511fa1b54e6STejun Heo if (!idr_pre_get(&worker_pool_idr, GFP_KERNEL)) 512fa1b54e6STejun Heo return -ENOMEM; 5139daf9e67STejun Heo ret = idr_get_new(&worker_pool_idr, pool, &pool->id); 514fa1b54e6STejun Heo } while (ret == -EAGAIN); 5159daf9e67STejun Heo 5169daf9e67STejun Heo return ret; 5179daf9e67STejun Heo } 5189daf9e67STejun Heo 51976af4d93STejun Heo /** 52076af4d93STejun Heo * first_pwq - return the first pool_workqueue of the specified workqueue 52176af4d93STejun Heo * @wq: the target workqueue 52276af4d93STejun Heo * 523b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 524794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 525794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 52676af4d93STejun Heo */ 5277fb98ea7STejun Heo static struct pool_workqueue *first_pwq(struct workqueue_struct *wq) 528a848e3b6SOleg Nesterov { 529b09f4fd3SLai Jiangshan assert_rcu_or_wq_mutex(wq); 53076af4d93STejun Heo return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue, 53176af4d93STejun Heo pwqs_node); 532f3421797STejun Heo } 533a848e3b6SOleg Nesterov 534df2d5ae4STejun Heo /** 535df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 536df2d5ae4STejun Heo * @wq: the target workqueue 537df2d5ae4STejun Heo * @node: the node ID 538df2d5ae4STejun Heo * 539df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 540df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 541df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 542df2d5ae4STejun Heo */ 543df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 544df2d5ae4STejun Heo int node) 545df2d5ae4STejun Heo { 546df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 547df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 548df2d5ae4STejun Heo } 549df2d5ae4STejun Heo 55073f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 55173f53c4aSTejun Heo { 55273f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 55373f53c4aSTejun Heo } 55473f53c4aSTejun Heo 55573f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 55673f53c4aSTejun Heo { 55773f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 55873f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 55973f53c4aSTejun Heo } 56073f53c4aSTejun Heo 56173f53c4aSTejun Heo static int work_next_color(int color) 56273f53c4aSTejun Heo { 56373f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 564b1f4ec17SOleg Nesterov } 565b1f4ec17SOleg Nesterov 5664594bf15SDavid Howells /* 567112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 568112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5697c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5707a22ad75STejun Heo * 571112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 572112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 573bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 574bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5757a22ad75STejun Heo * 576112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5777c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 578112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5797c3eed5cSTejun Heo * available only while the work item is queued. 580bbb68dfaSTejun Heo * 581bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 582bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 583bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 584bbb68dfaSTejun Heo * try to steal the PENDING bit. 5854594bf15SDavid Howells */ 5867a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5877a22ad75STejun Heo unsigned long flags) 5887a22ad75STejun Heo { 5896183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5907a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5917a22ad75STejun Heo } 5927a22ad75STejun Heo 593112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5944690c4abSTejun Heo unsigned long extra_flags) 595365970a1SDavid Howells { 596112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 597112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 598365970a1SDavid Howells } 599365970a1SDavid Howells 6004468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6014468a00fSLai Jiangshan int pool_id) 6024468a00fSLai Jiangshan { 6034468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6044468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6054468a00fSLai Jiangshan } 6064468a00fSLai Jiangshan 6077c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6087c3eed5cSTejun Heo int pool_id) 6094d707b9fSOleg Nesterov { 61023657bb1STejun Heo /* 61123657bb1STejun Heo * The following wmb is paired with the implied mb in 61223657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 61323657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 61423657bb1STejun Heo * owner. 61523657bb1STejun Heo */ 61623657bb1STejun Heo smp_wmb(); 6177c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6184d707b9fSOleg Nesterov } 6194d707b9fSOleg Nesterov 6207a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 621365970a1SDavid Howells { 6227c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6237c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6247a22ad75STejun Heo } 6257a22ad75STejun Heo 626112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6277a22ad75STejun Heo { 628e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6297a22ad75STejun Heo 630112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 631e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 632e120153dSTejun Heo else 633e120153dSTejun Heo return NULL; 6347a22ad75STejun Heo } 6357a22ad75STejun Heo 6367c3eed5cSTejun Heo /** 6377c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6387c3eed5cSTejun Heo * @work: the work item of interest 6397c3eed5cSTejun Heo * 6407c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 641fa1b54e6STejun Heo * 64268e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 64368e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 64468e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 645fa1b54e6STejun Heo * 646fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 647fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 648fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 649fa1b54e6STejun Heo * returned pool is and stays online. 6507c3eed5cSTejun Heo */ 6517c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6527a22ad75STejun Heo { 653e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6547c3eed5cSTejun Heo int pool_id; 6557a22ad75STejun Heo 65668e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 657fa1b54e6STejun Heo 658112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 659112202d9STejun Heo return ((struct pool_workqueue *) 6607c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6617a22ad75STejun Heo 6627c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6637c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6647a22ad75STejun Heo return NULL; 6657a22ad75STejun Heo 666fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6677c3eed5cSTejun Heo } 6687c3eed5cSTejun Heo 6697c3eed5cSTejun Heo /** 6707c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6717c3eed5cSTejun Heo * @work: the work item of interest 6727c3eed5cSTejun Heo * 6737c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6747c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6757c3eed5cSTejun Heo */ 6767c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6777c3eed5cSTejun Heo { 67854d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6797c3eed5cSTejun Heo 680112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 681112202d9STejun Heo return ((struct pool_workqueue *) 68254d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 68354d5b7d0SLai Jiangshan 68454d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6857c3eed5cSTejun Heo } 6867c3eed5cSTejun Heo 687bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 688bbb68dfaSTejun Heo { 6897c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 690bbb68dfaSTejun Heo 6917c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6927c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 693bbb68dfaSTejun Heo } 694bbb68dfaSTejun Heo 695bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 696bbb68dfaSTejun Heo { 697bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 698bbb68dfaSTejun Heo 699112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 700bbb68dfaSTejun Heo } 701bbb68dfaSTejun Heo 702e22bee78STejun Heo /* 7033270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7043270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 705d565ed63STejun Heo * they're being called with pool->lock held. 706e22bee78STejun Heo */ 707e22bee78STejun Heo 70863d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 709649027d7STejun Heo { 710e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 711649027d7STejun Heo } 712649027d7STejun Heo 713e22bee78STejun Heo /* 714e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 715e22bee78STejun Heo * running workers. 716974271c4STejun Heo * 717974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 718706026c2STejun Heo * function will always return %true for unbound pools as long as the 719974271c4STejun Heo * worklist isn't empty. 720e22bee78STejun Heo */ 72163d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 722e22bee78STejun Heo { 72363d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 724e22bee78STejun Heo } 725e22bee78STejun Heo 726e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 72763d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 728e22bee78STejun Heo { 72963d95a91STejun Heo return pool->nr_idle; 730e22bee78STejun Heo } 731e22bee78STejun Heo 732e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 73363d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 734e22bee78STejun Heo { 735e19e397aSTejun Heo return !list_empty(&pool->worklist) && 736e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 737e22bee78STejun Heo } 738e22bee78STejun Heo 739e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 74063d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 741e22bee78STejun Heo { 74263d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 743e22bee78STejun Heo } 744e22bee78STejun Heo 745e22bee78STejun Heo /* Do I need to be the manager? */ 74663d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 747e22bee78STejun Heo { 74863d95a91STejun Heo return need_to_create_worker(pool) || 74911ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 750e22bee78STejun Heo } 751e22bee78STejun Heo 752e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 75363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 754e22bee78STejun Heo { 75534a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 75663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 75763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 758e22bee78STejun Heo 759ea1abd61SLai Jiangshan /* 760ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 761ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 762ea1abd61SLai Jiangshan */ 763ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 764ea1abd61SLai Jiangshan return false; 765ea1abd61SLai Jiangshan 766e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 767e22bee78STejun Heo } 768e22bee78STejun Heo 769e22bee78STejun Heo /* 770e22bee78STejun Heo * Wake up functions. 771e22bee78STejun Heo */ 772e22bee78STejun Heo 7737e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 77463d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7757e11629dSTejun Heo { 77663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7777e11629dSTejun Heo return NULL; 7787e11629dSTejun Heo 77963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7807e11629dSTejun Heo } 7817e11629dSTejun Heo 7827e11629dSTejun Heo /** 7837e11629dSTejun Heo * wake_up_worker - wake up an idle worker 78463d95a91STejun Heo * @pool: worker pool to wake worker from 7857e11629dSTejun Heo * 78663d95a91STejun Heo * Wake up the first idle worker of @pool. 7877e11629dSTejun Heo * 7887e11629dSTejun Heo * CONTEXT: 789d565ed63STejun Heo * spin_lock_irq(pool->lock). 7907e11629dSTejun Heo */ 79163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7927e11629dSTejun Heo { 79363d95a91STejun Heo struct worker *worker = first_worker(pool); 7947e11629dSTejun Heo 7957e11629dSTejun Heo if (likely(worker)) 7967e11629dSTejun Heo wake_up_process(worker->task); 7977e11629dSTejun Heo } 7987e11629dSTejun Heo 7994690c4abSTejun Heo /** 800e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 801e22bee78STejun Heo * @task: task waking up 802e22bee78STejun Heo * @cpu: CPU @task is waking up to 803e22bee78STejun Heo * 804e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 805e22bee78STejun Heo * being awoken. 806e22bee78STejun Heo * 807e22bee78STejun Heo * CONTEXT: 808e22bee78STejun Heo * spin_lock_irq(rq->lock) 809e22bee78STejun Heo */ 810d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 811e22bee78STejun Heo { 812e22bee78STejun Heo struct worker *worker = kthread_data(task); 813e22bee78STejun Heo 81436576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 815ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 816e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 817e22bee78STejun Heo } 81836576000SJoonsoo Kim } 819e22bee78STejun Heo 820e22bee78STejun Heo /** 821e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 822e22bee78STejun Heo * @task: task going to sleep 823e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 824e22bee78STejun Heo * 825e22bee78STejun Heo * This function is called during schedule() when a busy worker is 826e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 827e22bee78STejun Heo * returning pointer to its task. 828e22bee78STejun Heo * 829e22bee78STejun Heo * CONTEXT: 830e22bee78STejun Heo * spin_lock_irq(rq->lock) 831e22bee78STejun Heo * 832e22bee78STejun Heo * RETURNS: 833e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 834e22bee78STejun Heo */ 835d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 836e22bee78STejun Heo { 837e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 838111c225aSTejun Heo struct worker_pool *pool; 839e22bee78STejun Heo 840111c225aSTejun Heo /* 841111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 842111c225aSTejun Heo * workers, also reach here, let's not access anything before 843111c225aSTejun Heo * checking NOT_RUNNING. 844111c225aSTejun Heo */ 8452d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 846e22bee78STejun Heo return NULL; 847e22bee78STejun Heo 848111c225aSTejun Heo pool = worker->pool; 849111c225aSTejun Heo 850e22bee78STejun Heo /* this can only happen on the local cpu */ 8516183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8526183c009STejun Heo return NULL; 853e22bee78STejun Heo 854e22bee78STejun Heo /* 855e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 856e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 857e22bee78STejun Heo * Please read comment there. 858e22bee78STejun Heo * 859628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 860628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 861628c78e7STejun Heo * disabled, which in turn means that none else could be 862d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 863628c78e7STejun Heo * lock is safe. 864e22bee78STejun Heo */ 865e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 866e19e397aSTejun Heo !list_empty(&pool->worklist)) 86763d95a91STejun Heo to_wakeup = first_worker(pool); 868e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 869e22bee78STejun Heo } 870e22bee78STejun Heo 871e22bee78STejun Heo /** 872e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 873cb444766STejun Heo * @worker: self 874d302f017STejun Heo * @flags: flags to set 875d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 876d302f017STejun Heo * 877e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 878e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 879e22bee78STejun Heo * woken up. 880d302f017STejun Heo * 881cb444766STejun Heo * CONTEXT: 882d565ed63STejun Heo * spin_lock_irq(pool->lock) 883d302f017STejun Heo */ 884d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 885d302f017STejun Heo bool wakeup) 886d302f017STejun Heo { 887bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 888e22bee78STejun Heo 889cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 890cb444766STejun Heo 891e22bee78STejun Heo /* 892e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 893e22bee78STejun Heo * wake up an idle worker as necessary if requested by 894e22bee78STejun Heo * @wakeup. 895e22bee78STejun Heo */ 896e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 897e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 898e22bee78STejun Heo if (wakeup) { 899e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 900bd7bdd43STejun Heo !list_empty(&pool->worklist)) 90163d95a91STejun Heo wake_up_worker(pool); 902e22bee78STejun Heo } else 903e19e397aSTejun Heo atomic_dec(&pool->nr_running); 904e22bee78STejun Heo } 905e22bee78STejun Heo 906d302f017STejun Heo worker->flags |= flags; 907d302f017STejun Heo } 908d302f017STejun Heo 909d302f017STejun Heo /** 910e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 911cb444766STejun Heo * @worker: self 912d302f017STejun Heo * @flags: flags to clear 913d302f017STejun Heo * 914e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 915d302f017STejun Heo * 916cb444766STejun Heo * CONTEXT: 917d565ed63STejun Heo * spin_lock_irq(pool->lock) 918d302f017STejun Heo */ 919d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 920d302f017STejun Heo { 92163d95a91STejun Heo struct worker_pool *pool = worker->pool; 922e22bee78STejun Heo unsigned int oflags = worker->flags; 923e22bee78STejun Heo 924cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 925cb444766STejun Heo 926d302f017STejun Heo worker->flags &= ~flags; 927e22bee78STejun Heo 92842c025f3STejun Heo /* 92942c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 93042c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 93142c025f3STejun Heo * of multiple flags, not a single flag. 93242c025f3STejun Heo */ 933e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 934e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 935e19e397aSTejun Heo atomic_inc(&pool->nr_running); 936d302f017STejun Heo } 937d302f017STejun Heo 938d302f017STejun Heo /** 9398cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 940c9e7cf27STejun Heo * @pool: pool of interest 9418cca0eeaSTejun Heo * @work: work to find worker for 9428cca0eeaSTejun Heo * 943c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 944c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 945a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 946a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 947a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 948a2c1c57bSTejun Heo * being executed. 949a2c1c57bSTejun Heo * 950a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 951a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 952a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 953a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 954a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 955a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 956a2c1c57bSTejun Heo * 957c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 958c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 959c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 960c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 961c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 962c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9638cca0eeaSTejun Heo * 9648cca0eeaSTejun Heo * CONTEXT: 965d565ed63STejun Heo * spin_lock_irq(pool->lock). 9668cca0eeaSTejun Heo * 9678cca0eeaSTejun Heo * RETURNS: 9688cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9698cca0eeaSTejun Heo * otherwise. 9708cca0eeaSTejun Heo */ 971c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9728cca0eeaSTejun Heo struct work_struct *work) 9738cca0eeaSTejun Heo { 97442f8570fSSasha Levin struct worker *worker; 97542f8570fSSasha Levin 976b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 977a2c1c57bSTejun Heo (unsigned long)work) 978a2c1c57bSTejun Heo if (worker->current_work == work && 979a2c1c57bSTejun Heo worker->current_func == work->func) 98042f8570fSSasha Levin return worker; 98142f8570fSSasha Levin 98242f8570fSSasha Levin return NULL; 9838cca0eeaSTejun Heo } 9848cca0eeaSTejun Heo 9858cca0eeaSTejun Heo /** 986bf4ede01STejun Heo * move_linked_works - move linked works to a list 987bf4ede01STejun Heo * @work: start of series of works to be scheduled 988bf4ede01STejun Heo * @head: target list to append @work to 989bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 990bf4ede01STejun Heo * 991bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 992bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 993bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 994bf4ede01STejun Heo * 995bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 996bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 997bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 998bf4ede01STejun Heo * 999bf4ede01STejun Heo * CONTEXT: 1000d565ed63STejun Heo * spin_lock_irq(pool->lock). 1001bf4ede01STejun Heo */ 1002bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1003bf4ede01STejun Heo struct work_struct **nextp) 1004bf4ede01STejun Heo { 1005bf4ede01STejun Heo struct work_struct *n; 1006bf4ede01STejun Heo 1007bf4ede01STejun Heo /* 1008bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1009bf4ede01STejun Heo * use NULL for list head. 1010bf4ede01STejun Heo */ 1011bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1012bf4ede01STejun Heo list_move_tail(&work->entry, head); 1013bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1014bf4ede01STejun Heo break; 1015bf4ede01STejun Heo } 1016bf4ede01STejun Heo 1017bf4ede01STejun Heo /* 1018bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1019bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1020bf4ede01STejun Heo * needs to be updated. 