11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 20b11895c4SLibin * normal work items and the other for high priority ones) and some extra 21b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 22b11895c4SLibin * number of these backing pools is dynamic. 23c54fce6eSTejun Heo * 249a261491SBenjamin Peterson * Please read Documentation/core-api/workqueue.rst for details. 251da177e4SLinus Torvalds */ 261da177e4SLinus Torvalds 279984de1aSPaul Gortmaker #include <linux/export.h> 281da177e4SLinus Torvalds #include <linux/kernel.h> 291da177e4SLinus Torvalds #include <linux/sched.h> 301da177e4SLinus Torvalds #include <linux/init.h> 311da177e4SLinus Torvalds #include <linux/signal.h> 321da177e4SLinus Torvalds #include <linux/completion.h> 331da177e4SLinus Torvalds #include <linux/workqueue.h> 341da177e4SLinus Torvalds #include <linux/slab.h> 351da177e4SLinus Torvalds #include <linux/cpu.h> 361da177e4SLinus Torvalds #include <linux/notifier.h> 371da177e4SLinus Torvalds #include <linux/kthread.h> 381fa44ecaSJames Bottomley #include <linux/hardirq.h> 3946934023SChristoph Lameter #include <linux/mempolicy.h> 40341a5958SRafael J. Wysocki #include <linux/freezer.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 4429c91e99STejun Heo #include <linux/jhash.h> 4542f8570fSSasha Levin #include <linux/hashtable.h> 4676af4d93STejun Heo #include <linux/rculist.h> 47bce90380STejun Heo #include <linux/nodemask.h> 484c16bd32STejun Heo #include <linux/moduleparam.h> 493d1cb205STejun Heo #include <linux/uaccess.h> 50c98a9805STal Shorer #include <linux/sched/isolation.h> 5162635ea8SSergey Senozhatsky #include <linux/nmi.h> 52e22bee78STejun Heo 53ea138446STejun Heo #include "workqueue_internal.h" 541da177e4SLinus Torvalds 55c8e55f36STejun Heo enum { 56bc2ae0f5STejun Heo /* 5724647570STejun Heo * worker_pool flags 58bc2ae0f5STejun Heo * 5924647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 60bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 61bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 62bc2ae0f5STejun Heo * is in effect. 63bc2ae0f5STejun Heo * 64bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 65bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6624647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 67bc2ae0f5STejun Heo * 68bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 6992f9c5c4SLai Jiangshan * attach_mutex to avoid changing binding state while 704736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 71bc2ae0f5STejun Heo */ 72692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7324647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 74db7bccf4STejun Heo 75c8e55f36STejun Heo /* worker flags */ 76c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 77c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 78e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 79fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 80f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 81a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 82e22bee78STejun Heo 83a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 84a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 85db7bccf4STejun Heo 86e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 874ce62e9eSTejun Heo 8829c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 89c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 90db7bccf4STejun Heo 91e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 92e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 93e22bee78STejun Heo 943233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 953233cdbdSTejun Heo /* call for help after 10ms 963233cdbdSTejun Heo (min two ticks) */ 97e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 98e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 991da177e4SLinus Torvalds 1001da177e4SLinus Torvalds /* 101e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1028698a745SDongsheng Yang * all cpus. Give MIN_NICE. 103e22bee78STejun Heo */ 1048698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1058698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 106ecf6881fSTejun Heo 107ecf6881fSTejun Heo WQ_NAME_LEN = 24, 108c8e55f36STejun Heo }; 109c8e55f36STejun Heo 1101da177e4SLinus Torvalds /* 1114690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1124690c4abSTejun Heo * 113e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 114e41e704bSTejun Heo * everyone else. 1154690c4abSTejun Heo * 116e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 117e22bee78STejun Heo * only be modified and accessed from the local cpu. 118e22bee78STejun Heo * 119d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1204690c4abSTejun Heo * 121d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 122d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 123d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 124d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 125e22bee78STejun Heo * 12692f9c5c4SLai Jiangshan * A: pool->attach_mutex protected. 127822d8405STejun Heo * 12868e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 12976af4d93STejun Heo * 13068e13a67SLai Jiangshan * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads. 1315bcab335STejun Heo * 1325b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1335b95e1afSLai Jiangshan * 1345b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 1355b95e1afSLai Jiangshan * sched-RCU for reads. 1365b95e1afSLai Jiangshan * 1373c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1383c25a55dSLai Jiangshan * 139b5927605SLai Jiangshan * WR: wq->mutex protected for writes. Sched-RCU protected for reads. 1402e109a28STejun Heo * 1412e109a28STejun Heo * MD: wq_mayday_lock protected. 1424690c4abSTejun Heo */ 1434690c4abSTejun Heo 1442eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 145c34056a3STejun Heo 146bd7bdd43STejun Heo struct worker_pool { 147d565ed63STejun Heo spinlock_t lock; /* the pool lock */ 148d84ff051STejun Heo int cpu; /* I: the associated cpu */ 149f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1509daf9e67STejun Heo int id; /* I: pool ID */ 15111ebea50STejun Heo unsigned int flags; /* X: flags */ 152bd7bdd43STejun Heo 15382607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15482607adcSTejun Heo 155bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 156bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 157ea1abd61SLai Jiangshan 158ea1abd61SLai Jiangshan /* nr_idle includes the ones off idle_list for rebinding */ 159bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 160bd7bdd43STejun Heo 161bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 162bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 163bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 164bd7bdd43STejun Heo 165c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 166c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 167c9e7cf27STejun Heo /* L: hash of busy workers */ 168c9e7cf27STejun Heo 169bc3a1afcSTejun Heo /* see manage_workers() for details on the two manager mutexes */ 1702607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17192f9c5c4SLai Jiangshan struct mutex attach_mutex; /* attach/detach exclusion */ 17292f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17360f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 174e19e397aSTejun Heo 1757cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 176e19e397aSTejun Heo 1777a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17868e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17968e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1807a4e344cSTejun Heo 181e19e397aSTejun Heo /* 182e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 183e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 184e19e397aSTejun Heo * cacheline. 185e19e397aSTejun Heo */ 186e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18729c91e99STejun Heo 18829c91e99STejun Heo /* 18929c91e99STejun Heo * Destruction of pool is sched-RCU protected to allow dereferences 19029c91e99STejun Heo * from get_work_pool(). 19129c91e99STejun Heo */ 19229c91e99STejun Heo struct rcu_head rcu; 1938b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1948b03ae3cSTejun Heo 1958b03ae3cSTejun Heo /* 196112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 197112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 198112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 199112202d9STejun Heo * number of flag bits. 2001da177e4SLinus Torvalds */ 201112202d9STejun Heo struct pool_workqueue { 202bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2034690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20473f53c4aSTejun Heo int work_color; /* L: current color */ 20573f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2068864b4e5STejun Heo int refcnt; /* L: reference count */ 20773f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20873f53c4aSTejun Heo /* L: nr of in_flight works */ 2091e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 210a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2111e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2123c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2132e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2148864b4e5STejun Heo 2158864b4e5STejun Heo /* 2168864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2178864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 2188864b4e5STejun Heo * itself is also sched-RCU protected so that the first pwq can be 219b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2208864b4e5STejun Heo */ 2218864b4e5STejun Heo struct work_struct unbound_release_work; 2228864b4e5STejun Heo struct rcu_head rcu; 223e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2241da177e4SLinus Torvalds 2251da177e4SLinus Torvalds /* 22673f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22773f53c4aSTejun Heo */ 22873f53c4aSTejun Heo struct wq_flusher { 2293c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2303c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 23173f53c4aSTejun Heo struct completion done; /* flush completion */ 23273f53c4aSTejun Heo }; 2331da177e4SLinus Torvalds 234226223abSTejun Heo struct wq_device; 235226223abSTejun Heo 23673f53c4aSTejun Heo /* 237c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 238c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2391da177e4SLinus Torvalds */ 2401da177e4SLinus Torvalds struct workqueue_struct { 2413c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 242e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 24373f53c4aSTejun Heo 2443c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2453c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2463c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 247112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2483c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2493c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2503c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 25173f53c4aSTejun Heo 2522e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 253e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 254e22bee78STejun Heo 25587fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 256a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 257226223abSTejun Heo 2585b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2595b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2606029a918STejun Heo 261226223abSTejun Heo #ifdef CONFIG_SYSFS 262226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 263226223abSTejun Heo #endif 2644e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2654e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2664e6045f1SJohannes Berg #endif 267ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2682728fd2fSTejun Heo 269e2dca7adSTejun Heo /* 270e2dca7adSTejun Heo * Destruction of workqueue_struct is sched-RCU protected to allow 271e2dca7adSTejun Heo * walking the workqueues list without grabbing wq_pool_mutex. 272e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 273e2dca7adSTejun Heo */ 274e2dca7adSTejun Heo struct rcu_head rcu; 275e2dca7adSTejun Heo 2762728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2772728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2782728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2795b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2801da177e4SLinus Torvalds }; 2811da177e4SLinus Torvalds 282e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 283e904e6c2STejun Heo 284bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 285bce90380STejun Heo /* possible CPUs of each node */ 286bce90380STejun Heo 287d55262c4STejun Heo static bool wq_disable_numa; 288d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 289d55262c4STejun Heo 290cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 291552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 292cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 293cee22a15SViresh Kumar 294863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 2953347fa09STejun Heo 296bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 297bce90380STejun Heo 2984c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 2994c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3004c16bd32STejun Heo 30168e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3022e109a28STejun Heo static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 303692b4825STejun Heo static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */ 3045bcab335STejun Heo 305e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 30668e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3077d19c5ceSTejun Heo 308ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 309ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 310ef557180SMike Galbraith 311ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 312ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 313b05a7928SFrederic Weisbecker 314f303fccbSTejun Heo /* 315f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 316f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 317f303fccbSTejun Heo * to uncover usages which depend on it. 318f303fccbSTejun Heo */ 319f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 320f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 321f303fccbSTejun Heo #else 322f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 323f303fccbSTejun Heo #endif 324f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 325f303fccbSTejun Heo 3267d19c5ceSTejun Heo /* the per-cpu worker pools */ 32725528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3287d19c5ceSTejun Heo 32968e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3307d19c5ceSTejun Heo 33168e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 33229c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 33329c91e99STejun Heo 334c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 33529c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 33629c91e99STejun Heo 3378a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3388a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3398a2b7538STejun Heo 340d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 341ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 342044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3431aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 344044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 345d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 346044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 347f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 348044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 34924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3500668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3510668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3520668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3530668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 354d320c038STejun Heo 3557d19c5ceSTejun Heo static int worker_thread(void *__worker); 3566ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 3577d19c5ceSTejun Heo 35897bd2347STejun Heo #define CREATE_TRACE_POINTS 35997bd2347STejun Heo #include <trace/events/workqueue.h> 36097bd2347STejun Heo 36168e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 362f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 363f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 36468e13a67SLai Jiangshan "sched RCU or wq_pool_mutex should be held") 3655bcab335STejun Heo 366b09f4fd3SLai Jiangshan #define assert_rcu_or_wq_mutex(wq) \ 367f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 368f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex), \ 369b09f4fd3SLai Jiangshan "sched RCU or wq->mutex should be held") 37076af4d93STejun Heo 3715b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 372f78f5b90SPaul E. McKenney RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \ 373f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 374f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 3755b95e1afSLai Jiangshan "sched RCU, wq->mutex or wq_pool_mutex should be held") 3765b95e1afSLai Jiangshan 377f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 378f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 379f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3807a62c2c8STejun Heo (pool)++) 3814ce62e9eSTejun Heo 38249e3cf44STejun Heo /** 38317116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 38417116969STejun Heo * @pool: iteration cursor 385611c92a0STejun Heo * @pi: integer used for iteration 386fa1b54e6STejun Heo * 38768e13a67SLai Jiangshan * This must be called either with wq_pool_mutex held or sched RCU read 38868e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 38968e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 390fa1b54e6STejun Heo * 391fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 392fa1b54e6STejun Heo * ignored. 39317116969STejun Heo */ 394611c92a0STejun Heo #define for_each_pool(pool, pi) \ 395611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 39668e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 397fa1b54e6STejun Heo else 39817116969STejun Heo 39917116969STejun Heo /** 400822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 401822d8405STejun Heo * @worker: iteration cursor 402822d8405STejun Heo * @pool: worker_pool to iterate workers of 403822d8405STejun Heo * 40492f9c5c4SLai Jiangshan * This must be called with @pool->attach_mutex. 405822d8405STejun Heo * 406822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 407822d8405STejun Heo * ignored. 408822d8405STejun Heo */ 409da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 410da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 41192f9c5c4SLai Jiangshan if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \ 412822d8405STejun Heo else 413822d8405STejun Heo 414822d8405STejun Heo /** 41549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 41649e3cf44STejun Heo * @pwq: iteration cursor 41749e3cf44STejun Heo * @wq: the target workqueue 41876af4d93STejun Heo * 419b09f4fd3SLai Jiangshan * This must be called either with wq->mutex held or sched RCU read locked. 420794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 421794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 42276af4d93STejun Heo * 42376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 42476af4d93STejun Heo * ignored. 42549e3cf44STejun Heo */ 42649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 42776af4d93STejun Heo list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \ 428b09f4fd3SLai Jiangshan if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \ 42976af4d93STejun Heo else 430f3421797STejun Heo 431dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 432dc186ad7SThomas Gleixner 433dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 434dc186ad7SThomas Gleixner 43599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 43699777288SStanislaw Gruszka { 43799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 43899777288SStanislaw Gruszka } 43999777288SStanislaw Gruszka 440b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 441b9fdac7fSDu, Changbin { 442b9fdac7fSDu, Changbin struct work_struct *work = addr; 443b9fdac7fSDu, Changbin 444b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 445b9fdac7fSDu, Changbin } 446b9fdac7fSDu, Changbin 447dc186ad7SThomas Gleixner /* 448dc186ad7SThomas Gleixner * fixup_init is called when: 449dc186ad7SThomas Gleixner * - an active object is initialized 450dc186ad7SThomas Gleixner */ 45102a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 452dc186ad7SThomas Gleixner { 453dc186ad7SThomas Gleixner struct work_struct *work = addr; 454dc186ad7SThomas Gleixner 455dc186ad7SThomas Gleixner switch (state) { 456dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 457dc186ad7SThomas Gleixner cancel_work_sync(work); 458dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 45902a982a6SDu, Changbin return true; 460dc186ad7SThomas Gleixner default: 46102a982a6SDu, Changbin return false; 462dc186ad7SThomas Gleixner } 463dc186ad7SThomas Gleixner } 464dc186ad7SThomas Gleixner 465dc186ad7SThomas Gleixner /* 466dc186ad7SThomas Gleixner * fixup_free is called when: 467dc186ad7SThomas Gleixner * - an active object is freed 468dc186ad7SThomas Gleixner */ 46902a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 470dc186ad7SThomas Gleixner { 471dc186ad7SThomas Gleixner struct work_struct *work = addr; 472dc186ad7SThomas Gleixner 473dc186ad7SThomas Gleixner switch (state) { 474dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 475dc186ad7SThomas Gleixner cancel_work_sync(work); 476dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 47702a982a6SDu, Changbin return true; 478dc186ad7SThomas Gleixner default: 47902a982a6SDu, Changbin return false; 480dc186ad7SThomas Gleixner } 481dc186ad7SThomas Gleixner } 482dc186ad7SThomas Gleixner 483dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 484dc186ad7SThomas Gleixner .name = "work_struct", 48599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 486b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 487dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 488dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 489dc186ad7SThomas Gleixner }; 490dc186ad7SThomas Gleixner 491dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 492dc186ad7SThomas Gleixner { 493dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 494dc186ad7SThomas Gleixner } 495dc186ad7SThomas Gleixner 496dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 497dc186ad7SThomas Gleixner { 498dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 499dc186ad7SThomas Gleixner } 500dc186ad7SThomas Gleixner 501dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 502dc186ad7SThomas Gleixner { 503dc186ad7SThomas Gleixner if (onstack) 504dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 505dc186ad7SThomas Gleixner else 506dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 507dc186ad7SThomas Gleixner } 508dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 509dc186ad7SThomas Gleixner 510dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 511dc186ad7SThomas Gleixner { 512dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 513dc186ad7SThomas Gleixner } 514dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 515dc186ad7SThomas Gleixner 516ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 517ea2e64f2SThomas Gleixner { 518ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 519ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 520ea2e64f2SThomas Gleixner } 521ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 522ea2e64f2SThomas Gleixner 523dc186ad7SThomas Gleixner #else 524dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 525dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 526dc186ad7SThomas Gleixner #endif 527dc186ad7SThomas Gleixner 5284e8b22bdSLi Bin /** 5294e8b22bdSLi Bin * worker_pool_assign_id - allocate ID and assing it to @pool 5304e8b22bdSLi Bin * @pool: the pool pointer of interest 5314e8b22bdSLi Bin * 5324e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5334e8b22bdSLi Bin * successfully, -errno on failure. 5344e8b22bdSLi Bin */ 5359daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5369daf9e67STejun Heo { 5379daf9e67STejun Heo int ret; 5389daf9e67STejun Heo 53968e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5405bcab335STejun Heo 5414e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5424e8b22bdSLi Bin GFP_KERNEL); 543229641a6STejun Heo if (ret >= 0) { 544e68035fbSTejun Heo pool->id = ret; 545229641a6STejun Heo return 0; 546229641a6STejun Heo } 5479daf9e67STejun Heo return ret; 5489daf9e67STejun Heo } 5499daf9e67STejun Heo 55076af4d93STejun Heo /** 551df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 552df2d5ae4STejun Heo * @wq: the target workqueue 553df2d5ae4STejun Heo * @node: the node ID 554df2d5ae4STejun Heo * 5555b95e1afSLai Jiangshan * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU 5565b95e1afSLai Jiangshan * read locked. 557df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 558df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 559d185af30SYacine Belkadi * 560d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 561df2d5ae4STejun Heo */ 562df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 563df2d5ae4STejun Heo int node) 564df2d5ae4STejun Heo { 5655b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 566d6e022f1STejun Heo 567d6e022f1STejun Heo /* 568d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 569d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 570d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 571d6e022f1STejun Heo * happens, this workaround can be removed. 572d6e022f1STejun Heo */ 573d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 574d6e022f1STejun Heo return wq->dfl_pwq; 575d6e022f1STejun Heo 576df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 577df2d5ae4STejun Heo } 578df2d5ae4STejun Heo 57973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 58073f53c4aSTejun Heo { 58173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 58273f53c4aSTejun Heo } 58373f53c4aSTejun Heo 58473f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 58573f53c4aSTejun Heo { 58673f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 58773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 58873f53c4aSTejun Heo } 58973f53c4aSTejun Heo 59073f53c4aSTejun Heo static int work_next_color(int color) 59173f53c4aSTejun Heo { 59273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 593a848e3b6SOleg Nesterov } 594a848e3b6SOleg Nesterov 5954594bf15SDavid Howells /* 596112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 597112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 5987c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 5997a22ad75STejun Heo * 600112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 601112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 602bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 603bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6047a22ad75STejun Heo * 605112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6067c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 607112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6087c3eed5cSTejun Heo * available only while the work item is queued. 609bbb68dfaSTejun Heo * 610bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 611bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 612bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 613bbb68dfaSTejun Heo * try to steal the PENDING bit. 6144594bf15SDavid Howells */ 6157a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6167a22ad75STejun Heo unsigned long flags) 6177a22ad75STejun Heo { 6186183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6197a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6207a22ad75STejun Heo } 6217a22ad75STejun Heo 622112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6234690c4abSTejun Heo unsigned long extra_flags) 624365970a1SDavid Howells { 625112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 626112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 627365970a1SDavid Howells } 628365970a1SDavid Howells 6294468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6304468a00fSLai Jiangshan int pool_id) 6314468a00fSLai Jiangshan { 6324468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6334468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6344468a00fSLai Jiangshan } 6354468a00fSLai Jiangshan 6367c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6377c3eed5cSTejun Heo int pool_id) 6384d707b9fSOleg Nesterov { 63923657bb1STejun Heo /* 64023657bb1STejun Heo * The following wmb is paired with the implied mb in 64123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 64223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 64323657bb1STejun Heo * owner. 