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> 493d1cb205STejun Heo #include <linux/uaccess.h> 50e22bee78STejun Heo 51ea138446STejun Heo #include "workqueue_internal.h" 521da177e4SLinus Torvalds 53c8e55f36STejun Heo enum { 54bc2ae0f5STejun Heo /* 5524647570STejun Heo * worker_pool flags 56bc2ae0f5STejun Heo * 5724647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 58bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 59bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 60bc2ae0f5STejun Heo * is in effect. 61bc2ae0f5STejun Heo * 62bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 63bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6424647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 65bc2ae0f5STejun Heo * 66bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 67bc3a1afcSTejun Heo * manager_mutex to avoid changing binding state while 6824647570STejun Heo * create_worker() is in progress. 69bc2ae0f5STejun Heo */ 7011ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 7124647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 7235b6bb63STejun Heo POOL_FREEZING = 1 << 3, /* freeze in progress */ 73db7bccf4STejun Heo 74c8e55f36STejun Heo /* worker flags */ 75c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 76c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 77c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 78e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 79fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 80f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 81a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 82e22bee78STejun Heo 83a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 84a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 85db7bccf4STejun Heo 86e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 874ce62e9eSTejun Heo 8829c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 89c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 90db7bccf4STejun Heo 91e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 92e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 93e22bee78STejun Heo 943233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 953233cdbdSTejun Heo /* call for help after 10ms 963233cdbdSTejun Heo (min two ticks) */ 97e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 98e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 991da177e4SLinus Torvalds 1001da177e4SLinus Torvalds /* 101e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 102e22bee78STejun Heo * all cpus. Give -20. 103e22bee78STejun Heo */ 104e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1053270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 106ecf6881fSTejun Heo 107ecf6881fSTejun Heo WQ_NAME_LEN = 24, 108c8e55f36STejun Heo }; 109c8e55f36STejun Heo 1101da177e4SLinus Torvalds /* 1114690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1124690c4abSTejun Heo * 113e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 114e41e704bSTejun Heo * everyone else. 1154690c4abSTejun Heo * 116e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 117e22bee78STejun Heo * only be modified and accessed from the local cpu. 118e22bee78STejun Heo * 119d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1204690c4abSTejun Heo * 121d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 122d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 123d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 124d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 125e22bee78STejun Heo * 126822d8405STejun Heo * MG: pool->manager_mutex and pool->lock protected. Writes require both 127822d8405STejun Heo * locks. Reads can happen under either lock. 128822d8405STejun Heo * 12968e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13076af4d93STejun Heo * 13168e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1325bcab335STejun Heo * 1333c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1343c25a55dSLai Jiangshan * 135b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1362e109a28STejun Heo * 1372e109a28STejun Heo * MD: wq_mayday_lock protected. 1384690c4abSTejun Heo */ 1394690c4abSTejun Heo 1402eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 141c34056a3STejun Heo 142bd7bdd43STejun Heo struct worker_pool { 143d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 144d84ff051STejun Heo int cpu; /* I: the associated cpu */ 145f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1469daf9e67STejun Heo int id; /* I: pool ID */ 14711ebea50STejun Heo unsigned int flags; /* X: flags */ 148bd7bdd43STejun Heo 149bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 150bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 151ea1abd61SLai Jiangshan 152ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 153bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 154bd7bdd43STejun Heo 155bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 156bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 157bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 158bd7bdd43STejun Heo 159c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 160c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 161c9e7cf27STejun Heo /* L: hash of busy workers */ 162c9e7cf27STejun Heo 163bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 16434a06bd6STejun Heo struct mutex manager_arb; /* manager arbitration */ 165bc3a1afcSTejun Heo struct mutex manager_mutex; /* manager exclusion */ 166822d8405STejun Heo struct idr worker_idr; /* MG: worker IDs and iteration */ 167e19e397aSTejun Heo 1687a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 16968e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17068e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1717a4e344cSTejun Heo 172e19e397aSTejun Heo /* 173e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 174e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 175e19e397aSTejun Heo * cacheline. 176e19e397aSTejun Heo */ 177e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 17829c91e99STejun Heo 17929c91e99STejun Heo /* 18029c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 18129c91e99STejun Heo * from get_work_pool(). 18229c91e99STejun Heo */ 18329c91e99STejun Heo struct rcu_head rcu; 1848b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1858b03ae3cSTejun Heo 1868b03ae3cSTejun Heo /* 187112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 188112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 189112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 190112202d9STejun Heo * number of flag bits. 1911da177e4SLinus Torvalds */ 192112202d9STejun Heo struct pool_workqueue { 193bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 1944690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 19573f53c4aSTejun Heo int work_color; /* L: current color */ 19673f53c4aSTejun Heo int flush_color; /* L: flushing color */ 1978864b4e5STejun Heo int refcnt; /* L: reference count */ 19873f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 19973f53c4aSTejun Heo /* L: nr of in_flight works */ 2001e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 201a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2021e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2033c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2042e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2058864b4e5STejun Heo 2068864b4e5STejun Heo /* 2078864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2088864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2098864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 210b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2118864b4e5STejun Heo */ 2128864b4e5STejun Heo struct work_struct unbound_release_work; 2138864b4e5STejun Heo struct rcu_head rcu; 214e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* 21773f53c4aSTejun Heo * Structure used to wait for workqueue flush. 21873f53c4aSTejun Heo */ 21973f53c4aSTejun Heo struct wq_flusher { 2203c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2213c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 22273f53c4aSTejun Heo struct completion done; /* flush completion */ 22373f53c4aSTejun Heo }; 2241da177e4SLinus Torvalds 225226223abSTejun Heo struct wq_device; 226226223abSTejun Heo 22773f53c4aSTejun Heo /* 228c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 229c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2301da177e4SLinus Torvalds */ 2311da177e4SLinus Torvalds struct workqueue_struct { 2323c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 23368e13a67SLai Jiangshan struct list_head list; /* PL: list of all workqueues */ 23473f53c4aSTejun Heo 2353c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2363c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2373c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 238112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2393c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2403c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2413c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 24273f53c4aSTejun Heo 2432e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 244e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 245e22bee78STejun Heo 24687fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 247a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 248226223abSTejun Heo 2496029a918STejun Heo struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */ 2504c16bd32STejun Heo struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */ 2516029a918STejun Heo 252226223abSTejun Heo #ifdef CONFIG_SYSFS 253226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 254226223abSTejun Heo #endif 2554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2564e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2574e6045f1SJohannes Berg #endif 258ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2592728fd2fSTejun Heo 2602728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2612728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2622728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 263df2d5ae4STejun Heo struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */ 2641da177e4SLinus Torvalds }; 2651da177e4SLinus Torvalds 266e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 267e904e6c2STejun Heo 268bce90380STejun Heo static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */ 269bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 270bce90380STejun Heo /* possible CPUs of each node */ 271bce90380STejun Heo 272d55262c4STejun Heo static bool wq_disable_numa; 273d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 274d55262c4STejun Heo 275bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 276bce90380STejun Heo 2774c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2784c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 2794c16bd32STejun Heo 28068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 2812e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 2825bcab335STejun Heo 28368e13a67SLai Jiangshan static LIST_HEAD(workqueues); /* PL: list of all workqueues */ 28468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 2857d19c5ceSTejun Heo 2867d19c5ceSTejun Heo /* the per-cpu worker pools */ 2877d19c5ceSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], 2887d19c5ceSTejun Heo cpu_worker_pools); 2897d19c5ceSTejun Heo 29068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 2917d19c5ceSTejun Heo 29268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 29329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 29429c91e99STejun Heo 295c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 29629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 29729c91e99STejun Heo 298d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 299ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 300044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3011aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 302044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 303d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 304044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 305f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 306044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 30724d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 308d320c038STejun Heo 3097d19c5ceSTejun Heo static int worker_thread(void *__worker); 3107d19c5ceSTejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 3117d19c5ceSTejun Heo const struct workqueue_attrs *from); 3127d19c5ceSTejun Heo 31397bd2347STejun Heo #define CREATE_TRACE_POINTS 31497bd2347STejun Heo #include <trace/events/workqueue.h> 31597bd2347STejun Heo 31668e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 3175bcab335STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 31868e13a67SLai Jiangshan lockdep_is_held(&wq_pool_mutex), \ 31968e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3205bcab335STejun Heo 321b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 32276af4d93STejun Heo rcu_lockdep_assert(rcu_read_lock_sched_held() || \ 323b5927605SLai Jiangshan lockdep_is_held(&wq->mutex), \ 324b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 32576af4d93STejun Heo 326822d8405STejun Heo #ifdef CONFIG_LOCKDEP 327822d8405STejun Heo #define assert_manager_or_pool_lock(pool) \ 328519e3c11SLai Jiangshan WARN_ONCE(debug_locks && \ 329519e3c11SLai Jiangshan !lockdep_is_held(&(pool)->manager_mutex) && \ 330822d8405STejun Heo !lockdep_is_held(&(pool)->lock), \ 331822d8405STejun Heo "pool->manager_mutex or ->lock should be held") 332822d8405STejun Heo #else 333822d8405STejun Heo #define assert_manager_or_pool_lock(pool) do { } while (0) 334822d8405STejun Heo #endif 335822d8405STejun Heo 336f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 337f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 338f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3397a62c2c8STejun Heo (pool)++) 3404ce62e9eSTejun Heo 34149e3cf44STejun Heo /** 34217116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 34317116969STejun Heo * @pool: iteration cursor 344611c92a0STejun Heo * @pi: integer used for iteration 345fa1b54e6STejun Heo * 34668e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 34768e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 34868e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 349fa1b54e6STejun Heo * 350fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 351fa1b54e6STejun Heo * ignored. 35217116969STejun Heo */ 353611c92a0STejun Heo #define for_each_pool(pool, pi) \ 354611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 35568e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 356fa1b54e6STejun Heo else 35717116969STejun Heo 35817116969STejun Heo /** 359822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 360822d8405STejun Heo * @worker: iteration cursor 361822d8405STejun Heo * @wi: integer used for iteration 362822d8405STejun Heo * @pool: worker_pool to iterate workers of 363822d8405STejun Heo * 364822d8405STejun Heo * This must be called with either @pool->manager_mutex or ->lock held. 365822d8405STejun Heo * 366822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 367822d8405STejun Heo * ignored. 368822d8405STejun Heo */ 369822d8405STejun Heo #define for_each_pool_worker(worker, wi, pool) \ 370822d8405STejun Heo idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \ 371822d8405STejun Heo if (({ assert_manager_or_pool_lock((pool)); false; })) { } \ 372822d8405STejun Heo else 373822d8405STejun Heo 374822d8405STejun Heo /** 37549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 37649e3cf44STejun Heo * @pwq: iteration cursor 37749e3cf44STejun Heo * @wq: the target workqueue 37876af4d93STejun Heo * 379b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 380794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 381794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 38276af4d93STejun Heo * 38376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 38476af4d93STejun Heo * ignored. 38549e3cf44STejun Heo */ 38649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 38776af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 388b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 38976af4d93STejun Heo else 390f3421797STejun Heo 391dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 392dc186ad7SThomas Gleixner 393dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 394dc186ad7SThomas Gleixner 39599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 39699777288SStanislaw Gruszka { 39799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 39899777288SStanislaw Gruszka } 39999777288SStanislaw Gruszka 400dc186ad7SThomas Gleixner /* 401dc186ad7SThomas Gleixner * fixup_init is called when: 402dc186ad7SThomas Gleixner * - an active object is initialized 403dc186ad7SThomas Gleixner */ 404dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 405dc186ad7SThomas Gleixner { 406dc186ad7SThomas Gleixner struct work_struct *work = addr; 407dc186ad7SThomas Gleixner 408dc186ad7SThomas Gleixner switch (state) { 409dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 410dc186ad7SThomas Gleixner cancel_work_sync(work); 411dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 412dc186ad7SThomas Gleixner return 1; 413dc186ad7SThomas Gleixner default: 414dc186ad7SThomas Gleixner return 0; 415dc186ad7SThomas Gleixner } 416dc186ad7SThomas Gleixner } 417dc186ad7SThomas Gleixner 418dc186ad7SThomas Gleixner /* 419dc186ad7SThomas Gleixner * fixup_activate is called when: 420dc186ad7SThomas Gleixner * - an active object is activated 421dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 422dc186ad7SThomas Gleixner */ 423dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 424dc186ad7SThomas Gleixner { 425dc186ad7SThomas Gleixner struct work_struct *work = addr; 426dc186ad7SThomas Gleixner 427dc186ad7SThomas Gleixner switch (state) { 428dc186ad7SThomas Gleixner 429dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 430dc186ad7SThomas Gleixner /* 431dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 432dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 433dc186ad7SThomas Gleixner * is tracked in the object tracker. 434dc186ad7SThomas Gleixner */ 43522df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 436dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 437dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 438dc186ad7SThomas Gleixner return 0; 439dc186ad7SThomas Gleixner } 440dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 441dc186ad7SThomas Gleixner return 0; 442dc186ad7SThomas Gleixner 443dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 444dc186ad7SThomas Gleixner WARN_ON(1); 445dc186ad7SThomas Gleixner 446dc186ad7SThomas Gleixner default: 447dc186ad7SThomas Gleixner return 0; 448dc186ad7SThomas Gleixner } 449dc186ad7SThomas Gleixner } 450dc186ad7SThomas Gleixner 451dc186ad7SThomas Gleixner /* 452dc186ad7SThomas Gleixner * fixup_free is called when: 453dc186ad7SThomas Gleixner * - an active object is freed 454dc186ad7SThomas Gleixner */ 455dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 456dc186ad7SThomas Gleixner { 457dc186ad7SThomas Gleixner struct work_struct *work = addr; 458dc186ad7SThomas Gleixner 459dc186ad7SThomas Gleixner switch (state) { 460dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 461dc186ad7SThomas Gleixner cancel_work_sync(work); 462dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 463dc186ad7SThomas Gleixner return 1; 464dc186ad7SThomas Gleixner default: 465dc186ad7SThomas Gleixner return 0; 466dc186ad7SThomas Gleixner } 467dc186ad7SThomas Gleixner } 468dc186ad7SThomas Gleixner 469dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 470dc186ad7SThomas Gleixner .name = "work_struct", 47199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 472dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 473dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 474dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 475dc186ad7SThomas Gleixner }; 476dc186ad7SThomas Gleixner 477dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 478dc186ad7SThomas Gleixner { 479dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner 482dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 483dc186ad7SThomas Gleixner { 484dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 485dc186ad7SThomas Gleixner } 486dc186ad7SThomas Gleixner 487dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 488dc186ad7SThomas Gleixner { 489dc186ad7SThomas Gleixner if (onstack) 490dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 491dc186ad7SThomas Gleixner else 492dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 493dc186ad7SThomas Gleixner } 494dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 495dc186ad7SThomas Gleixner 496dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 497dc186ad7SThomas Gleixner { 498dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 499dc186ad7SThomas Gleixner } 500dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 501dc186ad7SThomas Gleixner 502dc186ad7SThomas Gleixner #else 503dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 504dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 505dc186ad7SThomas Gleixner #endif 506dc186ad7SThomas Gleixner 5079daf9e67STejun Heo /* allocate ID and assign it to @pool */ 5089daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5099daf9e67STejun Heo { 5109daf9e67STejun Heo int ret; 5119daf9e67STejun Heo 51268e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5135bcab335STejun Heo 514e68035fbSTejun Heo ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL); 515229641a6STejun Heo if (ret >= 0) { 516e68035fbSTejun Heo pool->id = ret; 517229641a6STejun Heo return 0; 518229641a6STejun Heo } 5199daf9e67STejun Heo return ret; 5209daf9e67STejun Heo } 5219daf9e67STejun Heo 52276af4d93STejun Heo /** 523df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 524df2d5ae4STejun Heo * @wq: the target workqueue 525df2d5ae4STejun Heo * @node: the node ID 526df2d5ae4STejun Heo * 527df2d5ae4STejun Heo * This must be called either with pwq_lock held or sched RCU read locked. 528df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 529df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 530df2d5ae4STejun Heo */ 531df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 532df2d5ae4STejun Heo int node) 533df2d5ae4STejun Heo { 534df2d5ae4STejun Heo assert_rcu_or_wq_mutex(wq); 535df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 536df2d5ae4STejun Heo } 537df2d5ae4STejun Heo 53873f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 53973f53c4aSTejun Heo { 54073f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 54173f53c4aSTejun Heo } 54273f53c4aSTejun Heo 54373f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 54473f53c4aSTejun Heo { 54573f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 54673f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 54773f53c4aSTejun Heo } 54873f53c4aSTejun Heo 54973f53c4aSTejun Heo static int work_next_color(int color) 55073f53c4aSTejun Heo { 55173f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 552a848e3b6SOleg Nesterov } 553a848e3b6SOleg Nesterov 5544594bf15SDavid Howells /* 555112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 556112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5577c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5587a22ad75STejun Heo * 559112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 560112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 561bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 562bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 5637a22ad75STejun Heo * 564112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 5657c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 566112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 5677c3eed5cSTejun Heo * available only while the work item is queued. 568bbb68dfaSTejun Heo * 569bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 570bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 571bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 572bbb68dfaSTejun Heo * try to steal the PENDING bit. 5734594bf15SDavid Howells */ 5747a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5757a22ad75STejun Heo unsigned long flags) 5767a22ad75STejun Heo { 5776183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 5787a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5797a22ad75STejun Heo } 5807a22ad75STejun Heo 581112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 5824690c4abSTejun Heo unsigned long extra_flags) 583365970a1SDavid Howells { 584112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 585112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 586365970a1SDavid Howells } 587365970a1SDavid Howells 5884468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 5894468a00fSLai Jiangshan int pool_id) 5904468a00fSLai Jiangshan { 5914468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 5924468a00fSLai Jiangshan WORK_STRUCT_PENDING); 5934468a00fSLai Jiangshan } 5944468a00fSLai Jiangshan 5957c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 5967c3eed5cSTejun Heo int pool_id) 5974d707b9fSOleg Nesterov { 59823657bb1STejun Heo /* 59923657bb1STejun Heo * The following wmb is paired with the implied mb in 60023657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 60123657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 60223657bb1STejun Heo * owner. 60323657bb1STejun Heo */ 60423657bb1STejun Heo smp_wmb(); 6057c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 6064d707b9fSOleg Nesterov } 6074d707b9fSOleg Nesterov 6087a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 609365970a1SDavid Howells { 6107c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6117c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6127a22ad75STejun Heo } 6137a22ad75STejun Heo 614112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6157a22ad75STejun Heo { 616e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6177a22ad75STejun Heo 618112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 619e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 620e120153dSTejun Heo else 621e120153dSTejun Heo return NULL; 6227a22ad75STejun Heo } 6237a22ad75STejun Heo 6247c3eed5cSTejun Heo /** 6257c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6267c3eed5cSTejun Heo * @work: the work item of interest 6277c3eed5cSTejun Heo * 6287c3eed5cSTejun Heo * Return the worker_pool @work was last associated with. %NULL if none. 629fa1b54e6STejun Heo * 63068e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 63168e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 63268e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 633fa1b54e6STejun Heo * 634fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 635fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 636fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 637fa1b54e6STejun Heo * returned pool is and stays online. 