1021bf4ede01STejun Heo */ 1022bf4ede01STejun Heo if (nextp) 1023bf4ede01STejun Heo *nextp = n; 1024bf4ede01STejun Heo } 1025bf4ede01STejun Heo 10268864b4e5STejun Heo /** 10278864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10288864b4e5STejun Heo * @pwq: pool_workqueue to get 10298864b4e5STejun Heo * 10308864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10318864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10328864b4e5STejun Heo */ 10338864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10348864b4e5STejun Heo { 10358864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10368864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10378864b4e5STejun Heo pwq->refcnt++; 10388864b4e5STejun Heo } 10398864b4e5STejun Heo 10408864b4e5STejun Heo /** 10418864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10428864b4e5STejun Heo * @pwq: pool_workqueue to put 10438864b4e5STejun Heo * 10448864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10458864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10468864b4e5STejun Heo */ 10478864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10488864b4e5STejun Heo { 10498864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10508864b4e5STejun Heo if (likely(--pwq->refcnt)) 10518864b4e5STejun Heo return; 10528864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10538864b4e5STejun Heo return; 10548864b4e5STejun Heo /* 10558864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10568864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10578864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10588864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10598864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10608864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10618864b4e5STejun Heo */ 10628864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10638864b4e5STejun Heo } 10648864b4e5STejun Heo 1065dce90d47STejun Heo /** 1066dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1067dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1068dce90d47STejun Heo * 1069dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1070dce90d47STejun Heo */ 1071dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1072dce90d47STejun Heo { 1073dce90d47STejun Heo if (pwq) { 1074dce90d47STejun Heo /* 1075dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1076dce90d47STejun Heo * following lock operations are safe. 1077dce90d47STejun Heo */ 1078dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1079dce90d47STejun Heo put_pwq(pwq); 1080dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1081dce90d47STejun Heo } 1082dce90d47STejun Heo } 1083dce90d47STejun Heo 1084112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1085bf4ede01STejun Heo { 1086112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1087bf4ede01STejun Heo 1088bf4ede01STejun Heo trace_workqueue_activate_work(work); 1089112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1090bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1091112202d9STejun Heo pwq->nr_active++; 1092bf4ede01STejun Heo } 1093bf4ede01STejun Heo 1094112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10953aa62497SLai Jiangshan { 1096112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10973aa62497SLai Jiangshan struct work_struct, entry); 10983aa62497SLai Jiangshan 1099112202d9STejun Heo pwq_activate_delayed_work(work); 11003aa62497SLai Jiangshan } 11013aa62497SLai Jiangshan 1102bf4ede01STejun Heo /** 1103112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1104112202d9STejun Heo * @pwq: pwq of interest 1105bf4ede01STejun Heo * @color: color of work which left the queue 1106bf4ede01STejun Heo * 1107bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1108112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1109bf4ede01STejun Heo * 1110bf4ede01STejun Heo * CONTEXT: 1111d565ed63STejun Heo * spin_lock_irq(pool->lock). 1112bf4ede01STejun Heo */ 1113112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1114bf4ede01STejun Heo { 11158864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1116bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11178864b4e5STejun Heo goto out_put; 1118bf4ede01STejun Heo 1119112202d9STejun Heo pwq->nr_in_flight[color]--; 1120bf4ede01STejun Heo 1121112202d9STejun Heo pwq->nr_active--; 1122112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1123bf4ede01STejun Heo /* one down, submit a delayed one */ 1124112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1125112202d9STejun Heo pwq_activate_first_delayed(pwq); 1126bf4ede01STejun Heo } 1127bf4ede01STejun Heo 1128bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1129112202d9STejun Heo if (likely(pwq->flush_color != color)) 11308864b4e5STejun Heo goto out_put; 1131bf4ede01STejun Heo 1132bf4ede01STejun Heo /* are there still in-flight works? */ 1133112202d9STejun Heo if (pwq->nr_in_flight[color]) 11348864b4e5STejun Heo goto out_put; 1135bf4ede01STejun Heo 1136112202d9STejun Heo /* this pwq is done, clear flush_color */ 1137112202d9STejun Heo pwq->flush_color = -1; 1138bf4ede01STejun Heo 1139bf4ede01STejun Heo /* 1140112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1141bf4ede01STejun Heo * will handle the rest. 1142bf4ede01STejun Heo */ 1143112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1144112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11458864b4e5STejun Heo out_put: 11468864b4e5STejun Heo put_pwq(pwq); 1147bf4ede01STejun Heo } 1148bf4ede01STejun Heo 114936e227d2STejun Heo /** 1150bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 115136e227d2STejun Heo * @work: work item to steal 115236e227d2STejun Heo * @is_dwork: @work is a delayed_work 1153bbb68dfaSTejun Heo * @flags: place to store irq state 115436e227d2STejun Heo * 115536e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 115636e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 115736e227d2STejun Heo * 115836e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 115936e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 116036e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1161bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1162bbb68dfaSTejun Heo * for arbitrarily long 116336e227d2STejun Heo * 1164bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1165e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1166e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1167e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1168bbb68dfaSTejun Heo * 1169bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1170bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1171bbb68dfaSTejun Heo * 1172e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1173bf4ede01STejun Heo */ 1174bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1175bbb68dfaSTejun Heo unsigned long *flags) 1176bf4ede01STejun Heo { 1177d565ed63STejun Heo struct worker_pool *pool; 1178112202d9STejun Heo struct pool_workqueue *pwq; 1179bf4ede01STejun Heo 1180bbb68dfaSTejun Heo local_irq_save(*flags); 1181bbb68dfaSTejun Heo 118236e227d2STejun Heo /* try to steal the timer if it exists */ 118336e227d2STejun Heo if (is_dwork) { 118436e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 118536e227d2STejun Heo 1186e0aecdd8STejun Heo /* 1187e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1188e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1189e0aecdd8STejun Heo * running on the local CPU. 1190e0aecdd8STejun Heo */ 119136e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 119236e227d2STejun Heo return 1; 119336e227d2STejun Heo } 119436e227d2STejun Heo 119536e227d2STejun Heo /* try to claim PENDING the normal way */ 1196bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1197bf4ede01STejun Heo return 0; 1198bf4ede01STejun Heo 1199bf4ede01STejun Heo /* 1200bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1201bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1202bf4ede01STejun Heo */ 1203d565ed63STejun Heo pool = get_work_pool(work); 1204d565ed63STejun Heo if (!pool) 1205bbb68dfaSTejun Heo goto fail; 1206bf4ede01STejun Heo 1207d565ed63STejun Heo spin_lock(&pool->lock); 1208bf4ede01STejun Heo /* 1209112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1210112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1211112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1212112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1213112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12140b3dae68SLai Jiangshan * item is currently queued on that pool. 1215bf4ede01STejun Heo */ 1216112202d9STejun Heo pwq = get_work_pwq(work); 1217112202d9STejun Heo if (pwq && pwq->pool == pool) { 1218bf4ede01STejun Heo debug_work_deactivate(work); 12193aa62497SLai Jiangshan 12203aa62497SLai Jiangshan /* 122116062836STejun Heo * A delayed work item cannot be grabbed directly because 122216062836STejun Heo * it might have linked NO_COLOR work items which, if left 1223112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 122416062836STejun Heo * management later on and cause stall. Make sure the work 122516062836STejun Heo * item is activated before grabbing. 12263aa62497SLai Jiangshan */ 12273aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1228112202d9STejun Heo pwq_activate_delayed_work(work); 12293aa62497SLai Jiangshan 1230bf4ede01STejun Heo list_del_init(&work->entry); 1231112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 123236e227d2STejun Heo 1233112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12344468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12354468a00fSLai Jiangshan 1236d565ed63STejun Heo spin_unlock(&pool->lock); 123736e227d2STejun Heo return 1; 1238bf4ede01STejun Heo } 1239d565ed63STejun Heo spin_unlock(&pool->lock); 1240bbb68dfaSTejun Heo fail: 1241bbb68dfaSTejun Heo local_irq_restore(*flags); 1242bbb68dfaSTejun Heo if (work_is_canceling(work)) 1243bbb68dfaSTejun Heo return -ENOENT; 1244bbb68dfaSTejun Heo cpu_relax(); 124536e227d2STejun Heo return -EAGAIN; 1246bf4ede01STejun Heo } 1247bf4ede01STejun Heo 1248bf4ede01STejun Heo /** 1249706026c2STejun Heo * insert_work - insert a work into a pool 1250112202d9STejun Heo * @pwq: pwq @work belongs to 12514690c4abSTejun Heo * @work: work to insert 12524690c4abSTejun Heo * @head: insertion point 12534690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12544690c4abSTejun Heo * 1255112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1256706026c2STejun Heo * work_struct flags. 12574690c4abSTejun Heo * 12584690c4abSTejun Heo * CONTEXT: 1259d565ed63STejun Heo * spin_lock_irq(pool->lock). 1260365970a1SDavid Howells */ 1261112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1262112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1263b89deed3SOleg Nesterov { 1264112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1265e1d8aa9fSFrederic Weisbecker 12664690c4abSTejun Heo /* we own @work, set data and link */ 1267112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12681a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12698864b4e5STejun Heo get_pwq(pwq); 1270e22bee78STejun Heo 1271e22bee78STejun Heo /* 1272c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1273c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1274c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1275e22bee78STejun Heo */ 1276e22bee78STejun Heo smp_mb(); 1277e22bee78STejun Heo 127863d95a91STejun Heo if (__need_more_worker(pool)) 127963d95a91STejun Heo wake_up_worker(pool); 1280b89deed3SOleg Nesterov } 1281b89deed3SOleg Nesterov 1282c8efcc25STejun Heo /* 1283c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12848d03ecfeSTejun Heo * same workqueue. 1285c8efcc25STejun Heo */ 1286c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1287c8efcc25STejun Heo { 1288c8efcc25STejun Heo struct worker *worker; 1289c8efcc25STejun Heo 12908d03ecfeSTejun Heo worker = current_wq_worker(); 1291c8efcc25STejun Heo /* 12928d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12938d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1294c8efcc25STejun Heo */ 1295112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1296c8efcc25STejun Heo } 1297c8efcc25STejun Heo 1298d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12991da177e4SLinus Torvalds struct work_struct *work) 13001da177e4SLinus Torvalds { 1301112202d9STejun Heo struct pool_workqueue *pwq; 1302c9178087STejun Heo struct worker_pool *last_pool; 13031e19ffc6STejun Heo struct list_head *worklist; 13048a2e8e5dSTejun Heo unsigned int work_flags; 1305b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13068930cabaSTejun Heo 13078930cabaSTejun Heo /* 13088930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13098930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13108930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13118930cabaSTejun Heo * happen with IRQ disabled. 13128930cabaSTejun Heo */ 13138930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13141da177e4SLinus Torvalds 1315dc186ad7SThomas Gleixner debug_work_activate(work); 13161e19ffc6STejun Heo 1317c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1318618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1319c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1320e41e704bSTejun Heo return; 13219e8cd2f5STejun Heo retry: 1322df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1323f3421797STejun Heo cpu = raw_smp_processor_id(); 1324df2d5ae4STejun Heo 1325df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1326df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1327c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1328df2d5ae4STejun Heo else 1329df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1330f3421797STejun Heo 133118aa9effSTejun Heo /* 1332c9178087STejun Heo * If @work was previously on a different pool, it might still be 1333c9178087STejun Heo * running there, in which case the work needs to be queued on that 1334c9178087STejun Heo * pool to guarantee non-reentrancy. 133518aa9effSTejun Heo */ 1336c9e7cf27STejun Heo last_pool = get_work_pool(work); 1337112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 133818aa9effSTejun Heo struct worker *worker; 133918aa9effSTejun Heo 1340d565ed63STejun Heo spin_lock(&last_pool->lock); 134118aa9effSTejun Heo 1342c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 134318aa9effSTejun Heo 1344112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1345c9178087STejun Heo pwq = worker->current_pwq; 13468594fadeSLai Jiangshan } else { 134718aa9effSTejun Heo /* meh... not running there, queue here */ 1348d565ed63STejun Heo spin_unlock(&last_pool->lock); 1349112202d9STejun Heo spin_lock(&pwq->pool->lock); 135018aa9effSTejun Heo } 13518930cabaSTejun Heo } else { 1352112202d9STejun Heo spin_lock(&pwq->pool->lock); 13538930cabaSTejun Heo } 1354502ca9d8STejun Heo 13559e8cd2f5STejun Heo /* 13569e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13579e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13589e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1359df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1360df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13619e8cd2f5STejun Heo * make forward-progress. 13629e8cd2f5STejun Heo */ 13639e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13649e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13659e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13669e8cd2f5STejun Heo cpu_relax(); 13679e8cd2f5STejun Heo goto retry; 13689e8cd2f5STejun Heo } 13699e8cd2f5STejun Heo /* oops */ 13709e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13719e8cd2f5STejun Heo wq->name, cpu); 13729e8cd2f5STejun Heo } 13739e8cd2f5STejun Heo 1374112202d9STejun Heo /* pwq determined, queue */ 1375112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1376502ca9d8STejun Heo 1377f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1378112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1379f5b2552bSDan Carpenter return; 1380f5b2552bSDan Carpenter } 13811e19ffc6STejun Heo 1382112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1383112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13841e19ffc6STejun Heo 1385112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1386cdadf009STejun Heo trace_workqueue_activate_work(work); 1387112202d9STejun Heo pwq->nr_active++; 1388112202d9STejun Heo worklist = &pwq->pool->worklist; 13898a2e8e5dSTejun Heo } else { 13908a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1391112202d9STejun Heo worklist = &pwq->delayed_works; 13928a2e8e5dSTejun Heo } 13931e19ffc6STejun Heo 1394112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13951e19ffc6STejun Heo 1396112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13971da177e4SLinus Torvalds } 13981da177e4SLinus Torvalds 13990fcb78c2SRolf Eike Beer /** 1400c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1401c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1402c1a220e7SZhang Rui * @wq: workqueue to use 1403c1a220e7SZhang Rui * @work: work to queue 1404c1a220e7SZhang Rui * 1405d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1406c1a220e7SZhang Rui * 1407c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1408c1a220e7SZhang Rui * can't go away. 1409c1a220e7SZhang Rui */ 1410d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1411d4283e93STejun Heo struct work_struct *work) 1412c1a220e7SZhang Rui { 1413d4283e93STejun Heo bool ret = false; 14148930cabaSTejun Heo unsigned long flags; 14158930cabaSTejun Heo 14168930cabaSTejun Heo local_irq_save(flags); 1417c1a220e7SZhang Rui 141822df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14194690c4abSTejun Heo __queue_work(cpu, wq, work); 1420d4283e93STejun Heo ret = true; 1421c1a220e7SZhang Rui } 14228930cabaSTejun Heo 14238930cabaSTejun Heo local_irq_restore(flags); 1424c1a220e7SZhang Rui return ret; 1425c1a220e7SZhang Rui } 1426c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1427c1a220e7SZhang Rui 1428d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14291da177e4SLinus Torvalds { 143052bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14311da177e4SLinus Torvalds 1432e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 143360c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14341da177e4SLinus Torvalds } 14351438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14361da177e4SLinus Torvalds 14377beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 143852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14391da177e4SLinus Torvalds { 14407beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14417beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14421da177e4SLinus Torvalds 14437beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14447beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1445fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1446fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14477beb2edfSTejun Heo 14488852aac2STejun Heo /* 14498852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14508852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14518852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14528852aac2STejun Heo * on that there's no such delay when @delay is 0. 14538852aac2STejun Heo */ 14548852aac2STejun Heo if (!delay) { 14558852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14568852aac2STejun Heo return; 14578852aac2STejun Heo } 14588852aac2STejun Heo 14597beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14607beb2edfSTejun Heo 146160c057bcSLai Jiangshan dwork->wq = wq; 14621265057fSTejun Heo dwork->cpu = cpu; 14637beb2edfSTejun Heo timer->expires = jiffies + delay; 14647beb2edfSTejun Heo 14657beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14667beb2edfSTejun Heo add_timer_on(timer, cpu); 14677beb2edfSTejun Heo else 14687beb2edfSTejun Heo add_timer(timer); 14697beb2edfSTejun Heo } 14701da177e4SLinus Torvalds 14710fcb78c2SRolf Eike Beer /** 14720fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14730fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14740fcb78c2SRolf Eike Beer * @wq: workqueue to use 1475af9997e4SRandy Dunlap * @dwork: work to queue 14760fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14770fcb78c2SRolf Eike Beer * 1478715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1479715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1480715f1300STejun Heo * execution. 14810fcb78c2SRolf Eike Beer */ 1482d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 148352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14847a6bc1cdSVenkatesh Pallipadi { 148552bad64dSDavid Howells struct work_struct *work = &dwork->work; 1486d4283e93STejun Heo bool ret = false; 14878930cabaSTejun Heo unsigned long flags; 14888930cabaSTejun Heo 14898930cabaSTejun Heo /* read the comment in __queue_work() */ 14908930cabaSTejun Heo local_irq_save(flags); 14917a6bc1cdSVenkatesh Pallipadi 149222df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14937beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1494d4283e93STejun Heo ret = true; 14957a6bc1cdSVenkatesh Pallipadi } 14968930cabaSTejun Heo 14978930cabaSTejun Heo local_irq_restore(flags); 14987a6bc1cdSVenkatesh Pallipadi return ret; 14997a6bc1cdSVenkatesh Pallipadi } 1500ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 15011da177e4SLinus Torvalds 1502c8e55f36STejun Heo /** 15038376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15048376fe22STejun Heo * @cpu: CPU number to execute work on 15058376fe22STejun Heo * @wq: workqueue to use 15068376fe22STejun Heo * @dwork: work to queue 15078376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15088376fe22STejun Heo * 15098376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15108376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15118376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15128376fe22STejun Heo * current state. 15138376fe22STejun Heo * 15148376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 15158376fe22STejun Heo * pending and its timer was modified. 15168376fe22STejun Heo * 1517e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15188376fe22STejun Heo * See try_to_grab_pending() for details. 15198376fe22STejun Heo */ 15208376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15218376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15228376fe22STejun Heo { 15238376fe22STejun Heo unsigned long flags; 15248376fe22STejun Heo int ret; 15258376fe22STejun Heo 15268376fe22STejun Heo do { 15278376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15288376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15298376fe22STejun Heo 15308376fe22STejun Heo if (likely(ret >= 0)) { 15318376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15328376fe22STejun Heo local_irq_restore(flags); 15338376fe22STejun Heo } 15348376fe22STejun Heo 15358376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15368376fe22STejun Heo return ret; 15378376fe22STejun Heo } 15388376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15398376fe22STejun Heo 15408376fe22STejun Heo /** 1541c8e55f36STejun Heo * worker_enter_idle - enter idle state 1542c8e55f36STejun Heo * @worker: worker which is entering idle state 1543c8e55f36STejun Heo * 1544c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1545c8e55f36STejun Heo * necessary. 1546c8e55f36STejun Heo * 1547c8e55f36STejun Heo * LOCKING: 1548d565ed63STejun Heo * spin_lock_irq(pool->lock). 1549c8e55f36STejun Heo */ 1550c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15511da177e4SLinus Torvalds { 1552bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1553c8e55f36STejun Heo 15546183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15556183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15566183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15576183c009STejun Heo return; 1558c8e55f36STejun Heo 1559cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1560cb444766STejun Heo worker->flags |= WORKER_IDLE; 1561bd7bdd43STejun Heo pool->nr_idle++; 1562e22bee78STejun Heo worker->last_active = jiffies; 1563c8e55f36STejun Heo 1564c8e55f36STejun Heo /* idle_list is LIFO */ 1565bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1566db7bccf4STejun Heo 156763d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1568628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1569cb444766STejun Heo 1570544ecf31STejun Heo /* 1571706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1572d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1573628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1574628c78e7STejun Heo * unbind is not in progress. 1575544ecf31STejun Heo */ 157624647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1577bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1578e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1579c8e55f36STejun Heo } 1580c8e55f36STejun Heo 1581c8e55f36STejun Heo /** 1582c8e55f36STejun Heo * worker_leave_idle - leave idle state 1583c8e55f36STejun Heo * @worker: worker which is leaving idle state 1584c8e55f36STejun Heo * 1585c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1586c8e55f36STejun Heo * 1587c8e55f36STejun Heo * LOCKING: 1588d565ed63STejun Heo * spin_lock_irq(pool->lock). 1589c8e55f36STejun Heo */ 1590c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1591c8e55f36STejun Heo { 1592bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1593c8e55f36STejun Heo 15946183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15956183c009STejun Heo return; 1596d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1597bd7bdd43STejun Heo pool->nr_idle--; 1598c8e55f36STejun Heo list_del_init(&worker->entry); 1599c8e55f36STejun Heo } 1600c8e55f36STejun Heo 1601e22bee78STejun Heo /** 1602f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1603f36dc67bSLai Jiangshan * @pool: target worker_pool 1604f36dc67bSLai Jiangshan * 1605f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1606e22bee78STejun Heo * 1607e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1608e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1609e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1610e22bee78STejun Heo * guaranteed to execute on the cpu. 1611e22bee78STejun Heo * 1612f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1613e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1614e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1615e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1616706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1617e22bee78STejun Heo * [dis]associated in the meantime. 