64423657bb1STejun Heo */ 64523657bb1STejun Heo smp_wmb(); 6467c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 647346c09f8SRoman Pen /* 648346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 649346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 650346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 651346c09f8SRoman Pen * reordering can lead to a missed execution on attempt to qeueue 652346c09f8SRoman Pen * the same @work. E.g. consider this case: 653346c09f8SRoman Pen * 654346c09f8SRoman Pen * CPU#0 CPU#1 655346c09f8SRoman Pen * ---------------------------- -------------------------------- 656346c09f8SRoman Pen * 657346c09f8SRoman Pen * 1 STORE event_indicated 658346c09f8SRoman Pen * 2 queue_work_on() { 659346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 660346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 661346c09f8SRoman Pen * 5 set_work_data() # clear bit 662346c09f8SRoman Pen * 6 smp_mb() 663346c09f8SRoman Pen * 7 work->current_func() { 664346c09f8SRoman Pen * 8 LOAD event_indicated 665346c09f8SRoman Pen * } 666346c09f8SRoman Pen * 667346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 668346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 669346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 670346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 671346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 672346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 673346c09f8SRoman Pen * before actual STORE. 674346c09f8SRoman Pen */ 675346c09f8SRoman Pen smp_mb(); 6764d707b9fSOleg Nesterov } 6774d707b9fSOleg Nesterov 6787a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 679365970a1SDavid Howells { 6807c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6817c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6827a22ad75STejun Heo } 6837a22ad75STejun Heo 684112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 6857a22ad75STejun Heo { 686e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 6877a22ad75STejun Heo 688112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 689e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 690e120153dSTejun Heo else 691e120153dSTejun Heo return NULL; 6927a22ad75STejun Heo } 6937a22ad75STejun Heo 6947c3eed5cSTejun Heo /** 6957c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 6967c3eed5cSTejun Heo * @work: the work item of interest 6977c3eed5cSTejun Heo * 69868e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 69968e13a67SLai Jiangshan * access under sched-RCU read lock. As such, this function should be 70068e13a67SLai Jiangshan * called under wq_pool_mutex or with preemption disabled. 701fa1b54e6STejun Heo * 702fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 703fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 704fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 705fa1b54e6STejun Heo * returned pool is and stays online. 706d185af30SYacine Belkadi * 707d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7087c3eed5cSTejun Heo */ 7097c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7107a22ad75STejun Heo { 711e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7127c3eed5cSTejun Heo int pool_id; 7137a22ad75STejun Heo 71468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 715fa1b54e6STejun Heo 716112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 717112202d9STejun Heo return ((struct pool_workqueue *) 7187c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7197a22ad75STejun Heo 7207c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7217c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7227a22ad75STejun Heo return NULL; 7237a22ad75STejun Heo 724fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7257c3eed5cSTejun Heo } 7267c3eed5cSTejun Heo 7277c3eed5cSTejun Heo /** 7287c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7297c3eed5cSTejun Heo * @work: the work item of interest 7307c3eed5cSTejun Heo * 731d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7327c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7337c3eed5cSTejun Heo */ 7347c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7357c3eed5cSTejun Heo { 73654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7377c3eed5cSTejun Heo 738112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 739112202d9STejun Heo return ((struct pool_workqueue *) 74054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 74154d5b7d0SLai Jiangshan 74254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7437c3eed5cSTejun Heo } 7447c3eed5cSTejun Heo 745bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 746bbb68dfaSTejun Heo { 7477c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 748bbb68dfaSTejun Heo 7497c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7507c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 751bbb68dfaSTejun Heo } 752bbb68dfaSTejun Heo 753bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 754bbb68dfaSTejun Heo { 755bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 756bbb68dfaSTejun Heo 757112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 758bbb68dfaSTejun Heo } 759bbb68dfaSTejun Heo 760e22bee78STejun Heo /* 7613270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7623270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 763d565ed63STejun Heo * they're being called with pool->lock held. 764e22bee78STejun Heo */ 765e22bee78STejun Heo 76663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 767649027d7STejun Heo { 768e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 769649027d7STejun Heo } 770649027d7STejun Heo 771e22bee78STejun Heo /* 772e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 773e22bee78STejun Heo * running workers. 774974271c4STejun Heo * 775974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 776706026c2STejun Heo * function will always return %true for unbound pools as long as the 777974271c4STejun Heo * worklist isn't empty. 778e22bee78STejun Heo */ 77963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 780e22bee78STejun Heo { 78163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 782e22bee78STejun Heo } 783e22bee78STejun Heo 784e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 78563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 786e22bee78STejun Heo { 78763d95a91STejun Heo return pool->nr_idle; 788e22bee78STejun Heo } 789e22bee78STejun Heo 790e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 79163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 792e22bee78STejun Heo { 793e19e397aSTejun Heo return !list_empty(&pool->worklist) && 794e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 795e22bee78STejun Heo } 796e22bee78STejun Heo 797e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 79863d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 799e22bee78STejun Heo { 80063d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 801e22bee78STejun Heo } 802e22bee78STejun Heo 803e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 80463d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 805e22bee78STejun Heo { 806692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 80763d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 80863d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 809e22bee78STejun Heo 810e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 811e22bee78STejun Heo } 812e22bee78STejun Heo 813e22bee78STejun Heo /* 814e22bee78STejun Heo * Wake up functions. 815e22bee78STejun Heo */ 816e22bee78STejun Heo 8171037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8181037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8197e11629dSTejun Heo { 82063d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8217e11629dSTejun Heo return NULL; 8227e11629dSTejun Heo 82363d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8247e11629dSTejun Heo } 8257e11629dSTejun Heo 8267e11629dSTejun Heo /** 8277e11629dSTejun Heo * wake_up_worker - wake up an idle worker 82863d95a91STejun Heo * @pool: worker pool to wake worker from 8297e11629dSTejun Heo * 83063d95a91STejun Heo * Wake up the first idle worker of @pool. 8317e11629dSTejun Heo * 8327e11629dSTejun Heo * CONTEXT: 833d565ed63STejun Heo * spin_lock_irq(pool->lock). 8347e11629dSTejun Heo */ 83563d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8367e11629dSTejun Heo { 8371037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8387e11629dSTejun Heo 8397e11629dSTejun Heo if (likely(worker)) 8407e11629dSTejun Heo wake_up_process(worker->task); 8417e11629dSTejun Heo } 8427e11629dSTejun Heo 8434690c4abSTejun Heo /** 844e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 845e22bee78STejun Heo * @task: task waking up 846e22bee78STejun Heo * @cpu: CPU @task is waking up to 847e22bee78STejun Heo * 848e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 849e22bee78STejun Heo * being awoken. 850e22bee78STejun Heo * 851e22bee78STejun Heo * CONTEXT: 852e22bee78STejun Heo * spin_lock_irq(rq->lock) 853e22bee78STejun Heo */ 854d84ff051STejun Heo void wq_worker_waking_up(struct task_struct *task, int cpu) 855e22bee78STejun Heo { 856e22bee78STejun Heo struct worker *worker = kthread_data(task); 857e22bee78STejun Heo 85836576000SJoonsoo Kim if (!(worker->flags & WORKER_NOT_RUNNING)) { 859ec22ca5eSTejun Heo WARN_ON_ONCE(worker->pool->cpu != cpu); 860e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 861e22bee78STejun Heo } 86236576000SJoonsoo Kim } 863e22bee78STejun Heo 864e22bee78STejun Heo /** 865e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 866e22bee78STejun Heo * @task: task going to sleep 867e22bee78STejun Heo * 868e22bee78STejun Heo * This function is called during schedule() when a busy worker is 869e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 870e22bee78STejun Heo * returning pointer to its task. 871e22bee78STejun Heo * 872e22bee78STejun Heo * CONTEXT: 873e22bee78STejun Heo * spin_lock_irq(rq->lock) 874e22bee78STejun Heo * 875d185af30SYacine Belkadi * Return: 876e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 877e22bee78STejun Heo */ 8789b7f6597SAlexander Gordeev struct task_struct *wq_worker_sleeping(struct task_struct *task) 879e22bee78STejun Heo { 880e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 881111c225aSTejun Heo struct worker_pool *pool; 882e22bee78STejun Heo 883111c225aSTejun Heo /* 884111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 885111c225aSTejun Heo * workers, also reach here, let's not access anything before 886111c225aSTejun Heo * checking NOT_RUNNING. 887111c225aSTejun Heo */ 8882d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 889e22bee78STejun Heo return NULL; 890e22bee78STejun Heo 891111c225aSTejun Heo pool = worker->pool; 892111c225aSTejun Heo 893e22bee78STejun Heo /* this can only happen on the local cpu */ 8949b7f6597SAlexander Gordeev if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id())) 8956183c009STejun Heo return NULL; 896e22bee78STejun Heo 897e22bee78STejun Heo /* 898e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 899e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 900e22bee78STejun Heo * Please read comment there. 901e22bee78STejun Heo * 902628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 903628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 904628c78e7STejun Heo * disabled, which in turn means that none else could be 905d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 906628c78e7STejun Heo * lock is safe. 907e22bee78STejun Heo */ 908e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 909e19e397aSTejun Heo !list_empty(&pool->worklist)) 9101037de36SLai Jiangshan to_wakeup = first_idle_worker(pool); 911e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 912e22bee78STejun Heo } 913e22bee78STejun Heo 914e22bee78STejun Heo /** 915e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 916cb444766STejun Heo * @worker: self 917d302f017STejun Heo * @flags: flags to set 918d302f017STejun Heo * 919228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 920d302f017STejun Heo * 921cb444766STejun Heo * CONTEXT: 922d565ed63STejun Heo * spin_lock_irq(pool->lock) 923d302f017STejun Heo */ 924228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 925d302f017STejun Heo { 926bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 927e22bee78STejun Heo 928cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 929cb444766STejun Heo 930228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 931e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 932e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 933e19e397aSTejun Heo atomic_dec(&pool->nr_running); 934e22bee78STejun Heo } 935e22bee78STejun Heo 936d302f017STejun Heo worker->flags |= flags; 937d302f017STejun Heo } 938d302f017STejun Heo 939d302f017STejun Heo /** 940e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 941cb444766STejun Heo * @worker: self 942d302f017STejun Heo * @flags: flags to clear 943d302f017STejun Heo * 944e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 945d302f017STejun Heo * 946cb444766STejun Heo * CONTEXT: 947d565ed63STejun Heo * spin_lock_irq(pool->lock) 948d302f017STejun Heo */ 949d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 950d302f017STejun Heo { 95163d95a91STejun Heo struct worker_pool *pool = worker->pool; 952e22bee78STejun Heo unsigned int oflags = worker->flags; 953e22bee78STejun Heo 954cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 955cb444766STejun Heo 956d302f017STejun Heo worker->flags &= ~flags; 957e22bee78STejun Heo 95842c025f3STejun Heo /* 95942c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 96042c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 96142c025f3STejun Heo * of multiple flags, not a single flag. 96242c025f3STejun Heo */ 963e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 964e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 965e19e397aSTejun Heo atomic_inc(&pool->nr_running); 966d302f017STejun Heo } 967d302f017STejun Heo 968d302f017STejun Heo /** 9698cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 970c9e7cf27STejun Heo * @pool: pool of interest 9718cca0eeaSTejun Heo * @work: work to find worker for 9728cca0eeaSTejun Heo * 973c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 974c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 975a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 976a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 977a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 978a2c1c57bSTejun Heo * being executed. 979a2c1c57bSTejun Heo * 980a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 981a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 982a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 983a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 984a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 985a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 986a2c1c57bSTejun Heo * 987c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 988c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 989c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 990c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 991c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 992c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 9938cca0eeaSTejun Heo * 9948cca0eeaSTejun Heo * CONTEXT: 995d565ed63STejun Heo * spin_lock_irq(pool->lock). 9968cca0eeaSTejun Heo * 997d185af30SYacine Belkadi * Return: 998d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 9998cca0eeaSTejun Heo * otherwise. 10008cca0eeaSTejun Heo */ 1001c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10028cca0eeaSTejun Heo struct work_struct *work) 10038cca0eeaSTejun Heo { 100442f8570fSSasha Levin struct worker *worker; 100542f8570fSSasha Levin 1006b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1007a2c1c57bSTejun Heo (unsigned long)work) 1008a2c1c57bSTejun Heo if (worker->current_work == work && 1009a2c1c57bSTejun Heo worker->current_func == work->func) 101042f8570fSSasha Levin return worker; 101142f8570fSSasha Levin 101242f8570fSSasha Levin return NULL; 10138cca0eeaSTejun Heo } 10148cca0eeaSTejun Heo 10158cca0eeaSTejun Heo /** 1016bf4ede01STejun Heo * move_linked_works - move linked works to a list 1017bf4ede01STejun Heo * @work: start of series of works to be scheduled 1018bf4ede01STejun Heo * @head: target list to append @work to 1019402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1020bf4ede01STejun Heo * 1021bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1022bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1023bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1024bf4ede01STejun Heo * 1025bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1026bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1027bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1028bf4ede01STejun Heo * 1029bf4ede01STejun Heo * CONTEXT: 1030d565ed63STejun Heo * spin_lock_irq(pool->lock). 1031bf4ede01STejun Heo */ 1032bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1033bf4ede01STejun Heo struct work_struct **nextp) 1034bf4ede01STejun Heo { 1035bf4ede01STejun Heo struct work_struct *n; 1036bf4ede01STejun Heo 1037bf4ede01STejun Heo /* 1038bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1039bf4ede01STejun Heo * use NULL for list head. 1040bf4ede01STejun Heo */ 1041bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1042bf4ede01STejun Heo list_move_tail(&work->entry, head); 1043bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1044bf4ede01STejun Heo break; 1045bf4ede01STejun Heo } 1046bf4ede01STejun Heo 1047bf4ede01STejun Heo /* 1048bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1049bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1050bf4ede01STejun Heo * needs to be updated. 1051bf4ede01STejun Heo */ 1052bf4ede01STejun Heo if (nextp) 1053bf4ede01STejun Heo *nextp = n; 1054bf4ede01STejun Heo } 1055bf4ede01STejun Heo 10568864b4e5STejun Heo /** 10578864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 10588864b4e5STejun Heo * @pwq: pool_workqueue to get 10598864b4e5STejun Heo * 10608864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 10618864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 10628864b4e5STejun Heo */ 10638864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 10648864b4e5STejun Heo { 10658864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10668864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 10678864b4e5STejun Heo pwq->refcnt++; 10688864b4e5STejun Heo } 10698864b4e5STejun Heo 10708864b4e5STejun Heo /** 10718864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 10728864b4e5STejun Heo * @pwq: pool_workqueue to put 10738864b4e5STejun Heo * 10748864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 10758864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 10768864b4e5STejun Heo */ 10778864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 10788864b4e5STejun Heo { 10798864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 10808864b4e5STejun Heo if (likely(--pwq->refcnt)) 10818864b4e5STejun Heo return; 10828864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 10838864b4e5STejun Heo return; 10848864b4e5STejun Heo /* 10858864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 10868864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 10878864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 10888864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 10898864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 10908864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 10918864b4e5STejun Heo */ 10928864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 10938864b4e5STejun Heo } 10948864b4e5STejun Heo 1095dce90d47STejun Heo /** 1096dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1097dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1098dce90d47STejun Heo * 1099dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1100dce90d47STejun Heo */ 1101dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1102dce90d47STejun Heo { 1103dce90d47STejun Heo if (pwq) { 1104dce90d47STejun Heo /* 1105dce90d47STejun Heo * As both pwqs and pools are sched-RCU protected, the 1106dce90d47STejun Heo * following lock operations are safe. 1107dce90d47STejun Heo */ 1108dce90d47STejun Heo spin_lock_irq(&pwq->pool->lock); 1109dce90d47STejun Heo put_pwq(pwq); 1110dce90d47STejun Heo spin_unlock_irq(&pwq->pool->lock); 1111dce90d47STejun Heo } 1112dce90d47STejun Heo } 1113dce90d47STejun Heo 1114112202d9STejun Heo static void pwq_activate_delayed_work(struct work_struct *work) 1115bf4ede01STejun Heo { 1116112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1117bf4ede01STejun Heo 1118bf4ede01STejun Heo trace_workqueue_activate_work(work); 111982607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 112082607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1121112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1122bf4ede01STejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 1123112202d9STejun Heo pwq->nr_active++; 1124bf4ede01STejun Heo } 1125bf4ede01STejun Heo 1126112202d9STejun Heo static void pwq_activate_first_delayed(struct pool_workqueue *pwq) 11273aa62497SLai Jiangshan { 1128112202d9STejun Heo struct work_struct *work = list_first_entry(&pwq->delayed_works, 11293aa62497SLai Jiangshan struct work_struct, entry); 11303aa62497SLai Jiangshan 1131112202d9STejun Heo pwq_activate_delayed_work(work); 11323aa62497SLai Jiangshan } 11333aa62497SLai Jiangshan 1134bf4ede01STejun Heo /** 1135112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1136112202d9STejun Heo * @pwq: pwq of interest 1137bf4ede01STejun Heo * @color: color of work which left the queue 1138bf4ede01STejun Heo * 1139bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1140112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1141bf4ede01STejun Heo * 1142bf4ede01STejun Heo * CONTEXT: 1143d565ed63STejun Heo * spin_lock_irq(pool->lock). 1144bf4ede01STejun Heo */ 1145112202d9STejun Heo static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color) 1146bf4ede01STejun Heo { 11478864b4e5STejun Heo /* uncolored work items don't participate in flushing or nr_active */ 1148bf4ede01STejun Heo if (color == WORK_NO_COLOR) 11498864b4e5STejun Heo goto out_put; 1150bf4ede01STejun Heo 1151112202d9STejun Heo pwq->nr_in_flight[color]--; 1152bf4ede01STejun Heo 1153112202d9STejun Heo pwq->nr_active--; 1154112202d9STejun Heo if (!list_empty(&pwq->delayed_works)) { 1155bf4ede01STejun Heo /* one down, submit a delayed one */ 1156112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1157112202d9STejun Heo pwq_activate_first_delayed(pwq); 1158bf4ede01STejun Heo } 1159bf4ede01STejun Heo 1160bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1161112202d9STejun Heo if (likely(pwq->flush_color != color)) 11628864b4e5STejun Heo goto out_put; 1163bf4ede01STejun Heo 1164bf4ede01STejun Heo /* are there still in-flight works? */ 1165112202d9STejun Heo if (pwq->nr_in_flight[color]) 11668864b4e5STejun Heo goto out_put; 1167bf4ede01STejun Heo 1168112202d9STejun Heo /* this pwq is done, clear flush_color */ 1169112202d9STejun Heo pwq->flush_color = -1; 1170bf4ede01STejun Heo 1171bf4ede01STejun Heo /* 1172112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1173bf4ede01STejun Heo * will handle the rest. 1174bf4ede01STejun Heo */ 1175112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1176112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 11778864b4e5STejun Heo out_put: 11788864b4e5STejun Heo put_pwq(pwq); 1179bf4ede01STejun Heo } 1180bf4ede01STejun Heo 118136e227d2STejun Heo /** 1182bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 118336e227d2STejun Heo * @work: work item to steal 118436e227d2STejun Heo * @is_dwork: @work is a delayed_work 1185bbb68dfaSTejun Heo * @flags: place to store irq state 118636e227d2STejun Heo * 118736e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1188d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 118936e227d2STejun Heo * 1190d185af30SYacine Belkadi * Return: 119136e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 119236e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 119336e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1194bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1195bbb68dfaSTejun Heo * for arbitrarily long 119636e227d2STejun Heo * 1197d185af30SYacine Belkadi * Note: 1198bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1199e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1200e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1201e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1202bbb68dfaSTejun Heo * 1203bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1204bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1205bbb68dfaSTejun Heo * 1206e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1207bf4ede01STejun Heo */ 1208bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1209bbb68dfaSTejun Heo unsigned long *flags) 1210bf4ede01STejun Heo { 1211d565ed63STejun Heo struct worker_pool *pool; 1212112202d9STejun Heo struct pool_workqueue *pwq; 1213bf4ede01STejun Heo 1214bbb68dfaSTejun Heo local_irq_save(*flags); 1215bbb68dfaSTejun Heo 121636e227d2STejun Heo /* try to steal the timer if it exists */ 121736e227d2STejun Heo if (is_dwork) { 121836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 121936e227d2STejun Heo 1220e0aecdd8STejun Heo /* 1221e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1222e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1223e0aecdd8STejun Heo * running on the local CPU. 1224e0aecdd8STejun Heo */ 122536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 122636e227d2STejun Heo return 1; 122736e227d2STejun Heo } 122836e227d2STejun Heo 122936e227d2STejun Heo /* try to claim PENDING the normal way */ 1230bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1231bf4ede01STejun Heo return 0; 1232bf4ede01STejun Heo 1233bf4ede01STejun Heo /* 1234bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1235bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1236bf4ede01STejun Heo */ 1237d565ed63STejun Heo pool = get_work_pool(work); 1238d565ed63STejun Heo if (!pool) 1239bbb68dfaSTejun Heo goto fail; 1240bf4ede01STejun Heo 1241d565ed63STejun Heo spin_lock(&pool->lock); 1242bf4ede01STejun Heo /* 1243112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1244112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1245112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1246112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1247112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12480b3dae68SLai Jiangshan * item is currently queued on that pool. 1249bf4ede01STejun Heo */ 1250112202d9STejun Heo pwq = get_work_pwq(work); 1251112202d9STejun Heo if (pwq && pwq->pool == pool) { 1252bf4ede01STejun Heo debug_work_deactivate(work); 12533aa62497SLai Jiangshan 12543aa62497SLai Jiangshan /* 125516062836STejun Heo * A delayed work item cannot be grabbed directly because 125616062836STejun Heo * it might have linked NO_COLOR work items which, if left 1257112202d9STejun Heo * on the delayed_list, will confuse pwq->nr_active 125816062836STejun Heo * management later on and cause stall. Make sure the work 125916062836STejun Heo * item is activated before grabbing. 12603aa62497SLai Jiangshan */ 12613aa62497SLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_DELAYED) 1262112202d9STejun Heo pwq_activate_delayed_work(work); 12633aa62497SLai Jiangshan 1264bf4ede01STejun Heo list_del_init(&work->entry); 12659c34a704SLai Jiangshan pwq_dec_nr_in_flight(pwq, get_work_color(work)); 126636e227d2STejun Heo 1267112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 12684468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 12694468a00fSLai Jiangshan 1270d565ed63STejun Heo spin_unlock(&pool->lock); 127136e227d2STejun Heo return 1; 1272bf4ede01STejun Heo } 1273d565ed63STejun Heo spin_unlock(&pool->lock); 1274bbb68dfaSTejun Heo fail: 1275bbb68dfaSTejun Heo local_irq_restore(*flags); 1276bbb68dfaSTejun Heo if (work_is_canceling(work)) 1277bbb68dfaSTejun Heo return -ENOENT; 1278bbb68dfaSTejun Heo cpu_relax(); 127936e227d2STejun Heo return -EAGAIN; 1280bf4ede01STejun Heo } 1281bf4ede01STejun Heo 1282bf4ede01STejun Heo /** 1283706026c2STejun Heo * insert_work - insert a work into a pool 1284112202d9STejun Heo * @pwq: pwq @work belongs to 12854690c4abSTejun Heo * @work: work to insert 12864690c4abSTejun Heo * @head: insertion point 12874690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 12884690c4abSTejun Heo * 1289112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1290706026c2STejun Heo * work_struct flags. 12914690c4abSTejun Heo * 12924690c4abSTejun Heo * CONTEXT: 1293d565ed63STejun Heo * spin_lock_irq(pool->lock). 1294365970a1SDavid Howells */ 1295112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1296112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1297b89deed3SOleg Nesterov { 1298112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1299e1d8aa9fSFrederic Weisbecker 13004690c4abSTejun Heo /* we own @work, set data and link */ 1301112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13021a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13038864b4e5STejun Heo get_pwq(pwq); 1304e22bee78STejun Heo 1305e22bee78STejun Heo /* 1306c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1307c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1308c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1309e22bee78STejun Heo */ 1310e22bee78STejun Heo smp_mb(); 1311e22bee78STejun Heo 131263d95a91STejun Heo if (__need_more_worker(pool)) 131363d95a91STejun Heo wake_up_worker(pool); 1314b89deed3SOleg Nesterov } 1315b89deed3SOleg Nesterov 1316c8efcc25STejun Heo /* 1317c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13188d03ecfeSTejun Heo * same workqueue. 