1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 3c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 41da177e4SLinus Torvalds * 5c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 81da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 9e1f8e874SFrancois Cami * Andrew Morton 101da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 111da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1289ada679SChristoph Lameter * 13cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 14c54fce6eSTejun Heo * 15c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 16c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 17c54fce6eSTejun Heo * 18c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 19c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 20b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 21b11895c4SLibin * normal work items and the other for high priority ones) and some extra 22b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 23b11895c4SLibin * number of these backing pools is dynamic. 24c54fce6eSTejun Heo * 259a261491SBenjamin Peterson * Please read Documentation/core-api/workqueue.rst for details. 261da177e4SLinus Torvalds */ 271da177e4SLinus Torvalds 289984de1aSPaul Gortmaker #include <linux/export.h> 291da177e4SLinus Torvalds #include <linux/kernel.h> 301da177e4SLinus Torvalds #include <linux/sched.h> 311da177e4SLinus Torvalds #include <linux/init.h> 321da177e4SLinus Torvalds #include <linux/signal.h> 331da177e4SLinus Torvalds #include <linux/completion.h> 341da177e4SLinus Torvalds #include <linux/workqueue.h> 351da177e4SLinus Torvalds #include <linux/slab.h> 361da177e4SLinus Torvalds #include <linux/cpu.h> 371da177e4SLinus Torvalds #include <linux/notifier.h> 381da177e4SLinus Torvalds #include <linux/kthread.h> 391fa44ecaSJames Bottomley #include <linux/hardirq.h> 4046934023SChristoph Lameter #include <linux/mempolicy.h> 41341a5958SRafael J. Wysocki #include <linux/freezer.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51c98a9805STal Shorer #include <linux/sched/isolation.h> 5262635ea8SSergey Senozhatsky #include <linux/nmi.h> 53940d71c6SSergey Senozhatsky #include <linux/kvm_para.h> 54e22bee78STejun Heo 55ea138446STejun Heo #include "workqueue_internal.h" 561da177e4SLinus Torvalds 57c8e55f36STejun Heo enum { 58bc2ae0f5STejun Heo /* 5924647570STejun Heo * worker_pool flags 60bc2ae0f5STejun Heo * 6124647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 62bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 63bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 64bc2ae0f5STejun Heo * is in effect. 65bc2ae0f5STejun Heo * 66bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 67bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6824647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 69bc2ae0f5STejun Heo * 70bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 711258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 724736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 73bc2ae0f5STejun Heo */ 74692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7524647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 76db7bccf4STejun Heo 77c8e55f36STejun Heo /* worker flags */ 78c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 79c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 80e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 81fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 82f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 83a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 84e22bee78STejun Heo 85a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 86a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 87db7bccf4STejun Heo 88e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 894ce62e9eSTejun Heo 9029c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 91c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 92db7bccf4STejun Heo 93e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 94e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 95e22bee78STejun Heo 963233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 973233cdbdSTejun Heo /* call for help after 10ms 983233cdbdSTejun Heo (min two ticks) */ 99e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 100e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* 103e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1048698a745SDongsheng Yang * all cpus. Give MIN_NICE. 105e22bee78STejun Heo */ 1068698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1078698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 108ecf6881fSTejun Heo 109ecf6881fSTejun Heo WQ_NAME_LEN = 24, 110c8e55f36STejun Heo }; 111c8e55f36STejun Heo 1121da177e4SLinus Torvalds /* 1134690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1144690c4abSTejun Heo * 115e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 116e41e704bSTejun Heo * everyone else. 1174690c4abSTejun Heo * 118e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 119e22bee78STejun Heo * only be modified and accessed from the local cpu. 120e22bee78STejun Heo * 121d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1224690c4abSTejun Heo * 123d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 124d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 125d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 126d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 127e22bee78STejun Heo * 1281258fae7STejun Heo * A: wq_pool_attach_mutex protected. 129822d8405STejun Heo * 13068e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13176af4d93STejun Heo * 13224acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1335bcab335STejun Heo * 1345b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1355b95e1afSLai Jiangshan * 1365b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 13724acfb71SThomas Gleixner * RCU for reads. 1385b95e1afSLai Jiangshan * 1393c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1403c25a55dSLai Jiangshan * 14124acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1422e109a28STejun Heo * 1432e109a28STejun Heo * MD: wq_mayday_lock protected. 1444690c4abSTejun Heo */ 1454690c4abSTejun Heo 1462eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 147c34056a3STejun Heo 148bd7bdd43STejun Heo struct worker_pool { 149a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 150d84ff051STejun Heo int cpu; /* I: the associated cpu */ 151f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1529daf9e67STejun Heo int id; /* I: pool ID */ 15311ebea50STejun Heo unsigned int flags; /* X: flags */ 154bd7bdd43STejun Heo 15582607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15682607adcSTejun Heo 157bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 158ea1abd61SLai Jiangshan 1595826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1605826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 161bd7bdd43STejun Heo 162bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 163bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 164bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 165bd7bdd43STejun Heo 166c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 167c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 168c9e7cf27STejun Heo /* L: hash of busy workers */ 169c9e7cf27STejun Heo 1702607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17192f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17260f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 173e19e397aSTejun Heo 1747cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 175e19e397aSTejun Heo 1767a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17768e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17868e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1797a4e344cSTejun Heo 180e19e397aSTejun Heo /* 181e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 182e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 183e19e397aSTejun Heo * cacheline. 184e19e397aSTejun Heo */ 185e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18629c91e99STejun Heo 18729c91e99STejun Heo /* 18824acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 18929c91e99STejun Heo * from get_work_pool(). 19029c91e99STejun Heo */ 19129c91e99STejun Heo struct rcu_head rcu; 1928b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1938b03ae3cSTejun Heo 1948b03ae3cSTejun Heo /* 195112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 196112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 197112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 198112202d9STejun Heo * number of flag bits. 1991da177e4SLinus Torvalds */ 200112202d9STejun Heo struct pool_workqueue { 201bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2024690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20373f53c4aSTejun Heo int work_color; /* L: current color */ 20473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2058864b4e5STejun Heo int refcnt; /* L: reference count */ 20673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20773f53c4aSTejun Heo /* L: nr of in_flight works */ 2081e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 209a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 210f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2113c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2122e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2138864b4e5STejun Heo 2148864b4e5STejun Heo /* 2158864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2168864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 21724acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 218b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2198864b4e5STejun Heo */ 2208864b4e5STejun Heo struct work_struct unbound_release_work; 2218864b4e5STejun Heo struct rcu_head rcu; 222e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2231da177e4SLinus Torvalds 2241da177e4SLinus Torvalds /* 22573f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22673f53c4aSTejun Heo */ 22773f53c4aSTejun Heo struct wq_flusher { 2283c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2293c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 23073f53c4aSTejun Heo struct completion done; /* flush completion */ 23173f53c4aSTejun Heo }; 2321da177e4SLinus Torvalds 233226223abSTejun Heo struct wq_device; 234226223abSTejun Heo 23573f53c4aSTejun Heo /* 236c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 237c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2381da177e4SLinus Torvalds */ 2391da177e4SLinus Torvalds struct workqueue_struct { 2403c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 241e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24273f53c4aSTejun Heo 2433c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2443c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2453c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 246112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2473c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2483c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2493c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 25073f53c4aSTejun Heo 2512e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 25230ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 253e22bee78STejun Heo 25487fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 255a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 256226223abSTejun Heo 2575b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2585b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2596029a918STejun Heo 260226223abSTejun Heo #ifdef CONFIG_SYSFS 261226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 262226223abSTejun Heo #endif 2634e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 264669de8bdSBart Van Assche char *lock_name; 265669de8bdSBart Van Assche struct lock_class_key key; 2664e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2674e6045f1SJohannes Berg #endif 268ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2692728fd2fSTejun Heo 270e2dca7adSTejun Heo /* 27124acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 27224acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 273e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 274e2dca7adSTejun Heo */ 275e2dca7adSTejun Heo struct rcu_head rcu; 276e2dca7adSTejun Heo 2772728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2782728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2792728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2805b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2811da177e4SLinus Torvalds }; 2821da177e4SLinus Torvalds 283e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 284e904e6c2STejun Heo 285bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 286bce90380STejun Heo /* possible CPUs of each node */ 287bce90380STejun Heo 288d55262c4STejun Heo static bool wq_disable_numa; 289d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 290d55262c4STejun Heo 291cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 292552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 293cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 294cee22a15SViresh Kumar 295863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 2963347fa09STejun Heo 297bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 298bce90380STejun Heo 2994c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3004c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3014c16bd32STejun Heo 30268e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3031258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 304a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 305d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 306d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3075bcab335STejun Heo 308e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30968e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3107d19c5ceSTejun Heo 311ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 312ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 313ef557180SMike Galbraith 314ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 315ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 316b05a7928SFrederic Weisbecker 317f303fccbSTejun Heo /* 318f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 319f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 320f303fccbSTejun Heo * to uncover usages which depend on it. 321f303fccbSTejun Heo */ 322f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 323f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 324f303fccbSTejun Heo #else 325f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 326f303fccbSTejun Heo #endif 327f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 328f303fccbSTejun Heo 3297d19c5ceSTejun Heo /* the per-cpu worker pools */ 33025528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3317d19c5ceSTejun Heo 33268e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3337d19c5ceSTejun Heo 33468e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33529c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33629c91e99STejun Heo 337c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33829c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33929c91e99STejun Heo 3408a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3418a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3428a2b7538STejun Heo 343d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 344ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 345044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3461aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 347044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 348d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 349044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 350f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 351044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 35224d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3530668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3540668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3550668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3560668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 357d320c038STejun Heo 3587d19c5ceSTejun Heo static int worker_thread(void *__worker); 3596ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 360c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 3617d19c5ceSTejun Heo 36297bd2347STejun Heo #define CREATE_TRACE_POINTS 36397bd2347STejun Heo #include <trace/events/workqueue.h> 36497bd2347STejun Heo 36568e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 36624acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 367f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36824acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 3695bcab335STejun Heo 3705b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 37124acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 372f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 373f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 37424acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 3755b95e1afSLai Jiangshan 376f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 377f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 378f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3797a62c2c8STejun Heo (pool)++) 3804ce62e9eSTejun Heo 38149e3cf44STejun Heo /** 38217116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38317116969STejun Heo * @pool: iteration cursor 384611c92a0STejun Heo * @pi: integer used for iteration 385fa1b54e6STejun Heo * 38624acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 38768e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38868e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 389fa1b54e6STejun Heo * 390fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 391fa1b54e6STejun Heo * ignored. 39217116969STejun Heo */ 393611c92a0STejun Heo #define for_each_pool(pool, pi) \ 394611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39568e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 396fa1b54e6STejun Heo else 39717116969STejun Heo 39817116969STejun Heo /** 399822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 400822d8405STejun Heo * @worker: iteration cursor 401822d8405STejun Heo * @pool: worker_pool to iterate workers of 402822d8405STejun Heo * 4031258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 404822d8405STejun Heo * 405822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 406822d8405STejun Heo * ignored. 407822d8405STejun Heo */ 408da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 409da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4101258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 411822d8405STejun Heo else 412822d8405STejun Heo 413822d8405STejun Heo /** 41449e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41549e3cf44STejun Heo * @pwq: iteration cursor 41649e3cf44STejun Heo * @wq: the target workqueue 41776af4d93STejun Heo * 41824acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 419794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 420794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42176af4d93STejun Heo * 42276af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42376af4d93STejun Heo * ignored. 42449e3cf44STejun Heo */ 42549e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42649e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4275a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 428f3421797STejun Heo 429dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 430dc186ad7SThomas Gleixner 431f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 432dc186ad7SThomas Gleixner 43399777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43499777288SStanislaw Gruszka { 43599777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43699777288SStanislaw Gruszka } 43799777288SStanislaw Gruszka 438b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 439b9fdac7fSDu, Changbin { 440b9fdac7fSDu, Changbin struct work_struct *work = addr; 441b9fdac7fSDu, Changbin 442b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 443b9fdac7fSDu, Changbin } 444b9fdac7fSDu, Changbin 445dc186ad7SThomas Gleixner /* 446dc186ad7SThomas Gleixner * fixup_init is called when: 447dc186ad7SThomas Gleixner * - an active object is initialized 448dc186ad7SThomas Gleixner */ 44902a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 450dc186ad7SThomas Gleixner { 451dc186ad7SThomas Gleixner struct work_struct *work = addr; 452dc186ad7SThomas Gleixner 453dc186ad7SThomas Gleixner switch (state) { 454dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 455dc186ad7SThomas Gleixner cancel_work_sync(work); 456dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45702a982a6SDu, Changbin return true; 458dc186ad7SThomas Gleixner default: 45902a982a6SDu, Changbin return false; 460dc186ad7SThomas Gleixner } 461dc186ad7SThomas Gleixner } 462dc186ad7SThomas Gleixner 463dc186ad7SThomas Gleixner /* 464dc186ad7SThomas Gleixner * fixup_free is called when: 465dc186ad7SThomas Gleixner * - an active object is freed 466dc186ad7SThomas Gleixner */ 46702a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner struct work_struct *work = addr; 470dc186ad7SThomas Gleixner 471dc186ad7SThomas Gleixner switch (state) { 472dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 473dc186ad7SThomas Gleixner cancel_work_sync(work); 474dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47502a982a6SDu, Changbin return true; 476dc186ad7SThomas Gleixner default: 47702a982a6SDu, Changbin return false; 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner 481f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 482dc186ad7SThomas Gleixner .name = "work_struct", 48399777288SStanislaw Gruszka .debug_hint = work_debug_hint, 484b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 485dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 486dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 487dc186ad7SThomas Gleixner }; 488dc186ad7SThomas Gleixner 489dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 490dc186ad7SThomas Gleixner { 491dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 492dc186ad7SThomas Gleixner } 493dc186ad7SThomas Gleixner 494dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 495dc186ad7SThomas Gleixner { 496dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner 499dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 500dc186ad7SThomas Gleixner { 501dc186ad7SThomas Gleixner if (onstack) 502dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 503dc186ad7SThomas Gleixner else 504dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 505dc186ad7SThomas Gleixner } 506dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 507dc186ad7SThomas Gleixner 508dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 511dc186ad7SThomas Gleixner } 512dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 513dc186ad7SThomas Gleixner 514ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 515ea2e64f2SThomas Gleixner { 516ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 517ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 518ea2e64f2SThomas Gleixner } 519ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 520ea2e64f2SThomas Gleixner 521dc186ad7SThomas Gleixner #else 522dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 523dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 524dc186ad7SThomas Gleixner #endif 525dc186ad7SThomas Gleixner 5264e8b22bdSLi Bin /** 52767dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5284e8b22bdSLi Bin * @pool: the pool pointer of interest 5294e8b22bdSLi Bin * 5304e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5314e8b22bdSLi Bin * successfully, -errno on failure. 5324e8b22bdSLi Bin */ 5339daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5349daf9e67STejun Heo { 5359daf9e67STejun Heo int ret; 5369daf9e67STejun Heo 53768e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5385bcab335STejun Heo 5394e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5404e8b22bdSLi Bin GFP_KERNEL); 541229641a6STejun Heo if (ret >= 0) { 542e68035fbSTejun Heo pool->id = ret; 543229641a6STejun Heo return 0; 544229641a6STejun Heo } 5459daf9e67STejun Heo return ret; 5469daf9e67STejun Heo } 5479daf9e67STejun Heo 54876af4d93STejun Heo /** 549df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 550df2d5ae4STejun Heo * @wq: the target workqueue 551df2d5ae4STejun Heo * @node: the node ID 552df2d5ae4STejun Heo * 55324acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5545b95e1afSLai Jiangshan * read locked. 555df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 556df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 557d185af30SYacine Belkadi * 558d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 559df2d5ae4STejun Heo */ 560df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 561df2d5ae4STejun Heo int node) 562df2d5ae4STejun Heo { 5635b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 564d6e022f1STejun Heo 565d6e022f1STejun Heo /* 566d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 567d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 568d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 569d6e022f1STejun Heo * happens, this workaround can be removed. 570d6e022f1STejun Heo */ 571d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 572d6e022f1STejun Heo return wq->dfl_pwq; 573d6e022f1STejun Heo 574df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 575df2d5ae4STejun Heo } 576df2d5ae4STejun Heo 57773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 57873f53c4aSTejun Heo { 57973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 58073f53c4aSTejun Heo } 58173f53c4aSTejun Heo 582*c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 58373f53c4aSTejun Heo { 584*c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 58573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58673f53c4aSTejun Heo } 58773f53c4aSTejun Heo 58873f53c4aSTejun Heo static int work_next_color(int color) 58973f53c4aSTejun Heo { 59073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 591a848e3b6SOleg Nesterov } 592a848e3b6SOleg Nesterov 5934594bf15SDavid Howells /* 594112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 595112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5967c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5977a22ad75STejun Heo * 598112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 599112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 600bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 601bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6027a22ad75STejun Heo * 603112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6047c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 605112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6067c3eed5cSTejun Heo * available only while the work item is queued. 607bbb68dfaSTejun Heo * 608bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 609bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 610bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 611bbb68dfaSTejun Heo * try to steal the PENDING bit. 6124594bf15SDavid Howells */ 6137a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6147a22ad75STejun Heo unsigned long flags) 6157a22ad75STejun Heo { 6166183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6177a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6187a22ad75STejun Heo } 6197a22ad75STejun Heo 620112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6214690c4abSTejun Heo unsigned long extra_flags) 622365970a1SDavid Howells { 623112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 624112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 625365970a1SDavid Howells } 626365970a1SDavid Howells 6274468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6284468a00fSLai Jiangshan int pool_id) 6294468a00fSLai Jiangshan { 6304468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6314468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6324468a00fSLai Jiangshan } 6334468a00fSLai Jiangshan 6347c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6357c3eed5cSTejun Heo int pool_id) 6364d707b9fSOleg Nesterov { 63723657bb1STejun Heo /* 63823657bb1STejun Heo * The following wmb is paired with the implied mb in 63923657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 64023657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64123657bb1STejun Heo * owner. 64223657bb1STejun Heo */ 64323657bb1STejun Heo smp_wmb(); 6447c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 645346c09f8SRoman Pen /* 646346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 647346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 648346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6498bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 650346c09f8SRoman Pen * the same @work. E.g. consider this case: 651346c09f8SRoman Pen * 652346c09f8SRoman Pen * CPU#0 CPU#1 653346c09f8SRoman Pen * ---------------------------- -------------------------------- 654346c09f8SRoman Pen * 655346c09f8SRoman Pen * 1 STORE event_indicated 656346c09f8SRoman Pen * 2 queue_work_on() { 657346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 658346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 659346c09f8SRoman Pen * 5 set_work_data() # clear bit 660346c09f8SRoman Pen * 6 smp_mb() 661346c09f8SRoman Pen * 7 work->current_func() { 662346c09f8SRoman Pen * 8 LOAD event_indicated 663346c09f8SRoman Pen * } 664346c09f8SRoman Pen * 665346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 666346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 667346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 668346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 669346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 670346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 671346c09f8SRoman Pen * before actual STORE. 672346c09f8SRoman Pen */ 673346c09f8SRoman Pen smp_mb(); 6744d707b9fSOleg Nesterov } 6754d707b9fSOleg Nesterov 6767a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 677365970a1SDavid Howells { 6787c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6797c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6807a22ad75STejun Heo } 6817a22ad75STejun Heo 682112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6837a22ad75STejun Heo { 684e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6857a22ad75STejun Heo 686112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 687e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 688e120153dSTejun Heo else 689e120153dSTejun Heo return NULL; 6907a22ad75STejun Heo } 6917a22ad75STejun Heo 6927c3eed5cSTejun Heo /** 6937c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6947c3eed5cSTejun Heo * @work: the work item of interest 6957c3eed5cSTejun Heo * 69668e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69724acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 69824acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 699fa1b54e6STejun Heo * 700fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 701fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 702fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 703fa1b54e6STejun Heo * returned pool is and stays online. 704d185af30SYacine Belkadi * 705d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7067c3eed5cSTejun Heo */ 7077c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7087a22ad75STejun Heo { 709e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7107c3eed5cSTejun Heo int pool_id; 7117a22ad75STejun Heo 71268e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 713fa1b54e6STejun Heo 714112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 715112202d9STejun Heo return ((struct pool_workqueue *) 7167c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7177a22ad75STejun Heo 7187c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7197c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7207a22ad75STejun Heo return NULL; 7217a22ad75STejun Heo 722fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7237c3eed5cSTejun Heo } 7247c3eed5cSTejun Heo 7257c3eed5cSTejun Heo /** 7267c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7277c3eed5cSTejun Heo * @work: the work item of interest 7287c3eed5cSTejun Heo * 729d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7307c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7317c3eed5cSTejun Heo */ 7327c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7337c3eed5cSTejun Heo { 73454d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7357c3eed5cSTejun Heo 736112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 737112202d9STejun Heo return ((struct pool_workqueue *) 73854d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 73954d5b7d0SLai Jiangshan 74054d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7417c3eed5cSTejun Heo } 7427c3eed5cSTejun Heo 743bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 744bbb68dfaSTejun Heo { 7457c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 746bbb68dfaSTejun Heo 7477c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7487c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 749bbb68dfaSTejun Heo } 750bbb68dfaSTejun Heo 751bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 752bbb68dfaSTejun Heo { 753bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 754bbb68dfaSTejun Heo 755112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 756bbb68dfaSTejun Heo } 757bbb68dfaSTejun Heo 758e22bee78STejun Heo /* 7593270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7603270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 761d565ed63STejun Heo * they're being called with pool->lock held. 762e22bee78STejun Heo */ 763e22bee78STejun Heo 76463d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 765649027d7STejun Heo { 766e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 767649027d7STejun Heo } 768649027d7STejun Heo 769e22bee78STejun Heo /* 770e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 771e22bee78STejun Heo * running workers. 