6387c3eed5cSTejun Heo */ 6397c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 6407a22ad75STejun Heo { 641e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6427c3eed5cSTejun Heo int pool_id; 6437a22ad75STejun Heo 64468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 645fa1b54e6STejun Heo 646112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 647112202d9STejun Heo return ((struct pool_workqueue *) 6487c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 6497a22ad75STejun Heo 6507c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 6517c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 6527a22ad75STejun Heo return NULL; 6537a22ad75STejun Heo 654fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 6557c3eed5cSTejun Heo } 6567c3eed5cSTejun Heo 6577c3eed5cSTejun Heo /** 6587c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 6597c3eed5cSTejun Heo * @work: the work item of interest 6607c3eed5cSTejun Heo * 6617c3eed5cSTejun Heo * Return the worker_pool ID @work was last associated with. 6627c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 6637c3eed5cSTejun Heo */ 6647c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 6657c3eed5cSTejun Heo { 66654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 6677c3eed5cSTejun Heo 668112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 669112202d9STejun Heo return ((struct pool_workqueue *) 67054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 67154d5b7d0SLai Jiangshan 67254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 6737c3eed5cSTejun Heo } 6747c3eed5cSTejun Heo 675bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 676bbb68dfaSTejun Heo { 6777c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 678bbb68dfaSTejun Heo 6797c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 6807c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 681bbb68dfaSTejun Heo } 682bbb68dfaSTejun Heo 683bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 684bbb68dfaSTejun Heo { 685bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 686bbb68dfaSTejun Heo 687112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 688bbb68dfaSTejun Heo } 689bbb68dfaSTejun Heo 690e22bee78STejun Heo /* 6913270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6923270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 693d565ed63STejun Heo * they're being called with pool->lock held. 694e22bee78STejun Heo */ 695e22bee78STejun Heo 69663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 697649027d7STejun Heo { 698e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 699649027d7STejun Heo } 700649027d7STejun Heo 701e22bee78STejun Heo /* 702e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 703e22bee78STejun Heo * running workers. 704974271c4STejun Heo * 705974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 706706026c2STejun Heo * function will always return %true for unbound pools as long as the 707974271c4STejun Heo * worklist isn't empty. 708e22bee78STejun Heo */ 70963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 710e22bee78STejun Heo { 71163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 712e22bee78STejun Heo } 713e22bee78STejun Heo 714e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 71563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 716e22bee78STejun Heo { 71763d95a91STejun Heo return pool->nr_idle; 718e22bee78STejun Heo } 719e22bee78STejun Heo 720e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 72163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 722e22bee78STejun Heo { 723e19e397aSTejun Heo return !list_empty(&pool->worklist) && 724e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 725e22bee78STejun Heo } 726e22bee78STejun Heo 727e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 72863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 729e22bee78STejun Heo { 73063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 731e22bee78STejun Heo } 732e22bee78STejun Heo 733e22bee78STejun Heo /* Do I need to be the manager? */ 73463d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 735e22bee78STejun Heo { 73663d95a91STejun Heo return need_to_create_worker(pool) || 73711ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 738e22bee78STejun Heo } 739e22bee78STejun Heo 740e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 74163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 742e22bee78STejun Heo { 74334a06bd6STejun Heo bool managing = mutex_is_locked(&pool->manager_arb); 74463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 74563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 746e22bee78STejun Heo 747ea1abd61SLai Jiangshan /* 748ea1abd61SLai Jiangshan * nr_idle and idle_list may disagree if idle rebinding is in 749ea1abd61SLai Jiangshan * progress. Never return %true if idle_list is empty. 750ea1abd61SLai Jiangshan */ 751ea1abd61SLai Jiangshan if (list_empty(&pool->idle_list)) 752ea1abd61SLai Jiangshan return false; 753ea1abd61SLai Jiangshan 754e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 755e22bee78STejun Heo } 756e22bee78STejun Heo 757e22bee78STejun Heo /* 758e22bee78STejun Heo * Wake up functions. 759e22bee78STejun Heo */ 760e22bee78STejun Heo 7617e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 76263d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 7637e11629dSTejun Heo { 76463d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 7657e11629dSTejun Heo return NULL; 7667e11629dSTejun Heo 76763d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 7687e11629dSTejun Heo } 7697e11629dSTejun Heo 7707e11629dSTejun Heo /** 7717e11629dSTejun Heo * wake_up_worker - wake up an idle worker 77263d95a91STejun Heo * @pool: worker pool to wake worker from 7737e11629dSTejun Heo * 77463d95a91STejun Heo * Wake up the first idle worker of @pool. 7757e11629dSTejun Heo * 7767e11629dSTejun Heo * CONTEXT: 777d565ed63STejun Heo * spin_lock_irq(pool->lock). 7787e11629dSTejun Heo */ 77963d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 7807e11629dSTejun Heo { 78163d95a91STejun Heo struct worker *worker = first_worker(pool); 7827e11629dSTejun Heo 7837e11629dSTejun Heo if (likely(worker)) 7847e11629dSTejun Heo wake_up_process(worker->task); 7857e11629dSTejun Heo } 7867e11629dSTejun Heo 7874690c4abSTejun Heo /** 788e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 789e22bee78STejun Heo * @task: task waking up 790e22bee78STejun Heo * @cpu: CPU @task is waking up to 791e22bee78STejun Heo * 792e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 793e22bee78STejun Heo * being awoken. 794e22bee78STejun Heo * 795e22bee78STejun Heo * CONTEXT: 796e22bee78STejun Heo * spin_lock_irq(rq->lock) 797e22bee78STejun Heo */ 798d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 799e22bee78STejun Heo { 800e22bee78STejun Heo struct worker *worker = kthread_data(task); 801e22bee78STejun Heo 80236576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 803ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 804e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 805e22bee78STejun Heo } 80636576000SJoonsoo Kim } 807e22bee78STejun Heo 808e22bee78STejun Heo /** 809e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 810e22bee78STejun Heo * @task: task going to sleep 811e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 812e22bee78STejun Heo * 813e22bee78STejun Heo * This function is called during schedule() when a busy worker is 814e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 815e22bee78STejun Heo * returning pointer to its task. 816e22bee78STejun Heo * 817e22bee78STejun Heo * CONTEXT: 818e22bee78STejun Heo * spin_lock_irq(rq->lock) 819e22bee78STejun Heo * 820e22bee78STejun Heo * RETURNS: 821e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 822e22bee78STejun Heo */ 823d84ff051STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu) 824e22bee78STejun Heo { 825e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 826111c225aSTejun Heo struct worker_pool *pool; 827e22bee78STejun Heo 828111c225aSTejun Heo /* 829111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 830111c225aSTejun Heo * workers, also reach here, let's not access anything before 831111c225aSTejun Heo * checking NOT_RUNNING. 832111c225aSTejun Heo */ 8332d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 834e22bee78STejun Heo return NULL; 835e22bee78STejun Heo 836111c225aSTejun Heo pool = worker->pool; 837111c225aSTejun Heo 838e22bee78STejun Heo /* this can only happen on the local cpu */ 8396183c009STejun Heo if (WARN_ON_ONCE(cpu != raw_smp_processor_id())) 8406183c009STejun Heo return NULL; 841e22bee78STejun Heo 842e22bee78STejun Heo /* 843e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 844e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 845e22bee78STejun Heo * Please read comment there. 846e22bee78STejun Heo * 847628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 848628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 849628c78e7STejun Heo * disabled, which in turn means that none else could be 850d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 851628c78e7STejun Heo * lock is safe. 852e22bee78STejun Heo */ 853e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 854e19e397aSTejun Heo !list_empty(&pool->worklist)) 85563d95a91STejun Heo to_wakeup = first_worker(pool); 856e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 857e22bee78STejun Heo } 858e22bee78STejun Heo 859e22bee78STejun Heo /** 860e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 861cb444766STejun Heo * @worker: self 862d302f017STejun Heo * @flags: flags to set 863d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 864d302f017STejun Heo * 865e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 866e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 867e22bee78STejun Heo * woken up. 868d302f017STejun Heo * 869cb444766STejun Heo * CONTEXT: 870d565ed63STejun Heo * spin_lock_irq(pool->lock) 871d302f017STejun Heo */ 872d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 873d302f017STejun Heo bool wakeup) 874d302f017STejun Heo { 875bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 876e22bee78STejun Heo 877cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 878cb444766STejun Heo 879e22bee78STejun Heo /* 880e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 881e22bee78STejun Heo * wake up an idle worker as necessary if requested by 882e22bee78STejun Heo * @wakeup. 883e22bee78STejun Heo */ 884e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 885e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 886e22bee78STejun Heo if (wakeup) { 887e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 888bd7bdd43STejun Heo !list_empty(&pool->worklist)) 88963d95a91STejun Heo wake_up_worker(pool); 890e22bee78STejun Heo } else 891e19e397aSTejun Heo atomic_dec(&pool->nr_running); 892e22bee78STejun Heo } 893e22bee78STejun Heo 894d302f017STejun Heo worker->flags |= flags; 895d302f017STejun Heo } 896d302f017STejun Heo 897d302f017STejun Heo /** 898e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 899cb444766STejun Heo * @worker: self 900d302f017STejun Heo * @flags: flags to clear 901d302f017STejun Heo * 902e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 903d302f017STejun Heo * 904cb444766STejun Heo * CONTEXT: 905d565ed63STejun Heo * spin_lock_irq(pool->lock) 906d302f017STejun Heo */ 907d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 908d302f017STejun Heo { 90963d95a91STejun Heo struct worker_pool *pool = worker->pool; 910e22bee78STejun Heo unsigned int oflags = worker->flags; 911e22bee78STejun Heo 912cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 913cb444766STejun Heo 914d302f017STejun Heo worker->flags &= ~flags; 915e22bee78STejun Heo 91642c025f3STejun Heo /* 91742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 91842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 91942c025f3STejun Heo * of multiple flags, not a single flag. 92042c025f3STejun Heo */ 921e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 922e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 923e19e397aSTejun Heo atomic_inc(&pool->nr_running); 924d302f017STejun Heo } 925d302f017STejun Heo 926d302f017STejun Heo /** 9278cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 928c9e7cf27STejun Heo * @pool: pool of interest 9298cca0eeaSTejun Heo * @work: work to find worker for 9308cca0eeaSTejun Heo * 931c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 932c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 933a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 934a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 935a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 936a2c1c57bSTejun Heo * being executed. 937a2c1c57bSTejun Heo * 938a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 939a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 940a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 941a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 942a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 943a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 944a2c1c57bSTejun Heo * 945c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 946c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 947c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 948c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 949c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 950c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9518cca0eeaSTejun Heo * 9528cca0eeaSTejun Heo * CONTEXT: 953d565ed63STejun Heo * spin_lock_irq(pool->lock). 9548cca0eeaSTejun Heo * 9558cca0eeaSTejun Heo * RETURNS: 9568cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9578cca0eeaSTejun Heo * otherwise. 9588cca0eeaSTejun Heo */ 959c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 9608cca0eeaSTejun Heo struct work_struct *work) 9618cca0eeaSTejun Heo { 96242f8570fSSasha Levin struct worker *worker; 96342f8570fSSasha Levin 964b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 965a2c1c57bSTejun Heo (unsigned long)work) 966a2c1c57bSTejun Heo if (worker->current_work == work && 967a2c1c57bSTejun Heo worker->current_func == work->func) 96842f8570fSSasha Levin return worker; 96942f8570fSSasha Levin 97042f8570fSSasha Levin return NULL; 9718cca0eeaSTejun Heo } 9728cca0eeaSTejun Heo 9738cca0eeaSTejun Heo /** 974bf4ede01STejun Heo * move_linked_works - move linked works to a list 975bf4ede01STejun Heo * @work: start of series of works to be scheduled 976bf4ede01STejun Heo * @head: target list to append @work to 977bf4ede01STejun Heo * @nextp: out paramter for nested worklist walking 978bf4ede01STejun Heo * 979bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 980bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 981bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 982bf4ede01STejun Heo * 983bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 984bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 985bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 986bf4ede01STejun Heo * 987bf4ede01STejun Heo * CONTEXT: 988d565ed63STejun Heo * spin_lock_irq(pool->lock). 989bf4ede01STejun Heo */ 990bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 991bf4ede01STejun Heo struct work_struct **nextp) 992bf4ede01STejun Heo { 993bf4ede01STejun Heo struct work_struct *n; 994bf4ede01STejun Heo 995bf4ede01STejun Heo /* 996bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 997bf4ede01STejun Heo * use NULL for list head. 998bf4ede01STejun Heo */ 999bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1000bf4ede01STejun Heo list_move_tail(&work->entry, head); 1001bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1002bf4ede01STejun Heo break; 1003bf4ede01STejun Heo } 1004bf4ede01STejun Heo 1005bf4ede01STejun Heo /* 1006bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1007bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1008bf4ede01STejun Heo * needs to be updated. 1009bf4ede01STejun Heo */ 1010bf4ede01STejun Heo if (nextp) 1011bf4ede01STejun Heo *nextp = n; 1012bf4ede01STejun Heo } 1013bf4ede01STejun Heo 10148864b4e5STejun Heo /** 10158864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10168864b4e5STejun Heo * @pwq: pool_workqueue to get 10178864b4e5STejun Heo * 10188864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10198864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10208864b4e5STejun Heo */ 10218864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10228864b4e5STejun Heo { 10238864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10248864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10258864b4e5STejun Heo pwq->refcnt++; 10268864b4e5STejun Heo } 10278864b4e5STejun Heo 10288864b4e5STejun Heo /** 10298864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10308864b4e5STejun Heo * @pwq: pool_workqueue to put 10318864b4e5STejun Heo * 10328864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10338864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10348864b4e5STejun Heo */ 10358864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10368864b4e5STejun Heo { 10378864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10388864b4e5STejun Heo if (likely(--pwq->refcnt)) 10398864b4e5STejun Heo return; 10408864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10418864b4e5STejun Heo return; 10428864b4e5STejun Heo /* 10438864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10448864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10458864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10468864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10478864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10488864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10498864b4e5STejun Heo */ 10508864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10518864b4e5STejun Heo } 10528864b4e5STejun Heo 1053dce90d47STejun Heo /** 1054dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1055dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1056dce90d47STejun Heo * 1057dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1058dce90d47STejun Heo */ 1059dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1060dce90d47STejun Heo { 1061dce90d47STejun Heo if (pwq) { 1062dce90d47STejun Heo /* 1063dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1064dce90d47STejun Heo * following lock operations are safe. 1065dce90d47STejun Heo */ 1066dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1067dce90d47STejun Heo put_pwq(pwq); 1068dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1069dce90d47STejun Heo } 1070dce90d47STejun Heo } 1071dce90d47STejun Heo 1072112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1073bf4ede01STejun Heo { 1074112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1075bf4ede01STejun Heo 1076bf4ede01STejun Heo trace_workqueue_activate_work(work); 1077112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1078bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1079112202d9STejun Heo pwq->nr_active++; 1080bf4ede01STejun Heo } 1081bf4ede01STejun Heo 1082112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 10833aa62497SLai Jiangshan { 1084112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 10853aa62497SLai Jiangshan struct work_struct, entry); 10863aa62497SLai Jiangshan 1087112202d9STejun Heo pwq_activate_delayed_work(work); 10883aa62497SLai Jiangshan } 10893aa62497SLai Jiangshan 1090bf4ede01STejun Heo /** 1091112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1092112202d9STejun Heo * @pwq: pwq of interest 1093bf4ede01STejun Heo * @color: color of work which left the queue 1094bf4ede01STejun Heo * 1095bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1096112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1097bf4ede01STejun Heo * 1098bf4ede01STejun Heo * CONTEXT: 1099d565ed63STejun Heo * spin_lock_irq(pool->lock). 1100bf4ede01STejun Heo */ 1101112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1102bf4ede01STejun Heo { 11038864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1104bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11058864b4e5STejun Heo goto out_put; 1106bf4ede01STejun Heo 1107112202d9STejun Heo pwq->nr_in_flight[color]--; 1108bf4ede01STejun Heo 1109112202d9STejun Heo pwq->nr_active--; 1110112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1111bf4ede01STejun Heo /* one down, submit a delayed one */ 1112112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1113112202d9STejun Heo pwq_activate_first_delayed(pwq); 1114bf4ede01STejun Heo } 1115bf4ede01STejun Heo 1116bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1117112202d9STejun Heo if (likely(pwq->flush_color != color)) 11188864b4e5STejun Heo goto out_put; 1119bf4ede01STejun Heo 1120bf4ede01STejun Heo /* are there still in-flight works? */ 1121112202d9STejun Heo if (pwq->nr_in_flight[color]) 11228864b4e5STejun Heo goto out_put; 1123bf4ede01STejun Heo 1124112202d9STejun Heo /* this pwq is done, clear flush_color */ 1125112202d9STejun Heo pwq->flush_color = -1; 1126bf4ede01STejun Heo 1127bf4ede01STejun Heo /* 1128112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1129bf4ede01STejun Heo * will handle the rest. 1130bf4ede01STejun Heo */ 1131112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1132112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11338864b4e5STejun Heo out_put: 11348864b4e5STejun Heo put_pwq(pwq); 1135bf4ede01STejun Heo } 1136bf4ede01STejun Heo 113736e227d2STejun Heo /** 1138bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 113936e227d2STejun Heo * @work: work item to steal 114036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1141bbb68dfaSTejun Heo * @flags: place to store irq state 114236e227d2STejun Heo * 114336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 114436e227d2STejun Heo * stable state - idle, on timer or on worklist. Return values are 114536e227d2STejun Heo * 114636e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 114736e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 114836e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1149bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1150bbb68dfaSTejun Heo * for arbitrarily long 115136e227d2STejun Heo * 1152bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1153e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1154e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1155e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1156bbb68dfaSTejun Heo * 1157bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1158bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1159bbb68dfaSTejun Heo * 1160e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1161bf4ede01STejun Heo */ 1162bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1163bbb68dfaSTejun Heo unsigned long *flags) 1164bf4ede01STejun Heo { 1165d565ed63STejun Heo struct worker_pool *pool; 1166112202d9STejun Heo struct pool_workqueue *pwq; 1167bf4ede01STejun Heo 1168bbb68dfaSTejun Heo local_irq_save(*flags); 1169bbb68dfaSTejun Heo 117036e227d2STejun Heo /* try to steal the timer if it exists */ 117136e227d2STejun Heo if (is_dwork) { 117236e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 117336e227d2STejun Heo 1174e0aecdd8STejun Heo /* 1175e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1176e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1177e0aecdd8STejun Heo * running on the local CPU. 1178e0aecdd8STejun Heo */ 117936e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 118036e227d2STejun Heo return 1; 118136e227d2STejun Heo } 118236e227d2STejun Heo 118336e227d2STejun Heo /* try to claim PENDING the normal way */ 1184bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1185bf4ede01STejun Heo return 0; 1186bf4ede01STejun Heo 1187bf4ede01STejun Heo /* 1188bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1189bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1190bf4ede01STejun Heo */ 1191d565ed63STejun Heo pool = get_work_pool(work); 1192d565ed63STejun Heo if (!pool) 1193bbb68dfaSTejun Heo goto fail; 1194bf4ede01STejun Heo 1195d565ed63STejun Heo spin_lock(&pool->lock); 1196bf4ede01STejun Heo /* 1197112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1198112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1199112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1200112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1201112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12020b3dae68SLai Jiangshan * item is currently queued on that pool. 1203bf4ede01STejun Heo */ 1204112202d9STejun Heo pwq = get_work_pwq(work); 1205112202d9STejun Heo if (pwq && pwq->pool == pool) { 1206bf4ede01STejun Heo debug_work_deactivate(work); 12073aa62497SLai Jiangshan 12083aa62497SLai Jiangshan /* 120916062836STejun Heo * A delayed work item cannot be grabbed directly because 121016062836STejun Heo * it might have linked NO_COLOR work items which, if left 1211112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 121216062836STejun Heo * management later on and cause stall. Make sure the work 121316062836STejun Heo * item is activated before grabbing. 12143aa62497SLai Jiangshan */ 12153aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1216112202d9STejun Heo pwq_activate_delayed_work(work); 12173aa62497SLai Jiangshan 1218bf4ede01STejun Heo list_del_init(&work->entry); 1219112202d9STejun Heo pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work)); 122036e227d2STejun Heo 1221112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12224468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12234468a00fSLai Jiangshan 1224d565ed63STejun Heo spin_unlock(&pool->lock); 122536e227d2STejun Heo return 1; 1226bf4ede01STejun Heo } 1227d565ed63STejun Heo spin_unlock(&pool->lock); 1228bbb68dfaSTejun Heo fail: 1229bbb68dfaSTejun Heo local_irq_restore(*flags); 1230bbb68dfaSTejun Heo if (work_is_canceling(work)) 1231bbb68dfaSTejun Heo return -ENOENT; 1232bbb68dfaSTejun Heo cpu_relax(); 123336e227d2STejun Heo return -EAGAIN; 1234bf4ede01STejun Heo } 1235bf4ede01STejun Heo 1236bf4ede01STejun Heo /** 1237706026c2STejun Heo * insert_work - insert a work into a pool 1238112202d9STejun Heo * @pwq: pwq @work belongs to 12394690c4abSTejun Heo * @work: work to insert 12404690c4abSTejun Heo * @head: insertion point 12414690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12424690c4abSTejun Heo * 1243112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1244706026c2STejun Heo * work_struct flags. 12454690c4abSTejun Heo * 12464690c4abSTejun Heo * CONTEXT: 1247d565ed63STejun Heo * spin_lock_irq(pool->lock). 1248365970a1SDavid Howells */ 1249112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1250112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1251b89deed3SOleg Nesterov { 1252112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1253e1d8aa9fSFrederic Weisbecker 12544690c4abSTejun Heo /* we own @work, set data and link */ 1255112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 12561a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 12578864b4e5STejun Heo get_pwq(pwq); 1258e22bee78STejun Heo 1259e22bee78STejun Heo /* 1260c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1261c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1262c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1263e22bee78STejun Heo */ 1264e22bee78STejun Heo smp_mb(); 1265e22bee78STejun Heo 126663d95a91STejun Heo if (__need_more_worker(pool)) 126763d95a91STejun Heo wake_up_worker(pool); 1268b89deed3SOleg Nesterov } 1269b89deed3SOleg Nesterov 1270c8efcc25STejun Heo /* 1271c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 12728d03ecfeSTejun Heo * same workqueue. 1273c8efcc25STejun Heo */ 1274c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1275c8efcc25STejun Heo { 1276c8efcc25STejun Heo struct worker *worker; 1277c8efcc25STejun Heo 12788d03ecfeSTejun Heo worker = current_wq_worker(); 1279c8efcc25STejun Heo /* 12808d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 12818d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1282c8efcc25STejun Heo */ 1283112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1284c8efcc25STejun Heo } 1285c8efcc25STejun Heo 1286d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 12871da177e4SLinus Torvalds struct work_struct *work) 12881da177e4SLinus Torvalds { 1289112202d9STejun Heo struct pool_workqueue *pwq; 1290c9178087STejun Heo struct worker_pool *last_pool; 12911e19ffc6STejun Heo struct list_head *worklist; 12928a2e8e5dSTejun Heo unsigned int work_flags; 1293b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 12948930cabaSTejun Heo 12958930cabaSTejun Heo /* 12968930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 12978930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 12988930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 12998930cabaSTejun Heo * happen with IRQ disabled. 