1319c8efcc25STejun Heo */ 1320c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1321c8efcc25STejun Heo { 1322c8efcc25STejun Heo struct worker *worker; 1323c8efcc25STejun Heo 13248d03ecfeSTejun Heo worker = current_wq_worker(); 1325c8efcc25STejun Heo /* 13268d03ecfeSTejun Heo * Return %true iff I'm a worker execuing a work item on @wq. If 13278d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1328c8efcc25STejun Heo */ 1329112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1330c8efcc25STejun Heo } 1331c8efcc25STejun Heo 1332ef557180SMike Galbraith /* 1333ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1334ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1335ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1336ef557180SMike Galbraith */ 1337ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1338ef557180SMike Galbraith { 1339f303fccbSTejun Heo static bool printed_dbg_warning; 1340ef557180SMike Galbraith int new_cpu; 1341ef557180SMike Galbraith 1342f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1343ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1344ef557180SMike Galbraith return cpu; 1345f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1346f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1347f303fccbSTejun Heo printed_dbg_warning = true; 1348f303fccbSTejun Heo } 1349f303fccbSTejun Heo 1350ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1351ef557180SMike Galbraith return cpu; 1352ef557180SMike Galbraith 1353ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1354ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1355ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1356ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1357ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1358ef557180SMike Galbraith return cpu; 1359ef557180SMike Galbraith } 1360ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1361ef557180SMike Galbraith 1362ef557180SMike Galbraith return new_cpu; 1363ef557180SMike Galbraith } 1364ef557180SMike Galbraith 1365d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 13661da177e4SLinus Torvalds struct work_struct *work) 13671da177e4SLinus Torvalds { 1368112202d9STejun Heo struct pool_workqueue *pwq; 1369c9178087STejun Heo struct worker_pool *last_pool; 13701e19ffc6STejun Heo struct list_head *worklist; 13718a2e8e5dSTejun Heo unsigned int work_flags; 1372b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 13738930cabaSTejun Heo 13748930cabaSTejun Heo /* 13758930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 13768930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 13778930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 13788930cabaSTejun Heo * happen with IRQ disabled. 13798930cabaSTejun Heo */ 13808e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 13811da177e4SLinus Torvalds 1382dc186ad7SThomas Gleixner debug_work_activate(work); 13831e19ffc6STejun Heo 13849ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1385618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1386c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1387e41e704bSTejun Heo return; 13889e8cd2f5STejun Heo retry: 1389df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1390ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1391df2d5ae4STejun Heo 1392df2d5ae4STejun Heo /* pwq which will be used unless @work is executing elsewhere */ 1393df2d5ae4STejun Heo if (!(wq->flags & WQ_UNBOUND)) 1394c9178087STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1395df2d5ae4STejun Heo else 1396df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1397f3421797STejun Heo 139818aa9effSTejun Heo /* 1399c9178087STejun Heo * If @work was previously on a different pool, it might still be 1400c9178087STejun Heo * running there, in which case the work needs to be queued on that 1401c9178087STejun Heo * pool to guarantee non-reentrancy. 140218aa9effSTejun Heo */ 1403c9e7cf27STejun Heo last_pool = get_work_pool(work); 1404112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 140518aa9effSTejun Heo struct worker *worker; 140618aa9effSTejun Heo 1407d565ed63STejun Heo spin_lock(&last_pool->lock); 140818aa9effSTejun Heo 1409c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 141018aa9effSTejun Heo 1411112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1412c9178087STejun Heo pwq = worker->current_pwq; 14138594fadeSLai Jiangshan } else { 141418aa9effSTejun Heo /* meh... not running there, queue here */ 1415d565ed63STejun Heo spin_unlock(&last_pool->lock); 1416112202d9STejun Heo spin_lock(&pwq->pool->lock); 141718aa9effSTejun Heo } 14188930cabaSTejun Heo } else { 1419112202d9STejun Heo spin_lock(&pwq->pool->lock); 14208930cabaSTejun Heo } 1421502ca9d8STejun Heo 14229e8cd2f5STejun Heo /* 14239e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14249e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14259e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1426df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1427df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14289e8cd2f5STejun Heo * make forward-progress. 14299e8cd2f5STejun Heo */ 14309e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14319e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 14329e8cd2f5STejun Heo spin_unlock(&pwq->pool->lock); 14339e8cd2f5STejun Heo cpu_relax(); 14349e8cd2f5STejun Heo goto retry; 14359e8cd2f5STejun Heo } 14369e8cd2f5STejun Heo /* oops */ 14379e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14389e8cd2f5STejun Heo wq->name, cpu); 14399e8cd2f5STejun Heo } 14409e8cd2f5STejun Heo 1441112202d9STejun Heo /* pwq determined, queue */ 1442112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1443502ca9d8STejun Heo 1444f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1445112202d9STejun Heo spin_unlock(&pwq->pool->lock); 1446f5b2552bSDan Carpenter return; 1447f5b2552bSDan Carpenter } 14481e19ffc6STejun Heo 1449112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1450112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 14511e19ffc6STejun Heo 1452112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1453cdadf009STejun Heo trace_workqueue_activate_work(work); 1454112202d9STejun Heo pwq->nr_active++; 1455112202d9STejun Heo worklist = &pwq->pool->worklist; 145682607adcSTejun Heo if (list_empty(worklist)) 145782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 14588a2e8e5dSTejun Heo } else { 14598a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 1460112202d9STejun Heo worklist = &pwq->delayed_works; 14618a2e8e5dSTejun Heo } 14621e19ffc6STejun Heo 1463112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 14641e19ffc6STejun Heo 1465112202d9STejun Heo spin_unlock(&pwq->pool->lock); 14661da177e4SLinus Torvalds } 14671da177e4SLinus Torvalds 14680fcb78c2SRolf Eike Beer /** 1469c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1470c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1471c1a220e7SZhang Rui * @wq: workqueue to use 1472c1a220e7SZhang Rui * @work: work to queue 1473c1a220e7SZhang Rui * 1474c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1475c1a220e7SZhang Rui * can't go away. 1476d185af30SYacine Belkadi * 1477d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1478c1a220e7SZhang Rui */ 1479d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1480d4283e93STejun Heo struct work_struct *work) 1481c1a220e7SZhang Rui { 1482d4283e93STejun Heo bool ret = false; 14838930cabaSTejun Heo unsigned long flags; 14848930cabaSTejun Heo 14858930cabaSTejun Heo local_irq_save(flags); 1486c1a220e7SZhang Rui 148722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 14884690c4abSTejun Heo __queue_work(cpu, wq, work); 1489d4283e93STejun Heo ret = true; 1490c1a220e7SZhang Rui } 14918930cabaSTejun Heo 14928930cabaSTejun Heo local_irq_restore(flags); 1493c1a220e7SZhang Rui return ret; 1494c1a220e7SZhang Rui } 1495ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1496c1a220e7SZhang Rui 14978c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 14981da177e4SLinus Torvalds { 14998c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 15001da177e4SLinus Torvalds 1501e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 150260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 15031da177e4SLinus Torvalds } 15041438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 15051da177e4SLinus Torvalds 15067beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 150752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15081da177e4SLinus Torvalds { 15097beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 15107beb2edfSTejun Heo struct work_struct *work = &dwork->work; 15111da177e4SLinus Torvalds 1512637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 1513841b86f3SKees Cook WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1514fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1515fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 15167beb2edfSTejun Heo 15178852aac2STejun Heo /* 15188852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 15198852aac2STejun Heo * both optimization and correctness. The earliest @timer can 15208852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 15218852aac2STejun Heo * on that there's no such delay when @delay is 0. 15228852aac2STejun Heo */ 15238852aac2STejun Heo if (!delay) { 15248852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 15258852aac2STejun Heo return; 15268852aac2STejun Heo } 15278852aac2STejun Heo 152860c057bcSLai Jiangshan dwork->wq = wq; 15291265057fSTejun Heo dwork->cpu = cpu; 15307beb2edfSTejun Heo timer->expires = jiffies + delay; 15317beb2edfSTejun Heo 1532041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 15337beb2edfSTejun Heo add_timer_on(timer, cpu); 1534041bd12eSTejun Heo else 1535041bd12eSTejun Heo add_timer(timer); 15367beb2edfSTejun Heo } 15371da177e4SLinus Torvalds 15380fcb78c2SRolf Eike Beer /** 15390fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 15400fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 15410fcb78c2SRolf Eike Beer * @wq: workqueue to use 1542af9997e4SRandy Dunlap * @dwork: work to queue 15430fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 15440fcb78c2SRolf Eike Beer * 1545d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1546715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1547715f1300STejun Heo * execution. 15480fcb78c2SRolf Eike Beer */ 1549d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 155052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 15517a6bc1cdSVenkatesh Pallipadi { 155252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1553d4283e93STejun Heo bool ret = false; 15548930cabaSTejun Heo unsigned long flags; 15558930cabaSTejun Heo 15568930cabaSTejun Heo /* read the comment in __queue_work() */ 15578930cabaSTejun Heo local_irq_save(flags); 15587a6bc1cdSVenkatesh Pallipadi 155922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15607beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1561d4283e93STejun Heo ret = true; 15627a6bc1cdSVenkatesh Pallipadi } 15638930cabaSTejun Heo 15648930cabaSTejun Heo local_irq_restore(flags); 15657a6bc1cdSVenkatesh Pallipadi return ret; 15667a6bc1cdSVenkatesh Pallipadi } 1567ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 15681da177e4SLinus Torvalds 1569c8e55f36STejun Heo /** 15708376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 15718376fe22STejun Heo * @cpu: CPU number to execute work on 15728376fe22STejun Heo * @wq: workqueue to use 15738376fe22STejun Heo * @dwork: work to queue 15748376fe22STejun Heo * @delay: number of jiffies to wait before queueing 15758376fe22STejun Heo * 15768376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 15778376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 15788376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 15798376fe22STejun Heo * current state. 15808376fe22STejun Heo * 1581d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 15828376fe22STejun Heo * pending and its timer was modified. 15838376fe22STejun Heo * 1584e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 15858376fe22STejun Heo * See try_to_grab_pending() for details. 15868376fe22STejun Heo */ 15878376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 15888376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 15898376fe22STejun Heo { 15908376fe22STejun Heo unsigned long flags; 15918376fe22STejun Heo int ret; 15928376fe22STejun Heo 15938376fe22STejun Heo do { 15948376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 15958376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 15968376fe22STejun Heo 15978376fe22STejun Heo if (likely(ret >= 0)) { 15988376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 15998376fe22STejun Heo local_irq_restore(flags); 16008376fe22STejun Heo } 16018376fe22STejun Heo 16028376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 16038376fe22STejun Heo return ret; 16048376fe22STejun Heo } 16058376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 16068376fe22STejun Heo 160705f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 160805f0fe6bSTejun Heo { 160905f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 161005f0fe6bSTejun Heo 161105f0fe6bSTejun Heo /* read the comment in __queue_work() */ 161205f0fe6bSTejun Heo local_irq_disable(); 161305f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 161405f0fe6bSTejun Heo local_irq_enable(); 161505f0fe6bSTejun Heo } 161605f0fe6bSTejun Heo 161705f0fe6bSTejun Heo /** 161805f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 161905f0fe6bSTejun Heo * @wq: workqueue to use 162005f0fe6bSTejun Heo * @rwork: work to queue 162105f0fe6bSTejun Heo * 162205f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 162305f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 162405f0fe6bSTejun Heo * While @rwork is guarnateed to be executed after a %false return, the 162505f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 162605f0fe6bSTejun Heo */ 162705f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 162805f0fe6bSTejun Heo { 162905f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 163005f0fe6bSTejun Heo 163105f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 163205f0fe6bSTejun Heo rwork->wq = wq; 163305f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 163405f0fe6bSTejun Heo return true; 163505f0fe6bSTejun Heo } 163605f0fe6bSTejun Heo 163705f0fe6bSTejun Heo return false; 163805f0fe6bSTejun Heo } 163905f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 164005f0fe6bSTejun Heo 16418376fe22STejun Heo /** 1642c8e55f36STejun Heo * worker_enter_idle - enter idle state 1643c8e55f36STejun Heo * @worker: worker which is entering idle state 1644c8e55f36STejun Heo * 1645c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1646c8e55f36STejun Heo * necessary. 1647c8e55f36STejun Heo * 1648c8e55f36STejun Heo * LOCKING: 1649d565ed63STejun Heo * spin_lock_irq(pool->lock). 1650c8e55f36STejun Heo */ 1651c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 16521da177e4SLinus Torvalds { 1653bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1654c8e55f36STejun Heo 16556183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 16566183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 16576183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 16586183c009STejun Heo return; 1659c8e55f36STejun Heo 1660051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1661cb444766STejun Heo worker->flags |= WORKER_IDLE; 1662bd7bdd43STejun Heo pool->nr_idle++; 1663e22bee78STejun Heo worker->last_active = jiffies; 1664c8e55f36STejun Heo 1665c8e55f36STejun Heo /* idle_list is LIFO */ 1666bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1667db7bccf4STejun Heo 166863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1669628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1670cb444766STejun Heo 1671544ecf31STejun Heo /* 1672e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1673d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1674628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1675628c78e7STejun Heo * unbind is not in progress. 1676544ecf31STejun Heo */ 167724647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1678bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1679e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1680c8e55f36STejun Heo } 1681c8e55f36STejun Heo 1682c8e55f36STejun Heo /** 1683c8e55f36STejun Heo * worker_leave_idle - leave idle state 1684c8e55f36STejun Heo * @worker: worker which is leaving idle state 1685c8e55f36STejun Heo * 1686c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1687c8e55f36STejun Heo * 1688c8e55f36STejun Heo * LOCKING: 1689d565ed63STejun Heo * spin_lock_irq(pool->lock). 1690c8e55f36STejun Heo */ 1691c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1692c8e55f36STejun Heo { 1693bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1694c8e55f36STejun Heo 16956183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 16966183c009STejun Heo return; 1697d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1698bd7bdd43STejun Heo pool->nr_idle--; 1699c8e55f36STejun Heo list_del_init(&worker->entry); 1700c8e55f36STejun Heo } 1701c8e55f36STejun Heo 1702f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1703c34056a3STejun Heo { 1704c34056a3STejun Heo struct worker *worker; 1705c34056a3STejun Heo 1706f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1707c8e55f36STejun Heo if (worker) { 1708c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1709affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1710da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1711e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1712e22bee78STejun Heo worker->flags = WORKER_PREP; 1713c8e55f36STejun Heo } 1714c34056a3STejun Heo return worker; 1715c34056a3STejun Heo } 1716c34056a3STejun Heo 1717c34056a3STejun Heo /** 17184736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 17194736cbf7SLai Jiangshan * @worker: worker to be attached 17204736cbf7SLai Jiangshan * @pool: the target pool 17214736cbf7SLai Jiangshan * 17224736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 17234736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 17244736cbf7SLai Jiangshan * cpu-[un]hotplugs. 17254736cbf7SLai Jiangshan */ 17264736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 17274736cbf7SLai Jiangshan struct worker_pool *pool) 17284736cbf7SLai Jiangshan { 17294736cbf7SLai Jiangshan mutex_lock(&pool->attach_mutex); 17304736cbf7SLai Jiangshan 17314736cbf7SLai Jiangshan /* 17324736cbf7SLai Jiangshan * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any 17334736cbf7SLai Jiangshan * online CPUs. It'll be re-applied when any of the CPUs come up. 17344736cbf7SLai Jiangshan */ 17354736cbf7SLai Jiangshan set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 17364736cbf7SLai Jiangshan 17374736cbf7SLai Jiangshan /* 17384736cbf7SLai Jiangshan * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains 17394736cbf7SLai Jiangshan * stable across this function. See the comments above the 17404736cbf7SLai Jiangshan * flag definition for details. 17414736cbf7SLai Jiangshan */ 17424736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 17434736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 17444736cbf7SLai Jiangshan 17454736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 17464736cbf7SLai Jiangshan 17474736cbf7SLai Jiangshan mutex_unlock(&pool->attach_mutex); 17484736cbf7SLai Jiangshan } 17494736cbf7SLai Jiangshan 17504736cbf7SLai Jiangshan /** 175160f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 175260f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 175360f5a4bcSLai Jiangshan * @pool: the pool @worker is attached to 175460f5a4bcSLai Jiangshan * 17554736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 17564736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 17574736cbf7SLai Jiangshan * other reference to the pool. 175860f5a4bcSLai Jiangshan */ 175960f5a4bcSLai Jiangshan static void worker_detach_from_pool(struct worker *worker, 176060f5a4bcSLai Jiangshan struct worker_pool *pool) 176160f5a4bcSLai Jiangshan { 176260f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 176360f5a4bcSLai Jiangshan 176492f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 1765da028469SLai Jiangshan list_del(&worker->node); 1766da028469SLai Jiangshan if (list_empty(&pool->workers)) 176760f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 176892f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 176960f5a4bcSLai Jiangshan 1770b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1771b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1772b62c0751SLai Jiangshan 177360f5a4bcSLai Jiangshan if (detach_completion) 177460f5a4bcSLai Jiangshan complete(detach_completion); 177560f5a4bcSLai Jiangshan } 177660f5a4bcSLai Jiangshan 177760f5a4bcSLai Jiangshan /** 1778c34056a3STejun Heo * create_worker - create a new workqueue worker 177963d95a91STejun Heo * @pool: pool the new worker will belong to 1780c34056a3STejun Heo * 1781051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1782c34056a3STejun Heo * 1783c34056a3STejun Heo * CONTEXT: 1784c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1785c34056a3STejun Heo * 1786d185af30SYacine Belkadi * Return: 1787c34056a3STejun Heo * Pointer to the newly created worker. 1788c34056a3STejun Heo */ 1789bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1790c34056a3STejun Heo { 1791c34056a3STejun Heo struct worker *worker = NULL; 1792f3421797STejun Heo int id = -1; 1793e3c916a4STejun Heo char id_buf[16]; 1794c34056a3STejun Heo 17957cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 17967cda9aaeSLai Jiangshan id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL); 1797822d8405STejun Heo if (id < 0) 1798c34056a3STejun Heo goto fail; 1799c34056a3STejun Heo 1800f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1801c34056a3STejun Heo if (!worker) 1802c34056a3STejun Heo goto fail; 1803c34056a3STejun Heo 1804bd7bdd43STejun Heo worker->pool = pool; 1805c34056a3STejun Heo worker->id = id; 1806c34056a3STejun Heo 180729c91e99STejun Heo if (pool->cpu >= 0) 1808e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1809e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1810f3421797STejun Heo else 1811e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1812e3c916a4STejun Heo 1813f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1814e3c916a4STejun Heo "kworker/%s", id_buf); 1815c34056a3STejun Heo if (IS_ERR(worker->task)) 1816c34056a3STejun Heo goto fail; 1817c34056a3STejun Heo 181891151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 181925834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 182091151228SOleg Nesterov 1821da028469SLai Jiangshan /* successful, attach the worker to the pool */ 18224736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1823822d8405STejun Heo 1824051e1850SLai Jiangshan /* start the newly created worker */ 1825051e1850SLai Jiangshan spin_lock_irq(&pool->lock); 1826051e1850SLai Jiangshan worker->pool->nr_workers++; 1827051e1850SLai Jiangshan worker_enter_idle(worker); 1828051e1850SLai Jiangshan wake_up_process(worker->task); 1829051e1850SLai Jiangshan spin_unlock_irq(&pool->lock); 1830051e1850SLai Jiangshan 1831c34056a3STejun Heo return worker; 1832822d8405STejun Heo 1833c34056a3STejun Heo fail: 18349625ab17SLai Jiangshan if (id >= 0) 18357cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, id); 1836c34056a3STejun Heo kfree(worker); 1837c34056a3STejun Heo return NULL; 1838c34056a3STejun Heo } 1839c34056a3STejun Heo 1840c34056a3STejun Heo /** 1841c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1842c34056a3STejun Heo * @worker: worker to be destroyed 1843c34056a3STejun Heo * 184473eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 184573eb7fe7SLai Jiangshan * be idle. 1846c8e55f36STejun Heo * 1847c8e55f36STejun Heo * CONTEXT: 184860f5a4bcSLai Jiangshan * spin_lock_irq(pool->lock). 1849c34056a3STejun Heo */ 1850c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1851c34056a3STejun Heo { 1852bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1853c34056a3STejun Heo 1854cd549687STejun Heo lockdep_assert_held(&pool->lock); 1855cd549687STejun Heo 1856c34056a3STejun Heo /* sanity check frenzy */ 18576183c009STejun Heo if (WARN_ON(worker->current_work) || 185873eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 185973eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 18606183c009STejun Heo return; 1861c34056a3STejun Heo 1862bd7bdd43STejun Heo pool->nr_workers--; 1863bd7bdd43STejun Heo pool->nr_idle--; 1864c8e55f36STejun Heo 1865c8e55f36STejun Heo list_del_init(&worker->entry); 1866cb444766STejun Heo worker->flags |= WORKER_DIE; 186760f5a4bcSLai Jiangshan wake_up_process(worker->task); 1868c34056a3STejun Heo } 1869c34056a3STejun Heo 187032a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 1871e22bee78STejun Heo { 187232a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 1873e22bee78STejun Heo 1874d565ed63STejun Heo spin_lock_irq(&pool->lock); 1875e22bee78STejun Heo 18763347fc9fSLai Jiangshan while (too_many_workers(pool)) { 1877e22bee78STejun Heo struct worker *worker; 1878e22bee78STejun Heo unsigned long expires; 1879e22bee78STejun Heo 1880e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 188163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1882e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1883e22bee78STejun Heo 18843347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 188563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 18863347fc9fSLai Jiangshan break; 1887e22bee78STejun Heo } 18883347fc9fSLai Jiangshan 18893347fc9fSLai Jiangshan destroy_worker(worker); 1890e22bee78STejun Heo } 1891e22bee78STejun Heo 1892d565ed63STejun Heo spin_unlock_irq(&pool->lock); 1893e22bee78STejun Heo } 1894e22bee78STejun Heo 1895493a1724STejun Heo static void send_mayday(struct work_struct *work) 1896e22bee78STejun Heo { 1897112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1898112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 1899493a1724STejun Heo 19002e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 1901e22bee78STejun Heo 1902493008a8STejun Heo if (!wq->rescuer) 1903493a1724STejun Heo return; 1904e22bee78STejun Heo 1905e22bee78STejun Heo /* mayday mayday mayday */ 1906493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 190777668c8bSLai Jiangshan /* 190877668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 190977668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 191077668c8bSLai Jiangshan * rescuer is done with it. 191177668c8bSLai Jiangshan */ 191277668c8bSLai Jiangshan get_pwq(pwq); 1913493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 1914e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1915493a1724STejun Heo } 1916e22bee78STejun Heo } 1917e22bee78STejun Heo 191832a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 1919e22bee78STejun Heo { 192032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 1921e22bee78STejun Heo struct work_struct *work; 1922e22bee78STejun Heo 1923b2d82909STejun Heo spin_lock_irq(&pool->lock); 1924b2d82909STejun Heo spin_lock(&wq_mayday_lock); /* for wq->maydays */ 1925e22bee78STejun Heo 192663d95a91STejun Heo if (need_to_create_worker(pool)) { 1927e22bee78STejun Heo /* 1928e22bee78STejun Heo * We've been trying to create a new worker but 1929e22bee78STejun Heo * haven't been successful. We might be hitting an 1930e22bee78STejun Heo * allocation deadlock. Send distress signals to 1931e22bee78STejun Heo * rescuers. 1932e22bee78STejun Heo */ 193363d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1934e22bee78STejun Heo send_mayday(work); 1935e22bee78STejun Heo } 1936e22bee78STejun Heo 1937b2d82909STejun Heo spin_unlock(&wq_mayday_lock); 1938b2d82909STejun Heo spin_unlock_irq(&pool->lock); 1939e22bee78STejun Heo 194063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1941e22bee78STejun Heo } 1942e22bee78STejun Heo 1943e22bee78STejun Heo /** 1944e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 194563d95a91STejun Heo * @pool: pool to create a new worker for 1946e22bee78STejun Heo * 194763d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1948e22bee78STejun Heo * have at least one idle worker on return from this function. If 1949e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 195063d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1951e22bee78STejun Heo * possible allocation deadlock. 1952e22bee78STejun Heo * 1953c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 1954c5aa87bbSTejun Heo * may_start_working() %true. 1955e22bee78STejun Heo * 1956e22bee78STejun Heo * LOCKING: 1957d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 1958e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1959e22bee78STejun Heo * manager. 