772974271c4STejun Heo * 773974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 774706026c2STejun Heo * function will always return %true for unbound pools as long as the 775974271c4STejun Heo * worklist isn't empty. 776e22bee78STejun Heo */ 77763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 778e22bee78STejun Heo { 77963d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 780e22bee78STejun Heo } 781e22bee78STejun Heo 782e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 784e22bee78STejun Heo { 78563d95a91STejun Heo return pool->nr_idle; 786e22bee78STejun Heo } 787e22bee78STejun Heo 788e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 78963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 790e22bee78STejun Heo { 791e19e397aSTejun Heo return !list_empty(&pool->worklist) && 792e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 793e22bee78STejun Heo } 794e22bee78STejun Heo 795e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 797e22bee78STejun Heo { 79863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 799e22bee78STejun Heo } 800e22bee78STejun Heo 801e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 803e22bee78STejun Heo { 804692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 80563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 807e22bee78STejun Heo 808e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 809e22bee78STejun Heo } 810e22bee78STejun Heo 811e22bee78STejun Heo /* 812e22bee78STejun Heo * Wake up functions. 813e22bee78STejun Heo */ 814e22bee78STejun Heo 8151037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8161037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8177e11629dSTejun Heo { 81863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8197e11629dSTejun Heo return NULL; 8207e11629dSTejun Heo 82163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8227e11629dSTejun Heo } 8237e11629dSTejun Heo 8247e11629dSTejun Heo /** 8257e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82663d95a91STejun Heo * @pool: worker pool to wake worker from 8277e11629dSTejun Heo * 82863d95a91STejun Heo * Wake up the first idle worker of @pool. 8297e11629dSTejun Heo * 8307e11629dSTejun Heo * CONTEXT: 831a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8327e11629dSTejun Heo */ 83363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8347e11629dSTejun Heo { 8351037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8367e11629dSTejun Heo 8377e11629dSTejun Heo if (likely(worker)) 8387e11629dSTejun Heo wake_up_process(worker->task); 8397e11629dSTejun Heo } 8407e11629dSTejun Heo 8414690c4abSTejun Heo /** 8426d25be57SThomas Gleixner * wq_worker_running - a worker is running again 843e22bee78STejun Heo * @task: task waking up 844e22bee78STejun Heo * 8456d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 846e22bee78STejun Heo */ 8476d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 848e22bee78STejun Heo { 849e22bee78STejun Heo struct worker *worker = kthread_data(task); 850e22bee78STejun Heo 8516d25be57SThomas Gleixner if (!worker->sleeping) 8526d25be57SThomas Gleixner return; 8536d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 854e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 8556d25be57SThomas Gleixner worker->sleeping = 0; 85636576000SJoonsoo Kim } 857e22bee78STejun Heo 858e22bee78STejun Heo /** 859e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 860e22bee78STejun Heo * @task: task going to sleep 861e22bee78STejun Heo * 8626d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 86362849a96SSebastian Andrzej Siewior * going to sleep. Preemption needs to be disabled to protect ->sleeping 86462849a96SSebastian Andrzej Siewior * assignment. 865e22bee78STejun Heo */ 8666d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 867e22bee78STejun Heo { 8686d25be57SThomas Gleixner struct worker *next, *worker = kthread_data(task); 869111c225aSTejun Heo struct worker_pool *pool; 870e22bee78STejun Heo 871111c225aSTejun Heo /* 872111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 873111c225aSTejun Heo * workers, also reach here, let's not access anything before 874111c225aSTejun Heo * checking NOT_RUNNING. 875111c225aSTejun Heo */ 8762d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 8776d25be57SThomas Gleixner return; 878e22bee78STejun Heo 879111c225aSTejun Heo pool = worker->pool; 880111c225aSTejun Heo 88162849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 88262849a96SSebastian Andrzej Siewior if (worker->sleeping) 8836d25be57SThomas Gleixner return; 8846d25be57SThomas Gleixner 8856d25be57SThomas Gleixner worker->sleeping = 1; 886a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 887e22bee78STejun Heo 888e22bee78STejun Heo /* 889e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 890e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 891e22bee78STejun Heo * Please read comment there. 892e22bee78STejun Heo * 893628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 894628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 895628c78e7STejun Heo * disabled, which in turn means that none else could be 896d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 897628c78e7STejun Heo * lock is safe. 898e22bee78STejun Heo */ 899e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 9006d25be57SThomas Gleixner !list_empty(&pool->worklist)) { 9016d25be57SThomas Gleixner next = first_idle_worker(pool); 9026d25be57SThomas Gleixner if (next) 9036d25be57SThomas Gleixner wake_up_process(next->task); 9046d25be57SThomas Gleixner } 905a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 906e22bee78STejun Heo } 907e22bee78STejun Heo 908e22bee78STejun Heo /** 9091b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9108194fe94SBart Van Assche * @task: Task to retrieve last work function of. 9111b69ac6bSJohannes Weiner * 9121b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9131b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9141b69ac6bSJohannes Weiner * 9151b69ac6bSJohannes Weiner * CONTEXT: 916a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 9171b69ac6bSJohannes Weiner * 9184b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9194b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9204b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9214b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9224b047002SJohannes Weiner * 9234b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9244b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9254b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9264b047002SJohannes Weiner * is guaranteed to not be processing any works. 9274b047002SJohannes Weiner * 9281b69ac6bSJohannes Weiner * Return: 9291b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9301b69ac6bSJohannes Weiner * hasn't executed any work yet. 9311b69ac6bSJohannes Weiner */ 9321b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9331b69ac6bSJohannes Weiner { 9341b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9351b69ac6bSJohannes Weiner 9361b69ac6bSJohannes Weiner return worker->last_func; 9371b69ac6bSJohannes Weiner } 9381b69ac6bSJohannes Weiner 9391b69ac6bSJohannes Weiner /** 940e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 941cb444766STejun Heo * @worker: self 942d302f017STejun Heo * @flags: flags to set 943d302f017STejun Heo * 944228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 945d302f017STejun Heo * 946cb444766STejun Heo * CONTEXT: 947a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 948d302f017STejun Heo */ 949228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 950d302f017STejun Heo { 951bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 952e22bee78STejun Heo 953cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 954cb444766STejun Heo 955228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 956e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 957e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 958e19e397aSTejun Heo atomic_dec(&pool->nr_running); 959e22bee78STejun Heo } 960e22bee78STejun Heo 961d302f017STejun Heo worker->flags |= flags; 962d302f017STejun Heo } 963d302f017STejun Heo 964d302f017STejun Heo /** 965e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 966cb444766STejun Heo * @worker: self 967d302f017STejun Heo * @flags: flags to clear 968d302f017STejun Heo * 969e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 970d302f017STejun Heo * 971cb444766STejun Heo * CONTEXT: 972a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 973d302f017STejun Heo */ 974d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 975d302f017STejun Heo { 97663d95a91STejun Heo struct worker_pool *pool = worker->pool; 977e22bee78STejun Heo unsigned int oflags = worker->flags; 978e22bee78STejun Heo 979cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 980cb444766STejun Heo 981d302f017STejun Heo worker->flags &= ~flags; 982e22bee78STejun Heo 98342c025f3STejun Heo /* 98442c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 98542c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 98642c025f3STejun Heo * of multiple flags, not a single flag. 98742c025f3STejun Heo */ 988e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 989e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 990e19e397aSTejun Heo atomic_inc(&pool->nr_running); 991d302f017STejun Heo } 992d302f017STejun Heo 993d302f017STejun Heo /** 9948cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 995c9e7cf27STejun Heo * @pool: pool of interest 9968cca0eeaSTejun Heo * @work: work to find worker for 9978cca0eeaSTejun Heo * 998c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 999c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1000a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1001a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1002a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1003a2c1c57bSTejun Heo * being executed. 1004a2c1c57bSTejun Heo * 1005a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1006a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1007a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1008a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1009a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1010a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1011a2c1c57bSTejun Heo * 1012c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1013c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1014c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1015c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1016c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1017c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10188cca0eeaSTejun Heo * 10198cca0eeaSTejun Heo * CONTEXT: 1020a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10218cca0eeaSTejun Heo * 1022d185af30SYacine Belkadi * Return: 1023d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10248cca0eeaSTejun Heo * otherwise. 10258cca0eeaSTejun Heo */ 1026c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10278cca0eeaSTejun Heo struct work_struct *work) 10288cca0eeaSTejun Heo { 102942f8570fSSasha Levin struct worker *worker; 103042f8570fSSasha Levin 1031b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1032a2c1c57bSTejun Heo (unsigned long)work) 1033a2c1c57bSTejun Heo if (worker->current_work == work && 1034a2c1c57bSTejun Heo worker->current_func == work->func) 103542f8570fSSasha Levin return worker; 103642f8570fSSasha Levin 103742f8570fSSasha Levin return NULL; 10388cca0eeaSTejun Heo } 10398cca0eeaSTejun Heo 10408cca0eeaSTejun Heo /** 1041bf4ede01STejun Heo * move_linked_works - move linked works to a list 1042bf4ede01STejun Heo * @work: start of series of works to be scheduled 1043bf4ede01STejun Heo * @head: target list to append @work to 1044402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1045bf4ede01STejun Heo * 1046bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1047bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1048bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1049bf4ede01STejun Heo * 1050bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1051bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1052bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1053bf4ede01STejun Heo * 1054bf4ede01STejun Heo * CONTEXT: 1055a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1056bf4ede01STejun Heo */ 1057bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1058bf4ede01STejun Heo struct work_struct **nextp) 1059bf4ede01STejun Heo { 1060bf4ede01STejun Heo struct work_struct *n; 1061bf4ede01STejun Heo 1062bf4ede01STejun Heo /* 1063bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1064bf4ede01STejun Heo * use NULL for list head. 1065bf4ede01STejun Heo */ 1066bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1067bf4ede01STejun Heo list_move_tail(&work->entry, head); 1068bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1069bf4ede01STejun Heo break; 1070bf4ede01STejun Heo } 1071bf4ede01STejun Heo 1072bf4ede01STejun Heo /* 1073bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1074bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1075bf4ede01STejun Heo * needs to be updated. 1076bf4ede01STejun Heo */ 1077bf4ede01STejun Heo if (nextp) 1078bf4ede01STejun Heo *nextp = n; 1079bf4ede01STejun Heo } 1080bf4ede01STejun Heo 10818864b4e5STejun Heo /** 10828864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10838864b4e5STejun Heo * @pwq: pool_workqueue to get 10848864b4e5STejun Heo * 10858864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10868864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10878864b4e5STejun Heo */ 10888864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10898864b4e5STejun Heo { 10908864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10918864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10928864b4e5STejun Heo pwq->refcnt++; 10938864b4e5STejun Heo } 10948864b4e5STejun Heo 10958864b4e5STejun Heo /** 10968864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10978864b4e5STejun Heo * @pwq: pool_workqueue to put 10988864b4e5STejun Heo * 10998864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11008864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11018864b4e5STejun Heo */ 11028864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11038864b4e5STejun Heo { 11048864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11058864b4e5STejun Heo if (likely(--pwq->refcnt)) 11068864b4e5STejun Heo return; 11078864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11088864b4e5STejun Heo return; 11098864b4e5STejun Heo /* 11108864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11118864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11128864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11138864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11148864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11158864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11168864b4e5STejun Heo */ 11178864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11188864b4e5STejun Heo } 11198864b4e5STejun Heo 1120dce90d47STejun Heo /** 1121dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1122dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1123dce90d47STejun Heo * 1124dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1125dce90d47STejun Heo */ 1126dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1127dce90d47STejun Heo { 1128dce90d47STejun Heo if (pwq) { 1129dce90d47STejun Heo /* 113024acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1131dce90d47STejun Heo * following lock operations are safe. 1132dce90d47STejun Heo */ 1133a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1134dce90d47STejun Heo put_pwq(pwq); 1135a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1136dce90d47STejun Heo } 1137dce90d47STejun Heo } 1138dce90d47STejun Heo 1139f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1140bf4ede01STejun Heo { 1141112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1142bf4ede01STejun Heo 1143bf4ede01STejun Heo trace_workqueue_activate_work(work); 114482607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 114582607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1146112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1147f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1148112202d9STejun Heo pwq->nr_active++; 1149bf4ede01STejun Heo } 1150bf4ede01STejun Heo 1151f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 11523aa62497SLai Jiangshan { 1153f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 11543aa62497SLai Jiangshan struct work_struct, entry); 11553aa62497SLai Jiangshan 1156f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 11573aa62497SLai Jiangshan } 11583aa62497SLai Jiangshan 1159bf4ede01STejun Heo /** 1160112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1161112202d9STejun Heo * @pwq: pwq of interest 1162*c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1163bf4ede01STejun Heo * 1164bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1165112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1166bf4ede01STejun Heo * 1167bf4ede01STejun Heo * CONTEXT: 1168a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1169bf4ede01STejun Heo */ 1170*c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1171bf4ede01STejun Heo { 1172*c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1173*c4560c2cSLai Jiangshan 11748864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1175bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11768864b4e5STejun Heo goto out_put; 1177bf4ede01STejun Heo 1178112202d9STejun Heo pwq->nr_in_flight[color]--; 1179bf4ede01STejun Heo 1180112202d9STejun Heo pwq->nr_active--; 1181f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1182f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1183112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1184f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1185bf4ede01STejun Heo } 1186bf4ede01STejun Heo 1187bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1188112202d9STejun Heo if (likely(pwq->flush_color != color)) 11898864b4e5STejun Heo goto out_put; 1190bf4ede01STejun Heo 1191bf4ede01STejun Heo /* are there still in-flight works? */ 1192112202d9STejun Heo if (pwq->nr_in_flight[color]) 11938864b4e5STejun Heo goto out_put; 1194bf4ede01STejun Heo 1195112202d9STejun Heo /* this pwq is done, clear flush_color */ 1196112202d9STejun Heo pwq->flush_color = -1; 1197bf4ede01STejun Heo 1198bf4ede01STejun Heo /* 1199112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1200bf4ede01STejun Heo * will handle the rest. 1201bf4ede01STejun Heo */ 1202112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1203112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12048864b4e5STejun Heo out_put: 12058864b4e5STejun Heo put_pwq(pwq); 1206bf4ede01STejun Heo } 1207bf4ede01STejun Heo 120836e227d2STejun Heo /** 1209bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 121036e227d2STejun Heo * @work: work item to steal 121136e227d2STejun Heo * @is_dwork: @work is a delayed_work 1212bbb68dfaSTejun Heo * @flags: place to store irq state 121336e227d2STejun Heo * 121436e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1215d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 121636e227d2STejun Heo * 1217d185af30SYacine Belkadi * Return: 12183eb6b31bSMauro Carvalho Chehab * 12193eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 122036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 122136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 122236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1223bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1224bbb68dfaSTejun Heo * for arbitrarily long 12253eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 122636e227d2STejun Heo * 1227d185af30SYacine Belkadi * Note: 1228bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1229e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1230e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1231e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1232bbb68dfaSTejun Heo * 1233bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1234bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1235bbb68dfaSTejun Heo * 1236e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1237bf4ede01STejun Heo */ 1238bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1239bbb68dfaSTejun Heo unsigned long *flags) 1240bf4ede01STejun Heo { 1241d565ed63STejun Heo struct worker_pool *pool; 1242112202d9STejun Heo struct pool_workqueue *pwq; 1243bf4ede01STejun Heo 1244bbb68dfaSTejun Heo local_irq_save(*flags); 1245bbb68dfaSTejun Heo 124636e227d2STejun Heo /* try to steal the timer if it exists */ 124736e227d2STejun Heo if (is_dwork) { 124836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 124936e227d2STejun Heo 1250e0aecdd8STejun Heo /* 1251e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1252e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1253e0aecdd8STejun Heo * running on the local CPU. 1254e0aecdd8STejun Heo */ 125536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 125636e227d2STejun Heo return 1; 125736e227d2STejun Heo } 125836e227d2STejun Heo 125936e227d2STejun Heo /* try to claim PENDING the normal way */ 1260bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1261bf4ede01STejun Heo return 0; 1262bf4ede01STejun Heo 126324acfb71SThomas Gleixner rcu_read_lock(); 1264bf4ede01STejun Heo /* 1265bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1266bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1267bf4ede01STejun Heo */ 1268d565ed63STejun Heo pool = get_work_pool(work); 1269d565ed63STejun Heo if (!pool) 1270bbb68dfaSTejun Heo goto fail; 1271bf4ede01STejun Heo 1272a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1273bf4ede01STejun Heo /* 1274112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1275112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1276112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1277112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1278112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12790b3dae68SLai Jiangshan * item is currently queued on that pool. 1280bf4ede01STejun Heo */ 1281112202d9STejun Heo pwq = get_work_pwq(work); 1282112202d9STejun Heo if (pwq && pwq->pool == pool) { 1283bf4ede01STejun Heo debug_work_deactivate(work); 12843aa62497SLai Jiangshan 12853aa62497SLai Jiangshan /* 1286f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 128716062836STejun Heo * it might have linked NO_COLOR work items which, if left 1288f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 128916062836STejun Heo * management later on and cause stall. Make sure the work 129016062836STejun Heo * item is activated before grabbing. 12913aa62497SLai Jiangshan */ 1292f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1293f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 12943aa62497SLai Jiangshan 1295bf4ede01STejun Heo list_del_init(&work->entry); 1296*c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 129736e227d2STejun Heo 1298112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12994468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 13004468a00fSLai Jiangshan 1301a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 130224acfb71SThomas Gleixner rcu_read_unlock(); 130336e227d2STejun Heo return 1; 1304bf4ede01STejun Heo } 1305a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1306bbb68dfaSTejun Heo fail: 130724acfb71SThomas Gleixner rcu_read_unlock(); 1308bbb68dfaSTejun Heo local_irq_restore(*flags); 1309bbb68dfaSTejun Heo if (work_is_canceling(work)) 1310bbb68dfaSTejun Heo return -ENOENT; 1311bbb68dfaSTejun Heo cpu_relax(); 131236e227d2STejun Heo return -EAGAIN; 1313bf4ede01STejun Heo } 1314bf4ede01STejun Heo 1315bf4ede01STejun Heo /** 1316706026c2STejun Heo * insert_work - insert a work into a pool 1317112202d9STejun Heo * @pwq: pwq @work belongs to 13184690c4abSTejun Heo * @work: work to insert 13194690c4abSTejun Heo * @head: insertion point 13204690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13214690c4abSTejun Heo * 1322112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1323706026c2STejun Heo * work_struct flags. 13244690c4abSTejun Heo * 13254690c4abSTejun Heo * CONTEXT: 1326a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1327365970a1SDavid Howells */ 1328112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1329112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1330b89deed3SOleg Nesterov { 1331112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1332e1d8aa9fSFrederic Weisbecker 1333e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1334e89a85d6SWalter Wu kasan_record_aux_stack(work); 1335e89a85d6SWalter Wu 13364690c4abSTejun Heo /* we own @work, set data and link */ 1337112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13381a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13398864b4e5STejun Heo get_pwq(pwq); 1340e22bee78STejun Heo 1341e22bee78STejun Heo /* 1342c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1343c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1344c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1345e22bee78STejun Heo */ 1346e22bee78STejun Heo smp_mb(); 1347e22bee78STejun Heo 134863d95a91STejun Heo if (__need_more_worker(pool)) 134963d95a91STejun Heo wake_up_worker(pool); 1350b89deed3SOleg Nesterov } 1351b89deed3SOleg Nesterov 1352c8efcc25STejun Heo /* 1353c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13548d03ecfeSTejun Heo * same workqueue. 1355c8efcc25STejun Heo */ 1356c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1357c8efcc25STejun Heo { 1358c8efcc25STejun Heo struct worker *worker; 1359c8efcc25STejun Heo 13608d03ecfeSTejun Heo worker = current_wq_worker(); 1361c8efcc25STejun Heo /* 1362bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 13638d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1364c8efcc25STejun Heo */ 1365112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1366c8efcc25STejun Heo } 1367c8efcc25STejun Heo 1368ef557180SMike Galbraith /* 1369ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1370ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1371ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1372ef557180SMike Galbraith */ 1373ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1374ef557180SMike Galbraith { 1375f303fccbSTejun Heo static bool printed_dbg_warning; 1376ef557180SMike Galbraith int new_cpu; 1377ef557180SMike Galbraith 1378f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1379ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1380ef557180SMike Galbraith return cpu; 1381f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1382f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1383f303fccbSTejun Heo printed_dbg_warning = true; 1384f303fccbSTejun Heo } 1385f303fccbSTejun Heo 1386ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1387ef557180SMike Galbraith return cpu; 1388ef557180SMike Galbraith 1389ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1390ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1391ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1392ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1393ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1394ef557180SMike Galbraith return cpu; 1395ef557180SMike Galbraith } 1396ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1397ef557180SMike Galbraith 1398ef557180SMike Galbraith return new_cpu; 1399ef557180SMike Galbraith } 1400ef557180SMike Galbraith 1401d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 14021da177e4SLinus Torvalds struct work_struct *work) 14031da177e4SLinus Torvalds { 1404112202d9STejun Heo struct pool_workqueue *pwq; 1405c9178087STejun Heo struct worker_pool *last_pool; 14061e19ffc6STejun Heo struct list_head *worklist; 14078a2e8e5dSTejun Heo unsigned int work_flags; 1408b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14098930cabaSTejun Heo 14108930cabaSTejun Heo /* 14118930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14128930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14138930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14148930cabaSTejun Heo * happen with IRQ disabled. 14158930cabaSTejun Heo */ 14168e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14171da177e4SLinus Torvalds 14181e19ffc6STejun Heo 14199ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1420618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1421c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1422e41e704bSTejun Heo return; 142324acfb71SThomas Gleixner rcu_read_lock(); 14249e8cd2f5STejun Heo retry: 1425aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1426aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1427df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1428ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1429df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1430aa202f1fSHillf Danton } else { 1431aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1432aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1433aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1434aa202f1fSHillf Danton } 1435f3421797STejun Heo 143618aa9effSTejun Heo /* 1437c9178087STejun Heo * If @work was previously on a different pool, it might still be 1438c9178087STejun Heo * running there, in which case the work needs to be queued on that 1439c9178087STejun Heo * pool to guarantee non-reentrancy. 144018aa9effSTejun Heo */ 1441c9e7cf27STejun Heo last_pool = get_work_pool(work); 1442112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 144318aa9effSTejun Heo struct worker *worker; 144418aa9effSTejun Heo 1445a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 144618aa9effSTejun Heo 1447c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 144818aa9effSTejun Heo 1449112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1450c9178087STejun Heo pwq = worker->current_pwq; 14518594fadeSLai Jiangshan } else { 145218aa9effSTejun Heo /* meh... not running there, queue here */ 1453a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1454a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 145518aa9effSTejun Heo } 14568930cabaSTejun Heo } else { 1457a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14588930cabaSTejun Heo } 1459502ca9d8STejun Heo 14609e8cd2f5STejun Heo /* 14619e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14629e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14639e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1464df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1465df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14669e8cd2f5STejun Heo * make forward-progress. 14679e8cd2f5STejun Heo */ 14689e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14699e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1470a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14719e8cd2f5STejun Heo cpu_relax(); 14729e8cd2f5STejun Heo goto retry; 14739e8cd2f5STejun Heo } 14749e8cd2f5STejun Heo /* oops */ 14759e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14769e8cd2f5STejun Heo wq->name, cpu); 14779e8cd2f5STejun Heo } 14789e8cd2f5STejun Heo 1479112202d9STejun Heo /* pwq determined, queue */ 1480112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1481502ca9d8STejun Heo 148224acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 148324acfb71SThomas Gleixner goto out; 14841e19ffc6STejun Heo 1485112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1486112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14871e19ffc6STejun Heo 1488112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1489cdadf009STejun Heo trace_workqueue_activate_work(work); 1490112202d9STejun Heo pwq->nr_active++; 1491112202d9STejun Heo worklist = &pwq->pool->worklist; 149282607adcSTejun Heo if (list_empty(worklist)) 149382607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14948a2e8e5dSTejun Heo } else { 1495f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1496f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 14978a2e8e5dSTejun Heo } 14981e19ffc6STejun Heo 14990687c66bSZqiang debug_work_activate(work); 1500112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15011e19ffc6STejun Heo 150224acfb71SThomas Gleixner out: 1503a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 150424acfb71SThomas Gleixner rcu_read_unlock(); 15051da177e4SLinus Torvalds } 15061da177e4SLinus Torvalds 15070fcb78c2SRolf Eike Beer /** 1508c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1509c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1510c1a220e7SZhang Rui * @wq: workqueue to use 1511c1a220e7SZhang Rui * @work: work to queue 1512c1a220e7SZhang Rui * 1513c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1514c1a220e7SZhang Rui * can't go away. 1515d185af30SYacine Belkadi * 1516d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1517c1a220e7SZhang Rui */ 1518d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1519d4283e93STejun Heo struct work_struct *work) 1520c1a220e7SZhang Rui { 1521d4283e93STejun Heo bool ret = false; 15228930cabaSTejun Heo unsigned long flags; 15238930cabaSTejun Heo 15248930cabaSTejun Heo local_irq_save(flags); 1525c1a220e7SZhang Rui 152622df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15274690c4abSTejun Heo __queue_work(cpu, wq, work); 1528d4283e93STejun Heo ret = true; 1529c1a220e7SZhang Rui } 15308930cabaSTejun Heo 15318930cabaSTejun Heo local_irq_restore(flags); 1532c1a220e7SZhang Rui return ret; 1533c1a220e7SZhang Rui } 1534ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1535c1a220e7SZhang Rui 15368204e0c1SAlexander Duyck /** 15378204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15388204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15398204e0c1SAlexander Duyck * 15408204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15418204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15428204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15438204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15448204e0c1SAlexander Duyck */ 15458204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15468204e0c1SAlexander Duyck { 15478204e0c1SAlexander Duyck int cpu; 15488204e0c1SAlexander Duyck 15498204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15508204e0c1SAlexander Duyck if (!wq_numa_enabled) 15518204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15528204e0c1SAlexander Duyck 15538204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15548204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15558204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15568204e0c1SAlexander Duyck 15578204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15588204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15598204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15608204e0c1SAlexander Duyck return cpu; 15618204e0c1SAlexander Duyck 15628204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15638204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15648204e0c1SAlexander Duyck 15658204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15668204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15678204e0c1SAlexander Duyck } 15688204e0c1SAlexander Duyck 15698204e0c1SAlexander Duyck /** 15708204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15718204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15728204e0c1SAlexander Duyck * @wq: workqueue to use 15738204e0c1SAlexander Duyck * @work: work to queue 15748204e0c1SAlexander Duyck * 15758204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15768204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15778204e0c1SAlexander Duyck * NUMA node. 15788204e0c1SAlexander Duyck * 15798204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 15808204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 15818204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 15828204e0c1SAlexander Duyck * 15838204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 15848204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 15858204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 15868204e0c1SAlexander Duyck * 15878204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 15888204e0c1SAlexander Duyck */ 15898204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 15908204e0c1SAlexander Duyck struct work_struct *work) 15918204e0c1SAlexander Duyck { 15928204e0c1SAlexander Duyck unsigned long flags; 15938204e0c1SAlexander Duyck bool ret = false; 15948204e0c1SAlexander Duyck 15958204e0c1SAlexander Duyck /* 15968204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 15978204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 15988204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 15998204e0c1SAlexander Duyck * 16008204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16018204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16028204e0c1SAlexander Duyck * some round robin type logic. 