13008930cabaSTejun Heo */ 13018930cabaSTejun Heo WARN_ON_ONCE(!irqs_disabled()); 13021da177e4SLinus Torvalds 1303dc186ad7SThomas Gleixner debug_work_activate(work); 13041e19ffc6STejun Heo 1305c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 1306618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1307c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1308e41e704bSTejun Heo return; 13099e8cd2f5STejun Heo retry: 1310df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1311f3421797STejun Heo cpu = raw_smp_processor_id(); 1312df2d5ae4STejun Heo 1313df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1314df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1315c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1316df2d5ae4STejun Heo else 1317df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1318f3421797STejun Heo 131918aa9effSTejun Heo /* 1320c9178087STejun Heo * If @work was previously on a different pool, it might still be 1321c9178087STejun Heo * running there, in which case the work needs to be queued on that 1322c9178087STejun Heo * pool to guarantee non-reentrancy. 132318aa9effSTejun Heo */ 1324c9e7cf27STejun Heo last_pool = get_work_pool(work); 1325112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 132618aa9effSTejun Heo struct worker *worker; 132718aa9effSTejun Heo 1328d565ed63STejun Heo spin_lock(&last_pool->lock); 132918aa9effSTejun Heo 1330c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 133118aa9effSTejun Heo 1332112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1333c9178087STejun Heo pwq = worker->current_pwq; 13348594fadeSLai Jiangshan } else { 133518aa9effSTejun Heo /* meh... not running there, queue here */ 1336d565ed63STejun Heo spin_unlock(&last_pool->lock); 1337112202d9STejun Heo spin_lock(&pwq->pool->lock); 133818aa9effSTejun Heo } 13398930cabaSTejun Heo } else { 1340112202d9STejun Heo spin_lock(&pwq->pool->lock); 13418930cabaSTejun Heo } 1342502ca9d8STejun Heo 13439e8cd2f5STejun Heo /* 13449e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 13459e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 13469e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1347df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1348df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 13499e8cd2f5STejun Heo * make forward-progress. 13509e8cd2f5STejun Heo */ 13519e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 13529e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 13539e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 13549e8cd2f5STejun Heo cpu_relax(); 13559e8cd2f5STejun Heo goto retry; 13569e8cd2f5STejun Heo } 13579e8cd2f5STejun Heo /* oops */ 13589e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 13599e8cd2f5STejun Heo wq->name, cpu); 13609e8cd2f5STejun Heo } 13619e8cd2f5STejun Heo 1362112202d9STejun Heo /* pwq determined, queue */ 1363112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1364502ca9d8STejun Heo 1365f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1366112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1367f5b2552bSDan Carpenter return; 1368f5b2552bSDan Carpenter } 13691e19ffc6STejun Heo 1370112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1371112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 13721e19ffc6STejun Heo 1373112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1374cdadf009STejun Heo trace_workqueue_activate_work(work); 1375112202d9STejun Heo pwq->nr_active++; 1376112202d9STejun Heo worklist = &pwq->pool->worklist; 13778a2e8e5dSTejun Heo } else { 13788a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1379112202d9STejun Heo worklist = &pwq->delayed_works; 13808a2e8e5dSTejun Heo } 13811e19ffc6STejun Heo 1382112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 13831e19ffc6STejun Heo 1384112202d9STejun Heo spin_unlock(&pwq->pool->lock); 13851da177e4SLinus Torvalds } 13861da177e4SLinus Torvalds 13870fcb78c2SRolf Eike Beer /** 1388c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1389c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1390c1a220e7SZhang Rui * @wq: workqueue to use 1391c1a220e7SZhang Rui * @work: work to queue 1392c1a220e7SZhang Rui * 1393d4283e93STejun Heo * Returns %false if @work was already on a queue, %true otherwise. 1394c1a220e7SZhang Rui * 1395c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1396c1a220e7SZhang Rui * can't go away. 1397c1a220e7SZhang Rui */ 1398d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1399d4283e93STejun Heo struct work_struct *work) 1400c1a220e7SZhang Rui { 1401d4283e93STejun Heo bool ret = false; 14028930cabaSTejun Heo unsigned long flags; 14038930cabaSTejun Heo 14048930cabaSTejun Heo local_irq_save(flags); 1405c1a220e7SZhang Rui 140622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14074690c4abSTejun Heo __queue_work(cpu, wq, work); 1408d4283e93STejun Heo ret = true; 1409c1a220e7SZhang Rui } 14108930cabaSTejun Heo 14118930cabaSTejun Heo local_irq_restore(flags); 1412c1a220e7SZhang Rui return ret; 1413c1a220e7SZhang Rui } 1414ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1415c1a220e7SZhang Rui 1416d8e794dfSTejun Heo void delayed_work_timer_fn(unsigned long __data) 14171da177e4SLinus Torvalds { 141852bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 14191da177e4SLinus Torvalds 1420e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 142160c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 14221da177e4SLinus Torvalds } 14231438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 14241da177e4SLinus Torvalds 14257beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 142652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14271da177e4SLinus Torvalds { 14287beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 14297beb2edfSTejun Heo struct work_struct *work = &dwork->work; 14301da177e4SLinus Torvalds 14317beb2edfSTejun Heo WARN_ON_ONCE(timer->function != delayed_work_timer_fn || 14327beb2edfSTejun Heo timer->data != (unsigned long)dwork); 1433fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1434fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 14357beb2edfSTejun Heo 14368852aac2STejun Heo /* 14378852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 14388852aac2STejun Heo * both optimization and correctness. The earliest @timer can 14398852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 14408852aac2STejun Heo * on that there's no such delay when @delay is 0. 14418852aac2STejun Heo */ 14428852aac2STejun Heo if (!delay) { 14438852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 14448852aac2STejun Heo return; 14458852aac2STejun Heo } 14468852aac2STejun Heo 14477beb2edfSTejun Heo timer_stats_timer_set_start_info(&dwork->timer); 14487beb2edfSTejun Heo 144960c057bcSLai Jiangshan dwork->wq = wq; 14501265057fSTejun Heo dwork->cpu = cpu; 14517beb2edfSTejun Heo timer->expires = jiffies + delay; 14527beb2edfSTejun Heo 14537beb2edfSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 14547beb2edfSTejun Heo add_timer_on(timer, cpu); 14557beb2edfSTejun Heo else 14567beb2edfSTejun Heo add_timer(timer); 14577beb2edfSTejun Heo } 14581da177e4SLinus Torvalds 14590fcb78c2SRolf Eike Beer /** 14600fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 14610fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 14620fcb78c2SRolf Eike Beer * @wq: workqueue to use 1463af9997e4SRandy Dunlap * @dwork: work to queue 14640fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 14650fcb78c2SRolf Eike Beer * 1466715f1300STejun Heo * Returns %false if @work was already on a queue, %true otherwise. If 1467715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1468715f1300STejun Heo * execution. 14690fcb78c2SRolf Eike Beer */ 1470d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 147152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 14727a6bc1cdSVenkatesh Pallipadi { 147352bad64dSDavid Howells struct work_struct *work = &dwork->work; 1474d4283e93STejun Heo bool ret = false; 14758930cabaSTejun Heo unsigned long flags; 14768930cabaSTejun Heo 14778930cabaSTejun Heo /* read the comment in __queue_work() */ 14788930cabaSTejun Heo local_irq_save(flags); 14797a6bc1cdSVenkatesh Pallipadi 148022df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14817beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1482d4283e93STejun Heo ret = true; 14837a6bc1cdSVenkatesh Pallipadi } 14848930cabaSTejun Heo 14858930cabaSTejun Heo local_irq_restore(flags); 14867a6bc1cdSVenkatesh Pallipadi return ret; 14877a6bc1cdSVenkatesh Pallipadi } 1488ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 14891da177e4SLinus Torvalds 1490c8e55f36STejun Heo /** 14918376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 14928376fe22STejun Heo * @cpu: CPU number to execute work on 14938376fe22STejun Heo * @wq: workqueue to use 14948376fe22STejun Heo * @dwork: work to queue 14958376fe22STejun Heo * @delay: number of jiffies to wait before queueing 14968376fe22STejun Heo * 14978376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 14988376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 14998376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15008376fe22STejun Heo * current state. 15018376fe22STejun Heo * 15028376fe22STejun Heo * Returns %false if @dwork was idle and queued, %true if @dwork was 15038376fe22STejun Heo * pending and its timer was modified. 15048376fe22STejun Heo * 1505e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15068376fe22STejun Heo * See try_to_grab_pending() for details. 15078376fe22STejun Heo */ 15088376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15098376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15108376fe22STejun Heo { 15118376fe22STejun Heo unsigned long flags; 15128376fe22STejun Heo int ret; 15138376fe22STejun Heo 15148376fe22STejun Heo do { 15158376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15168376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15178376fe22STejun Heo 15188376fe22STejun Heo if (likely(ret >= 0)) { 15198376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15208376fe22STejun Heo local_irq_restore(flags); 15218376fe22STejun Heo } 15228376fe22STejun Heo 15238376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 15248376fe22STejun Heo return ret; 15258376fe22STejun Heo } 15268376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 15278376fe22STejun Heo 15288376fe22STejun Heo /** 1529c8e55f36STejun Heo * worker_enter_idle - enter idle state 1530c8e55f36STejun Heo * @worker: worker which is entering idle state 1531c8e55f36STejun Heo * 1532c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1533c8e55f36STejun Heo * necessary. 1534c8e55f36STejun Heo * 1535c8e55f36STejun Heo * LOCKING: 1536d565ed63STejun Heo * spin_lock_irq(pool->lock). 1537c8e55f36STejun Heo */ 1538c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 15391da177e4SLinus Torvalds { 1540bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1541c8e55f36STejun Heo 15426183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 15436183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 15446183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 15456183c009STejun Heo return; 1546c8e55f36STejun Heo 1547cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1548cb444766STejun Heo worker->flags |= WORKER_IDLE; 1549bd7bdd43STejun Heo pool->nr_idle++; 1550e22bee78STejun Heo worker->last_active = jiffies; 1551c8e55f36STejun Heo 1552c8e55f36STejun Heo /* idle_list is LIFO */ 1553bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1554db7bccf4STejun Heo 155563d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1556628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1557cb444766STejun Heo 1558544ecf31STejun Heo /* 1559706026c2STejun Heo * Sanity check nr_running. Because wq_unbind_fn() releases 1560d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1561628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1562628c78e7STejun Heo * unbind is not in progress. 1563544ecf31STejun Heo */ 156424647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1565bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1566e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1567c8e55f36STejun Heo } 1568c8e55f36STejun Heo 1569c8e55f36STejun Heo /** 1570c8e55f36STejun Heo * worker_leave_idle - leave idle state 1571c8e55f36STejun Heo * @worker: worker which is leaving idle state 1572c8e55f36STejun Heo * 1573c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1574c8e55f36STejun Heo * 1575c8e55f36STejun Heo * LOCKING: 1576d565ed63STejun Heo * spin_lock_irq(pool->lock). 1577c8e55f36STejun Heo */ 1578c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1579c8e55f36STejun Heo { 1580bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1581c8e55f36STejun Heo 15826183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 15836183c009STejun Heo return; 1584d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1585bd7bdd43STejun Heo pool->nr_idle--; 1586c8e55f36STejun Heo list_del_init(&worker->entry); 1587c8e55f36STejun Heo } 1588c8e55f36STejun Heo 1589e22bee78STejun Heo /** 1590f36dc67bSLai Jiangshan * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it 1591f36dc67bSLai Jiangshan * @pool: target worker_pool 1592f36dc67bSLai Jiangshan * 1593f36dc67bSLai Jiangshan * Bind %current to the cpu of @pool if it is associated and lock @pool. 1594e22bee78STejun Heo * 1595e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1596e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1597e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1598e22bee78STejun Heo * guaranteed to execute on the cpu. 1599e22bee78STejun Heo * 1600f5faa077SLai Jiangshan * This function is to be used by unbound workers and rescuers to bind 1601e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1602e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1603e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1604706026c2STejun Heo * verbatim as it's best effort and blocking and pool may be 1605e22bee78STejun Heo * [dis]associated in the meantime. 1606e22bee78STejun Heo * 1607706026c2STejun Heo * This function tries set_cpus_allowed() and locks pool and verifies the 160824647570STejun Heo * binding against %POOL_DISASSOCIATED which is set during 1609f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1610f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1611f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1612e22bee78STejun Heo * 1613e22bee78STejun Heo * CONTEXT: 1614d565ed63STejun Heo * Might sleep. Called without any lock but returns with pool->lock 1615e22bee78STejun Heo * held. 1616e22bee78STejun Heo * 1617e22bee78STejun Heo * RETURNS: 1618706026c2STejun Heo * %true if the associated pool is online (@worker is successfully 1619e22bee78STejun Heo * bound), %false if offline. 1620e22bee78STejun Heo */ 1621f36dc67bSLai Jiangshan static bool worker_maybe_bind_and_lock(struct worker_pool *pool) 1622d565ed63STejun Heo __acquires(&pool->lock) 1623e22bee78STejun Heo { 1624e22bee78STejun Heo while (true) { 1625e22bee78STejun Heo /* 1626e22bee78STejun Heo * The following call may fail, succeed or succeed 1627e22bee78STejun Heo * without actually migrating the task to the cpu if 1628e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 162924647570STejun Heo * against POOL_DISASSOCIATED. 1630e22bee78STejun Heo */ 163124647570STejun Heo if (!(pool->flags & POOL_DISASSOCIATED)) 16327a4e344cSTejun Heo set_cpus_allowed_ptr(current, pool->attrs->cpumask); 1633e22bee78STejun Heo 1634d565ed63STejun Heo spin_lock_irq(&pool->lock); 163524647570STejun Heo if (pool->flags & POOL_DISASSOCIATED) 1636e22bee78STejun Heo return false; 1637f5faa077SLai Jiangshan if (task_cpu(current) == pool->cpu && 16387a4e344cSTejun Heo cpumask_equal(¤t->cpus_allowed, pool->attrs->cpumask)) 1639e22bee78STejun Heo return true; 1640d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1641e22bee78STejun Heo 16425035b20fSTejun Heo /* 16435035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 16445035b20fSTejun Heo * and retry migration. cond_resched() is required here; 16455035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 16465035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 16475035b20fSTejun Heo */ 1648e22bee78STejun Heo cpu_relax(); 16495035b20fSTejun Heo cond_resched(); 1650e22bee78STejun Heo } 1651e22bee78STejun Heo } 1652e22bee78STejun Heo 1653c34056a3STejun Heo static struct worker *alloc_worker(void) 1654c34056a3STejun Heo { 1655c34056a3STejun Heo struct worker *worker; 1656c34056a3STejun Heo 1657c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1658c8e55f36STejun Heo if (worker) { 1659c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1660affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1661e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1662e22bee78STejun Heo worker->flags = WORKER_PREP; 1663c8e55f36STejun Heo } 1664c34056a3STejun Heo return worker; 1665c34056a3STejun Heo } 1666c34056a3STejun Heo 1667c34056a3STejun Heo /** 1668c34056a3STejun Heo * create_worker - create a new workqueue worker 166963d95a91STejun Heo * @pool: pool the new worker will belong to 1670c34056a3STejun Heo * 167163d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1672c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1673c34056a3STejun Heo * destroy_worker(). 1674c34056a3STejun Heo * 1675c34056a3STejun Heo * CONTEXT: 1676c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1677c34056a3STejun Heo * 1678c34056a3STejun Heo * RETURNS: 1679c34056a3STejun Heo * Pointer to the newly created worker. 1680c34056a3STejun Heo */ 1681bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1682c34056a3STejun Heo { 1683c34056a3STejun Heo struct worker *worker = NULL; 1684f3421797STejun Heo int id = -1; 1685e3c916a4STejun Heo char id_buf[16]; 1686c34056a3STejun Heo 1687cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1688cd549687STejun Heo 1689822d8405STejun Heo /* 1690822d8405STejun Heo * ID is needed to determine kthread name. Allocate ID first 1691822d8405STejun Heo * without installing the pointer. 1692822d8405STejun Heo */ 1693822d8405STejun Heo idr_preload(GFP_KERNEL); 1694d565ed63STejun Heo spin_lock_irq(&pool->lock); 1695822d8405STejun Heo 1696822d8405STejun Heo id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT); 1697822d8405STejun Heo 1698d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1699822d8405STejun Heo idr_preload_end(); 1700822d8405STejun Heo if (id < 0) 1701c34056a3STejun Heo goto fail; 1702c34056a3STejun Heo 1703c34056a3STejun Heo worker = alloc_worker(); 1704c34056a3STejun Heo if (!worker) 1705c34056a3STejun Heo goto fail; 1706c34056a3STejun Heo 1707bd7bdd43STejun Heo worker->pool = pool; 1708c34056a3STejun Heo worker->id = id; 1709c34056a3STejun Heo 171029c91e99STejun Heo if (pool->cpu >= 0) 1711e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1712e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1713f3421797STejun Heo else 1714e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1715e3c916a4STejun Heo 1716f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1717e3c916a4STejun Heo "kworker/%s", id_buf); 1718c34056a3STejun Heo if (IS_ERR(worker->task)) 1719c34056a3STejun Heo goto fail; 1720c34056a3STejun Heo 1721c5aa87bbSTejun Heo /* 1722c5aa87bbSTejun Heo * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 1723c5aa87bbSTejun Heo * online CPUs. It'll be re-applied when any of the CPUs come up. 1724c5aa87bbSTejun Heo */ 17257a4e344cSTejun Heo set_user_nice(worker->task, pool->attrs->nice); 17267a4e344cSTejun Heo set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17273270476aSTejun Heo 172814a40ffcSTejun Heo /* prevent userland from meddling with cpumask of workqueue workers */ 172914a40ffcSTejun Heo worker->task->flags |= PF_NO_SETAFFINITY; 17307a4e344cSTejun Heo 17317a4e344cSTejun Heo /* 17327a4e344cSTejun Heo * The caller is responsible for ensuring %POOL_DISASSOCIATED 17337a4e344cSTejun Heo * remains stable across this function. See the comments above the 17347a4e344cSTejun Heo * flag definition for details. 17357a4e344cSTejun Heo */ 17367a4e344cSTejun Heo if (pool->flags & POOL_DISASSOCIATED) 1737f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1738c34056a3STejun Heo 1739822d8405STejun Heo /* successful, commit the pointer to idr */ 1740822d8405STejun Heo spin_lock_irq(&pool->lock); 1741822d8405STejun Heo idr_replace(&pool->worker_idr, worker, worker->id); 1742822d8405STejun Heo spin_unlock_irq(&pool->lock); 1743822d8405STejun Heo 1744c34056a3STejun Heo return worker; 1745822d8405STejun Heo 1746c34056a3STejun Heo fail: 1747c34056a3STejun Heo if (id >= 0) { 1748d565ed63STejun Heo spin_lock_irq(&pool->lock); 1749822d8405STejun Heo idr_remove(&pool->worker_idr, id); 1750d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1751c34056a3STejun Heo } 1752c34056a3STejun Heo kfree(worker); 1753c34056a3STejun Heo return NULL; 1754c34056a3STejun Heo } 1755c34056a3STejun Heo 1756c34056a3STejun Heo /** 1757c34056a3STejun Heo * start_worker - start a newly created worker 1758c34056a3STejun Heo * @worker: worker to start 1759c34056a3STejun Heo * 1760706026c2STejun Heo * Make the pool aware of @worker and start it. 1761c34056a3STejun Heo * 1762c34056a3STejun Heo * CONTEXT: 1763d565ed63STejun Heo * spin_lock_irq(pool->lock). 1764c34056a3STejun Heo */ 1765c34056a3STejun Heo static void start_worker(struct worker *worker) 1766c34056a3STejun Heo { 1767cb444766STejun Heo worker->flags |= WORKER_STARTED; 1768bd7bdd43STejun Heo worker->pool->nr_workers++; 1769c8e55f36STejun Heo worker_enter_idle(worker); 1770c34056a3STejun Heo wake_up_process(worker->task); 1771c34056a3STejun Heo } 1772c34056a3STejun Heo 1773c34056a3STejun Heo /** 1774ebf44d16STejun Heo * create_and_start_worker - create and start a worker for a pool 1775ebf44d16STejun Heo * @pool: the target pool 1776ebf44d16STejun Heo * 1777cd549687STejun Heo * Grab the managership of @pool and create and start a new worker for it. 1778ebf44d16STejun Heo */ 1779ebf44d16STejun Heo static int create_and_start_worker(struct worker_pool *pool) 1780ebf44d16STejun Heo { 1781ebf44d16STejun Heo struct worker *worker; 1782ebf44d16STejun Heo 1783cd549687STejun Heo mutex_lock(&pool->manager_mutex); 1784cd549687STejun Heo 1785ebf44d16STejun Heo worker = create_worker(pool); 1786ebf44d16STejun Heo if (worker) { 1787ebf44d16STejun Heo spin_lock_irq(&pool->lock); 1788ebf44d16STejun Heo start_worker(worker); 1789ebf44d16STejun Heo spin_unlock_irq(&pool->lock); 1790ebf44d16STejun Heo } 1791ebf44d16STejun Heo 1792cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 1793cd549687STejun Heo 1794ebf44d16STejun Heo return worker ? 0 : -ENOMEM; 1795ebf44d16STejun Heo } 1796ebf44d16STejun Heo 1797ebf44d16STejun Heo /** 1798c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1799c34056a3STejun Heo * @worker: worker to be destroyed 1800c34056a3STejun Heo * 1801706026c2STejun Heo * Destroy @worker and adjust @pool stats accordingly. 1802c8e55f36STejun Heo * 1803c8e55f36STejun Heo * CONTEXT: 1804d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 1805c34056a3STejun Heo */ 1806c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1807c34056a3STejun Heo { 1808bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1809c34056a3STejun Heo 1810cd549687STejun Heo lockdep_assert_held(&pool->manager_mutex); 1811cd549687STejun Heo lockdep_assert_held(&pool->lock); 1812cd549687STejun Heo 1813c34056a3STejun Heo /* sanity check frenzy */ 18146183c009STejun Heo if (WARN_ON(worker->current_work) || 18156183c009STejun Heo WARN_ON(!list_empty(&worker->scheduled))) 18166183c009STejun Heo return; 1817c34056a3STejun Heo 1818c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1819bd7bdd43STejun Heo pool->nr_workers--; 1820c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1821bd7bdd43STejun Heo pool->nr_idle--; 1822c8e55f36STejun Heo 1823c8e55f36STejun Heo list_del_init(&worker->entry); 1824cb444766STejun Heo worker->flags |= WORKER_DIE; 1825c8e55f36STejun Heo 1826822d8405STejun Heo idr_remove(&pool->worker_idr, worker->id); 1827822d8405STejun Heo 1828d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1829c8e55f36STejun Heo 1830c34056a3STejun Heo kthread_stop(worker->task); 1831c34056a3STejun Heo kfree(worker); 1832c34056a3STejun Heo 1833d565ed63STejun Heo spin_lock_irq(&pool->lock); 1834c34056a3STejun Heo } 1835c34056a3STejun Heo 183663d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1837e22bee78STejun Heo { 183863d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1839e22bee78STejun Heo 1840d565ed63STejun Heo spin_lock_irq(&pool->lock); 1841e22bee78STejun Heo 184263d95a91STejun Heo if (too_many_workers(pool)) { 1843e22bee78STejun Heo struct worker *worker; 1844e22bee78STejun Heo unsigned long expires; 1845e22bee78STejun Heo 1846e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 184763d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1848e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1849e22bee78STejun Heo 1850e22bee78STejun Heo if (time_before(jiffies, expires)) 185163d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1852e22bee78STejun Heo else { 1853e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 185411ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 185563d95a91STejun Heo wake_up_worker(pool); 1856e22bee78STejun Heo } 1857e22bee78STejun Heo } 1858e22bee78STejun Heo 1859d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1860e22bee78STejun Heo } 1861e22bee78STejun Heo 1862493a1724STejun Heo static void send_mayday(struct work_struct *work) 1863e22bee78STejun Heo { 1864112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1865112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1866493a1724STejun Heo 18672e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1868e22bee78STejun Heo 1869493008a8STejun Heo if (!wq->rescuer) 1870493a1724STejun Heo return; 1871e22bee78STejun Heo 1872e22bee78STejun Heo /* mayday mayday mayday */ 1873493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 1874493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1875e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1876493a1724STejun Heo } 1877e22bee78STejun Heo } 1878e22bee78STejun Heo 1879706026c2STejun Heo static void pool_mayday_timeout(unsigned long __pool) 1880e22bee78STejun Heo { 188163d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 1882e22bee78STejun Heo struct work_struct *work; 1883e22bee78STejun Heo 18842e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */ 1885493a1724STejun Heo spin_lock(&pool->lock); 1886e22bee78STejun Heo 188763d95a91STejun Heo if (need_to_create_worker(pool)) { 1888e22bee78STejun Heo /* 1889e22bee78STejun Heo * We've been trying to create a new worker but 1890e22bee78STejun Heo * haven't been successful. We might be hitting an 1891e22bee78STejun Heo * allocation deadlock. Send distress signals to 1892e22bee78STejun Heo * rescuers. 1893e22bee78STejun Heo */ 189463d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1895e22bee78STejun Heo send_mayday(work); 1896e22bee78STejun Heo } 1897e22bee78STejun Heo 1898493a1724STejun Heo spin_unlock(&pool->lock); 18992e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 1900e22bee78STejun Heo 190163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1902e22bee78STejun Heo } 1903e22bee78STejun Heo 1904e22bee78STejun Heo /** 1905e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 190663d95a91STejun Heo * @pool: pool to create a new worker for 1907e22bee78STejun Heo * 190863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1909e22bee78STejun Heo * have at least one idle worker on return from this function. If 1910e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 191163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1912e22bee78STejun Heo * possible allocation deadlock. 