1618e22bee78STejun Heo * 1619706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 162024647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1621f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1622f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1623f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1624e22bee78STejun Heo * 1625e22bee78STejun Heo * CONTEXT: 1626d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1627e22bee78STejun Heo * held. 1628e22bee78STejun Heo * 1629e22bee78STejun Heo * RETURNS: 1630706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1631e22bee78STejun Heo * bound), %false if offline. 1632e22bee78STejun Heo */ 1633f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1634d565ed63STejun Heo __acquires(&pool->lock) 1635e22bee78STejun Heo { 1636e22bee78STejun Heo while (true) { 1637e22bee78STejun Heo /* 1638e22bee78STejun Heo * The following call may fail, succeed or succeed 1639e22bee78STejun Heo * without actually migrating the task to the cpu if 1640e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 164124647570STejun Heo * against POOL_DISASSOCIATED. 1642e22bee78STejun Heo */ 164324647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16447a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1645e22bee78STejun Heo 1646d565ed63STejun Heo spin_lock_irq(&pool->lock); 164724647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1648e22bee78STejun Heo return false; 1649f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16507a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1651e22bee78STejun Heo return true; 1652d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1653e22bee78STejun Heo 16545035b20fSTejun Heo /* 16555035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16565035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16575035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16585035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16595035b20fSTejun Heo */ 1660e22bee78STejun Heo cpu_relax(); 16615035b20fSTejun Heo cond_resched(); 1662e22bee78STejun Heo } 1663e22bee78STejun Heo } 1664e22bee78STejun Heo 1665c34056a3STejun Heo static struct worker *alloc_worker(void) 1666c34056a3STejun Heo { 1667c34056a3STejun Heo struct worker *worker; 1668c34056a3STejun Heo 1669c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1670c8e55f36STejun Heo if (worker) { 1671c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1672affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1673e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1674e22bee78STejun Heo worker->flags = WORKER_PREP; 1675c8e55f36STejun Heo } 1676c34056a3STejun Heo return worker; 1677c34056a3STejun Heo } 1678c34056a3STejun Heo 1679c34056a3STejun Heo /** 1680c34056a3STejun Heo * create_worker - create a new workqueue worker 168163d95a91STejun Heo * @pool: pool the new worker will belong to 1682c34056a3STejun Heo * 168363d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1684c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1685c34056a3STejun Heo * destroy_worker(). 1686c34056a3STejun Heo * 1687c34056a3STejun Heo * CONTEXT: 1688c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1689c34056a3STejun Heo * 1690c34056a3STejun Heo * RETURNS: 1691c34056a3STejun Heo * Pointer to the newly created worker. 1692c34056a3STejun Heo */ 1693bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1694c34056a3STejun Heo { 1695c34056a3STejun Heo struct worker *worker = NULL; 1696f3421797STejun Heo int id = -1; 1697e3c916a4STejun Heo char id_buf[16]; 1698c34056a3STejun Heo 1699cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1700cd549687STejun Heo 1701822d8405STejun Heo /* 1702822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1703822d8405STejun Heo * without installing the pointer. 1704822d8405STejun Heo */ 1705822d8405STejun Heo idr_preload(GFP_KERNEL); 1706d565ed63STejun Heo spin_lock_irq(&pool->lock); 1707822d8405STejun Heo 1708822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1709822d8405STejun Heo 1710d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1711822d8405STejun Heo idr_preload_end(); 1712822d8405STejun Heo if (id < 0) 1713c34056a3STejun Heo goto fail; 1714c34056a3STejun Heo 1715c34056a3STejun Heo worker = alloc_worker(); 1716c34056a3STejun Heo if (!worker) 1717c34056a3STejun Heo goto fail; 1718c34056a3STejun Heo 1719bd7bdd43STejun Heo worker->pool = pool; 1720c34056a3STejun Heo worker->id = id; 1721c34056a3STejun Heo 172229c91e99STejun Heo if (pool->cpu >= 0) 1723e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1724e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1725f3421797STejun Heo else 1726e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1727e3c916a4STejun Heo 1728f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1729e3c916a4STejun Heo "kworker/%s", id_buf); 1730c34056a3STejun Heo if (IS_ERR(worker->task)) 1731c34056a3STejun Heo goto fail; 1732c34056a3STejun Heo 1733c5aa87bbSTejun Heo /* 1734c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1735c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1736c5aa87bbSTejun Heo */ 17377a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 17387a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17393270476aSTejun Heo 174014a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 174114a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 17427a4e344cSTejun Heo 17437a4e344cSTejun Heo /* 17447a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17457a4e344cSTejun Heo * remains stable across this function. See the comments above the 17467a4e344cSTejun Heo * flag definition for details. 17477a4e344cSTejun Heo */ 17487a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1749f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1750c34056a3STejun Heo 1751822d8405STejun Heo /* successful, commit the pointer to idr */ 1752822d8405STejun Heo spin_lock_irq(&pool->lock); 1753822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1754822d8405STejun Heo spin_unlock_irq(&pool->lock); 1755822d8405STejun Heo 1756c34056a3STejun Heo return worker; 1757822d8405STejun Heo 1758c34056a3STejun Heo fail: 1759c34056a3STejun Heo if (id >= 0) { 1760d565ed63STejun Heo spin_lock_irq(&pool->lock); 1761822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1762d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1763c34056a3STejun Heo } 1764c34056a3STejun Heo kfree(worker); 1765c34056a3STejun Heo return NULL; 1766c34056a3STejun Heo } 1767c34056a3STejun Heo 1768c34056a3STejun Heo /** 1769c34056a3STejun Heo * start_worker - start a newly created worker 1770c34056a3STejun Heo * @worker: worker to start 1771c34056a3STejun Heo * 1772706026c2STejun Heo * Make the pool aware of @worker and start it. 1773c34056a3STejun Heo * 1774c34056a3STejun Heo * CONTEXT: 1775d565ed63STejun Heo * spin_lock_irq(pool->lock). 1776c34056a3STejun Heo */ 1777c34056a3STejun Heo static void start_worker(struct worker *worker) 1778c34056a3STejun Heo { 1779cb444766STejun Heo worker->flags |= WORKER_STARTED; 1780bd7bdd43STejun Heo worker->pool->nr_workers++; 1781c8e55f36STejun Heo worker_enter_idle(worker); 1782c34056a3STejun Heo wake_up_process(worker->task); 1783c34056a3STejun Heo } 1784c34056a3STejun Heo 1785c34056a3STejun Heo /** 1786ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1787ebf44d16STejun Heo * @pool: the target pool 1788ebf44d16STejun Heo * 1789cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1790ebf44d16STejun Heo */ 1791ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1792ebf44d16STejun Heo { 1793ebf44d16STejun Heo struct worker *worker; 1794ebf44d16STejun Heo 1795cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1796cd549687STejun Heo 1797ebf44d16STejun Heo worker = create_worker(pool); 1798ebf44d16STejun Heo if (worker) { 1799ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1800ebf44d16STejun Heo start_worker(worker); 1801ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1802ebf44d16STejun Heo } 1803ebf44d16STejun Heo 1804cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1805cd549687STejun Heo 1806ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1807ebf44d16STejun Heo } 1808ebf44d16STejun Heo 1809ebf44d16STejun Heo /** 1810c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1811c34056a3STejun Heo * @worker: worker to be destroyed 1812c34056a3STejun Heo * 1813706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1814c8e55f36STejun Heo * 1815c8e55f36STejun Heo * CONTEXT: 1816d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1817c34056a3STejun Heo */ 1818c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1819c34056a3STejun Heo { 1820bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1821c34056a3STejun Heo 1822cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1823cd549687STejun Heo lockdep_assert_held(&pool->lock); 1824cd549687STejun Heo 1825c34056a3STejun Heo /* sanity check frenzy */ 18266183c009STejun Heo if (WARN_ON(worker->current_work) || 18276183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18286183c009STejun Heo return; 1829c34056a3STejun Heo 1830c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1831bd7bdd43STejun Heo pool->nr_workers--; 1832c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1833bd7bdd43STejun Heo pool->nr_idle--; 1834c8e55f36STejun Heo 1835c8e55f36STejun Heo list_del_init(&worker->entry); 1836cb444766STejun Heo worker->flags |= WORKER_DIE; 1837c8e55f36STejun Heo 1838822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1839822d8405STejun Heo 1840d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1841c8e55f36STejun Heo 1842c34056a3STejun Heo kthread_stop(worker->task); 1843c34056a3STejun Heo kfree(worker); 1844c34056a3STejun Heo 1845d565ed63STejun Heo spin_lock_irq(&pool->lock); 1846c34056a3STejun Heo } 1847c34056a3STejun Heo 184863d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1849e22bee78STejun Heo { 185063d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1851e22bee78STejun Heo 1852d565ed63STejun Heo spin_lock_irq(&pool->lock); 1853e22bee78STejun Heo 185463d95a91STejun Heo if (too_many_workers(pool)) { 1855e22bee78STejun Heo struct worker *worker; 1856e22bee78STejun Heo unsigned long expires; 1857e22bee78STejun Heo 1858e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 185963d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1860e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1861e22bee78STejun Heo 1862e22bee78STejun Heo if (time_before(jiffies, expires)) 186363d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1864e22bee78STejun Heo else { 1865e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 186611ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 186763d95a91STejun Heo wake_up_worker(pool); 1868e22bee78STejun Heo } 1869e22bee78STejun Heo } 1870e22bee78STejun Heo 1871d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1872e22bee78STejun Heo } 1873e22bee78STejun Heo 1874493a1724STejun Heo static void send_mayday(struct work_struct *work) 1875e22bee78STejun Heo { 1876112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1877112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1878493a1724STejun Heo 18792e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1880e22bee78STejun Heo 1881493008a8STejun Heo if (!wq->rescuer) 1882493a1724STejun Heo return; 1883e22bee78STejun Heo 1884e22bee78STejun Heo /* mayday mayday mayday */ 1885493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1886493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1887e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1888493a1724STejun Heo } 1889e22bee78STejun Heo } 1890e22bee78STejun Heo 1891706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1892e22bee78STejun Heo { 189363d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1894e22bee78STejun Heo struct work_struct *work; 1895e22bee78STejun Heo 18962e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1897493a1724STejun Heo spin_lock(&pool->lock); 1898e22bee78STejun Heo 189963d95a91STejun Heo if (need_to_create_worker(pool)) { 1900e22bee78STejun Heo /* 1901e22bee78STejun Heo * We've been trying to create a new worker but 1902e22bee78STejun Heo * haven't been successful. We might be hitting an 1903e22bee78STejun Heo * allocation deadlock. Send distress signals to 1904e22bee78STejun Heo * rescuers. 1905e22bee78STejun Heo */ 190663d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1907e22bee78STejun Heo send_mayday(work); 1908e22bee78STejun Heo } 1909e22bee78STejun Heo 1910493a1724STejun Heo spin_unlock(&pool->lock); 19112e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1912e22bee78STejun Heo 191363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1914e22bee78STejun Heo } 1915e22bee78STejun Heo 1916e22bee78STejun Heo /** 1917e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 191863d95a91STejun Heo * @pool: pool to create a new worker for 1919e22bee78STejun Heo * 192063d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1921e22bee78STejun Heo * have at least one idle worker on return from this function. If 1922e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 192363d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1924e22bee78STejun Heo * possible allocation deadlock. 1925e22bee78STejun Heo * 1926c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1927c5aa87bbSTejun Heo * may_start_working() %true. 1928e22bee78STejun Heo * 1929e22bee78STejun Heo * LOCKING: 1930d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1931e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1932e22bee78STejun Heo * manager. 1933e22bee78STejun Heo * 1934e22bee78STejun Heo * RETURNS: 1935c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1936e22bee78STejun Heo * otherwise. 1937e22bee78STejun Heo */ 193863d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1939d565ed63STejun Heo __releases(&pool->lock) 1940d565ed63STejun Heo __acquires(&pool->lock) 1941e22bee78STejun Heo { 194263d95a91STejun Heo if (!need_to_create_worker(pool)) 1943e22bee78STejun Heo return false; 1944e22bee78STejun Heo restart: 1945d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19469f9c2364STejun Heo 1947e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 194863d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1949e22bee78STejun Heo 1950e22bee78STejun Heo while (true) { 1951e22bee78STejun Heo struct worker *worker; 1952e22bee78STejun Heo 1953bc2ae0f5STejun Heo worker = create_worker(pool); 1954e22bee78STejun Heo if (worker) { 195563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1956d565ed63STejun Heo spin_lock_irq(&pool->lock); 1957e22bee78STejun Heo start_worker(worker); 19586183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19596183c009STejun Heo goto restart; 1960e22bee78STejun Heo return true; 1961e22bee78STejun Heo } 1962e22bee78STejun Heo 196363d95a91STejun Heo if (!need_to_create_worker(pool)) 1964e22bee78STejun Heo break; 1965e22bee78STejun Heo 1966e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1967e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19689f9c2364STejun Heo 196963d95a91STejun Heo if (!need_to_create_worker(pool)) 1970e22bee78STejun Heo break; 1971e22bee78STejun Heo } 1972e22bee78STejun Heo 197363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1974d565ed63STejun Heo spin_lock_irq(&pool->lock); 197563d95a91STejun Heo if (need_to_create_worker(pool)) 1976e22bee78STejun Heo goto restart; 1977e22bee78STejun Heo return true; 1978e22bee78STejun Heo } 1979e22bee78STejun Heo 1980e22bee78STejun Heo /** 1981e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 198263d95a91STejun Heo * @pool: pool to destroy workers for 1983e22bee78STejun Heo * 198463d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1985e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1986e22bee78STejun Heo * 1987e22bee78STejun Heo * LOCKING: 1988d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1989e22bee78STejun Heo * multiple times. Called only from manager. 1990e22bee78STejun Heo * 1991e22bee78STejun Heo * RETURNS: 1992c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1993e22bee78STejun Heo * otherwise. 1994e22bee78STejun Heo */ 199563d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1996e22bee78STejun Heo { 1997e22bee78STejun Heo bool ret = false; 1998e22bee78STejun Heo 199963d95a91STejun Heo while (too_many_workers(pool)) { 2000e22bee78STejun Heo struct worker *worker; 2001e22bee78STejun Heo unsigned long expires; 2002e22bee78STejun Heo 200363d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2004e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2005e22bee78STejun Heo 2006e22bee78STejun Heo if (time_before(jiffies, expires)) { 200763d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 2008e22bee78STejun Heo break; 2009e22bee78STejun Heo } 2010e22bee78STejun Heo 2011e22bee78STejun Heo destroy_worker(worker); 2012e22bee78STejun Heo ret = true; 2013e22bee78STejun Heo } 2014e22bee78STejun Heo 2015e22bee78STejun Heo return ret; 2016e22bee78STejun Heo } 2017e22bee78STejun Heo 2018e22bee78STejun Heo /** 2019e22bee78STejun Heo * manage_workers - manage worker pool 2020e22bee78STejun Heo * @worker: self 2021e22bee78STejun Heo * 2022706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2023e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2024706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2025e22bee78STejun Heo * 2026e22bee78STejun Heo * The caller can safely start processing works on false return. On 2027e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2028e22bee78STejun Heo * and may_start_working() is true. 2029e22bee78STejun Heo * 2030e22bee78STejun Heo * CONTEXT: 2031d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2032e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2033e22bee78STejun Heo * 2034e22bee78STejun Heo * RETURNS: 2035d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2036d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2037e22bee78STejun Heo */ 2038e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2039e22bee78STejun Heo { 204063d95a91STejun Heo struct worker_pool *pool = worker->pool; 2041e22bee78STejun Heo bool ret = false; 2042e22bee78STejun Heo 2043bc3a1afcSTejun Heo /* 2044bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2045bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2046bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2047bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2048bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2049bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2050bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2051bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2052bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2053bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2054bc3a1afcSTejun Heo * 2055bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2056bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2057bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2058bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2059bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2060bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2061bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2062bc3a1afcSTejun Heo * manager_mutex. 2063bc3a1afcSTejun Heo */ 206434a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2065e22bee78STejun Heo return ret; 2066e22bee78STejun Heo 2067ee378aa4SLai Jiangshan /* 2068bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2069bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2070ee378aa4SLai Jiangshan */ 2071bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2072d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2073bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 2074ee378aa4SLai Jiangshan ret = true; 2075ee378aa4SLai Jiangshan } 2076ee378aa4SLai Jiangshan 207711ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2078e22bee78STejun Heo 2079e22bee78STejun Heo /* 2080e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2081e22bee78STejun Heo * on return. 2082e22bee78STejun Heo */ 208363d95a91STejun Heo ret |= maybe_destroy_workers(pool); 208463d95a91STejun Heo ret |= maybe_create_worker(pool); 2085e22bee78STejun Heo 2086bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 208734a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2088e22bee78STejun Heo return ret; 2089e22bee78STejun Heo } 2090e22bee78STejun Heo 2091a62428c0STejun Heo /** 2092a62428c0STejun Heo * process_one_work - process single work 2093c34056a3STejun Heo * @worker: self 2094a62428c0STejun Heo * @work: work to process 2095a62428c0STejun Heo * 2096a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2097a62428c0STejun Heo * process a single work including synchronization against and 2098a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2099a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2100a62428c0STejun Heo * call this function to process a work. 2101a62428c0STejun Heo * 2102a62428c0STejun Heo * CONTEXT: 2103d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2104a62428c0STejun Heo */ 2105c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2106d565ed63STejun Heo __releases(&pool->lock) 2107d565ed63STejun Heo __acquires(&pool->lock) 21081da177e4SLinus Torvalds { 2109112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2110bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2111112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 211273f53c4aSTejun Heo int work_color; 21137e11629dSTejun Heo struct worker *collision; 21144e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21154e6045f1SJohannes Berg /* 2116a62428c0STejun Heo * It is permissible to free the struct work_struct from 2117a62428c0STejun Heo * inside the function that is called from it, this we need to 2118a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2119a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2120a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21214e6045f1SJohannes Berg */ 21224d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21234d82a1deSPeter Zijlstra 21244d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21254e6045f1SJohannes Berg #endif 21266fec10a1STejun Heo /* 21276fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21286fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 212924647570STejun Heo * unbound or a disassociated pool. 21306fec10a1STejun Heo */ 21315f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 213224647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2133ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 213425511a47STejun Heo 21357e11629dSTejun Heo /* 21367e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21377e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21387e11629dSTejun Heo * already processing the work. If so, defer the work to the 21397e11629dSTejun Heo * currently executing one. 21407e11629dSTejun Heo */ 2141c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21427e11629dSTejun Heo if (unlikely(collision)) { 21437e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21447e11629dSTejun Heo return; 21457e11629dSTejun Heo } 21461da177e4SLinus Torvalds 21478930cabaSTejun Heo /* claim and dequeue */ 21481da177e4SLinus Torvalds debug_work_deactivate(work); 2149c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2150c34056a3STejun Heo worker->current_work = work; 2151a2c1c57bSTejun Heo worker->current_func = work->func; 2152112202d9STejun Heo worker->current_pwq = pwq; 215373f53c4aSTejun Heo work_color = get_work_color(work); 21547a22ad75STejun Heo 2155a62428c0STejun Heo list_del_init(&work->entry); 2156a62428c0STejun Heo 2157649027d7STejun Heo /* 2158fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2159fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2160fb0e7bebSTejun Heo */ 2161fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2162fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2163fb0e7bebSTejun Heo 2164974271c4STejun Heo /* 2165d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2166974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2167974271c4STejun Heo */ 216863d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 216963d95a91STejun Heo wake_up_worker(pool); 2170974271c4STejun Heo 21718930cabaSTejun Heo /* 21727c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2173d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 217423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 217523657bb1STejun Heo * disabled. 