1960e22bee78STejun Heo */ 196129187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 1962d565ed63STejun Heo __releases(&pool->lock) 1963d565ed63STejun Heo __acquires(&pool->lock) 1964e22bee78STejun Heo { 1965e22bee78STejun Heo restart: 1966d565ed63STejun Heo spin_unlock_irq(&pool->lock); 19679f9c2364STejun Heo 1968e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 196963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1970e22bee78STejun Heo 1971e22bee78STejun Heo while (true) { 1972051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 1973e22bee78STejun Heo break; 1974e22bee78STejun Heo 1975e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 19769f9c2364STejun Heo 197763d95a91STejun Heo if (!need_to_create_worker(pool)) 1978e22bee78STejun Heo break; 1979e22bee78STejun Heo } 1980e22bee78STejun Heo 198163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1982d565ed63STejun Heo spin_lock_irq(&pool->lock); 1983051e1850SLai Jiangshan /* 1984051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 1985051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 1986051e1850SLai Jiangshan * already become busy. 1987051e1850SLai Jiangshan */ 198863d95a91STejun Heo if (need_to_create_worker(pool)) 1989e22bee78STejun Heo goto restart; 1990e22bee78STejun Heo } 1991e22bee78STejun Heo 1992e22bee78STejun Heo /** 1993e22bee78STejun Heo * manage_workers - manage worker pool 1994e22bee78STejun Heo * @worker: self 1995e22bee78STejun Heo * 1996706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 1997e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1998706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 1999e22bee78STejun Heo * 2000e22bee78STejun Heo * The caller can safely start processing works on false return. On 2001e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2002e22bee78STejun Heo * and may_start_working() is true. 2003e22bee78STejun Heo * 2004e22bee78STejun Heo * CONTEXT: 2005d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2006e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2007e22bee78STejun Heo * 2008d185af30SYacine Belkadi * Return: 200929187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 201029187a9eSTejun Heo * start processing works, %true if management function was performed and 201129187a9eSTejun Heo * the conditions that the caller verified before calling the function may 201229187a9eSTejun Heo * no longer be true. 2013e22bee78STejun Heo */ 2014e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2015e22bee78STejun Heo { 201663d95a91STejun Heo struct worker_pool *pool = worker->pool; 2017e22bee78STejun Heo 2018692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 201929187a9eSTejun Heo return false; 2020692b4825STejun Heo 2021692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 20222607d7a6STejun Heo pool->manager = worker; 2023e22bee78STejun Heo 202429187a9eSTejun Heo maybe_create_worker(pool); 2025e22bee78STejun Heo 20262607d7a6STejun Heo pool->manager = NULL; 2027692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2028692b4825STejun Heo wake_up(&wq_manager_wait); 202929187a9eSTejun Heo return true; 2030e22bee78STejun Heo } 2031e22bee78STejun Heo 2032a62428c0STejun Heo /** 2033a62428c0STejun Heo * process_one_work - process single work 2034c34056a3STejun Heo * @worker: self 2035a62428c0STejun Heo * @work: work to process 2036a62428c0STejun Heo * 2037a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2038a62428c0STejun Heo * process a single work including synchronization against and 2039a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2040a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2041a62428c0STejun Heo * call this function to process a work. 2042a62428c0STejun Heo * 2043a62428c0STejun Heo * CONTEXT: 2044d565ed63STejun Heo * spin_lock_irq(pool->lock) which is released and regrabbed. 2045a62428c0STejun Heo */ 2046c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2047d565ed63STejun Heo __releases(&pool->lock) 2048d565ed63STejun Heo __acquires(&pool->lock) 20491da177e4SLinus Torvalds { 2050112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2051bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2052112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 205373f53c4aSTejun Heo int work_color; 20547e11629dSTejun Heo struct worker *collision; 20554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 20564e6045f1SJohannes Berg /* 2057a62428c0STejun Heo * It is permissible to free the struct work_struct from 2058a62428c0STejun Heo * inside the function that is called from it, this we need to 2059a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2060a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2061a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 20624e6045f1SJohannes Berg */ 20634d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 20644d82a1deSPeter Zijlstra 20654d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 20664e6045f1SJohannes Berg #endif 2067807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 206885327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2069ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 207025511a47STejun Heo 20717e11629dSTejun Heo /* 20727e11629dSTejun Heo * A single work shouldn't be executed concurrently by 20737e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 20747e11629dSTejun Heo * already processing the work. If so, defer the work to the 20757e11629dSTejun Heo * currently executing one. 20767e11629dSTejun Heo */ 2077c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 20787e11629dSTejun Heo if (unlikely(collision)) { 20797e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 20807e11629dSTejun Heo return; 20817e11629dSTejun Heo } 20821da177e4SLinus Torvalds 20838930cabaSTejun Heo /* claim and dequeue */ 20841da177e4SLinus Torvalds debug_work_deactivate(work); 2085c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2086c34056a3STejun Heo worker->current_work = work; 2087a2c1c57bSTejun Heo worker->current_func = work->func; 2088112202d9STejun Heo worker->current_pwq = pwq; 208973f53c4aSTejun Heo work_color = get_work_color(work); 20907a22ad75STejun Heo 2091a62428c0STejun Heo list_del_init(&work->entry); 2092a62428c0STejun Heo 2093649027d7STejun Heo /* 2094228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2095228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2096228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2097228f1d00SLai Jiangshan * execution of the pending work items. 2098fb0e7bebSTejun Heo */ 2099fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2100228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2101fb0e7bebSTejun Heo 2102974271c4STejun Heo /* 2103a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2104a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2105a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2106a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2107228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2108974271c4STejun Heo */ 2109a489a03eSLai Jiangshan if (need_more_worker(pool)) 211063d95a91STejun Heo wake_up_worker(pool); 2111974271c4STejun Heo 21128930cabaSTejun Heo /* 21137c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2114d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 211523657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 211623657bb1STejun Heo * disabled. 21178930cabaSTejun Heo */ 21187c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 21191da177e4SLinus Torvalds 2120d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2121365970a1SDavid Howells 2122a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 21233295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2124e6f3faa7SPeter Zijlstra /* 2125f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2126f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2127e6f3faa7SPeter Zijlstra * 2128e6f3faa7SPeter Zijlstra * However, that would result in: 2129e6f3faa7SPeter Zijlstra * 2130e6f3faa7SPeter Zijlstra * A(W1) 2131e6f3faa7SPeter Zijlstra * WFC(C) 2132e6f3faa7SPeter Zijlstra * A(W1) 2133e6f3faa7SPeter Zijlstra * C(C) 2134e6f3faa7SPeter Zijlstra * 2135e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2136e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2137e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2138f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2139e6f3faa7SPeter Zijlstra * these locks. 2140e6f3faa7SPeter Zijlstra * 2141e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2142e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2143e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2144e6f3faa7SPeter Zijlstra */ 2145f52be570SPeter Zijlstra lockdep_invariant_state(true); 2146e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2147a2c1c57bSTejun Heo worker->current_func(work); 2148e36c886aSArjan van de Ven /* 2149e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2150e36c886aSArjan van de Ven * point will only record its address. 2151e36c886aSArjan van de Ven */ 2152e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 21533295f0efSIngo Molnar lock_map_release(&lockdep_map); 2154112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 21551da177e4SLinus Torvalds 2156d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2157044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2158044c782cSValentin Ilie " last function: %pf\n", 2159a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2160a2c1c57bSTejun Heo worker->current_func); 2161d5abe669SPeter Zijlstra debug_show_held_locks(current); 2162d5abe669SPeter Zijlstra dump_stack(); 2163d5abe669SPeter Zijlstra } 2164d5abe669SPeter Zijlstra 2165b22ce278STejun Heo /* 2166b22ce278STejun Heo * The following prevents a kworker from hogging CPU on !PREEMPT 2167b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2168b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2169b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2170789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2171789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2172b22ce278STejun Heo */ 2173a7e6425eSPaul E. McKenney cond_resched(); 2174b22ce278STejun Heo 2175d565ed63STejun Heo spin_lock_irq(&pool->lock); 2176a62428c0STejun Heo 2177fb0e7bebSTejun Heo /* clear cpu intensive status */ 2178fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2179fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2180fb0e7bebSTejun Heo 2181a62428c0STejun Heo /* we're done with it, release */ 218242f8570fSSasha Levin hash_del(&worker->hentry); 2183c34056a3STejun Heo worker->current_work = NULL; 2184a2c1c57bSTejun Heo worker->current_func = NULL; 2185112202d9STejun Heo worker->current_pwq = NULL; 21863d1cb205STejun Heo worker->desc_valid = false; 2187112202d9STejun Heo pwq_dec_nr_in_flight(pwq, work_color); 21881da177e4SLinus Torvalds } 21891da177e4SLinus Torvalds 2190affee4b2STejun Heo /** 2191affee4b2STejun Heo * process_scheduled_works - process scheduled works 2192affee4b2STejun Heo * @worker: self 2193affee4b2STejun Heo * 2194affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2195affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2196affee4b2STejun Heo * fetches a work from the top and executes it. 2197affee4b2STejun Heo * 2198affee4b2STejun Heo * CONTEXT: 2199d565ed63STejun Heo * spin_lock_irq(pool->lock) which may be released and regrabbed 2200affee4b2STejun Heo * multiple times. 2201affee4b2STejun Heo */ 2202affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 22031da177e4SLinus Torvalds { 2204affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2205affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2206a62428c0STejun Heo struct work_struct, entry); 2207c34056a3STejun Heo process_one_work(worker, work); 2208a62428c0STejun Heo } 22091da177e4SLinus Torvalds } 22101da177e4SLinus Torvalds 22114690c4abSTejun Heo /** 22124690c4abSTejun Heo * worker_thread - the worker thread function 2213c34056a3STejun Heo * @__worker: self 22144690c4abSTejun Heo * 2215c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2216c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2217c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2218c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2219c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2220d185af30SYacine Belkadi * 2221d185af30SYacine Belkadi * Return: 0 22224690c4abSTejun Heo */ 2223c34056a3STejun Heo static int worker_thread(void *__worker) 22241da177e4SLinus Torvalds { 2225c34056a3STejun Heo struct worker *worker = __worker; 2226bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 22271da177e4SLinus Torvalds 2228e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2229e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2230c8e55f36STejun Heo woke_up: 2231d565ed63STejun Heo spin_lock_irq(&pool->lock); 2232affee4b2STejun Heo 2233a9ab775bSTejun Heo /* am I supposed to die? */ 2234a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2235d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2236a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2237e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 223860f5a4bcSLai Jiangshan 223960f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 22407cda9aaeSLai Jiangshan ida_simple_remove(&pool->worker_ida, worker->id); 224160f5a4bcSLai Jiangshan worker_detach_from_pool(worker, pool); 224260f5a4bcSLai Jiangshan kfree(worker); 2243c8e55f36STejun Heo return 0; 2244c8e55f36STejun Heo } 2245c8e55f36STejun Heo 2246c8e55f36STejun Heo worker_leave_idle(worker); 2247db7bccf4STejun Heo recheck: 2248e22bee78STejun Heo /* no more worker necessary? */ 224963d95a91STejun Heo if (!need_more_worker(pool)) 2250e22bee78STejun Heo goto sleep; 2251e22bee78STejun Heo 2252e22bee78STejun Heo /* do we need to manage? */ 225363d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2254e22bee78STejun Heo goto recheck; 2255e22bee78STejun Heo 2256c8e55f36STejun Heo /* 2257c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2258c8e55f36STejun Heo * preparing to process a work or actually processing it. 2259c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2260c8e55f36STejun Heo */ 22616183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2262c8e55f36STejun Heo 2263e22bee78STejun Heo /* 2264a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2265a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2266a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2267a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2268a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2269e22bee78STejun Heo */ 2270a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2271e22bee78STejun Heo 2272e22bee78STejun Heo do { 2273affee4b2STejun Heo struct work_struct *work = 2274bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2275affee4b2STejun Heo struct work_struct, entry); 2276affee4b2STejun Heo 227782607adcSTejun Heo pool->watchdog_ts = jiffies; 227882607adcSTejun Heo 2279c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2280affee4b2STejun Heo /* optimization path, not strictly necessary */ 2281affee4b2STejun Heo process_one_work(worker, work); 2282affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2283affee4b2STejun Heo process_scheduled_works(worker); 2284affee4b2STejun Heo } else { 2285c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2286affee4b2STejun Heo process_scheduled_works(worker); 2287affee4b2STejun Heo } 228863d95a91STejun Heo } while (keep_working(pool)); 2289affee4b2STejun Heo 2290228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2291d313dd85STejun Heo sleep: 2292c8e55f36STejun Heo /* 2293d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2294d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2295d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2296d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2297d565ed63STejun Heo * event. 2298c8e55f36STejun Heo */ 2299c8e55f36STejun Heo worker_enter_idle(worker); 2300c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2301d565ed63STejun Heo spin_unlock_irq(&pool->lock); 23021da177e4SLinus Torvalds schedule(); 2303c8e55f36STejun Heo goto woke_up; 23041da177e4SLinus Torvalds } 23051da177e4SLinus Torvalds 2306e22bee78STejun Heo /** 2307e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2308111c225aSTejun Heo * @__rescuer: self 2309e22bee78STejun Heo * 2310e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2311493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2312e22bee78STejun Heo * 2313706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2314e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2315e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2316e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2317e22bee78STejun Heo * the problem rescuer solves. 2318e22bee78STejun Heo * 2319706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2320706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2321e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2322e22bee78STejun Heo * 2323e22bee78STejun Heo * This should happen rarely. 2324d185af30SYacine Belkadi * 2325d185af30SYacine Belkadi * Return: 0 2326e22bee78STejun Heo */ 2327111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2328e22bee78STejun Heo { 2329111c225aSTejun Heo struct worker *rescuer = __rescuer; 2330111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2331e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 23324d595b86SLai Jiangshan bool should_stop; 2333e22bee78STejun Heo 2334e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2335111c225aSTejun Heo 2336111c225aSTejun Heo /* 2337111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2338111c225aSTejun Heo * doesn't participate in concurrency management. 2339111c225aSTejun Heo */ 2340111c225aSTejun Heo rescuer->task->flags |= PF_WQ_WORKER; 2341e22bee78STejun Heo repeat: 2342c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 23431da177e4SLinus Torvalds 23444d595b86SLai Jiangshan /* 23454d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 23464d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 23474d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 23484d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 23494d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 23504d595b86SLai Jiangshan * list is always empty on exit. 23514d595b86SLai Jiangshan */ 23524d595b86SLai Jiangshan should_stop = kthread_should_stop(); 23531da177e4SLinus Torvalds 2354493a1724STejun Heo /* see whether any pwq is asking for help */ 23552e109a28STejun Heo spin_lock_irq(&wq_mayday_lock); 2356493a1724STejun Heo 2357493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2358493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2359493a1724STejun Heo struct pool_workqueue, mayday_node); 2360112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2361e22bee78STejun Heo struct work_struct *work, *n; 236282607adcSTejun Heo bool first = true; 2363e22bee78STejun Heo 2364e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2365493a1724STejun Heo list_del_init(&pwq->mayday_node); 2366493a1724STejun Heo 23672e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2368e22bee78STejun Heo 236951697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 237051697d39SLai Jiangshan 237151697d39SLai Jiangshan spin_lock_irq(&pool->lock); 2372b3104104SLai Jiangshan rescuer->pool = pool; 2373e22bee78STejun Heo 2374e22bee78STejun Heo /* 2375e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2376e22bee78STejun Heo * process'em. 2377e22bee78STejun Heo */ 23780479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 237982607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 238082607adcSTejun Heo if (get_work_pwq(work) == pwq) { 238182607adcSTejun Heo if (first) 238282607adcSTejun Heo pool->watchdog_ts = jiffies; 2383e22bee78STejun Heo move_linked_works(work, scheduled, &n); 238482607adcSTejun Heo } 238582607adcSTejun Heo first = false; 238682607adcSTejun Heo } 2387e22bee78STejun Heo 2388008847f6SNeilBrown if (!list_empty(scheduled)) { 2389e22bee78STejun Heo process_scheduled_works(rescuer); 23907576958aSTejun Heo 23917576958aSTejun Heo /* 2392008847f6SNeilBrown * The above execution of rescued work items could 2393008847f6SNeilBrown * have created more to rescue through 2394008847f6SNeilBrown * pwq_activate_first_delayed() or chained 2395008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2396008847f6SNeilBrown * that such back-to-back work items, which may be 2397008847f6SNeilBrown * being used to relieve memory pressure, don't 2398008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2399008847f6SNeilBrown */ 2400008847f6SNeilBrown if (need_to_create_worker(pool)) { 2401008847f6SNeilBrown spin_lock(&wq_mayday_lock); 2402008847f6SNeilBrown get_pwq(pwq); 2403008847f6SNeilBrown list_move_tail(&pwq->mayday_node, &wq->maydays); 2404008847f6SNeilBrown spin_unlock(&wq_mayday_lock); 2405008847f6SNeilBrown } 2406008847f6SNeilBrown } 2407008847f6SNeilBrown 2408008847f6SNeilBrown /* 240977668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 241013b1d625SLai Jiangshan * go away while we're still attached to it. 241177668c8bSLai Jiangshan */ 241277668c8bSLai Jiangshan put_pwq(pwq); 241377668c8bSLai Jiangshan 241477668c8bSLai Jiangshan /* 2415d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 24167576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 24177576958aSTejun Heo * and stalling the execution. 24187576958aSTejun Heo */ 2419d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 242063d95a91STejun Heo wake_up_worker(pool); 24217576958aSTejun Heo 2422b3104104SLai Jiangshan rescuer->pool = NULL; 242313b1d625SLai Jiangshan spin_unlock_irq(&pool->lock); 242413b1d625SLai Jiangshan 242513b1d625SLai Jiangshan worker_detach_from_pool(rescuer, pool); 242613b1d625SLai Jiangshan 242713b1d625SLai Jiangshan spin_lock_irq(&wq_mayday_lock); 24281da177e4SLinus Torvalds } 24291da177e4SLinus Torvalds 24302e109a28STejun Heo spin_unlock_irq(&wq_mayday_lock); 2431493a1724STejun Heo 24324d595b86SLai Jiangshan if (should_stop) { 24334d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 24344d595b86SLai Jiangshan rescuer->task->flags &= ~PF_WQ_WORKER; 24354d595b86SLai Jiangshan return 0; 24364d595b86SLai Jiangshan } 24374d595b86SLai Jiangshan 2438111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2439111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2440e22bee78STejun Heo schedule(); 2441e22bee78STejun Heo goto repeat; 24421da177e4SLinus Torvalds } 24431da177e4SLinus Torvalds 2444fca839c0STejun Heo /** 2445fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2446fca839c0STejun Heo * @target_wq: workqueue being flushed 2447fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2448fca839c0STejun Heo * 2449fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2450fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2451fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2452fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2453fca839c0STejun Heo * a deadlock. 2454fca839c0STejun Heo */ 2455fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2456fca839c0STejun Heo struct work_struct *target_work) 2457fca839c0STejun Heo { 2458fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2459fca839c0STejun Heo struct worker *worker; 2460fca839c0STejun Heo 2461fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2462fca839c0STejun Heo return; 2463fca839c0STejun Heo 2464fca839c0STejun Heo worker = current_wq_worker(); 2465fca839c0STejun Heo 2466fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2467fca839c0STejun Heo "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf", 2468fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 246923d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 247023d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2471fca839c0STejun Heo "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf", 2472fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2473fca839c0STejun Heo target_wq->name, target_func); 2474fca839c0STejun Heo } 2475fca839c0STejun Heo 2476fc2e4d70SOleg Nesterov struct wq_barrier { 2477fc2e4d70SOleg Nesterov struct work_struct work; 2478fc2e4d70SOleg Nesterov struct completion done; 24792607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2480fc2e4d70SOleg Nesterov }; 2481fc2e4d70SOleg Nesterov 2482fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2483fc2e4d70SOleg Nesterov { 2484fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2485fc2e4d70SOleg Nesterov complete(&barr->done); 2486fc2e4d70SOleg Nesterov } 2487fc2e4d70SOleg Nesterov 24884690c4abSTejun Heo /** 24894690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2490112202d9STejun Heo * @pwq: pwq to insert barrier into 24914690c4abSTejun Heo * @barr: wq_barrier to insert 2492affee4b2STejun Heo * @target: target work to attach @barr to 2493affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 24944690c4abSTejun Heo * 2495affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2496affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2497affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2498affee4b2STejun Heo * cpu. 2499affee4b2STejun Heo * 2500affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2501affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2502affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2503affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2504affee4b2STejun Heo * after a work with LINKED flag set. 2505affee4b2STejun Heo * 2506affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2507112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 25084690c4abSTejun Heo * 25094690c4abSTejun Heo * CONTEXT: 2510d565ed63STejun Heo * spin_lock_irq(pool->lock). 25114690c4abSTejun Heo */ 2512112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2513affee4b2STejun Heo struct wq_barrier *barr, 2514affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2515fc2e4d70SOleg Nesterov { 2516affee4b2STejun Heo struct list_head *head; 2517affee4b2STejun Heo unsigned int linked = 0; 2518affee4b2STejun Heo 2519dc186ad7SThomas Gleixner /* 2520d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2521dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2522dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2523dc186ad7SThomas Gleixner * might deadlock. 2524dc186ad7SThomas Gleixner */ 2525ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 252622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 252752fa5bc5SBoqun Feng 2528fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2529fd1a5b04SByungchul Park 25302607d7a6STejun Heo barr->task = current; 253183c22520SOleg Nesterov 2532affee4b2STejun Heo /* 2533affee4b2STejun Heo * If @target is currently being executed, schedule the 2534affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2535affee4b2STejun Heo */ 2536affee4b2STejun Heo if (worker) 2537affee4b2STejun Heo head = worker->scheduled.next; 2538affee4b2STejun Heo else { 2539affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2540affee4b2STejun Heo 2541affee4b2STejun Heo head = target->entry.next; 2542affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2543affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2544affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2545affee4b2STejun Heo } 2546affee4b2STejun Heo 2547dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2548112202d9STejun Heo insert_work(pwq, &barr->work, head, 2549affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2550fc2e4d70SOleg Nesterov } 2551fc2e4d70SOleg Nesterov 255273f53c4aSTejun Heo /** 2553112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 255473f53c4aSTejun Heo * @wq: workqueue being flushed 255573f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 255673f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 255773f53c4aSTejun Heo * 2558112202d9STejun Heo * Prepare pwqs for workqueue flushing. 255973f53c4aSTejun Heo * 2560112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2561112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2562112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2563112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2564112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 256573f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 256673f53c4aSTejun Heo * 256773f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 256873f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 256973f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 257073f53c4aSTejun Heo * is returned. 257173f53c4aSTejun Heo * 2572112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 257373f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 257473f53c4aSTejun Heo * advanced to @work_color. 257573f53c4aSTejun Heo * 257673f53c4aSTejun Heo * CONTEXT: 25773c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 257873f53c4aSTejun Heo * 2579d185af30SYacine Belkadi * Return: 258073f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 258173f53c4aSTejun Heo * otherwise. 