1913e22bee78STejun Heo * 1914c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1915c5aa87bbSTejun Heo * may_start_working() %true. 1916e22bee78STejun Heo * 1917e22bee78STejun Heo * LOCKING: 1918d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1919e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1920e22bee78STejun Heo * manager. 1921e22bee78STejun Heo * 1922e22bee78STejun Heo * RETURNS: 1923c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1924e22bee78STejun Heo * otherwise. 1925e22bee78STejun Heo */ 192663d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 1927d565ed63STejun Heo __releases(&pool->lock) 1928d565ed63STejun Heo __acquires(&pool->lock) 1929e22bee78STejun Heo { 193063d95a91STejun Heo if (!need_to_create_worker(pool)) 1931e22bee78STejun Heo return false; 1932e22bee78STejun Heo restart: 1933d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19349f9c2364STejun Heo 1935e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 193663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1937e22bee78STejun Heo 1938e22bee78STejun Heo while (true) { 1939e22bee78STejun Heo struct worker *worker; 1940e22bee78STejun Heo 1941bc2ae0f5STejun Heo worker = create_worker(pool); 1942e22bee78STejun Heo if (worker) { 194363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1944d565ed63STejun Heo spin_lock_irq(&pool->lock); 1945e22bee78STejun Heo start_worker(worker); 19466183c009STejun Heo if (WARN_ON_ONCE(need_to_create_worker(pool))) 19476183c009STejun Heo goto restart; 1948e22bee78STejun Heo return true; 1949e22bee78STejun Heo } 1950e22bee78STejun Heo 195163d95a91STejun Heo if (!need_to_create_worker(pool)) 1952e22bee78STejun Heo break; 1953e22bee78STejun Heo 1954e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1955e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 19569f9c2364STejun Heo 195763d95a91STejun Heo if (!need_to_create_worker(pool)) 1958e22bee78STejun Heo break; 1959e22bee78STejun Heo } 1960e22bee78STejun Heo 196163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1962d565ed63STejun Heo spin_lock_irq(&pool->lock); 196363d95a91STejun Heo if (need_to_create_worker(pool)) 1964e22bee78STejun Heo goto restart; 1965e22bee78STejun Heo return true; 1966e22bee78STejun Heo } 1967e22bee78STejun Heo 1968e22bee78STejun Heo /** 1969e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 197063d95a91STejun Heo * @pool: pool to destroy workers for 1971e22bee78STejun Heo * 197263d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1973e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1974e22bee78STejun Heo * 1975e22bee78STejun Heo * LOCKING: 1976d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1977e22bee78STejun Heo * multiple times. Called only from manager. 1978e22bee78STejun Heo * 1979e22bee78STejun Heo * RETURNS: 1980c5aa87bbSTejun Heo * %false if no action was taken and pool->lock stayed locked, %true 1981e22bee78STejun Heo * otherwise. 1982e22bee78STejun Heo */ 198363d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1984e22bee78STejun Heo { 1985e22bee78STejun Heo bool ret = false; 1986e22bee78STejun Heo 198763d95a91STejun Heo while (too_many_workers(pool)) { 1988e22bee78STejun Heo struct worker *worker; 1989e22bee78STejun Heo unsigned long expires; 1990e22bee78STejun Heo 199163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1992e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1993e22bee78STejun Heo 1994e22bee78STejun Heo if (time_before(jiffies, expires)) { 199563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1996e22bee78STejun Heo break; 1997e22bee78STejun Heo } 1998e22bee78STejun Heo 1999e22bee78STejun Heo destroy_worker(worker); 2000e22bee78STejun Heo ret = true; 2001e22bee78STejun Heo } 2002e22bee78STejun Heo 2003e22bee78STejun Heo return ret; 2004e22bee78STejun Heo } 2005e22bee78STejun Heo 2006e22bee78STejun Heo /** 2007e22bee78STejun Heo * manage_workers - manage worker pool 2008e22bee78STejun Heo * @worker: self 2009e22bee78STejun Heo * 2010706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2011e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2012706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2013e22bee78STejun Heo * 2014e22bee78STejun Heo * The caller can safely start processing works on false return. On 2015e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2016e22bee78STejun Heo * and may_start_working() is true. 2017e22bee78STejun Heo * 2018e22bee78STejun Heo * CONTEXT: 2019d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2020e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2021e22bee78STejun Heo * 2022e22bee78STejun Heo * RETURNS: 2023d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2024d565ed63STejun Heo * multiple times. Does GFP_KERNEL allocations. 2025e22bee78STejun Heo */ 2026e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2027e22bee78STejun Heo { 202863d95a91STejun Heo struct worker_pool *pool = worker->pool; 2029e22bee78STejun Heo bool ret = false; 2030e22bee78STejun Heo 2031bc3a1afcSTejun Heo /* 2032bc3a1afcSTejun Heo * Managership is governed by two mutexes - manager_arb and 2033bc3a1afcSTejun Heo * manager_mutex. manager_arb handles arbitration of manager role. 2034bc3a1afcSTejun Heo * Anyone who successfully grabs manager_arb wins the arbitration 2035bc3a1afcSTejun Heo * and becomes the manager. mutex_trylock() on pool->manager_arb 2036bc3a1afcSTejun Heo * failure while holding pool->lock reliably indicates that someone 2037bc3a1afcSTejun Heo * else is managing the pool and the worker which failed trylock 2038bc3a1afcSTejun Heo * can proceed to executing work items. This means that anyone 2039bc3a1afcSTejun Heo * grabbing manager_arb is responsible for actually performing 2040bc3a1afcSTejun Heo * manager duties. If manager_arb is grabbed and released without 2041bc3a1afcSTejun Heo * actual management, the pool may stall indefinitely. 2042bc3a1afcSTejun Heo * 2043bc3a1afcSTejun Heo * manager_mutex is used for exclusion of actual management 2044bc3a1afcSTejun Heo * operations. The holder of manager_mutex can be sure that none 2045bc3a1afcSTejun Heo * of management operations, including creation and destruction of 2046bc3a1afcSTejun Heo * workers, won't take place until the mutex is released. Because 2047bc3a1afcSTejun Heo * manager_mutex doesn't interfere with manager role arbitration, 2048bc3a1afcSTejun Heo * it is guaranteed that the pool's management, while may be 2049bc3a1afcSTejun Heo * delayed, won't be disturbed by someone else grabbing 2050bc3a1afcSTejun Heo * manager_mutex. 2051bc3a1afcSTejun Heo */ 205234a06bd6STejun Heo if (!mutex_trylock(&pool->manager_arb)) 2053e22bee78STejun Heo return ret; 2054e22bee78STejun Heo 2055ee378aa4SLai Jiangshan /* 2056bc3a1afcSTejun Heo * With manager arbitration won, manager_mutex would be free in 2057bc3a1afcSTejun Heo * most cases. trylock first without dropping @pool->lock. 2058ee378aa4SLai Jiangshan */ 2059bc3a1afcSTejun Heo if (unlikely(!mutex_trylock(&pool->manager_mutex))) { 2060d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2061bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 20628f174b11SJoonsoo Kim spin_lock_irq(&pool->lock); 2063ee378aa4SLai Jiangshan ret = true; 2064ee378aa4SLai Jiangshan } 2065ee378aa4SLai Jiangshan 206611ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 2067e22bee78STejun Heo 2068e22bee78STejun Heo /* 2069e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 2070e22bee78STejun Heo * on return. 2071e22bee78STejun Heo */ 207263d95a91STejun Heo ret |= maybe_destroy_workers(pool); 207363d95a91STejun Heo ret |= maybe_create_worker(pool); 2074e22bee78STejun Heo 2075bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 207634a06bd6STejun Heo mutex_unlock(&pool->manager_arb); 2077e22bee78STejun Heo return ret; 2078e22bee78STejun Heo } 2079e22bee78STejun Heo 2080a62428c0STejun Heo /** 2081a62428c0STejun Heo * process_one_work - process single work 2082c34056a3STejun Heo * @worker: self 2083a62428c0STejun Heo * @work: work to process 2084a62428c0STejun Heo * 2085a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2086a62428c0STejun Heo * process a single work including synchronization against and 2087a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2088a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2089a62428c0STejun Heo * call this function to process a work. 2090a62428c0STejun Heo * 2091a62428c0STejun Heo * CONTEXT: 2092d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2093a62428c0STejun Heo */ 2094c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2095d565ed63STejun Heo __releases(&pool->lock) 2096d565ed63STejun Heo __acquires(&pool->lock) 20971da177e4SLinus Torvalds { 2098112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2099bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2100112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 210173f53c4aSTejun Heo int work_color; 21027e11629dSTejun Heo struct worker *collision; 21034e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21044e6045f1SJohannes Berg /* 2105a62428c0STejun Heo * It is permissible to free the struct work_struct from 2106a62428c0STejun Heo * inside the function that is called from it, this we need to 2107a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2108a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2109a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21104e6045f1SJohannes Berg */ 21114d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21124d82a1deSPeter Zijlstra 21134d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21144e6045f1SJohannes Berg #endif 21156fec10a1STejun Heo /* 21166fec10a1STejun Heo * Ensure we're on the correct CPU. DISASSOCIATED test is 21176fec10a1STejun Heo * necessary to avoid spurious warnings from rescuers servicing the 211824647570STejun Heo * unbound or a disassociated pool. 21196fec10a1STejun Heo */ 21205f7dabfdSLai Jiangshan WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) && 212124647570STejun Heo !(pool->flags & POOL_DISASSOCIATED) && 2122ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 212325511a47STejun Heo 21247e11629dSTejun Heo /* 21257e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21267e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21277e11629dSTejun Heo * already processing the work. If so, defer the work to the 21287e11629dSTejun Heo * currently executing one. 21297e11629dSTejun Heo */ 2130c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 21317e11629dSTejun Heo if (unlikely(collision)) { 21327e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 21337e11629dSTejun Heo return; 21347e11629dSTejun Heo } 21351da177e4SLinus Torvalds 21368930cabaSTejun Heo /* claim and dequeue */ 21371da177e4SLinus Torvalds debug_work_deactivate(work); 2138c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2139c34056a3STejun Heo worker->current_work = work; 2140a2c1c57bSTejun Heo worker->current_func = work->func; 2141112202d9STejun Heo worker->current_pwq = pwq; 214273f53c4aSTejun Heo work_color = get_work_color(work); 21437a22ad75STejun Heo 2144a62428c0STejun Heo list_del_init(&work->entry); 2145a62428c0STejun Heo 2146649027d7STejun Heo /* 2147fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 2148fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 2149fb0e7bebSTejun Heo */ 2150fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2151fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 2152fb0e7bebSTejun Heo 2153974271c4STejun Heo /* 2154d565ed63STejun Heo * Unbound pool isn't concurrency managed and work items should be 2155974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2156974271c4STejun Heo */ 215763d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 215863d95a91STejun Heo wake_up_worker(pool); 2159974271c4STejun Heo 21608930cabaSTejun Heo /* 21617c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2162d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 216323657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 216423657bb1STejun Heo * disabled. 21658930cabaSTejun Heo */ 21667c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21671da177e4SLinus Torvalds 2168d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2169365970a1SDavid Howells 2170112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 21713295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2172e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2173a2c1c57bSTejun Heo worker->current_func(work); 2174e36c886aSArjan van de Ven /* 2175e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2176e36c886aSArjan van de Ven * point will only record its address. 2177e36c886aSArjan van de Ven */ 2178e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21793295f0efSIngo Molnar lock_map_release(&lockdep_map); 2180112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21811da177e4SLinus Torvalds 2182d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2183044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2184044c782cSValentin Ilie " last function: %pf\n", 2185a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2186a2c1c57bSTejun Heo worker->current_func); 2187d5abe669SPeter Zijlstra debug_show_held_locks(current); 2188d5abe669SPeter Zijlstra dump_stack(); 2189d5abe669SPeter Zijlstra } 2190d5abe669SPeter Zijlstra 2191d565ed63STejun Heo spin_lock_irq(&pool->lock); 2192a62428c0STejun Heo 2193fb0e7bebSTejun Heo /* clear cpu intensive status */ 2194fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2195fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2196fb0e7bebSTejun Heo 2197a62428c0STejun Heo /* we're done with it, release */ 219842f8570fSSasha Levin hash_del(&worker->hentry); 2199c34056a3STejun Heo worker->current_work = NULL; 2200a2c1c57bSTejun Heo worker->current_func = NULL; 2201112202d9STejun Heo worker->current_pwq = NULL; 22023d1cb205STejun Heo worker->desc_valid = false; 2203112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 22041da177e4SLinus Torvalds } 22051da177e4SLinus Torvalds 2206affee4b2STejun Heo /** 2207affee4b2STejun Heo * process_scheduled_works - process scheduled works 2208affee4b2STejun Heo * @worker: self 2209affee4b2STejun Heo * 2210affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2211affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2212affee4b2STejun Heo * fetches a work from the top and executes it. 2213affee4b2STejun Heo * 2214affee4b2STejun Heo * CONTEXT: 2215d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2216affee4b2STejun Heo * multiple times. 2217affee4b2STejun Heo */ 2218affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22191da177e4SLinus Torvalds { 2220affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2221affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2222a62428c0STejun Heo struct work_struct, entry); 2223c34056a3STejun Heo process_one_work(worker, work); 2224a62428c0STejun Heo } 22251da177e4SLinus Torvalds } 22261da177e4SLinus Torvalds 22274690c4abSTejun Heo /** 22284690c4abSTejun Heo * worker_thread - the worker thread function 2229c34056a3STejun Heo * @__worker: self 22304690c4abSTejun Heo * 2231c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2232c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2233c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2234c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2235c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 22364690c4abSTejun Heo */ 2237c34056a3STejun Heo static int worker_thread(void *__worker) 22381da177e4SLinus Torvalds { 2239c34056a3STejun Heo struct worker *worker = __worker; 2240bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22411da177e4SLinus Torvalds 2242e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2243e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2244c8e55f36STejun Heo woke_up: 2245d565ed63STejun Heo spin_lock_irq(&pool->lock); 2246affee4b2STejun Heo 2247a9ab775bSTejun Heo /* am I supposed to die? */ 2248a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2249d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2250a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2251e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2252c8e55f36STejun Heo return 0; 2253c8e55f36STejun Heo } 2254c8e55f36STejun Heo 2255c8e55f36STejun Heo worker_leave_idle(worker); 2256db7bccf4STejun Heo recheck: 2257e22bee78STejun Heo /* no more worker necessary? */ 225863d95a91STejun Heo if (!need_more_worker(pool)) 2259e22bee78STejun Heo goto sleep; 2260e22bee78STejun Heo 2261e22bee78STejun Heo /* do we need to manage? */ 226263d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2263e22bee78STejun Heo goto recheck; 2264e22bee78STejun Heo 2265c8e55f36STejun Heo /* 2266c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2267c8e55f36STejun Heo * preparing to process a work or actually processing it. 2268c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2269c8e55f36STejun Heo */ 22706183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2271c8e55f36STejun Heo 2272e22bee78STejun Heo /* 2273a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2274a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2275a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2276a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2277a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2278e22bee78STejun Heo */ 2279a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2280e22bee78STejun Heo 2281e22bee78STejun Heo do { 2282affee4b2STejun Heo struct work_struct *work = 2283bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2284affee4b2STejun Heo struct work_struct, entry); 2285affee4b2STejun Heo 2286c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2287affee4b2STejun Heo /* optimization path, not strictly necessary */ 2288affee4b2STejun Heo process_one_work(worker, work); 2289affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2290affee4b2STejun Heo process_scheduled_works(worker); 2291affee4b2STejun Heo } else { 2292c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2293affee4b2STejun Heo process_scheduled_works(worker); 2294affee4b2STejun Heo } 229563d95a91STejun Heo } while (keep_working(pool)); 2296affee4b2STejun Heo 2297e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2298d313dd85STejun Heo sleep: 229963d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2300e22bee78STejun Heo goto recheck; 2301d313dd85STejun Heo 2302c8e55f36STejun Heo /* 2303d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2304d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2305d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2306d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2307d565ed63STejun Heo * event. 2308c8e55f36STejun Heo */ 2309c8e55f36STejun Heo worker_enter_idle(worker); 2310c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 2311d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23121da177e4SLinus Torvalds schedule(); 2313c8e55f36STejun Heo goto woke_up; 23141da177e4SLinus Torvalds } 23151da177e4SLinus Torvalds 2316e22bee78STejun Heo /** 2317e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2318111c225aSTejun Heo * @__rescuer: self 2319e22bee78STejun Heo * 2320e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2321493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2322e22bee78STejun Heo * 2323706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2324e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2325e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2326e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2327e22bee78STejun Heo * the problem rescuer solves. 2328e22bee78STejun Heo * 2329706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2330706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2331e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2332e22bee78STejun Heo * 2333e22bee78STejun Heo * This should happen rarely. 2334e22bee78STejun Heo */ 2335111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2336e22bee78STejun Heo { 2337111c225aSTejun Heo struct worker *rescuer = __rescuer; 2338111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2339e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2340e22bee78STejun Heo 2341e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2342111c225aSTejun Heo 2343111c225aSTejun Heo /* 2344111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2345111c225aSTejun Heo * doesn't participate in concurrency management. 2346111c225aSTejun Heo */ 2347111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2348e22bee78STejun Heo repeat: 2349e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 23501da177e4SLinus Torvalds 2351412d32e6SMike Galbraith if (kthread_should_stop()) { 2352412d32e6SMike Galbraith __set_current_state(TASK_RUNNING); 2353111c225aSTejun Heo rescuer->task->flags &= ~PF_WQ_WORKER; 2354e22bee78STejun Heo return 0; 2355412d32e6SMike Galbraith } 23561da177e4SLinus Torvalds 2357493a1724STejun Heo /* see whether any pwq is asking for help */ 23582e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2359493a1724STejun Heo 2360493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2361493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2362493a1724STejun Heo struct pool_workqueue, mayday_node); 2363112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2364e22bee78STejun Heo struct work_struct *work, *n; 2365e22bee78STejun Heo 2366e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2367493a1724STejun Heo list_del_init(&pwq->mayday_node); 2368493a1724STejun Heo 23692e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2370e22bee78STejun Heo 2371e22bee78STejun Heo /* migrate to the target cpu if possible */ 2372f36dc67bSLai Jiangshan worker_maybe_bind_and_lock(pool); 2373b3104104SLai Jiangshan rescuer->pool = pool; 2374e22bee78STejun Heo 2375e22bee78STejun Heo /* 2376e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2377e22bee78STejun Heo * process'em. 2378e22bee78STejun Heo */ 23796183c009STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 2380bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2381112202d9STejun Heo if (get_work_pwq(work) == pwq) 2382e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2383e22bee78STejun Heo 2384e22bee78STejun Heo process_scheduled_works(rescuer); 23857576958aSTejun Heo 23867576958aSTejun Heo /* 2387d565ed63STejun Heo * Leave this pool. If keep_working() is %true, notify a 23887576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 23897576958aSTejun Heo * and stalling the execution. 23907576958aSTejun Heo */ 239163d95a91STejun Heo if (keep_working(pool)) 239263d95a91STejun Heo wake_up_worker(pool); 23937576958aSTejun Heo 2394b3104104SLai Jiangshan rescuer->pool = NULL; 2395493a1724STejun Heo spin_unlock(&pool->lock); 23962e109a28STejun Heo spin_lock(&wq_mayday_lock); 23971da177e4SLinus Torvalds } 23981da177e4SLinus Torvalds 23992e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2400493a1724STejun Heo 2401111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2402111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2403e22bee78STejun Heo schedule(); 2404e22bee78STejun Heo goto repeat; 24051da177e4SLinus Torvalds } 24061da177e4SLinus Torvalds 2407fc2e4d70SOleg Nesterov struct wq_barrier { 2408fc2e4d70SOleg Nesterov struct work_struct work; 2409fc2e4d70SOleg Nesterov struct completion done; 2410fc2e4d70SOleg Nesterov }; 2411fc2e4d70SOleg Nesterov 2412fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2413fc2e4d70SOleg Nesterov { 2414fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2415fc2e4d70SOleg Nesterov complete(&barr->done); 2416fc2e4d70SOleg Nesterov } 2417fc2e4d70SOleg Nesterov 24184690c4abSTejun Heo /** 24194690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2420112202d9STejun Heo * @pwq: pwq to insert barrier into 24214690c4abSTejun Heo * @barr: wq_barrier to insert 2422affee4b2STejun Heo * @target: target work to attach @barr to 2423affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24244690c4abSTejun Heo * 2425affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2426affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2427affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2428affee4b2STejun Heo * cpu. 2429affee4b2STejun Heo * 2430affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2431affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2432affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2433affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2434affee4b2STejun Heo * after a work with LINKED flag set. 2435affee4b2STejun Heo * 2436affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2437112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 24384690c4abSTejun Heo * 24394690c4abSTejun Heo * CONTEXT: 2440d565ed63STejun Heo * spin_lock_irq(pool->lock). 24414690c4abSTejun Heo */ 2442112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2443affee4b2STejun Heo struct wq_barrier *barr, 2444affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2445fc2e4d70SOleg Nesterov { 2446affee4b2STejun Heo struct list_head *head; 2447affee4b2STejun Heo unsigned int linked = 0; 2448affee4b2STejun Heo 2449dc186ad7SThomas Gleixner /* 2450d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2451dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2452dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2453dc186ad7SThomas Gleixner * might deadlock. 2454dc186ad7SThomas Gleixner */ 2455ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 245622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2457fc2e4d70SOleg Nesterov init_completion(&barr->done); 245883c22520SOleg Nesterov 2459affee4b2STejun Heo /* 2460affee4b2STejun Heo * If @target is currently being executed, schedule the 2461affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2462affee4b2STejun Heo */ 2463affee4b2STejun Heo if (worker) 2464affee4b2STejun Heo head = worker->scheduled.next; 2465affee4b2STejun Heo else { 2466affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2467affee4b2STejun Heo 2468affee4b2STejun Heo head = target->entry.next; 2469affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2470affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2471affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2472affee4b2STejun Heo } 2473affee4b2STejun Heo 2474dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2475112202d9STejun Heo insert_work(pwq, &barr->work, head, 2476affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2477fc2e4d70SOleg Nesterov } 2478fc2e4d70SOleg Nesterov 247973f53c4aSTejun Heo /** 2480112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 248173f53c4aSTejun Heo * @wq: workqueue being flushed 248273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 248373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 248473f53c4aSTejun Heo * 2485112202d9STejun Heo * Prepare pwqs for workqueue flushing. 248673f53c4aSTejun Heo * 2487112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2488112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2489112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2490112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2491112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 249273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 249373f53c4aSTejun Heo * 249473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 249573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 249673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 249773f53c4aSTejun Heo * is returned. 249873f53c4aSTejun Heo * 2499112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 250073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 250173f53c4aSTejun Heo * advanced to @work_color. 250273f53c4aSTejun Heo * 250373f53c4aSTejun Heo * CONTEXT: 25043c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 250573f53c4aSTejun Heo * 250673f53c4aSTejun Heo * RETURNS: 250773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 250873f53c4aSTejun Heo * otherwise. 