21768930cabaSTejun Heo */ 21777c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21781da177e4SLinus Torvalds 2179d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2180365970a1SDavid Howells 2181112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21823295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2183e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2184a2c1c57bSTejun Heo worker->current_func(work); 2185e36c886aSArjan van de Ven /* 2186e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2187e36c886aSArjan van de Ven * point will only record its address. 2188e36c886aSArjan van de Ven */ 2189e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21903295f0efSIngo Molnar lock_map_release(&lockdep_map); 2191112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21921da177e4SLinus Torvalds 2193d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2194044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2195044c782cSValentin Ilie " last function: %pf\n", 2196a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2197a2c1c57bSTejun Heo worker->current_func); 2198d5abe669SPeter Zijlstra debug_show_held_locks(current); 2199d5abe669SPeter Zijlstra dump_stack(); 2200d5abe669SPeter Zijlstra } 2201d5abe669SPeter Zijlstra 2202d565ed63STejun Heo spin_lock_irq(&pool->lock); 2203a62428c0STejun Heo 2204fb0e7bebSTejun Heo /* clear cpu intensive status */ 2205fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2206fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2207fb0e7bebSTejun Heo 2208a62428c0STejun Heo /* we're done with it, release */ 220942f8570fSSasha Levin hash_del(&worker->hentry); 2210c34056a3STejun Heo worker->current_work = NULL; 2211a2c1c57bSTejun Heo worker->current_func = NULL; 2212112202d9STejun Heo worker->current_pwq = NULL; 2213112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22141da177e4SLinus Torvalds } 22151da177e4SLinus Torvalds 2216affee4b2STejun Heo /** 2217affee4b2STejun Heo * process_scheduled_works - process scheduled works 2218affee4b2STejun Heo * @worker: self 2219affee4b2STejun Heo * 2220affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2221affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2222affee4b2STejun Heo * fetches a work from the top and executes it. 2223affee4b2STejun Heo * 2224affee4b2STejun Heo * CONTEXT: 2225d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2226affee4b2STejun Heo * multiple times. 2227affee4b2STejun Heo */ 2228affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22291da177e4SLinus Torvalds { 2230affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2231affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2232a62428c0STejun Heo struct work_struct, entry); 2233c34056a3STejun Heo process_one_work(worker, work); 2234a62428c0STejun Heo } 22351da177e4SLinus Torvalds } 22361da177e4SLinus Torvalds 22374690c4abSTejun Heo /** 22384690c4abSTejun Heo * worker_thread - the worker thread function 2239c34056a3STejun Heo * @__worker: self 22404690c4abSTejun Heo * 2241c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2242c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2243c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2244c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2245c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22464690c4abSTejun Heo */ 2247c34056a3STejun Heo static int worker_thread(void *__worker) 22481da177e4SLinus Torvalds { 2249c34056a3STejun Heo struct worker *worker = __worker; 2250bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22511da177e4SLinus Torvalds 2252e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2253e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2254c8e55f36STejun Heo woke_up: 2255d565ed63STejun Heo spin_lock_irq(&pool->lock); 2256affee4b2STejun Heo 2257a9ab775bSTejun Heo /* am I supposed to die? */ 2258a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2259d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2260a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2261e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2262c8e55f36STejun Heo return 0; 2263c8e55f36STejun Heo } 2264c8e55f36STejun Heo 2265c8e55f36STejun Heo worker_leave_idle(worker); 2266db7bccf4STejun Heo recheck: 2267e22bee78STejun Heo /* no more worker necessary? */ 226863d95a91STejun Heo if (!need_more_worker(pool)) 2269e22bee78STejun Heo goto sleep; 2270e22bee78STejun Heo 2271e22bee78STejun Heo /* do we need to manage? */ 227263d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2273e22bee78STejun Heo goto recheck; 2274e22bee78STejun Heo 2275c8e55f36STejun Heo /* 2276c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2277c8e55f36STejun Heo * preparing to process a work or actually processing it. 2278c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2279c8e55f36STejun Heo */ 22806183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2281c8e55f36STejun Heo 2282e22bee78STejun Heo /* 2283a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2284a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2285a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2286a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2287a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2288e22bee78STejun Heo */ 2289a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2290e22bee78STejun Heo 2291e22bee78STejun Heo do { 2292affee4b2STejun Heo struct work_struct *work = 2293bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2294affee4b2STejun Heo struct work_struct, entry); 2295affee4b2STejun Heo 2296c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2297affee4b2STejun Heo /* optimization path, not strictly necessary */ 2298affee4b2STejun Heo process_one_work(worker, work); 2299affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2300affee4b2STejun Heo process_scheduled_works(worker); 2301affee4b2STejun Heo } else { 2302c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2303affee4b2STejun Heo process_scheduled_works(worker); 2304affee4b2STejun Heo } 230563d95a91STejun Heo } while (keep_working(pool)); 2306affee4b2STejun Heo 2307e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2308d313dd85STejun Heo sleep: 230963d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2310e22bee78STejun Heo goto recheck; 2311d313dd85STejun Heo 2312c8e55f36STejun Heo /* 2313d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2314d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2315d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2316d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2317d565ed63STejun Heo * event. 2318c8e55f36STejun Heo */ 2319c8e55f36STejun Heo worker_enter_idle(worker); 2320c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2321d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23221da177e4SLinus Torvalds schedule(); 2323c8e55f36STejun Heo goto woke_up; 23241da177e4SLinus Torvalds } 23251da177e4SLinus Torvalds 2326e22bee78STejun Heo /** 2327e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2328111c225aSTejun Heo * @__rescuer: self 2329e22bee78STejun Heo * 2330e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2331493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2332e22bee78STejun Heo * 2333706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2334e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2335e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2336e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2337e22bee78STejun Heo * the problem rescuer solves. 2338e22bee78STejun Heo * 2339706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2340706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2341e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2342e22bee78STejun Heo * 2343e22bee78STejun Heo * This should happen rarely. 2344e22bee78STejun Heo */ 2345111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2346e22bee78STejun Heo { 2347111c225aSTejun Heo struct worker *rescuer = __rescuer; 2348111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2349e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2350e22bee78STejun Heo 2351e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2352111c225aSTejun Heo 2353111c225aSTejun Heo /* 2354111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2355111c225aSTejun Heo * doesn't participate in concurrency management. 2356111c225aSTejun Heo */ 2357111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2358e22bee78STejun Heo repeat: 2359e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23601da177e4SLinus Torvalds 2361412d32e6SMike Galbraith if (kthread_should_stop()) { 2362412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2363111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2364e22bee78STejun Heo return 0; 2365412d32e6SMike Galbraith } 23661da177e4SLinus Torvalds 2367493a1724STejun Heo /* see whether any pwq is asking for help */ 23682e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2369493a1724STejun Heo 2370493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2371493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2372493a1724STejun Heo struct pool_workqueue, mayday_node); 2373112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2374e22bee78STejun Heo struct work_struct *work, *n; 2375e22bee78STejun Heo 2376e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2377493a1724STejun Heo list_del_init(&pwq->mayday_node); 2378493a1724STejun Heo 23792e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2380e22bee78STejun Heo 2381e22bee78STejun Heo /* migrate to the target cpu if possible */ 2382f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2383b3104104SLai Jiangshan rescuer->pool = pool; 2384e22bee78STejun Heo 2385e22bee78STejun Heo /* 2386e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2387e22bee78STejun Heo * process'em. 2388e22bee78STejun Heo */ 23896183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2390bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2391112202d9STejun Heo if (get_work_pwq(work) == pwq) 2392e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2393e22bee78STejun Heo 2394e22bee78STejun Heo process_scheduled_works(rescuer); 23957576958aSTejun Heo 23967576958aSTejun Heo /* 2397d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23987576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23997576958aSTejun Heo * and stalling the execution. 24007576958aSTejun Heo */ 240163d95a91STejun Heo if (keep_working(pool)) 240263d95a91STejun Heo wake_up_worker(pool); 24037576958aSTejun Heo 2404b3104104SLai Jiangshan rescuer->pool = NULL; 2405493a1724STejun Heo spin_unlock(&pool->lock); 24062e109a28STejun Heo spin_lock(&wq_mayday_lock); 24071da177e4SLinus Torvalds } 24081da177e4SLinus Torvalds 24092e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2410493a1724STejun Heo 2411111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2412111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2413e22bee78STejun Heo schedule(); 2414e22bee78STejun Heo goto repeat; 24151da177e4SLinus Torvalds } 24161da177e4SLinus Torvalds 2417fc2e4d70SOleg Nesterov struct wq_barrier { 2418fc2e4d70SOleg Nesterov struct work_struct work; 2419fc2e4d70SOleg Nesterov struct completion done; 2420fc2e4d70SOleg Nesterov }; 2421fc2e4d70SOleg Nesterov 2422fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2423fc2e4d70SOleg Nesterov { 2424fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2425fc2e4d70SOleg Nesterov complete(&barr->done); 2426fc2e4d70SOleg Nesterov } 2427fc2e4d70SOleg Nesterov 24284690c4abSTejun Heo /** 24294690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2430112202d9STejun Heo * @pwq: pwq to insert barrier into 24314690c4abSTejun Heo * @barr: wq_barrier to insert 2432affee4b2STejun Heo * @target: target work to attach @barr to 2433affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24344690c4abSTejun Heo * 2435affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2436affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2437affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2438affee4b2STejun Heo * cpu. 2439affee4b2STejun Heo * 2440affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2441affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2442affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2443affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2444affee4b2STejun Heo * after a work with LINKED flag set. 2445affee4b2STejun Heo * 2446affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2447112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24484690c4abSTejun Heo * 24494690c4abSTejun Heo * CONTEXT: 2450d565ed63STejun Heo * spin_lock_irq(pool->lock). 24514690c4abSTejun Heo */ 2452112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2453affee4b2STejun Heo struct wq_barrier *barr, 2454affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2455fc2e4d70SOleg Nesterov { 2456affee4b2STejun Heo struct list_head *head; 2457affee4b2STejun Heo unsigned int linked = 0; 2458affee4b2STejun Heo 2459dc186ad7SThomas Gleixner /* 2460d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2461dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2462dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2463dc186ad7SThomas Gleixner * might deadlock. 2464dc186ad7SThomas Gleixner */ 2465ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 246622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2467fc2e4d70SOleg Nesterov init_completion(&barr->done); 246883c22520SOleg Nesterov 2469affee4b2STejun Heo /* 2470affee4b2STejun Heo * If @target is currently being executed, schedule the 2471affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2472affee4b2STejun Heo */ 2473affee4b2STejun Heo if (worker) 2474affee4b2STejun Heo head = worker->scheduled.next; 2475affee4b2STejun Heo else { 2476affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2477affee4b2STejun Heo 2478affee4b2STejun Heo head = target->entry.next; 2479affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2480affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2481affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2482affee4b2STejun Heo } 2483affee4b2STejun Heo 2484dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2485112202d9STejun Heo insert_work(pwq, &barr->work, head, 2486affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2487fc2e4d70SOleg Nesterov } 2488fc2e4d70SOleg Nesterov 248973f53c4aSTejun Heo /** 2490112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 249173f53c4aSTejun Heo * @wq: workqueue being flushed 249273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 249373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 249473f53c4aSTejun Heo * 2495112202d9STejun Heo * Prepare pwqs for workqueue flushing. 249673f53c4aSTejun Heo * 2497112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2498112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2499112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2500112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2501112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 250273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 250373f53c4aSTejun Heo * 250473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 250573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 250673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 250773f53c4aSTejun Heo * is returned. 250873f53c4aSTejun Heo * 2509112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 251073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 251173f53c4aSTejun Heo * advanced to @work_color. 251273f53c4aSTejun Heo * 251373f53c4aSTejun Heo * CONTEXT: 25143c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 251573f53c4aSTejun Heo * 251673f53c4aSTejun Heo * RETURNS: 251773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 251873f53c4aSTejun Heo * otherwise. 251973f53c4aSTejun Heo */ 2520112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 252173f53c4aSTejun Heo int flush_color, int work_color) 25221da177e4SLinus Torvalds { 252373f53c4aSTejun Heo bool wait = false; 252449e3cf44STejun Heo struct pool_workqueue *pwq; 25251da177e4SLinus Torvalds 252673f53c4aSTejun Heo if (flush_color >= 0) { 25276183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2528112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2529dc186ad7SThomas Gleixner } 253014441960SOleg Nesterov 253149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2532112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25331da177e4SLinus Torvalds 2534b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 253573f53c4aSTejun Heo 253673f53c4aSTejun Heo if (flush_color >= 0) { 25376183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 253873f53c4aSTejun Heo 2539112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2540112202d9STejun Heo pwq->flush_color = flush_color; 2541112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 254273f53c4aSTejun Heo wait = true; 25431da177e4SLinus Torvalds } 254473f53c4aSTejun Heo } 254573f53c4aSTejun Heo 254673f53c4aSTejun Heo if (work_color >= 0) { 25476183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2548112202d9STejun Heo pwq->work_color = work_color; 254973f53c4aSTejun Heo } 255073f53c4aSTejun Heo 2551b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25521da177e4SLinus Torvalds } 25531da177e4SLinus Torvalds 2554112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 255573f53c4aSTejun Heo complete(&wq->first_flusher->done); 255673f53c4aSTejun Heo 255773f53c4aSTejun Heo return wait; 255883c22520SOleg Nesterov } 25591da177e4SLinus Torvalds 25600fcb78c2SRolf Eike Beer /** 25611da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25620fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25631da177e4SLinus Torvalds * 2564c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2565c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25661da177e4SLinus Torvalds */ 25677ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25681da177e4SLinus Torvalds { 256973f53c4aSTejun Heo struct wq_flusher this_flusher = { 257073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 257173f53c4aSTejun Heo .flush_color = -1, 257273f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 257373f53c4aSTejun Heo }; 257473f53c4aSTejun Heo int next_color; 2575b1f4ec17SOleg Nesterov 25763295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25773295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 257873f53c4aSTejun Heo 25793c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 258073f53c4aSTejun Heo 258173f53c4aSTejun Heo /* 258273f53c4aSTejun Heo * Start-to-wait phase 258373f53c4aSTejun Heo */ 258473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 258573f53c4aSTejun Heo 258673f53c4aSTejun Heo if (next_color != wq->flush_color) { 258773f53c4aSTejun Heo /* 258873f53c4aSTejun Heo * Color space is not full. The current work_color 258973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 259073f53c4aSTejun Heo * by one. 259173f53c4aSTejun Heo */ 25926183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 259373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 259473f53c4aSTejun Heo wq->work_color = next_color; 259573f53c4aSTejun Heo 259673f53c4aSTejun Heo if (!wq->first_flusher) { 259773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25986183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 259973f53c4aSTejun Heo 260073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 260173f53c4aSTejun Heo 2602112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 260373f53c4aSTejun Heo wq->work_color)) { 260473f53c4aSTejun Heo /* nothing to flush, done */ 260573f53c4aSTejun Heo wq->flush_color = next_color; 260673f53c4aSTejun Heo wq->first_flusher = NULL; 260773f53c4aSTejun Heo goto out_unlock; 260873f53c4aSTejun Heo } 260973f53c4aSTejun Heo } else { 261073f53c4aSTejun Heo /* wait in queue */ 26116183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 261273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2613112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 261473f53c4aSTejun Heo } 261573f53c4aSTejun Heo } else { 261673f53c4aSTejun Heo /* 261773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 261873f53c4aSTejun Heo * The next flush completion will assign us 261973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 262073f53c4aSTejun Heo */ 262173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 262273f53c4aSTejun Heo } 262373f53c4aSTejun Heo 26243c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 262573f53c4aSTejun Heo 262673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 262773f53c4aSTejun Heo 262873f53c4aSTejun Heo /* 262973f53c4aSTejun Heo * Wake-up-and-cascade phase 263073f53c4aSTejun Heo * 263173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 263273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 263373f53c4aSTejun Heo */ 263473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 263573f53c4aSTejun Heo return; 263673f53c4aSTejun Heo 26373c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 263873f53c4aSTejun Heo 26394ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26404ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26414ce48b37STejun Heo goto out_unlock; 26424ce48b37STejun Heo 264373f53c4aSTejun Heo wq->first_flusher = NULL; 264473f53c4aSTejun Heo 26456183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26466183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 264773f53c4aSTejun Heo 264873f53c4aSTejun Heo while (true) { 264973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 265073f53c4aSTejun Heo 265173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 265273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 265373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 265473f53c4aSTejun Heo break; 265573f53c4aSTejun Heo list_del_init(&next->list); 265673f53c4aSTejun Heo complete(&next->done); 265773f53c4aSTejun Heo } 265873f53c4aSTejun Heo 26596183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 266073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 266173f53c4aSTejun Heo 266273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 266373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 266473f53c4aSTejun Heo 266573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 266673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 266773f53c4aSTejun Heo /* 266873f53c4aSTejun Heo * Assign the same color to all overflowed 266973f53c4aSTejun Heo * flushers, advance work_color and append to 267073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 267173f53c4aSTejun Heo * phase for these overflowed flushers. 