258273f53c4aSTejun Heo */ 2583112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 258473f53c4aSTejun Heo int flush_color, int work_color) 25851da177e4SLinus Torvalds { 258673f53c4aSTejun Heo bool wait = false; 258749e3cf44STejun Heo struct pool_workqueue *pwq; 25881da177e4SLinus Torvalds 258973f53c4aSTejun Heo if (flush_color >= 0) { 25906183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2591112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2592dc186ad7SThomas Gleixner } 259314441960SOleg Nesterov 259449e3cf44STejun Heo for_each_pwq(pwq, wq) { 2595112202d9STejun Heo struct worker_pool *pool = pwq->pool; 25961da177e4SLinus Torvalds 2597b09f4fd3SLai Jiangshan spin_lock_irq(&pool->lock); 259873f53c4aSTejun Heo 259973f53c4aSTejun Heo if (flush_color >= 0) { 26006183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 260173f53c4aSTejun Heo 2602112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2603112202d9STejun Heo pwq->flush_color = flush_color; 2604112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 260573f53c4aSTejun Heo wait = true; 26061da177e4SLinus Torvalds } 260773f53c4aSTejun Heo } 260873f53c4aSTejun Heo 260973f53c4aSTejun Heo if (work_color >= 0) { 26106183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2611112202d9STejun Heo pwq->work_color = work_color; 261273f53c4aSTejun Heo } 261373f53c4aSTejun Heo 2614b09f4fd3SLai Jiangshan spin_unlock_irq(&pool->lock); 26151da177e4SLinus Torvalds } 26161da177e4SLinus Torvalds 2617112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 261873f53c4aSTejun Heo complete(&wq->first_flusher->done); 261973f53c4aSTejun Heo 262073f53c4aSTejun Heo return wait; 262183c22520SOleg Nesterov } 26221da177e4SLinus Torvalds 26230fcb78c2SRolf Eike Beer /** 26241da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 26250fcb78c2SRolf Eike Beer * @wq: workqueue to flush 26261da177e4SLinus Torvalds * 2627c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2628c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 26291da177e4SLinus Torvalds */ 26307ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 26311da177e4SLinus Torvalds { 263273f53c4aSTejun Heo struct wq_flusher this_flusher = { 263373f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 263473f53c4aSTejun Heo .flush_color = -1, 2635fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 263673f53c4aSTejun Heo }; 263773f53c4aSTejun Heo int next_color; 2638b1f4ec17SOleg Nesterov 26393347fa09STejun Heo if (WARN_ON(!wq_online)) 26403347fa09STejun Heo return; 26413347fa09STejun Heo 26423c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 264373f53c4aSTejun Heo 264473f53c4aSTejun Heo /* 264573f53c4aSTejun Heo * Start-to-wait phase 264673f53c4aSTejun Heo */ 264773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 264873f53c4aSTejun Heo 264973f53c4aSTejun Heo if (next_color != wq->flush_color) { 265073f53c4aSTejun Heo /* 265173f53c4aSTejun Heo * Color space is not full. The current work_color 265273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 265373f53c4aSTejun Heo * by one. 265473f53c4aSTejun Heo */ 26556183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 265673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 265773f53c4aSTejun Heo wq->work_color = next_color; 265873f53c4aSTejun Heo 265973f53c4aSTejun Heo if (!wq->first_flusher) { 266073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 26616183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 266273f53c4aSTejun Heo 266373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 266473f53c4aSTejun Heo 2665112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 266673f53c4aSTejun Heo wq->work_color)) { 266773f53c4aSTejun Heo /* nothing to flush, done */ 266873f53c4aSTejun Heo wq->flush_color = next_color; 266973f53c4aSTejun Heo wq->first_flusher = NULL; 267073f53c4aSTejun Heo goto out_unlock; 267173f53c4aSTejun Heo } 267273f53c4aSTejun Heo } else { 267373f53c4aSTejun Heo /* wait in queue */ 26746183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 267573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2676112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 267773f53c4aSTejun Heo } 267873f53c4aSTejun Heo } else { 267973f53c4aSTejun Heo /* 268073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 268173f53c4aSTejun Heo * The next flush completion will assign us 268273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 268373f53c4aSTejun Heo */ 268473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 268573f53c4aSTejun Heo } 268673f53c4aSTejun Heo 2687fca839c0STejun Heo check_flush_dependency(wq, NULL); 2688fca839c0STejun Heo 26893c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 269073f53c4aSTejun Heo 269173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 269273f53c4aSTejun Heo 269373f53c4aSTejun Heo /* 269473f53c4aSTejun Heo * Wake-up-and-cascade phase 269573f53c4aSTejun Heo * 269673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 269773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 269873f53c4aSTejun Heo */ 269973f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 270073f53c4aSTejun Heo return; 270173f53c4aSTejun Heo 27023c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 270373f53c4aSTejun Heo 27044ce48b37STejun Heo /* we might have raced, check again with mutex held */ 27054ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 27064ce48b37STejun Heo goto out_unlock; 27074ce48b37STejun Heo 270873f53c4aSTejun Heo wq->first_flusher = NULL; 270973f53c4aSTejun Heo 27106183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 27116183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 271273f53c4aSTejun Heo 271373f53c4aSTejun Heo while (true) { 271473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 271573f53c4aSTejun Heo 271673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 271773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 271873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 271973f53c4aSTejun Heo break; 272073f53c4aSTejun Heo list_del_init(&next->list); 272173f53c4aSTejun Heo complete(&next->done); 272273f53c4aSTejun Heo } 272373f53c4aSTejun Heo 27246183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 272573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 272673f53c4aSTejun Heo 272773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 272873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 272973f53c4aSTejun Heo 273073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 273173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 273273f53c4aSTejun Heo /* 273373f53c4aSTejun Heo * Assign the same color to all overflowed 273473f53c4aSTejun Heo * flushers, advance work_color and append to 273573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 273673f53c4aSTejun Heo * phase for these overflowed flushers. 273773f53c4aSTejun Heo */ 273873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 273973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 274073f53c4aSTejun Heo 274173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 274273f53c4aSTejun Heo 274373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 274473f53c4aSTejun Heo &wq->flusher_queue); 2745112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 274673f53c4aSTejun Heo } 274773f53c4aSTejun Heo 274873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 27496183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 275073f53c4aSTejun Heo break; 275173f53c4aSTejun Heo } 275273f53c4aSTejun Heo 275373f53c4aSTejun Heo /* 275473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2755112202d9STejun Heo * the new first flusher and arm pwqs. 275673f53c4aSTejun Heo */ 27576183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 27586183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 275973f53c4aSTejun Heo 276073f53c4aSTejun Heo list_del_init(&next->list); 276173f53c4aSTejun Heo wq->first_flusher = next; 276273f53c4aSTejun Heo 2763112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 276473f53c4aSTejun Heo break; 276573f53c4aSTejun Heo 276673f53c4aSTejun Heo /* 276773f53c4aSTejun Heo * Meh... this color is already done, clear first 276873f53c4aSTejun Heo * flusher and repeat cascading. 276973f53c4aSTejun Heo */ 277073f53c4aSTejun Heo wq->first_flusher = NULL; 277173f53c4aSTejun Heo } 277273f53c4aSTejun Heo 277373f53c4aSTejun Heo out_unlock: 27743c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 27751da177e4SLinus Torvalds } 27761dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 27771da177e4SLinus Torvalds 27789c5a2ba7STejun Heo /** 27799c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 27809c5a2ba7STejun Heo * @wq: workqueue to drain 27819c5a2ba7STejun Heo * 27829c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 27839c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 27849c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2785b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 27869c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 27879c5a2ba7STejun Heo * takes too long. 27889c5a2ba7STejun Heo */ 27899c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 27909c5a2ba7STejun Heo { 27919c5a2ba7STejun Heo unsigned int flush_cnt = 0; 279249e3cf44STejun Heo struct pool_workqueue *pwq; 27939c5a2ba7STejun Heo 27949c5a2ba7STejun Heo /* 27959c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 27969c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2797618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 27989c5a2ba7STejun Heo */ 279987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 28009c5a2ba7STejun Heo if (!wq->nr_drainers++) 2801618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 280287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28039c5a2ba7STejun Heo reflush: 28049c5a2ba7STejun Heo flush_workqueue(wq); 28059c5a2ba7STejun Heo 2806b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 280776af4d93STejun Heo 280849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2809fa2563e4SThomas Tuttle bool drained; 28109c5a2ba7STejun Heo 2811b09f4fd3SLai Jiangshan spin_lock_irq(&pwq->pool->lock); 2812112202d9STejun Heo drained = !pwq->nr_active && list_empty(&pwq->delayed_works); 2813b09f4fd3SLai Jiangshan spin_unlock_irq(&pwq->pool->lock); 2814fa2563e4SThomas Tuttle 2815fa2563e4SThomas Tuttle if (drained) 28169c5a2ba7STejun Heo continue; 28179c5a2ba7STejun Heo 28189c5a2ba7STejun Heo if (++flush_cnt == 10 || 28199c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 2820c5aa87bbSTejun Heo pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n", 28219c5a2ba7STejun Heo wq->name, flush_cnt); 282276af4d93STejun Heo 2823b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 28249c5a2ba7STejun Heo goto reflush; 28259c5a2ba7STejun Heo } 28269c5a2ba7STejun Heo 28279c5a2ba7STejun Heo if (!--wq->nr_drainers) 2828618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 282987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 28309c5a2ba7STejun Heo } 28319c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 28329c5a2ba7STejun Heo 2833606a5020STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr) 2834baf59022STejun Heo { 2835baf59022STejun Heo struct worker *worker = NULL; 2836c9e7cf27STejun Heo struct worker_pool *pool; 2837112202d9STejun Heo struct pool_workqueue *pwq; 2838baf59022STejun Heo 2839baf59022STejun Heo might_sleep(); 2840baf59022STejun Heo 2841fa1b54e6STejun Heo local_irq_disable(); 2842fa1b54e6STejun Heo pool = get_work_pool(work); 2843fa1b54e6STejun Heo if (!pool) { 2844fa1b54e6STejun Heo local_irq_enable(); 2845fa1b54e6STejun Heo return false; 2846fa1b54e6STejun Heo } 2847fa1b54e6STejun Heo 2848fa1b54e6STejun Heo spin_lock(&pool->lock); 28490b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 2850112202d9STejun Heo pwq = get_work_pwq(work); 2851112202d9STejun Heo if (pwq) { 2852112202d9STejun Heo if (unlikely(pwq->pool != pool)) 2853baf59022STejun Heo goto already_gone; 2854606a5020STejun Heo } else { 2855c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 2856baf59022STejun Heo if (!worker) 2857baf59022STejun Heo goto already_gone; 2858112202d9STejun Heo pwq = worker->current_pwq; 2859606a5020STejun Heo } 2860baf59022STejun Heo 2861fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 2862fca839c0STejun Heo 2863112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 2864d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2865baf59022STejun Heo 2866e159489bSTejun Heo /* 2867a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 2868a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 2869a1d14934SPeter Zijlstra * 2870a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 2871a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 2872a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 2873a1d14934SPeter Zijlstra * forward progress. 2874e159489bSTejun Heo */ 2875a1d14934SPeter Zijlstra if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) { 2876112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 2877112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 2878a1d14934SPeter Zijlstra } 2879e159489bSTejun Heo 2880baf59022STejun Heo return true; 2881baf59022STejun Heo already_gone: 2882d565ed63STejun Heo spin_unlock_irq(&pool->lock); 2883baf59022STejun Heo return false; 2884baf59022STejun Heo } 2885baf59022STejun Heo 2886db700897SOleg Nesterov /** 2887401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2888401a8d04STejun Heo * @work: the work to flush 2889db700897SOleg Nesterov * 2890606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 2891606a5020STejun Heo * on return if it hasn't been requeued since flush started. 2892401a8d04STejun Heo * 2893d185af30SYacine Belkadi * Return: 2894401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2895401a8d04STejun Heo * %false if it was already idle. 2896db700897SOleg Nesterov */ 2897401a8d04STejun Heo bool flush_work(struct work_struct *work) 2898db700897SOleg Nesterov { 289912997d1aSBjorn Helgaas struct wq_barrier barr; 290012997d1aSBjorn Helgaas 29013347fa09STejun Heo if (WARN_ON(!wq_online)) 29023347fa09STejun Heo return false; 29033347fa09STejun Heo 290412997d1aSBjorn Helgaas if (start_flush_work(work, &barr)) { 290512997d1aSBjorn Helgaas wait_for_completion(&barr.done); 290612997d1aSBjorn Helgaas destroy_work_on_stack(&barr.work); 290712997d1aSBjorn Helgaas return true; 290812997d1aSBjorn Helgaas } else { 290912997d1aSBjorn Helgaas return false; 291012997d1aSBjorn Helgaas } 2911606a5020STejun Heo } 2912db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2913db700897SOleg Nesterov 29148603e1b3STejun Heo struct cwt_wait { 2915ac6424b9SIngo Molnar wait_queue_entry_t wait; 29168603e1b3STejun Heo struct work_struct *work; 29178603e1b3STejun Heo }; 29188603e1b3STejun Heo 2919ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 29208603e1b3STejun Heo { 29218603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 29228603e1b3STejun Heo 29238603e1b3STejun Heo if (cwait->work != key) 29248603e1b3STejun Heo return 0; 29258603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 29268603e1b3STejun Heo } 29278603e1b3STejun Heo 292836e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 2929401a8d04STejun Heo { 29308603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 2931bbb68dfaSTejun Heo unsigned long flags; 29321f1f642eSOleg Nesterov int ret; 29331f1f642eSOleg Nesterov 29341f1f642eSOleg Nesterov do { 2935bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 2936bbb68dfaSTejun Heo /* 29378603e1b3STejun Heo * If someone else is already canceling, wait for it to 29388603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 29398603e1b3STejun Heo * because we may get scheduled between @work's completion 29408603e1b3STejun Heo * and the other canceling task resuming and clearing 29418603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 29428603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 29438603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 29448603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 29458603e1b3STejun Heo * we're hogging the CPU. 29468603e1b3STejun Heo * 29478603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 29488603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 29498603e1b3STejun Heo * wake function which matches @work along with exclusive 29508603e1b3STejun Heo * wait and wakeup. 2951bbb68dfaSTejun Heo */ 29528603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 29538603e1b3STejun Heo struct cwt_wait cwait; 29548603e1b3STejun Heo 29558603e1b3STejun Heo init_wait(&cwait.wait); 29568603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 29578603e1b3STejun Heo cwait.work = work; 29588603e1b3STejun Heo 29598603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 29608603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 29618603e1b3STejun Heo if (work_is_canceling(work)) 29628603e1b3STejun Heo schedule(); 29638603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 29648603e1b3STejun Heo } 29651f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 29661f1f642eSOleg Nesterov 2967bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 2968bbb68dfaSTejun Heo mark_work_canceling(work); 2969bbb68dfaSTejun Heo local_irq_restore(flags); 2970bbb68dfaSTejun Heo 29713347fa09STejun Heo /* 29723347fa09STejun Heo * This allows canceling during early boot. We know that @work 29733347fa09STejun Heo * isn't executing. 29743347fa09STejun Heo */ 29753347fa09STejun Heo if (wq_online) 2976606a5020STejun Heo flush_work(work); 29773347fa09STejun Heo 29787a22ad75STejun Heo clear_work_data(work); 29798603e1b3STejun Heo 29808603e1b3STejun Heo /* 29818603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 29828603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 29838603e1b3STejun Heo * visible there. 29848603e1b3STejun Heo */ 29858603e1b3STejun Heo smp_mb(); 29868603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 29878603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 29888603e1b3STejun Heo 29891f1f642eSOleg Nesterov return ret; 29901f1f642eSOleg Nesterov } 29911f1f642eSOleg Nesterov 29926e84d644SOleg Nesterov /** 2993401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2994401a8d04STejun Heo * @work: the work to cancel 29956e84d644SOleg Nesterov * 2996401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2997401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2998401a8d04STejun Heo * another workqueue. On return from this function, @work is 2999401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 30001f1f642eSOleg Nesterov * 3001401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3002401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 30036e84d644SOleg Nesterov * 3004401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 30056e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3006401a8d04STejun Heo * 3007d185af30SYacine Belkadi * Return: 3008401a8d04STejun Heo * %true if @work was pending, %false otherwise. 30096e84d644SOleg Nesterov */ 3010401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 30116e84d644SOleg Nesterov { 301236e227d2STejun Heo return __cancel_work_timer(work, false); 3013b89deed3SOleg Nesterov } 301428e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3015b89deed3SOleg Nesterov 30166e84d644SOleg Nesterov /** 3017401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3018401a8d04STejun Heo * @dwork: the delayed work to flush 30196e84d644SOleg Nesterov * 3020401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3021401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3022401a8d04STejun Heo * considers the last queueing instance of @dwork. 30231f1f642eSOleg Nesterov * 3024d185af30SYacine Belkadi * Return: 3025401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3026401a8d04STejun Heo * %false if it was already idle. 30276e84d644SOleg Nesterov */ 3028401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3029401a8d04STejun Heo { 30308930cabaSTejun Heo local_irq_disable(); 3031401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 303260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 30338930cabaSTejun Heo local_irq_enable(); 3034401a8d04STejun Heo return flush_work(&dwork->work); 3035401a8d04STejun Heo } 3036401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3037401a8d04STejun Heo 303805f0fe6bSTejun Heo /** 303905f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 304005f0fe6bSTejun Heo * @rwork: the rcu work to flush 304105f0fe6bSTejun Heo * 304205f0fe6bSTejun Heo * Return: 304305f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 304405f0fe6bSTejun Heo * %false if it was already idle. 304505f0fe6bSTejun Heo */ 304605f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 304705f0fe6bSTejun Heo { 304805f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 304905f0fe6bSTejun Heo rcu_barrier(); 305005f0fe6bSTejun Heo flush_work(&rwork->work); 305105f0fe6bSTejun Heo return true; 305205f0fe6bSTejun Heo } else { 305305f0fe6bSTejun Heo return flush_work(&rwork->work); 305405f0fe6bSTejun Heo } 305505f0fe6bSTejun Heo } 305605f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 305705f0fe6bSTejun Heo 3058f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3059f72b8792SJens Axboe { 3060f72b8792SJens Axboe unsigned long flags; 3061f72b8792SJens Axboe int ret; 3062f72b8792SJens Axboe 3063f72b8792SJens Axboe do { 3064f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3065f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3066f72b8792SJens Axboe 3067f72b8792SJens Axboe if (unlikely(ret < 0)) 3068f72b8792SJens Axboe return false; 3069f72b8792SJens Axboe 3070f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3071f72b8792SJens Axboe local_irq_restore(flags); 3072f72b8792SJens Axboe return ret; 3073f72b8792SJens Axboe } 3074f72b8792SJens Axboe 3075f72b8792SJens Axboe /* 3076f72b8792SJens Axboe * See cancel_delayed_work() 3077f72b8792SJens Axboe */ 3078f72b8792SJens Axboe bool cancel_work(struct work_struct *work) 3079f72b8792SJens Axboe { 3080f72b8792SJens Axboe return __cancel_work(work, false); 3081f72b8792SJens Axboe } 3082f72b8792SJens Axboe 3083401a8d04STejun Heo /** 308457b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 308557b30ae7STejun Heo * @dwork: delayed_work to cancel 308609383498STejun Heo * 3087d185af30SYacine Belkadi * Kill off a pending delayed_work. 3088d185af30SYacine Belkadi * 3089d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3090d185af30SYacine Belkadi * pending. 3091d185af30SYacine Belkadi * 3092d185af30SYacine Belkadi * Note: 3093d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3094d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3095d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 309609383498STejun Heo * 309757b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 309809383498STejun Heo */ 309957b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 310009383498STejun Heo { 3101f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 310209383498STejun Heo } 310357b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 310409383498STejun Heo 310509383498STejun Heo /** 3106401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3107401a8d04STejun Heo * @dwork: the delayed work cancel 3108401a8d04STejun Heo * 3109401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3110401a8d04STejun Heo * 3111d185af30SYacine Belkadi * Return: 3112401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3113401a8d04STejun Heo */ 3114401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 31156e84d644SOleg Nesterov { 311636e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 31176e84d644SOleg Nesterov } 3118f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 31191da177e4SLinus Torvalds 31200fcb78c2SRolf Eike Beer /** 312131ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3122b6136773SAndrew Morton * @func: the function to call 3123b6136773SAndrew Morton * 312431ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 312531ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3126b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 312731ddd871STejun Heo * 3128d185af30SYacine Belkadi * Return: 312931ddd871STejun Heo * 0 on success, -errno on failure. 3130b6136773SAndrew Morton */ 313165f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 313215316ba8SChristoph Lameter { 313315316ba8SChristoph Lameter int cpu; 313438f51568SNamhyung Kim struct work_struct __percpu *works; 313515316ba8SChristoph Lameter 3136b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3137b6136773SAndrew Morton if (!works) 313815316ba8SChristoph Lameter return -ENOMEM; 3139b6136773SAndrew Morton 314095402b38SGautham R Shenoy get_online_cpus(); 314193981800STejun Heo 314215316ba8SChristoph Lameter for_each_online_cpu(cpu) { 31439bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 31449bfb1839SIngo Molnar 31459bfb1839SIngo Molnar INIT_WORK(work, func); 31468de6d308SOleg Nesterov schedule_work_on(cpu, work); 314715316ba8SChristoph Lameter } 314893981800STejun Heo 314993981800STejun Heo for_each_online_cpu(cpu) 31508616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 315193981800STejun Heo 315295402b38SGautham R Shenoy put_online_cpus(); 3153b6136773SAndrew Morton free_percpu(works); 315415316ba8SChristoph Lameter return 0; 315515316ba8SChristoph Lameter } 315615316ba8SChristoph Lameter 3157eef6a7d5SAlan Stern /** 31581fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 31591fa44ecaSJames Bottomley * @fn: the function to execute 31601fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 31611fa44ecaSJames Bottomley * be available when the work executes) 31621fa44ecaSJames Bottomley * 31631fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 31641fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 31651fa44ecaSJames Bottomley * 3166d185af30SYacine Belkadi * Return: 0 - function was executed 31671fa44ecaSJames Bottomley * 1 - function was scheduled for execution 31681fa44ecaSJames Bottomley */ 316965f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 31701fa44ecaSJames Bottomley { 31711fa44ecaSJames Bottomley if (!in_interrupt()) { 317265f27f38SDavid Howells fn(&ew->work); 31731fa44ecaSJames Bottomley return 0; 31741fa44ecaSJames Bottomley } 31751fa44ecaSJames Bottomley 317665f27f38SDavid Howells INIT_WORK(&ew->work, fn); 31771fa44ecaSJames Bottomley schedule_work(&ew->work); 31781fa44ecaSJames Bottomley 31791fa44ecaSJames Bottomley return 1; 31801fa44ecaSJames Bottomley } 31811fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 31821fa44ecaSJames Bottomley 31837a4e344cSTejun Heo /** 31847a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 31857a4e344cSTejun Heo * @attrs: workqueue_attrs to free 31867a4e344cSTejun Heo * 31877a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 31887a4e344cSTejun Heo */ 31897a4e344cSTejun Heo void free_workqueue_attrs(struct workqueue_attrs *attrs) 31907a4e344cSTejun Heo { 31917a4e344cSTejun Heo if (attrs) { 31927a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 31937a4e344cSTejun Heo kfree(attrs); 31947a4e344cSTejun Heo } 31957a4e344cSTejun Heo } 31967a4e344cSTejun Heo 31977a4e344cSTejun Heo /** 31987a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 31997a4e344cSTejun Heo * @gfp_mask: allocation mask to use 32007a4e344cSTejun Heo * 32017a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3202d185af30SYacine Belkadi * return it. 3203d185af30SYacine Belkadi * 3204d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 32057a4e344cSTejun Heo */ 32067a4e344cSTejun Heo struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask) 32077a4e344cSTejun Heo { 32087a4e344cSTejun Heo struct workqueue_attrs *attrs; 32097a4e344cSTejun Heo 32107a4e344cSTejun Heo attrs = kzalloc(sizeof(*attrs), gfp_mask); 32117a4e344cSTejun Heo if (!attrs) 32127a4e344cSTejun Heo goto fail; 32137a4e344cSTejun Heo if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask)) 32147a4e344cSTejun Heo goto fail; 32157a4e344cSTejun Heo 321613e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 32177a4e344cSTejun Heo return attrs; 32187a4e344cSTejun Heo fail: 32197a4e344cSTejun Heo free_workqueue_attrs(attrs); 32207a4e344cSTejun Heo return NULL; 32217a4e344cSTejun Heo } 32227a4e344cSTejun Heo 322329c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 322429c91e99STejun Heo const struct workqueue_attrs *from) 322529c91e99STejun Heo { 322629c91e99STejun Heo to->nice = from->nice; 322729c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 32282865a8fbSShaohua Li /* 32292865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 32302865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 32312865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 32322865a8fbSShaohua Li */ 32332865a8fbSShaohua Li to->no_numa = from->no_numa; 323429c91e99STejun Heo } 323529c91e99STejun Heo 323629c91e99STejun Heo /* hash value of the content of @attr */ 323729c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 323829c91e99STejun Heo { 323929c91e99STejun Heo u32 hash = 0; 324029c91e99STejun Heo 324129c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 324213e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 324313e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 324429c91e99STejun Heo return hash; 324529c91e99STejun Heo } 324629c91e99STejun Heo 324729c91e99STejun Heo /* content equality test */ 324829c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 324929c91e99STejun Heo const struct workqueue_attrs *b) 325029c91e99STejun Heo { 325129c91e99STejun Heo if (a->nice != b->nice) 325229c91e99STejun Heo return false; 325329c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 325429c91e99STejun Heo return false; 325529c91e99STejun Heo return true; 325629c91e99STejun Heo } 325729c91e99STejun Heo 32587a4e344cSTejun Heo /** 32597a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 