16038204e0c1SAlexander Duyck */ 16048204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16058204e0c1SAlexander Duyck 16068204e0c1SAlexander Duyck local_irq_save(flags); 16078204e0c1SAlexander Duyck 16088204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16098204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16108204e0c1SAlexander Duyck 16118204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16128204e0c1SAlexander Duyck ret = true; 16138204e0c1SAlexander Duyck } 16148204e0c1SAlexander Duyck 16158204e0c1SAlexander Duyck local_irq_restore(flags); 16168204e0c1SAlexander Duyck return ret; 16178204e0c1SAlexander Duyck } 16188204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16198204e0c1SAlexander Duyck 16208c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16211da177e4SLinus Torvalds { 16228c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16231da177e4SLinus Torvalds 1624e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 162560c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16261da177e4SLinus Torvalds } 16271438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16281da177e4SLinus Torvalds 16297beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 163052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16311da177e4SLinus Torvalds { 16327beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16337beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16341da177e4SLinus Torvalds 1635637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 163698173112SSami Tolvanen WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn); 1637fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1638fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16397beb2edfSTejun Heo 16408852aac2STejun Heo /* 16418852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16428852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16438852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16448852aac2STejun Heo * on that there's no such delay when @delay is 0. 16458852aac2STejun Heo */ 16468852aac2STejun Heo if (!delay) { 16478852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16488852aac2STejun Heo return; 16498852aac2STejun Heo } 16508852aac2STejun Heo 165160c057bcSLai Jiangshan dwork->wq = wq; 16521265057fSTejun Heo dwork->cpu = cpu; 16537beb2edfSTejun Heo timer->expires = jiffies + delay; 16547beb2edfSTejun Heo 1655041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16567beb2edfSTejun Heo add_timer_on(timer, cpu); 1657041bd12eSTejun Heo else 1658041bd12eSTejun Heo add_timer(timer); 16597beb2edfSTejun Heo } 16601da177e4SLinus Torvalds 16610fcb78c2SRolf Eike Beer /** 16620fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16630fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16640fcb78c2SRolf Eike Beer * @wq: workqueue to use 1665af9997e4SRandy Dunlap * @dwork: work to queue 16660fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16670fcb78c2SRolf Eike Beer * 1668d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1669715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1670715f1300STejun Heo * execution. 16710fcb78c2SRolf Eike Beer */ 1672d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 167352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16747a6bc1cdSVenkatesh Pallipadi { 167552bad64dSDavid Howells struct work_struct *work = &dwork->work; 1676d4283e93STejun Heo bool ret = false; 16778930cabaSTejun Heo unsigned long flags; 16788930cabaSTejun Heo 16798930cabaSTejun Heo /* read the comment in __queue_work() */ 16808930cabaSTejun Heo local_irq_save(flags); 16817a6bc1cdSVenkatesh Pallipadi 168222df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16837beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1684d4283e93STejun Heo ret = true; 16857a6bc1cdSVenkatesh Pallipadi } 16868930cabaSTejun Heo 16878930cabaSTejun Heo local_irq_restore(flags); 16887a6bc1cdSVenkatesh Pallipadi return ret; 16897a6bc1cdSVenkatesh Pallipadi } 1690ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 16911da177e4SLinus Torvalds 1692c8e55f36STejun Heo /** 16938376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 16948376fe22STejun Heo * @cpu: CPU number to execute work on 16958376fe22STejun Heo * @wq: workqueue to use 16968376fe22STejun Heo * @dwork: work to queue 16978376fe22STejun Heo * @delay: number of jiffies to wait before queueing 16988376fe22STejun Heo * 16998376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17008376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17018376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17028376fe22STejun Heo * current state. 17038376fe22STejun Heo * 1704d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17058376fe22STejun Heo * pending and its timer was modified. 17068376fe22STejun Heo * 1707e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17088376fe22STejun Heo * See try_to_grab_pending() for details. 17098376fe22STejun Heo */ 17108376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17118376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17128376fe22STejun Heo { 17138376fe22STejun Heo unsigned long flags; 17148376fe22STejun Heo int ret; 17158376fe22STejun Heo 17168376fe22STejun Heo do { 17178376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17188376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17198376fe22STejun Heo 17208376fe22STejun Heo if (likely(ret >= 0)) { 17218376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17228376fe22STejun Heo local_irq_restore(flags); 17238376fe22STejun Heo } 17248376fe22STejun Heo 17258376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17268376fe22STejun Heo return ret; 17278376fe22STejun Heo } 17288376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17298376fe22STejun Heo 173005f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 173105f0fe6bSTejun Heo { 173205f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 173305f0fe6bSTejun Heo 173405f0fe6bSTejun Heo /* read the comment in __queue_work() */ 173505f0fe6bSTejun Heo local_irq_disable(); 173605f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 173705f0fe6bSTejun Heo local_irq_enable(); 173805f0fe6bSTejun Heo } 173905f0fe6bSTejun Heo 174005f0fe6bSTejun Heo /** 174105f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 174205f0fe6bSTejun Heo * @wq: workqueue to use 174305f0fe6bSTejun Heo * @rwork: work to queue 174405f0fe6bSTejun Heo * 174505f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 174605f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1747bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 174805f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 174905f0fe6bSTejun Heo */ 175005f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 175105f0fe6bSTejun Heo { 175205f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 175305f0fe6bSTejun Heo 175405f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 175505f0fe6bSTejun Heo rwork->wq = wq; 175605f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 175705f0fe6bSTejun Heo return true; 175805f0fe6bSTejun Heo } 175905f0fe6bSTejun Heo 176005f0fe6bSTejun Heo return false; 176105f0fe6bSTejun Heo } 176205f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 176305f0fe6bSTejun Heo 17648376fe22STejun Heo /** 1765c8e55f36STejun Heo * worker_enter_idle - enter idle state 1766c8e55f36STejun Heo * @worker: worker which is entering idle state 1767c8e55f36STejun Heo * 1768c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1769c8e55f36STejun Heo * necessary. 1770c8e55f36STejun Heo * 1771c8e55f36STejun Heo * LOCKING: 1772a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1773c8e55f36STejun Heo */ 1774c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17751da177e4SLinus Torvalds { 1776bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1777c8e55f36STejun Heo 17786183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 17796183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 17806183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 17816183c009STejun Heo return; 1782c8e55f36STejun Heo 1783051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1784cb444766STejun Heo worker->flags |= WORKER_IDLE; 1785bd7bdd43STejun Heo pool->nr_idle++; 1786e22bee78STejun Heo worker->last_active = jiffies; 1787c8e55f36STejun Heo 1788c8e55f36STejun Heo /* idle_list is LIFO */ 1789bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1790db7bccf4STejun Heo 179163d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1792628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1793cb444766STejun Heo 1794544ecf31STejun Heo /* 1795e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1796d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1797628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1798628c78e7STejun Heo * unbind is not in progress. 1799544ecf31STejun Heo */ 180024647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1801bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1802e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1803c8e55f36STejun Heo } 1804c8e55f36STejun Heo 1805c8e55f36STejun Heo /** 1806c8e55f36STejun Heo * worker_leave_idle - leave idle state 1807c8e55f36STejun Heo * @worker: worker which is leaving idle state 1808c8e55f36STejun Heo * 1809c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1810c8e55f36STejun Heo * 1811c8e55f36STejun Heo * LOCKING: 1812a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1813c8e55f36STejun Heo */ 1814c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1815c8e55f36STejun Heo { 1816bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1817c8e55f36STejun Heo 18186183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18196183c009STejun Heo return; 1820d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1821bd7bdd43STejun Heo pool->nr_idle--; 1822c8e55f36STejun Heo list_del_init(&worker->entry); 1823c8e55f36STejun Heo } 1824c8e55f36STejun Heo 1825f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1826c34056a3STejun Heo { 1827c34056a3STejun Heo struct worker *worker; 1828c34056a3STejun Heo 1829f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1830c8e55f36STejun Heo if (worker) { 1831c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1832affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1833da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1834e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1835e22bee78STejun Heo worker->flags = WORKER_PREP; 1836c8e55f36STejun Heo } 1837c34056a3STejun Heo return worker; 1838c34056a3STejun Heo } 1839c34056a3STejun Heo 1840c34056a3STejun Heo /** 18414736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18424736cbf7SLai Jiangshan * @worker: worker to be attached 18434736cbf7SLai Jiangshan * @pool: the target pool 18444736cbf7SLai Jiangshan * 18454736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18464736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18474736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18484736cbf7SLai Jiangshan */ 18494736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18504736cbf7SLai Jiangshan struct worker_pool *pool) 18514736cbf7SLai Jiangshan { 18521258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18534736cbf7SLai Jiangshan 18544736cbf7SLai Jiangshan /* 18551258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18561258fae7STejun Heo * stable across this function. See the comments above the flag 18571258fae7STejun Heo * definition for details. 18584736cbf7SLai Jiangshan */ 18594736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18604736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18615c25b5ffSPeter Zijlstra else 18625c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 18634736cbf7SLai Jiangshan 1864640f17c8SPeter Zijlstra if (worker->rescue_wq) 1865640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1866640f17c8SPeter Zijlstra 18674736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1868a2d812a2STejun Heo worker->pool = pool; 18694736cbf7SLai Jiangshan 18701258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18714736cbf7SLai Jiangshan } 18724736cbf7SLai Jiangshan 18734736cbf7SLai Jiangshan /** 187460f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 187560f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 187660f5a4bcSLai Jiangshan * 18774736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18784736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 18794736cbf7SLai Jiangshan * other reference to the pool. 188060f5a4bcSLai Jiangshan */ 1881a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 188260f5a4bcSLai Jiangshan { 1883a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 188460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 188560f5a4bcSLai Jiangshan 18861258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1887a2d812a2STejun Heo 18885c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 1889da028469SLai Jiangshan list_del(&worker->node); 1890a2d812a2STejun Heo worker->pool = NULL; 1891a2d812a2STejun Heo 1892da028469SLai Jiangshan if (list_empty(&pool->workers)) 189360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 18941258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 189560f5a4bcSLai Jiangshan 1896b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1897b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1898b62c0751SLai Jiangshan 189960f5a4bcSLai Jiangshan if (detach_completion) 190060f5a4bcSLai Jiangshan complete(detach_completion); 190160f5a4bcSLai Jiangshan } 190260f5a4bcSLai Jiangshan 190360f5a4bcSLai Jiangshan /** 1904c34056a3STejun Heo * create_worker - create a new workqueue worker 190563d95a91STejun Heo * @pool: pool the new worker will belong to 1906c34056a3STejun Heo * 1907051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1908c34056a3STejun Heo * 1909c34056a3STejun Heo * CONTEXT: 1910c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1911c34056a3STejun Heo * 1912d185af30SYacine Belkadi * Return: 1913c34056a3STejun Heo * Pointer to the newly created worker. 1914c34056a3STejun Heo */ 1915bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1916c34056a3STejun Heo { 1917e441b56fSZhen Lei struct worker *worker; 1918e441b56fSZhen Lei int id; 1919e3c916a4STejun Heo char id_buf[16]; 1920c34056a3STejun Heo 19217cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 1922e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 1923822d8405STejun Heo if (id < 0) 1924e441b56fSZhen Lei return NULL; 1925c34056a3STejun Heo 1926f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1927c34056a3STejun Heo if (!worker) 1928c34056a3STejun Heo goto fail; 1929c34056a3STejun Heo 1930c34056a3STejun Heo worker->id = id; 1931c34056a3STejun Heo 193229c91e99STejun Heo if (pool->cpu >= 0) 1933e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1934e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1935f3421797STejun Heo else 1936e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1937e3c916a4STejun Heo 1938f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1939e3c916a4STejun Heo "kworker/%s", id_buf); 1940c34056a3STejun Heo if (IS_ERR(worker->task)) 1941c34056a3STejun Heo goto fail; 1942c34056a3STejun Heo 194391151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 194425834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 194591151228SOleg Nesterov 1946da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19474736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1948822d8405STejun Heo 1949051e1850SLai Jiangshan /* start the newly created worker */ 1950a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1951051e1850SLai Jiangshan worker->pool->nr_workers++; 1952051e1850SLai Jiangshan worker_enter_idle(worker); 1953051e1850SLai Jiangshan wake_up_process(worker->task); 1954a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1955051e1850SLai Jiangshan 1956c34056a3STejun Heo return worker; 1957822d8405STejun Heo 1958c34056a3STejun Heo fail: 1959e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 1960c34056a3STejun Heo kfree(worker); 1961c34056a3STejun Heo return NULL; 1962c34056a3STejun Heo } 1963c34056a3STejun Heo 1964c34056a3STejun Heo /** 1965c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1966c34056a3STejun Heo * @worker: worker to be destroyed 1967c34056a3STejun Heo * 196873eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 196973eb7fe7SLai Jiangshan * be idle. 1970c8e55f36STejun Heo * 1971c8e55f36STejun Heo * CONTEXT: 1972a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1973c34056a3STejun Heo */ 1974c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1975c34056a3STejun Heo { 1976bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1977c34056a3STejun Heo 1978cd549687STejun Heo lockdep_assert_held(&pool->lock); 1979cd549687STejun Heo 1980c34056a3STejun Heo /* sanity check frenzy */ 19816183c009STejun Heo if (WARN_ON(worker->current_work) || 198273eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 198373eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 19846183c009STejun Heo return; 1985c34056a3STejun Heo 1986bd7bdd43STejun Heo pool->nr_workers--; 1987bd7bdd43STejun Heo pool->nr_idle--; 1988c8e55f36STejun Heo 1989c8e55f36STejun Heo list_del_init(&worker->entry); 1990cb444766STejun Heo worker->flags |= WORKER_DIE; 199160f5a4bcSLai Jiangshan wake_up_process(worker->task); 1992c34056a3STejun Heo } 1993c34056a3STejun Heo 199432a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1995e22bee78STejun Heo { 199632a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1997e22bee78STejun Heo 1998a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1999e22bee78STejun Heo 20003347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2001e22bee78STejun Heo struct worker *worker; 2002e22bee78STejun Heo unsigned long expires; 2003e22bee78STejun Heo 2004e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 200563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2006e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2007e22bee78STejun Heo 20083347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 200963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 20103347fc9fSLai Jiangshan break; 2011e22bee78STejun Heo } 20123347fc9fSLai Jiangshan 20133347fc9fSLai Jiangshan destroy_worker(worker); 2014e22bee78STejun Heo } 2015e22bee78STejun Heo 2016a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2017e22bee78STejun Heo } 2018e22bee78STejun Heo 2019493a1724STejun Heo static void send_mayday(struct work_struct *work) 2020e22bee78STejun Heo { 2021112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2022112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2023493a1724STejun Heo 20242e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2025e22bee78STejun Heo 2026493008a8STejun Heo if (!wq->rescuer) 2027493a1724STejun Heo return; 2028e22bee78STejun Heo 2029e22bee78STejun Heo /* mayday mayday mayday */ 2030493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 203177668c8bSLai Jiangshan /* 203277668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 203377668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 203477668c8bSLai Jiangshan * rescuer is done with it. 203577668c8bSLai Jiangshan */ 203677668c8bSLai Jiangshan get_pwq(pwq); 2037493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2038e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2039493a1724STejun Heo } 2040e22bee78STejun Heo } 2041e22bee78STejun Heo 204232a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2043e22bee78STejun Heo { 204432a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2045e22bee78STejun Heo struct work_struct *work; 2046e22bee78STejun Heo 2047a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2048a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2049e22bee78STejun Heo 205063d95a91STejun Heo if (need_to_create_worker(pool)) { 2051e22bee78STejun Heo /* 2052e22bee78STejun Heo * We've been trying to create a new worker but 2053e22bee78STejun Heo * haven't been successful. We might be hitting an 2054e22bee78STejun Heo * allocation deadlock. Send distress signals to 2055e22bee78STejun Heo * rescuers. 2056e22bee78STejun Heo */ 205763d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2058e22bee78STejun Heo send_mayday(work); 2059e22bee78STejun Heo } 2060e22bee78STejun Heo 2061a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2062a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2063e22bee78STejun Heo 206463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2065e22bee78STejun Heo } 2066e22bee78STejun Heo 2067e22bee78STejun Heo /** 2068e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 206963d95a91STejun Heo * @pool: pool to create a new worker for 2070e22bee78STejun Heo * 207163d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2072e22bee78STejun Heo * have at least one idle worker on return from this function. If 2073e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 207463d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2075e22bee78STejun Heo * possible allocation deadlock. 2076e22bee78STejun Heo * 2077c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2078c5aa87bbSTejun Heo * may_start_working() %true. 2079e22bee78STejun Heo * 2080e22bee78STejun Heo * LOCKING: 2081a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2082e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2083e22bee78STejun Heo * manager. 2084e22bee78STejun Heo */ 208529187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2086d565ed63STejun Heo __releases(&pool->lock) 2087d565ed63STejun Heo __acquires(&pool->lock) 2088e22bee78STejun Heo { 2089e22bee78STejun Heo restart: 2090a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 20919f9c2364STejun Heo 2092e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 209363d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2094e22bee78STejun Heo 2095e22bee78STejun Heo while (true) { 2096051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2097e22bee78STejun Heo break; 2098e22bee78STejun Heo 2099e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 21009f9c2364STejun Heo 210163d95a91STejun Heo if (!need_to_create_worker(pool)) 2102e22bee78STejun Heo break; 2103e22bee78STejun Heo } 2104e22bee78STejun Heo 210563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2106a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2107051e1850SLai Jiangshan /* 2108051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2109051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2110051e1850SLai Jiangshan * already become busy. 2111051e1850SLai Jiangshan */ 211263d95a91STejun Heo if (need_to_create_worker(pool)) 2113e22bee78STejun Heo goto restart; 2114e22bee78STejun Heo } 2115e22bee78STejun Heo 2116e22bee78STejun Heo /** 2117e22bee78STejun Heo * manage_workers - manage worker pool 2118e22bee78STejun Heo * @worker: self 2119e22bee78STejun Heo * 2120706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2121e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2122706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2123e22bee78STejun Heo * 2124e22bee78STejun Heo * The caller can safely start processing works on false return. On 2125e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2126e22bee78STejun Heo * and may_start_working() is true. 2127e22bee78STejun Heo * 2128e22bee78STejun Heo * CONTEXT: 2129a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2130e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2131e22bee78STejun Heo * 2132d185af30SYacine Belkadi * Return: 213329187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 213429187a9eSTejun Heo * start processing works, %true if management function was performed and 213529187a9eSTejun Heo * the conditions that the caller verified before calling the function may 213629187a9eSTejun Heo * no longer be true. 2137e22bee78STejun Heo */ 2138e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2139e22bee78STejun Heo { 214063d95a91STejun Heo struct worker_pool *pool = worker->pool; 2141e22bee78STejun Heo 2142692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 214329187a9eSTejun Heo return false; 2144692b4825STejun Heo 2145692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 21462607d7a6STejun Heo pool->manager = worker; 2147e22bee78STejun Heo 214829187a9eSTejun Heo maybe_create_worker(pool); 2149e22bee78STejun Heo 21502607d7a6STejun Heo pool->manager = NULL; 2151692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2152d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 215329187a9eSTejun Heo return true; 2154e22bee78STejun Heo } 2155e22bee78STejun Heo 2156a62428c0STejun Heo /** 2157a62428c0STejun Heo * process_one_work - process single work 2158c34056a3STejun Heo * @worker: self 2159a62428c0STejun Heo * @work: work to process 2160a62428c0STejun Heo * 2161a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2162a62428c0STejun Heo * process a single work including synchronization against and 2163a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2164a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2165a62428c0STejun Heo * call this function to process a work. 2166a62428c0STejun Heo * 2167a62428c0STejun Heo * CONTEXT: 2168a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2169a62428c0STejun Heo */ 2170c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2171d565ed63STejun Heo __releases(&pool->lock) 2172d565ed63STejun Heo __acquires(&pool->lock) 21731da177e4SLinus Torvalds { 2174112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2175bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2176112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2177*c4560c2cSLai Jiangshan unsigned long work_data; 21787e11629dSTejun Heo struct worker *collision; 21794e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 21804e6045f1SJohannes Berg /* 2181a62428c0STejun Heo * It is permissible to free the struct work_struct from 2182a62428c0STejun Heo * inside the function that is called from it, this we need to 2183a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2184a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2185a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 21864e6045f1SJohannes Berg */ 21874d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 21884d82a1deSPeter Zijlstra 21894d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 21904e6045f1SJohannes Berg #endif 2191807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 219285327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2193ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 219425511a47STejun Heo 21957e11629dSTejun Heo /* 21967e11629dSTejun Heo * A single work shouldn't be executed concurrently by 21977e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 21987e11629dSTejun Heo * already processing the work. If so, defer the work to the 21997e11629dSTejun Heo * currently executing one. 22007e11629dSTejun Heo */ 2201c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 22027e11629dSTejun Heo if (unlikely(collision)) { 22037e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 22047e11629dSTejun Heo return; 22057e11629dSTejun Heo } 22061da177e4SLinus Torvalds 22078930cabaSTejun Heo /* claim and dequeue */ 22081da177e4SLinus Torvalds debug_work_deactivate(work); 2209c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2210c34056a3STejun Heo worker->current_work = work; 2211a2c1c57bSTejun Heo worker->current_func = work->func; 2212112202d9STejun Heo worker->current_pwq = pwq; 2213*c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 22147a22ad75STejun Heo 22158bf89593STejun Heo /* 22168bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 22178bf89593STejun Heo * overridden through set_worker_desc(). 22188bf89593STejun Heo */ 22198bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 22208bf89593STejun Heo 2221a62428c0STejun Heo list_del_init(&work->entry); 2222a62428c0STejun Heo 2223649027d7STejun Heo /* 2224228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2225228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2226228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2227228f1d00SLai Jiangshan * execution of the pending work items. 2228fb0e7bebSTejun Heo */ 2229fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2230228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2231fb0e7bebSTejun Heo 2232974271c4STejun Heo /* 2233a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2234a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2235a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2236a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2237228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2238974271c4STejun Heo */ 2239a489a03eSLai Jiangshan if (need_more_worker(pool)) 224063d95a91STejun Heo wake_up_worker(pool); 2241974271c4STejun Heo 22428930cabaSTejun Heo /* 22437c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2244d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 224523657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 224623657bb1STejun Heo * disabled. 22478930cabaSTejun Heo */ 22487c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22491da177e4SLinus Torvalds 2250a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2251365970a1SDavid Howells 2252a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 22533295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2254e6f3faa7SPeter Zijlstra /* 2255f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2256f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2257e6f3faa7SPeter Zijlstra * 2258e6f3faa7SPeter Zijlstra * However, that would result in: 2259e6f3faa7SPeter Zijlstra * 2260e6f3faa7SPeter Zijlstra * A(W1) 2261e6f3faa7SPeter Zijlstra * WFC(C) 2262e6f3faa7SPeter Zijlstra * A(W1) 2263e6f3faa7SPeter Zijlstra * C(C) 2264e6f3faa7SPeter Zijlstra * 2265e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2266e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2267e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2268f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2269e6f3faa7SPeter Zijlstra * these locks. 2270e6f3faa7SPeter Zijlstra * 2271e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2272e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2273e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2274e6f3faa7SPeter Zijlstra */ 2275f52be570SPeter Zijlstra lockdep_invariant_state(true); 2276e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2277a2c1c57bSTejun Heo worker->current_func(work); 2278e36c886aSArjan van de Ven /* 2279e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2280e36c886aSArjan van de Ven * point will only record its address. 