250973f53c4aSTejun Heo */ 2510112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 251173f53c4aSTejun Heo int flush_color, int work_color) 25121da177e4SLinus Torvalds { 251373f53c4aSTejun Heo bool wait = false; 251449e3cf44STejun Heo struct pool_workqueue *pwq; 25151da177e4SLinus Torvalds 251673f53c4aSTejun Heo if (flush_color >= 0) { 25176183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2518112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2519dc186ad7SThomas Gleixner } 252014441960SOleg Nesterov 252149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2522112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25231da177e4SLinus Torvalds 2524b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 252573f53c4aSTejun Heo 252673f53c4aSTejun Heo if (flush_color >= 0) { 25276183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 252873f53c4aSTejun Heo 2529112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2530112202d9STejun Heo pwq->flush_color = flush_color; 2531112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 253273f53c4aSTejun Heo wait = true; 25331da177e4SLinus Torvalds } 253473f53c4aSTejun Heo } 253573f53c4aSTejun Heo 253673f53c4aSTejun Heo if (work_color >= 0) { 25376183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2538112202d9STejun Heo pwq->work_color = work_color; 253973f53c4aSTejun Heo } 254073f53c4aSTejun Heo 2541b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 25421da177e4SLinus Torvalds } 25431da177e4SLinus Torvalds 2544112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 254573f53c4aSTejun Heo complete(&wq->first_flusher->done); 254673f53c4aSTejun Heo 254773f53c4aSTejun Heo return wait; 254883c22520SOleg Nesterov } 25491da177e4SLinus Torvalds 25500fcb78c2SRolf Eike Beer /** 25511da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 25520fcb78c2SRolf Eike Beer * @wq: workqueue to flush 25531da177e4SLinus Torvalds * 2554c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2555c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 25561da177e4SLinus Torvalds */ 25577ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 25581da177e4SLinus Torvalds { 255973f53c4aSTejun Heo struct wq_flusher this_flusher = { 256073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 256173f53c4aSTejun Heo .flush_color = -1, 256273f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 256373f53c4aSTejun Heo }; 256473f53c4aSTejun Heo int next_color; 2565b1f4ec17SOleg Nesterov 25663295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 25673295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 256873f53c4aSTejun Heo 25693c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 257073f53c4aSTejun Heo 257173f53c4aSTejun Heo /* 257273f53c4aSTejun Heo * Start-to-wait phase 257373f53c4aSTejun Heo */ 257473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 257573f53c4aSTejun Heo 257673f53c4aSTejun Heo if (next_color != wq->flush_color) { 257773f53c4aSTejun Heo /* 257873f53c4aSTejun Heo * Color space is not full. The current work_color 257973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 258073f53c4aSTejun Heo * by one. 258173f53c4aSTejun Heo */ 25826183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 258373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 258473f53c4aSTejun Heo wq->work_color = next_color; 258573f53c4aSTejun Heo 258673f53c4aSTejun Heo if (!wq->first_flusher) { 258773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 25886183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 258973f53c4aSTejun Heo 259073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 259173f53c4aSTejun Heo 2592112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 259373f53c4aSTejun Heo wq->work_color)) { 259473f53c4aSTejun Heo /* nothing to flush, done */ 259573f53c4aSTejun Heo wq->flush_color = next_color; 259673f53c4aSTejun Heo wq->first_flusher = NULL; 259773f53c4aSTejun Heo goto out_unlock; 259873f53c4aSTejun Heo } 259973f53c4aSTejun Heo } else { 260073f53c4aSTejun Heo /* wait in queue */ 26016183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 260273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2603112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 260473f53c4aSTejun Heo } 260573f53c4aSTejun Heo } else { 260673f53c4aSTejun Heo /* 260773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 260873f53c4aSTejun Heo * The next flush completion will assign us 260973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 261073f53c4aSTejun Heo */ 261173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 261273f53c4aSTejun Heo } 261373f53c4aSTejun Heo 26143c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 261573f53c4aSTejun Heo 261673f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 261773f53c4aSTejun Heo 261873f53c4aSTejun Heo /* 261973f53c4aSTejun Heo * Wake-up-and-cascade phase 262073f53c4aSTejun Heo * 262173f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 262273f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 262373f53c4aSTejun Heo */ 262473f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 262573f53c4aSTejun Heo return; 262673f53c4aSTejun Heo 26273c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 262873f53c4aSTejun Heo 26294ce48b37STejun Heo /* we might have raced, check again with mutex held */ 26304ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 26314ce48b37STejun Heo goto out_unlock; 26324ce48b37STejun Heo 263373f53c4aSTejun Heo wq->first_flusher = NULL; 263473f53c4aSTejun Heo 26356183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 26366183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 263773f53c4aSTejun Heo 263873f53c4aSTejun Heo while (true) { 263973f53c4aSTejun Heo struct wq_flusher *next, *tmp; 264073f53c4aSTejun Heo 264173f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 264273f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 264373f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 264473f53c4aSTejun Heo break; 264573f53c4aSTejun Heo list_del_init(&next->list); 264673f53c4aSTejun Heo complete(&next->done); 264773f53c4aSTejun Heo } 264873f53c4aSTejun Heo 26496183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 265073f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 265173f53c4aSTejun Heo 265273f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 265373f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 265473f53c4aSTejun Heo 265573f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 265673f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 265773f53c4aSTejun Heo /* 265873f53c4aSTejun Heo * Assign the same color to all overflowed 265973f53c4aSTejun Heo * flushers, advance work_color and append to 266073f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 266173f53c4aSTejun Heo * phase for these overflowed flushers. 266273f53c4aSTejun Heo */ 266373f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 266473f53c4aSTejun Heo tmp->flush_color = wq->work_color; 266573f53c4aSTejun Heo 266673f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 266773f53c4aSTejun Heo 266873f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 266973f53c4aSTejun Heo &wq->flusher_queue); 2670112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 267173f53c4aSTejun Heo } 267273f53c4aSTejun Heo 267373f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 26746183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 267573f53c4aSTejun Heo break; 267673f53c4aSTejun Heo } 267773f53c4aSTejun Heo 267873f53c4aSTejun Heo /* 267973f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2680112202d9STejun Heo * the new first flusher and arm pwqs. 268173f53c4aSTejun Heo */ 26826183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 26836183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 268473f53c4aSTejun Heo 268573f53c4aSTejun Heo list_del_init(&next->list); 268673f53c4aSTejun Heo wq->first_flusher = next; 268773f53c4aSTejun Heo 2688112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 268973f53c4aSTejun Heo break; 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo /* 269273f53c4aSTejun Heo * Meh... this color is already done, clear first 269373f53c4aSTejun Heo * flusher and repeat cascading. 269473f53c4aSTejun Heo */ 269573f53c4aSTejun Heo wq->first_flusher = NULL; 269673f53c4aSTejun Heo } 269773f53c4aSTejun Heo 269873f53c4aSTejun Heo out_unlock: 26993c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27001da177e4SLinus Torvalds } 2701ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 27021da177e4SLinus Torvalds 27039c5a2ba7STejun Heo /** 27049c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27059c5a2ba7STejun Heo * @wq: workqueue to drain 27069c5a2ba7STejun Heo * 27079c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27089c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27099c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 27109c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 27119c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27129c5a2ba7STejun Heo * takes too long. 27139c5a2ba7STejun Heo */ 27149c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27159c5a2ba7STejun Heo { 27169c5a2ba7STejun Heo unsigned int flush_cnt = 0; 271749e3cf44STejun Heo struct pool_workqueue *pwq; 27189c5a2ba7STejun Heo 27199c5a2ba7STejun Heo /* 27209c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27219c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2722618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27239c5a2ba7STejun Heo */ 272487fc741eSLai Jiangshan mutex_lock(&wq->mutex); 27259c5a2ba7STejun Heo if (!wq->nr_drainers++) 2726618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 272787fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27289c5a2ba7STejun Heo reflush: 27299c5a2ba7STejun Heo flush_workqueue(wq); 27309c5a2ba7STejun Heo 2731b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 273276af4d93STejun Heo 273349e3cf44STejun Heo for_each_pwq(pwq, wq) { 2734fa2563e4SThomas Tuttle bool drained; 27359c5a2ba7STejun Heo 2736b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2737112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2738b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2739fa2563e4SThomas Tuttle 2740fa2563e4SThomas Tuttle if (drained) 27419c5a2ba7STejun Heo continue; 27429c5a2ba7STejun Heo 27439c5a2ba7STejun Heo if (++flush_cnt == 10 || 27449c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2745c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 27469c5a2ba7STejun Heo wq->name, flush_cnt); 274776af4d93STejun Heo 2748b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 27499c5a2ba7STejun Heo goto reflush; 27509c5a2ba7STejun Heo } 27519c5a2ba7STejun Heo 27529c5a2ba7STejun Heo if (!--wq->nr_drainers) 2753618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 275487fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 27559c5a2ba7STejun Heo } 27569c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 27579c5a2ba7STejun Heo 2758606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2759baf59022STejun Heo { 2760baf59022STejun Heo struct worker *worker = NULL; 2761c9e7cf27STejun Heo struct worker_pool *pool; 2762112202d9STejun Heo struct pool_workqueue *pwq; 2763baf59022STejun Heo 2764baf59022STejun Heo might_sleep(); 2765baf59022STejun Heo 2766fa1b54e6STejun Heo local_irq_disable(); 2767fa1b54e6STejun Heo pool = get_work_pool(work); 2768fa1b54e6STejun Heo if (!pool) { 2769fa1b54e6STejun Heo local_irq_enable(); 2770fa1b54e6STejun Heo return false; 2771fa1b54e6STejun Heo } 2772fa1b54e6STejun Heo 2773fa1b54e6STejun Heo spin_lock(&pool->lock); 27740b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2775112202d9STejun Heo pwq = get_work_pwq(work); 2776112202d9STejun Heo if (pwq) { 2777112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2778baf59022STejun Heo goto already_gone; 2779606a5020STejun Heo } else { 2780c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2781baf59022STejun Heo if (!worker) 2782baf59022STejun Heo goto already_gone; 2783112202d9STejun Heo pwq = worker->current_pwq; 2784606a5020STejun Heo } 2785baf59022STejun Heo 2786112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2787d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2788baf59022STejun Heo 2789e159489bSTejun Heo /* 2790e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2791e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2792e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2793e159489bSTejun Heo * access. 2794e159489bSTejun Heo */ 2795493008a8STejun Heo if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) 2796112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2797e159489bSTejun Heo else 2798112202d9STejun Heo lock_map_acquire_read(&pwq->wq->lockdep_map); 2799112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2800e159489bSTejun Heo 2801baf59022STejun Heo return true; 2802baf59022STejun Heo already_gone: 2803d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2804baf59022STejun Heo return false; 2805baf59022STejun Heo } 2806baf59022STejun Heo 2807db700897SOleg Nesterov /** 2808401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2809401a8d04STejun Heo * @work: the work to flush 2810db700897SOleg Nesterov * 2811606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2812606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2813401a8d04STejun Heo * 2814401a8d04STejun Heo * RETURNS: 2815401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2816401a8d04STejun Heo * %false if it was already idle. 2817db700897SOleg Nesterov */ 2818401a8d04STejun Heo bool flush_work(struct work_struct *work) 2819db700897SOleg Nesterov { 2820db700897SOleg Nesterov struct wq_barrier barr; 2821db700897SOleg Nesterov 28220976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 28230976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 28240976dfc1SStephen Boyd 2825606a5020STejun Heo if (start_flush_work(work, &barr)) { 2826db700897SOleg Nesterov wait_for_completion(&barr.done); 2827dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2828401a8d04STejun Heo return true; 2829606a5020STejun Heo } else { 2830401a8d04STejun Heo return false; 2831db700897SOleg Nesterov } 2832606a5020STejun Heo } 2833db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2834db700897SOleg Nesterov 283536e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2836401a8d04STejun Heo { 2837bbb68dfaSTejun Heo unsigned long flags; 28381f1f642eSOleg Nesterov int ret; 28391f1f642eSOleg Nesterov 28401f1f642eSOleg Nesterov do { 2841bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2842bbb68dfaSTejun Heo /* 2843bbb68dfaSTejun Heo * If someone else is canceling, wait for the same event it 2844bbb68dfaSTejun Heo * would be waiting for before retrying. 2845bbb68dfaSTejun Heo */ 2846bbb68dfaSTejun Heo if (unlikely(ret == -ENOENT)) 2847606a5020STejun Heo flush_work(work); 28481f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28491f1f642eSOleg Nesterov 2850bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2851bbb68dfaSTejun Heo mark_work_canceling(work); 2852bbb68dfaSTejun Heo local_irq_restore(flags); 2853bbb68dfaSTejun Heo 2854606a5020STejun Heo flush_work(work); 28557a22ad75STejun Heo clear_work_data(work); 28561f1f642eSOleg Nesterov return ret; 28571f1f642eSOleg Nesterov } 28581f1f642eSOleg Nesterov 28596e84d644SOleg Nesterov /** 2860401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2861401a8d04STejun Heo * @work: the work to cancel 28626e84d644SOleg Nesterov * 2863401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2864401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2865401a8d04STejun Heo * another workqueue. On return from this function, @work is 2866401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28671f1f642eSOleg Nesterov * 2868401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2869401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28706e84d644SOleg Nesterov * 2871401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28726e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2873401a8d04STejun Heo * 2874401a8d04STejun Heo * RETURNS: 2875401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28766e84d644SOleg Nesterov */ 2877401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28786e84d644SOleg Nesterov { 287936e227d2STejun Heo return __cancel_work_timer(work, false); 2880b89deed3SOleg Nesterov } 288128e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2882b89deed3SOleg Nesterov 28836e84d644SOleg Nesterov /** 2884401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2885401a8d04STejun Heo * @dwork: the delayed work to flush 28866e84d644SOleg Nesterov * 2887401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2888401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2889401a8d04STejun Heo * considers the last queueing instance of @dwork. 28901f1f642eSOleg Nesterov * 2891401a8d04STejun Heo * RETURNS: 2892401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2893401a8d04STejun Heo * %false if it was already idle. 28946e84d644SOleg Nesterov */ 2895401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2896401a8d04STejun Heo { 28978930cabaSTejun Heo local_irq_disable(); 2898401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 289960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 29008930cabaSTejun Heo local_irq_enable(); 2901401a8d04STejun Heo return flush_work(&dwork->work); 2902401a8d04STejun Heo } 2903401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2904401a8d04STejun Heo 2905401a8d04STejun Heo /** 290657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 290757b30ae7STejun Heo * @dwork: delayed_work to cancel 290809383498STejun Heo * 290957b30ae7STejun Heo * Kill off a pending delayed_work. Returns %true if @dwork was pending 291057b30ae7STejun Heo * and canceled; %false if wasn't pending. Note that the work callback 291157b30ae7STejun Heo * function may still be running on return, unless it returns %true and the 291257b30ae7STejun Heo * work doesn't re-arm itself. Explicitly flush or use 291357b30ae7STejun Heo * cancel_delayed_work_sync() to wait on it. 291409383498STejun Heo * 291557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 291609383498STejun Heo */ 291757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 291809383498STejun Heo { 291957b30ae7STejun Heo unsigned long flags; 292057b30ae7STejun Heo int ret; 292157b30ae7STejun Heo 292257b30ae7STejun Heo do { 292357b30ae7STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 292457b30ae7STejun Heo } while (unlikely(ret == -EAGAIN)); 292557b30ae7STejun Heo 292657b30ae7STejun Heo if (unlikely(ret < 0)) 292757b30ae7STejun Heo return false; 292857b30ae7STejun Heo 29297c3eed5cSTejun Heo set_work_pool_and_clear_pending(&dwork->work, 29307c3eed5cSTejun Heo get_work_pool_id(&dwork->work)); 293157b30ae7STejun Heo local_irq_restore(flags); 2932c0158ca6SDan Magenheimer return ret; 293309383498STejun Heo } 293457b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 293509383498STejun Heo 293609383498STejun Heo /** 2937401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2938401a8d04STejun Heo * @dwork: the delayed work cancel 2939401a8d04STejun Heo * 2940401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2941401a8d04STejun Heo * 2942401a8d04STejun Heo * RETURNS: 2943401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2944401a8d04STejun Heo */ 2945401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 29466e84d644SOleg Nesterov { 294736e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 29486e84d644SOleg Nesterov } 2949f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 29501da177e4SLinus Torvalds 29510fcb78c2SRolf Eike Beer /** 295231ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2953b6136773SAndrew Morton * @func: the function to call 2954b6136773SAndrew Morton * 295531ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 295631ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2957b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 295831ddd871STejun Heo * 295931ddd871STejun Heo * RETURNS: 296031ddd871STejun Heo * 0 on success, -errno on failure. 2961b6136773SAndrew Morton */ 296265f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 296315316ba8SChristoph Lameter { 296415316ba8SChristoph Lameter int cpu; 296538f51568SNamhyung Kim struct work_struct __percpu *works; 296615316ba8SChristoph Lameter 2967b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2968b6136773SAndrew Morton if (!works) 296915316ba8SChristoph Lameter return -ENOMEM; 2970b6136773SAndrew Morton 297195402b38SGautham R Shenoy get_online_cpus(); 297293981800STejun Heo 297315316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29749bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29759bfb1839SIngo Molnar 29769bfb1839SIngo Molnar INIT_WORK(work, func); 29778de6d308SOleg Nesterov schedule_work_on(cpu, work); 297815316ba8SChristoph Lameter } 297993981800STejun Heo 298093981800STejun Heo for_each_online_cpu(cpu) 29818616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 298293981800STejun Heo 298395402b38SGautham R Shenoy put_online_cpus(); 2984b6136773SAndrew Morton free_percpu(works); 298515316ba8SChristoph Lameter return 0; 298615316ba8SChristoph Lameter } 298715316ba8SChristoph Lameter 2988eef6a7d5SAlan Stern /** 2989eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2990eef6a7d5SAlan Stern * 2991eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2992eef6a7d5SAlan Stern * completion. 2993eef6a7d5SAlan Stern * 2994eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2995eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2996eef6a7d5SAlan Stern * will lead to deadlock: 2997eef6a7d5SAlan Stern * 2998eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2999eef6a7d5SAlan Stern * a lock held by your code or its caller. 3000eef6a7d5SAlan Stern * 3001eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3002eef6a7d5SAlan Stern * 3003eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3004eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3005eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3006eef6a7d5SAlan Stern * 3007eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3008eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3009eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3010eef6a7d5SAlan Stern * cancel_work_sync() instead. 3011eef6a7d5SAlan Stern */ 30121da177e4SLinus Torvalds void flush_scheduled_work(void) 30131da177e4SLinus Torvalds { 3014d320c038STejun Heo flush_workqueue(system_wq); 30151da177e4SLinus Torvalds } 3016ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30171da177e4SLinus Torvalds 30181da177e4SLinus Torvalds /** 30191fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30201fa44ecaSJames Bottomley * @fn: the function to execute 30211fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30221fa44ecaSJames Bottomley * be available when the work executes) 30231fa44ecaSJames Bottomley * 30241fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30251fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30261fa44ecaSJames Bottomley * 30271fa44ecaSJames Bottomley * Returns: 0 - function was executed 30281fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30291fa44ecaSJames Bottomley */ 303065f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30311fa44ecaSJames Bottomley { 30321fa44ecaSJames Bottomley if (!in_interrupt()) { 303365f27f38SDavid Howells fn(&ew->work); 30341fa44ecaSJames Bottomley return 0; 30351fa44ecaSJames Bottomley } 30361fa44ecaSJames Bottomley 303765f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30381fa44ecaSJames Bottomley schedule_work(&ew->work); 30391fa44ecaSJames Bottomley 30401fa44ecaSJames Bottomley return 1; 30411fa44ecaSJames Bottomley } 30421fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30431fa44ecaSJames Bottomley 3044226223abSTejun Heo #ifdef CONFIG_SYSFS 3045226223abSTejun Heo /* 3046226223abSTejun Heo * Workqueues with WQ_SYSFS flag set is visible to userland via 3047226223abSTejun Heo * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 3048226223abSTejun Heo * following attributes. 3049226223abSTejun Heo * 3050226223abSTejun Heo * per_cpu RO bool : whether the workqueue is per-cpu or unbound 3051226223abSTejun Heo * max_active RW int : maximum number of in-flight work items 3052226223abSTejun Heo * 3053226223abSTejun Heo * Unbound workqueues have the following extra attributes. 