32607a4e344cSTejun Heo * @pool: worker_pool to initialize 32617a4e344cSTejun Heo * 3262402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3263d185af30SYacine Belkadi * 3264d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 326529c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 326629c91e99STejun Heo * on @pool safely to release it. 32677a4e344cSTejun Heo */ 32687a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 32694e1a1f9aSTejun Heo { 32704e1a1f9aSTejun Heo spin_lock_init(&pool->lock); 327129c91e99STejun Heo pool->id = -1; 327229c91e99STejun Heo pool->cpu = -1; 3273f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 32744e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 327582607adcSTejun Heo pool->watchdog_ts = jiffies; 32764e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 32774e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 32784e1a1f9aSTejun Heo hash_init(pool->busy_hash); 32794e1a1f9aSTejun Heo 328032a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 32814e1a1f9aSTejun Heo 328232a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 32834e1a1f9aSTejun Heo 328492f9c5c4SLai Jiangshan mutex_init(&pool->attach_mutex); 3285da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 32867a4e344cSTejun Heo 32877cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 328829c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 328929c91e99STejun Heo pool->refcnt = 1; 329029c91e99STejun Heo 329129c91e99STejun Heo /* shouldn't fail above this point */ 32927a4e344cSTejun Heo pool->attrs = alloc_workqueue_attrs(GFP_KERNEL); 32937a4e344cSTejun Heo if (!pool->attrs) 32947a4e344cSTejun Heo return -ENOMEM; 32957a4e344cSTejun Heo return 0; 32964e1a1f9aSTejun Heo } 32974e1a1f9aSTejun Heo 3298e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3299e2dca7adSTejun Heo { 3300e2dca7adSTejun Heo struct workqueue_struct *wq = 3301e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3302e2dca7adSTejun Heo 3303e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3304e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3305e2dca7adSTejun Heo else 3306e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3307e2dca7adSTejun Heo 3308e2dca7adSTejun Heo kfree(wq->rescuer); 3309e2dca7adSTejun Heo kfree(wq); 3310e2dca7adSTejun Heo } 3311e2dca7adSTejun Heo 331229c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 331329c91e99STejun Heo { 331429c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 331529c91e99STejun Heo 33167cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 331729c91e99STejun Heo free_workqueue_attrs(pool->attrs); 331829c91e99STejun Heo kfree(pool); 331929c91e99STejun Heo } 332029c91e99STejun Heo 332129c91e99STejun Heo /** 332229c91e99STejun Heo * put_unbound_pool - put a worker_pool 332329c91e99STejun Heo * @pool: worker_pool to put 332429c91e99STejun Heo * 332529c91e99STejun Heo * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU 3326c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3327c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3328c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3329a892caccSTejun Heo * 3330a892caccSTejun Heo * Should be called with wq_pool_mutex held. 333129c91e99STejun Heo */ 333229c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 333329c91e99STejun Heo { 333460f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 333529c91e99STejun Heo struct worker *worker; 333629c91e99STejun Heo 3337a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3338a892caccSTejun Heo 3339a892caccSTejun Heo if (--pool->refcnt) 334029c91e99STejun Heo return; 334129c91e99STejun Heo 334229c91e99STejun Heo /* sanity checks */ 334361d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3344a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 334529c91e99STejun Heo return; 334629c91e99STejun Heo 334729c91e99STejun Heo /* release id and unhash */ 334829c91e99STejun Heo if (pool->id >= 0) 334929c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 335029c91e99STejun Heo hash_del(&pool->hash_node); 335129c91e99STejun Heo 3352c5aa87bbSTejun Heo /* 3353692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3354692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3355692b4825STejun Heo * manager and @pool gets freed with the flag set. 3356c5aa87bbSTejun Heo */ 335760f5a4bcSLai Jiangshan spin_lock_irq(&pool->lock); 3358692b4825STejun Heo wait_event_lock_irq(wq_manager_wait, 3359692b4825STejun Heo !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock); 3360692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3361692b4825STejun Heo 33621037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 336329c91e99STejun Heo destroy_worker(worker); 336429c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 336529c91e99STejun Heo spin_unlock_irq(&pool->lock); 336660f5a4bcSLai Jiangshan 336792f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 3368da028469SLai Jiangshan if (!list_empty(&pool->workers)) 336960f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 337092f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 337160f5a4bcSLai Jiangshan 337260f5a4bcSLai Jiangshan if (pool->detach_completion) 337360f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 337460f5a4bcSLai Jiangshan 337529c91e99STejun Heo /* shut down the timers */ 337629c91e99STejun Heo del_timer_sync(&pool->idle_timer); 337729c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 337829c91e99STejun Heo 337929c91e99STejun Heo /* sched-RCU protected to allow dereferences from get_work_pool() */ 338029c91e99STejun Heo call_rcu_sched(&pool->rcu, rcu_free_pool); 338129c91e99STejun Heo } 338229c91e99STejun Heo 338329c91e99STejun Heo /** 338429c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 338529c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 338629c91e99STejun Heo * 338729c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 338829c91e99STejun Heo * reference count and return it. If there already is a matching 338929c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3390d185af30SYacine Belkadi * create a new one. 3391a892caccSTejun Heo * 3392a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3393d185af30SYacine Belkadi * 3394d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3395d185af30SYacine Belkadi * On failure, %NULL. 339629c91e99STejun Heo */ 339729c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 339829c91e99STejun Heo { 339929c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 340029c91e99STejun Heo struct worker_pool *pool; 3401f3f90ad4STejun Heo int node; 3402e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 340329c91e99STejun Heo 3404a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 340529c91e99STejun Heo 340629c91e99STejun Heo /* do we already have a matching pool? */ 340729c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 340829c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 340929c91e99STejun Heo pool->refcnt++; 34103fb1823cSLai Jiangshan return pool; 341129c91e99STejun Heo } 341229c91e99STejun Heo } 341329c91e99STejun Heo 3414e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3415e2273584SXunlei Pang if (wq_numa_enabled) { 3416e2273584SXunlei Pang for_each_node(node) { 3417e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3418e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3419e2273584SXunlei Pang target_node = node; 3420e2273584SXunlei Pang break; 3421e2273584SXunlei Pang } 3422e2273584SXunlei Pang } 3423e2273584SXunlei Pang } 3424e2273584SXunlei Pang 342529c91e99STejun Heo /* nope, create a new one */ 3426e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 342729c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 342829c91e99STejun Heo goto fail; 342929c91e99STejun Heo 34308864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 343129c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3432e2273584SXunlei Pang pool->node = target_node; 343329c91e99STejun Heo 34342865a8fbSShaohua Li /* 34352865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 34362865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 34372865a8fbSShaohua Li */ 34382865a8fbSShaohua Li pool->attrs->no_numa = false; 34392865a8fbSShaohua Li 344029c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 344129c91e99STejun Heo goto fail; 344229c91e99STejun Heo 344329c91e99STejun Heo /* create and start the initial worker */ 34443347fa09STejun Heo if (wq_online && !create_worker(pool)) 344529c91e99STejun Heo goto fail; 344629c91e99STejun Heo 344729c91e99STejun Heo /* install */ 344829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 34493fb1823cSLai Jiangshan 345029c91e99STejun Heo return pool; 345129c91e99STejun Heo fail: 345229c91e99STejun Heo if (pool) 345329c91e99STejun Heo put_unbound_pool(pool); 345429c91e99STejun Heo return NULL; 345529c91e99STejun Heo } 345629c91e99STejun Heo 34578864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 34588864b4e5STejun Heo { 34598864b4e5STejun Heo kmem_cache_free(pwq_cache, 34608864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 34618864b4e5STejun Heo } 34628864b4e5STejun Heo 34638864b4e5STejun Heo /* 34648864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 34658864b4e5STejun Heo * and needs to be destroyed. 34668864b4e5STejun Heo */ 34678864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 34688864b4e5STejun Heo { 34698864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 34708864b4e5STejun Heo unbound_release_work); 34718864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 34728864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3473bc0caf09STejun Heo bool is_last; 34748864b4e5STejun Heo 34758864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 34768864b4e5STejun Heo return; 34778864b4e5STejun Heo 34783c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 34798864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3480bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 34813c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 34828864b4e5STejun Heo 3483a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 34848864b4e5STejun Heo put_unbound_pool(pool); 3485a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3486a892caccSTejun Heo 34878864b4e5STejun Heo call_rcu_sched(&pwq->rcu, rcu_free_pwq); 34888864b4e5STejun Heo 34898864b4e5STejun Heo /* 34908864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3491e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 34928864b4e5STejun Heo */ 3493e2dca7adSTejun Heo if (is_last) 3494e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 34956029a918STejun Heo } 34968864b4e5STejun Heo 34970fbd95aaSTejun Heo /** 3498699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 34990fbd95aaSTejun Heo * @pwq: target pool_workqueue 35000fbd95aaSTejun Heo * 3501699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3502699ce097STejun Heo * workqueue's saved_max_active and activate delayed work items 3503699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 35040fbd95aaSTejun Heo */ 3505699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 35060fbd95aaSTejun Heo { 3507699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3508699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 35093347fa09STejun Heo unsigned long flags; 3510699ce097STejun Heo 3511699ce097STejun Heo /* for @wq->saved_max_active */ 3512a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3513699ce097STejun Heo 3514699ce097STejun Heo /* fast exit for non-freezable wqs */ 3515699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3516699ce097STejun Heo return; 3517699ce097STejun Heo 35183347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 35193347fa09STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 3520699ce097STejun Heo 352174b414eaSLai Jiangshan /* 352274b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 352374b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 352474b414eaSLai Jiangshan * is updated and visible. 352574b414eaSLai Jiangshan */ 352674b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 3527699ce097STejun Heo pwq->max_active = wq->saved_max_active; 35280fbd95aaSTejun Heo 35290fbd95aaSTejun Heo while (!list_empty(&pwq->delayed_works) && 35300fbd95aaSTejun Heo pwq->nr_active < pwq->max_active) 35310fbd95aaSTejun Heo pwq_activate_first_delayed(pwq); 3532951a078aSLai Jiangshan 3533951a078aSLai Jiangshan /* 3534951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 3535951a078aSLai Jiangshan * max_active is bumped. It's a slow path. Do it always. 3536951a078aSLai Jiangshan */ 3537951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3538699ce097STejun Heo } else { 3539699ce097STejun Heo pwq->max_active = 0; 3540699ce097STejun Heo } 3541699ce097STejun Heo 35423347fa09STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 35430fbd95aaSTejun Heo } 35440fbd95aaSTejun Heo 3545e50aba9aSTejun Heo /* initialize newly alloced @pwq which is associated with @wq and @pool */ 3546f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3547f147f29eSTejun Heo struct worker_pool *pool) 3548d2c1d404STejun Heo { 3549d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3550d2c1d404STejun Heo 3551e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3552e50aba9aSTejun Heo 3553d2c1d404STejun Heo pwq->pool = pool; 3554d2c1d404STejun Heo pwq->wq = wq; 3555d2c1d404STejun Heo pwq->flush_color = -1; 35568864b4e5STejun Heo pwq->refcnt = 1; 3557d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->delayed_works); 35581befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3559d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 35608864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3561f147f29eSTejun Heo } 3562d2c1d404STejun Heo 3563f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 35641befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3565f147f29eSTejun Heo { 3566f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3567f147f29eSTejun Heo 3568f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 356975ccf595STejun Heo 35701befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 35711befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 35721befcf30STejun Heo return; 35731befcf30STejun Heo 357429b1cb41SLai Jiangshan /* set the matching work_color */ 357575ccf595STejun Heo pwq->work_color = wq->work_color; 3576983ca25eSTejun Heo 3577983ca25eSTejun Heo /* sync max_active to the current setting */ 3578983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3579983ca25eSTejun Heo 3580983ca25eSTejun Heo /* link in @pwq */ 35819e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3582df2d5ae4STejun Heo } 35836029a918STejun Heo 3584f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3585f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3586f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3587f147f29eSTejun Heo { 3588f147f29eSTejun Heo struct worker_pool *pool; 3589f147f29eSTejun Heo struct pool_workqueue *pwq; 3590f147f29eSTejun Heo 3591f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3592f147f29eSTejun Heo 3593f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3594f147f29eSTejun Heo if (!pool) 3595f147f29eSTejun Heo return NULL; 3596f147f29eSTejun Heo 3597e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3598f147f29eSTejun Heo if (!pwq) { 3599f147f29eSTejun Heo put_unbound_pool(pool); 3600f147f29eSTejun Heo return NULL; 3601f147f29eSTejun Heo } 3602f147f29eSTejun Heo 3603f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3604f147f29eSTejun Heo return pwq; 3605d2c1d404STejun Heo } 3606d2c1d404STejun Heo 36074c16bd32STejun Heo /** 360830186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3609042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 36104c16bd32STejun Heo * @node: the target NUMA node 36114c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 36124c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 36134c16bd32STejun Heo * 36144c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 36154c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3616d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 36174c16bd32STejun Heo * 36184c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 36194c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 36204c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 36214c16bd32STejun Heo * @attrs->cpumask. 36224c16bd32STejun Heo * 36234c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 36244c16bd32STejun Heo * stable. 3625d185af30SYacine Belkadi * 3626d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3627d185af30SYacine Belkadi * %false if equal. 36284c16bd32STejun Heo */ 36294c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 36304c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 36314c16bd32STejun Heo { 3632d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 36334c16bd32STejun Heo goto use_dfl; 36344c16bd32STejun Heo 36354c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 36364c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 36374c16bd32STejun Heo if (cpu_going_down >= 0) 36384c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 36394c16bd32STejun Heo 36404c16bd32STejun Heo if (cpumask_empty(cpumask)) 36414c16bd32STejun Heo goto use_dfl; 36424c16bd32STejun Heo 36434c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 36444c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 36451ad0f0a7SMichael Bringmann 36461ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 36471ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 36481ad0f0a7SMichael Bringmann "possible intersect\n"); 36491ad0f0a7SMichael Bringmann return false; 36501ad0f0a7SMichael Bringmann } 36511ad0f0a7SMichael Bringmann 36524c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 36534c16bd32STejun Heo 36544c16bd32STejun Heo use_dfl: 36554c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 36564c16bd32STejun Heo return false; 36574c16bd32STejun Heo } 36584c16bd32STejun Heo 36591befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 36601befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 36611befcf30STejun Heo int node, 36621befcf30STejun Heo struct pool_workqueue *pwq) 36631befcf30STejun Heo { 36641befcf30STejun Heo struct pool_workqueue *old_pwq; 36651befcf30STejun Heo 36665b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 36671befcf30STejun Heo lockdep_assert_held(&wq->mutex); 36681befcf30STejun Heo 36691befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 36701befcf30STejun Heo link_pwq(pwq); 36711befcf30STejun Heo 36721befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 36731befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 36741befcf30STejun Heo return old_pwq; 36751befcf30STejun Heo } 36761befcf30STejun Heo 36772d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 36782d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 36792d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 36802d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3681042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 36822d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 36832d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 36842d5f0764SLai Jiangshan }; 36852d5f0764SLai Jiangshan 36862d5f0764SLai Jiangshan /* free the resources after success or abort */ 36872d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 36882d5f0764SLai Jiangshan { 36892d5f0764SLai Jiangshan if (ctx) { 36902d5f0764SLai Jiangshan int node; 36912d5f0764SLai Jiangshan 36922d5f0764SLai Jiangshan for_each_node(node) 36932d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 36942d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 36952d5f0764SLai Jiangshan 36962d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 36972d5f0764SLai Jiangshan 36982d5f0764SLai Jiangshan kfree(ctx); 36992d5f0764SLai Jiangshan } 37002d5f0764SLai Jiangshan } 37012d5f0764SLai Jiangshan 37022d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 37032d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 37042d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 37052d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 37062d5f0764SLai Jiangshan { 37072d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 37082d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 37092d5f0764SLai Jiangshan int node; 37102d5f0764SLai Jiangshan 37112d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 37122d5f0764SLai Jiangshan 37132d5f0764SLai Jiangshan ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]), 37142d5f0764SLai Jiangshan GFP_KERNEL); 37152d5f0764SLai Jiangshan 37162d5f0764SLai Jiangshan new_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37172d5f0764SLai Jiangshan tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL); 37182d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 37192d5f0764SLai Jiangshan goto out_free; 37202d5f0764SLai Jiangshan 3721042f7df1SLai Jiangshan /* 3722042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3723042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3724042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3725042f7df1SLai Jiangshan */ 37262d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3727b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3728042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3729042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 37302d5f0764SLai Jiangshan 37312d5f0764SLai Jiangshan /* 37322d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 37332d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 37342d5f0764SLai Jiangshan * pools. 37352d5f0764SLai Jiangshan */ 37362d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 37372d5f0764SLai Jiangshan 37382d5f0764SLai Jiangshan /* 37392d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 37402d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 37412d5f0764SLai Jiangshan * it even if we don't use it immediately. 37422d5f0764SLai Jiangshan */ 37432d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 37442d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 37452d5f0764SLai Jiangshan goto out_free; 37462d5f0764SLai Jiangshan 37472d5f0764SLai Jiangshan for_each_node(node) { 3748042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 37492d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 37502d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 37512d5f0764SLai Jiangshan goto out_free; 37522d5f0764SLai Jiangshan } else { 37532d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 37542d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 37552d5f0764SLai Jiangshan } 37562d5f0764SLai Jiangshan } 37572d5f0764SLai Jiangshan 3758042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 3759042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3760042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 37612d5f0764SLai Jiangshan ctx->attrs = new_attrs; 3762042f7df1SLai Jiangshan 37632d5f0764SLai Jiangshan ctx->wq = wq; 37642d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37652d5f0764SLai Jiangshan return ctx; 37662d5f0764SLai Jiangshan 37672d5f0764SLai Jiangshan out_free: 37682d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 37692d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 37702d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 37712d5f0764SLai Jiangshan return NULL; 37722d5f0764SLai Jiangshan } 37732d5f0764SLai Jiangshan 37742d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 37752d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 37762d5f0764SLai Jiangshan { 37772d5f0764SLai Jiangshan int node; 37782d5f0764SLai Jiangshan 37792d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 37802d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 37812d5f0764SLai Jiangshan 37822d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 37832d5f0764SLai Jiangshan 37842d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 37852d5f0764SLai Jiangshan for_each_node(node) 37862d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 37872d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 37882d5f0764SLai Jiangshan 37892d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 37902d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 37912d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 37922d5f0764SLai Jiangshan 37932d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 37942d5f0764SLai Jiangshan } 37952d5f0764SLai Jiangshan 3796a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 3797a0111cf6SLai Jiangshan { 3798a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 3799a0111cf6SLai Jiangshan get_online_cpus(); 3800a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 3801a0111cf6SLai Jiangshan } 3802a0111cf6SLai Jiangshan 3803a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 3804a0111cf6SLai Jiangshan { 3805a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 3806a0111cf6SLai Jiangshan put_online_cpus(); 3807a0111cf6SLai Jiangshan } 3808a0111cf6SLai Jiangshan 3809a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 3810a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 3811a0111cf6SLai Jiangshan { 3812a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 3813a0111cf6SLai Jiangshan 3814a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 3815a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 3816a0111cf6SLai Jiangshan return -EINVAL; 3817a0111cf6SLai Jiangshan 3818a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 38190a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 38200a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 3821a0111cf6SLai Jiangshan return -EINVAL; 3822a0111cf6SLai Jiangshan 38230a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 38240a94efb5STejun Heo } 38250a94efb5STejun Heo 3826a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 38276201171eSwanghaibin if (!ctx) 38286201171eSwanghaibin return -ENOMEM; 3829a0111cf6SLai Jiangshan 3830a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 3831a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 3832a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 3833a0111cf6SLai Jiangshan 38346201171eSwanghaibin return 0; 3835a0111cf6SLai Jiangshan } 3836a0111cf6SLai Jiangshan 38379e8cd2f5STejun Heo /** 38389e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 38399e8cd2f5STejun Heo * @wq: the target workqueue 38409e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 38419e8cd2f5STejun Heo * 38424c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 38434c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 38444c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 38454c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 38464c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 38474c16bd32STejun Heo * back-to-back will stay on its current pwq. 38489e8cd2f5STejun Heo * 3849d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 3850d185af30SYacine Belkadi * 3851d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 38529e8cd2f5STejun Heo */ 38539e8cd2f5STejun Heo int apply_workqueue_attrs(struct workqueue_struct *wq, 38549e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 38559e8cd2f5STejun Heo { 3856a0111cf6SLai Jiangshan int ret; 38579e8cd2f5STejun Heo 3858a0111cf6SLai Jiangshan apply_wqattrs_lock(); 3859a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 3860a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 38612d5f0764SLai Jiangshan 38622d5f0764SLai Jiangshan return ret; 38639e8cd2f5STejun Heo } 38646106c0f8SNeilBrown EXPORT_SYMBOL_GPL(apply_workqueue_attrs); 38659e8cd2f5STejun Heo 38664c16bd32STejun Heo /** 38674c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 38684c16bd32STejun Heo * @wq: the target workqueue 38694c16bd32STejun Heo * @cpu: the CPU coming up or going down 38704c16bd32STejun Heo * @online: whether @cpu is coming up or going down 38714c16bd32STejun Heo * 38724c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 38734c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 38744c16bd32STejun Heo * @wq accordingly. 38754c16bd32STejun Heo * 38764c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 38774c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 38784c16bd32STejun Heo * correct. 38794c16bd32STejun Heo * 38804c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 38814c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 38824c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 38834c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 38844c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 38854c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 38864c16bd32STejun Heo * CPU_DOWN_PREPARE. 38874c16bd32STejun Heo */ 38884c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 38894c16bd32STejun Heo bool online) 38904c16bd32STejun Heo { 38914c16bd32STejun Heo int node = cpu_to_node(cpu); 38924c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 38934c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 38944c16bd32STejun Heo struct workqueue_attrs *target_attrs; 38954c16bd32STejun Heo cpumask_t *cpumask; 38964c16bd32STejun Heo 38974c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 38984c16bd32STejun Heo 3899f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 3900f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 39014c16bd32STejun Heo return; 39024c16bd32STejun Heo 39034c16bd32STejun Heo /* 39044c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 39054c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 39064c16bd32STejun Heo * CPU hotplug exclusion. 