2281e36c886aSArjan van de Ven */ 22821c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 22833295f0efSIngo Molnar lock_map_release(&lockdep_map); 2284112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 22851da177e4SLinus Torvalds 2286d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2287044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2288d75f773cSSakari Ailus " last function: %ps\n", 2289a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2290a2c1c57bSTejun Heo worker->current_func); 2291d5abe669SPeter Zijlstra debug_show_held_locks(current); 2292d5abe669SPeter Zijlstra dump_stack(); 2293d5abe669SPeter Zijlstra } 2294d5abe669SPeter Zijlstra 2295b22ce278STejun Heo /* 2296025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2297b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2298b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2299b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2300789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2301789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2302b22ce278STejun Heo */ 2303a7e6425eSPaul E. McKenney cond_resched(); 2304b22ce278STejun Heo 2305a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2306a62428c0STejun Heo 2307fb0e7bebSTejun Heo /* clear cpu intensive status */ 2308fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2309fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2310fb0e7bebSTejun Heo 23111b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 23121b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 23131b69ac6bSJohannes Weiner 2314a62428c0STejun Heo /* we're done with it, release */ 231542f8570fSSasha Levin hash_del(&worker->hentry); 2316c34056a3STejun Heo worker->current_work = NULL; 2317a2c1c57bSTejun Heo worker->current_func = NULL; 2318112202d9STejun Heo worker->current_pwq = NULL; 2319*c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 23201da177e4SLinus Torvalds } 23211da177e4SLinus Torvalds 2322affee4b2STejun Heo /** 2323affee4b2STejun Heo * process_scheduled_works - process scheduled works 2324affee4b2STejun Heo * @worker: self 2325affee4b2STejun Heo * 2326affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2327affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2328affee4b2STejun Heo * fetches a work from the top and executes it. 2329affee4b2STejun Heo * 2330affee4b2STejun Heo * CONTEXT: 2331a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2332affee4b2STejun Heo * multiple times. 2333affee4b2STejun Heo */ 2334affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 23351da177e4SLinus Torvalds { 2336affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2337affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2338a62428c0STejun Heo struct work_struct, entry); 2339c34056a3STejun Heo process_one_work(worker, work); 2340a62428c0STejun Heo } 23411da177e4SLinus Torvalds } 23421da177e4SLinus Torvalds 2343197f6accSTejun Heo static void set_pf_worker(bool val) 2344197f6accSTejun Heo { 2345197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2346197f6accSTejun Heo if (val) 2347197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2348197f6accSTejun Heo else 2349197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2350197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2351197f6accSTejun Heo } 2352197f6accSTejun Heo 23534690c4abSTejun Heo /** 23544690c4abSTejun Heo * worker_thread - the worker thread function 2355c34056a3STejun Heo * @__worker: self 23564690c4abSTejun Heo * 2357c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2358c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2359c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2360c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2361c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2362d185af30SYacine Belkadi * 2363d185af30SYacine Belkadi * Return: 0 23644690c4abSTejun Heo */ 2365c34056a3STejun Heo static int worker_thread(void *__worker) 23661da177e4SLinus Torvalds { 2367c34056a3STejun Heo struct worker *worker = __worker; 2368bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23691da177e4SLinus Torvalds 2370e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2371197f6accSTejun Heo set_pf_worker(true); 2372c8e55f36STejun Heo woke_up: 2373a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2374affee4b2STejun Heo 2375a9ab775bSTejun Heo /* am I supposed to die? */ 2376a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2377a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2378a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2379197f6accSTejun Heo set_pf_worker(false); 238060f5a4bcSLai Jiangshan 238160f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2382e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2383a2d812a2STejun Heo worker_detach_from_pool(worker); 238460f5a4bcSLai Jiangshan kfree(worker); 2385c8e55f36STejun Heo return 0; 2386c8e55f36STejun Heo } 2387c8e55f36STejun Heo 2388c8e55f36STejun Heo worker_leave_idle(worker); 2389db7bccf4STejun Heo recheck: 2390e22bee78STejun Heo /* no more worker necessary? */ 239163d95a91STejun Heo if (!need_more_worker(pool)) 2392e22bee78STejun Heo goto sleep; 2393e22bee78STejun Heo 2394e22bee78STejun Heo /* do we need to manage? */ 239563d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2396e22bee78STejun Heo goto recheck; 2397e22bee78STejun Heo 2398c8e55f36STejun Heo /* 2399c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2400c8e55f36STejun Heo * preparing to process a work or actually processing it. 2401c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2402c8e55f36STejun Heo */ 24036183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2404c8e55f36STejun Heo 2405e22bee78STejun Heo /* 2406a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2407a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2408a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2409a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2410a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2411e22bee78STejun Heo */ 2412a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2413e22bee78STejun Heo 2414e22bee78STejun Heo do { 2415affee4b2STejun Heo struct work_struct *work = 2416bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2417affee4b2STejun Heo struct work_struct, entry); 2418affee4b2STejun Heo 241982607adcSTejun Heo pool->watchdog_ts = jiffies; 242082607adcSTejun Heo 2421c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2422affee4b2STejun Heo /* optimization path, not strictly necessary */ 2423affee4b2STejun Heo process_one_work(worker, work); 2424affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2425affee4b2STejun Heo process_scheduled_works(worker); 2426affee4b2STejun Heo } else { 2427c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2428affee4b2STejun Heo process_scheduled_works(worker); 2429affee4b2STejun Heo } 243063d95a91STejun Heo } while (keep_working(pool)); 2431affee4b2STejun Heo 2432228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2433d313dd85STejun Heo sleep: 2434c8e55f36STejun Heo /* 2435d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2436d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2437d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2438d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2439d565ed63STejun Heo * event. 2440c8e55f36STejun Heo */ 2441c8e55f36STejun Heo worker_enter_idle(worker); 2442c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2443a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24441da177e4SLinus Torvalds schedule(); 2445c8e55f36STejun Heo goto woke_up; 24461da177e4SLinus Torvalds } 24471da177e4SLinus Torvalds 2448e22bee78STejun Heo /** 2449e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2450111c225aSTejun Heo * @__rescuer: self 2451e22bee78STejun Heo * 2452e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2453493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2454e22bee78STejun Heo * 2455706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2456e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2457e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2458e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2459e22bee78STejun Heo * the problem rescuer solves. 2460e22bee78STejun Heo * 2461706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2462706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2463e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2464e22bee78STejun Heo * 2465e22bee78STejun Heo * This should happen rarely. 2466d185af30SYacine Belkadi * 2467d185af30SYacine Belkadi * Return: 0 2468e22bee78STejun Heo */ 2469111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2470e22bee78STejun Heo { 2471111c225aSTejun Heo struct worker *rescuer = __rescuer; 2472111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2473e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 24744d595b86SLai Jiangshan bool should_stop; 2475e22bee78STejun Heo 2476e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2477111c225aSTejun Heo 2478111c225aSTejun Heo /* 2479111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2480111c225aSTejun Heo * doesn't participate in concurrency management. 2481111c225aSTejun Heo */ 2482197f6accSTejun Heo set_pf_worker(true); 2483e22bee78STejun Heo repeat: 2484c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 24851da177e4SLinus Torvalds 24864d595b86SLai Jiangshan /* 24874d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 24884d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 24894d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 24904d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 24914d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 24924d595b86SLai Jiangshan * list is always empty on exit. 24934d595b86SLai Jiangshan */ 24944d595b86SLai Jiangshan should_stop = kthread_should_stop(); 24951da177e4SLinus Torvalds 2496493a1724STejun Heo /* see whether any pwq is asking for help */ 2497a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2498493a1724STejun Heo 2499493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2500493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2501493a1724STejun Heo struct pool_workqueue, mayday_node); 2502112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2503e22bee78STejun Heo struct work_struct *work, *n; 250482607adcSTejun Heo bool first = true; 2505e22bee78STejun Heo 2506e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2507493a1724STejun Heo list_del_init(&pwq->mayday_node); 2508493a1724STejun Heo 2509a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2510e22bee78STejun Heo 251151697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 251251697d39SLai Jiangshan 2513a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2514e22bee78STejun Heo 2515e22bee78STejun Heo /* 2516e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2517e22bee78STejun Heo * process'em. 2518e22bee78STejun Heo */ 25190479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 252082607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 252182607adcSTejun Heo if (get_work_pwq(work) == pwq) { 252282607adcSTejun Heo if (first) 252382607adcSTejun Heo pool->watchdog_ts = jiffies; 2524e22bee78STejun Heo move_linked_works(work, scheduled, &n); 252582607adcSTejun Heo } 252682607adcSTejun Heo first = false; 252782607adcSTejun Heo } 2528e22bee78STejun Heo 2529008847f6SNeilBrown if (!list_empty(scheduled)) { 2530e22bee78STejun Heo process_scheduled_works(rescuer); 25317576958aSTejun Heo 25327576958aSTejun Heo /* 2533008847f6SNeilBrown * The above execution of rescued work items could 2534008847f6SNeilBrown * have created more to rescue through 2535f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2536008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2537008847f6SNeilBrown * that such back-to-back work items, which may be 2538008847f6SNeilBrown * being used to relieve memory pressure, don't 2539008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2540008847f6SNeilBrown */ 25414f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2542a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2543e66b39afSTejun Heo /* 2544e66b39afSTejun Heo * Queue iff we aren't racing destruction 2545e66b39afSTejun Heo * and somebody else hasn't queued it already. 2546e66b39afSTejun Heo */ 2547e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2548008847f6SNeilBrown get_pwq(pwq); 2549e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2550e66b39afSTejun Heo } 2551a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2552008847f6SNeilBrown } 2553008847f6SNeilBrown } 2554008847f6SNeilBrown 2555008847f6SNeilBrown /* 255677668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 255713b1d625SLai Jiangshan * go away while we're still attached to it. 255877668c8bSLai Jiangshan */ 255977668c8bSLai Jiangshan put_pwq(pwq); 256077668c8bSLai Jiangshan 256177668c8bSLai Jiangshan /* 2562d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 25637576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 25647576958aSTejun Heo * and stalling the execution. 25657576958aSTejun Heo */ 2566d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 256763d95a91STejun Heo wake_up_worker(pool); 25687576958aSTejun Heo 2569a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 257013b1d625SLai Jiangshan 2571a2d812a2STejun Heo worker_detach_from_pool(rescuer); 257213b1d625SLai Jiangshan 2573a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 25741da177e4SLinus Torvalds } 25751da177e4SLinus Torvalds 2576a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2577493a1724STejun Heo 25784d595b86SLai Jiangshan if (should_stop) { 25794d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2580197f6accSTejun Heo set_pf_worker(false); 25814d595b86SLai Jiangshan return 0; 25824d595b86SLai Jiangshan } 25834d595b86SLai Jiangshan 2584111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2585111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2586e22bee78STejun Heo schedule(); 2587e22bee78STejun Heo goto repeat; 25881da177e4SLinus Torvalds } 25891da177e4SLinus Torvalds 2590fca839c0STejun Heo /** 2591fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2592fca839c0STejun Heo * @target_wq: workqueue being flushed 2593fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2594fca839c0STejun Heo * 2595fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2596fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2597fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2598fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2599fca839c0STejun Heo * a deadlock. 2600fca839c0STejun Heo */ 2601fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2602fca839c0STejun Heo struct work_struct *target_work) 2603fca839c0STejun Heo { 2604fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2605fca839c0STejun Heo struct worker *worker; 2606fca839c0STejun Heo 2607fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2608fca839c0STejun Heo return; 2609fca839c0STejun Heo 2610fca839c0STejun Heo worker = current_wq_worker(); 2611fca839c0STejun Heo 2612fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2613d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2614fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 261523d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 261623d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2617d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2618fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2619fca839c0STejun Heo target_wq->name, target_func); 2620fca839c0STejun Heo } 2621fca839c0STejun Heo 2622fc2e4d70SOleg Nesterov struct wq_barrier { 2623fc2e4d70SOleg Nesterov struct work_struct work; 2624fc2e4d70SOleg Nesterov struct completion done; 26252607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2626fc2e4d70SOleg Nesterov }; 2627fc2e4d70SOleg Nesterov 2628fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2629fc2e4d70SOleg Nesterov { 2630fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2631fc2e4d70SOleg Nesterov complete(&barr->done); 2632fc2e4d70SOleg Nesterov } 2633fc2e4d70SOleg Nesterov 26344690c4abSTejun Heo /** 26354690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2636112202d9STejun Heo * @pwq: pwq to insert barrier into 26374690c4abSTejun Heo * @barr: wq_barrier to insert 2638affee4b2STejun Heo * @target: target work to attach @barr to 2639affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 26404690c4abSTejun Heo * 2641affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2642affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2643affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2644affee4b2STejun Heo * cpu. 2645affee4b2STejun Heo * 2646affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2647affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2648affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2649affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2650affee4b2STejun Heo * after a work with LINKED flag set. 2651affee4b2STejun Heo * 2652affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2653112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 26544690c4abSTejun Heo * 26554690c4abSTejun Heo * CONTEXT: 2656a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 26574690c4abSTejun Heo */ 2658112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2659affee4b2STejun Heo struct wq_barrier *barr, 2660affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2661fc2e4d70SOleg Nesterov { 2662affee4b2STejun Heo struct list_head *head; 2663affee4b2STejun Heo unsigned int linked = 0; 2664affee4b2STejun Heo 2665dc186ad7SThomas Gleixner /* 2666d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2667dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2668dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2669dc186ad7SThomas Gleixner * might deadlock. 2670dc186ad7SThomas Gleixner */ 2671ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 267222df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 267352fa5bc5SBoqun Feng 2674fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2675fd1a5b04SByungchul Park 26762607d7a6STejun Heo barr->task = current; 267783c22520SOleg Nesterov 2678affee4b2STejun Heo /* 2679affee4b2STejun Heo * If @target is currently being executed, schedule the 2680affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2681affee4b2STejun Heo */ 2682affee4b2STejun Heo if (worker) 2683affee4b2STejun Heo head = worker->scheduled.next; 2684affee4b2STejun Heo else { 2685affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2686affee4b2STejun Heo 2687affee4b2STejun Heo head = target->entry.next; 2688affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2689affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2690affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2691affee4b2STejun Heo } 2692affee4b2STejun Heo 2693dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2694112202d9STejun Heo insert_work(pwq, &barr->work, head, 2695affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2696fc2e4d70SOleg Nesterov } 2697fc2e4d70SOleg Nesterov 269873f53c4aSTejun Heo /** 2699112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 270073f53c4aSTejun Heo * @wq: workqueue being flushed 270173f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 270273f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 270373f53c4aSTejun Heo * 2704112202d9STejun Heo * Prepare pwqs for workqueue flushing. 270573f53c4aSTejun Heo * 2706112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2707112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2708112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2709112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2710112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 271173f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 271273f53c4aSTejun Heo * 271373f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 271473f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 271573f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 271673f53c4aSTejun Heo * is returned. 271773f53c4aSTejun Heo * 2718112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 271973f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 272073f53c4aSTejun Heo * advanced to @work_color. 272173f53c4aSTejun Heo * 272273f53c4aSTejun Heo * CONTEXT: 27233c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 272473f53c4aSTejun Heo * 2725d185af30SYacine Belkadi * Return: 272673f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 272773f53c4aSTejun Heo * otherwise. 272873f53c4aSTejun Heo */ 2729112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 273073f53c4aSTejun Heo int flush_color, int work_color) 27311da177e4SLinus Torvalds { 273273f53c4aSTejun Heo bool wait = false; 273349e3cf44STejun Heo struct pool_workqueue *pwq; 27341da177e4SLinus Torvalds 273573f53c4aSTejun Heo if (flush_color >= 0) { 27366183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2737112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2738dc186ad7SThomas Gleixner } 273914441960SOleg Nesterov 274049e3cf44STejun Heo for_each_pwq(pwq, wq) { 2741112202d9STejun Heo struct worker_pool *pool = pwq->pool; 27421da177e4SLinus Torvalds 2743a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 274473f53c4aSTejun Heo 274573f53c4aSTejun Heo if (flush_color >= 0) { 27466183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 274773f53c4aSTejun Heo 2748112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2749112202d9STejun Heo pwq->flush_color = flush_color; 2750112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 275173f53c4aSTejun Heo wait = true; 27521da177e4SLinus Torvalds } 275373f53c4aSTejun Heo } 275473f53c4aSTejun Heo 275573f53c4aSTejun Heo if (work_color >= 0) { 27566183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2757112202d9STejun Heo pwq->work_color = work_color; 275873f53c4aSTejun Heo } 275973f53c4aSTejun Heo 2760a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27611da177e4SLinus Torvalds } 27621da177e4SLinus Torvalds 2763112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 276473f53c4aSTejun Heo complete(&wq->first_flusher->done); 276573f53c4aSTejun Heo 276673f53c4aSTejun Heo return wait; 276783c22520SOleg Nesterov } 27681da177e4SLinus Torvalds 27690fcb78c2SRolf Eike Beer /** 27701da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 27710fcb78c2SRolf Eike Beer * @wq: workqueue to flush 27721da177e4SLinus Torvalds * 2773c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2774c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 27751da177e4SLinus Torvalds */ 27767ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 27771da177e4SLinus Torvalds { 277873f53c4aSTejun Heo struct wq_flusher this_flusher = { 277973f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 278073f53c4aSTejun Heo .flush_color = -1, 2781fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 278273f53c4aSTejun Heo }; 278373f53c4aSTejun Heo int next_color; 2784b1f4ec17SOleg Nesterov 27853347fa09STejun Heo if (WARN_ON(!wq_online)) 27863347fa09STejun Heo return; 27873347fa09STejun Heo 278887915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 278987915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 279087915adcSJohannes Berg 27913c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 279273f53c4aSTejun Heo 279373f53c4aSTejun Heo /* 279473f53c4aSTejun Heo * Start-to-wait phase 279573f53c4aSTejun Heo */ 279673f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 279773f53c4aSTejun Heo 279873f53c4aSTejun Heo if (next_color != wq->flush_color) { 279973f53c4aSTejun Heo /* 280073f53c4aSTejun Heo * Color space is not full. The current work_color 280173f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 280273f53c4aSTejun Heo * by one. 280373f53c4aSTejun Heo */ 28046183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 280573f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 280673f53c4aSTejun Heo wq->work_color = next_color; 280773f53c4aSTejun Heo 280873f53c4aSTejun Heo if (!wq->first_flusher) { 280973f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 28106183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 281173f53c4aSTejun Heo 281273f53c4aSTejun Heo wq->first_flusher = &this_flusher; 281373f53c4aSTejun Heo 2814112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 281573f53c4aSTejun Heo wq->work_color)) { 281673f53c4aSTejun Heo /* nothing to flush, done */ 281773f53c4aSTejun Heo wq->flush_color = next_color; 281873f53c4aSTejun Heo wq->first_flusher = NULL; 281973f53c4aSTejun Heo goto out_unlock; 282073f53c4aSTejun Heo } 282173f53c4aSTejun Heo } else { 282273f53c4aSTejun Heo /* wait in queue */ 28236183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 282473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2825112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 282673f53c4aSTejun Heo } 282773f53c4aSTejun Heo } else { 282873f53c4aSTejun Heo /* 282973f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 283073f53c4aSTejun Heo * The next flush completion will assign us 283173f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 283273f53c4aSTejun Heo */ 283373f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 283473f53c4aSTejun Heo } 283573f53c4aSTejun Heo 2836fca839c0STejun Heo check_flush_dependency(wq, NULL); 2837fca839c0STejun Heo 28383c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 283973f53c4aSTejun Heo 284073f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 284173f53c4aSTejun Heo 284273f53c4aSTejun Heo /* 284373f53c4aSTejun Heo * Wake-up-and-cascade phase 284473f53c4aSTejun Heo * 284573f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 284673f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 284773f53c4aSTejun Heo */ 284800d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 284973f53c4aSTejun Heo return; 285073f53c4aSTejun Heo 28513c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 285273f53c4aSTejun Heo 28534ce48b37STejun Heo /* we might have raced, check again with mutex held */ 28544ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 28554ce48b37STejun Heo goto out_unlock; 28564ce48b37STejun Heo 285700d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 285873f53c4aSTejun Heo 28596183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 28606183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 286173f53c4aSTejun Heo 286273f53c4aSTejun Heo while (true) { 286373f53c4aSTejun Heo struct wq_flusher *next, *tmp; 286473f53c4aSTejun Heo 286573f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 286673f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 286773f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 286873f53c4aSTejun Heo break; 286973f53c4aSTejun Heo list_del_init(&next->list); 287073f53c4aSTejun Heo complete(&next->done); 287173f53c4aSTejun Heo } 287273f53c4aSTejun Heo 28736183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 287473f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 287573f53c4aSTejun Heo 287673f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 287773f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 287873f53c4aSTejun Heo 287973f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 288073f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 288173f53c4aSTejun Heo /* 288273f53c4aSTejun Heo * Assign the same color to all overflowed 288373f53c4aSTejun Heo * flushers, advance work_color and append to 288473f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 288573f53c4aSTejun Heo * phase for these overflowed flushers. 288673f53c4aSTejun Heo */ 288773f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 288873f53c4aSTejun Heo tmp->flush_color = wq->work_color; 288973f53c4aSTejun Heo 289073f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 289173f53c4aSTejun Heo 289273f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 289373f53c4aSTejun Heo &wq->flusher_queue); 2894112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 289573f53c4aSTejun Heo } 289673f53c4aSTejun Heo 289773f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 28986183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 289973f53c4aSTejun Heo break; 290073f53c4aSTejun Heo } 290173f53c4aSTejun Heo 290273f53c4aSTejun Heo /* 290373f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2904112202d9STejun Heo * the new first flusher and arm pwqs. 290573f53c4aSTejun Heo */ 29066183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 29076183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 290873f53c4aSTejun Heo 290973f53c4aSTejun Heo list_del_init(&next->list); 291073f53c4aSTejun Heo wq->first_flusher = next; 291173f53c4aSTejun Heo 2912112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 291373f53c4aSTejun Heo break; 291473f53c4aSTejun Heo 291573f53c4aSTejun Heo /* 291673f53c4aSTejun Heo * Meh... this color is already done, clear first 291773f53c4aSTejun Heo * flusher and repeat cascading. 291873f53c4aSTejun Heo */ 291973f53c4aSTejun Heo wq->first_flusher = NULL; 292073f53c4aSTejun Heo } 292173f53c4aSTejun Heo 292273f53c4aSTejun Heo out_unlock: 29233c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 29241da177e4SLinus Torvalds } 29251dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 29261da177e4SLinus Torvalds 29279c5a2ba7STejun Heo /** 29289c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 29299c5a2ba7STejun Heo * @wq: workqueue to drain 29309c5a2ba7STejun Heo * 29319c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 29329c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 29339c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2934b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 29359c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 29369c5a2ba7STejun Heo * takes too long. 29379c5a2ba7STejun Heo */ 29389c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 29399c5a2ba7STejun Heo { 29409c5a2ba7STejun Heo unsigned int flush_cnt = 0; 294149e3cf44STejun Heo struct pool_workqueue *pwq; 29429c5a2ba7STejun Heo 29439c5a2ba7STejun Heo /* 29449c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 29459c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2946618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 29479c5a2ba7STejun Heo */ 294887fc741eSLai Jiangshan mutex_lock(&wq->mutex); 29499c5a2ba7STejun Heo if (!wq->nr_drainers++) 2950618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 295187fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29529c5a2ba7STejun Heo reflush: 29539c5a2ba7STejun Heo flush_workqueue(wq); 29549c5a2ba7STejun Heo 2955b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 295676af4d93STejun Heo 295749e3cf44STejun Heo for_each_pwq(pwq, wq) { 2958fa2563e4SThomas Tuttle bool drained; 29599c5a2ba7STejun Heo 2960a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 2961f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 2962a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 2963fa2563e4SThomas Tuttle 2964fa2563e4SThomas Tuttle if (drained) 29659c5a2ba7STejun Heo continue; 29669c5a2ba7STejun Heo 29679c5a2ba7STejun Heo if (++flush_cnt == 10 || 29689c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2969e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 2970e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 297176af4d93STejun Heo 2972b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 29739c5a2ba7STejun Heo goto reflush; 29749c5a2ba7STejun Heo } 29759c5a2ba7STejun Heo 29769c5a2ba7STejun Heo if (!--wq->nr_drainers) 2977618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 297887fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29799c5a2ba7STejun Heo } 29809c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 29819c5a2ba7STejun Heo 2982d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2983d6e89786SJohannes Berg bool from_cancel) 2984baf59022STejun Heo { 2985baf59022STejun Heo struct worker *worker = NULL; 2986c9e7cf27STejun Heo struct worker_pool *pool; 2987112202d9STejun Heo struct pool_workqueue *pwq; 2988baf59022STejun Heo 2989baf59022STejun Heo might_sleep(); 2990baf59022STejun Heo 299124acfb71SThomas Gleixner rcu_read_lock(); 2992fa1b54e6STejun Heo pool = get_work_pool(work); 2993fa1b54e6STejun Heo if (!pool) { 299424acfb71SThomas Gleixner rcu_read_unlock(); 2995fa1b54e6STejun Heo return false; 2996fa1b54e6STejun Heo } 2997fa1b54e6STejun Heo 2998a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 29990b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3000112202d9STejun Heo pwq = get_work_pwq(work); 3001112202d9STejun Heo if (pwq) { 3002112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3003baf59022STejun Heo goto already_gone; 3004606a5020STejun Heo } else { 3005c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3006baf59022STejun Heo if (!worker) 3007baf59022STejun Heo goto already_gone; 3008112202d9STejun Heo pwq = worker->current_pwq; 3009606a5020STejun Heo } 3010baf59022STejun Heo 3011fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3012fca839c0STejun Heo 3013112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3014a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3015baf59022STejun Heo 3016e159489bSTejun Heo /* 3017a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3018a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3019a1d14934SPeter Zijlstra * 3020a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3021a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3022a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3023a1d14934SPeter Zijlstra * forward progress. 