3054226223abSTejun Heo * 3055226223abSTejun Heo * id RO int : the associated pool ID 3056226223abSTejun Heo * nice RW int : nice value of the workers 3057226223abSTejun Heo * cpumask RW mask : bitmask of allowed CPUs for the workers 3058226223abSTejun Heo */ 3059226223abSTejun Heo struct wq_device { 3060226223abSTejun Heo struct workqueue_struct *wq; 3061226223abSTejun Heo struct device dev; 3062226223abSTejun Heo }; 3063226223abSTejun Heo 3064226223abSTejun Heo static struct workqueue_struct *dev_to_wq(struct device *dev) 3065226223abSTejun Heo { 3066226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3067226223abSTejun Heo 3068226223abSTejun Heo return wq_dev->wq; 3069226223abSTejun Heo } 3070226223abSTejun Heo 3071226223abSTejun Heo static ssize_t wq_per_cpu_show(struct device *dev, 3072226223abSTejun Heo struct device_attribute *attr, char *buf) 3073226223abSTejun Heo { 3074226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3075226223abSTejun Heo 3076226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 3077226223abSTejun Heo } 3078226223abSTejun Heo 3079226223abSTejun Heo static ssize_t wq_max_active_show(struct device *dev, 3080226223abSTejun Heo struct device_attribute *attr, char *buf) 3081226223abSTejun Heo { 3082226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3083226223abSTejun Heo 3084226223abSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 3085226223abSTejun Heo } 3086226223abSTejun Heo 3087226223abSTejun Heo static ssize_t wq_max_active_store(struct device *dev, 3088226223abSTejun Heo struct device_attribute *attr, 3089226223abSTejun Heo const char *buf, size_t count) 3090226223abSTejun Heo { 3091226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3092226223abSTejun Heo int val; 3093226223abSTejun Heo 3094226223abSTejun Heo if (sscanf(buf, "%d", &val) != 1 || val <= 0) 3095226223abSTejun Heo return -EINVAL; 3096226223abSTejun Heo 3097226223abSTejun Heo workqueue_set_max_active(wq, val); 3098226223abSTejun Heo return count; 3099226223abSTejun Heo } 3100226223abSTejun Heo 3101226223abSTejun Heo static struct device_attribute wq_sysfs_attrs[] = { 3102226223abSTejun Heo __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL), 3103226223abSTejun Heo __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store), 3104226223abSTejun Heo __ATTR_NULL, 3105226223abSTejun Heo }; 3106226223abSTejun Heo 3107d55262c4STejun Heo static ssize_t wq_pool_ids_show(struct device *dev, 3108226223abSTejun Heo struct device_attribute *attr, char *buf) 3109226223abSTejun Heo { 3110226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3111d55262c4STejun Heo const char *delim = ""; 3112d55262c4STejun Heo int node, written = 0; 3113226223abSTejun Heo 3114226223abSTejun Heo rcu_read_lock_sched(); 3115d55262c4STejun Heo for_each_node(node) { 3116d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, 3117d55262c4STejun Heo "%s%d:%d", delim, node, 3118d55262c4STejun Heo unbound_pwq_by_node(wq, node)->pool->id); 3119d55262c4STejun Heo delim = " "; 3120d55262c4STejun Heo } 3121d55262c4STejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3122226223abSTejun Heo rcu_read_unlock_sched(); 3123226223abSTejun Heo 3124226223abSTejun Heo return written; 3125226223abSTejun Heo } 3126226223abSTejun Heo 3127226223abSTejun Heo static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 3128226223abSTejun Heo char *buf) 3129226223abSTejun Heo { 3130226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3131226223abSTejun Heo int written; 3132226223abSTejun Heo 31336029a918STejun Heo mutex_lock(&wq->mutex); 31346029a918STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 31356029a918STejun Heo mutex_unlock(&wq->mutex); 3136226223abSTejun Heo 3137226223abSTejun Heo return written; 3138226223abSTejun Heo } 3139226223abSTejun Heo 3140226223abSTejun Heo /* prepare workqueue_attrs for sysfs store operations */ 3141226223abSTejun Heo static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 3142226223abSTejun Heo { 3143226223abSTejun Heo struct workqueue_attrs *attrs; 3144226223abSTejun Heo 3145226223abSTejun Heo attrs = alloc_workqueue_attrs(GFP_KERNEL); 3146226223abSTejun Heo if (!attrs) 3147226223abSTejun Heo return NULL; 3148226223abSTejun Heo 31496029a918STejun Heo mutex_lock(&wq->mutex); 31506029a918STejun Heo copy_workqueue_attrs(attrs, wq->unbound_attrs); 31516029a918STejun Heo mutex_unlock(&wq->mutex); 3152226223abSTejun Heo return attrs; 3153226223abSTejun Heo } 3154226223abSTejun Heo 3155226223abSTejun Heo static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 3156226223abSTejun Heo const char *buf, size_t count) 3157226223abSTejun Heo { 3158226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3159226223abSTejun Heo struct workqueue_attrs *attrs; 3160226223abSTejun Heo int ret; 3161226223abSTejun Heo 3162226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3163226223abSTejun Heo if (!attrs) 3164226223abSTejun Heo return -ENOMEM; 3165226223abSTejun Heo 3166226223abSTejun Heo if (sscanf(buf, "%d", &attrs->nice) == 1 && 3167226223abSTejun Heo attrs->nice >= -20 && attrs->nice <= 19) 3168226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3169226223abSTejun Heo else 3170226223abSTejun Heo ret = -EINVAL; 3171226223abSTejun Heo 3172226223abSTejun Heo free_workqueue_attrs(attrs); 3173226223abSTejun Heo return ret ?: count; 3174226223abSTejun Heo } 3175226223abSTejun Heo 3176226223abSTejun Heo static ssize_t wq_cpumask_show(struct device *dev, 3177226223abSTejun Heo struct device_attribute *attr, char *buf) 3178226223abSTejun Heo { 3179226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3180226223abSTejun Heo int written; 3181226223abSTejun Heo 31826029a918STejun Heo mutex_lock(&wq->mutex); 31836029a918STejun Heo written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask); 31846029a918STejun Heo mutex_unlock(&wq->mutex); 3185226223abSTejun Heo 3186226223abSTejun Heo written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 3187226223abSTejun Heo return written; 3188226223abSTejun Heo } 3189226223abSTejun Heo 3190226223abSTejun Heo static ssize_t wq_cpumask_store(struct device *dev, 3191226223abSTejun Heo struct device_attribute *attr, 3192226223abSTejun Heo const char *buf, size_t count) 3193226223abSTejun Heo { 3194226223abSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3195226223abSTejun Heo struct workqueue_attrs *attrs; 3196226223abSTejun Heo int ret; 3197226223abSTejun Heo 3198226223abSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 3199226223abSTejun Heo if (!attrs) 3200226223abSTejun Heo return -ENOMEM; 3201226223abSTejun Heo 3202226223abSTejun Heo ret = cpumask_parse(buf, attrs->cpumask); 3203226223abSTejun Heo if (!ret) 3204226223abSTejun Heo ret = apply_workqueue_attrs(wq, attrs); 3205226223abSTejun Heo 3206226223abSTejun Heo free_workqueue_attrs(attrs); 3207226223abSTejun Heo return ret ?: count; 3208226223abSTejun Heo } 3209226223abSTejun Heo 3210d55262c4STejun Heo static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 3211d55262c4STejun Heo char *buf) 3212d55262c4STejun Heo { 3213d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3214d55262c4STejun Heo int written; 3215d55262c4STejun Heo 3216d55262c4STejun Heo mutex_lock(&wq->mutex); 3217d55262c4STejun Heo written = scnprintf(buf, PAGE_SIZE, "%d\n", 3218d55262c4STejun Heo !wq->unbound_attrs->no_numa); 3219d55262c4STejun Heo mutex_unlock(&wq->mutex); 3220d55262c4STejun Heo 3221d55262c4STejun Heo return written; 3222d55262c4STejun Heo } 3223d55262c4STejun Heo 3224d55262c4STejun Heo static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 3225d55262c4STejun Heo const char *buf, size_t count) 3226d55262c4STejun Heo { 3227d55262c4STejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 3228d55262c4STejun Heo struct workqueue_attrs *attrs; 3229d55262c4STejun Heo int v, ret; 3230d55262c4STejun Heo 3231d55262c4STejun Heo attrs = wq_sysfs_prep_attrs(wq); 3232d55262c4STejun Heo if (!attrs) 3233d55262c4STejun Heo return -ENOMEM; 3234d55262c4STejun Heo 3235d55262c4STejun Heo ret = -EINVAL; 3236d55262c4STejun Heo if (sscanf(buf, "%d", &v) == 1) { 3237d55262c4STejun Heo attrs->no_numa = !v; 3238d55262c4STejun Heo ret = apply_workqueue_attrs(wq, attrs); 3239d55262c4STejun Heo } 3240d55262c4STejun Heo 3241d55262c4STejun Heo free_workqueue_attrs(attrs); 3242d55262c4STejun Heo return ret ?: count; 3243d55262c4STejun Heo } 3244d55262c4STejun Heo 3245226223abSTejun Heo static struct device_attribute wq_sysfs_unbound_attrs[] = { 3246d55262c4STejun Heo __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 3247226223abSTejun Heo __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 3248226223abSTejun Heo __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 3249d55262c4STejun Heo __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 3250226223abSTejun Heo __ATTR_NULL, 3251226223abSTejun Heo }; 3252226223abSTejun Heo 3253226223abSTejun Heo static struct bus_type wq_subsys = { 3254226223abSTejun Heo .name = "workqueue", 3255226223abSTejun Heo .dev_attrs = wq_sysfs_attrs, 3256226223abSTejun Heo }; 3257226223abSTejun Heo 3258226223abSTejun Heo static int __init wq_sysfs_init(void) 3259226223abSTejun Heo { 3260226223abSTejun Heo return subsys_virtual_register(&wq_subsys, NULL); 3261226223abSTejun Heo } 3262226223abSTejun Heo core_initcall(wq_sysfs_init); 3263226223abSTejun Heo 3264226223abSTejun Heo static void wq_device_release(struct device *dev) 3265226223abSTejun Heo { 3266226223abSTejun Heo struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 3267226223abSTejun Heo 3268226223abSTejun Heo kfree(wq_dev); 3269226223abSTejun Heo } 3270226223abSTejun Heo 3271226223abSTejun Heo /** 3272226223abSTejun Heo * workqueue_sysfs_register - make a workqueue visible in sysfs 3273226223abSTejun Heo * @wq: the workqueue to register 3274226223abSTejun Heo * 3275226223abSTejun Heo * Expose @wq in sysfs under /sys/bus/workqueue/devices. 3276226223abSTejun Heo * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 3277226223abSTejun Heo * which is the preferred method. 3278226223abSTejun Heo * 3279226223abSTejun Heo * Workqueue user should use this function directly iff it wants to apply 3280226223abSTejun Heo * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 3281226223abSTejun Heo * apply_workqueue_attrs() may race against userland updating the 3282226223abSTejun Heo * attributes. 3283226223abSTejun Heo * 3284226223abSTejun Heo * Returns 0 on success, -errno on failure. 3285226223abSTejun Heo */ 3286226223abSTejun Heo int workqueue_sysfs_register(struct workqueue_struct *wq) 3287226223abSTejun Heo { 3288226223abSTejun Heo struct wq_device *wq_dev; 3289226223abSTejun Heo int ret; 3290226223abSTejun Heo 3291226223abSTejun Heo /* 3292226223abSTejun Heo * Adjusting max_active or creating new pwqs by applyting 3293226223abSTejun Heo * attributes breaks ordering guarantee. Disallow exposing ordered 3294226223abSTejun Heo * workqueues. 3295226223abSTejun Heo */ 3296226223abSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 3297226223abSTejun Heo return -EINVAL; 3298226223abSTejun Heo 3299226223abSTejun Heo wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 3300226223abSTejun Heo if (!wq_dev) 3301226223abSTejun Heo return -ENOMEM; 3302226223abSTejun Heo 3303226223abSTejun Heo wq_dev->wq = wq; 3304226223abSTejun Heo wq_dev->dev.bus = &wq_subsys; 3305226223abSTejun Heo wq_dev->dev.init_name = wq->name; 3306226223abSTejun Heo wq_dev->dev.release = wq_device_release; 3307226223abSTejun Heo 3308226223abSTejun Heo /* 3309226223abSTejun Heo * unbound_attrs are created separately. Suppress uevent until 3310226223abSTejun Heo * everything is ready. 3311226223abSTejun Heo */ 3312226223abSTejun Heo dev_set_uevent_suppress(&wq_dev->dev, true); 3313226223abSTejun Heo 3314226223abSTejun Heo ret = device_register(&wq_dev->dev); 3315226223abSTejun Heo if (ret) { 3316226223abSTejun Heo kfree(wq_dev); 3317226223abSTejun Heo wq->wq_dev = NULL; 3318226223abSTejun Heo return ret; 3319226223abSTejun Heo } 3320226223abSTejun Heo 3321226223abSTejun Heo if (wq->flags & WQ_UNBOUND) { 3322226223abSTejun Heo struct device_attribute *attr; 3323226223abSTejun Heo 3324226223abSTejun Heo for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 3325226223abSTejun Heo ret = device_create_file(&wq_dev->dev, attr); 3326226223abSTejun Heo if (ret) { 3327226223abSTejun Heo device_unregister(&wq_dev->dev); 3328226223abSTejun Heo wq->wq_dev = NULL; 3329226223abSTejun Heo return ret; 3330226223abSTejun Heo } 3331226223abSTejun Heo } 3332226223abSTejun Heo } 3333226223abSTejun Heo 3334226223abSTejun Heo kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 3335226223abSTejun Heo return 0; 3336226223abSTejun Heo } 3337226223abSTejun Heo 3338226223abSTejun Heo /** 3339226223abSTejun Heo * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 3340226223abSTejun Heo * @wq: the workqueue to unregister 3341226223abSTejun Heo * 3342226223abSTejun Heo * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 3343226223abSTejun Heo */ 3344226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 3345226223abSTejun Heo { 3346226223abSTejun Heo struct wq_device *wq_dev = wq->wq_dev; 3347226223abSTejun Heo 3348226223abSTejun Heo if (!wq->wq_dev) 3349226223abSTejun Heo return; 3350226223abSTejun Heo 3351226223abSTejun Heo wq->wq_dev = NULL; 3352226223abSTejun Heo device_unregister(&wq_dev->dev); 3353226223abSTejun Heo } 3354226223abSTejun Heo #else /* CONFIG_SYSFS */ 3355226223abSTejun Heo static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 3356226223abSTejun Heo #endif /* CONFIG_SYSFS */ 3357226223abSTejun Heo 33587a4e344cSTejun Heo /** 33597a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33607a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33617a4e344cSTejun Heo * 33627a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33637a4e344cSTejun Heo */ 33647a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 33657a4e344cSTejun Heo { 33667a4e344cSTejun Heo if (attrs) { 33677a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33687a4e344cSTejun Heo kfree(attrs); 33697a4e344cSTejun Heo } 33707a4e344cSTejun Heo } 33717a4e344cSTejun Heo 33727a4e344cSTejun Heo /** 33737a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33747a4e344cSTejun Heo * @gfp_mask: allocation mask to use 33757a4e344cSTejun Heo * 33767a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 33777a4e344cSTejun Heo * return it. Returns NULL on failure. 33787a4e344cSTejun Heo */ 33797a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 33807a4e344cSTejun Heo { 33817a4e344cSTejun Heo struct workqueue_attrs *attrs; 33827a4e344cSTejun Heo 33837a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 33847a4e344cSTejun Heo if (!attrs) 33857a4e344cSTejun Heo goto fail; 33867a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 33877a4e344cSTejun Heo goto fail; 33887a4e344cSTejun Heo 338913e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33907a4e344cSTejun Heo return attrs; 33917a4e344cSTejun Heo fail: 33927a4e344cSTejun Heo free_workqueue_attrs(attrs); 33937a4e344cSTejun Heo return NULL; 33947a4e344cSTejun Heo } 33957a4e344cSTejun Heo 339629c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 339729c91e99STejun Heo const struct workqueue_attrs *from) 339829c91e99STejun Heo { 339929c91e99STejun Heo to->nice = from->nice; 340029c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 340129c91e99STejun Heo } 340229c91e99STejun Heo 340329c91e99STejun Heo /* hash value of the content of @attr */ 340429c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 340529c91e99STejun Heo { 340629c91e99STejun Heo u32 hash = 0; 340729c91e99STejun Heo 340829c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 340913e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 341013e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 341129c91e99STejun Heo return hash; 341229c91e99STejun Heo } 341329c91e99STejun Heo 341429c91e99STejun Heo /* content equality test */ 341529c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 341629c91e99STejun Heo const struct workqueue_attrs *b) 341729c91e99STejun Heo { 341829c91e99STejun Heo if (a->nice != b->nice) 341929c91e99STejun Heo return false; 342029c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 342129c91e99STejun Heo return false; 342229c91e99STejun Heo return true; 342329c91e99STejun Heo } 342429c91e99STejun Heo 34257a4e344cSTejun Heo /** 34267a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34277a4e344cSTejun Heo * @pool: worker_pool to initialize 34287a4e344cSTejun Heo * 34297a4e344cSTejun Heo * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs. 343029c91e99STejun Heo * Returns 0 on success, -errno on failure. Even on failure, all fields 343129c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 343229c91e99STejun Heo * on @pool safely to release it. 34337a4e344cSTejun Heo */ 34347a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34354e1a1f9aSTejun Heo { 34364e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 343729c91e99STejun Heo pool->id = -1; 343829c91e99STejun Heo pool->cpu = -1; 3439f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34404e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 34414e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34424e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34434e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34444e1a1f9aSTejun Heo 34454e1a1f9aSTejun Heo init_timer_deferrable(&pool->idle_timer); 34464e1a1f9aSTejun Heo pool->idle_timer.function = idle_worker_timeout; 34474e1a1f9aSTejun Heo pool->idle_timer.data = (unsigned long)pool; 34484e1a1f9aSTejun Heo 34494e1a1f9aSTejun Heo setup_timer(&pool->mayday_timer, pool_mayday_timeout, 34504e1a1f9aSTejun Heo (unsigned long)pool); 34514e1a1f9aSTejun Heo 34524e1a1f9aSTejun Heo mutex_init(&pool->manager_arb); 3453bc3a1afcSTejun Heo mutex_init(&pool->manager_mutex); 3454822d8405STejun Heo idr_init(&pool->worker_idr); 34557a4e344cSTejun Heo 345629c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 345729c91e99STejun Heo pool->refcnt = 1; 345829c91e99STejun Heo 345929c91e99STejun Heo /* shouldn't fail above this point */ 34607a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 34617a4e344cSTejun Heo if (!pool->attrs) 34627a4e344cSTejun Heo return -ENOMEM; 34637a4e344cSTejun Heo return 0; 34644e1a1f9aSTejun Heo } 34654e1a1f9aSTejun Heo 346629c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 346729c91e99STejun Heo { 346829c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 346929c91e99STejun Heo 3470822d8405STejun Heo idr_destroy(&pool->worker_idr); 347129c91e99STejun Heo free_workqueue_attrs(pool->attrs); 347229c91e99STejun Heo kfree(pool); 347329c91e99STejun Heo } 347429c91e99STejun Heo 347529c91e99STejun Heo /** 347629c91e99STejun Heo * put_unbound_pool - put a worker_pool 347729c91e99STejun Heo * @pool: worker_pool to put 347829c91e99STejun Heo * 347929c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3480c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3481c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3482c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3483a892caccSTejun Heo * 3484a892caccSTejun Heo * Should be called with wq_pool_mutex held. 348529c91e99STejun Heo */ 348629c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 348729c91e99STejun Heo { 348829c91e99STejun Heo struct worker *worker; 348929c91e99STejun Heo 3490a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3491a892caccSTejun Heo 3492a892caccSTejun Heo if (--pool->refcnt) 349329c91e99STejun Heo return; 349429c91e99STejun Heo 349529c91e99STejun Heo /* sanity checks */ 349629c91e99STejun Heo if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) || 3497a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 349829c91e99STejun Heo return; 349929c91e99STejun Heo 350029c91e99STejun Heo /* release id and unhash */ 350129c91e99STejun Heo if (pool->id >= 0) 350229c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 350329c91e99STejun Heo hash_del(&pool->hash_node); 350429c91e99STejun Heo 3505c5aa87bbSTejun Heo /* 3506c5aa87bbSTejun Heo * Become the manager and destroy all workers. Grabbing 3507c5aa87bbSTejun Heo * manager_arb prevents @pool's workers from blocking on 3508c5aa87bbSTejun Heo * manager_mutex. 3509c5aa87bbSTejun Heo */ 351029c91e99STejun Heo mutex_lock(&pool->manager_arb); 3511cd549687STejun Heo mutex_lock(&pool->manager_mutex); 351229c91e99STejun Heo spin_lock_irq(&pool->lock); 351329c91e99STejun Heo 351429c91e99STejun Heo while ((worker = first_worker(pool))) 351529c91e99STejun Heo destroy_worker(worker); 351629c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 351729c91e99STejun Heo 351829c91e99STejun Heo spin_unlock_irq(&pool->lock); 3519cd549687STejun Heo mutex_unlock(&pool->manager_mutex); 352029c91e99STejun Heo mutex_unlock(&pool->manager_arb); 352129c91e99STejun Heo 352229c91e99STejun Heo /* shut down the timers */ 352329c91e99STejun Heo del_timer_sync(&pool->idle_timer); 352429c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 352529c91e99STejun Heo 352629c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 352729c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 352829c91e99STejun Heo } 352929c91e99STejun Heo 353029c91e99STejun Heo /** 353129c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 353229c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 353329c91e99STejun Heo * 353429c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 353529c91e99STejun Heo * reference count and return it. If there already is a matching 353629c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 353729c91e99STejun Heo * create a new one. On failure, returns NULL. 3538a892caccSTejun Heo * 3539a892caccSTejun Heo * Should be called with wq_pool_mutex held. 354029c91e99STejun Heo */ 354129c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 354229c91e99STejun Heo { 354329c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 354429c91e99STejun Heo struct worker_pool *pool; 3545f3f90ad4STejun Heo int node; 354629c91e99STejun Heo 3547a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 354829c91e99STejun Heo 354929c91e99STejun Heo /* do we already have a matching pool? */ 355029c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 355129c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 355229c91e99STejun Heo pool->refcnt++; 355329c91e99STejun Heo goto out_unlock; 355429c91e99STejun Heo } 355529c91e99STejun Heo } 355629c91e99STejun Heo 355729c91e99STejun Heo /* nope, create a new one */ 355829c91e99STejun Heo pool = kzalloc(sizeof(*pool), GFP_KERNEL); 355929c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 356029c91e99STejun Heo goto fail; 356129c91e99STejun Heo 356212ee4fc6SLai Jiangshan if (workqueue_freezing) 356312ee4fc6SLai Jiangshan pool->flags |= POOL_FREEZING; 356412ee4fc6SLai Jiangshan 35658864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 356629c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 356729c91e99STejun Heo 3568f3f90ad4STejun Heo /* if cpumask is contained inside a NUMA node, we belong to that node */ 3569f3f90ad4STejun Heo if (wq_numa_enabled) { 3570f3f90ad4STejun Heo for_each_node(node) { 3571f3f90ad4STejun Heo if (cpumask_subset(pool->attrs->cpumask, 3572f3f90ad4STejun Heo wq_numa_possible_cpumask[node])) { 3573f3f90ad4STejun Heo pool->node = node; 3574f3f90ad4STejun Heo break; 3575f3f90ad4STejun Heo } 3576f3f90ad4STejun Heo } 3577f3f90ad4STejun Heo } 3578f3f90ad4STejun Heo 357929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 358029c91e99STejun Heo goto fail; 358129c91e99STejun Heo 358229c91e99STejun Heo /* create and start the initial worker */ 3583ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 358429c91e99STejun Heo goto fail; 358529c91e99STejun Heo 358629c91e99STejun Heo /* install */ 358729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 358829c91e99STejun Heo out_unlock: 358929c91e99STejun Heo return pool; 359029c91e99STejun Heo fail: 359129c91e99STejun Heo if (pool) 359229c91e99STejun Heo put_unbound_pool(pool); 359329c91e99STejun Heo return NULL; 359429c91e99STejun Heo } 359529c91e99STejun Heo 35968864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 35978864b4e5STejun Heo { 35988864b4e5STejun Heo kmem_cache_free(pwq_cache, 35998864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36008864b4e5STejun Heo } 36018864b4e5STejun Heo 36028864b4e5STejun Heo /* 36038864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36048864b4e5STejun Heo * and needs to be destroyed. 36058864b4e5STejun Heo */ 36068864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36078864b4e5STejun Heo { 36088864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36098864b4e5STejun Heo unbound_release_work); 36108864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36118864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3612bc0caf09STejun Heo bool is_last; 36138864b4e5STejun Heo 36148864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36158864b4e5STejun Heo return; 36168864b4e5STejun Heo 361775ccf595STejun Heo /* 36183c25a55dSLai Jiangshan * Unlink @pwq. Synchronization against wq->mutex isn't strictly 361975ccf595STejun Heo * necessary on release but do it anyway. It's easier to verify 362075ccf595STejun Heo * and consistent with the linking path. 362175ccf595STejun Heo */ 36223c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36238864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3624bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36253c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 36268864b4e5STejun Heo 3627a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36288864b4e5STejun Heo put_unbound_pool(pool); 3629a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3630a892caccSTejun Heo 36318864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 36328864b4e5STejun Heo 36338864b4e5STejun Heo /* 36348864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 36358864b4e5STejun Heo * is gonna access it anymore. Free it. 36368864b4e5STejun Heo */ 36376029a918STejun Heo if (is_last) { 36386029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 36398864b4e5STejun Heo kfree(wq); 36408864b4e5STejun Heo } 36416029a918STejun Heo } 36428864b4e5STejun Heo 36430fbd95aaSTejun Heo /** 3644699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 36450fbd95aaSTejun Heo * @pwq: target pool_workqueue 36460fbd95aaSTejun Heo * 3647699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3648699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3649699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 36500fbd95aaSTejun Heo */ 3651699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 36520fbd95aaSTejun Heo { 3653699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3654699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 3655699ce097STejun Heo 3656699ce097STejun Heo /* for @wq->saved_max_active */ 3657a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3658699ce097STejun Heo 3659699ce097STejun Heo /* fast exit for non-freezable wqs */ 3660699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3661699ce097STejun Heo return; 3662699ce097STejun Heo 3663a357fc03SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 3664699ce097STejun Heo 3665699ce097STejun Heo if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) { 3666699ce097STejun Heo pwq->max_active = wq->saved_max_active; 36670fbd95aaSTejun Heo 36680fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 36690fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 36700fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3671951a078aSLai Jiangshan 3672951a078aSLai Jiangshan /* 3673951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3674951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3675951a078aSLai Jiangshan */ 3676951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3677699ce097STejun Heo } else { 3678699ce097STejun Heo pwq->max_active = 0; 3679699ce097STejun Heo } 3680699ce097STejun Heo 3681a357fc03SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 36820fbd95aaSTejun Heo } 36830fbd95aaSTejun Heo 3684e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3685f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3686f147f29eSTejun Heo struct worker_pool *pool) 3687d2c1d404STejun Heo { 3688d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3689d2c1d404STejun Heo 3690e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3691e50aba9aSTejun Heo 3692d2c1d404STejun Heo pwq->pool = pool; 3693d2c1d404STejun Heo pwq->wq = wq; 3694d2c1d404STejun Heo pwq->flush_color = -1; 36958864b4e5STejun Heo pwq->refcnt = 1; 3696d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 36971befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3698d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 36998864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3700f147f29eSTejun Heo } 3701d2c1d404STejun Heo 3702f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37031befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3704f147f29eSTejun Heo { 3705f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3706f147f29eSTejun Heo 3707f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 370875ccf595STejun Heo 37091befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37101befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37111befcf30STejun Heo return; 37121befcf30STejun Heo 3713983ca25eSTejun Heo /* 3714983ca25eSTejun Heo * Set the matching work_color. This is synchronized with 37153c25a55dSLai Jiangshan * wq->mutex to avoid confusing flush_workqueue(). 3716983ca25eSTejun Heo */ 371775ccf595STejun Heo pwq->work_color = wq->work_color; 3718983ca25eSTejun Heo 3719983ca25eSTejun Heo /* sync max_active to the current setting */ 3720983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3721983ca25eSTejun Heo 3722983ca25eSTejun Heo /* link in @pwq */ 37239e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3724df2d5ae4STejun Heo } 37256029a918STejun Heo 3726f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3727f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3728f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3729f147f29eSTejun Heo { 3730f147f29eSTejun Heo struct worker_pool *pool; 3731f147f29eSTejun Heo struct pool_workqueue *pwq; 3732f147f29eSTejun Heo 3733f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3734f147f29eSTejun Heo 3735f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3736f147f29eSTejun Heo if (!pool) 3737f147f29eSTejun Heo return NULL; 3738f147f29eSTejun Heo 3739e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3740f147f29eSTejun Heo if (!pwq) { 3741f147f29eSTejun Heo put_unbound_pool(pool); 3742f147f29eSTejun Heo return NULL; 3743f147f29eSTejun Heo } 3744f147f29eSTejun Heo 3745f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3746f147f29eSTejun Heo return pwq; 3747d2c1d404STejun Heo } 3748d2c1d404STejun Heo 37494c16bd32STejun Heo /* undo alloc_unbound_pwq(), used only in the error path */ 37504c16bd32STejun Heo static void free_unbound_pwq(struct pool_workqueue *pwq) 37514c16bd32STejun Heo { 37524c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 37534c16bd32STejun Heo 37544c16bd32STejun Heo if (pwq) { 37554c16bd32STejun Heo put_unbound_pool(pwq->pool); 3756cece95dfSWei Yongjun kmem_cache_free(pwq_cache, pwq); 37574c16bd32STejun Heo } 37584c16bd32STejun Heo } 37594c16bd32STejun Heo 37604c16bd32STejun Heo /** 37614c16bd32STejun Heo * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node 37624c16bd32STejun Heo * @attrs: the wq_attrs of interest 37634c16bd32STejun Heo * @node: the target NUMA node 37644c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 37654c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 37664c16bd32STejun Heo * 37674c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 37684c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 37694c16bd32STejun Heo * calculation. The result is stored in @cpumask. This function returns 37704c16bd32STejun Heo * %true if the resulting @cpumask is different from @attrs->cpumask, 37714c16bd32STejun Heo * %false if equal. 