39074c16bd32STejun Heo */ 39084c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 39094c16bd32STejun Heo cpumask = target_attrs->cpumask; 39104c16bd32STejun Heo 39114c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 39124c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 39134c16bd32STejun Heo 39144c16bd32STejun Heo /* 39154c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 3916042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 3917042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 3918042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 39194c16bd32STejun Heo */ 3920042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 39214c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 3922f7142ed4SLai Jiangshan return; 39234c16bd32STejun Heo } else { 39244c16bd32STejun Heo goto use_dfl_pwq; 39254c16bd32STejun Heo } 39264c16bd32STejun Heo 39274c16bd32STejun Heo /* create a new pwq */ 39284c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 39294c16bd32STejun Heo if (!pwq) { 39302d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 39314c16bd32STejun Heo wq->name); 393277f300b1SDaeseok Youn goto use_dfl_pwq; 39334c16bd32STejun Heo } 39344c16bd32STejun Heo 3935f7142ed4SLai Jiangshan /* Install the new pwq. */ 39364c16bd32STejun Heo mutex_lock(&wq->mutex); 39374c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 39384c16bd32STejun Heo goto out_unlock; 39394c16bd32STejun Heo 39404c16bd32STejun Heo use_dfl_pwq: 3941f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 39424c16bd32STejun Heo spin_lock_irq(&wq->dfl_pwq->pool->lock); 39434c16bd32STejun Heo get_pwq(wq->dfl_pwq); 39444c16bd32STejun Heo spin_unlock_irq(&wq->dfl_pwq->pool->lock); 39454c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 39464c16bd32STejun Heo out_unlock: 39474c16bd32STejun Heo mutex_unlock(&wq->mutex); 39484c16bd32STejun Heo put_pwq_unlocked(old_pwq); 39494c16bd32STejun Heo } 39504c16bd32STejun Heo 395130cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 39521da177e4SLinus Torvalds { 395349e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 39548a2b7538STejun Heo int cpu, ret; 3955e1d8aa9fSFrederic Weisbecker 395630cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 3957420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 3958420c0ddbSTejun Heo if (!wq->cpu_pwqs) 395930cdf249STejun Heo return -ENOMEM; 396030cdf249STejun Heo 396130cdf249STejun Heo for_each_possible_cpu(cpu) { 39627fb98ea7STejun Heo struct pool_workqueue *pwq = 39637fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 39647a62c2c8STejun Heo struct worker_pool *cpu_pools = 3965f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 396630cdf249STejun Heo 3967f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 3968f147f29eSTejun Heo 3969f147f29eSTejun Heo mutex_lock(&wq->mutex); 39701befcf30STejun Heo link_pwq(pwq); 3971f147f29eSTejun Heo mutex_unlock(&wq->mutex); 397230cdf249STejun Heo } 397330cdf249STejun Heo return 0; 39748a2b7538STejun Heo } else if (wq->flags & __WQ_ORDERED) { 39758a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 39768a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 39778a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 39788a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 39798a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 39808a2b7538STejun Heo return ret; 39819e8cd2f5STejun Heo } else { 39829e8cd2f5STejun Heo return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 39839e8cd2f5STejun Heo } 39840f900049STejun Heo } 39850f900049STejun Heo 3986f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3987f3421797STejun Heo const char *name) 3988b71ab8c2STejun Heo { 3989f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3990f3421797STejun Heo 3991f3421797STejun Heo if (max_active < 1 || max_active > lim) 3992044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 3993f3421797STejun Heo max_active, name, 1, lim); 3994b71ab8c2STejun Heo 3995f3421797STejun Heo return clamp_val(max_active, 1, lim); 3996b71ab8c2STejun Heo } 3997b71ab8c2STejun Heo 3998983c7515STejun Heo /* 3999983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4000983c7515STejun Heo * to guarantee forward progress. 4001983c7515STejun Heo */ 4002983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4003983c7515STejun Heo { 4004983c7515STejun Heo struct worker *rescuer; 4005983c7515STejun Heo int ret; 4006983c7515STejun Heo 4007983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4008983c7515STejun Heo return 0; 4009983c7515STejun Heo 4010983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4011983c7515STejun Heo if (!rescuer) 4012983c7515STejun Heo return -ENOMEM; 4013983c7515STejun Heo 4014983c7515STejun Heo rescuer->rescue_wq = wq; 4015983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4016983c7515STejun Heo ret = PTR_ERR_OR_ZERO(rescuer->task); 4017983c7515STejun Heo if (ret) { 4018983c7515STejun Heo kfree(rescuer); 4019983c7515STejun Heo return ret; 4020983c7515STejun Heo } 4021983c7515STejun Heo 4022983c7515STejun Heo wq->rescuer = rescuer; 4023983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4024983c7515STejun Heo wake_up_process(rescuer->task); 4025983c7515STejun Heo 4026983c7515STejun Heo return 0; 4027983c7515STejun Heo } 4028983c7515STejun Heo 4029b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 403097e37d7bSTejun Heo unsigned int flags, 40311e19ffc6STejun Heo int max_active, 4032eb13ba87SJohannes Berg struct lock_class_key *key, 4033b196be89STejun Heo const char *lock_name, ...) 40343af24433SOleg Nesterov { 4035df2d5ae4STejun Heo size_t tbl_size = 0; 4036ecf6881fSTejun Heo va_list args; 40373af24433SOleg Nesterov struct workqueue_struct *wq; 403849e3cf44STejun Heo struct pool_workqueue *pwq; 4039b196be89STejun Heo 40405c0338c6STejun Heo /* 40415c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 40425c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 40435c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 40445c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 40455c0338c6STejun Heo * on NUMA. 40465c0338c6STejun Heo */ 40475c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 40485c0338c6STejun Heo flags |= __WQ_ORDERED; 40495c0338c6STejun Heo 4050cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4051cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4052cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4053cee22a15SViresh Kumar 4054ecf6881fSTejun Heo /* allocate wq and format name */ 4055df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4056ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4057df2d5ae4STejun Heo 4058df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4059b196be89STejun Heo if (!wq) 4060d2c1d404STejun Heo return NULL; 4061b196be89STejun Heo 40626029a918STejun Heo if (flags & WQ_UNBOUND) { 40636029a918STejun Heo wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL); 40646029a918STejun Heo if (!wq->unbound_attrs) 40656029a918STejun Heo goto err_free_wq; 40666029a918STejun Heo } 40676029a918STejun Heo 4068ecf6881fSTejun Heo va_start(args, lock_name); 4069ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4070b196be89STejun Heo va_end(args); 40713af24433SOleg Nesterov 4072d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4073b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 40743af24433SOleg Nesterov 4075b196be89STejun Heo /* init wq */ 407697e37d7bSTejun Heo wq->flags = flags; 4077a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 40783c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4079112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 408030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 408173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 408273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4083493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 40843af24433SOleg Nesterov 4085eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 4086cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 40873af24433SOleg Nesterov 408830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 4089d2c1d404STejun Heo goto err_free_wq; 40901537663fSTejun Heo 409140c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4092d2c1d404STejun Heo goto err_destroy; 4093e22bee78STejun Heo 4094226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4095226223abSTejun Heo goto err_destroy; 4096226223abSTejun Heo 40976af8bf3dSOleg Nesterov /* 409868e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 409968e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 410068e13a67SLai Jiangshan * list. 41016af8bf3dSOleg Nesterov */ 410268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4103a0a1a5fdSTejun Heo 4104a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 410549e3cf44STejun Heo for_each_pwq(pwq, wq) 4106699ce097STejun Heo pwq_adjust_max_active(pwq); 4107a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4108a0a1a5fdSTejun Heo 4109e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4110a0a1a5fdSTejun Heo 411168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41123af24433SOleg Nesterov 41133af24433SOleg Nesterov return wq; 4114d2c1d404STejun Heo 4115d2c1d404STejun Heo err_free_wq: 41166029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 41174690c4abSTejun Heo kfree(wq); 4118d2c1d404STejun Heo return NULL; 4119d2c1d404STejun Heo err_destroy: 4120d2c1d404STejun Heo destroy_workqueue(wq); 41214690c4abSTejun Heo return NULL; 41221da177e4SLinus Torvalds } 4123d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 41241da177e4SLinus Torvalds 41253af24433SOleg Nesterov /** 41263af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 41273af24433SOleg Nesterov * @wq: target workqueue 41283af24433SOleg Nesterov * 41293af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 41303af24433SOleg Nesterov */ 41313af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 41323af24433SOleg Nesterov { 413349e3cf44STejun Heo struct pool_workqueue *pwq; 41344c16bd32STejun Heo int node; 41353af24433SOleg Nesterov 41369c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 41379c5a2ba7STejun Heo drain_workqueue(wq); 4138c8efcc25STejun Heo 41396183c009STejun Heo /* sanity checks */ 4140b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 414149e3cf44STejun Heo for_each_pwq(pwq, wq) { 41426183c009STejun Heo int i; 41436183c009STejun Heo 414476af4d93STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) { 414576af4d93STejun Heo if (WARN_ON(pwq->nr_in_flight[i])) { 4146b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4147fa07fb6aSTejun Heo show_workqueue_state(); 41486183c009STejun Heo return; 414976af4d93STejun Heo } 415076af4d93STejun Heo } 415176af4d93STejun Heo 41525c529597SLai Jiangshan if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) || 41538864b4e5STejun Heo WARN_ON(pwq->nr_active) || 415476af4d93STejun Heo WARN_ON(!list_empty(&pwq->delayed_works))) { 4155b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4156fa07fb6aSTejun Heo show_workqueue_state(); 41576183c009STejun Heo return; 41586183c009STejun Heo } 415976af4d93STejun Heo } 4160b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 41616183c009STejun Heo 4162a0a1a5fdSTejun Heo /* 4163a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4164a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4165a0a1a5fdSTejun Heo */ 416668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4167e2dca7adSTejun Heo list_del_rcu(&wq->list); 416868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 41693af24433SOleg Nesterov 4170226223abSTejun Heo workqueue_sysfs_unregister(wq); 4171226223abSTejun Heo 4172e2dca7adSTejun Heo if (wq->rescuer) 4173e22bee78STejun Heo kthread_stop(wq->rescuer->task); 4174e22bee78STejun Heo 41758864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 417629c91e99STejun Heo /* 41778864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4178e2dca7adSTejun Heo * schedule RCU free. 417929c91e99STejun Heo */ 4180e2dca7adSTejun Heo call_rcu_sched(&wq->rcu, rcu_free_wq); 41818864b4e5STejun Heo } else { 41828864b4e5STejun Heo /* 41838864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 41844c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 41854c16bd32STejun Heo * @wq will be freed when the last pwq is released. 41868864b4e5STejun Heo */ 41874c16bd32STejun Heo for_each_node(node) { 41884c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 41894c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 41904c16bd32STejun Heo put_pwq_unlocked(pwq); 41914c16bd32STejun Heo } 41924c16bd32STejun Heo 41934c16bd32STejun Heo /* 41944c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 41954c16bd32STejun Heo * put. Don't access it afterwards. 41964c16bd32STejun Heo */ 41974c16bd32STejun Heo pwq = wq->dfl_pwq; 41984c16bd32STejun Heo wq->dfl_pwq = NULL; 4199dce90d47STejun Heo put_pwq_unlocked(pwq); 420029c91e99STejun Heo } 42013af24433SOleg Nesterov } 42023af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 42033af24433SOleg Nesterov 4204dcd989cbSTejun Heo /** 4205dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4206dcd989cbSTejun Heo * @wq: target workqueue 4207dcd989cbSTejun Heo * @max_active: new max_active value. 4208dcd989cbSTejun Heo * 4209dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4210dcd989cbSTejun Heo * 4211dcd989cbSTejun Heo * CONTEXT: 4212dcd989cbSTejun Heo * Don't call from IRQ context. 4213dcd989cbSTejun Heo */ 4214dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4215dcd989cbSTejun Heo { 421649e3cf44STejun Heo struct pool_workqueue *pwq; 4217dcd989cbSTejun Heo 42188719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 42190a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 42208719dceaSTejun Heo return; 42218719dceaSTejun Heo 4222f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4223dcd989cbSTejun Heo 4224a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4225dcd989cbSTejun Heo 42260a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4227dcd989cbSTejun Heo wq->saved_max_active = max_active; 4228dcd989cbSTejun Heo 4229699ce097STejun Heo for_each_pwq(pwq, wq) 4230699ce097STejun Heo pwq_adjust_max_active(pwq); 4231dcd989cbSTejun Heo 4232a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4233dcd989cbSTejun Heo } 4234dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4235dcd989cbSTejun Heo 4236dcd989cbSTejun Heo /** 423727d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 423827d4ee03SLukas Wunner * 423927d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 424027d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 424127d4ee03SLukas Wunner * 424227d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 424327d4ee03SLukas Wunner */ 424427d4ee03SLukas Wunner struct work_struct *current_work(void) 424527d4ee03SLukas Wunner { 424627d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 424727d4ee03SLukas Wunner 424827d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 424927d4ee03SLukas Wunner } 425027d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 425127d4ee03SLukas Wunner 425227d4ee03SLukas Wunner /** 4253e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4254e6267616STejun Heo * 4255e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4256e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4257d185af30SYacine Belkadi * 4258d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4259e6267616STejun Heo */ 4260e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4261e6267616STejun Heo { 4262e6267616STejun Heo struct worker *worker = current_wq_worker(); 4263e6267616STejun Heo 42646a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4265e6267616STejun Heo } 4266e6267616STejun Heo 4267e6267616STejun Heo /** 4268dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4269dcd989cbSTejun Heo * @cpu: CPU in question 4270dcd989cbSTejun Heo * @wq: target workqueue 4271dcd989cbSTejun Heo * 4272dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4273dcd989cbSTejun Heo * no synchronization around this function and the test result is 4274dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4275dcd989cbSTejun Heo * 4276d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4277d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4278d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4279d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4280d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4281d3251859STejun Heo * 4282d185af30SYacine Belkadi * Return: 4283dcd989cbSTejun Heo * %true if congested, %false otherwise. 4284dcd989cbSTejun Heo */ 4285d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4286dcd989cbSTejun Heo { 42877fb98ea7STejun Heo struct pool_workqueue *pwq; 428876af4d93STejun Heo bool ret; 428976af4d93STejun Heo 429088109453SLai Jiangshan rcu_read_lock_sched(); 42917fb98ea7STejun Heo 4292d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4293d3251859STejun Heo cpu = smp_processor_id(); 4294d3251859STejun Heo 42957fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 42967fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 42977fb98ea7STejun Heo else 4298df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4299dcd989cbSTejun Heo 430076af4d93STejun Heo ret = !list_empty(&pwq->delayed_works); 430188109453SLai Jiangshan rcu_read_unlock_sched(); 430276af4d93STejun Heo 430376af4d93STejun Heo return ret; 4304dcd989cbSTejun Heo } 4305dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4306dcd989cbSTejun Heo 4307dcd989cbSTejun Heo /** 4308dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4309dcd989cbSTejun Heo * @work: the work to be tested 4310dcd989cbSTejun Heo * 4311dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4312dcd989cbSTejun Heo * synchronization around this function and the test result is 4313dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4314dcd989cbSTejun Heo * 4315d185af30SYacine Belkadi * Return: 4316dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4317dcd989cbSTejun Heo */ 4318dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4319dcd989cbSTejun Heo { 4320fa1b54e6STejun Heo struct worker_pool *pool; 4321dcd989cbSTejun Heo unsigned long flags; 4322dcd989cbSTejun Heo unsigned int ret = 0; 4323dcd989cbSTejun Heo 4324dcd989cbSTejun Heo if (work_pending(work)) 4325dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4326038366c5SLai Jiangshan 4327fa1b54e6STejun Heo local_irq_save(flags); 4328fa1b54e6STejun Heo pool = get_work_pool(work); 4329038366c5SLai Jiangshan if (pool) { 4330fa1b54e6STejun Heo spin_lock(&pool->lock); 4331c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4332dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4333fa1b54e6STejun Heo spin_unlock(&pool->lock); 4334038366c5SLai Jiangshan } 4335fa1b54e6STejun Heo local_irq_restore(flags); 4336dcd989cbSTejun Heo 4337dcd989cbSTejun Heo return ret; 4338dcd989cbSTejun Heo } 4339dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4340dcd989cbSTejun Heo 43413d1cb205STejun Heo /** 43423d1cb205STejun Heo * set_worker_desc - set description for the current work item 43433d1cb205STejun Heo * @fmt: printf-style format string 43443d1cb205STejun Heo * @...: arguments for the format string 43453d1cb205STejun Heo * 43463d1cb205STejun Heo * This function can be called by a running work function to describe what 43473d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 43483d1cb205STejun Heo * information will be printed out together to help debugging. The 43493d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 43503d1cb205STejun Heo */ 43513d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 43523d1cb205STejun Heo { 43533d1cb205STejun Heo struct worker *worker = current_wq_worker(); 43543d1cb205STejun Heo va_list args; 43553d1cb205STejun Heo 43563d1cb205STejun Heo if (worker) { 43573d1cb205STejun Heo va_start(args, fmt); 43583d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 43593d1cb205STejun Heo va_end(args); 43603d1cb205STejun Heo worker->desc_valid = true; 43613d1cb205STejun Heo } 43623d1cb205STejun Heo } 43633d1cb205STejun Heo 43643d1cb205STejun Heo /** 43653d1cb205STejun Heo * print_worker_info - print out worker information and description 43663d1cb205STejun Heo * @log_lvl: the log level to use when printing 43673d1cb205STejun Heo * @task: target task 43683d1cb205STejun Heo * 43693d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 43703d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 43713d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 43723d1cb205STejun Heo * 43733d1cb205STejun Heo * This function can be safely called on any task as long as the 43743d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 43753d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 43763d1cb205STejun Heo */ 43773d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 43783d1cb205STejun Heo { 43793d1cb205STejun Heo work_func_t *fn = NULL; 43803d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 43813d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 43823d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 43833d1cb205STejun Heo struct workqueue_struct *wq = NULL; 43843d1cb205STejun Heo bool desc_valid = false; 43853d1cb205STejun Heo struct worker *worker; 43863d1cb205STejun Heo 43873d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 43883d1cb205STejun Heo return; 43893d1cb205STejun Heo 43903d1cb205STejun Heo /* 43913d1cb205STejun Heo * This function is called without any synchronization and @task 43923d1cb205STejun Heo * could be in any state. Be careful with dereferences. 43933d1cb205STejun Heo */ 4394e700591aSPetr Mladek worker = kthread_probe_data(task); 43953d1cb205STejun Heo 43963d1cb205STejun Heo /* 43973d1cb205STejun Heo * Carefully copy the associated workqueue's workfn and name. Keep 43983d1cb205STejun Heo * the original last '\0' in case the original contains garbage. 43993d1cb205STejun Heo */ 44003d1cb205STejun Heo probe_kernel_read(&fn, &worker->current_func, sizeof(fn)); 44013d1cb205STejun Heo probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq)); 44023d1cb205STejun Heo probe_kernel_read(&wq, &pwq->wq, sizeof(wq)); 44033d1cb205STejun Heo probe_kernel_read(name, wq->name, sizeof(name) - 1); 44043d1cb205STejun Heo 44053d1cb205STejun Heo /* copy worker description */ 44063d1cb205STejun Heo probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid)); 44073d1cb205STejun Heo if (desc_valid) 44083d1cb205STejun Heo probe_kernel_read(desc, worker->desc, sizeof(desc) - 1); 44093d1cb205STejun Heo 44103d1cb205STejun Heo if (fn || name[0] || desc[0]) { 44113d1cb205STejun Heo printk("%sWorkqueue: %s %pf", log_lvl, name, fn); 44123d1cb205STejun Heo if (desc[0]) 44133d1cb205STejun Heo pr_cont(" (%s)", desc); 44143d1cb205STejun Heo pr_cont("\n"); 44153d1cb205STejun Heo } 44163d1cb205STejun Heo } 44173d1cb205STejun Heo 44183494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 44193494fc30STejun Heo { 44203494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 44213494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 44223494fc30STejun Heo pr_cont(" node=%d", pool->node); 44233494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 44243494fc30STejun Heo } 44253494fc30STejun Heo 44263494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 44273494fc30STejun Heo { 44283494fc30STejun Heo if (work->func == wq_barrier_func) { 44293494fc30STejun Heo struct wq_barrier *barr; 44303494fc30STejun Heo 44313494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 44323494fc30STejun Heo 44333494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 44343494fc30STejun Heo task_pid_nr(barr->task)); 44353494fc30STejun Heo } else { 44363494fc30STejun Heo pr_cont("%s %pf", comma ? "," : "", work->func); 44373494fc30STejun Heo } 44383494fc30STejun Heo } 44393494fc30STejun Heo 44403494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 44413494fc30STejun Heo { 44423494fc30STejun Heo struct worker_pool *pool = pwq->pool; 44433494fc30STejun Heo struct work_struct *work; 44443494fc30STejun Heo struct worker *worker; 44453494fc30STejun Heo bool has_in_flight = false, has_pending = false; 44463494fc30STejun Heo int bkt; 44473494fc30STejun Heo 44483494fc30STejun Heo pr_info(" pwq %d:", pool->id); 44493494fc30STejun Heo pr_cont_pool_info(pool); 44503494fc30STejun Heo 44513494fc30STejun Heo pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active, 44523494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 44533494fc30STejun Heo 44543494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44553494fc30STejun Heo if (worker->current_pwq == pwq) { 44563494fc30STejun Heo has_in_flight = true; 44573494fc30STejun Heo break; 44583494fc30STejun Heo } 44593494fc30STejun Heo } 44603494fc30STejun Heo if (has_in_flight) { 44613494fc30STejun Heo bool comma = false; 44623494fc30STejun Heo 44633494fc30STejun Heo pr_info(" in-flight:"); 44643494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 44653494fc30STejun Heo if (worker->current_pwq != pwq) 44663494fc30STejun Heo continue; 44673494fc30STejun Heo 44683494fc30STejun Heo pr_cont("%s %d%s:%pf", comma ? "," : "", 44693494fc30STejun Heo task_pid_nr(worker->task), 44703494fc30STejun Heo worker == pwq->wq->rescuer ? "(RESCUER)" : "", 44713494fc30STejun Heo worker->current_func); 44723494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 44733494fc30STejun Heo pr_cont_work(false, work); 44743494fc30STejun Heo comma = true; 44753494fc30STejun Heo } 44763494fc30STejun Heo pr_cont("\n"); 44773494fc30STejun Heo } 44783494fc30STejun Heo 44793494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44803494fc30STejun Heo if (get_work_pwq(work) == pwq) { 44813494fc30STejun Heo has_pending = true; 44823494fc30STejun Heo break; 44833494fc30STejun Heo } 44843494fc30STejun Heo } 44853494fc30STejun Heo if (has_pending) { 44863494fc30STejun Heo bool comma = false; 44873494fc30STejun Heo 44883494fc30STejun Heo pr_info(" pending:"); 44893494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 44903494fc30STejun Heo if (get_work_pwq(work) != pwq) 44913494fc30STejun Heo continue; 44923494fc30STejun Heo 44933494fc30STejun Heo pr_cont_work(comma, work); 44943494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 44953494fc30STejun Heo } 44963494fc30STejun Heo pr_cont("\n"); 44973494fc30STejun Heo } 44983494fc30STejun Heo 44993494fc30STejun Heo if (!list_empty(&pwq->delayed_works)) { 45003494fc30STejun Heo bool comma = false; 45013494fc30STejun Heo 45023494fc30STejun Heo pr_info(" delayed:"); 45033494fc30STejun Heo list_for_each_entry(work, &pwq->delayed_works, entry) { 45043494fc30STejun Heo pr_cont_work(comma, work); 45053494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 45063494fc30STejun Heo } 45073494fc30STejun Heo pr_cont("\n"); 45083494fc30STejun Heo } 45093494fc30STejun Heo } 45103494fc30STejun Heo 45113494fc30STejun Heo /** 45123494fc30STejun Heo * show_workqueue_state - dump workqueue state 45133494fc30STejun Heo * 45147b776af6SRoger Lu * Called from a sysrq handler or try_to_freeze_tasks() and prints out 45157b776af6SRoger Lu * all busy workqueues and pools. 45163494fc30STejun Heo */ 45173494fc30STejun Heo void show_workqueue_state(void) 45183494fc30STejun Heo { 45193494fc30STejun Heo struct workqueue_struct *wq; 45203494fc30STejun Heo struct worker_pool *pool; 45213494fc30STejun Heo unsigned long flags; 45223494fc30STejun Heo int pi; 45233494fc30STejun Heo 45243494fc30STejun Heo rcu_read_lock_sched(); 45253494fc30STejun Heo 45263494fc30STejun Heo pr_info("Showing busy workqueues and worker pools:\n"); 45273494fc30STejun Heo 45283494fc30STejun Heo list_for_each_entry_rcu(wq, &workqueues, list) { 45293494fc30STejun Heo struct pool_workqueue *pwq; 45303494fc30STejun Heo bool idle = true; 45313494fc30STejun Heo 45323494fc30STejun Heo for_each_pwq(pwq, wq) { 45333494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) { 45343494fc30STejun Heo idle = false; 45353494fc30STejun Heo break; 45363494fc30STejun Heo } 45373494fc30STejun Heo } 45383494fc30STejun Heo if (idle) 45393494fc30STejun Heo continue; 45403494fc30STejun Heo 45413494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 45423494fc30STejun Heo 45433494fc30STejun Heo for_each_pwq(pwq, wq) { 45443494fc30STejun Heo spin_lock_irqsave(&pwq->pool->lock, flags); 45453494fc30STejun Heo if (pwq->nr_active || !list_empty(&pwq->delayed_works)) 45463494fc30STejun Heo show_pwq(pwq); 45473494fc30STejun Heo spin_unlock_irqrestore(&pwq->pool->lock, flags); 454862635ea8SSergey Senozhatsky /* 454962635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 455062635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 455162635ea8SSergey Senozhatsky * hard lockup. 