267273f53c4aSTejun Heo */ 267373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 267473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 267573f53c4aSTejun Heo 267673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 267773f53c4aSTejun Heo 267873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 267973f53c4aSTejun Heo &wq->flusher_queue); 2680112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 268173f53c4aSTejun Heo } 268273f53c4aSTejun Heo 268373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26846183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 268573f53c4aSTejun Heo break; 268673f53c4aSTejun Heo } 268773f53c4aSTejun Heo 268873f53c4aSTejun Heo /* 268973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2690112202d9STejun Heo * the new first flusher and arm pwqs. 269173f53c4aSTejun Heo */ 26926183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26936183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 269473f53c4aSTejun Heo 269573f53c4aSTejun Heo list_del_init(&next->list); 269673f53c4aSTejun Heo wq->first_flusher = next; 269773f53c4aSTejun Heo 2698112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 269973f53c4aSTejun Heo break; 270073f53c4aSTejun Heo 270173f53c4aSTejun Heo /* 270273f53c4aSTejun Heo * Meh... this color is already done, clear first 270373f53c4aSTejun Heo * flusher and repeat cascading. 270473f53c4aSTejun Heo */ 270573f53c4aSTejun Heo wq->first_flusher = NULL; 270673f53c4aSTejun Heo } 270773f53c4aSTejun Heo 270873f53c4aSTejun Heo out_unlock: 27093c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27101da177e4SLinus Torvalds } 2711ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27121da177e4SLinus Torvalds 27139c5a2ba7STejun Heo /** 27149c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27159c5a2ba7STejun Heo * @wq: workqueue to drain 27169c5a2ba7STejun Heo * 27179c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27189c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27199c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27209c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27219c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27229c5a2ba7STejun Heo * takes too long. 27239c5a2ba7STejun Heo */ 27249c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27259c5a2ba7STejun Heo { 27269c5a2ba7STejun Heo unsigned int flush_cnt = 0; 272749e3cf44STejun Heo struct pool_workqueue *pwq; 27289c5a2ba7STejun Heo 27299c5a2ba7STejun Heo /* 27309c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27319c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2732618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27339c5a2ba7STejun Heo */ 273487fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27359c5a2ba7STejun Heo if (!wq->nr_drainers++) 2736618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 273787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27389c5a2ba7STejun Heo reflush: 27399c5a2ba7STejun Heo flush_workqueue(wq); 27409c5a2ba7STejun Heo 2741b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 274276af4d93STejun Heo 274349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2744fa2563e4SThomas Tuttle bool drained; 27459c5a2ba7STejun Heo 2746b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2747112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2748b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2749fa2563e4SThomas Tuttle 2750fa2563e4SThomas Tuttle if (drained) 27519c5a2ba7STejun Heo continue; 27529c5a2ba7STejun Heo 27539c5a2ba7STejun Heo if (++flush_cnt == 10 || 27549c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2755c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27569c5a2ba7STejun Heo wq->name, flush_cnt); 275776af4d93STejun Heo 2758b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27599c5a2ba7STejun Heo goto reflush; 27609c5a2ba7STejun Heo } 27619c5a2ba7STejun Heo 27629c5a2ba7STejun Heo if (!--wq->nr_drainers) 2763618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 276487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27659c5a2ba7STejun Heo } 27669c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27679c5a2ba7STejun Heo 2768606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2769baf59022STejun Heo { 2770baf59022STejun Heo struct worker *worker = NULL; 2771c9e7cf27STejun Heo struct worker_pool *pool; 2772112202d9STejun Heo struct pool_workqueue *pwq; 2773baf59022STejun Heo 2774baf59022STejun Heo might_sleep(); 2775baf59022STejun Heo 2776fa1b54e6STejun Heo local_irq_disable(); 2777fa1b54e6STejun Heo pool = get_work_pool(work); 2778fa1b54e6STejun Heo if (!pool) { 2779fa1b54e6STejun Heo local_irq_enable(); 2780fa1b54e6STejun Heo return false; 2781fa1b54e6STejun Heo } 2782fa1b54e6STejun Heo 2783fa1b54e6STejun Heo spin_lock(&pool->lock); 27840b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2785112202d9STejun Heo pwq = get_work_pwq(work); 2786112202d9STejun Heo if (pwq) { 2787112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2788baf59022STejun Heo goto already_gone; 2789606a5020STejun Heo } else { 2790c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2791baf59022STejun Heo if (!worker) 2792baf59022STejun Heo goto already_gone; 2793112202d9STejun Heo pwq = worker->current_pwq; 2794606a5020STejun Heo } 2795baf59022STejun Heo 2796112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2797d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2798baf59022STejun Heo 2799e159489bSTejun Heo /* 2800e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2801e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2802e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2803e159489bSTejun Heo * access. 2804e159489bSTejun Heo */ 2805493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2806112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2807e159489bSTejun Heo else 2808112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2809112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2810e159489bSTejun Heo 2811baf59022STejun Heo return true; 2812baf59022STejun Heo already_gone: 2813d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2814baf59022STejun Heo return false; 2815baf59022STejun Heo } 2816baf59022STejun Heo 2817db700897SOleg Nesterov /** 2818401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2819401a8d04STejun Heo * @work: the work to flush 2820db700897SOleg Nesterov * 2821606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2822606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2823401a8d04STejun Heo * 2824401a8d04STejun Heo * RETURNS: 2825401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2826401a8d04STejun Heo * %false if it was already idle. 2827db700897SOleg Nesterov */ 2828401a8d04STejun Heo bool flush_work(struct work_struct *work) 2829db700897SOleg Nesterov { 2830db700897SOleg Nesterov struct wq_barrier barr; 2831db700897SOleg Nesterov 28320976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28330976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28340976dfc1SStephen Boyd 2835606a5020STejun Heo if (start_flush_work(work, &barr)) { 2836db700897SOleg Nesterov wait_for_completion(&barr.done); 2837dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2838401a8d04STejun Heo return true; 2839606a5020STejun Heo } else { 2840401a8d04STejun Heo return false; 2841db700897SOleg Nesterov } 2842606a5020STejun Heo } 2843db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2844db700897SOleg Nesterov 284536e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2846401a8d04STejun Heo { 2847bbb68dfaSTejun Heo unsigned long flags; 28481f1f642eSOleg Nesterov int ret; 28491f1f642eSOleg Nesterov 28501f1f642eSOleg Nesterov do { 2851bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2852bbb68dfaSTejun Heo /* 2853bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2854bbb68dfaSTejun Heo * would be waiting for before retrying. 2855bbb68dfaSTejun Heo */ 2856bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2857606a5020STejun Heo flush_work(work); 28581f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28591f1f642eSOleg Nesterov 2860bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2861bbb68dfaSTejun Heo mark_work_canceling(work); 2862bbb68dfaSTejun Heo local_irq_restore(flags); 2863bbb68dfaSTejun Heo 2864606a5020STejun Heo flush_work(work); 28657a22ad75STejun Heo clear_work_data(work); 28661f1f642eSOleg Nesterov return ret; 28671f1f642eSOleg Nesterov } 28681f1f642eSOleg Nesterov 28696e84d644SOleg Nesterov /** 2870401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2871401a8d04STejun Heo * @work: the work to cancel 28726e84d644SOleg Nesterov * 2873401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2874401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2875401a8d04STejun Heo * another workqueue. On return from this function, @work is 2876401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28771f1f642eSOleg Nesterov * 2878401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2879401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28806e84d644SOleg Nesterov * 2881401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28826e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2883401a8d04STejun Heo * 2884401a8d04STejun Heo * RETURNS: 2885401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28866e84d644SOleg Nesterov */ 2887401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28886e84d644SOleg Nesterov { 288936e227d2STejun Heo return __cancel_work_timer(work, false); 2890b89deed3SOleg Nesterov } 289128e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2892b89deed3SOleg Nesterov 28936e84d644SOleg Nesterov /** 2894401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2895401a8d04STejun Heo * @dwork: the delayed work to flush 28966e84d644SOleg Nesterov * 2897401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2898401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2899401a8d04STejun Heo * considers the last queueing instance of @dwork. 29001f1f642eSOleg Nesterov * 2901401a8d04STejun Heo * RETURNS: 2902401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2903401a8d04STejun Heo * %false if it was already idle. 29046e84d644SOleg Nesterov */ 2905401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2906401a8d04STejun Heo { 29078930cabaSTejun Heo local_irq_disable(); 2908401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 290960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29108930cabaSTejun Heo local_irq_enable(); 2911401a8d04STejun Heo return flush_work(&dwork->work); 2912401a8d04STejun Heo } 2913401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2914401a8d04STejun Heo 2915401a8d04STejun Heo /** 291657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 291757b30ae7STejun Heo * @dwork: delayed_work to cancel 291809383498STejun Heo * 291957b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 292057b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 292157b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 292257b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 292357b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 292409383498STejun Heo * 292557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 292609383498STejun Heo */ 292757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 292809383498STejun Heo { 292957b30ae7STejun Heo unsigned long flags; 293057b30ae7STejun Heo int ret; 293157b30ae7STejun Heo 293257b30ae7STejun Heo do { 293357b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 293457b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 293557b30ae7STejun Heo 293657b30ae7STejun Heo if (unlikely(ret < 0)) 293757b30ae7STejun Heo return false; 293857b30ae7STejun Heo 29397c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29407c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 294157b30ae7STejun Heo local_irq_restore(flags); 2942c0158ca6SDan Magenheimer return ret; 294309383498STejun Heo } 294457b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 294509383498STejun Heo 294609383498STejun Heo /** 2947401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2948401a8d04STejun Heo * @dwork: the delayed work cancel 2949401a8d04STejun Heo * 2950401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2951401a8d04STejun Heo * 2952401a8d04STejun Heo * RETURNS: 2953401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2954401a8d04STejun Heo */ 2955401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29566e84d644SOleg Nesterov { 295736e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29586e84d644SOleg Nesterov } 2959f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29601da177e4SLinus Torvalds 29610fcb78c2SRolf Eike Beer /** 296231ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2963b6136773SAndrew Morton * @func: the function to call 2964b6136773SAndrew Morton * 296531ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 296631ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2967b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 296831ddd871STejun Heo * 296931ddd871STejun Heo * RETURNS: 297031ddd871STejun Heo * 0 on success, -errno on failure. 2971b6136773SAndrew Morton */ 297265f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 297315316ba8SChristoph Lameter { 297415316ba8SChristoph Lameter int cpu; 297538f51568SNamhyung Kim struct work_struct __percpu *works; 297615316ba8SChristoph Lameter 2977b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2978b6136773SAndrew Morton if (!works) 297915316ba8SChristoph Lameter return -ENOMEM; 2980b6136773SAndrew Morton 298195402b38SGautham R Shenoy get_online_cpus(); 298293981800STejun Heo 298315316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29849bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29859bfb1839SIngo Molnar 29869bfb1839SIngo Molnar INIT_WORK(work, func); 29878de6d308SOleg Nesterov schedule_work_on(cpu, work); 298815316ba8SChristoph Lameter } 298993981800STejun Heo 299093981800STejun Heo for_each_online_cpu(cpu) 29918616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 299293981800STejun Heo 299395402b38SGautham R Shenoy put_online_cpus(); 2994b6136773SAndrew Morton free_percpu(works); 299515316ba8SChristoph Lameter return 0; 299615316ba8SChristoph Lameter } 299715316ba8SChristoph Lameter 2998eef6a7d5SAlan Stern /** 2999eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 3000eef6a7d5SAlan Stern * 3001eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 3002eef6a7d5SAlan Stern * completion. 3003eef6a7d5SAlan Stern * 3004eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 3005eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 3006eef6a7d5SAlan Stern * will lead to deadlock: 3007eef6a7d5SAlan Stern * 3008eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 3009eef6a7d5SAlan Stern * a lock held by your code or its caller. 3010eef6a7d5SAlan Stern * 3011eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3012eef6a7d5SAlan Stern * 3013eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3014eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3015eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3016eef6a7d5SAlan Stern * 3017eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3018eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3019eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3020eef6a7d5SAlan Stern * cancel_work_sync() instead. 3021eef6a7d5SAlan Stern */ 30221da177e4SLinus Torvalds void flush_scheduled_work(void) 30231da177e4SLinus Torvalds { 3024d320c038STejun Heo flush_workqueue(system_wq); 30251da177e4SLinus Torvalds } 3026ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30271da177e4SLinus Torvalds 30281da177e4SLinus Torvalds /** 30291fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30301fa44ecaSJames Bottomley * @fn: the function to execute 30311fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30321fa44ecaSJames Bottomley * be available when the work executes) 30331fa44ecaSJames Bottomley * 30341fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30351fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30361fa44ecaSJames Bottomley * 30371fa44ecaSJames Bottomley * Returns: 0 - function was executed 30381fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30391fa44ecaSJames Bottomley */ 304065f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30411fa44ecaSJames Bottomley { 30421fa44ecaSJames Bottomley if (!in_interrupt()) { 304365f27f38SDavid Howells fn(&ew->work); 30441fa44ecaSJames Bottomley return 0; 30451fa44ecaSJames Bottomley } 30461fa44ecaSJames Bottomley 304765f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30481fa44ecaSJames Bottomley schedule_work(&ew->work); 30491fa44ecaSJames Bottomley 30501fa44ecaSJames Bottomley return 1; 30511fa44ecaSJames Bottomley } 30521fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30531fa44ecaSJames Bottomley 3054226223abSTejun Heo #ifdef CONFIG_SYSFS 3055226223abSTejun Heo /* 3056226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3057226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3058226223abSTejun Heo * following attributes. 3059226223abSTejun Heo * 3060226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3061226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3062226223abSTejun Heo * 3063226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3064226223abSTejun Heo * 3065226223abSTejun Heo * id RO int : the associated pool ID 3066226223abSTejun Heo * nice RW int : nice value of the workers 3067226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3068226223abSTejun Heo */ 3069226223abSTejun Heo struct wq_device { 3070226223abSTejun Heo struct workqueue_struct *wq; 3071226223abSTejun Heo struct device dev; 3072226223abSTejun Heo }; 3073226223abSTejun Heo 3074226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3075226223abSTejun Heo { 3076226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3077226223abSTejun Heo 3078226223abSTejun Heo return wq_dev->wq; 3079226223abSTejun Heo } 3080226223abSTejun Heo 3081226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3082226223abSTejun Heo struct device_attribute *attr, char *buf) 3083226223abSTejun Heo { 3084226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3085226223abSTejun Heo 3086226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3087226223abSTejun Heo } 3088226223abSTejun Heo 3089226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3090226223abSTejun Heo struct device_attribute *attr, char *buf) 3091226223abSTejun Heo { 3092226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3093226223abSTejun Heo 3094226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3095226223abSTejun Heo } 3096226223abSTejun Heo 3097226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3098226223abSTejun Heo struct device_attribute *attr, 3099226223abSTejun Heo const char *buf, size_t count) 3100226223abSTejun Heo { 3101226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3102226223abSTejun Heo int val; 3103226223abSTejun Heo 3104226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3105226223abSTejun Heo return -EINVAL; 3106226223abSTejun Heo 3107226223abSTejun Heo workqueue_set_max_active(wq, val); 3108226223abSTejun Heo return count; 3109226223abSTejun Heo } 3110226223abSTejun Heo 3111226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3112226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3113226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3114226223abSTejun Heo __ATTR_NULL, 3115226223abSTejun Heo }; 3116226223abSTejun Heo 3117226223abSTejun Heo static ssize_t wq_pool_id_show(struct device *dev, 3118226223abSTejun Heo struct device_attribute *attr, char *buf) 3119226223abSTejun Heo { 3120226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3121226223abSTejun Heo struct worker_pool *pool; 3122226223abSTejun Heo int written; 3123226223abSTejun Heo 3124226223abSTejun Heo rcu_read_lock_sched(); 3125226223abSTejun Heo pool = first_pwq(wq)->pool; 3126226223abSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", pool->id); 3127226223abSTejun Heo rcu_read_unlock_sched(); 3128226223abSTejun Heo 3129226223abSTejun Heo return written; 3130226223abSTejun Heo } 3131226223abSTejun Heo 3132226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3133226223abSTejun Heo char *buf) 3134226223abSTejun Heo { 3135226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3136226223abSTejun Heo int written; 3137226223abSTejun Heo 31386029a918STejun Heo mutex_lock(&wq->mutex); 31396029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 31406029a918STejun Heo mutex_unlock(&wq->mutex); 3141226223abSTejun Heo 3142226223abSTejun Heo return written; 3143226223abSTejun Heo } 3144226223abSTejun Heo 3145226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3146226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3147226223abSTejun Heo { 3148226223abSTejun Heo struct workqueue_attrs *attrs; 3149226223abSTejun Heo 3150226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3151226223abSTejun Heo if (!attrs) 3152226223abSTejun Heo return NULL; 3153226223abSTejun Heo 31546029a918STejun Heo mutex_lock(&wq->mutex); 31556029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31566029a918STejun Heo mutex_unlock(&wq->mutex); 3157226223abSTejun Heo return attrs; 3158226223abSTejun Heo } 3159226223abSTejun Heo 3160226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3161226223abSTejun Heo const char *buf, size_t count) 3162226223abSTejun Heo { 3163226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3164226223abSTejun Heo struct workqueue_attrs *attrs; 3165226223abSTejun Heo int ret; 3166226223abSTejun Heo 3167226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3168226223abSTejun Heo if (!attrs) 3169226223abSTejun Heo return -ENOMEM; 3170226223abSTejun Heo 3171226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3172226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3173226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3174226223abSTejun Heo else 3175226223abSTejun Heo ret = -EINVAL; 3176226223abSTejun Heo 3177226223abSTejun Heo free_workqueue_attrs(attrs); 3178226223abSTejun Heo return ret ?: count; 3179226223abSTejun Heo } 3180226223abSTejun Heo 3181226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3182226223abSTejun Heo struct device_attribute *attr, char *buf) 3183226223abSTejun Heo { 3184226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3185226223abSTejun Heo int written; 3186226223abSTejun Heo 31876029a918STejun Heo mutex_lock(&wq->mutex); 31886029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 31896029a918STejun Heo mutex_unlock(&wq->mutex); 3190226223abSTejun Heo 3191226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3192226223abSTejun Heo return written; 3193226223abSTejun Heo } 3194226223abSTejun Heo 3195226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3196226223abSTejun Heo struct device_attribute *attr, 3197226223abSTejun Heo const char *buf, size_t count) 3198226223abSTejun Heo { 3199226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3200226223abSTejun Heo struct workqueue_attrs *attrs; 3201226223abSTejun Heo int ret; 3202226223abSTejun Heo 3203226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3204226223abSTejun Heo if (!attrs) 3205226223abSTejun Heo return -ENOMEM; 3206226223abSTejun Heo 3207226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3208226223abSTejun Heo if (!ret) 3209226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3210226223abSTejun Heo 3211226223abSTejun Heo free_workqueue_attrs(attrs); 3212226223abSTejun Heo return ret ?: count; 3213226223abSTejun Heo } 3214226223abSTejun Heo 3215226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3216226223abSTejun Heo __ATTR(pool_id, 0444, wq_pool_id_show, NULL), 3217226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3218226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3219226223abSTejun Heo __ATTR_NULL, 3220226223abSTejun Heo }; 3221226223abSTejun Heo 3222226223abSTejun Heo static struct bus_type wq_subsys = { 3223226223abSTejun Heo .name = "workqueue", 3224226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3225226223abSTejun Heo }; 3226226223abSTejun Heo 3227226223abSTejun Heo static int __init wq_sysfs_init(void) 3228226223abSTejun Heo { 3229226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3230226223abSTejun Heo } 3231226223abSTejun Heo core_initcall(wq_sysfs_init); 3232226223abSTejun Heo 3233226223abSTejun Heo static void wq_device_release(struct device *dev) 3234226223abSTejun Heo { 3235226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3236226223abSTejun Heo 3237226223abSTejun Heo kfree(wq_dev); 3238226223abSTejun Heo } 3239226223abSTejun Heo 3240226223abSTejun Heo /** 3241226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3242226223abSTejun Heo * @wq: the workqueue to register 3243226223abSTejun Heo * 3244226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3245226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3246226223abSTejun Heo * which is the preferred method. 