37724c16bd32STejun Heo * 37734c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 37744c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 37754c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 37764c16bd32STejun Heo * @attrs->cpumask. 37774c16bd32STejun Heo * 37784c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 37794c16bd32STejun Heo * stable. 37804c16bd32STejun Heo */ 37814c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 37824c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 37834c16bd32STejun Heo { 3784d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 37854c16bd32STejun Heo goto use_dfl; 37864c16bd32STejun Heo 37874c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 37884c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 37894c16bd32STejun Heo if (cpu_going_down >= 0) 37904c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 37914c16bd32STejun Heo 37924c16bd32STejun Heo if (cpumask_empty(cpumask)) 37934c16bd32STejun Heo goto use_dfl; 37944c16bd32STejun Heo 37954c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 37964c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 37974c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 37984c16bd32STejun Heo 37994c16bd32STejun Heo use_dfl: 38004c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38014c16bd32STejun Heo return false; 38024c16bd32STejun Heo } 38034c16bd32STejun Heo 38041befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38051befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38061befcf30STejun Heo int node, 38071befcf30STejun Heo struct pool_workqueue *pwq) 38081befcf30STejun Heo { 38091befcf30STejun Heo struct pool_workqueue *old_pwq; 38101befcf30STejun Heo 38111befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38121befcf30STejun Heo 38131befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38141befcf30STejun Heo link_pwq(pwq); 38151befcf30STejun Heo 38161befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38171befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38181befcf30STejun Heo return old_pwq; 38191befcf30STejun Heo } 38201befcf30STejun Heo 38219e8cd2f5STejun Heo /** 38229e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38239e8cd2f5STejun Heo * @wq: the target workqueue 38249e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38259e8cd2f5STejun Heo * 38264c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38274c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38284c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38294c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38304c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38314c16bd32STejun Heo * back-to-back will stay on its current pwq. 38329e8cd2f5STejun Heo * 38339e8cd2f5STejun Heo * Performs GFP_KERNEL allocations. Returns 0 on success and -errno on 38349e8cd2f5STejun Heo * failure. 38359e8cd2f5STejun Heo */ 38369e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38379e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38389e8cd2f5STejun Heo { 38394c16bd32STejun Heo struct workqueue_attrs *new_attrs, *tmp_attrs; 38404c16bd32STejun Heo struct pool_workqueue **pwq_tbl, *dfl_pwq; 3841f147f29eSTejun Heo int node, ret; 38429e8cd2f5STejun Heo 38438719dceaSTejun Heo /* only unbound workqueues can change attributes */ 38449e8cd2f5STejun Heo if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 38459e8cd2f5STejun Heo return -EINVAL; 38469e8cd2f5STejun Heo 38478719dceaSTejun Heo /* creating multiple pwqs breaks ordering guarantee */ 38488719dceaSTejun Heo if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))) 38498719dceaSTejun Heo return -EINVAL; 38508719dceaSTejun Heo 38514c16bd32STejun Heo pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL); 385213e2e556STejun Heo new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38534c16bd32STejun Heo tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 38544c16bd32STejun Heo if (!pwq_tbl || !new_attrs || !tmp_attrs) 385513e2e556STejun Heo goto enomem; 385613e2e556STejun Heo 38574c16bd32STejun Heo /* make a copy of @attrs and sanitize it */ 385813e2e556STejun Heo copy_workqueue_attrs(new_attrs, attrs); 385913e2e556STejun Heo cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 386013e2e556STejun Heo 38614c16bd32STejun Heo /* 38624c16bd32STejun Heo * We may create multiple pwqs with differing cpumasks. Make a 38634c16bd32STejun Heo * copy of @new_attrs which will be modified and used to obtain 38644c16bd32STejun Heo * pools. 38654c16bd32STejun Heo */ 38664c16bd32STejun Heo copy_workqueue_attrs(tmp_attrs, new_attrs); 38679e8cd2f5STejun Heo 38684c16bd32STejun Heo /* 38694c16bd32STejun Heo * CPUs should stay stable across pwq creations and installations. 38704c16bd32STejun Heo * Pin CPUs, determine the target cpumask for each node and create 38714c16bd32STejun Heo * pwqs accordingly. 38724c16bd32STejun Heo */ 38734c16bd32STejun Heo get_online_cpus(); 38744c16bd32STejun Heo 38754c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 38764c16bd32STejun Heo 38774c16bd32STejun Heo /* 38784c16bd32STejun Heo * If something goes wrong during CPU up/down, we'll fall back to 38794c16bd32STejun Heo * the default pwq covering whole @attrs->cpumask. Always create 38804c16bd32STejun Heo * it even if we don't use it immediately. 38814c16bd32STejun Heo */ 38824c16bd32STejun Heo dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 38834c16bd32STejun Heo if (!dfl_pwq) 38844c16bd32STejun Heo goto enomem_pwq; 38854c16bd32STejun Heo 38864c16bd32STejun Heo for_each_node(node) { 38874c16bd32STejun Heo if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) { 38884c16bd32STejun Heo pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 38894c16bd32STejun Heo if (!pwq_tbl[node]) 38904c16bd32STejun Heo goto enomem_pwq; 38914c16bd32STejun Heo } else { 38924c16bd32STejun Heo dfl_pwq->refcnt++; 38934c16bd32STejun Heo pwq_tbl[node] = dfl_pwq; 38944c16bd32STejun Heo } 38954c16bd32STejun Heo } 38964c16bd32STejun Heo 38974c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 38984c16bd32STejun Heo 38994c16bd32STejun Heo /* all pwqs have been created successfully, let's install'em */ 3900f147f29eSTejun Heo mutex_lock(&wq->mutex); 3901a892caccSTejun Heo 3902f147f29eSTejun Heo copy_workqueue_attrs(wq->unbound_attrs, new_attrs); 39034c16bd32STejun Heo 39044c16bd32STejun Heo /* save the previous pwq and install the new one */ 3905f147f29eSTejun Heo for_each_node(node) 39064c16bd32STejun Heo pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]); 39074c16bd32STejun Heo 39084c16bd32STejun Heo /* @dfl_pwq might not have been used, ensure it's linked */ 39094c16bd32STejun Heo link_pwq(dfl_pwq); 39104c16bd32STejun Heo swap(wq->dfl_pwq, dfl_pwq); 3911f147f29eSTejun Heo 3912f147f29eSTejun Heo mutex_unlock(&wq->mutex); 3913f147f29eSTejun Heo 39144c16bd32STejun Heo /* put the old pwqs */ 39154c16bd32STejun Heo for_each_node(node) 39164c16bd32STejun Heo put_pwq_unlocked(pwq_tbl[node]); 39174c16bd32STejun Heo put_pwq_unlocked(dfl_pwq); 39184c16bd32STejun Heo 39194c16bd32STejun Heo put_online_cpus(); 39204862125bSTejun Heo ret = 0; 39214862125bSTejun Heo /* fall through */ 39224862125bSTejun Heo out_free: 39234c16bd32STejun Heo free_workqueue_attrs(tmp_attrs); 39244862125bSTejun Heo free_workqueue_attrs(new_attrs); 39254c16bd32STejun Heo kfree(pwq_tbl); 39264862125bSTejun Heo return ret; 392713e2e556STejun Heo 39284c16bd32STejun Heo enomem_pwq: 39294c16bd32STejun Heo free_unbound_pwq(dfl_pwq); 39304c16bd32STejun Heo for_each_node(node) 39314c16bd32STejun Heo if (pwq_tbl && pwq_tbl[node] != dfl_pwq) 39324c16bd32STejun Heo free_unbound_pwq(pwq_tbl[node]); 39334c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 39344c16bd32STejun Heo put_online_cpus(); 393513e2e556STejun Heo enomem: 39364862125bSTejun Heo ret = -ENOMEM; 39374862125bSTejun Heo goto out_free; 39389e8cd2f5STejun Heo } 39399e8cd2f5STejun Heo 39404c16bd32STejun Heo /** 39414c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 39424c16bd32STejun Heo * @wq: the target workqueue 39434c16bd32STejun Heo * @cpu: the CPU coming up or going down 39444c16bd32STejun Heo * @online: whether @cpu is coming up or going down 39454c16bd32STejun Heo * 39464c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 39474c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 39484c16bd32STejun Heo * @wq accordingly. 39494c16bd32STejun Heo * 39504c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 39514c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 39524c16bd32STejun Heo * correct. 39534c16bd32STejun Heo * 39544c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 39554c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 39564c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 39574c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 39584c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 39594c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 39604c16bd32STejun Heo * CPU_DOWN_PREPARE. 39614c16bd32STejun Heo */ 39624c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 39634c16bd32STejun Heo bool online) 39644c16bd32STejun Heo { 39654c16bd32STejun Heo int node = cpu_to_node(cpu); 39664c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 39674c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 39684c16bd32STejun Heo struct workqueue_attrs *target_attrs; 39694c16bd32STejun Heo cpumask_t *cpumask; 39704c16bd32STejun Heo 39714c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 39724c16bd32STejun Heo 39734c16bd32STejun Heo if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND)) 39744c16bd32STejun Heo return; 39754c16bd32STejun Heo 39764c16bd32STejun Heo /* 39774c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39784c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39794c16bd32STejun Heo * CPU hotplug exclusion. 39804c16bd32STejun Heo */ 39814c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39824c16bd32STejun Heo cpumask = target_attrs->cpumask; 39834c16bd32STejun Heo 39844c16bd32STejun Heo mutex_lock(&wq->mutex); 3985d55262c4STejun Heo if (wq->unbound_attrs->no_numa) 3986d55262c4STejun Heo goto out_unlock; 39874c16bd32STejun Heo 39884c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39894c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39904c16bd32STejun Heo 39914c16bd32STejun Heo /* 39924c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 39934c16bd32STejun Heo * different from wq's, we need to compare it to @pwq's and create 39944c16bd32STejun Heo * a new one if they don't match. If the target cpumask equals 39954c16bd32STejun Heo * wq's, the default pwq should be used. If @pwq is already the 39964c16bd32STejun Heo * default one, nothing to do; otherwise, install the default one. 39974c16bd32STejun Heo */ 39984c16bd32STejun Heo if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) { 39994c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 40004c16bd32STejun Heo goto out_unlock; 40014c16bd32STejun Heo } else { 40024c16bd32STejun Heo if (pwq == wq->dfl_pwq) 40034c16bd32STejun Heo goto out_unlock; 40044c16bd32STejun Heo else 40054c16bd32STejun Heo goto use_dfl_pwq; 40064c16bd32STejun Heo } 40074c16bd32STejun Heo 40084c16bd32STejun Heo mutex_unlock(&wq->mutex); 40094c16bd32STejun Heo 40104c16bd32STejun Heo /* create a new pwq */ 40114c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 40124c16bd32STejun Heo if (!pwq) { 40134c16bd32STejun Heo pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 40144c16bd32STejun Heo wq->name); 40154c16bd32STejun Heo goto out_unlock; 40164c16bd32STejun Heo } 40174c16bd32STejun Heo 40184c16bd32STejun Heo /* 40194c16bd32STejun Heo * Install the new pwq. As this function is called only from CPU 40204c16bd32STejun Heo * hotplug callbacks and applying a new attrs is wrapped with 40214c16bd32STejun Heo * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed 40224c16bd32STejun Heo * inbetween. 40234c16bd32STejun Heo */ 40244c16bd32STejun Heo mutex_lock(&wq->mutex); 40254c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 40264c16bd32STejun Heo goto out_unlock; 40274c16bd32STejun Heo 40284c16bd32STejun Heo use_dfl_pwq: 40294c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 40304c16bd32STejun Heo get_pwq(wq->dfl_pwq); 40314c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 40324c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 40334c16bd32STejun Heo out_unlock: 40344c16bd32STejun Heo mutex_unlock(&wq->mutex); 40354c16bd32STejun Heo put_pwq_unlocked(old_pwq); 40364c16bd32STejun Heo } 40374c16bd32STejun Heo 403830cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 40391da177e4SLinus Torvalds { 404049e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 404130cdf249STejun Heo int cpu; 4042e1d8aa9fSFrederic Weisbecker 404330cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4044420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4045420c0ddbSTejun Heo if (!wq->cpu_pwqs) 404630cdf249STejun Heo return -ENOMEM; 404730cdf249STejun Heo 404830cdf249STejun Heo for_each_possible_cpu(cpu) { 40497fb98ea7STejun Heo struct pool_workqueue *pwq = 40507fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 40517a62c2c8STejun Heo struct worker_pool *cpu_pools = 4052f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 405330cdf249STejun Heo 4054f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4055f147f29eSTejun Heo 4056f147f29eSTejun Heo mutex_lock(&wq->mutex); 40571befcf30STejun Heo link_pwq(pwq); 4058f147f29eSTejun Heo mutex_unlock(&wq->mutex); 405930cdf249STejun Heo } 406030cdf249STejun Heo return 0; 40619e8cd2f5STejun Heo } else { 40629e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 40639e8cd2f5STejun Heo } 40640f900049STejun Heo } 40650f900049STejun Heo 4066f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4067f3421797STejun Heo const char *name) 4068b71ab8c2STejun Heo { 4069f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4070f3421797STejun Heo 4071f3421797STejun Heo if (max_active < 1 || max_active > lim) 4072044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4073f3421797STejun Heo max_active, name, 1, lim); 4074b71ab8c2STejun Heo 4075f3421797STejun Heo return clamp_val(max_active, 1, lim); 4076b71ab8c2STejun Heo } 4077b71ab8c2STejun Heo 4078b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 407997e37d7bSTejun Heo unsigned int flags, 40801e19ffc6STejun Heo int max_active, 4081eb13ba87SJohannes Berg struct lock_class_key *key, 4082b196be89STejun Heo const char *lock_name, ...) 40833af24433SOleg Nesterov { 4084df2d5ae4STejun Heo size_t tbl_size = 0; 4085ecf6881fSTejun Heo va_list args; 40863af24433SOleg Nesterov struct workqueue_struct *wq; 408749e3cf44STejun Heo struct pool_workqueue *pwq; 4088b196be89STejun Heo 4089ecf6881fSTejun Heo /* allocate wq and format name */ 4090df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4091df2d5ae4STejun Heo tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]); 4092df2d5ae4STejun Heo 4093df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4094b196be89STejun Heo if (!wq) 4095d2c1d404STejun Heo return NULL; 4096b196be89STejun Heo 40976029a918STejun Heo if (flags & WQ_UNBOUND) { 40986029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40996029a918STejun Heo if (!wq->unbound_attrs) 41006029a918STejun Heo goto err_free_wq; 41016029a918STejun Heo } 41026029a918STejun Heo 4103ecf6881fSTejun Heo va_start(args, lock_name); 4104ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4105b196be89STejun Heo va_end(args); 41063af24433SOleg Nesterov 4107d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4108b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 41093af24433SOleg Nesterov 4110b196be89STejun Heo /* init wq */ 411197e37d7bSTejun Heo wq->flags = flags; 4112a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 41133c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4114112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 411530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 411673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 411773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4118493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 41193af24433SOleg Nesterov 4120eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4121cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 41223af24433SOleg Nesterov 412330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4124d2c1d404STejun Heo goto err_free_wq; 41251537663fSTejun Heo 4126493008a8STejun Heo /* 4127493008a8STejun Heo * Workqueues which may be used during memory reclaim should 4128493008a8STejun Heo * have a rescuer to guarantee forward progress. 4129493008a8STejun Heo */ 4130493008a8STejun Heo if (flags & WQ_MEM_RECLAIM) { 4131e22bee78STejun Heo struct worker *rescuer; 4132e22bee78STejun Heo 4133d2c1d404STejun Heo rescuer = alloc_worker(); 4134e22bee78STejun Heo if (!rescuer) 4135d2c1d404STejun Heo goto err_destroy; 4136e22bee78STejun Heo 4137111c225aSTejun Heo rescuer->rescue_wq = wq; 4138111c225aSTejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", 4139b196be89STejun Heo wq->name); 4140d2c1d404STejun Heo if (IS_ERR(rescuer->task)) { 4141d2c1d404STejun Heo kfree(rescuer); 4142d2c1d404STejun Heo goto err_destroy; 4143d2c1d404STejun Heo } 4144e22bee78STejun Heo 4145d2c1d404STejun Heo wq->rescuer = rescuer; 414614a40ffcSTejun Heo rescuer->task->flags |= PF_NO_SETAFFINITY; 4147e22bee78STejun Heo wake_up_process(rescuer->task); 41483af24433SOleg Nesterov } 41491537663fSTejun Heo 4150226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4151226223abSTejun Heo goto err_destroy; 4152226223abSTejun Heo 41533af24433SOleg Nesterov /* 415468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 415568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 415668e13a67SLai Jiangshan * list. 41573af24433SOleg Nesterov */ 415868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4159a0a1a5fdSTejun Heo 4160a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 416149e3cf44STejun Heo for_each_pwq(pwq, wq) 4162699ce097STejun Heo pwq_adjust_max_active(pwq); 4163a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4164a0a1a5fdSTejun Heo 41653af24433SOleg Nesterov list_add(&wq->list, &workqueues); 4166a0a1a5fdSTejun Heo 416768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41683af24433SOleg Nesterov 41693af24433SOleg Nesterov return wq; 4170d2c1d404STejun Heo 4171d2c1d404STejun Heo err_free_wq: 41726029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41734690c4abSTejun Heo kfree(wq); 4174d2c1d404STejun Heo return NULL; 4175d2c1d404STejun Heo err_destroy: 4176d2c1d404STejun Heo destroy_workqueue(wq); 41774690c4abSTejun Heo return NULL; 41781da177e4SLinus Torvalds } 4179d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41801da177e4SLinus Torvalds 41813af24433SOleg Nesterov /** 41823af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41833af24433SOleg Nesterov * @wq: target workqueue 41843af24433SOleg Nesterov * 41853af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41863af24433SOleg Nesterov */ 41873af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41883af24433SOleg Nesterov { 418949e3cf44STejun Heo struct pool_workqueue *pwq; 41904c16bd32STejun Heo int node; 41913af24433SOleg Nesterov 41929c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41939c5a2ba7STejun Heo drain_workqueue(wq); 4194c8efcc25STejun Heo 41956183c009STejun Heo /* sanity checks */ 4196b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 419749e3cf44STejun Heo for_each_pwq(pwq, wq) { 41986183c009STejun Heo int i; 41996183c009STejun Heo 420076af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 420176af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4202b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42036183c009STejun Heo return; 420476af4d93STejun Heo } 420576af4d93STejun Heo } 420676af4d93STejun Heo 42075c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 42088864b4e5STejun Heo WARN_ON(pwq->nr_active) || 420976af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4210b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42116183c009STejun Heo return; 42126183c009STejun Heo } 421376af4d93STejun Heo } 4214b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 42156183c009STejun Heo 4216a0a1a5fdSTejun Heo /* 4217a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4218a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4219a0a1a5fdSTejun Heo */ 422068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4221d2c1d404STejun Heo list_del_init(&wq->list); 422268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 42233af24433SOleg Nesterov 4224226223abSTejun Heo workqueue_sysfs_unregister(wq); 4225226223abSTejun Heo 4226493008a8STejun Heo if (wq->rescuer) { 4227e22bee78STejun Heo kthread_stop(wq->rescuer->task); 42288d9df9f0SXiaotian Feng kfree(wq->rescuer); 4229493008a8STejun Heo wq->rescuer = NULL; 4230e22bee78STejun Heo } 4231e22bee78STejun Heo 42328864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 423329c91e99STejun Heo /* 42348864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 42358864b4e5STejun Heo * free the pwqs and wq. 423629c91e99STejun Heo */ 42378864b4e5STejun Heo free_percpu(wq->cpu_pwqs); 42388864b4e5STejun Heo kfree(wq); 42398864b4e5STejun Heo } else { 42408864b4e5STejun Heo /* 42418864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 42424c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 42434c16bd32STejun Heo * @wq will be freed when the last pwq is released. 42448864b4e5STejun Heo */ 42454c16bd32STejun Heo for_each_node(node) { 42464c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42474c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 42484c16bd32STejun Heo put_pwq_unlocked(pwq); 42494c16bd32STejun Heo } 42504c16bd32STejun Heo 42514c16bd32STejun Heo /* 42524c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 42534c16bd32STejun Heo * put. Don't access it afterwards. 42544c16bd32STejun Heo */ 42554c16bd32STejun Heo pwq = wq->dfl_pwq; 42564c16bd32STejun Heo wq->dfl_pwq = NULL; 4257dce90d47STejun Heo put_pwq_unlocked(pwq); 425829c91e99STejun Heo } 42593af24433SOleg Nesterov } 42603af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42613af24433SOleg Nesterov 4262dcd989cbSTejun Heo /** 4263dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4264dcd989cbSTejun Heo * @wq: target workqueue 4265dcd989cbSTejun Heo * @max_active: new max_active value. 4266dcd989cbSTejun Heo * 4267dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4268dcd989cbSTejun Heo * 4269dcd989cbSTejun Heo * CONTEXT: 4270dcd989cbSTejun Heo * Don't call from IRQ context. 4271dcd989cbSTejun Heo */ 4272dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4273dcd989cbSTejun Heo { 427449e3cf44STejun Heo struct pool_workqueue *pwq; 4275dcd989cbSTejun Heo 42768719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42778719dceaSTejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED)) 42788719dceaSTejun Heo return; 42798719dceaSTejun Heo 4280f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4281dcd989cbSTejun Heo 4282a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4283dcd989cbSTejun Heo 4284dcd989cbSTejun Heo wq->saved_max_active = max_active; 4285dcd989cbSTejun Heo 4286699ce097STejun Heo for_each_pwq(pwq, wq) 4287699ce097STejun Heo pwq_adjust_max_active(pwq); 4288dcd989cbSTejun Heo 4289a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4290dcd989cbSTejun Heo } 4291dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4292dcd989cbSTejun Heo 4293dcd989cbSTejun Heo /** 4294e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4295e6267616STejun Heo * 4296e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4297e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4298e6267616STejun Heo */ 4299e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4300e6267616STejun Heo { 4301e6267616STejun Heo struct worker *worker = current_wq_worker(); 4302e6267616STejun Heo 43036a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4304e6267616STejun Heo } 4305e6267616STejun Heo 4306e6267616STejun Heo /** 4307dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4308dcd989cbSTejun Heo * @cpu: CPU in question 4309dcd989cbSTejun Heo * @wq: target workqueue 4310dcd989cbSTejun Heo * 4311dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4312dcd989cbSTejun Heo * no synchronization around this function and the test result is 4313dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4314dcd989cbSTejun Heo * 4315d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4316d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4317d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4318d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4319d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4320d3251859STejun Heo * 4321dcd989cbSTejun Heo * RETURNS: 4322dcd989cbSTejun Heo * %true if congested, %false otherwise. 