3024e159489bSTejun Heo */ 3025d6e89786SJohannes Berg if (!from_cancel && 3026d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3027112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3028112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3029a1d14934SPeter Zijlstra } 303024acfb71SThomas Gleixner rcu_read_unlock(); 3031baf59022STejun Heo return true; 3032baf59022STejun Heo already_gone: 3033a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 303424acfb71SThomas Gleixner rcu_read_unlock(); 3035baf59022STejun Heo return false; 3036baf59022STejun Heo } 3037baf59022STejun Heo 3038d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3039d6e89786SJohannes Berg { 3040d6e89786SJohannes Berg struct wq_barrier barr; 3041d6e89786SJohannes Berg 3042d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3043d6e89786SJohannes Berg return false; 3044d6e89786SJohannes Berg 30454d43d395STetsuo Handa if (WARN_ON(!work->func)) 30464d43d395STetsuo Handa return false; 30474d43d395STetsuo Handa 304887915adcSJohannes Berg if (!from_cancel) { 304987915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 305087915adcSJohannes Berg lock_map_release(&work->lockdep_map); 305187915adcSJohannes Berg } 305287915adcSJohannes Berg 3053d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3054d6e89786SJohannes Berg wait_for_completion(&barr.done); 3055d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3056d6e89786SJohannes Berg return true; 3057d6e89786SJohannes Berg } else { 3058d6e89786SJohannes Berg return false; 3059d6e89786SJohannes Berg } 3060d6e89786SJohannes Berg } 3061d6e89786SJohannes Berg 3062db700897SOleg Nesterov /** 3063401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3064401a8d04STejun Heo * @work: the work to flush 3065db700897SOleg Nesterov * 3066606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3067606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3068401a8d04STejun Heo * 3069d185af30SYacine Belkadi * Return: 3070401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3071401a8d04STejun Heo * %false if it was already idle. 3072db700897SOleg Nesterov */ 3073401a8d04STejun Heo bool flush_work(struct work_struct *work) 3074db700897SOleg Nesterov { 3075d6e89786SJohannes Berg return __flush_work(work, false); 3076606a5020STejun Heo } 3077db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3078db700897SOleg Nesterov 30798603e1b3STejun Heo struct cwt_wait { 3080ac6424b9SIngo Molnar wait_queue_entry_t wait; 30818603e1b3STejun Heo struct work_struct *work; 30828603e1b3STejun Heo }; 30838603e1b3STejun Heo 3084ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 30858603e1b3STejun Heo { 30868603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 30878603e1b3STejun Heo 30888603e1b3STejun Heo if (cwait->work != key) 30898603e1b3STejun Heo return 0; 30908603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 30918603e1b3STejun Heo } 30928603e1b3STejun Heo 309336e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3094401a8d04STejun Heo { 30958603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3096bbb68dfaSTejun Heo unsigned long flags; 30971f1f642eSOleg Nesterov int ret; 30981f1f642eSOleg Nesterov 30991f1f642eSOleg Nesterov do { 3100bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3101bbb68dfaSTejun Heo /* 31028603e1b3STejun Heo * If someone else is already canceling, wait for it to 31038603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 31048603e1b3STejun Heo * because we may get scheduled between @work's completion 31058603e1b3STejun Heo * and the other canceling task resuming and clearing 31068603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 31078603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 31088603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 31098603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 31108603e1b3STejun Heo * we're hogging the CPU. 31118603e1b3STejun Heo * 31128603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 31138603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 31148603e1b3STejun Heo * wake function which matches @work along with exclusive 31158603e1b3STejun Heo * wait and wakeup. 3116bbb68dfaSTejun Heo */ 31178603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31188603e1b3STejun Heo struct cwt_wait cwait; 31198603e1b3STejun Heo 31208603e1b3STejun Heo init_wait(&cwait.wait); 31218603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31228603e1b3STejun Heo cwait.work = work; 31238603e1b3STejun Heo 31248603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31258603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31268603e1b3STejun Heo if (work_is_canceling(work)) 31278603e1b3STejun Heo schedule(); 31288603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31298603e1b3STejun Heo } 31301f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31311f1f642eSOleg Nesterov 3132bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3133bbb68dfaSTejun Heo mark_work_canceling(work); 3134bbb68dfaSTejun Heo local_irq_restore(flags); 3135bbb68dfaSTejun Heo 31363347fa09STejun Heo /* 31373347fa09STejun Heo * This allows canceling during early boot. We know that @work 31383347fa09STejun Heo * isn't executing. 31393347fa09STejun Heo */ 31403347fa09STejun Heo if (wq_online) 3141d6e89786SJohannes Berg __flush_work(work, true); 31423347fa09STejun Heo 31437a22ad75STejun Heo clear_work_data(work); 31448603e1b3STejun Heo 31458603e1b3STejun Heo /* 31468603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31478603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31488603e1b3STejun Heo * visible there. 31498603e1b3STejun Heo */ 31508603e1b3STejun Heo smp_mb(); 31518603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31528603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31538603e1b3STejun Heo 31541f1f642eSOleg Nesterov return ret; 31551f1f642eSOleg Nesterov } 31561f1f642eSOleg Nesterov 31576e84d644SOleg Nesterov /** 3158401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3159401a8d04STejun Heo * @work: the work to cancel 31606e84d644SOleg Nesterov * 3161401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3162401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3163401a8d04STejun Heo * another workqueue. On return from this function, @work is 3164401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31651f1f642eSOleg Nesterov * 3166401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3167401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31686e84d644SOleg Nesterov * 3169401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 31706e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3171401a8d04STejun Heo * 3172d185af30SYacine Belkadi * Return: 3173401a8d04STejun Heo * %true if @work was pending, %false otherwise. 31746e84d644SOleg Nesterov */ 3175401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 31766e84d644SOleg Nesterov { 317736e227d2STejun Heo return __cancel_work_timer(work, false); 3178b89deed3SOleg Nesterov } 317928e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3180b89deed3SOleg Nesterov 31816e84d644SOleg Nesterov /** 3182401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3183401a8d04STejun Heo * @dwork: the delayed work to flush 31846e84d644SOleg Nesterov * 3185401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3186401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3187401a8d04STejun Heo * considers the last queueing instance of @dwork. 31881f1f642eSOleg Nesterov * 3189d185af30SYacine Belkadi * Return: 3190401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3191401a8d04STejun Heo * %false if it was already idle. 31926e84d644SOleg Nesterov */ 3193401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3194401a8d04STejun Heo { 31958930cabaSTejun Heo local_irq_disable(); 3196401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 319760c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 31988930cabaSTejun Heo local_irq_enable(); 3199401a8d04STejun Heo return flush_work(&dwork->work); 3200401a8d04STejun Heo } 3201401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3202401a8d04STejun Heo 320305f0fe6bSTejun Heo /** 320405f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 320505f0fe6bSTejun Heo * @rwork: the rcu work to flush 320605f0fe6bSTejun Heo * 320705f0fe6bSTejun Heo * Return: 320805f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 320905f0fe6bSTejun Heo * %false if it was already idle. 321005f0fe6bSTejun Heo */ 321105f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 321205f0fe6bSTejun Heo { 321305f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 321405f0fe6bSTejun Heo rcu_barrier(); 321505f0fe6bSTejun Heo flush_work(&rwork->work); 321605f0fe6bSTejun Heo return true; 321705f0fe6bSTejun Heo } else { 321805f0fe6bSTejun Heo return flush_work(&rwork->work); 321905f0fe6bSTejun Heo } 322005f0fe6bSTejun Heo } 322105f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 322205f0fe6bSTejun Heo 3223f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3224f72b8792SJens Axboe { 3225f72b8792SJens Axboe unsigned long flags; 3226f72b8792SJens Axboe int ret; 3227f72b8792SJens Axboe 3228f72b8792SJens Axboe do { 3229f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3230f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3231f72b8792SJens Axboe 3232f72b8792SJens Axboe if (unlikely(ret < 0)) 3233f72b8792SJens Axboe return false; 3234f72b8792SJens Axboe 3235f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3236f72b8792SJens Axboe local_irq_restore(flags); 3237f72b8792SJens Axboe return ret; 3238f72b8792SJens Axboe } 3239f72b8792SJens Axboe 3240401a8d04STejun Heo /** 324157b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 324257b30ae7STejun Heo * @dwork: delayed_work to cancel 324309383498STejun Heo * 3244d185af30SYacine Belkadi * Kill off a pending delayed_work. 3245d185af30SYacine Belkadi * 3246d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3247d185af30SYacine Belkadi * pending. 3248d185af30SYacine Belkadi * 3249d185af30SYacine Belkadi * Note: 3250d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3251d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3252d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 325309383498STejun Heo * 325457b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 325509383498STejun Heo */ 325657b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 325709383498STejun Heo { 3258f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 325909383498STejun Heo } 326057b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 326109383498STejun Heo 326209383498STejun Heo /** 3263401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3264401a8d04STejun Heo * @dwork: the delayed work cancel 3265401a8d04STejun Heo * 3266401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3267401a8d04STejun Heo * 3268d185af30SYacine Belkadi * Return: 3269401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3270401a8d04STejun Heo */ 3271401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 32726e84d644SOleg Nesterov { 327336e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 32746e84d644SOleg Nesterov } 3275f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 32761da177e4SLinus Torvalds 32770fcb78c2SRolf Eike Beer /** 327831ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3279b6136773SAndrew Morton * @func: the function to call 3280b6136773SAndrew Morton * 328131ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 328231ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3283b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 328431ddd871STejun Heo * 3285d185af30SYacine Belkadi * Return: 328631ddd871STejun Heo * 0 on success, -errno on failure. 3287b6136773SAndrew Morton */ 328865f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 328915316ba8SChristoph Lameter { 329015316ba8SChristoph Lameter int cpu; 329138f51568SNamhyung Kim struct work_struct __percpu *works; 329215316ba8SChristoph Lameter 3293b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3294b6136773SAndrew Morton if (!works) 329515316ba8SChristoph Lameter return -ENOMEM; 3296b6136773SAndrew Morton 3297ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 329893981800STejun Heo 329915316ba8SChristoph Lameter for_each_online_cpu(cpu) { 33009bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 33019bfb1839SIngo Molnar 33029bfb1839SIngo Molnar INIT_WORK(work, func); 33038de6d308SOleg Nesterov schedule_work_on(cpu, work); 330415316ba8SChristoph Lameter } 330593981800STejun Heo 330693981800STejun Heo for_each_online_cpu(cpu) 33078616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 330893981800STejun Heo 3309ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3310b6136773SAndrew Morton free_percpu(works); 331115316ba8SChristoph Lameter return 0; 331215316ba8SChristoph Lameter } 331315316ba8SChristoph Lameter 3314eef6a7d5SAlan Stern /** 33151fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33161fa44ecaSJames Bottomley * @fn: the function to execute 33171fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33181fa44ecaSJames Bottomley * be available when the work executes) 33191fa44ecaSJames Bottomley * 33201fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33211fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33221fa44ecaSJames Bottomley * 3323d185af30SYacine Belkadi * Return: 0 - function was executed 33241fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33251fa44ecaSJames Bottomley */ 332665f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33271fa44ecaSJames Bottomley { 33281fa44ecaSJames Bottomley if (!in_interrupt()) { 332965f27f38SDavid Howells fn(&ew->work); 33301fa44ecaSJames Bottomley return 0; 33311fa44ecaSJames Bottomley } 33321fa44ecaSJames Bottomley 333365f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33341fa44ecaSJames Bottomley schedule_work(&ew->work); 33351fa44ecaSJames Bottomley 33361fa44ecaSJames Bottomley return 1; 33371fa44ecaSJames Bottomley } 33381fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33391fa44ecaSJames Bottomley 33407a4e344cSTejun Heo /** 33417a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33427a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33437a4e344cSTejun Heo * 33447a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33457a4e344cSTejun Heo */ 3346513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 33477a4e344cSTejun Heo { 33487a4e344cSTejun Heo if (attrs) { 33497a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33507a4e344cSTejun Heo kfree(attrs); 33517a4e344cSTejun Heo } 33527a4e344cSTejun Heo } 33537a4e344cSTejun Heo 33547a4e344cSTejun Heo /** 33557a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33567a4e344cSTejun Heo * 33577a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3358d185af30SYacine Belkadi * return it. 3359d185af30SYacine Belkadi * 3360d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33617a4e344cSTejun Heo */ 3362513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 33637a4e344cSTejun Heo { 33647a4e344cSTejun Heo struct workqueue_attrs *attrs; 33657a4e344cSTejun Heo 3366be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 33677a4e344cSTejun Heo if (!attrs) 33687a4e344cSTejun Heo goto fail; 3369be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 33707a4e344cSTejun Heo goto fail; 33717a4e344cSTejun Heo 337213e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 33737a4e344cSTejun Heo return attrs; 33747a4e344cSTejun Heo fail: 33757a4e344cSTejun Heo free_workqueue_attrs(attrs); 33767a4e344cSTejun Heo return NULL; 33777a4e344cSTejun Heo } 33787a4e344cSTejun Heo 337929c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 338029c91e99STejun Heo const struct workqueue_attrs *from) 338129c91e99STejun Heo { 338229c91e99STejun Heo to->nice = from->nice; 338329c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 33842865a8fbSShaohua Li /* 33852865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 33862865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 33872865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 33882865a8fbSShaohua Li */ 33892865a8fbSShaohua Li to->no_numa = from->no_numa; 339029c91e99STejun Heo } 339129c91e99STejun Heo 339229c91e99STejun Heo /* hash value of the content of @attr */ 339329c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 339429c91e99STejun Heo { 339529c91e99STejun Heo u32 hash = 0; 339629c91e99STejun Heo 339729c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 339813e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 339913e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 340029c91e99STejun Heo return hash; 340129c91e99STejun Heo } 340229c91e99STejun Heo 340329c91e99STejun Heo /* content equality test */ 340429c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 340529c91e99STejun Heo const struct workqueue_attrs *b) 340629c91e99STejun Heo { 340729c91e99STejun Heo if (a->nice != b->nice) 340829c91e99STejun Heo return false; 340929c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 341029c91e99STejun Heo return false; 341129c91e99STejun Heo return true; 341229c91e99STejun Heo } 341329c91e99STejun Heo 34147a4e344cSTejun Heo /** 34157a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34167a4e344cSTejun Heo * @pool: worker_pool to initialize 34177a4e344cSTejun Heo * 3418402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3419d185af30SYacine Belkadi * 3420d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 342129c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 342229c91e99STejun Heo * on @pool safely to release it. 34237a4e344cSTejun Heo */ 34247a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34254e1a1f9aSTejun Heo { 3426a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 342729c91e99STejun Heo pool->id = -1; 342829c91e99STejun Heo pool->cpu = -1; 3429f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34304e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 343182607adcSTejun Heo pool->watchdog_ts = jiffies; 34324e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34334e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34344e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34354e1a1f9aSTejun Heo 343632a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34374e1a1f9aSTejun Heo 343832a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34394e1a1f9aSTejun Heo 3440da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34417a4e344cSTejun Heo 34427cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 344329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 344429c91e99STejun Heo pool->refcnt = 1; 344529c91e99STejun Heo 344629c91e99STejun Heo /* shouldn't fail above this point */ 3447be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 34487a4e344cSTejun Heo if (!pool->attrs) 34497a4e344cSTejun Heo return -ENOMEM; 34507a4e344cSTejun Heo return 0; 34514e1a1f9aSTejun Heo } 34524e1a1f9aSTejun Heo 3453669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3454669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3455669de8bdSBart Van Assche { 3456669de8bdSBart Van Assche char *lock_name; 3457669de8bdSBart Van Assche 3458669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3459669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3460669de8bdSBart Van Assche if (!lock_name) 3461669de8bdSBart Van Assche lock_name = wq->name; 346269a106c0SQian Cai 346369a106c0SQian Cai wq->lock_name = lock_name; 3464669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3465669de8bdSBart Van Assche } 3466669de8bdSBart Van Assche 3467669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3468669de8bdSBart Van Assche { 3469669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3470669de8bdSBart Van Assche } 3471669de8bdSBart Van Assche 3472669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3473669de8bdSBart Van Assche { 3474669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3475669de8bdSBart Van Assche kfree(wq->lock_name); 3476669de8bdSBart Van Assche } 3477669de8bdSBart Van Assche #else 3478669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3479669de8bdSBart Van Assche { 3480669de8bdSBart Van Assche } 3481669de8bdSBart Van Assche 3482669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3483669de8bdSBart Van Assche { 3484669de8bdSBart Van Assche } 3485669de8bdSBart Van Assche 3486669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3487669de8bdSBart Van Assche { 3488669de8bdSBart Van Assche } 3489669de8bdSBart Van Assche #endif 3490669de8bdSBart Van Assche 3491e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3492e2dca7adSTejun Heo { 3493e2dca7adSTejun Heo struct workqueue_struct *wq = 3494e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3495e2dca7adSTejun Heo 3496669de8bdSBart Van Assche wq_free_lockdep(wq); 3497669de8bdSBart Van Assche 3498e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3499e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3500e2dca7adSTejun Heo else 3501e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3502e2dca7adSTejun Heo 3503e2dca7adSTejun Heo kfree(wq); 3504e2dca7adSTejun Heo } 3505e2dca7adSTejun Heo 350629c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 350729c91e99STejun Heo { 350829c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 350929c91e99STejun Heo 35107cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 351129c91e99STejun Heo free_workqueue_attrs(pool->attrs); 351229c91e99STejun Heo kfree(pool); 351329c91e99STejun Heo } 351429c91e99STejun Heo 3515d8bb65abSSebastian Andrzej Siewior /* This returns with the lock held on success (pool manager is inactive). */ 3516d8bb65abSSebastian Andrzej Siewior static bool wq_manager_inactive(struct worker_pool *pool) 3517d8bb65abSSebastian Andrzej Siewior { 3518a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 3519d8bb65abSSebastian Andrzej Siewior 3520d8bb65abSSebastian Andrzej Siewior if (pool->flags & POOL_MANAGER_ACTIVE) { 3521a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3522d8bb65abSSebastian Andrzej Siewior return false; 3523d8bb65abSSebastian Andrzej Siewior } 3524d8bb65abSSebastian Andrzej Siewior return true; 3525d8bb65abSSebastian Andrzej Siewior } 3526d8bb65abSSebastian Andrzej Siewior 352729c91e99STejun Heo /** 352829c91e99STejun Heo * put_unbound_pool - put a worker_pool 352929c91e99STejun Heo * @pool: worker_pool to put 353029c91e99STejun Heo * 353124acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3532c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3533c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3534c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3535a892caccSTejun Heo * 3536a892caccSTejun Heo * Should be called with wq_pool_mutex held. 353729c91e99STejun Heo */ 353829c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 353929c91e99STejun Heo { 354060f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 354129c91e99STejun Heo struct worker *worker; 354229c91e99STejun Heo 3543a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3544a892caccSTejun Heo 3545a892caccSTejun Heo if (--pool->refcnt) 354629c91e99STejun Heo return; 354729c91e99STejun Heo 354829c91e99STejun Heo /* sanity checks */ 354961d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3550a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 355129c91e99STejun Heo return; 355229c91e99STejun Heo 355329c91e99STejun Heo /* release id and unhash */ 355429c91e99STejun Heo if (pool->id >= 0) 355529c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 355629c91e99STejun Heo hash_del(&pool->hash_node); 355729c91e99STejun Heo 3558c5aa87bbSTejun Heo /* 3559692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3560692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3561692b4825STejun Heo * manager and @pool gets freed with the flag set. 3562d8bb65abSSebastian Andrzej Siewior * Because of how wq_manager_inactive() works, we will hold the 3563d8bb65abSSebastian Andrzej Siewior * spinlock after a successful wait. 3564c5aa87bbSTejun Heo */ 3565d8bb65abSSebastian Andrzej Siewior rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), 3566d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3567692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3568692b4825STejun Heo 35691037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 357029c91e99STejun Heo destroy_worker(worker); 357129c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3572a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 357360f5a4bcSLai Jiangshan 35741258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3575da028469SLai Jiangshan if (!list_empty(&pool->workers)) 357660f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 35771258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 357860f5a4bcSLai Jiangshan 357960f5a4bcSLai Jiangshan if (pool->detach_completion) 358060f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 358160f5a4bcSLai Jiangshan 358229c91e99STejun Heo /* shut down the timers */ 358329c91e99STejun Heo del_timer_sync(&pool->idle_timer); 358429c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 358529c91e99STejun Heo 358624acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 358725b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 358829c91e99STejun Heo } 358929c91e99STejun Heo 359029c91e99STejun Heo /** 359129c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 359229c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 359329c91e99STejun Heo * 359429c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 359529c91e99STejun Heo * reference count and return it. If there already is a matching 359629c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3597d185af30SYacine Belkadi * create a new one. 3598a892caccSTejun Heo * 3599a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3600d185af30SYacine Belkadi * 3601d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3602d185af30SYacine Belkadi * On failure, %NULL. 360329c91e99STejun Heo */ 360429c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 360529c91e99STejun Heo { 360629c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 360729c91e99STejun Heo struct worker_pool *pool; 3608f3f90ad4STejun Heo int node; 3609e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 361029c91e99STejun Heo 3611a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 361229c91e99STejun Heo 361329c91e99STejun Heo /* do we already have a matching pool? */ 361429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 361529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 361629c91e99STejun Heo pool->refcnt++; 36173fb1823cSLai Jiangshan return pool; 361829c91e99STejun Heo } 361929c91e99STejun Heo } 362029c91e99STejun Heo 3621e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3622e2273584SXunlei Pang if (wq_numa_enabled) { 3623e2273584SXunlei Pang for_each_node(node) { 3624e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3625e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3626e2273584SXunlei Pang target_node = node; 3627e2273584SXunlei Pang break; 3628e2273584SXunlei Pang } 3629e2273584SXunlei Pang } 3630e2273584SXunlei Pang } 3631e2273584SXunlei Pang 363229c91e99STejun Heo /* nope, create a new one */ 3633e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 363429c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 363529c91e99STejun Heo goto fail; 363629c91e99STejun Heo 36378864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 363829c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3639e2273584SXunlei Pang pool->node = target_node; 364029c91e99STejun Heo 36412865a8fbSShaohua Li /* 36422865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36432865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36442865a8fbSShaohua Li */ 36452865a8fbSShaohua Li pool->attrs->no_numa = false; 36462865a8fbSShaohua Li 364729c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 364829c91e99STejun Heo goto fail; 364929c91e99STejun Heo 365029c91e99STejun Heo /* create and start the initial worker */ 36513347fa09STejun Heo if (wq_online && !create_worker(pool)) 365229c91e99STejun Heo goto fail; 365329c91e99STejun Heo 365429c91e99STejun Heo /* install */ 365529c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36563fb1823cSLai Jiangshan 365729c91e99STejun Heo return pool; 365829c91e99STejun Heo fail: 365929c91e99STejun Heo if (pool) 366029c91e99STejun Heo put_unbound_pool(pool); 366129c91e99STejun Heo return NULL; 366229c91e99STejun Heo } 366329c91e99STejun Heo 36648864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36658864b4e5STejun Heo { 36668864b4e5STejun Heo kmem_cache_free(pwq_cache, 36678864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36688864b4e5STejun Heo } 36698864b4e5STejun Heo 36708864b4e5STejun Heo /* 36718864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 36728864b4e5STejun Heo * and needs to be destroyed. 36738864b4e5STejun Heo */ 36748864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 36758864b4e5STejun Heo { 36768864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 36778864b4e5STejun Heo unbound_release_work); 36788864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 36798864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3680b42b0bddSYang Yingliang bool is_last = false; 36818864b4e5STejun Heo 3682b42b0bddSYang Yingliang /* 3683b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3684b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3685b42b0bddSYang Yingliang */ 3686b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 36878864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 36888864b4e5STejun Heo return; 36898864b4e5STejun Heo 36903c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 36918864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3692bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 36933c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3694b42b0bddSYang Yingliang } 36958864b4e5STejun Heo 3696a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 36978864b4e5STejun Heo put_unbound_pool(pool); 3698a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3699a892caccSTejun Heo 370025b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 37018864b4e5STejun Heo 37028864b4e5STejun Heo /* 37038864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3704e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 37058864b4e5STejun Heo */ 3706669de8bdSBart Van Assche if (is_last) { 3707669de8bdSBart Van Assche wq_unregister_lockdep(wq); 370825b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 37096029a918STejun Heo } 3710669de8bdSBart Van Assche } 37118864b4e5STejun Heo 37120fbd95aaSTejun Heo /** 3713699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37140fbd95aaSTejun Heo * @pwq: target pool_workqueue 37150fbd95aaSTejun Heo * 3716699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3717f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3718699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37190fbd95aaSTejun Heo */ 3720699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37210fbd95aaSTejun Heo { 3722699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3723699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 37243347fa09STejun Heo unsigned long flags; 3725699ce097STejun Heo 3726699ce097STejun Heo /* for @wq->saved_max_active */ 3727a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3728699ce097STejun Heo 3729699ce097STejun Heo /* fast exit for non-freezable wqs */ 3730699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3731699ce097STejun Heo return; 3732699ce097STejun Heo 37333347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3734a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3735699ce097STejun Heo 373674b414eaSLai Jiangshan /* 373774b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 373874b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 373974b414eaSLai Jiangshan * is updated and visible. 374074b414eaSLai Jiangshan */ 374174b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 374201341fbdSYunfeng Ye bool kick = false; 374301341fbdSYunfeng Ye 3744699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37450fbd95aaSTejun Heo 3746f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 374701341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3748f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 374901341fbdSYunfeng Ye kick = true; 375001341fbdSYunfeng Ye } 3751951a078aSLai Jiangshan 3752951a078aSLai Jiangshan /* 3753951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 375401341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 375501341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 375601341fbdSYunfeng Ye * let's kick iff work items were activated. 3757951a078aSLai Jiangshan */ 375801341fbdSYunfeng Ye if (kick) 3759951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3760699ce097STejun Heo } else { 3761699ce097STejun Heo pwq->max_active = 0; 3762699ce097STejun Heo } 3763699ce097STejun Heo 3764a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 37650fbd95aaSTejun Heo } 37660fbd95aaSTejun Heo 376767dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3768f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3769f147f29eSTejun Heo struct worker_pool *pool) 3770d2c1d404STejun Heo { 3771d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3772d2c1d404STejun Heo 3773e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3774e50aba9aSTejun Heo 3775d2c1d404STejun Heo pwq->pool = pool; 3776d2c1d404STejun Heo pwq->wq = wq; 3777d2c1d404STejun Heo pwq->flush_color = -1; 37788864b4e5STejun Heo pwq->refcnt = 1; 3779f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 37801befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3781d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 37828864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3783f147f29eSTejun Heo } 3784d2c1d404STejun Heo 3785f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 37861befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3787f147f29eSTejun Heo { 3788f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3789f147f29eSTejun Heo 3790f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 379175ccf595STejun Heo 37921befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 37931befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 37941befcf30STejun Heo return; 37951befcf30STejun Heo 379629b1cb41SLai Jiangshan /* set the matching work_color */ 379775ccf595STejun Heo pwq->work_color = wq->work_color; 3798983ca25eSTejun Heo 3799983ca25eSTejun Heo /* sync max_active to the current setting */ 3800983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3801983ca25eSTejun Heo 3802983ca25eSTejun Heo /* link in @pwq */ 38039e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3804df2d5ae4STejun Heo } 38056029a918STejun Heo 3806f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3807f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3808f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3809f147f29eSTejun Heo { 3810f147f29eSTejun Heo struct worker_pool *pool; 3811f147f29eSTejun Heo struct pool_workqueue *pwq; 3812f147f29eSTejun Heo 3813f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3814f147f29eSTejun Heo 3815f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3816f147f29eSTejun Heo if (!pool) 3817f147f29eSTejun Heo return NULL; 3818f147f29eSTejun Heo 3819e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3820f147f29eSTejun Heo if (!pwq) { 3821f147f29eSTejun Heo put_unbound_pool(pool); 3822f147f29eSTejun Heo return NULL; 3823f147f29eSTejun Heo } 3824f147f29eSTejun Heo 3825f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3826f147f29eSTejun Heo return pwq; 3827d2c1d404STejun Heo } 3828d2c1d404STejun Heo 38294c16bd32STejun Heo /** 383030186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3831042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 38324c16bd32STejun Heo * @node: the target NUMA node 38334c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38344c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38354c16bd32STejun Heo * 38364c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38374c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3838d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38394c16bd32STejun Heo * 38404c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38414c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38424c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38434c16bd32STejun Heo * @attrs->cpumask. 