3247226223abSTejun Heo * 3248226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3249226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3250226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3251226223abSTejun Heo * attributes. 3252226223abSTejun Heo * 3253226223abSTejun Heo * Returns 0 on success, -errno on failure. 3254226223abSTejun Heo */ 3255226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3256226223abSTejun Heo { 3257226223abSTejun Heo struct wq_device *wq_dev; 3258226223abSTejun Heo int ret; 3259226223abSTejun Heo 3260226223abSTejun Heo /* 3261226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3262226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3263226223abSTejun Heo * workqueues. 3264226223abSTejun Heo */ 3265226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3266226223abSTejun Heo return -EINVAL; 3267226223abSTejun Heo 3268226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3269226223abSTejun Heo if (!wq_dev) 3270226223abSTejun Heo return -ENOMEM; 3271226223abSTejun Heo 3272226223abSTejun Heo wq_dev->wq = wq; 3273226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3274226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3275226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3276226223abSTejun Heo 3277226223abSTejun Heo /* 3278226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3279226223abSTejun Heo * everything is ready. 3280226223abSTejun Heo */ 3281226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3282226223abSTejun Heo 3283226223abSTejun Heo ret = device_register(&wq_dev->dev); 3284226223abSTejun Heo if (ret) { 3285226223abSTejun Heo kfree(wq_dev); 3286226223abSTejun Heo wq->wq_dev = NULL; 3287226223abSTejun Heo return ret; 3288226223abSTejun Heo } 3289226223abSTejun Heo 3290226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3291226223abSTejun Heo struct device_attribute *attr; 3292226223abSTejun Heo 3293226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3294226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3295226223abSTejun Heo if (ret) { 3296226223abSTejun Heo device_unregister(&wq_dev->dev); 3297226223abSTejun Heo wq->wq_dev = NULL; 3298226223abSTejun Heo return ret; 3299226223abSTejun Heo } 3300226223abSTejun Heo } 3301226223abSTejun Heo } 3302226223abSTejun Heo 3303226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3304226223abSTejun Heo return 0; 3305226223abSTejun Heo } 3306226223abSTejun Heo 3307226223abSTejun Heo /** 3308226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3309226223abSTejun Heo * @wq: the workqueue to unregister 3310226223abSTejun Heo * 3311226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3312226223abSTejun Heo */ 3313226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3314226223abSTejun Heo { 3315226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3316226223abSTejun Heo 3317226223abSTejun Heo if (!wq->wq_dev) 3318226223abSTejun Heo return; 3319226223abSTejun Heo 3320226223abSTejun Heo wq->wq_dev = NULL; 3321226223abSTejun Heo device_unregister(&wq_dev->dev); 3322226223abSTejun Heo } 3323226223abSTejun Heo #else /* CONFIG_SYSFS */ 3324226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3325226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3326226223abSTejun Heo 33277a4e344cSTejun Heo /** 33287a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33297a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33307a4e344cSTejun Heo * 33317a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33327a4e344cSTejun Heo */ 33337a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33347a4e344cSTejun Heo { 33357a4e344cSTejun Heo if (attrs) { 33367a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33377a4e344cSTejun Heo kfree(attrs); 33387a4e344cSTejun Heo } 33397a4e344cSTejun Heo } 33407a4e344cSTejun Heo 33417a4e344cSTejun Heo /** 33427a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33437a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33447a4e344cSTejun Heo * 33457a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33467a4e344cSTejun Heo * return it. Returns NULL on failure. 33477a4e344cSTejun Heo */ 33487a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33497a4e344cSTejun Heo { 33507a4e344cSTejun Heo struct workqueue_attrs *attrs; 33517a4e344cSTejun Heo 33527a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33537a4e344cSTejun Heo if (!attrs) 33547a4e344cSTejun Heo goto fail; 33557a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33567a4e344cSTejun Heo goto fail; 33577a4e344cSTejun Heo 335813e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33597a4e344cSTejun Heo return attrs; 33607a4e344cSTejun Heo fail: 33617a4e344cSTejun Heo free_workqueue_attrs(attrs); 33627a4e344cSTejun Heo return NULL; 33637a4e344cSTejun Heo } 33647a4e344cSTejun Heo 336529c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 336629c91e99STejun Heo const struct workqueue_attrs *from) 336729c91e99STejun Heo { 336829c91e99STejun Heo to->nice = from->nice; 336929c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 337029c91e99STejun Heo } 337129c91e99STejun Heo 337229c91e99STejun Heo /* hash value of the content of @attr */ 337329c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 337429c91e99STejun Heo { 337529c91e99STejun Heo u32 hash = 0; 337629c91e99STejun Heo 337729c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 337813e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 337913e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 338029c91e99STejun Heo return hash; 338129c91e99STejun Heo } 338229c91e99STejun Heo 338329c91e99STejun Heo /* content equality test */ 338429c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 338529c91e99STejun Heo const struct workqueue_attrs *b) 338629c91e99STejun Heo { 338729c91e99STejun Heo if (a->nice != b->nice) 338829c91e99STejun Heo return false; 338929c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 339029c91e99STejun Heo return false; 339129c91e99STejun Heo return true; 339229c91e99STejun Heo } 339329c91e99STejun Heo 33947a4e344cSTejun Heo /** 33957a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 33967a4e344cSTejun Heo * @pool: worker_pool to initialize 33977a4e344cSTejun Heo * 33987a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 339929c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 340029c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 340129c91e99STejun Heo * on @pool safely to release it. 34027a4e344cSTejun Heo */ 34037a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34044e1a1f9aSTejun Heo { 34054e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 340629c91e99STejun Heo pool->id = -1; 340729c91e99STejun Heo pool->cpu = -1; 3408f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34094e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 34104e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34114e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34124e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34134e1a1f9aSTejun Heo 34144e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 34154e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 34164e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 34174e1a1f9aSTejun Heo 34184e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34194e1a1f9aSTejun Heo (unsigned long)pool); 34204e1a1f9aSTejun Heo 34214e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3422bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3423822d8405STejun Heo idr_init(&pool->worker_idr); 34247a4e344cSTejun Heo 342529c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 342629c91e99STejun Heo pool->refcnt = 1; 342729c91e99STejun Heo 342829c91e99STejun Heo /* shouldn't fail above this point */ 34297a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34307a4e344cSTejun Heo if (!pool->attrs) 34317a4e344cSTejun Heo return -ENOMEM; 34327a4e344cSTejun Heo return 0; 34334e1a1f9aSTejun Heo } 34344e1a1f9aSTejun Heo 343529c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 343629c91e99STejun Heo { 343729c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 343829c91e99STejun Heo 3439822d8405STejun Heo idr_destroy(&pool->worker_idr); 344029c91e99STejun Heo free_workqueue_attrs(pool->attrs); 344129c91e99STejun Heo kfree(pool); 344229c91e99STejun Heo } 344329c91e99STejun Heo 344429c91e99STejun Heo /** 344529c91e99STejun Heo * put_unbound_pool - put a worker_pool 344629c91e99STejun Heo * @pool: worker_pool to put 344729c91e99STejun Heo * 344829c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3449c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3450c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3451c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3452a892caccSTejun Heo * 3453a892caccSTejun Heo * Should be called with wq_pool_mutex held. 345429c91e99STejun Heo */ 345529c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 345629c91e99STejun Heo { 345729c91e99STejun Heo struct worker *worker; 345829c91e99STejun Heo 3459a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3460a892caccSTejun Heo 3461a892caccSTejun Heo if (--pool->refcnt) 346229c91e99STejun Heo return; 346329c91e99STejun Heo 346429c91e99STejun Heo /* sanity checks */ 346529c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3466a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 346729c91e99STejun Heo return; 346829c91e99STejun Heo 346929c91e99STejun Heo /* release id and unhash */ 347029c91e99STejun Heo if (pool->id >= 0) 347129c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 347229c91e99STejun Heo hash_del(&pool->hash_node); 347329c91e99STejun Heo 3474c5aa87bbSTejun Heo /* 3475c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3476c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3477c5aa87bbSTejun Heo * manager_mutex. 3478c5aa87bbSTejun Heo */ 347929c91e99STejun Heo mutex_lock(&pool->manager_arb); 3480cd549687STejun Heo mutex_lock(&pool->manager_mutex); 348129c91e99STejun Heo spin_lock_irq(&pool->lock); 348229c91e99STejun Heo 348329c91e99STejun Heo while ((worker = first_worker(pool))) 348429c91e99STejun Heo destroy_worker(worker); 348529c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 348629c91e99STejun Heo 348729c91e99STejun Heo spin_unlock_irq(&pool->lock); 3488cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 348929c91e99STejun Heo mutex_unlock(&pool->manager_arb); 349029c91e99STejun Heo 349129c91e99STejun Heo /* shut down the timers */ 349229c91e99STejun Heo del_timer_sync(&pool->idle_timer); 349329c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 349429c91e99STejun Heo 349529c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 349629c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 349729c91e99STejun Heo } 349829c91e99STejun Heo 349929c91e99STejun Heo /** 350029c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 350129c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 350229c91e99STejun Heo * 350329c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 350429c91e99STejun Heo * reference count and return it. If there already is a matching 350529c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 350629c91e99STejun Heo * create a new one. On failure, returns NULL. 3507a892caccSTejun Heo * 3508a892caccSTejun Heo * Should be called with wq_pool_mutex held. 350929c91e99STejun Heo */ 351029c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 351129c91e99STejun Heo { 351229c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 351329c91e99STejun Heo struct worker_pool *pool; 3514f3f90ad4STejun Heo int node; 351529c91e99STejun Heo 3516a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 351729c91e99STejun Heo 351829c91e99STejun Heo /* do we already have a matching pool? */ 351929c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 352029c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 352129c91e99STejun Heo pool->refcnt++; 352229c91e99STejun Heo goto out_unlock; 352329c91e99STejun Heo } 352429c91e99STejun Heo } 352529c91e99STejun Heo 352629c91e99STejun Heo /* nope, create a new one */ 352729c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 352829c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 352929c91e99STejun Heo goto fail; 353029c91e99STejun Heo 353112ee4fc6SLai Jiangshan if (workqueue_freezing) 353212ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 353312ee4fc6SLai Jiangshan 35348864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 353529c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 353629c91e99STejun Heo 3537f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3538f3f90ad4STejun Heo if (wq_numa_enabled) { 3539f3f90ad4STejun Heo for_each_node(node) { 3540f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3541f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3542f3f90ad4STejun Heo pool->node = node; 3543f3f90ad4STejun Heo break; 3544f3f90ad4STejun Heo } 3545f3f90ad4STejun Heo } 3546f3f90ad4STejun Heo } 3547f3f90ad4STejun Heo 354829c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 354929c91e99STejun Heo goto fail; 355029c91e99STejun Heo 355129c91e99STejun Heo /* create and start the initial worker */ 3552ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 355329c91e99STejun Heo goto fail; 355429c91e99STejun Heo 355529c91e99STejun Heo /* install */ 355629c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 355729c91e99STejun Heo out_unlock: 355829c91e99STejun Heo return pool; 355929c91e99STejun Heo fail: 356029c91e99STejun Heo if (pool) 356129c91e99STejun Heo put_unbound_pool(pool); 356229c91e99STejun Heo return NULL; 356329c91e99STejun Heo } 356429c91e99STejun Heo 35658864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35668864b4e5STejun Heo { 35678864b4e5STejun Heo kmem_cache_free(pwq_cache, 35688864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 35698864b4e5STejun Heo } 35708864b4e5STejun Heo 35718864b4e5STejun Heo /* 35728864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 35738864b4e5STejun Heo * and needs to be destroyed. 35748864b4e5STejun Heo */ 35758864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 35768864b4e5STejun Heo { 35778864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 35788864b4e5STejun Heo unbound_release_work); 35798864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 35808864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3581bc0caf09STejun Heo bool is_last; 35828864b4e5STejun Heo 35838864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 35848864b4e5STejun Heo return; 35858864b4e5STejun Heo 358675ccf595STejun Heo /* 35873c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 358875ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 358975ccf595STejun Heo * and consistent with the linking path. 359075ccf595STejun Heo */ 35913c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 35928864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3593bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 35943c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 35958864b4e5STejun Heo 3596a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 35978864b4e5STejun Heo put_unbound_pool(pool); 3598a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3599a892caccSTejun Heo 36008864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 36018864b4e5STejun Heo 36028864b4e5STejun Heo /* 36038864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 36048864b4e5STejun Heo * is gonna access it anymore. Free it. 36058864b4e5STejun Heo */ 36066029a918STejun Heo if (is_last) { 36076029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 36088864b4e5STejun Heo kfree(wq); 36098864b4e5STejun Heo } 36106029a918STejun Heo } 36118864b4e5STejun Heo 36120fbd95aaSTejun Heo /** 3613699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36140fbd95aaSTejun Heo * @pwq: target pool_workqueue 36150fbd95aaSTejun Heo * 3616699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3617699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3618699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36190fbd95aaSTejun Heo */ 3620699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36210fbd95aaSTejun Heo { 3622699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3623699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3624699ce097STejun Heo 3625699ce097STejun Heo /* for @wq->saved_max_active */ 3626a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3627699ce097STejun Heo 3628699ce097STejun Heo /* fast exit for non-freezable wqs */ 3629699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3630699ce097STejun Heo return; 3631699ce097STejun Heo 3632a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3633699ce097STejun Heo 3634699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3635699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36360fbd95aaSTejun Heo 36370fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36380fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36390fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3640951a078aSLai Jiangshan 3641951a078aSLai Jiangshan /* 3642951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3643951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3644951a078aSLai Jiangshan */ 3645951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3646699ce097STejun Heo } else { 3647699ce097STejun Heo pwq->max_active = 0; 3648699ce097STejun Heo } 3649699ce097STejun Heo 3650a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36510fbd95aaSTejun Heo } 36520fbd95aaSTejun Heo 3653e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3654f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3655f147f29eSTejun Heo struct worker_pool *pool) 3656d2c1d404STejun Heo { 3657d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3658d2c1d404STejun Heo 3659e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3660e50aba9aSTejun Heo 3661d2c1d404STejun Heo pwq->pool = pool; 3662d2c1d404STejun Heo pwq->wq = wq; 3663d2c1d404STejun Heo pwq->flush_color = -1; 36648864b4e5STejun Heo pwq->refcnt = 1; 3665d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36661befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3667d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36688864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3669f147f29eSTejun Heo } 3670d2c1d404STejun Heo 3671f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 36721befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3673f147f29eSTejun Heo { 3674f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3675f147f29eSTejun Heo 3676f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 367775ccf595STejun Heo 36781befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 36791befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 36801befcf30STejun Heo return; 36811befcf30STejun Heo 3682983ca25eSTejun Heo /* 3683983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 36843c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3685983ca25eSTejun Heo */ 368675ccf595STejun Heo pwq->work_color = wq->work_color; 3687983ca25eSTejun Heo 3688983ca25eSTejun Heo /* sync max_active to the current setting */ 3689983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3690983ca25eSTejun Heo 3691983ca25eSTejun Heo /* link in @pwq */ 36929e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3693df2d5ae4STejun Heo } 36946029a918STejun Heo 3695f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3696f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3697f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3698f147f29eSTejun Heo { 3699f147f29eSTejun Heo struct worker_pool *pool; 3700f147f29eSTejun Heo struct pool_workqueue *pwq; 3701f147f29eSTejun Heo 3702f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3703f147f29eSTejun Heo 3704f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3705f147f29eSTejun Heo if (!pool) 3706f147f29eSTejun Heo return NULL; 3707f147f29eSTejun Heo 3708e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3709f147f29eSTejun Heo if (!pwq) { 3710f147f29eSTejun Heo put_unbound_pool(pool); 3711f147f29eSTejun Heo return NULL; 3712f147f29eSTejun Heo } 3713f147f29eSTejun Heo 3714f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3715f147f29eSTejun Heo return pwq; 3716d2c1d404STejun Heo } 3717d2c1d404STejun Heo 37184c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 37194c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 37204c16bd32STejun Heo { 37214c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37224c16bd32STejun Heo 37234c16bd32STejun Heo if (pwq) { 37244c16bd32STejun Heo put_unbound_pool(pwq->pool); 37254c16bd32STejun Heo kfree(pwq); 37264c16bd32STejun Heo } 37274c16bd32STejun Heo } 37284c16bd32STejun Heo 37294c16bd32STejun Heo /** 37304c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 37314c16bd32STejun Heo * @attrs: the wq_attrs of interest 37324c16bd32STejun Heo * @node: the target NUMA node 37334c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 37344c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 37354c16bd32STejun Heo * 37364c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 37374c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 37384c16bd32STejun Heo * calculation. The result is stored in @cpumask. This function returns 37394c16bd32STejun Heo * %true if the resulting @cpumask is different from @attrs->cpumask, 37404c16bd32STejun Heo * %false if equal. 37414c16bd32STejun Heo * 37424c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 37434c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 37444c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 37454c16bd32STejun Heo * @attrs->cpumask. 37464c16bd32STejun Heo * 37474c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 37484c16bd32STejun Heo * stable. 37494c16bd32STejun Heo */ 37504c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 37514c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 37524c16bd32STejun Heo { 37534c16bd32STejun Heo if (!wq_numa_enabled) 37544c16bd32STejun Heo goto use_dfl; 37554c16bd32STejun Heo 37564c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 37574c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 37584c16bd32STejun Heo if (cpu_going_down >= 0) 37594c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 37604c16bd32STejun Heo 37614c16bd32STejun Heo if (cpumask_empty(cpumask)) 37624c16bd32STejun Heo goto use_dfl; 37634c16bd32STejun Heo 37644c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 37654c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 37664c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 37674c16bd32STejun Heo 37684c16bd32STejun Heo use_dfl: 37694c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 37704c16bd32STejun Heo return false; 37714c16bd32STejun Heo } 37724c16bd32STejun Heo 37731befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 37741befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 37751befcf30STejun Heo int node, 37761befcf30STejun Heo struct pool_workqueue *pwq) 37771befcf30STejun Heo { 37781befcf30STejun Heo struct pool_workqueue *old_pwq; 37791befcf30STejun Heo 37801befcf30STejun Heo lockdep_assert_held(&wq->mutex); 37811befcf30STejun Heo 37821befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 37831befcf30STejun Heo link_pwq(pwq); 37841befcf30STejun Heo 37851befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 37861befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 37871befcf30STejun Heo return old_pwq; 37881befcf30STejun Heo } 37891befcf30STejun Heo 37909e8cd2f5STejun Heo /** 37919e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 37929e8cd2f5STejun Heo * @wq: the target workqueue 37939e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 37949e8cd2f5STejun Heo * 37954c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 37964c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 37974c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 37984c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 37994c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38004c16bd32STejun Heo * back-to-back will stay on its current pwq. 38019e8cd2f5STejun Heo * 38029e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 38039e8cd2f5STejun Heo * failure. 