4323dcd989cbSTejun Heo */ 4324d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4325dcd989cbSTejun Heo { 43267fb98ea7STejun Heo struct pool_workqueue *pwq; 432776af4d93STejun Heo bool ret; 432876af4d93STejun Heo 432988109453SLai Jiangshan rcu_read_lock_sched(); 43307fb98ea7STejun Heo 4331d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4332d3251859STejun Heo cpu = smp_processor_id(); 4333d3251859STejun Heo 43347fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 43357fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 43367fb98ea7STejun Heo else 4337df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4338dcd989cbSTejun Heo 433976af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 434088109453SLai Jiangshan rcu_read_unlock_sched(); 434176af4d93STejun Heo 434276af4d93STejun Heo return ret; 4343dcd989cbSTejun Heo } 4344dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4345dcd989cbSTejun Heo 4346dcd989cbSTejun Heo /** 4347dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4348dcd989cbSTejun Heo * @work: the work to be tested 4349dcd989cbSTejun Heo * 4350dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4351dcd989cbSTejun Heo * synchronization around this function and the test result is 4352dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4353dcd989cbSTejun Heo * 4354dcd989cbSTejun Heo * RETURNS: 4355dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4356dcd989cbSTejun Heo */ 4357dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4358dcd989cbSTejun Heo { 4359fa1b54e6STejun Heo struct worker_pool *pool; 4360dcd989cbSTejun Heo unsigned long flags; 4361dcd989cbSTejun Heo unsigned int ret = 0; 4362dcd989cbSTejun Heo 4363dcd989cbSTejun Heo if (work_pending(work)) 4364dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4365038366c5SLai Jiangshan 4366fa1b54e6STejun Heo local_irq_save(flags); 4367fa1b54e6STejun Heo pool = get_work_pool(work); 4368038366c5SLai Jiangshan if (pool) { 4369fa1b54e6STejun Heo spin_lock(&pool->lock); 4370c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4371dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4372fa1b54e6STejun Heo spin_unlock(&pool->lock); 4373038366c5SLai Jiangshan } 4374fa1b54e6STejun Heo local_irq_restore(flags); 4375dcd989cbSTejun Heo 4376dcd989cbSTejun Heo return ret; 4377dcd989cbSTejun Heo } 4378dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4379dcd989cbSTejun Heo 43803d1cb205STejun Heo /** 43813d1cb205STejun Heo * set_worker_desc - set description for the current work item 43823d1cb205STejun Heo * @fmt: printf-style format string 43833d1cb205STejun Heo * @...: arguments for the format string 43843d1cb205STejun Heo * 43853d1cb205STejun Heo * This function can be called by a running work function to describe what 43863d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43873d1cb205STejun Heo * information will be printed out together to help debugging. The 43883d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43893d1cb205STejun Heo */ 43903d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43913d1cb205STejun Heo { 43923d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43933d1cb205STejun Heo va_list args; 43943d1cb205STejun Heo 43953d1cb205STejun Heo if (worker) { 43963d1cb205STejun Heo va_start(args, fmt); 43973d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43983d1cb205STejun Heo va_end(args); 43993d1cb205STejun Heo worker->desc_valid = true; 44003d1cb205STejun Heo } 44013d1cb205STejun Heo } 44023d1cb205STejun Heo 44033d1cb205STejun Heo /** 44043d1cb205STejun Heo * print_worker_info - print out worker information and description 44053d1cb205STejun Heo * @log_lvl: the log level to use when printing 44063d1cb205STejun Heo * @task: target task 44073d1cb205STejun Heo * 44083d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 44093d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 44103d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 44113d1cb205STejun Heo * 44123d1cb205STejun Heo * This function can be safely called on any task as long as the 44133d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 44143d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 44153d1cb205STejun Heo */ 44163d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 44173d1cb205STejun Heo { 44183d1cb205STejun Heo work_func_t *fn = NULL; 44193d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 44203d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 44213d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 44223d1cb205STejun Heo struct workqueue_struct *wq = NULL; 44233d1cb205STejun Heo bool desc_valid = false; 44243d1cb205STejun Heo struct worker *worker; 44253d1cb205STejun Heo 44263d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 44273d1cb205STejun Heo return; 44283d1cb205STejun Heo 44293d1cb205STejun Heo /* 44303d1cb205STejun Heo * This function is called without any synchronization and @task 44313d1cb205STejun Heo * could be in any state. Be careful with dereferences. 44323d1cb205STejun Heo */ 44333d1cb205STejun Heo worker = probe_kthread_data(task); 44343d1cb205STejun Heo 44353d1cb205STejun Heo /* 44363d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 44373d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 44383d1cb205STejun Heo */ 44393d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44403d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44413d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44423d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44433d1cb205STejun Heo 44443d1cb205STejun Heo /* copy worker description */ 44453d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 44463d1cb205STejun Heo if (desc_valid) 44473d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44483d1cb205STejun Heo 44493d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44503d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44513d1cb205STejun Heo if (desc[0]) 44523d1cb205STejun Heo pr_cont(" (%s)", desc); 44533d1cb205STejun Heo pr_cont("\n"); 44543d1cb205STejun Heo } 44553d1cb205STejun Heo } 44563d1cb205STejun Heo 4457db7bccf4STejun Heo /* 4458db7bccf4STejun Heo * CPU hotplug. 4459db7bccf4STejun Heo * 4460e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4461112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4462706026c2STejun Heo * pool which make migrating pending and scheduled works very 4463e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 446494cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4465e22bee78STejun Heo * blocked draining impractical. 4466e22bee78STejun Heo * 446724647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4468628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4469628c78e7STejun Heo * cpu comes back online. 4470db7bccf4STejun Heo */ 4471db7bccf4STejun Heo 4472706026c2STejun Heo static void wq_unbind_fn(struct work_struct *work) 4473db7bccf4STejun Heo { 447438db41d9STejun Heo int cpu = smp_processor_id(); 44754ce62e9eSTejun Heo struct worker_pool *pool; 4476db7bccf4STejun Heo struct worker *worker; 4477a9ab775bSTejun Heo int wi; 4478db7bccf4STejun Heo 4479f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 44806183c009STejun Heo WARN_ON_ONCE(cpu != smp_processor_id()); 4481db7bccf4STejun Heo 4482bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 448394cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4484e22bee78STejun Heo 4485f2d5a0eeSTejun Heo /* 4486bc3a1afcSTejun Heo * We've blocked all manager operations. Make all workers 448794cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 448894cf58bbSTejun Heo * except for the ones which are still executing works from 448994cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 449094cf58bbSTejun Heo * this, they may become diasporas. 4491f2d5a0eeSTejun Heo */ 4492a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4493403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4494db7bccf4STejun Heo 449524647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4496f2d5a0eeSTejun Heo 449794cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 4498bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 4499e22bee78STejun Heo 4500e22bee78STejun Heo /* 4501eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4502eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4503eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4504eb283428SLai Jiangshan * from other cpus. 4505628c78e7STejun Heo */ 4506628c78e7STejun Heo schedule(); 4507628c78e7STejun Heo 4508628c78e7STejun Heo /* 4509eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4510eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4511eb283428SLai Jiangshan * and keep_working() are always true as long as the 4512eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4513eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4514eb283428SLai Jiangshan * are served by workers tied to the pool. 4515e22bee78STejun Heo */ 4516e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4517eb283428SLai Jiangshan 4518eb283428SLai Jiangshan /* 4519eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4520eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4521eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4522eb283428SLai Jiangshan */ 4523eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4524eb283428SLai Jiangshan wake_up_worker(pool); 4525eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4526eb283428SLai Jiangshan } 4527db7bccf4STejun Heo } 4528db7bccf4STejun Heo 4529bd7c089eSTejun Heo /** 4530bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4531bd7c089eSTejun Heo * @pool: pool of interest 4532bd7c089eSTejun Heo * 4533a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4534bd7c089eSTejun Heo */ 4535bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4536bd7c089eSTejun Heo { 4537a9ab775bSTejun Heo struct worker *worker; 4538a9ab775bSTejun Heo int wi; 4539bd7c089eSTejun Heo 4540bd7c089eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 4541bd7c089eSTejun Heo 4542bd7c089eSTejun Heo /* 4543a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4544a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4545a9ab775bSTejun Heo * wake-ups for concurrency management happen, restore CPU affinty 4546a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4547a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4548bd7c089eSTejun Heo */ 4549a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) 4550a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4551a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4552a9ab775bSTejun Heo 4553a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4554a9ab775bSTejun Heo 4555a9ab775bSTejun Heo for_each_pool_worker(worker, wi, pool) { 4556a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4557a9ab775bSTejun Heo 4558a9ab775bSTejun Heo /* 4559a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4560a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4561a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4562a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4563a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4564a9ab775bSTejun Heo * be bound before @pool->lock is released. 4565a9ab775bSTejun Heo */ 4566a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4567bd7c089eSTejun Heo wake_up_process(worker->task); 4568bd7c089eSTejun Heo 4569bd7c089eSTejun Heo /* 4570a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4571a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4572a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4573a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4574a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4575a9ab775bSTejun Heo * concurrency management. Note that when or whether 4576a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4577a9ab775bSTejun Heo * 4578a9ab775bSTejun Heo * ACCESS_ONCE() is necessary because @worker->flags may be 4579a9ab775bSTejun Heo * tested without holding any lock in 4580a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4581a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4582a9ab775bSTejun Heo * management operations. 4583bd7c089eSTejun Heo */ 4584a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4585a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4586a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4587a9ab775bSTejun Heo ACCESS_ONCE(worker->flags) = worker_flags; 4588bd7c089eSTejun Heo } 4589a9ab775bSTejun Heo 4590a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4591bd7c089eSTejun Heo } 4592bd7c089eSTejun Heo 45937dbc725eSTejun Heo /** 45947dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 45957dbc725eSTejun Heo * @pool: unbound pool of interest 45967dbc725eSTejun Heo * @cpu: the CPU which is coming up 45977dbc725eSTejun Heo * 45987dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 45997dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 46007dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 46017dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 46027dbc725eSTejun Heo */ 46037dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 46047dbc725eSTejun Heo { 46057dbc725eSTejun Heo static cpumask_t cpumask; 46067dbc725eSTejun Heo struct worker *worker; 46077dbc725eSTejun Heo int wi; 46087dbc725eSTejun Heo 46097dbc725eSTejun Heo lockdep_assert_held(&pool->manager_mutex); 46107dbc725eSTejun Heo 46117dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 46127dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 46137dbc725eSTejun Heo return; 46147dbc725eSTejun Heo 46157dbc725eSTejun Heo /* is @cpu the only online CPU? */ 46167dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 46177dbc725eSTejun Heo if (cpumask_weight(&cpumask) != 1) 46187dbc725eSTejun Heo return; 46197dbc725eSTejun Heo 46207dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 46217dbc725eSTejun Heo for_each_pool_worker(worker, wi, pool) 46227dbc725eSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 46237dbc725eSTejun Heo pool->attrs->cpumask) < 0); 46247dbc725eSTejun Heo } 46257dbc725eSTejun Heo 46268db25e78STejun Heo /* 46278db25e78STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 46288db25e78STejun Heo * This will be registered high priority CPU notifier. 46298db25e78STejun Heo */ 46309fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb, 46311da177e4SLinus Torvalds unsigned long action, 46321da177e4SLinus Torvalds void *hcpu) 46331da177e4SLinus Torvalds { 4634d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46354ce62e9eSTejun Heo struct worker_pool *pool; 46364c16bd32STejun Heo struct workqueue_struct *wq; 46377dbc725eSTejun Heo int pi; 46381da177e4SLinus Torvalds 46398db25e78STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 46403af24433SOleg Nesterov case CPU_UP_PREPARE: 4641f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 46423ce63377STejun Heo if (pool->nr_workers) 46433ce63377STejun Heo continue; 4644ebf44d16STejun Heo if (create_and_start_worker(pool) < 0) 46453ce63377STejun Heo return NOTIFY_BAD; 46463af24433SOleg Nesterov } 46471da177e4SLinus Torvalds break; 46481da177e4SLinus Torvalds 464965758202STejun Heo case CPU_DOWN_FAILED: 465065758202STejun Heo case CPU_ONLINE: 465168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 46527dbc725eSTejun Heo 46537dbc725eSTejun Heo for_each_pool(pool, pi) { 4654bc3a1afcSTejun Heo mutex_lock(&pool->manager_mutex); 465594cf58bbSTejun Heo 46567dbc725eSTejun Heo if (pool->cpu == cpu) { 4657a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 465824647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4659a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4660a9ab775bSTejun Heo 466194cf58bbSTejun Heo rebind_workers(pool); 46627dbc725eSTejun Heo } else if (pool->cpu < 0) { 46637dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 46647dbc725eSTejun Heo } 466594cf58bbSTejun Heo 4666bc3a1afcSTejun Heo mutex_unlock(&pool->manager_mutex); 466794cf58bbSTejun Heo } 46687dbc725eSTejun Heo 46694c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46704c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 46714c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 46724c16bd32STejun Heo 467368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46748db25e78STejun Heo break; 467565758202STejun Heo } 467665758202STejun Heo return NOTIFY_OK; 467765758202STejun Heo } 467865758202STejun Heo 467965758202STejun Heo /* 468065758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 468165758202STejun Heo * This will be registered as low priority CPU notifier. 468265758202STejun Heo */ 46839fdf9b73SLai Jiangshan static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb, 468465758202STejun Heo unsigned long action, 468565758202STejun Heo void *hcpu) 468665758202STejun Heo { 4687d84ff051STejun Heo int cpu = (unsigned long)hcpu; 46888db25e78STejun Heo struct work_struct unbind_work; 46894c16bd32STejun Heo struct workqueue_struct *wq; 46908db25e78STejun Heo 469165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 469265758202STejun Heo case CPU_DOWN_PREPARE: 46934c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4694706026c2STejun Heo INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn); 46957635d2fdSJoonsoo Kim queue_work_on(cpu, system_highpri_wq, &unbind_work); 46964c16bd32STejun Heo 46974c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 46984c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 46994c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47004c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 47014c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 47024c16bd32STejun Heo 47034c16bd32STejun Heo /* wait for per-cpu unbinding to finish */ 47048db25e78STejun Heo flush_work(&unbind_work); 47058db25e78STejun Heo break; 470665758202STejun Heo } 470765758202STejun Heo return NOTIFY_OK; 470865758202STejun Heo } 470965758202STejun Heo 47102d3854a3SRusty Russell #ifdef CONFIG_SMP 47118ccad40dSRusty Russell 47122d3854a3SRusty Russell struct work_for_cpu { 4713ed48ece2STejun Heo struct work_struct work; 47142d3854a3SRusty Russell long (*fn)(void *); 47152d3854a3SRusty Russell void *arg; 47162d3854a3SRusty Russell long ret; 47172d3854a3SRusty Russell }; 47182d3854a3SRusty Russell 4719ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 47202d3854a3SRusty Russell { 4721ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4722ed48ece2STejun Heo 47232d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 47242d3854a3SRusty Russell } 47252d3854a3SRusty Russell 47262d3854a3SRusty Russell /** 47272d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 47282d3854a3SRusty Russell * @cpu: the cpu to run on 47292d3854a3SRusty Russell * @fn: the function to run 47302d3854a3SRusty Russell * @arg: the function arg 47312d3854a3SRusty Russell * 473231ad9081SRusty Russell * This will return the value @fn returns. 473331ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 47346b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 47352d3854a3SRusty Russell */ 4736d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 47372d3854a3SRusty Russell { 4738ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 47392d3854a3SRusty Russell 4740ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4741ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 4742ed48ece2STejun Heo flush_work(&wfc.work); 47432d3854a3SRusty Russell return wfc.ret; 47442d3854a3SRusty Russell } 47452d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 47462d3854a3SRusty Russell #endif /* CONFIG_SMP */ 47472d3854a3SRusty Russell 4748a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4749e7577c50SRusty Russell 4750a0a1a5fdSTejun Heo /** 4751a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4752a0a1a5fdSTejun Heo * 475358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4754c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4755706026c2STejun Heo * pool->worklist. 4756a0a1a5fdSTejun Heo * 4757a0a1a5fdSTejun Heo * CONTEXT: 4758a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4759a0a1a5fdSTejun Heo */ 4760a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4761a0a1a5fdSTejun Heo { 476217116969STejun Heo struct worker_pool *pool; 476324b8a847STejun Heo struct workqueue_struct *wq; 476424b8a847STejun Heo struct pool_workqueue *pwq; 4765611c92a0STejun Heo int pi; 4766a0a1a5fdSTejun Heo 476768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4768a0a1a5fdSTejun Heo 47696183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4770a0a1a5fdSTejun Heo workqueue_freezing = true; 4771a0a1a5fdSTejun Heo 477224b8a847STejun Heo /* set FREEZING */ 4773611c92a0STejun Heo for_each_pool(pool, pi) { 47745bcab335STejun Heo spin_lock_irq(&pool->lock); 477535b6bb63STejun Heo WARN_ON_ONCE(pool->flags & POOL_FREEZING); 477635b6bb63STejun Heo pool->flags |= POOL_FREEZING; 47775bcab335STejun Heo spin_unlock_irq(&pool->lock); 47781da177e4SLinus Torvalds } 47798b03ae3cSTejun Heo 478024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4781a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4782699ce097STejun Heo for_each_pwq(pwq, wq) 4783699ce097STejun Heo pwq_adjust_max_active(pwq); 4784a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4785a1056305STejun Heo } 47865bcab335STejun Heo 478768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4788a0a1a5fdSTejun Heo } 4789a0a1a5fdSTejun Heo 4790a0a1a5fdSTejun Heo /** 479158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4792a0a1a5fdSTejun Heo * 4793a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4794a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4795a0a1a5fdSTejun Heo * 4796a0a1a5fdSTejun Heo * CONTEXT: 479768e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4798a0a1a5fdSTejun Heo * 4799a0a1a5fdSTejun Heo * RETURNS: 480058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 480158a69cb4STejun Heo * is complete. 4802a0a1a5fdSTejun Heo */ 4803a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4804a0a1a5fdSTejun Heo { 4805a0a1a5fdSTejun Heo bool busy = false; 480624b8a847STejun Heo struct workqueue_struct *wq; 480724b8a847STejun Heo struct pool_workqueue *pwq; 4808a0a1a5fdSTejun Heo 480968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4810a0a1a5fdSTejun Heo 48116183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4812a0a1a5fdSTejun Heo 481324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 481424b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 481524b8a847STejun Heo continue; 4816a0a1a5fdSTejun Heo /* 4817a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4818a0a1a5fdSTejun Heo * to peek without lock. 4819a0a1a5fdSTejun Heo */ 482088109453SLai Jiangshan rcu_read_lock_sched(); 482124b8a847STejun Heo for_each_pwq(pwq, wq) { 48226183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4823112202d9STejun Heo if (pwq->nr_active) { 4824a0a1a5fdSTejun Heo busy = true; 482588109453SLai Jiangshan rcu_read_unlock_sched(); 4826a0a1a5fdSTejun Heo goto out_unlock; 4827a0a1a5fdSTejun Heo } 4828a0a1a5fdSTejun Heo } 482988109453SLai Jiangshan rcu_read_unlock_sched(); 4830a0a1a5fdSTejun Heo } 4831a0a1a5fdSTejun Heo out_unlock: 483268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4833a0a1a5fdSTejun Heo return busy; 4834a0a1a5fdSTejun Heo } 4835a0a1a5fdSTejun Heo 4836a0a1a5fdSTejun Heo /** 4837a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4838a0a1a5fdSTejun Heo * 4839a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4840706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4841a0a1a5fdSTejun Heo * 4842a0a1a5fdSTejun Heo * CONTEXT: 4843a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4844a0a1a5fdSTejun Heo */ 4845a0a1a5fdSTejun Heo void thaw_workqueues(void) 4846a0a1a5fdSTejun Heo { 484724b8a847STejun Heo struct workqueue_struct *wq; 484824b8a847STejun Heo struct pool_workqueue *pwq; 484924b8a847STejun Heo struct worker_pool *pool; 4850611c92a0STejun Heo int pi; 4851a0a1a5fdSTejun Heo 485268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4853a0a1a5fdSTejun Heo 4854a0a1a5fdSTejun Heo if (!workqueue_freezing) 4855a0a1a5fdSTejun Heo goto out_unlock; 4856a0a1a5fdSTejun Heo 485724b8a847STejun Heo /* clear FREEZING */ 4858611c92a0STejun Heo for_each_pool(pool, pi) { 48595bcab335STejun Heo spin_lock_irq(&pool->lock); 486035b6bb63STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_FREEZING)); 486135b6bb63STejun Heo pool->flags &= ~POOL_FREEZING; 48625bcab335STejun Heo spin_unlock_irq(&pool->lock); 4863d565ed63STejun Heo } 486424b8a847STejun Heo 486524b8a847STejun Heo /* restore max_active and repopulate worklist */ 486624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4867a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4868699ce097STejun Heo for_each_pwq(pwq, wq) 4869699ce097STejun Heo pwq_adjust_max_active(pwq); 4870a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 487124b8a847STejun Heo } 487224b8a847STejun Heo 4873a0a1a5fdSTejun Heo workqueue_freezing = false; 4874a0a1a5fdSTejun Heo out_unlock: 487568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4876a0a1a5fdSTejun Heo } 4877a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4878a0a1a5fdSTejun Heo 4879bce90380STejun Heo static void __init wq_numa_init(void) 4880bce90380STejun Heo { 4881bce90380STejun Heo cpumask_var_t *tbl; 4882bce90380STejun Heo int node, cpu; 4883bce90380STejun Heo 4884bce90380STejun Heo /* determine NUMA pwq table len - highest node id + 1 */ 4885bce90380STejun Heo for_each_node(node) 4886bce90380STejun Heo wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1); 4887bce90380STejun Heo 4888bce90380STejun Heo if (num_possible_nodes() <= 1) 4889bce90380STejun Heo return; 4890bce90380STejun Heo 4891d55262c4STejun Heo if (wq_disable_numa) { 4892d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 4893d55262c4STejun Heo return; 4894d55262c4STejun Heo } 4895d55262c4STejun Heo 48964c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 48974c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 48984c16bd32STejun Heo 4899bce90380STejun Heo /* 4900bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 4901bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 4902bce90380STejun Heo * fully initialized by now. 4903bce90380STejun Heo */ 4904bce90380STejun Heo tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL); 4905bce90380STejun Heo BUG_ON(!tbl); 4906bce90380STejun Heo 4907bce90380STejun Heo for_each_node(node) 4908bce90380STejun Heo BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL, node)); 4909bce90380STejun Heo 4910bce90380STejun Heo for_each_possible_cpu(cpu) { 4911bce90380STejun Heo node = cpu_to_node(cpu); 4912bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 4913bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 4914bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 4915bce90380STejun Heo return; 4916bce90380STejun Heo } 4917bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 4918bce90380STejun Heo } 4919bce90380STejun Heo 4920bce90380STejun Heo wq_numa_possible_cpumask = tbl; 4921bce90380STejun Heo wq_numa_enabled = true; 4922bce90380STejun Heo } 4923bce90380STejun Heo 49246ee0578bSSuresh Siddha static int __init init_workqueues(void) 49251da177e4SLinus Torvalds { 49267a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 49277a4e344cSTejun Heo int i, cpu; 4928c34056a3STejun Heo 49297c3eed5cSTejun Heo /* make sure we have enough bits for OFFQ pool ID */ 49307c3eed5cSTejun Heo BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) < 49316be19588SLai Jiangshan WORK_CPU_END * NR_STD_WORKER_POOLS); 4932b5490077STejun Heo 4933e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 4934e904e6c2STejun Heo 4935e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 4936e904e6c2STejun Heo 493765758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 4938a5b4e57dSLai Jiangshan hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 49398b03ae3cSTejun Heo 4940bce90380STejun Heo wq_numa_init(); 4941bce90380STejun Heo 4942706026c2STejun Heo /* initialize CPU pools */ 494329c91e99STejun Heo for_each_possible_cpu(cpu) { 49444ce62e9eSTejun Heo struct worker_pool *pool; 49458b03ae3cSTejun Heo 49467a4e344cSTejun Heo i = 0; 4947f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49487a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 4949ec22ca5eSTejun Heo pool->cpu = cpu; 49507a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 49517a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 4952f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 49537a4e344cSTejun Heo 49549daf9e67STejun Heo /* alloc pool ID */ 495568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 49569daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 495768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 49584ce62e9eSTejun Heo } 49598b03ae3cSTejun Heo } 49608b03ae3cSTejun Heo 4961e22bee78STejun Heo /* create the initial worker */ 496229c91e99STejun Heo for_each_online_cpu(cpu) { 49634ce62e9eSTejun Heo struct worker_pool *pool; 4964e22bee78STejun Heo 4965f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 496624647570STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 4967ebf44d16STejun Heo BUG_ON(create_and_start_worker(pool) < 0); 4968e22bee78STejun Heo } 49694ce62e9eSTejun Heo } 4970e22bee78STejun Heo 497129c91e99STejun Heo /* create default unbound wq attrs */ 497229c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 497329c91e99STejun Heo struct workqueue_attrs *attrs; 497429c91e99STejun Heo 497529c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 497629c91e99STejun Heo attrs->nice = std_nice[i]; 497729c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 497829c91e99STejun Heo } 497929c91e99STejun Heo 4980d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 49811aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 4982d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 4983f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 4984f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 498524d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 498624d51addSTejun Heo WQ_FREEZABLE, 0); 49871aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 4988ae930e0fSTejun Heo !system_unbound_wq || !system_freezable_wq); 49896ee0578bSSuresh Siddha return 0; 49901da177e4SLinus Torvalds } 49916ee0578bSSuresh Siddha early_initcall(init_workqueues); 4992