38444c16bd32STejun Heo * 38454c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38464c16bd32STejun Heo * stable. 3847d185af30SYacine Belkadi * 3848d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3849d185af30SYacine Belkadi * %false if equal. 38504c16bd32STejun Heo */ 38514c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38524c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38534c16bd32STejun Heo { 3854d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38554c16bd32STejun Heo goto use_dfl; 38564c16bd32STejun Heo 38574c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38584c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38594c16bd32STejun Heo if (cpu_going_down >= 0) 38604c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38614c16bd32STejun Heo 38624c16bd32STejun Heo if (cpumask_empty(cpumask)) 38634c16bd32STejun Heo goto use_dfl; 38644c16bd32STejun Heo 38654c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38664c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38671ad0f0a7SMichael Bringmann 38681ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 38691ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 38701ad0f0a7SMichael Bringmann "possible intersect\n"); 38711ad0f0a7SMichael Bringmann return false; 38721ad0f0a7SMichael Bringmann } 38731ad0f0a7SMichael Bringmann 38744c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 38754c16bd32STejun Heo 38764c16bd32STejun Heo use_dfl: 38774c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 38784c16bd32STejun Heo return false; 38794c16bd32STejun Heo } 38804c16bd32STejun Heo 38811befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 38821befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 38831befcf30STejun Heo int node, 38841befcf30STejun Heo struct pool_workqueue *pwq) 38851befcf30STejun Heo { 38861befcf30STejun Heo struct pool_workqueue *old_pwq; 38871befcf30STejun Heo 38885b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 38891befcf30STejun Heo lockdep_assert_held(&wq->mutex); 38901befcf30STejun Heo 38911befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 38921befcf30STejun Heo link_pwq(pwq); 38931befcf30STejun Heo 38941befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 38951befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 38961befcf30STejun Heo return old_pwq; 38971befcf30STejun Heo } 38981befcf30STejun Heo 38992d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 39002d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 39012d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 39022d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3903042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 39042d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 39052d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 39062d5f0764SLai Jiangshan }; 39072d5f0764SLai Jiangshan 39082d5f0764SLai Jiangshan /* free the resources after success or abort */ 39092d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 39102d5f0764SLai Jiangshan { 39112d5f0764SLai Jiangshan if (ctx) { 39122d5f0764SLai Jiangshan int node; 39132d5f0764SLai Jiangshan 39142d5f0764SLai Jiangshan for_each_node(node) 39152d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 39162d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 39172d5f0764SLai Jiangshan 39182d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 39192d5f0764SLai Jiangshan 39202d5f0764SLai Jiangshan kfree(ctx); 39212d5f0764SLai Jiangshan } 39222d5f0764SLai Jiangshan } 39232d5f0764SLai Jiangshan 39242d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 39252d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 39262d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 39272d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 39282d5f0764SLai Jiangshan { 39292d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 39302d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 39312d5f0764SLai Jiangshan int node; 39322d5f0764SLai Jiangshan 39332d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39342d5f0764SLai Jiangshan 3935acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 39362d5f0764SLai Jiangshan 3937be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 3938be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 39392d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39402d5f0764SLai Jiangshan goto out_free; 39412d5f0764SLai Jiangshan 3942042f7df1SLai Jiangshan /* 3943042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3944042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3945042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3946042f7df1SLai Jiangshan */ 39472d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3948b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3949042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3950042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39512d5f0764SLai Jiangshan 39522d5f0764SLai Jiangshan /* 39532d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39542d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39552d5f0764SLai Jiangshan * pools. 39562d5f0764SLai Jiangshan */ 39572d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39582d5f0764SLai Jiangshan 39592d5f0764SLai Jiangshan /* 39602d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39612d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39622d5f0764SLai Jiangshan * it even if we don't use it immediately. 39632d5f0764SLai Jiangshan */ 39642d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39652d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39662d5f0764SLai Jiangshan goto out_free; 39672d5f0764SLai Jiangshan 39682d5f0764SLai Jiangshan for_each_node(node) { 3969042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 39702d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 39712d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 39722d5f0764SLai Jiangshan goto out_free; 39732d5f0764SLai Jiangshan } else { 39742d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 39752d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 39762d5f0764SLai Jiangshan } 39772d5f0764SLai Jiangshan } 39782d5f0764SLai Jiangshan 3979042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3980042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3981042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 39822d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3983042f7df1SLai Jiangshan 39842d5f0764SLai Jiangshan ctx->wq = wq; 39852d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 39862d5f0764SLai Jiangshan return ctx; 39872d5f0764SLai Jiangshan 39882d5f0764SLai Jiangshan out_free: 39892d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 39902d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 39912d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 39922d5f0764SLai Jiangshan return NULL; 39932d5f0764SLai Jiangshan } 39942d5f0764SLai Jiangshan 39952d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 39962d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 39972d5f0764SLai Jiangshan { 39982d5f0764SLai Jiangshan int node; 39992d5f0764SLai Jiangshan 40002d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 40012d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 40022d5f0764SLai Jiangshan 40032d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 40042d5f0764SLai Jiangshan 40052d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 40062d5f0764SLai Jiangshan for_each_node(node) 40072d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 40082d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 40092d5f0764SLai Jiangshan 40102d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 40112d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 40122d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 40132d5f0764SLai Jiangshan 40142d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 40152d5f0764SLai Jiangshan } 40162d5f0764SLai Jiangshan 4017a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4018a0111cf6SLai Jiangshan { 4019a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4020ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4021a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4022a0111cf6SLai Jiangshan } 4023a0111cf6SLai Jiangshan 4024a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4025a0111cf6SLai Jiangshan { 4026a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4027ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4028a0111cf6SLai Jiangshan } 4029a0111cf6SLai Jiangshan 4030a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4031a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4032a0111cf6SLai Jiangshan { 4033a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4034a0111cf6SLai Jiangshan 4035a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4036a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4037a0111cf6SLai Jiangshan return -EINVAL; 4038a0111cf6SLai Jiangshan 4039a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40400a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40410a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4042a0111cf6SLai Jiangshan return -EINVAL; 4043a0111cf6SLai Jiangshan 40440a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40450a94efb5STejun Heo } 40460a94efb5STejun Heo 4047a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40486201171eSwanghaibin if (!ctx) 40496201171eSwanghaibin return -ENOMEM; 4050a0111cf6SLai Jiangshan 4051a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4052a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4053a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4054a0111cf6SLai Jiangshan 40556201171eSwanghaibin return 0; 4056a0111cf6SLai Jiangshan } 4057a0111cf6SLai Jiangshan 40589e8cd2f5STejun Heo /** 40599e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40609e8cd2f5STejun Heo * @wq: the target workqueue 40619e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40629e8cd2f5STejun Heo * 40634c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40644c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40654c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40664c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40674c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 40684c16bd32STejun Heo * back-to-back will stay on its current pwq. 40699e8cd2f5STejun Heo * 4070d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4071d185af30SYacine Belkadi * 4072ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4073509b3204SDaniel Jordan * 4074d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 40759e8cd2f5STejun Heo */ 4076513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 40779e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 40789e8cd2f5STejun Heo { 4079a0111cf6SLai Jiangshan int ret; 40809e8cd2f5STejun Heo 4081509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4082509b3204SDaniel Jordan 4083509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4084a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4085509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 40862d5f0764SLai Jiangshan 40872d5f0764SLai Jiangshan return ret; 40889e8cd2f5STejun Heo } 40899e8cd2f5STejun Heo 40904c16bd32STejun Heo /** 40914c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 40924c16bd32STejun Heo * @wq: the target workqueue 40934c16bd32STejun Heo * @cpu: the CPU coming up or going down 40944c16bd32STejun Heo * @online: whether @cpu is coming up or going down 40954c16bd32STejun Heo * 40964c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 40974c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 40984c16bd32STejun Heo * @wq accordingly. 40994c16bd32STejun Heo * 41004c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 41014c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 41024c16bd32STejun Heo * correct. 41034c16bd32STejun Heo * 41044c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 41054c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 41064c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 41074c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 41084c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 41094c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 41104c16bd32STejun Heo * CPU_DOWN_PREPARE. 41114c16bd32STejun Heo */ 41124c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 41134c16bd32STejun Heo bool online) 41144c16bd32STejun Heo { 41154c16bd32STejun Heo int node = cpu_to_node(cpu); 41164c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 41174c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 41184c16bd32STejun Heo struct workqueue_attrs *target_attrs; 41194c16bd32STejun Heo cpumask_t *cpumask; 41204c16bd32STejun Heo 41214c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 41224c16bd32STejun Heo 4123f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4124f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 41254c16bd32STejun Heo return; 41264c16bd32STejun Heo 41274c16bd32STejun Heo /* 41284c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 41294c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 41304c16bd32STejun Heo * CPU hotplug exclusion. 41314c16bd32STejun Heo */ 41324c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 41334c16bd32STejun Heo cpumask = target_attrs->cpumask; 41344c16bd32STejun Heo 41354c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 41364c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 41374c16bd32STejun Heo 41384c16bd32STejun Heo /* 41394c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4140042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4141042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4142042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41434c16bd32STejun Heo */ 4144042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41454c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4146f7142ed4SLai Jiangshan return; 41474c16bd32STejun Heo } else { 41484c16bd32STejun Heo goto use_dfl_pwq; 41494c16bd32STejun Heo } 41504c16bd32STejun Heo 41514c16bd32STejun Heo /* create a new pwq */ 41524c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41534c16bd32STejun Heo if (!pwq) { 41542d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41554c16bd32STejun Heo wq->name); 415677f300b1SDaeseok Youn goto use_dfl_pwq; 41574c16bd32STejun Heo } 41584c16bd32STejun Heo 4159f7142ed4SLai Jiangshan /* Install the new pwq. */ 41604c16bd32STejun Heo mutex_lock(&wq->mutex); 41614c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41624c16bd32STejun Heo goto out_unlock; 41634c16bd32STejun Heo 41644c16bd32STejun Heo use_dfl_pwq: 4165f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4166a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 41674c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4168a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 41694c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 41704c16bd32STejun Heo out_unlock: 41714c16bd32STejun Heo mutex_unlock(&wq->mutex); 41724c16bd32STejun Heo put_pwq_unlocked(old_pwq); 41734c16bd32STejun Heo } 41744c16bd32STejun Heo 417530cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 41761da177e4SLinus Torvalds { 417749e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 41788a2b7538STejun Heo int cpu, ret; 4179e1d8aa9fSFrederic Weisbecker 418030cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4181420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4182420c0ddbSTejun Heo if (!wq->cpu_pwqs) 418330cdf249STejun Heo return -ENOMEM; 418430cdf249STejun Heo 418530cdf249STejun Heo for_each_possible_cpu(cpu) { 41867fb98ea7STejun Heo struct pool_workqueue *pwq = 41877fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 41887a62c2c8STejun Heo struct worker_pool *cpu_pools = 4189f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 419030cdf249STejun Heo 4191f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4192f147f29eSTejun Heo 4193f147f29eSTejun Heo mutex_lock(&wq->mutex); 41941befcf30STejun Heo link_pwq(pwq); 4195f147f29eSTejun Heo mutex_unlock(&wq->mutex); 419630cdf249STejun Heo } 419730cdf249STejun Heo return 0; 4198509b3204SDaniel Jordan } 4199509b3204SDaniel Jordan 4200ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4201509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 42028a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 42038a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 42048a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 42058a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 42068a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 42079e8cd2f5STejun Heo } else { 4208509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 42099e8cd2f5STejun Heo } 4210ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4211509b3204SDaniel Jordan 4212509b3204SDaniel Jordan return ret; 42130f900049STejun Heo } 42140f900049STejun Heo 4215f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4216f3421797STejun Heo const char *name) 4217b71ab8c2STejun Heo { 4218f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4219f3421797STejun Heo 4220f3421797STejun Heo if (max_active < 1 || max_active > lim) 4221044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4222f3421797STejun Heo max_active, name, 1, lim); 4223b71ab8c2STejun Heo 4224f3421797STejun Heo return clamp_val(max_active, 1, lim); 4225b71ab8c2STejun Heo } 4226b71ab8c2STejun Heo 4227983c7515STejun Heo /* 4228983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4229983c7515STejun Heo * to guarantee forward progress. 4230983c7515STejun Heo */ 4231983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4232983c7515STejun Heo { 4233983c7515STejun Heo struct worker *rescuer; 4234b92b36eaSDan Carpenter int ret; 4235983c7515STejun Heo 4236983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4237983c7515STejun Heo return 0; 4238983c7515STejun Heo 4239983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4240983c7515STejun Heo if (!rescuer) 4241983c7515STejun Heo return -ENOMEM; 4242983c7515STejun Heo 4243983c7515STejun Heo rescuer->rescue_wq = wq; 4244983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4245f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4246b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4247983c7515STejun Heo kfree(rescuer); 4248b92b36eaSDan Carpenter return ret; 4249983c7515STejun Heo } 4250983c7515STejun Heo 4251983c7515STejun Heo wq->rescuer = rescuer; 4252983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4253983c7515STejun Heo wake_up_process(rescuer->task); 4254983c7515STejun Heo 4255983c7515STejun Heo return 0; 4256983c7515STejun Heo } 4257983c7515STejun Heo 4258a2775bbcSMathieu Malaterre __printf(1, 4) 4259669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 426097e37d7bSTejun Heo unsigned int flags, 4261669de8bdSBart Van Assche int max_active, ...) 42623af24433SOleg Nesterov { 4263df2d5ae4STejun Heo size_t tbl_size = 0; 4264ecf6881fSTejun Heo va_list args; 42653af24433SOleg Nesterov struct workqueue_struct *wq; 426649e3cf44STejun Heo struct pool_workqueue *pwq; 4267b196be89STejun Heo 42685c0338c6STejun Heo /* 42695c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 42705c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 42715c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 42725c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 42735c0338c6STejun Heo * on NUMA. 42745c0338c6STejun Heo */ 42755c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 42765c0338c6STejun Heo flags |= __WQ_ORDERED; 42775c0338c6STejun Heo 4278cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4279cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4280cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4281cee22a15SViresh Kumar 4282ecf6881fSTejun Heo /* allocate wq and format name */ 4283df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4284ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4285df2d5ae4STejun Heo 4286df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4287b196be89STejun Heo if (!wq) 4288d2c1d404STejun Heo return NULL; 4289b196be89STejun Heo 42906029a918STejun Heo if (flags & WQ_UNBOUND) { 4291be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 42926029a918STejun Heo if (!wq->unbound_attrs) 42936029a918STejun Heo goto err_free_wq; 42946029a918STejun Heo } 42956029a918STejun Heo 4296669de8bdSBart Van Assche va_start(args, max_active); 4297ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4298b196be89STejun Heo va_end(args); 42993af24433SOleg Nesterov 4300d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4301b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 43023af24433SOleg Nesterov 4303b196be89STejun Heo /* init wq */ 430497e37d7bSTejun Heo wq->flags = flags; 4305a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 43063c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4307112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 430830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 430973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 431073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4311493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 43123af24433SOleg Nesterov 4313669de8bdSBart Van Assche wq_init_lockdep(wq); 4314cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 43153af24433SOleg Nesterov 431630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 431782efcab3SBart Van Assche goto err_unreg_lockdep; 43181537663fSTejun Heo 431940c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4320d2c1d404STejun Heo goto err_destroy; 4321e22bee78STejun Heo 4322226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4323226223abSTejun Heo goto err_destroy; 4324226223abSTejun Heo 43256af8bf3dSOleg Nesterov /* 432668e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 432768e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 432868e13a67SLai Jiangshan * list. 43296af8bf3dSOleg Nesterov */ 433068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4331a0a1a5fdSTejun Heo 4332a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 433349e3cf44STejun Heo for_each_pwq(pwq, wq) 4334699ce097STejun Heo pwq_adjust_max_active(pwq); 4335a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4336a0a1a5fdSTejun Heo 4337e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4338a0a1a5fdSTejun Heo 433968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43403af24433SOleg Nesterov 43413af24433SOleg Nesterov return wq; 4342d2c1d404STejun Heo 434382efcab3SBart Van Assche err_unreg_lockdep: 4344009bb421SBart Van Assche wq_unregister_lockdep(wq); 4345009bb421SBart Van Assche wq_free_lockdep(wq); 434682efcab3SBart Van Assche err_free_wq: 43476029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43484690c4abSTejun Heo kfree(wq); 4349d2c1d404STejun Heo return NULL; 4350d2c1d404STejun Heo err_destroy: 4351d2c1d404STejun Heo destroy_workqueue(wq); 43524690c4abSTejun Heo return NULL; 43531da177e4SLinus Torvalds } 4354669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43551da177e4SLinus Torvalds 4356c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4357c29eb853STejun Heo { 4358c29eb853STejun Heo int i; 4359c29eb853STejun Heo 4360c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4361c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4362c29eb853STejun Heo return true; 4363c29eb853STejun Heo 4364c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4365c29eb853STejun Heo return true; 4366f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4367c29eb853STejun Heo return true; 4368c29eb853STejun Heo 4369c29eb853STejun Heo return false; 4370c29eb853STejun Heo } 4371c29eb853STejun Heo 43723af24433SOleg Nesterov /** 43733af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 43743af24433SOleg Nesterov * @wq: target workqueue 43753af24433SOleg Nesterov * 43763af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 43773af24433SOleg Nesterov */ 43783af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 43793af24433SOleg Nesterov { 438049e3cf44STejun Heo struct pool_workqueue *pwq; 43814c16bd32STejun Heo int node; 43823af24433SOleg Nesterov 4383def98c84STejun Heo /* 4384def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4385def98c84STejun Heo * lead to sysfs name conflicts. 4386def98c84STejun Heo */ 4387def98c84STejun Heo workqueue_sysfs_unregister(wq); 4388def98c84STejun Heo 43899c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 43909c5a2ba7STejun Heo drain_workqueue(wq); 4391c8efcc25STejun Heo 4392def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4393def98c84STejun Heo if (wq->rescuer) { 4394def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4395def98c84STejun Heo 4396def98c84STejun Heo /* this prevents new queueing */ 4397a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4398def98c84STejun Heo wq->rescuer = NULL; 4399a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4400def98c84STejun Heo 4401def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4402def98c84STejun Heo kthread_stop(rescuer->task); 44038efe1223STejun Heo kfree(rescuer); 4404def98c84STejun Heo } 4405def98c84STejun Heo 4406c29eb853STejun Heo /* 4407c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4408c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4409c29eb853STejun Heo */ 4410c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4411b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 441249e3cf44STejun Heo for_each_pwq(pwq, wq) { 4413a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4414c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 44151d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4416e66b39afSTejun Heo __func__, wq->name); 4417c29eb853STejun Heo show_pwq(pwq); 4418a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4419b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4420c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 4421fa07fb6aSTejun Heo show_workqueue_state(); 44226183c009STejun Heo return; 442376af4d93STejun Heo } 4424a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 442576af4d93STejun Heo } 4426b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 44276183c009STejun Heo 4428a0a1a5fdSTejun Heo /* 4429a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4430a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4431a0a1a5fdSTejun Heo */ 4432e2dca7adSTejun Heo list_del_rcu(&wq->list); 443368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44343af24433SOleg Nesterov 44358864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4436669de8bdSBart Van Assche wq_unregister_lockdep(wq); 443729c91e99STejun Heo /* 44388864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4439e2dca7adSTejun Heo * schedule RCU free. 444029c91e99STejun Heo */ 444125b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 44428864b4e5STejun Heo } else { 44438864b4e5STejun Heo /* 44448864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 44454c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 44464c16bd32STejun Heo * @wq will be freed when the last pwq is released. 44478864b4e5STejun Heo */ 44484c16bd32STejun Heo for_each_node(node) { 44494c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 44504c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 44514c16bd32STejun Heo put_pwq_unlocked(pwq); 44524c16bd32STejun Heo } 44534c16bd32STejun Heo 44544c16bd32STejun Heo /* 44554c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 44564c16bd32STejun Heo * put. Don't access it afterwards. 44574c16bd32STejun Heo */ 44584c16bd32STejun Heo pwq = wq->dfl_pwq; 44594c16bd32STejun Heo wq->dfl_pwq = NULL; 4460dce90d47STejun Heo put_pwq_unlocked(pwq); 446129c91e99STejun Heo } 44623af24433SOleg Nesterov } 44633af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 44643af24433SOleg Nesterov 4465dcd989cbSTejun Heo /** 4466dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4467dcd989cbSTejun Heo * @wq: target workqueue 4468dcd989cbSTejun Heo * @max_active: new max_active value. 4469dcd989cbSTejun Heo * 4470dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4471dcd989cbSTejun Heo * 4472dcd989cbSTejun Heo * CONTEXT: 4473dcd989cbSTejun Heo * Don't call from IRQ context. 4474dcd989cbSTejun Heo */ 4475dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4476dcd989cbSTejun Heo { 447749e3cf44STejun Heo struct pool_workqueue *pwq; 4478dcd989cbSTejun Heo 44798719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 44800a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 44818719dceaSTejun Heo return; 44828719dceaSTejun Heo 4483f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4484dcd989cbSTejun Heo 4485a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4486dcd989cbSTejun Heo 44870a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4488dcd989cbSTejun Heo wq->saved_max_active = max_active; 4489dcd989cbSTejun Heo 4490699ce097STejun Heo for_each_pwq(pwq, wq) 4491699ce097STejun Heo pwq_adjust_max_active(pwq); 4492dcd989cbSTejun Heo 4493a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4494dcd989cbSTejun Heo } 4495dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4496dcd989cbSTejun Heo 4497dcd989cbSTejun Heo /** 449827d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 449927d4ee03SLukas Wunner * 450027d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 450127d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 450227d4ee03SLukas Wunner * 450327d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 450427d4ee03SLukas Wunner */ 450527d4ee03SLukas Wunner struct work_struct *current_work(void) 450627d4ee03SLukas Wunner { 450727d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 450827d4ee03SLukas Wunner 450927d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 451027d4ee03SLukas Wunner } 451127d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 451227d4ee03SLukas Wunner 451327d4ee03SLukas Wunner /** 4514e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4515e6267616STejun Heo * 4516e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4517e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4518d185af30SYacine Belkadi * 4519d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4520e6267616STejun Heo */ 4521e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4522e6267616STejun Heo { 4523e6267616STejun Heo struct worker *worker = current_wq_worker(); 4524e6267616STejun Heo 45256a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4526e6267616STejun Heo } 4527e6267616STejun Heo 4528e6267616STejun Heo /** 4529dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4530dcd989cbSTejun Heo * @cpu: CPU in question 4531dcd989cbSTejun Heo * @wq: target workqueue 4532dcd989cbSTejun Heo * 4533dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4534dcd989cbSTejun Heo * no synchronization around this function and the test result is 4535dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4536dcd989cbSTejun Heo * 4537d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4538d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4539d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4540d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4541d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4542d3251859STejun Heo * 4543d185af30SYacine Belkadi * Return: 4544dcd989cbSTejun Heo * %true if congested, %false otherwise. 