38049e8cd2f5STejun Heo */ 38059e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38069e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38079e8cd2f5STejun Heo { 38084c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 38094c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3810f147f29eSTejun Heo int node, ret; 38119e8cd2f5STejun Heo 38128719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38139e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38149e8cd2f5STejun Heo return -EINVAL; 38159e8cd2f5STejun Heo 38168719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38178719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38188719dceaSTejun Heo return -EINVAL; 38198719dceaSTejun Heo 38204c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 382113e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38224c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38234c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 382413e2e556STejun Heo goto enomem; 382513e2e556STejun Heo 38264c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 382713e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 382813e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 382913e2e556STejun Heo 38304c16bd32STejun Heo /* 38314c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 38324c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 38334c16bd32STejun Heo * pools. 38344c16bd32STejun Heo */ 38354c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 38369e8cd2f5STejun Heo 38374c16bd32STejun Heo /* 38384c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 38394c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 38404c16bd32STejun Heo * pwqs accordingly. 38414c16bd32STejun Heo */ 38424c16bd32STejun Heo get_online_cpus(); 38434c16bd32STejun Heo 38444c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 38454c16bd32STejun Heo 38464c16bd32STejun Heo /* 38474c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 38484c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 38494c16bd32STejun Heo * it even if we don't use it immediately. 38504c16bd32STejun Heo */ 38514c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 38524c16bd32STejun Heo if (!dfl_pwq) 38534c16bd32STejun Heo goto enomem_pwq; 38544c16bd32STejun Heo 38554c16bd32STejun Heo for_each_node(node) { 38564c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 38574c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 38584c16bd32STejun Heo if (!pwq_tbl[node]) 38594c16bd32STejun Heo goto enomem_pwq; 38604c16bd32STejun Heo } else { 38614c16bd32STejun Heo dfl_pwq->refcnt++; 38624c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 38634c16bd32STejun Heo } 38644c16bd32STejun Heo } 38654c16bd32STejun Heo 38664c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38674c16bd32STejun Heo 38684c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3869f147f29eSTejun Heo mutex_lock(&wq->mutex); 3870a892caccSTejun Heo 3871f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 38724c16bd32STejun Heo 38734c16bd32STejun Heo /* save the previous pwq and install the new one */ 3874f147f29eSTejun Heo for_each_node(node) 38754c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 38764c16bd32STejun Heo 38774c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 38784c16bd32STejun Heo link_pwq(dfl_pwq); 38794c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3880f147f29eSTejun Heo 3881f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3882f147f29eSTejun Heo 38834c16bd32STejun Heo /* put the old pwqs */ 38844c16bd32STejun Heo for_each_node(node) 38854c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 38864c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 38874c16bd32STejun Heo 38884c16bd32STejun Heo put_online_cpus(); 38894862125bSTejun Heo ret = 0; 38904862125bSTejun Heo /* fall through */ 38914862125bSTejun Heo out_free: 38924c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 38934862125bSTejun Heo free_workqueue_attrs(new_attrs); 38944c16bd32STejun Heo kfree(pwq_tbl); 38954862125bSTejun Heo return ret; 389613e2e556STejun Heo 38974c16bd32STejun Heo enomem_pwq: 38984c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 38994c16bd32STejun Heo for_each_node(node) 39004c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 39014c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 39024c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39034c16bd32STejun Heo put_online_cpus(); 390413e2e556STejun Heo enomem: 39054862125bSTejun Heo ret = -ENOMEM; 39064862125bSTejun Heo goto out_free; 39079e8cd2f5STejun Heo } 39089e8cd2f5STejun Heo 39094c16bd32STejun Heo /** 39104c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 39114c16bd32STejun Heo * @wq: the target workqueue 39124c16bd32STejun Heo * @cpu: the CPU coming up or going down 39134c16bd32STejun Heo * @online: whether @cpu is coming up or going down 39144c16bd32STejun Heo * 39154c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 39164c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 39174c16bd32STejun Heo * @wq accordingly. 39184c16bd32STejun Heo * 39194c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 39204c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 39214c16bd32STejun Heo * correct. 39224c16bd32STejun Heo * 39234c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 39244c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39254c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39264c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39274c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39284c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39294c16bd32STejun Heo * CPU_DOWN_PREPARE. 39304c16bd32STejun Heo */ 39314c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 39324c16bd32STejun Heo bool online) 39334c16bd32STejun Heo { 39344c16bd32STejun Heo int node = cpu_to_node(cpu); 39354c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 39364c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 39374c16bd32STejun Heo struct workqueue_attrs *target_attrs; 39384c16bd32STejun Heo cpumask_t *cpumask; 39394c16bd32STejun Heo 39404c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 39414c16bd32STejun Heo 39424c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 39434c16bd32STejun Heo return; 39444c16bd32STejun Heo 39454c16bd32STejun Heo /* 39464c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39474c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39484c16bd32STejun Heo * CPU hotplug exclusion. 39494c16bd32STejun Heo */ 39504c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39514c16bd32STejun Heo cpumask = target_attrs->cpumask; 39524c16bd32STejun Heo 39534c16bd32STejun Heo mutex_lock(&wq->mutex); 39544c16bd32STejun Heo 39554c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39564c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39574c16bd32STejun Heo 39584c16bd32STejun Heo /* 39594c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 39604c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 39614c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 39624c16bd32STejun Heo * wq's, the default pwq should be used. If @pwq is already the 39634c16bd32STejun Heo * default one, nothing to do; otherwise, install the default one. 39644c16bd32STejun Heo */ 39654c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 39664c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 39674c16bd32STejun Heo goto out_unlock; 39684c16bd32STejun Heo } else { 39694c16bd32STejun Heo if (pwq == wq->dfl_pwq) 39704c16bd32STejun Heo goto out_unlock; 39714c16bd32STejun Heo else 39724c16bd32STejun Heo goto use_dfl_pwq; 39734c16bd32STejun Heo } 39744c16bd32STejun Heo 39754c16bd32STejun Heo mutex_unlock(&wq->mutex); 39764c16bd32STejun Heo 39774c16bd32STejun Heo /* create a new pwq */ 39784c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39794c16bd32STejun Heo if (!pwq) { 39804c16bd32STejun Heo pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39814c16bd32STejun Heo wq->name); 39824c16bd32STejun Heo goto out_unlock; 39834c16bd32STejun Heo } 39844c16bd32STejun Heo 39854c16bd32STejun Heo /* 39864c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 39874c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 39884c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 39894c16bd32STejun Heo * inbetween. 39904c16bd32STejun Heo */ 39914c16bd32STejun Heo mutex_lock(&wq->mutex); 39924c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 39934c16bd32STejun Heo goto out_unlock; 39944c16bd32STejun Heo 39954c16bd32STejun Heo use_dfl_pwq: 39964c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 39974c16bd32STejun Heo get_pwq(wq->dfl_pwq); 39984c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 39994c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 40004c16bd32STejun Heo out_unlock: 40014c16bd32STejun Heo mutex_unlock(&wq->mutex); 40024c16bd32STejun Heo put_pwq_unlocked(old_pwq); 40034c16bd32STejun Heo } 40044c16bd32STejun Heo 400530cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 40061da177e4SLinus Torvalds { 400749e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 400830cdf249STejun Heo int cpu; 4009e1d8aa9fSFrederic Weisbecker 401030cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4011420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4012420c0ddbSTejun Heo if (!wq->cpu_pwqs) 401330cdf249STejun Heo return -ENOMEM; 401430cdf249STejun Heo 401530cdf249STejun Heo for_each_possible_cpu(cpu) { 40167fb98ea7STejun Heo struct pool_workqueue *pwq = 40177fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 40187a62c2c8STejun Heo struct worker_pool *cpu_pools = 4019f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 402030cdf249STejun Heo 4021f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4022f147f29eSTejun Heo 4023f147f29eSTejun Heo mutex_lock(&wq->mutex); 40241befcf30STejun Heo link_pwq(pwq); 4025f147f29eSTejun Heo mutex_unlock(&wq->mutex); 402630cdf249STejun Heo } 402730cdf249STejun Heo return 0; 40289e8cd2f5STejun Heo } else { 40299e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 40309e8cd2f5STejun Heo } 40310f900049STejun Heo } 40320f900049STejun Heo 4033f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4034f3421797STejun Heo const char *name) 4035b71ab8c2STejun Heo { 4036f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4037f3421797STejun Heo 4038f3421797STejun Heo if (max_active < 1 || max_active > lim) 4039044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4040f3421797STejun Heo max_active, name, 1, lim); 4041b71ab8c2STejun Heo 4042f3421797STejun Heo return clamp_val(max_active, 1, lim); 4043b71ab8c2STejun Heo } 4044b71ab8c2STejun Heo 4045b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 404697e37d7bSTejun Heo unsigned int flags, 40471e19ffc6STejun Heo int max_active, 4048eb13ba87SJohannes Berg struct lock_class_key *key, 4049b196be89STejun Heo const char *lock_name, ...) 40503af24433SOleg Nesterov { 4051df2d5ae4STejun Heo size_t tbl_size = 0; 4052ecf6881fSTejun Heo va_list args; 40533af24433SOleg Nesterov struct workqueue_struct *wq; 405449e3cf44STejun Heo struct pool_workqueue *pwq; 4055b196be89STejun Heo 4056ecf6881fSTejun Heo /* allocate wq and format name */ 4057df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4058df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4059df2d5ae4STejun Heo 4060df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4061b196be89STejun Heo if (!wq) 4062d2c1d404STejun Heo return NULL; 4063b196be89STejun Heo 40646029a918STejun Heo if (flags & WQ_UNBOUND) { 40656029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40666029a918STejun Heo if (!wq->unbound_attrs) 40676029a918STejun Heo goto err_free_wq; 40686029a918STejun Heo } 40696029a918STejun Heo 4070ecf6881fSTejun Heo va_start(args, lock_name); 4071ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4072b196be89STejun Heo va_end(args); 40733af24433SOleg Nesterov 4074d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4075b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40763af24433SOleg Nesterov 4077b196be89STejun Heo /* init wq */ 407897e37d7bSTejun Heo wq->flags = flags; 4079a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40803c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4081112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 408230cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 408373f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 408473f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4085493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 40863af24433SOleg Nesterov 4087eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4088cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 40893af24433SOleg Nesterov 409030cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4091d2c1d404STejun Heo goto err_free_wq; 40921537663fSTejun Heo 4093493008a8STejun Heo /* 4094493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4095493008a8STejun Heo * have a rescuer to guarantee forward progress. 4096493008a8STejun Heo */ 4097493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4098e22bee78STejun Heo struct worker *rescuer; 4099e22bee78STejun Heo 4100d2c1d404STejun Heo rescuer = alloc_worker(); 4101e22bee78STejun Heo if (!rescuer) 4102d2c1d404STejun Heo goto err_destroy; 4103e22bee78STejun Heo 4104111c225aSTejun Heo rescuer->rescue_wq = wq; 4105111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4106b196be89STejun Heo wq->name); 4107d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4108d2c1d404STejun Heo kfree(rescuer); 4109d2c1d404STejun Heo goto err_destroy; 4110d2c1d404STejun Heo } 4111e22bee78STejun Heo 4112d2c1d404STejun Heo wq->rescuer = rescuer; 411314a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4114e22bee78STejun Heo wake_up_process(rescuer->task); 41153af24433SOleg Nesterov } 41161537663fSTejun Heo 4117226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4118226223abSTejun Heo goto err_destroy; 4119226223abSTejun Heo 41203af24433SOleg Nesterov /* 412168e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 412268e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 412368e13a67SLai Jiangshan * list. 41243af24433SOleg Nesterov */ 412568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4126a0a1a5fdSTejun Heo 4127a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 412849e3cf44STejun Heo for_each_pwq(pwq, wq) 4129699ce097STejun Heo pwq_adjust_max_active(pwq); 4130a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4131a0a1a5fdSTejun Heo 41323af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4133a0a1a5fdSTejun Heo 413468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41353af24433SOleg Nesterov 41363af24433SOleg Nesterov return wq; 4137d2c1d404STejun Heo 4138d2c1d404STejun Heo err_free_wq: 41396029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41404690c4abSTejun Heo kfree(wq); 4141d2c1d404STejun Heo return NULL; 4142d2c1d404STejun Heo err_destroy: 4143d2c1d404STejun Heo destroy_workqueue(wq); 41444690c4abSTejun Heo return NULL; 41451da177e4SLinus Torvalds } 4146d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41471da177e4SLinus Torvalds 41483af24433SOleg Nesterov /** 41493af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41503af24433SOleg Nesterov * @wq: target workqueue 41513af24433SOleg Nesterov * 41523af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41533af24433SOleg Nesterov */ 41543af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41553af24433SOleg Nesterov { 415649e3cf44STejun Heo struct pool_workqueue *pwq; 41574c16bd32STejun Heo int node; 41583af24433SOleg Nesterov 41599c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41609c5a2ba7STejun Heo drain_workqueue(wq); 4161c8efcc25STejun Heo 41626183c009STejun Heo /* sanity checks */ 4163b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 416449e3cf44STejun Heo for_each_pwq(pwq, wq) { 41656183c009STejun Heo int i; 41666183c009STejun Heo 416776af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 416876af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4169b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41706183c009STejun Heo return; 417176af4d93STejun Heo } 417276af4d93STejun Heo } 417376af4d93STejun Heo 41748864b4e5STejun Heo if (WARN_ON(pwq->refcnt > 1) || 41758864b4e5STejun Heo WARN_ON(pwq->nr_active) || 417676af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4177b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41786183c009STejun Heo return; 41796183c009STejun Heo } 418076af4d93STejun Heo } 4181b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41826183c009STejun Heo 4183a0a1a5fdSTejun Heo /* 4184a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4185a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4186a0a1a5fdSTejun Heo */ 418768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4188d2c1d404STejun Heo list_del_init(&wq->list); 418968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41903af24433SOleg Nesterov 4191226223abSTejun Heo workqueue_sysfs_unregister(wq); 4192226223abSTejun Heo 4193493008a8STejun Heo if (wq->rescuer) { 4194e22bee78STejun Heo kthread_stop(wq->rescuer->task); 41958d9df9f0SXiaotian Feng kfree(wq->rescuer); 4196493008a8STejun Heo wq->rescuer = NULL; 4197e22bee78STejun Heo } 4198e22bee78STejun Heo 41998864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 420029c91e99STejun Heo /* 42018864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 42028864b4e5STejun Heo * free the pwqs and wq. 420329c91e99STejun Heo */ 42048864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 42058864b4e5STejun Heo kfree(wq); 42068864b4e5STejun Heo } else { 42078864b4e5STejun Heo /* 42088864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42094c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42104c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42118864b4e5STejun Heo */ 42124c16bd32STejun Heo for_each_node(node) { 42134c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42144c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42154c16bd32STejun Heo put_pwq_unlocked(pwq); 42164c16bd32STejun Heo } 42174c16bd32STejun Heo 42184c16bd32STejun Heo /* 42194c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42204c16bd32STejun Heo * put. Don't access it afterwards. 42214c16bd32STejun Heo */ 42224c16bd32STejun Heo pwq = wq->dfl_pwq; 42234c16bd32STejun Heo wq->dfl_pwq = NULL; 4224dce90d47STejun Heo put_pwq_unlocked(pwq); 422529c91e99STejun Heo } 42263af24433SOleg Nesterov } 42273af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42283af24433SOleg Nesterov 4229dcd989cbSTejun Heo /** 4230dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4231dcd989cbSTejun Heo * @wq: target workqueue 4232dcd989cbSTejun Heo * @max_active: new max_active value. 4233dcd989cbSTejun Heo * 4234dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4235dcd989cbSTejun Heo * 4236dcd989cbSTejun Heo * CONTEXT: 4237dcd989cbSTejun Heo * Don't call from IRQ context. 4238dcd989cbSTejun Heo */ 4239dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4240dcd989cbSTejun Heo { 424149e3cf44STejun Heo struct pool_workqueue *pwq; 4242dcd989cbSTejun Heo 42438719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42448719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 42458719dceaSTejun Heo return; 42468719dceaSTejun Heo 4247f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4248dcd989cbSTejun Heo 4249a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4250dcd989cbSTejun Heo 4251dcd989cbSTejun Heo wq->saved_max_active = max_active; 4252dcd989cbSTejun Heo 4253699ce097STejun Heo for_each_pwq(pwq, wq) 4254699ce097STejun Heo pwq_adjust_max_active(pwq); 4255dcd989cbSTejun Heo 4256a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4257dcd989cbSTejun Heo } 4258dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4259dcd989cbSTejun Heo 4260dcd989cbSTejun Heo /** 4261e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4262e6267616STejun Heo * 4263e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4264e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4265e6267616STejun Heo */ 4266e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4267e6267616STejun Heo { 4268e6267616STejun Heo struct worker *worker = current_wq_worker(); 4269e6267616STejun Heo 42706a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4271e6267616STejun Heo } 4272e6267616STejun Heo 4273e6267616STejun Heo /** 4274dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4275dcd989cbSTejun Heo * @cpu: CPU in question 4276dcd989cbSTejun Heo * @wq: target workqueue 4277dcd989cbSTejun Heo * 4278dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4279dcd989cbSTejun Heo * no synchronization around this function and the test result is 4280dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4281dcd989cbSTejun Heo * 4282dcd989cbSTejun Heo * RETURNS: 4283dcd989cbSTejun Heo * %true if congested, %false otherwise. 4284dcd989cbSTejun Heo */ 4285d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4286dcd989cbSTejun Heo { 42877fb98ea7STejun Heo struct pool_workqueue *pwq; 428876af4d93STejun Heo bool ret; 428976af4d93STejun Heo 429088109453SLai Jiangshan rcu_read_lock_sched(); 42917fb98ea7STejun Heo 42927fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 42937fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 42947fb98ea7STejun Heo else 4295df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4296dcd989cbSTejun Heo 429776af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 429888109453SLai Jiangshan rcu_read_unlock_sched(); 429976af4d93STejun Heo 430076af4d93STejun Heo return ret; 4301dcd989cbSTejun Heo } 4302dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4303dcd989cbSTejun Heo 4304dcd989cbSTejun Heo /** 4305dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4306dcd989cbSTejun Heo * @work: the work to be tested 4307dcd989cbSTejun Heo * 4308dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4309dcd989cbSTejun Heo * synchronization around this function and the test result is 4310dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4311dcd989cbSTejun Heo * 4312dcd989cbSTejun Heo * RETURNS: 4313dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4314dcd989cbSTejun Heo */ 4315dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4316dcd989cbSTejun Heo { 4317fa1b54e6STejun Heo struct worker_pool *pool; 4318dcd989cbSTejun Heo unsigned long flags; 4319dcd989cbSTejun Heo unsigned int ret = 0; 4320dcd989cbSTejun Heo 4321dcd989cbSTejun Heo if (work_pending(work)) 4322dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4323038366c5SLai Jiangshan 4324fa1b54e6STejun Heo local_irq_save(flags); 4325fa1b54e6STejun Heo pool = get_work_pool(work); 4326038366c5SLai Jiangshan if (pool) { 4327fa1b54e6STejun Heo spin_lock(&pool->lock); 4328c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4329dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4330fa1b54e6STejun Heo spin_unlock(&pool->lock); 4331038366c5SLai Jiangshan } 4332fa1b54e6STejun Heo local_irq_restore(flags); 4333dcd989cbSTejun Heo 4334dcd989cbSTejun Heo return ret; 4335dcd989cbSTejun Heo } 4336dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4337dcd989cbSTejun Heo 4338db7bccf4STejun Heo /* 4339db7bccf4STejun Heo * CPU hotplug. 4340db7bccf4STejun Heo * 4341e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4342112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4343706026c2STejun Heo * pool which make migrating pending and scheduled works very 4344e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 434594cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4346e22bee78STejun Heo * blocked draining impractical. 4347e22bee78STejun Heo * 434824647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4349628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4350628c78e7STejun Heo * cpu comes back online. 4351db7bccf4STejun Heo */ 4352db7bccf4STejun Heo 4353706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4354db7bccf4STejun Heo { 435538db41d9STejun Heo int cpu = smp_processor_id(); 43564ce62e9eSTejun Heo struct worker_pool *pool; 4357db7bccf4STejun Heo struct worker *worker; 4358a9ab775bSTejun Heo int wi; 4359db7bccf4STejun Heo 4360f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 43616183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4362db7bccf4STejun Heo 4363bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 436494cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4365e22bee78STejun Heo 4366f2d5a0eeSTejun Heo /* 4367bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 436894cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 436994cf58bbSTejun Heo * except for the ones which are still executing works from 437094cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 437194cf58bbSTejun Heo * this, they may become diasporas. 4372f2d5a0eeSTejun Heo */ 4373a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4374403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4375db7bccf4STejun Heo 437624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4377f2d5a0eeSTejun Heo 437894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4379bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 438094cf58bbSTejun Heo } 4381e22bee78STejun Heo 4382e22bee78STejun Heo /* 4383628c78e7STejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 4384628c78e7STejun Heo * sched callbacks see the %WORKER_UNBOUND flag. This is necessary 4385628c78e7STejun Heo * as scheduler callbacks may be invoked from other cpus. 4386628c78e7STejun Heo */ 4387628c78e7STejun Heo schedule(); 4388628c78e7STejun Heo 4389628c78e7STejun Heo /* 4390628c78e7STejun Heo * Sched callbacks are disabled now. Zap nr_running. After this, 4391628c78e7STejun Heo * nr_running stays zero and need_more_worker() and keep_working() 439238db41d9STejun Heo * are always true as long as the worklist is not empty. Pools on 439338db41d9STejun Heo * @cpu now behave as unbound (in terms of concurrency management) 439438db41d9STejun Heo * pools which are served by workers tied to the CPU. 4395628c78e7STejun Heo * 4396628c78e7STejun Heo * On return from this function, the current worker would trigger 4397628c78e7STejun Heo * unbound chain execution of pending work items if other workers 4398628c78e7STejun Heo * didn't already. 4399e22bee78STejun Heo */ 4400f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) 4401e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4402db7bccf4STejun Heo } 4403db7bccf4STejun Heo 4404bd7c089eSTejun Heo /** 4405bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4406bd7c089eSTejun Heo * @pool: pool of interest 4407bd7c089eSTejun Heo * 4408a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4409bd7c089eSTejun Heo */ 4410bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4411bd7c089eSTejun Heo { 4412a9ab775bSTejun Heo struct worker *worker; 4413a9ab775bSTejun Heo int wi; 4414bd7c089eSTejun Heo 4415bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4416bd7c089eSTejun Heo 4417bd7c089eSTejun Heo /* 4418a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4419a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4420a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4421a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4422a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4423bd7c089eSTejun Heo */ 4424a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4425a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4426a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4427a9ab775bSTejun Heo 4428a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4429a9ab775bSTejun Heo 4430a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4431a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4432a9ab775bSTejun Heo 4433a9ab775bSTejun Heo /* 4434a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4435a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4436a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4437a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4438a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4439a9ab775bSTejun Heo * be bound before @pool->lock is released. 4440a9ab775bSTejun Heo */ 4441a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4442bd7c089eSTejun Heo wake_up_process(worker->task); 4443bd7c089eSTejun Heo 4444bd7c089eSTejun Heo /* 4445a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4446a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4447a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4448a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4449a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4450a9ab775bSTejun Heo * concurrency management. Note that when or whether 4451a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4452a9ab775bSTejun Heo * 4453a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4454a9ab775bSTejun Heo * tested without holding any lock in 4455a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4456a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4457a9ab775bSTejun Heo * management operations. 4458bd7c089eSTejun Heo */ 4459a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4460a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4461a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4462a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4463bd7c089eSTejun Heo } 4464a9ab775bSTejun Heo 4465a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4466bd7c089eSTejun Heo } 4467bd7c089eSTejun Heo 44687dbc725eSTejun Heo /** 44697dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 44707dbc725eSTejun Heo * @pool: unbound pool of interest 44717dbc725eSTejun Heo * @cpu: the CPU which is coming up 44727dbc725eSTejun Heo * 44737dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 44747dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 44757dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 44767dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 44777dbc725eSTejun Heo */ 44787dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 44797dbc725eSTejun Heo { 44807dbc725eSTejun Heo static cpumask_t cpumask; 44817dbc725eSTejun Heo struct worker *worker; 44827dbc725eSTejun Heo int wi; 44837dbc725eSTejun Heo 44847dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 44857dbc725eSTejun Heo 44867dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 44877dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 44887dbc725eSTejun Heo return; 44897dbc725eSTejun Heo 44907dbc725eSTejun Heo /* is @cpu the only online CPU? */ 44917dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 44927dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 44937dbc725eSTejun Heo return; 44947dbc725eSTejun Heo 44957dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 44967dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 44977dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 44987dbc725eSTejun Heo pool->attrs->cpumask) < 0); 44997dbc725eSTejun Heo } 45007dbc725eSTejun Heo 45018db25e78STejun Heo /* 45028db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 45038db25e78STejun Heo * This will be registered high priority CPU notifier. 45048db25e78STejun Heo */ 45059fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 45061da177e4SLinus Torvalds unsigned long action, 45071da177e4SLinus Torvalds void *hcpu) 45081da177e4SLinus Torvalds { 4509d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45104ce62e9eSTejun Heo struct worker_pool *pool; 45114c16bd32STejun Heo struct workqueue_struct *wq; 45127dbc725eSTejun Heo int pi; 45131da177e4SLinus Torvalds 45148db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 45153af24433SOleg Nesterov case CPU_UP_PREPARE: 4516f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 45173ce63377STejun Heo if (pool->nr_workers) 45183ce63377STejun Heo continue; 4519ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 45203ce63377STejun Heo return NOTIFY_BAD; 45213af24433SOleg Nesterov } 45221da177e4SLinus Torvalds break; 45231da177e4SLinus Torvalds 452465758202STejun Heo case CPU_DOWN_FAILED: 452565758202STejun Heo case CPU_ONLINE: 452668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 45277dbc725eSTejun Heo 45287dbc725eSTejun Heo for_each_pool(pool, pi) { 4529bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 453094cf58bbSTejun Heo 45317dbc725eSTejun Heo if (pool->cpu == cpu) { 4532a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 453324647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4534a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4535a9ab775bSTejun Heo 453694cf58bbSTejun Heo rebind_workers(pool); 45377dbc725eSTejun Heo } else if (pool->cpu < 0) { 45387dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 45397dbc725eSTejun Heo } 454094cf58bbSTejun Heo 4541bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 454294cf58bbSTejun Heo } 45437dbc725eSTejun Heo 45444c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45454c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45464c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 45474c16bd32STejun Heo 454868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45498db25e78STejun Heo break; 455065758202STejun Heo } 455165758202STejun Heo return NOTIFY_OK; 455265758202STejun Heo } 455365758202STejun Heo 455465758202STejun Heo /* 455565758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 455665758202STejun Heo * This will be registered as low priority CPU notifier. 455765758202STejun Heo */ 45589fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 455965758202STejun Heo unsigned long action, 456065758202STejun Heo void *hcpu) 456165758202STejun Heo { 4562d84ff051STejun Heo int cpu = (unsigned long)hcpu; 45638db25e78STejun Heo struct work_struct unbind_work; 45644c16bd32STejun Heo struct workqueue_struct *wq; 45658db25e78STejun Heo 456665758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 456765758202STejun Heo case CPU_DOWN_PREPARE: 45684c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4569706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 45707635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 45714c16bd32STejun Heo 45724c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 45734c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 45744c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 45754c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 45764c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 45774c16bd32STejun Heo 45784c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 45798db25e78STejun Heo flush_work(&unbind_work); 45808db25e78STejun Heo break; 458165758202STejun Heo } 458265758202STejun Heo return NOTIFY_OK; 458365758202STejun Heo } 458465758202STejun Heo 45852d3854a3SRusty Russell #ifdef CONFIG_SMP 45868ccad40dSRusty Russell 45872d3854a3SRusty Russell struct work_for_cpu { 4588ed48ece2STejun Heo struct work_struct work; 45892d3854a3SRusty Russell long (*fn)(void *); 45902d3854a3SRusty Russell void *arg; 45912d3854a3SRusty Russell long ret; 45922d3854a3SRusty Russell }; 45932d3854a3SRusty Russell 4594ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 45952d3854a3SRusty Russell { 4596ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4597ed48ece2STejun Heo 45982d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 45992d3854a3SRusty Russell } 46002d3854a3SRusty Russell 46012d3854a3SRusty Russell /** 46022d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 46032d3854a3SRusty Russell * @cpu: the cpu to run on 46042d3854a3SRusty Russell * @fn: the function to run 46052d3854a3SRusty Russell * @arg: the function arg 46062d3854a3SRusty Russell * 460731ad9081SRusty Russell * This will return the value @fn returns. 460831ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 46096b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 46102d3854a3SRusty Russell */ 4611d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 46122d3854a3SRusty Russell { 4613ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 46142d3854a3SRusty Russell 4615ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4616ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4617ed48ece2STejun Heo flush_work(&wfc.work); 46182d3854a3SRusty Russell return wfc.ret; 46192d3854a3SRusty Russell } 46202d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 46212d3854a3SRusty Russell #endif /* CONFIG_SMP */ 46222d3854a3SRusty Russell 4623a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4624e7577c50SRusty Russell 4625a0a1a5fdSTejun Heo /** 4626a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4627a0a1a5fdSTejun Heo * 462858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4629c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4630706026c2STejun Heo * pool->worklist. 4631a0a1a5fdSTejun Heo * 4632a0a1a5fdSTejun Heo * CONTEXT: 4633a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4634a0a1a5fdSTejun Heo */ 4635a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4636a0a1a5fdSTejun Heo { 463717116969STejun Heo struct worker_pool *pool; 463824b8a847STejun Heo struct workqueue_struct *wq; 463924b8a847STejun Heo struct pool_workqueue *pwq; 4640611c92a0STejun Heo int pi; 4641a0a1a5fdSTejun Heo 464268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4643a0a1a5fdSTejun Heo 46446183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4645a0a1a5fdSTejun Heo workqueue_freezing = true; 4646a0a1a5fdSTejun Heo 464724b8a847STejun Heo /* set FREEZING */ 4648611c92a0STejun Heo for_each_pool(pool, pi) { 46495bcab335STejun Heo spin_lock_irq(&pool->lock); 465035b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 465135b6bb63STejun Heo pool->flags |= POOL_FREEZING; 46525bcab335STejun Heo spin_unlock_irq(&pool->lock); 46531da177e4SLinus Torvalds } 46548b03ae3cSTejun Heo 465524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4656a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4657699ce097STejun Heo for_each_pwq(pwq, wq) 4658699ce097STejun Heo pwq_adjust_max_active(pwq); 4659a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4660a1056305STejun Heo } 46615bcab335STejun Heo 466268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4663a0a1a5fdSTejun Heo } 4664a0a1a5fdSTejun Heo 4665a0a1a5fdSTejun Heo /** 466658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4667a0a1a5fdSTejun Heo * 4668a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4669a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4670a0a1a5fdSTejun Heo * 4671a0a1a5fdSTejun Heo * CONTEXT: 467268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4673a0a1a5fdSTejun Heo * 4674a0a1a5fdSTejun Heo * RETURNS: 467558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 467658a69cb4STejun Heo * is complete. 4677a0a1a5fdSTejun Heo */ 4678a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4679a0a1a5fdSTejun Heo { 4680a0a1a5fdSTejun Heo bool busy = false; 468124b8a847STejun Heo struct workqueue_struct *wq; 468224b8a847STejun Heo struct pool_workqueue *pwq; 4683a0a1a5fdSTejun Heo 468468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4685a0a1a5fdSTejun Heo 46866183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4687a0a1a5fdSTejun Heo 468824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 468924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 469024b8a847STejun Heo continue; 4691a0a1a5fdSTejun Heo /* 4692a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4693a0a1a5fdSTejun Heo * to peek without lock. 4694a0a1a5fdSTejun Heo */ 469588109453SLai Jiangshan rcu_read_lock_sched(); 469624b8a847STejun Heo for_each_pwq(pwq, wq) { 46976183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4698112202d9STejun Heo if (pwq->nr_active) { 4699a0a1a5fdSTejun Heo busy = true; 470088109453SLai Jiangshan rcu_read_unlock_sched(); 4701a0a1a5fdSTejun Heo goto out_unlock; 4702a0a1a5fdSTejun Heo } 4703a0a1a5fdSTejun Heo } 470488109453SLai Jiangshan rcu_read_unlock_sched(); 4705a0a1a5fdSTejun Heo } 4706a0a1a5fdSTejun Heo out_unlock: 470768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4708a0a1a5fdSTejun Heo return busy; 4709a0a1a5fdSTejun Heo } 4710a0a1a5fdSTejun Heo 4711a0a1a5fdSTejun Heo /** 4712a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4713a0a1a5fdSTejun Heo * 4714a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4715706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4716a0a1a5fdSTejun Heo * 4717a0a1a5fdSTejun Heo * CONTEXT: 4718a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4719a0a1a5fdSTejun Heo */ 4720a0a1a5fdSTejun Heo void thaw_workqueues(void) 4721a0a1a5fdSTejun Heo { 472224b8a847STejun Heo struct workqueue_struct *wq; 472324b8a847STejun Heo struct pool_workqueue *pwq; 472424b8a847STejun Heo struct worker_pool *pool; 4725611c92a0STejun Heo int pi; 4726a0a1a5fdSTejun Heo 472768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4728a0a1a5fdSTejun Heo 4729a0a1a5fdSTejun Heo if (!workqueue_freezing) 4730a0a1a5fdSTejun Heo goto out_unlock; 4731a0a1a5fdSTejun Heo 473224b8a847STejun Heo /* clear FREEZING */ 4733611c92a0STejun Heo for_each_pool(pool, pi) { 47345bcab335STejun Heo spin_lock_irq(&pool->lock); 473535b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 473635b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 47375bcab335STejun Heo spin_unlock_irq(&pool->lock); 4738d565ed63STejun Heo } 473924b8a847STejun Heo 474024b8a847STejun Heo /* restore max_active and repopulate worklist */ 474124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4742a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4743699ce097STejun Heo for_each_pwq(pwq, wq) 4744699ce097STejun Heo pwq_adjust_max_active(pwq); 4745a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 474624b8a847STejun Heo } 474724b8a847STejun Heo 4748a0a1a5fdSTejun Heo workqueue_freezing = false; 4749a0a1a5fdSTejun Heo out_unlock: 475068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4751a0a1a5fdSTejun Heo } 4752a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4753a0a1a5fdSTejun Heo 4754bce90380STejun Heo static void __init wq_numa_init(void) 4755bce90380STejun Heo { 4756bce90380STejun Heo cpumask_var_t *tbl; 4757bce90380STejun Heo int node, cpu; 4758bce90380STejun Heo 4759bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4760bce90380STejun Heo for_each_node(node) 4761bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4762bce90380STejun Heo 4763bce90380STejun Heo if (num_possible_nodes() <= 1) 4764bce90380STejun Heo return; 4765bce90380STejun Heo 47664c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 47674c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 47684c16bd32STejun Heo 4769bce90380STejun Heo /* 4770bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4771bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4772bce90380STejun Heo * fully initialized by now. 4773bce90380STejun Heo */ 4774bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4775bce90380STejun Heo BUG_ON(!tbl); 4776bce90380STejun Heo 4777bce90380STejun Heo for_each_node(node) 4778bce90380STejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node)); 4779bce90380STejun Heo 4780bce90380STejun Heo for_each_possible_cpu(cpu) { 4781bce90380STejun Heo node = cpu_to_node(cpu); 4782bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4783bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4784bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4785bce90380STejun Heo return; 4786bce90380STejun Heo } 4787bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4788bce90380STejun Heo } 4789bce90380STejun Heo 4790bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4791bce90380STejun Heo wq_numa_enabled = true; 4792bce90380STejun Heo } 4793bce90380STejun Heo 47946ee0578bSSuresh Siddha static int __init init_workqueues(void) 47951da177e4SLinus Torvalds { 47967a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 47977a4e344cSTejun Heo int i, cpu; 4798c34056a3STejun Heo 47997c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 48007c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 48016be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4802b5490077STejun Heo 4803e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4804e904e6c2STejun Heo 4805e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4806e904e6c2STejun Heo 480765758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4808a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 48098b03ae3cSTejun Heo 4810bce90380STejun Heo wq_numa_init(); 4811bce90380STejun Heo 4812706026c2STejun Heo /* initialize CPU pools */ 481329c91e99STejun Heo for_each_possible_cpu(cpu) { 48144ce62e9eSTejun Heo struct worker_pool *pool; 48158b03ae3cSTejun Heo 48167a4e344cSTejun Heo i = 0; 4817f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 48187a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4819ec22ca5eSTejun Heo pool->cpu = cpu; 48207a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 48217a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4822f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 48237a4e344cSTejun Heo 48249daf9e67STejun Heo /* alloc pool ID */ 482568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 48269daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 482768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 48284ce62e9eSTejun Heo } 48298b03ae3cSTejun Heo } 48308b03ae3cSTejun Heo 4831e22bee78STejun Heo /* create the initial worker */ 483229c91e99STejun Heo for_each_online_cpu(cpu) { 48334ce62e9eSTejun Heo struct worker_pool *pool; 4834e22bee78STejun Heo 4835f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 483624647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4837ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4838e22bee78STejun Heo } 48394ce62e9eSTejun Heo } 4840e22bee78STejun Heo 484129c91e99STejun Heo /* create default unbound wq attrs */ 484229c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 484329c91e99STejun Heo struct workqueue_attrs *attrs; 484429c91e99STejun Heo 484529c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 484629c91e99STejun Heo attrs->nice = std_nice[i]; 484729c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 484829c91e99STejun Heo } 484929c91e99STejun Heo 4850d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 48511aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4852d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4853f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4854f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 485524d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 485624d51addSTejun Heo WQ_FREEZABLE, 0); 48571aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4858ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 48596ee0578bSSuresh Siddha return 0; 48601da177e4SLinus Torvalds } 48616ee0578bSSuresh Siddha early_initcall(init_workqueues); 4862