4545dcd989cbSTejun Heo */ 4546d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4547dcd989cbSTejun Heo { 45487fb98ea7STejun Heo struct pool_workqueue *pwq; 454976af4d93STejun Heo bool ret; 455076af4d93STejun Heo 455124acfb71SThomas Gleixner rcu_read_lock(); 455224acfb71SThomas Gleixner preempt_disable(); 45537fb98ea7STejun Heo 4554d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4555d3251859STejun Heo cpu = smp_processor_id(); 4556d3251859STejun Heo 45577fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 45587fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 45597fb98ea7STejun Heo else 4560df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4561dcd989cbSTejun Heo 4562f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 456324acfb71SThomas Gleixner preempt_enable(); 456424acfb71SThomas Gleixner rcu_read_unlock(); 456576af4d93STejun Heo 456676af4d93STejun Heo return ret; 4567dcd989cbSTejun Heo } 4568dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4569dcd989cbSTejun Heo 4570dcd989cbSTejun Heo /** 4571dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4572dcd989cbSTejun Heo * @work: the work to be tested 4573dcd989cbSTejun Heo * 4574dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4575dcd989cbSTejun Heo * synchronization around this function and the test result is 4576dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4577dcd989cbSTejun Heo * 4578d185af30SYacine Belkadi * Return: 4579dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4580dcd989cbSTejun Heo */ 4581dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4582dcd989cbSTejun Heo { 4583fa1b54e6STejun Heo struct worker_pool *pool; 4584dcd989cbSTejun Heo unsigned long flags; 4585dcd989cbSTejun Heo unsigned int ret = 0; 4586dcd989cbSTejun Heo 4587dcd989cbSTejun Heo if (work_pending(work)) 4588dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4589038366c5SLai Jiangshan 459024acfb71SThomas Gleixner rcu_read_lock(); 4591fa1b54e6STejun Heo pool = get_work_pool(work); 4592038366c5SLai Jiangshan if (pool) { 4593a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4594c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4595dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4596a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4597038366c5SLai Jiangshan } 459824acfb71SThomas Gleixner rcu_read_unlock(); 4599dcd989cbSTejun Heo 4600dcd989cbSTejun Heo return ret; 4601dcd989cbSTejun Heo } 4602dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4603dcd989cbSTejun Heo 46043d1cb205STejun Heo /** 46053d1cb205STejun Heo * set_worker_desc - set description for the current work item 46063d1cb205STejun Heo * @fmt: printf-style format string 46073d1cb205STejun Heo * @...: arguments for the format string 46083d1cb205STejun Heo * 46093d1cb205STejun Heo * This function can be called by a running work function to describe what 46103d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 46113d1cb205STejun Heo * information will be printed out together to help debugging. The 46123d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 46133d1cb205STejun Heo */ 46143d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 46153d1cb205STejun Heo { 46163d1cb205STejun Heo struct worker *worker = current_wq_worker(); 46173d1cb205STejun Heo va_list args; 46183d1cb205STejun Heo 46193d1cb205STejun Heo if (worker) { 46203d1cb205STejun Heo va_start(args, fmt); 46213d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 46223d1cb205STejun Heo va_end(args); 46233d1cb205STejun Heo } 46243d1cb205STejun Heo } 46255c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 46263d1cb205STejun Heo 46273d1cb205STejun Heo /** 46283d1cb205STejun Heo * print_worker_info - print out worker information and description 46293d1cb205STejun Heo * @log_lvl: the log level to use when printing 46303d1cb205STejun Heo * @task: target task 46313d1cb205STejun Heo * 46323d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 46333d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 46343d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 46353d1cb205STejun Heo * 46363d1cb205STejun Heo * This function can be safely called on any task as long as the 46373d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 46383d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 46393d1cb205STejun Heo */ 46403d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 46413d1cb205STejun Heo { 46423d1cb205STejun Heo work_func_t *fn = NULL; 46433d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 46443d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 46453d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 46463d1cb205STejun Heo struct workqueue_struct *wq = NULL; 46473d1cb205STejun Heo struct worker *worker; 46483d1cb205STejun Heo 46493d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 46503d1cb205STejun Heo return; 46513d1cb205STejun Heo 46523d1cb205STejun Heo /* 46533d1cb205STejun Heo * This function is called without any synchronization and @task 46543d1cb205STejun Heo * could be in any state. Be careful with dereferences. 46553d1cb205STejun Heo */ 4656e700591aSPetr Mladek worker = kthread_probe_data(task); 46573d1cb205STejun Heo 46583d1cb205STejun Heo /* 46598bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 46608bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 46613d1cb205STejun Heo */ 4662fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4663fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4664fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4665fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4666fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 46673d1cb205STejun Heo 46683d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4669d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 46708bf89593STejun Heo if (strcmp(name, desc)) 46713d1cb205STejun Heo pr_cont(" (%s)", desc); 46723d1cb205STejun Heo pr_cont("\n"); 46733d1cb205STejun Heo } 46743d1cb205STejun Heo } 46753d1cb205STejun Heo 46763494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 46773494fc30STejun Heo { 46783494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 46793494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 46803494fc30STejun Heo pr_cont(" node=%d", pool->node); 46813494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 46823494fc30STejun Heo } 46833494fc30STejun Heo 46843494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 46853494fc30STejun Heo { 46863494fc30STejun Heo if (work->func == wq_barrier_func) { 46873494fc30STejun Heo struct wq_barrier *barr; 46883494fc30STejun Heo 46893494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 46903494fc30STejun Heo 46913494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 46923494fc30STejun Heo task_pid_nr(barr->task)); 46933494fc30STejun Heo } else { 4694d75f773cSSakari Ailus pr_cont("%s %ps", comma ? "," : "", work->func); 46953494fc30STejun Heo } 46963494fc30STejun Heo } 46973494fc30STejun Heo 46983494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 46993494fc30STejun Heo { 47003494fc30STejun Heo struct worker_pool *pool = pwq->pool; 47013494fc30STejun Heo struct work_struct *work; 47023494fc30STejun Heo struct worker *worker; 47033494fc30STejun Heo bool has_in_flight = false, has_pending = false; 47043494fc30STejun Heo int bkt; 47053494fc30STejun Heo 47063494fc30STejun Heo pr_info(" pwq %d:", pool->id); 47073494fc30STejun Heo pr_cont_pool_info(pool); 47083494fc30STejun Heo 4709e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4710e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 47113494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 47123494fc30STejun Heo 47133494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47143494fc30STejun Heo if (worker->current_pwq == pwq) { 47153494fc30STejun Heo has_in_flight = true; 47163494fc30STejun Heo break; 47173494fc30STejun Heo } 47183494fc30STejun Heo } 47193494fc30STejun Heo if (has_in_flight) { 47203494fc30STejun Heo bool comma = false; 47213494fc30STejun Heo 47223494fc30STejun Heo pr_info(" in-flight:"); 47233494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47243494fc30STejun Heo if (worker->current_pwq != pwq) 47253494fc30STejun Heo continue; 47263494fc30STejun Heo 4727d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 47283494fc30STejun Heo task_pid_nr(worker->task), 472930ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 47303494fc30STejun Heo worker->current_func); 47313494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 47323494fc30STejun Heo pr_cont_work(false, work); 47333494fc30STejun Heo comma = true; 47343494fc30STejun Heo } 47353494fc30STejun Heo pr_cont("\n"); 47363494fc30STejun Heo } 47373494fc30STejun Heo 47383494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47393494fc30STejun Heo if (get_work_pwq(work) == pwq) { 47403494fc30STejun Heo has_pending = true; 47413494fc30STejun Heo break; 47423494fc30STejun Heo } 47433494fc30STejun Heo } 47443494fc30STejun Heo if (has_pending) { 47453494fc30STejun Heo bool comma = false; 47463494fc30STejun Heo 47473494fc30STejun Heo pr_info(" pending:"); 47483494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47493494fc30STejun Heo if (get_work_pwq(work) != pwq) 47503494fc30STejun Heo continue; 47513494fc30STejun Heo 47523494fc30STejun Heo pr_cont_work(comma, work); 47533494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47543494fc30STejun Heo } 47553494fc30STejun Heo pr_cont("\n"); 47563494fc30STejun Heo } 47573494fc30STejun Heo 4758f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 47593494fc30STejun Heo bool comma = false; 47603494fc30STejun Heo 4761f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4762f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 47633494fc30STejun Heo pr_cont_work(comma, work); 47643494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47653494fc30STejun Heo } 47663494fc30STejun Heo pr_cont("\n"); 47673494fc30STejun Heo } 47683494fc30STejun Heo } 47693494fc30STejun Heo 47703494fc30STejun Heo /** 47713494fc30STejun Heo * show_workqueue_state - dump workqueue state 47723494fc30STejun Heo * 47737b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 47747b776af6SRoger Lu * all busy workqueues and pools. 47753494fc30STejun Heo */ 47763494fc30STejun Heo void show_workqueue_state(void) 47773494fc30STejun Heo { 47783494fc30STejun Heo struct workqueue_struct *wq; 47793494fc30STejun Heo struct worker_pool *pool; 47803494fc30STejun Heo unsigned long flags; 47813494fc30STejun Heo int pi; 47823494fc30STejun Heo 478324acfb71SThomas Gleixner rcu_read_lock(); 47843494fc30STejun Heo 47853494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 47863494fc30STejun Heo 47873494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 47883494fc30STejun Heo struct pool_workqueue *pwq; 47893494fc30STejun Heo bool idle = true; 47903494fc30STejun Heo 47913494fc30STejun Heo for_each_pwq(pwq, wq) { 4792f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 47933494fc30STejun Heo idle = false; 47943494fc30STejun Heo break; 47953494fc30STejun Heo } 47963494fc30STejun Heo } 47973494fc30STejun Heo if (idle) 47983494fc30STejun Heo continue; 47993494fc30STejun Heo 48003494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 48013494fc30STejun Heo 48023494fc30STejun Heo for_each_pwq(pwq, wq) { 4803a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4804f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 48053494fc30STejun Heo show_pwq(pwq); 4806a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 480762635ea8SSergey Senozhatsky /* 480862635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 480962635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 481062635ea8SSergey Senozhatsky * hard lockup. 481162635ea8SSergey Senozhatsky */ 481262635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48133494fc30STejun Heo } 48143494fc30STejun Heo } 48153494fc30STejun Heo 48163494fc30STejun Heo for_each_pool(pool, pi) { 48173494fc30STejun Heo struct worker *worker; 48183494fc30STejun Heo bool first = true; 48193494fc30STejun Heo 4820a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 48213494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 48223494fc30STejun Heo goto next_pool; 48233494fc30STejun Heo 48243494fc30STejun Heo pr_info("pool %d:", pool->id); 48253494fc30STejun Heo pr_cont_pool_info(pool); 482682607adcSTejun Heo pr_cont(" hung=%us workers=%d", 482782607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 482882607adcSTejun Heo pool->nr_workers); 48293494fc30STejun Heo if (pool->manager) 48303494fc30STejun Heo pr_cont(" manager: %d", 48313494fc30STejun Heo task_pid_nr(pool->manager->task)); 48323494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 48333494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 48343494fc30STejun Heo task_pid_nr(worker->task)); 48353494fc30STejun Heo first = false; 48363494fc30STejun Heo } 48373494fc30STejun Heo pr_cont("\n"); 48383494fc30STejun Heo next_pool: 4839a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 484062635ea8SSergey Senozhatsky /* 484162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 484262635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 484362635ea8SSergey Senozhatsky * hard lockup. 484462635ea8SSergey Senozhatsky */ 484562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48463494fc30STejun Heo } 48473494fc30STejun Heo 484824acfb71SThomas Gleixner rcu_read_unlock(); 48493494fc30STejun Heo } 48503494fc30STejun Heo 48516b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 48526b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 48536b59808bSTejun Heo { 48546b59808bSTejun Heo int off; 48556b59808bSTejun Heo 48566b59808bSTejun Heo /* always show the actual comm */ 48576b59808bSTejun Heo off = strscpy(buf, task->comm, size); 48586b59808bSTejun Heo if (off < 0) 48596b59808bSTejun Heo return; 48606b59808bSTejun Heo 4861197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 48626b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 48636b59808bSTejun Heo 4864197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4865197f6accSTejun Heo struct worker *worker = kthread_data(task); 4866197f6accSTejun Heo struct worker_pool *pool = worker->pool; 48676b59808bSTejun Heo 48686b59808bSTejun Heo if (pool) { 4869a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 48706b59808bSTejun Heo /* 4871197f6accSTejun Heo * ->desc tracks information (wq name or 4872197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4873197f6accSTejun Heo * current, prepend '+', otherwise '-'. 48746b59808bSTejun Heo */ 48756b59808bSTejun Heo if (worker->desc[0] != '\0') { 48766b59808bSTejun Heo if (worker->current_work) 48776b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 48786b59808bSTejun Heo worker->desc); 48796b59808bSTejun Heo else 48806b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 48816b59808bSTejun Heo worker->desc); 48826b59808bSTejun Heo } 4883a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 48846b59808bSTejun Heo } 4885197f6accSTejun Heo } 48866b59808bSTejun Heo 48876b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 48886b59808bSTejun Heo } 48896b59808bSTejun Heo 489066448bc2SMathieu Malaterre #ifdef CONFIG_SMP 489166448bc2SMathieu Malaterre 4892db7bccf4STejun Heo /* 4893db7bccf4STejun Heo * CPU hotplug. 4894db7bccf4STejun Heo * 4895e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4896112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4897706026c2STejun Heo * pool which make migrating pending and scheduled works very 4898e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 489994cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4900e22bee78STejun Heo * blocked draining impractical. 4901e22bee78STejun Heo * 490224647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4903628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4904628c78e7STejun Heo * cpu comes back online. 4905db7bccf4STejun Heo */ 4906db7bccf4STejun Heo 4907e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4908db7bccf4STejun Heo { 49094ce62e9eSTejun Heo struct worker_pool *pool; 4910db7bccf4STejun Heo struct worker *worker; 4911db7bccf4STejun Heo 4912f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49131258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 4914a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4915e22bee78STejun Heo 4916f2d5a0eeSTejun Heo /* 491792f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 491894cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 491994cf58bbSTejun Heo * except for the ones which are still executing works from 492094cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 492194cf58bbSTejun Heo * this, they may become diasporas. 4922f2d5a0eeSTejun Heo */ 4923da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4924403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4925db7bccf4STejun Heo 492624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4927f2d5a0eeSTejun Heo 4928a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 492906249738SPeter Zijlstra 49305c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 49315c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 4932547a77d0SLai Jiangshan WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 49335c25b5ffSPeter Zijlstra } 493406249738SPeter Zijlstra 49351258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 4936e22bee78STejun Heo 4937e22bee78STejun Heo /* 4938eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4939eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4940eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4941eb283428SLai Jiangshan * from other cpus. 4942628c78e7STejun Heo */ 4943628c78e7STejun Heo schedule(); 4944628c78e7STejun Heo 4945628c78e7STejun Heo /* 4946eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4947eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4948eb283428SLai Jiangshan * and keep_working() are always true as long as the 4949eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4950eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4951eb283428SLai Jiangshan * are served by workers tied to the pool. 4952e22bee78STejun Heo */ 4953e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4954eb283428SLai Jiangshan 4955eb283428SLai Jiangshan /* 4956eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4957eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4958eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4959eb283428SLai Jiangshan */ 4960a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4961eb283428SLai Jiangshan wake_up_worker(pool); 4962a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 4963eb283428SLai Jiangshan } 4964db7bccf4STejun Heo } 4965db7bccf4STejun Heo 4966bd7c089eSTejun Heo /** 4967bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4968bd7c089eSTejun Heo * @pool: pool of interest 4969bd7c089eSTejun Heo * 4970a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4971bd7c089eSTejun Heo */ 4972bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4973bd7c089eSTejun Heo { 4974a9ab775bSTejun Heo struct worker *worker; 4975bd7c089eSTejun Heo 49761258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 4977bd7c089eSTejun Heo 4978bd7c089eSTejun Heo /* 4979a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4980a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4981402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4982a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4983a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4984bd7c089eSTejun Heo */ 49855c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 49865c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 4987a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4988a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 49895c25b5ffSPeter Zijlstra } 4990a9ab775bSTejun Heo 4991a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4992f7c17d26SWanpeng Li 49933de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4994a9ab775bSTejun Heo 4995da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4996a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4997a9ab775bSTejun Heo 4998a9ab775bSTejun Heo /* 4999a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 5000a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 5001a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 5002a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 5003a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 5004a9ab775bSTejun Heo * be bound before @pool->lock is released. 5005a9ab775bSTejun Heo */ 5006a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 5007bd7c089eSTejun Heo wake_up_process(worker->task); 5008bd7c089eSTejun Heo 5009bd7c089eSTejun Heo /* 5010a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5011a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5012a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5013a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5014a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5015a9ab775bSTejun Heo * concurrency management. Note that when or whether 5016a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5017a9ab775bSTejun Heo * 5018c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5019a9ab775bSTejun Heo * tested without holding any lock in 50206d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5021a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5022a9ab775bSTejun Heo * management operations. 5023bd7c089eSTejun Heo */ 5024a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5025a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5026a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5027c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5028bd7c089eSTejun Heo } 5029a9ab775bSTejun Heo 5030a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5031bd7c089eSTejun Heo } 5032bd7c089eSTejun Heo 50337dbc725eSTejun Heo /** 50347dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 50357dbc725eSTejun Heo * @pool: unbound pool of interest 50367dbc725eSTejun Heo * @cpu: the CPU which is coming up 50377dbc725eSTejun Heo * 50387dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 50397dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 50407dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 50417dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 50427dbc725eSTejun Heo */ 50437dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 50447dbc725eSTejun Heo { 50457dbc725eSTejun Heo static cpumask_t cpumask; 50467dbc725eSTejun Heo struct worker *worker; 50477dbc725eSTejun Heo 50481258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 50497dbc725eSTejun Heo 50507dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 50517dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 50527dbc725eSTejun Heo return; 50537dbc725eSTejun Heo 50547dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 50557dbc725eSTejun Heo 50567dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5057da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5058d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 50597dbc725eSTejun Heo } 50607dbc725eSTejun Heo 50617ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 50621da177e4SLinus Torvalds { 50634ce62e9eSTejun Heo struct worker_pool *pool; 50641da177e4SLinus Torvalds 5065f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 50663ce63377STejun Heo if (pool->nr_workers) 50673ce63377STejun Heo continue; 5068051e1850SLai Jiangshan if (!create_worker(pool)) 50697ee681b2SThomas Gleixner return -ENOMEM; 50703af24433SOleg Nesterov } 50717ee681b2SThomas Gleixner return 0; 50727ee681b2SThomas Gleixner } 50731da177e4SLinus Torvalds 50747ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 50757ee681b2SThomas Gleixner { 50767ee681b2SThomas Gleixner struct worker_pool *pool; 50777ee681b2SThomas Gleixner struct workqueue_struct *wq; 50787ee681b2SThomas Gleixner int pi; 50797ee681b2SThomas Gleixner 508068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 50817dbc725eSTejun Heo 50827dbc725eSTejun Heo for_each_pool(pool, pi) { 50831258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 508494cf58bbSTejun Heo 5085f05b558dSLai Jiangshan if (pool->cpu == cpu) 508694cf58bbSTejun Heo rebind_workers(pool); 5087f05b558dSLai Jiangshan else if (pool->cpu < 0) 50887dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 508994cf58bbSTejun Heo 50901258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 509194cf58bbSTejun Heo } 50927dbc725eSTejun Heo 50934c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 50944c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 50954c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 50964c16bd32STejun Heo 509768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 50987ee681b2SThomas Gleixner return 0; 509965758202STejun Heo } 510065758202STejun Heo 51017ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 510265758202STejun Heo { 51034c16bd32STejun Heo struct workqueue_struct *wq; 51048db25e78STejun Heo 51054c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5106e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5107e8b3f8dbSLai Jiangshan return -1; 5108e8b3f8dbSLai Jiangshan 5109e8b3f8dbSLai Jiangshan unbind_workers(cpu); 51104c16bd32STejun Heo 51114c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51124c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 51134c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51144c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 51154c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 51164c16bd32STejun Heo 51177ee681b2SThomas Gleixner return 0; 511865758202STejun Heo } 511965758202STejun Heo 51202d3854a3SRusty Russell struct work_for_cpu { 5121ed48ece2STejun Heo struct work_struct work; 51222d3854a3SRusty Russell long (*fn)(void *); 51232d3854a3SRusty Russell void *arg; 51242d3854a3SRusty Russell long ret; 51252d3854a3SRusty Russell }; 51262d3854a3SRusty Russell 5127ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 51282d3854a3SRusty Russell { 5129ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5130ed48ece2STejun Heo 51312d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 51322d3854a3SRusty Russell } 51332d3854a3SRusty Russell 51342d3854a3SRusty Russell /** 513522aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 51362d3854a3SRusty Russell * @cpu: the cpu to run on 51372d3854a3SRusty Russell * @fn: the function to run 51382d3854a3SRusty Russell * @arg: the function arg 51392d3854a3SRusty Russell * 514031ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 51416b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5142d185af30SYacine Belkadi * 5143d185af30SYacine Belkadi * Return: The value @fn returns. 51442d3854a3SRusty Russell */ 5145d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 51462d3854a3SRusty Russell { 5147ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 51482d3854a3SRusty Russell 5149ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5150ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 515112997d1aSBjorn Helgaas flush_work(&wfc.work); 5152440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 51532d3854a3SRusty Russell return wfc.ret; 51542d3854a3SRusty Russell } 51552d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 51560e8d6a93SThomas Gleixner 51570e8d6a93SThomas Gleixner /** 51580e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 51590e8d6a93SThomas Gleixner * @cpu: the cpu to run on 51600e8d6a93SThomas Gleixner * @fn: the function to run 51610e8d6a93SThomas Gleixner * @arg: the function argument 51620e8d6a93SThomas Gleixner * 51630e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 51640e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 51650e8d6a93SThomas Gleixner * 51660e8d6a93SThomas Gleixner * Return: The value @fn returns. 51670e8d6a93SThomas Gleixner */ 51680e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 51690e8d6a93SThomas Gleixner { 51700e8d6a93SThomas Gleixner long ret = -ENODEV; 51710e8d6a93SThomas Gleixner 5172ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 51730e8d6a93SThomas Gleixner if (cpu_online(cpu)) 51740e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5175ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 51760e8d6a93SThomas Gleixner return ret; 51770e8d6a93SThomas Gleixner } 51780e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 51792d3854a3SRusty Russell #endif /* CONFIG_SMP */ 51802d3854a3SRusty Russell 5181a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5182e7577c50SRusty Russell 5183a0a1a5fdSTejun Heo /** 5184a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5185a0a1a5fdSTejun Heo * 518658a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5187f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5188706026c2STejun Heo * pool->worklist. 5189a0a1a5fdSTejun Heo * 5190a0a1a5fdSTejun Heo * CONTEXT: 5191a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5192a0a1a5fdSTejun Heo */ 5193a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5194a0a1a5fdSTejun Heo { 519524b8a847STejun Heo struct workqueue_struct *wq; 519624b8a847STejun Heo struct pool_workqueue *pwq; 5197a0a1a5fdSTejun Heo 519868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5199a0a1a5fdSTejun Heo 52006183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5201a0a1a5fdSTejun Heo workqueue_freezing = true; 5202a0a1a5fdSTejun Heo 520324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5204a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5205699ce097STejun Heo for_each_pwq(pwq, wq) 5206699ce097STejun Heo pwq_adjust_max_active(pwq); 5207a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5208a1056305STejun Heo } 52095bcab335STejun Heo 521068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5211a0a1a5fdSTejun Heo } 5212a0a1a5fdSTejun Heo 5213a0a1a5fdSTejun Heo /** 521458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5215a0a1a5fdSTejun Heo * 5216a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5217a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5218a0a1a5fdSTejun Heo * 5219a0a1a5fdSTejun Heo * CONTEXT: 522068e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5221a0a1a5fdSTejun Heo * 5222d185af30SYacine Belkadi * Return: 522358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 522458a69cb4STejun Heo * is complete. 5225a0a1a5fdSTejun Heo */ 5226a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5227a0a1a5fdSTejun Heo { 5228a0a1a5fdSTejun Heo bool busy = false; 522924b8a847STejun Heo struct workqueue_struct *wq; 523024b8a847STejun Heo struct pool_workqueue *pwq; 5231a0a1a5fdSTejun Heo 523268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5233a0a1a5fdSTejun Heo 52346183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5235a0a1a5fdSTejun Heo 523624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 523724b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 523824b8a847STejun Heo continue; 5239a0a1a5fdSTejun Heo /* 5240a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5241a0a1a5fdSTejun Heo * to peek without lock. 5242a0a1a5fdSTejun Heo */ 524324acfb71SThomas Gleixner rcu_read_lock(); 524424b8a847STejun Heo for_each_pwq(pwq, wq) { 52456183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5246112202d9STejun Heo if (pwq->nr_active) { 5247a0a1a5fdSTejun Heo busy = true; 524824acfb71SThomas Gleixner rcu_read_unlock(); 5249a0a1a5fdSTejun Heo goto out_unlock; 5250a0a1a5fdSTejun Heo } 5251a0a1a5fdSTejun Heo } 525224acfb71SThomas Gleixner rcu_read_unlock(); 5253a0a1a5fdSTejun Heo } 5254a0a1a5fdSTejun Heo out_unlock: 525568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5256a0a1a5fdSTejun Heo return busy; 5257a0a1a5fdSTejun Heo } 5258a0a1a5fdSTejun Heo 5259a0a1a5fdSTejun Heo /** 5260a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5261a0a1a5fdSTejun Heo * 5262a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5263706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5264a0a1a5fdSTejun Heo * 5265a0a1a5fdSTejun Heo * CONTEXT: 5266a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5267a0a1a5fdSTejun Heo */ 5268a0a1a5fdSTejun Heo void thaw_workqueues(void) 5269a0a1a5fdSTejun Heo { 527024b8a847STejun Heo struct workqueue_struct *wq; 527124b8a847STejun Heo struct pool_workqueue *pwq; 5272a0a1a5fdSTejun Heo 527368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5274a0a1a5fdSTejun Heo 5275a0a1a5fdSTejun Heo if (!workqueue_freezing) 5276a0a1a5fdSTejun Heo goto out_unlock; 5277a0a1a5fdSTejun Heo 527874b414eaSLai Jiangshan workqueue_freezing = false; 527924b8a847STejun Heo 528024b8a847STejun Heo /* restore max_active and repopulate worklist */ 528124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5282a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5283699ce097STejun Heo for_each_pwq(pwq, wq) 5284699ce097STejun Heo pwq_adjust_max_active(pwq); 5285a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 528624b8a847STejun Heo } 528724b8a847STejun Heo 5288a0a1a5fdSTejun Heo out_unlock: 528968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5290a0a1a5fdSTejun Heo } 5291a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5292a0a1a5fdSTejun Heo 5293042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5294042f7df1SLai Jiangshan { 5295042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5296042f7df1SLai Jiangshan int ret = 0; 5297042f7df1SLai Jiangshan struct workqueue_struct *wq; 5298042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5299042f7df1SLai Jiangshan 5300042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5301042f7df1SLai Jiangshan 5302042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5303042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5304042f7df1SLai Jiangshan continue; 5305042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5306042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5307042f7df1SLai Jiangshan continue; 5308042f7df1SLai Jiangshan 5309042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5310042f7df1SLai Jiangshan if (!ctx) { 5311042f7df1SLai Jiangshan ret = -ENOMEM; 5312042f7df1SLai Jiangshan break; 5313042f7df1SLai Jiangshan } 5314042f7df1SLai Jiangshan 5315042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5316042f7df1SLai Jiangshan } 5317042f7df1SLai Jiangshan 5318042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5319042f7df1SLai Jiangshan if (!ret) 5320042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5321042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5322042f7df1SLai Jiangshan } 5323042f7df1SLai Jiangshan 5324042f7df1SLai Jiangshan return ret; 5325042f7df1SLai Jiangshan } 5326042f7df1SLai Jiangshan 5327042f7df1SLai Jiangshan /** 5328042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5329042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5330042f7df1SLai Jiangshan * 5331042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5332042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5333042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5334042f7df1SLai Jiangshan * 533567dc8325SCai Huoqing * Return: 0 - Success 5336042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5337042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5338042f7df1SLai Jiangshan */ 5339042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5340042f7df1SLai Jiangshan { 5341042f7df1SLai Jiangshan int ret = -EINVAL; 5342042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5343042f7df1SLai Jiangshan 5344042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5345042f7df1SLai Jiangshan return -ENOMEM; 5346042f7df1SLai Jiangshan 5347c98a9805STal Shorer /* 5348c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5349c98a9805STal Shorer * If the user wishes to include them, we allow that. 5350c98a9805STal Shorer */ 5351042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5352042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5353a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5354042f7df1SLai Jiangshan 5355042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5356042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5357042f7df1SLai Jiangshan 5358042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5359042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5360042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5361042f7df1SLai Jiangshan 5362042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5363042f7df1SLai Jiangshan if (ret < 0) 5364042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5365042f7df1SLai Jiangshan 5366a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5367042f7df1SLai Jiangshan } 5368042f7df1SLai Jiangshan 5369042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 5370042f7df1SLai Jiangshan return ret; 5371042f7df1SLai Jiangshan } 5372042f7df1SLai Jiangshan 53736ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 53746ba94429SFrederic Weisbecker /* 53756ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 53766ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 53776ba94429SFrederic Weisbecker * following attributes. 53786ba94429SFrederic Weisbecker * 53796ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 53806ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 53816ba94429SFrederic Weisbecker * 53826ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 53836ba94429SFrederic Weisbecker * 53849a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 53856ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 53866ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 53879a19b463SWang Long * numa RW bool : whether enable NUMA affinity 53886ba94429SFrederic Weisbecker */ 53896ba94429SFrederic Weisbecker struct wq_device { 53906ba94429SFrederic Weisbecker struct workqueue_struct *wq; 53916ba94429SFrederic Weisbecker struct device dev; 53926ba94429SFrederic Weisbecker }; 53936ba94429SFrederic Weisbecker 53946ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 53956ba94429SFrederic Weisbecker { 53966ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 53976ba94429SFrederic Weisbecker 53986ba94429SFrederic Weisbecker return wq_dev->wq; 53996ba94429SFrederic Weisbecker } 54006ba94429SFrederic Weisbecker 54016ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 54026ba94429SFrederic Weisbecker char *buf) 54036ba94429SFrederic Weisbecker { 54046ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54056ba94429SFrederic Weisbecker 54066ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 54076ba94429SFrederic Weisbecker } 54086ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 54096ba94429SFrederic Weisbecker 54106ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 54116ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54126ba94429SFrederic Weisbecker { 54136ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54146ba94429SFrederic Weisbecker 54156ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 54166ba94429SFrederic Weisbecker } 54176ba94429SFrederic Weisbecker 54186ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 54196ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 54206ba94429SFrederic Weisbecker size_t count) 54216ba94429SFrederic Weisbecker { 54226ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54236ba94429SFrederic Weisbecker int val; 54246ba94429SFrederic Weisbecker 54256ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 54266ba94429SFrederic Weisbecker return -EINVAL; 54276ba94429SFrederic Weisbecker 54286ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 54296ba94429SFrederic Weisbecker return count; 54306ba94429SFrederic Weisbecker } 54316ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 54326ba94429SFrederic Weisbecker 54336ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 54346ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 54356ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 54366ba94429SFrederic Weisbecker NULL, 54376ba94429SFrederic Weisbecker }; 54386ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 54396ba94429SFrederic Weisbecker 54406ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 54416ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54426ba94429SFrederic Weisbecker { 54436ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54446ba94429SFrederic Weisbecker const char *delim = ""; 54456ba94429SFrederic Weisbecker int node, written = 0; 54466ba94429SFrederic Weisbecker 5447ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 544824acfb71SThomas Gleixner rcu_read_lock(); 54496ba94429SFrederic Weisbecker for_each_node(node) { 54506ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 54516ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 54526ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 54536ba94429SFrederic Weisbecker delim = " "; 54546ba94429SFrederic Weisbecker } 54556ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 545624acfb71SThomas Gleixner rcu_read_unlock(); 5457ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 54586ba94429SFrederic Weisbecker 54596ba94429SFrederic Weisbecker return written; 54606ba94429SFrederic Weisbecker } 54616ba94429SFrederic Weisbecker 54626ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 54636ba94429SFrederic Weisbecker char *buf) 54646ba94429SFrederic Weisbecker { 54656ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54666ba94429SFrederic Weisbecker int written; 54676ba94429SFrederic Weisbecker 54686ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 54696ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 54706ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 54716ba94429SFrederic Weisbecker 54726ba94429SFrederic Weisbecker return written; 54736ba94429SFrederic Weisbecker } 54746ba94429SFrederic Weisbecker 54756ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 54766ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 54776ba94429SFrederic Weisbecker { 54786ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 54796ba94429SFrederic Weisbecker 5480899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5481899a94feSLai Jiangshan 5482be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 54836ba94429SFrederic Weisbecker if (!attrs) 54846ba94429SFrederic Weisbecker return NULL; 54856ba94429SFrederic Weisbecker 54866ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 54876ba94429SFrederic Weisbecker return attrs; 54886ba94429SFrederic Weisbecker } 54896ba94429SFrederic Weisbecker 54906ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 54916ba94429SFrederic Weisbecker const char *buf, size_t count) 54926ba94429SFrederic Weisbecker { 54936ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54946ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5495d4d3e257SLai Jiangshan int ret = -ENOMEM; 5496d4d3e257SLai Jiangshan 5497d4d3e257SLai Jiangshan apply_wqattrs_lock(); 54986ba94429SFrederic Weisbecker 54996ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55006ba94429SFrederic Weisbecker if (!attrs) 5501d4d3e257SLai Jiangshan goto out_unlock; 55026ba94429SFrederic Weisbecker 55036ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 55046ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5505d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55066ba94429SFrederic Weisbecker else 55076ba94429SFrederic Weisbecker ret = -EINVAL; 55086ba94429SFrederic Weisbecker 5509d4d3e257SLai Jiangshan out_unlock: 5510d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55116ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55126ba94429SFrederic Weisbecker return ret ?: count; 55136ba94429SFrederic Weisbecker } 55146ba94429SFrederic Weisbecker 55156ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 55166ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55176ba94429SFrederic Weisbecker { 55186ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55196ba94429SFrederic Weisbecker int written; 55206ba94429SFrederic Weisbecker 55216ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55226ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 55236ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 55246ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55256ba94429SFrederic Weisbecker return written; 55266ba94429SFrederic Weisbecker } 55276ba94429SFrederic Weisbecker 55286ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 55296ba94429SFrederic Weisbecker struct device_attribute *attr, 55306ba94429SFrederic Weisbecker const char *buf, size_t count) 55316ba94429SFrederic Weisbecker { 55326ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55336ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5534d4d3e257SLai Jiangshan int ret = -ENOMEM; 5535d4d3e257SLai Jiangshan 5536d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55376ba94429SFrederic Weisbecker 55386ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55396ba94429SFrederic Weisbecker if (!attrs) 5540d4d3e257SLai Jiangshan goto out_unlock; 55416ba94429SFrederic Weisbecker 55426ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 55436ba94429SFrederic Weisbecker if (!ret) 5544d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55456ba94429SFrederic Weisbecker 5546d4d3e257SLai Jiangshan out_unlock: 5547d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55486ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55496ba94429SFrederic Weisbecker return ret ?: count; 55506ba94429SFrederic Weisbecker } 55516ba94429SFrederic Weisbecker 55526ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 55536ba94429SFrederic Weisbecker char *buf) 55546ba94429SFrederic Weisbecker { 55556ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55566ba94429SFrederic Weisbecker int written; 55576ba94429SFrederic Weisbecker 55586ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55596ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 55606ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 55616ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55626ba94429SFrederic Weisbecker 55636ba94429SFrederic Weisbecker return written; 55646ba94429SFrederic Weisbecker } 55656ba94429SFrederic Weisbecker 55666ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 55676ba94429SFrederic Weisbecker const char *buf, size_t count) 55686ba94429SFrederic Weisbecker { 55696ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55706ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5571d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5572d4d3e257SLai Jiangshan 5573d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55746ba94429SFrederic Weisbecker 55756ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55766ba94429SFrederic Weisbecker if (!attrs) 5577d4d3e257SLai Jiangshan goto out_unlock; 55786ba94429SFrederic Weisbecker 55796ba94429SFrederic Weisbecker ret = -EINVAL; 55806ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 55816ba94429SFrederic Weisbecker attrs->no_numa = !v; 5582d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55836ba94429SFrederic Weisbecker } 55846ba94429SFrederic Weisbecker 5585d4d3e257SLai Jiangshan out_unlock: 5586d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55876ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55886ba94429SFrederic Weisbecker return ret ?: count; 55896ba94429SFrederic Weisbecker } 55906ba94429SFrederic Weisbecker 55916ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 55926ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 55936ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 55946ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 55956ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 55966ba94429SFrederic Weisbecker __ATTR_NULL, 55976ba94429SFrederic Weisbecker }; 55986ba94429SFrederic Weisbecker 55996ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 56006ba94429SFrederic Weisbecker .name = "workqueue", 56016ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 56026ba94429SFrederic Weisbecker }; 56036ba94429SFrederic Weisbecker 5604b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5605b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5606b05a7928SFrederic Weisbecker { 5607b05a7928SFrederic Weisbecker int written; 5608b05a7928SFrederic Weisbecker 5609042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5610b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5611b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5612042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5613b05a7928SFrederic Weisbecker 5614b05a7928SFrederic Weisbecker return written; 5615b05a7928SFrederic Weisbecker } 5616b05a7928SFrederic Weisbecker 5617042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5618042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5619042f7df1SLai Jiangshan { 5620042f7df1SLai Jiangshan cpumask_var_t cpumask; 5621042f7df1SLai Jiangshan int ret; 5622042f7df1SLai Jiangshan 5623042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5624042f7df1SLai Jiangshan return -ENOMEM; 5625042f7df1SLai Jiangshan 5626042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5627042f7df1SLai Jiangshan if (!ret) 5628042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5629042f7df1SLai Jiangshan 5630042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5631042f7df1SLai Jiangshan return ret ? ret : count; 5632042f7df1SLai Jiangshan } 5633042f7df1SLai Jiangshan 5634b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5635042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5636042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5637b05a7928SFrederic Weisbecker 56386ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 56396ba94429SFrederic Weisbecker { 5640b05a7928SFrederic Weisbecker int err; 5641b05a7928SFrederic Weisbecker 5642b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5643b05a7928SFrederic Weisbecker if (err) 5644b05a7928SFrederic Weisbecker return err; 5645b05a7928SFrederic Weisbecker 5646b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 56476ba94429SFrederic Weisbecker } 56486ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 56496ba94429SFrederic Weisbecker 56506ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 56516ba94429SFrederic Weisbecker { 56526ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 56536ba94429SFrederic Weisbecker 56546ba94429SFrederic Weisbecker kfree(wq_dev); 56556ba94429SFrederic Weisbecker } 56566ba94429SFrederic Weisbecker 56576ba94429SFrederic Weisbecker /** 56586ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 56596ba94429SFrederic Weisbecker * @wq: the workqueue to register 56606ba94429SFrederic Weisbecker * 56616ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 56626ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 56636ba94429SFrederic Weisbecker * which is the preferred method. 56646ba94429SFrederic Weisbecker * 56656ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 56666ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 56676ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 56686ba94429SFrederic Weisbecker * attributes. 56696ba94429SFrederic Weisbecker * 56706ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 56716ba94429SFrederic Weisbecker */ 56726ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 56736ba94429SFrederic Weisbecker { 56746ba94429SFrederic Weisbecker struct wq_device *wq_dev; 56756ba94429SFrederic Weisbecker int ret; 56766ba94429SFrederic Weisbecker 56776ba94429SFrederic Weisbecker /* 5678402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 56796ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 56806ba94429SFrederic Weisbecker * workqueues. 56816ba94429SFrederic Weisbecker */ 56820a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 56836ba94429SFrederic Weisbecker return -EINVAL; 56846ba94429SFrederic Weisbecker 56856ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 56866ba94429SFrederic Weisbecker if (!wq_dev) 56876ba94429SFrederic Weisbecker return -ENOMEM; 56886ba94429SFrederic Weisbecker 56896ba94429SFrederic Weisbecker wq_dev->wq = wq; 56906ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 56916ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 569223217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 56936ba94429SFrederic Weisbecker 56946ba94429SFrederic Weisbecker /* 56956ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 56966ba94429SFrederic Weisbecker * everything is ready. 56976ba94429SFrederic Weisbecker */ 56986ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 56996ba94429SFrederic Weisbecker 57006ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 57016ba94429SFrederic Weisbecker if (ret) { 5702537f4146SArvind Yadav put_device(&wq_dev->dev); 57036ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57046ba94429SFrederic Weisbecker return ret; 57056ba94429SFrederic Weisbecker } 57066ba94429SFrederic Weisbecker 57076ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 57086ba94429SFrederic Weisbecker struct device_attribute *attr; 57096ba94429SFrederic Weisbecker 57106ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 57116ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 57126ba94429SFrederic Weisbecker if (ret) { 57136ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57146ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57156ba94429SFrederic Weisbecker return ret; 57166ba94429SFrederic Weisbecker } 57176ba94429SFrederic Weisbecker } 57186ba94429SFrederic Weisbecker } 57196ba94429SFrederic Weisbecker 57206ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 57216ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 57226ba94429SFrederic Weisbecker return 0; 57236ba94429SFrederic Weisbecker } 57246ba94429SFrederic Weisbecker 57256ba94429SFrederic Weisbecker /** 57266ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 57276ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 57286ba94429SFrederic Weisbecker * 57296ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 57306ba94429SFrederic Weisbecker */ 57316ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 57326ba94429SFrederic Weisbecker { 57336ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 57346ba94429SFrederic Weisbecker 57356ba94429SFrederic Weisbecker if (!wq->wq_dev) 57366ba94429SFrederic Weisbecker return; 57376ba94429SFrederic Weisbecker 57386ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57396ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57406ba94429SFrederic Weisbecker } 57416ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 57426ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 57436ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 57446ba94429SFrederic Weisbecker 574582607adcSTejun Heo /* 574682607adcSTejun Heo * Workqueue watchdog. 574782607adcSTejun Heo * 574882607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 574982607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 575082607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 575182607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 575282607adcSTejun Heo * largely opaque. 575382607adcSTejun Heo * 575482607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 575582607adcSTejun Heo * state if some pools failed to make forward progress for a while where 575682607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 575782607adcSTejun Heo * 575882607adcSTejun Heo * This mechanism is controlled through the kernel parameter 575982607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 576082607adcSTejun Heo * corresponding sysfs parameter file. 576182607adcSTejun Heo */ 576282607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 576382607adcSTejun Heo 576482607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 57655cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 576682607adcSTejun Heo 576782607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 576882607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 576982607adcSTejun Heo 577082607adcSTejun Heo static void wq_watchdog_reset_touched(void) 577182607adcSTejun Heo { 577282607adcSTejun Heo int cpu; 577382607adcSTejun Heo 577482607adcSTejun Heo wq_watchdog_touched = jiffies; 577582607adcSTejun Heo for_each_possible_cpu(cpu) 577682607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 577782607adcSTejun Heo } 577882607adcSTejun Heo 57795cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 578082607adcSTejun Heo { 578182607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 578282607adcSTejun Heo bool lockup_detected = false; 5783940d71c6SSergey Senozhatsky unsigned long now = jiffies; 578482607adcSTejun Heo struct worker_pool *pool; 578582607adcSTejun Heo int pi; 578682607adcSTejun Heo 578782607adcSTejun Heo if (!thresh) 578882607adcSTejun Heo return; 578982607adcSTejun Heo 579082607adcSTejun Heo rcu_read_lock(); 579182607adcSTejun Heo 579282607adcSTejun Heo for_each_pool(pool, pi) { 579382607adcSTejun Heo unsigned long pool_ts, touched, ts; 579482607adcSTejun Heo 579582607adcSTejun Heo if (list_empty(&pool->worklist)) 579682607adcSTejun Heo continue; 579782607adcSTejun Heo 5798940d71c6SSergey Senozhatsky /* 5799940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 5800940d71c6SSergey Senozhatsky * the watchdog like a stall. 5801940d71c6SSergey Senozhatsky */ 5802940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 5803940d71c6SSergey Senozhatsky 580482607adcSTejun Heo /* get the latest of pool and touched timestamps */ 580589e28ce6SWang Qing if (pool->cpu >= 0) 580689e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 580789e28ce6SWang Qing else 580882607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 580989e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 581082607adcSTejun Heo 581182607adcSTejun Heo if (time_after(pool_ts, touched)) 581282607adcSTejun Heo ts = pool_ts; 581382607adcSTejun Heo else 581482607adcSTejun Heo ts = touched; 581582607adcSTejun Heo 581682607adcSTejun Heo /* did we stall? */ 5817940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 581882607adcSTejun Heo lockup_detected = true; 581982607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 582082607adcSTejun Heo pr_cont_pool_info(pool); 582182607adcSTejun Heo pr_cont(" stuck for %us!\n", 5822940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 582382607adcSTejun Heo } 582482607adcSTejun Heo } 582582607adcSTejun Heo 582682607adcSTejun Heo rcu_read_unlock(); 582782607adcSTejun Heo 582882607adcSTejun Heo if (lockup_detected) 582982607adcSTejun Heo show_workqueue_state(); 583082607adcSTejun Heo 583182607adcSTejun Heo wq_watchdog_reset_touched(); 583282607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 583382607adcSTejun Heo } 583482607adcSTejun Heo 5835cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 583682607adcSTejun Heo { 583782607adcSTejun Heo if (cpu >= 0) 583882607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 583989e28ce6SWang Qing 584082607adcSTejun Heo wq_watchdog_touched = jiffies; 584182607adcSTejun Heo } 584282607adcSTejun Heo 584382607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 584482607adcSTejun Heo { 584582607adcSTejun Heo wq_watchdog_thresh = 0; 584682607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 584782607adcSTejun Heo 584882607adcSTejun Heo if (thresh) { 584982607adcSTejun Heo wq_watchdog_thresh = thresh; 585082607adcSTejun Heo wq_watchdog_reset_touched(); 585182607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 585282607adcSTejun Heo } 585382607adcSTejun Heo } 585482607adcSTejun Heo 585582607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 585682607adcSTejun Heo const struct kernel_param *kp) 585782607adcSTejun Heo { 585882607adcSTejun Heo unsigned long thresh; 585982607adcSTejun Heo int ret; 586082607adcSTejun Heo 586182607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 586282607adcSTejun Heo if (ret) 586382607adcSTejun Heo return ret; 586482607adcSTejun Heo 586582607adcSTejun Heo if (system_wq) 586682607adcSTejun Heo wq_watchdog_set_thresh(thresh); 586782607adcSTejun Heo else 586882607adcSTejun Heo wq_watchdog_thresh = thresh; 586982607adcSTejun Heo 587082607adcSTejun Heo return 0; 587182607adcSTejun Heo } 587282607adcSTejun Heo 587382607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 587482607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 587582607adcSTejun Heo .get = param_get_ulong, 587682607adcSTejun Heo }; 587782607adcSTejun Heo 587882607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 587982607adcSTejun Heo 0644); 588082607adcSTejun Heo 588182607adcSTejun Heo static void wq_watchdog_init(void) 588282607adcSTejun Heo { 58835cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 588482607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 588582607adcSTejun Heo } 588682607adcSTejun Heo 588782607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 588882607adcSTejun Heo 588982607adcSTejun Heo static inline void wq_watchdog_init(void) { } 589082607adcSTejun Heo 589182607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 589282607adcSTejun Heo 5893bce90380STejun Heo static void __init wq_numa_init(void) 5894bce90380STejun Heo { 5895bce90380STejun Heo cpumask_var_t *tbl; 5896bce90380STejun Heo int node, cpu; 5897bce90380STejun Heo 5898bce90380STejun Heo if (num_possible_nodes() <= 1) 5899bce90380STejun Heo return; 5900bce90380STejun Heo 5901d55262c4STejun Heo if (wq_disable_numa) { 5902d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5903d55262c4STejun Heo return; 5904d55262c4STejun Heo } 5905d55262c4STejun Heo 5906f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 5907f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 5908f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5909f728c4a9SZhen Lei return; 5910f728c4a9SZhen Lei } 5911f728c4a9SZhen Lei } 5912f728c4a9SZhen Lei 5913be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 59144c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 59154c16bd32STejun Heo 5916bce90380STejun Heo /* 5917bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5918bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5919bce90380STejun Heo * fully initialized by now. 5920bce90380STejun Heo */ 59216396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5922bce90380STejun Heo BUG_ON(!tbl); 5923bce90380STejun Heo 5924bce90380STejun Heo for_each_node(node) 59255a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 59261be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5927bce90380STejun Heo 5928bce90380STejun Heo for_each_possible_cpu(cpu) { 5929bce90380STejun Heo node = cpu_to_node(cpu); 5930bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5931bce90380STejun Heo } 5932bce90380STejun Heo 5933bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5934bce90380STejun Heo wq_numa_enabled = true; 5935bce90380STejun Heo } 5936bce90380STejun Heo 59373347fa09STejun Heo /** 59383347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 59393347fa09STejun Heo * 59403347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 59413347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 59423347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 59433347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 59443347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 59453347fa09STejun Heo * created and scheduled right before early initcalls. 59463347fa09STejun Heo */ 59472333e829SYu Chen void __init workqueue_init_early(void) 59481da177e4SLinus Torvalds { 59497a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 59501bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 59517a4e344cSTejun Heo int i, cpu; 5952c34056a3STejun Heo 595310cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5954e904e6c2STejun Heo 5955b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 59561bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 5957b05a7928SFrederic Weisbecker 5958e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5959e904e6c2STejun Heo 5960706026c2STejun Heo /* initialize CPU pools */ 596129c91e99STejun Heo for_each_possible_cpu(cpu) { 59624ce62e9eSTejun Heo struct worker_pool *pool; 59638b03ae3cSTejun Heo 59647a4e344cSTejun Heo i = 0; 5965f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 59667a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5967ec22ca5eSTejun Heo pool->cpu = cpu; 59687a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 59697a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5970f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 59717a4e344cSTejun Heo 59729daf9e67STejun Heo /* alloc pool ID */ 597368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 59749daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 597568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 59764ce62e9eSTejun Heo } 59778b03ae3cSTejun Heo } 59788b03ae3cSTejun Heo 59798a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 598029c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 598129c91e99STejun Heo struct workqueue_attrs *attrs; 598229c91e99STejun Heo 5983be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 598429c91e99STejun Heo attrs->nice = std_nice[i]; 598529c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 59868a2b7538STejun Heo 59878a2b7538STejun Heo /* 59888a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 59898a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 59908a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 59918a2b7538STejun Heo */ 5992be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 59938a2b7538STejun Heo attrs->nice = std_nice[i]; 59948a2b7538STejun Heo attrs->no_numa = true; 59958a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 599629c91e99STejun Heo } 599729c91e99STejun Heo 5998d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 59991aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6000d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6001f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6002f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 600324d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 600424d51addSTejun Heo WQ_FREEZABLE, 0); 60050668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 60060668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 60070668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 60080668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 60090668106cSViresh Kumar 0); 60101aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 60110668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 60120668106cSViresh Kumar !system_power_efficient_wq || 60130668106cSViresh Kumar !system_freezable_power_efficient_wq); 60143347fa09STejun Heo } 60153347fa09STejun Heo 60163347fa09STejun Heo /** 60173347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 60183347fa09STejun Heo * 60193347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 60203347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 60213347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 60223347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 60233347fa09STejun Heo * with the initial workers and enable future kworker creations. 60243347fa09STejun Heo */ 60252333e829SYu Chen void __init workqueue_init(void) 60263347fa09STejun Heo { 60272186d9f9STejun Heo struct workqueue_struct *wq; 60283347fa09STejun Heo struct worker_pool *pool; 60293347fa09STejun Heo int cpu, bkt; 60303347fa09STejun Heo 60312186d9f9STejun Heo /* 60322186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 60332186d9f9STejun Heo * CPU to node mapping may not be available that early on some 60342186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 60352186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 60362186d9f9STejun Heo * affinity, fix them up. 603740c17f75STejun Heo * 603840c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 60392186d9f9STejun Heo */ 60402186d9f9STejun Heo wq_numa_init(); 60412186d9f9STejun Heo 60422186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 60432186d9f9STejun Heo 60442186d9f9STejun Heo for_each_possible_cpu(cpu) { 60452186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 60462186d9f9STejun Heo pool->node = cpu_to_node(cpu); 60472186d9f9STejun Heo } 60482186d9f9STejun Heo } 60492186d9f9STejun Heo 605040c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 60512186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 605240c17f75STejun Heo WARN(init_rescuer(wq), 605340c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 605440c17f75STejun Heo wq->name); 605540c17f75STejun Heo } 60562186d9f9STejun Heo 60572186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 60582186d9f9STejun Heo 60593347fa09STejun Heo /* create the initial workers */ 60603347fa09STejun Heo for_each_online_cpu(cpu) { 60613347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 60623347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 60633347fa09STejun Heo BUG_ON(!create_worker(pool)); 60643347fa09STejun Heo } 60653347fa09STejun Heo } 60663347fa09STejun Heo 60673347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 60683347fa09STejun Heo BUG_ON(!create_worker(pool)); 60693347fa09STejun Heo 60703347fa09STejun Heo wq_online = true; 607182607adcSTejun Heo wq_watchdog_init(); 60721da177e4SLinus Torvalds } 6073