455262635ea8SSergey Senozhatsky */ 455362635ea8SSergey Senozhatsky touch_nmi_watchdog(); 45543494fc30STejun Heo } 45553494fc30STejun Heo } 45563494fc30STejun Heo 45573494fc30STejun Heo for_each_pool(pool, pi) { 45583494fc30STejun Heo struct worker *worker; 45593494fc30STejun Heo bool first = true; 45603494fc30STejun Heo 45613494fc30STejun Heo spin_lock_irqsave(&pool->lock, flags); 45623494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 45633494fc30STejun Heo goto next_pool; 45643494fc30STejun Heo 45653494fc30STejun Heo pr_info("pool %d:", pool->id); 45663494fc30STejun Heo pr_cont_pool_info(pool); 456782607adcSTejun Heo pr_cont(" hung=%us workers=%d", 456882607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 456982607adcSTejun Heo pool->nr_workers); 45703494fc30STejun Heo if (pool->manager) 45713494fc30STejun Heo pr_cont(" manager: %d", 45723494fc30STejun Heo task_pid_nr(pool->manager->task)); 45733494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 45743494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 45753494fc30STejun Heo task_pid_nr(worker->task)); 45763494fc30STejun Heo first = false; 45773494fc30STejun Heo } 45783494fc30STejun Heo pr_cont("\n"); 45793494fc30STejun Heo next_pool: 45803494fc30STejun Heo spin_unlock_irqrestore(&pool->lock, flags); 458162635ea8SSergey Senozhatsky /* 458262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 458362635ea8SSergey Senozhatsky * sysrq-t -> show_workqueue_state(). Avoid triggering 458462635ea8SSergey Senozhatsky * hard lockup. 458562635ea8SSergey Senozhatsky */ 458662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 45873494fc30STejun Heo } 45883494fc30STejun Heo 45893494fc30STejun Heo rcu_read_unlock_sched(); 45903494fc30STejun Heo } 45913494fc30STejun Heo 4592db7bccf4STejun Heo /* 4593db7bccf4STejun Heo * CPU hotplug. 4594db7bccf4STejun Heo * 4595e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4596112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4597706026c2STejun Heo * pool which make migrating pending and scheduled works very 4598e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 459994cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4600e22bee78STejun Heo * blocked draining impractical. 4601e22bee78STejun Heo * 460224647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4603628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4604628c78e7STejun Heo * cpu comes back online. 4605db7bccf4STejun Heo */ 4606db7bccf4STejun Heo 4607e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4608db7bccf4STejun Heo { 46094ce62e9eSTejun Heo struct worker_pool *pool; 4610db7bccf4STejun Heo struct worker *worker; 4611db7bccf4STejun Heo 4612f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 461392f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 461494cf58bbSTejun Heo spin_lock_irq(&pool->lock); 4615e22bee78STejun Heo 4616f2d5a0eeSTejun Heo /* 461792f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 461894cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 461994cf58bbSTejun Heo * except for the ones which are still executing works from 462094cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 462194cf58bbSTejun Heo * this, they may become diasporas. 4622f2d5a0eeSTejun Heo */ 4623da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4624403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4625db7bccf4STejun Heo 462624647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4627f2d5a0eeSTejun Heo 462894cf58bbSTejun Heo spin_unlock_irq(&pool->lock); 462992f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 4630e22bee78STejun Heo 4631e22bee78STejun Heo /* 4632eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 4633eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 4634eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 4635eb283428SLai Jiangshan * from other cpus. 4636628c78e7STejun Heo */ 4637628c78e7STejun Heo schedule(); 4638628c78e7STejun Heo 4639628c78e7STejun Heo /* 4640eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 4641eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 4642eb283428SLai Jiangshan * and keep_working() are always true as long as the 4643eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 4644eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 4645eb283428SLai Jiangshan * are served by workers tied to the pool. 4646e22bee78STejun Heo */ 4647e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 4648eb283428SLai Jiangshan 4649eb283428SLai Jiangshan /* 4650eb283428SLai Jiangshan * With concurrency management just turned off, a busy 4651eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 4652eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 4653eb283428SLai Jiangshan */ 4654eb283428SLai Jiangshan spin_lock_irq(&pool->lock); 4655eb283428SLai Jiangshan wake_up_worker(pool); 4656eb283428SLai Jiangshan spin_unlock_irq(&pool->lock); 4657eb283428SLai Jiangshan } 4658db7bccf4STejun Heo } 4659db7bccf4STejun Heo 4660bd7c089eSTejun Heo /** 4661bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 4662bd7c089eSTejun Heo * @pool: pool of interest 4663bd7c089eSTejun Heo * 4664a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 4665bd7c089eSTejun Heo */ 4666bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 4667bd7c089eSTejun Heo { 4668a9ab775bSTejun Heo struct worker *worker; 4669bd7c089eSTejun Heo 467092f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 4671bd7c089eSTejun Heo 4672bd7c089eSTejun Heo /* 4673a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 4674a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 4675402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 4676a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 4677a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 4678bd7c089eSTejun Heo */ 4679da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4680a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 4681a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 4682a9ab775bSTejun Heo 4683a9ab775bSTejun Heo spin_lock_irq(&pool->lock); 4684f7c17d26SWanpeng Li 46853de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 4686a9ab775bSTejun Heo 4687da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 4688a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 4689a9ab775bSTejun Heo 4690a9ab775bSTejun Heo /* 4691a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 4692a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 4693a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 4694a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 4695a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 4696a9ab775bSTejun Heo * be bound before @pool->lock is released. 4697a9ab775bSTejun Heo */ 4698a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 4699bd7c089eSTejun Heo wake_up_process(worker->task); 4700bd7c089eSTejun Heo 4701bd7c089eSTejun Heo /* 4702a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 4703a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 4704a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 4705a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 4706a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 4707a9ab775bSTejun Heo * concurrency management. Note that when or whether 4708a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 4709a9ab775bSTejun Heo * 4710c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 4711a9ab775bSTejun Heo * tested without holding any lock in 4712a9ab775bSTejun Heo * wq_worker_waking_up(). Without it, NOT_RUNNING test may 4713a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 4714a9ab775bSTejun Heo * management operations. 4715bd7c089eSTejun Heo */ 4716a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 4717a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 4718a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 4719c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 4720bd7c089eSTejun Heo } 4721a9ab775bSTejun Heo 4722a9ab775bSTejun Heo spin_unlock_irq(&pool->lock); 4723bd7c089eSTejun Heo } 4724bd7c089eSTejun Heo 47257dbc725eSTejun Heo /** 47267dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 47277dbc725eSTejun Heo * @pool: unbound pool of interest 47287dbc725eSTejun Heo * @cpu: the CPU which is coming up 47297dbc725eSTejun Heo * 47307dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 47317dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 47327dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 47337dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 47347dbc725eSTejun Heo */ 47357dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 47367dbc725eSTejun Heo { 47377dbc725eSTejun Heo static cpumask_t cpumask; 47387dbc725eSTejun Heo struct worker *worker; 47397dbc725eSTejun Heo 474092f9c5c4SLai Jiangshan lockdep_assert_held(&pool->attach_mutex); 47417dbc725eSTejun Heo 47427dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 47437dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 47447dbc725eSTejun Heo return; 47457dbc725eSTejun Heo 47467dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 47477dbc725eSTejun Heo 47487dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 4749da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4750d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 47517dbc725eSTejun Heo } 47527dbc725eSTejun Heo 47537ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 47541da177e4SLinus Torvalds { 47554ce62e9eSTejun Heo struct worker_pool *pool; 47561da177e4SLinus Torvalds 4757f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 47583ce63377STejun Heo if (pool->nr_workers) 47593ce63377STejun Heo continue; 4760051e1850SLai Jiangshan if (!create_worker(pool)) 47617ee681b2SThomas Gleixner return -ENOMEM; 47623af24433SOleg Nesterov } 47637ee681b2SThomas Gleixner return 0; 47647ee681b2SThomas Gleixner } 47651da177e4SLinus Torvalds 47667ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 47677ee681b2SThomas Gleixner { 47687ee681b2SThomas Gleixner struct worker_pool *pool; 47697ee681b2SThomas Gleixner struct workqueue_struct *wq; 47707ee681b2SThomas Gleixner int pi; 47717ee681b2SThomas Gleixner 477268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 47737dbc725eSTejun Heo 47747dbc725eSTejun Heo for_each_pool(pool, pi) { 477592f9c5c4SLai Jiangshan mutex_lock(&pool->attach_mutex); 477694cf58bbSTejun Heo 4777f05b558dSLai Jiangshan if (pool->cpu == cpu) 477894cf58bbSTejun Heo rebind_workers(pool); 4779f05b558dSLai Jiangshan else if (pool->cpu < 0) 47807dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 478194cf58bbSTejun Heo 478292f9c5c4SLai Jiangshan mutex_unlock(&pool->attach_mutex); 478394cf58bbSTejun Heo } 47847dbc725eSTejun Heo 47854c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 47864c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 47874c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 47884c16bd32STejun Heo 478968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47907ee681b2SThomas Gleixner return 0; 479165758202STejun Heo } 479265758202STejun Heo 47937ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 479465758202STejun Heo { 47954c16bd32STejun Heo struct workqueue_struct *wq; 47968db25e78STejun Heo 47974c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 4798e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 4799e8b3f8dbSLai Jiangshan return -1; 4800e8b3f8dbSLai Jiangshan 4801e8b3f8dbSLai Jiangshan unbind_workers(cpu); 48024c16bd32STejun Heo 48034c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 48044c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 48054c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 48064c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 48074c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 48084c16bd32STejun Heo 48097ee681b2SThomas Gleixner return 0; 481065758202STejun Heo } 481165758202STejun Heo 48122d3854a3SRusty Russell #ifdef CONFIG_SMP 48138ccad40dSRusty Russell 48142d3854a3SRusty Russell struct work_for_cpu { 4815ed48ece2STejun Heo struct work_struct work; 48162d3854a3SRusty Russell long (*fn)(void *); 48172d3854a3SRusty Russell void *arg; 48182d3854a3SRusty Russell long ret; 48192d3854a3SRusty Russell }; 48202d3854a3SRusty Russell 4821ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 48222d3854a3SRusty Russell { 4823ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 4824ed48ece2STejun Heo 48252d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 48262d3854a3SRusty Russell } 48272d3854a3SRusty Russell 48282d3854a3SRusty Russell /** 482922aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 48302d3854a3SRusty Russell * @cpu: the cpu to run on 48312d3854a3SRusty Russell * @fn: the function to run 48322d3854a3SRusty Russell * @arg: the function arg 48332d3854a3SRusty Russell * 483431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 48356b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 4836d185af30SYacine Belkadi * 4837d185af30SYacine Belkadi * Return: The value @fn returns. 48382d3854a3SRusty Russell */ 4839d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 48402d3854a3SRusty Russell { 4841ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 48422d3854a3SRusty Russell 4843ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 4844ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 484512997d1aSBjorn Helgaas flush_work(&wfc.work); 4846440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 48472d3854a3SRusty Russell return wfc.ret; 48482d3854a3SRusty Russell } 48492d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 48500e8d6a93SThomas Gleixner 48510e8d6a93SThomas Gleixner /** 48520e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 48530e8d6a93SThomas Gleixner * @cpu: the cpu to run on 48540e8d6a93SThomas Gleixner * @fn: the function to run 48550e8d6a93SThomas Gleixner * @arg: the function argument 48560e8d6a93SThomas Gleixner * 48570e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 48580e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 48590e8d6a93SThomas Gleixner * 48600e8d6a93SThomas Gleixner * Return: The value @fn returns. 48610e8d6a93SThomas Gleixner */ 48620e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 48630e8d6a93SThomas Gleixner { 48640e8d6a93SThomas Gleixner long ret = -ENODEV; 48650e8d6a93SThomas Gleixner 48660e8d6a93SThomas Gleixner get_online_cpus(); 48670e8d6a93SThomas Gleixner if (cpu_online(cpu)) 48680e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 48690e8d6a93SThomas Gleixner put_online_cpus(); 48700e8d6a93SThomas Gleixner return ret; 48710e8d6a93SThomas Gleixner } 48720e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 48732d3854a3SRusty Russell #endif /* CONFIG_SMP */ 48742d3854a3SRusty Russell 4875a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 4876e7577c50SRusty Russell 4877a0a1a5fdSTejun Heo /** 4878a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 4879a0a1a5fdSTejun Heo * 488058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 4881c5aa87bbSTejun Heo * workqueues will queue new works to their delayed_works list instead of 4882706026c2STejun Heo * pool->worklist. 4883a0a1a5fdSTejun Heo * 4884a0a1a5fdSTejun Heo * CONTEXT: 4885a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4886a0a1a5fdSTejun Heo */ 4887a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 4888a0a1a5fdSTejun Heo { 488924b8a847STejun Heo struct workqueue_struct *wq; 489024b8a847STejun Heo struct pool_workqueue *pwq; 4891a0a1a5fdSTejun Heo 489268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4893a0a1a5fdSTejun Heo 48946183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 4895a0a1a5fdSTejun Heo workqueue_freezing = true; 4896a0a1a5fdSTejun Heo 489724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4898a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4899699ce097STejun Heo for_each_pwq(pwq, wq) 4900699ce097STejun Heo pwq_adjust_max_active(pwq); 4901a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4902a1056305STejun Heo } 49035bcab335STejun Heo 490468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4905a0a1a5fdSTejun Heo } 4906a0a1a5fdSTejun Heo 4907a0a1a5fdSTejun Heo /** 490858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 4909a0a1a5fdSTejun Heo * 4910a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 4911a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 4912a0a1a5fdSTejun Heo * 4913a0a1a5fdSTejun Heo * CONTEXT: 491468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 4915a0a1a5fdSTejun Heo * 4916d185af30SYacine Belkadi * Return: 491758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 491858a69cb4STejun Heo * is complete. 4919a0a1a5fdSTejun Heo */ 4920a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 4921a0a1a5fdSTejun Heo { 4922a0a1a5fdSTejun Heo bool busy = false; 492324b8a847STejun Heo struct workqueue_struct *wq; 492424b8a847STejun Heo struct pool_workqueue *pwq; 4925a0a1a5fdSTejun Heo 492668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4927a0a1a5fdSTejun Heo 49286183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 4929a0a1a5fdSTejun Heo 493024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 493124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 493224b8a847STejun Heo continue; 4933a0a1a5fdSTejun Heo /* 4934a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 4935a0a1a5fdSTejun Heo * to peek without lock. 4936a0a1a5fdSTejun Heo */ 493788109453SLai Jiangshan rcu_read_lock_sched(); 493824b8a847STejun Heo for_each_pwq(pwq, wq) { 49396183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 4940112202d9STejun Heo if (pwq->nr_active) { 4941a0a1a5fdSTejun Heo busy = true; 494288109453SLai Jiangshan rcu_read_unlock_sched(); 4943a0a1a5fdSTejun Heo goto out_unlock; 4944a0a1a5fdSTejun Heo } 4945a0a1a5fdSTejun Heo } 494688109453SLai Jiangshan rcu_read_unlock_sched(); 4947a0a1a5fdSTejun Heo } 4948a0a1a5fdSTejun Heo out_unlock: 494968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4950a0a1a5fdSTejun Heo return busy; 4951a0a1a5fdSTejun Heo } 4952a0a1a5fdSTejun Heo 4953a0a1a5fdSTejun Heo /** 4954a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 4955a0a1a5fdSTejun Heo * 4956a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 4957706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 4958a0a1a5fdSTejun Heo * 4959a0a1a5fdSTejun Heo * CONTEXT: 4960a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 4961a0a1a5fdSTejun Heo */ 4962a0a1a5fdSTejun Heo void thaw_workqueues(void) 4963a0a1a5fdSTejun Heo { 496424b8a847STejun Heo struct workqueue_struct *wq; 496524b8a847STejun Heo struct pool_workqueue *pwq; 4966a0a1a5fdSTejun Heo 496768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4968a0a1a5fdSTejun Heo 4969a0a1a5fdSTejun Heo if (!workqueue_freezing) 4970a0a1a5fdSTejun Heo goto out_unlock; 4971a0a1a5fdSTejun Heo 497274b414eaSLai Jiangshan workqueue_freezing = false; 497324b8a847STejun Heo 497424b8a847STejun Heo /* restore max_active and repopulate worklist */ 497524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 4976a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4977699ce097STejun Heo for_each_pwq(pwq, wq) 4978699ce097STejun Heo pwq_adjust_max_active(pwq); 4979a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 498024b8a847STejun Heo } 498124b8a847STejun Heo 4982a0a1a5fdSTejun Heo out_unlock: 498368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4984a0a1a5fdSTejun Heo } 4985a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 4986a0a1a5fdSTejun Heo 4987042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 4988042f7df1SLai Jiangshan { 4989042f7df1SLai Jiangshan LIST_HEAD(ctxs); 4990042f7df1SLai Jiangshan int ret = 0; 4991042f7df1SLai Jiangshan struct workqueue_struct *wq; 4992042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 4993042f7df1SLai Jiangshan 4994042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 4995042f7df1SLai Jiangshan 4996042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 4997042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 4998042f7df1SLai Jiangshan continue; 4999042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5000042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5001042f7df1SLai Jiangshan continue; 5002042f7df1SLai Jiangshan 5003042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5004042f7df1SLai Jiangshan if (!ctx) { 5005042f7df1SLai Jiangshan ret = -ENOMEM; 5006042f7df1SLai Jiangshan break; 5007042f7df1SLai Jiangshan } 5008042f7df1SLai Jiangshan 5009042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5010042f7df1SLai Jiangshan } 5011042f7df1SLai Jiangshan 5012042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5013042f7df1SLai Jiangshan if (!ret) 5014042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5015042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5016042f7df1SLai Jiangshan } 5017042f7df1SLai Jiangshan 5018042f7df1SLai Jiangshan return ret; 5019042f7df1SLai Jiangshan } 5020042f7df1SLai Jiangshan 5021042f7df1SLai Jiangshan /** 5022042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5023042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5024042f7df1SLai Jiangshan * 5025042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5026042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5027042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5028042f7df1SLai Jiangshan * 5029042f7df1SLai Jiangshan * Retun: 0 - Success 5030042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5031042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5032042f7df1SLai Jiangshan */ 5033042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5034042f7df1SLai Jiangshan { 5035042f7df1SLai Jiangshan int ret = -EINVAL; 5036042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5037042f7df1SLai Jiangshan 5038042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) 5039042f7df1SLai Jiangshan return -ENOMEM; 5040042f7df1SLai Jiangshan 5041c98a9805STal Shorer /* 5042c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5043c98a9805STal Shorer * If the user wishes to include them, we allow that. 5044c98a9805STal Shorer */ 5045042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5046042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5047a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5048042f7df1SLai Jiangshan 5049042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5050042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5051042f7df1SLai Jiangshan 5052042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5053042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5054042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5055042f7df1SLai Jiangshan 5056042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5057042f7df1SLai Jiangshan if (ret < 0) 5058042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5059042f7df1SLai Jiangshan 5060a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5061042f7df1SLai Jiangshan } 5062042f7df1SLai Jiangshan 5063042f7df1SLai Jiangshan free_cpumask_var(saved_cpumask); 5064042f7df1SLai Jiangshan return ret; 5065042f7df1SLai Jiangshan } 5066042f7df1SLai Jiangshan 50676ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 50686ba94429SFrederic Weisbecker /* 50696ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 50706ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 50716ba94429SFrederic Weisbecker * following attributes. 50726ba94429SFrederic Weisbecker * 50736ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 50746ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 50756ba94429SFrederic Weisbecker * 50766ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 50776ba94429SFrederic Weisbecker * 50789a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 50796ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 50806ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 50819a19b463SWang Long * numa RW bool : whether enable NUMA affinity 50826ba94429SFrederic Weisbecker */ 50836ba94429SFrederic Weisbecker struct wq_device { 50846ba94429SFrederic Weisbecker struct workqueue_struct *wq; 50856ba94429SFrederic Weisbecker struct device dev; 50866ba94429SFrederic Weisbecker }; 50876ba94429SFrederic Weisbecker 50886ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 50896ba94429SFrederic Weisbecker { 50906ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 50916ba94429SFrederic Weisbecker 50926ba94429SFrederic Weisbecker return wq_dev->wq; 50936ba94429SFrederic Weisbecker } 50946ba94429SFrederic Weisbecker 50956ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 50966ba94429SFrederic Weisbecker char *buf) 50976ba94429SFrederic Weisbecker { 50986ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 50996ba94429SFrederic Weisbecker 51006ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 51016ba94429SFrederic Weisbecker } 51026ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 51036ba94429SFrederic Weisbecker 51046ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 51056ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51066ba94429SFrederic Weisbecker { 51076ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51086ba94429SFrederic Weisbecker 51096ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 51106ba94429SFrederic Weisbecker } 51116ba94429SFrederic Weisbecker 51126ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 51136ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 51146ba94429SFrederic Weisbecker size_t count) 51156ba94429SFrederic Weisbecker { 51166ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51176ba94429SFrederic Weisbecker int val; 51186ba94429SFrederic Weisbecker 51196ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 51206ba94429SFrederic Weisbecker return -EINVAL; 51216ba94429SFrederic Weisbecker 51226ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 51236ba94429SFrederic Weisbecker return count; 51246ba94429SFrederic Weisbecker } 51256ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 51266ba94429SFrederic Weisbecker 51276ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 51286ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 51296ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 51306ba94429SFrederic Weisbecker NULL, 51316ba94429SFrederic Weisbecker }; 51326ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 51336ba94429SFrederic Weisbecker 51346ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 51356ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 51366ba94429SFrederic Weisbecker { 51376ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51386ba94429SFrederic Weisbecker const char *delim = ""; 51396ba94429SFrederic Weisbecker int node, written = 0; 51406ba94429SFrederic Weisbecker 51416ba94429SFrederic Weisbecker rcu_read_lock_sched(); 51426ba94429SFrederic Weisbecker for_each_node(node) { 51436ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 51446ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 51456ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 51466ba94429SFrederic Weisbecker delim = " "; 51476ba94429SFrederic Weisbecker } 51486ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 51496ba94429SFrederic Weisbecker rcu_read_unlock_sched(); 51506ba94429SFrederic Weisbecker 51516ba94429SFrederic Weisbecker return written; 51526ba94429SFrederic Weisbecker } 51536ba94429SFrederic Weisbecker 51546ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 51556ba94429SFrederic Weisbecker char *buf) 51566ba94429SFrederic Weisbecker { 51576ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51586ba94429SFrederic Weisbecker int written; 51596ba94429SFrederic Weisbecker 51606ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 51616ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 51626ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 51636ba94429SFrederic Weisbecker 51646ba94429SFrederic Weisbecker return written; 51656ba94429SFrederic Weisbecker } 51666ba94429SFrederic Weisbecker 51676ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 51686ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 51696ba94429SFrederic Weisbecker { 51706ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 51716ba94429SFrederic Weisbecker 5172899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5173899a94feSLai Jiangshan 51746ba94429SFrederic Weisbecker attrs = alloc_workqueue_attrs(GFP_KERNEL); 51756ba94429SFrederic Weisbecker if (!attrs) 51766ba94429SFrederic Weisbecker return NULL; 51776ba94429SFrederic Weisbecker 51786ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 51796ba94429SFrederic Weisbecker return attrs; 51806ba94429SFrederic Weisbecker } 51816ba94429SFrederic Weisbecker 51826ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 51836ba94429SFrederic Weisbecker const char *buf, size_t count) 51846ba94429SFrederic Weisbecker { 51856ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 51866ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5187d4d3e257SLai Jiangshan int ret = -ENOMEM; 5188d4d3e257SLai Jiangshan 5189d4d3e257SLai Jiangshan apply_wqattrs_lock(); 51906ba94429SFrederic Weisbecker 51916ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 51926ba94429SFrederic Weisbecker if (!attrs) 5193d4d3e257SLai Jiangshan goto out_unlock; 51946ba94429SFrederic Weisbecker 51956ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 51966ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5197d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 51986ba94429SFrederic Weisbecker else 51996ba94429SFrederic Weisbecker ret = -EINVAL; 52006ba94429SFrederic Weisbecker 5201d4d3e257SLai Jiangshan out_unlock: 5202d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52036ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52046ba94429SFrederic Weisbecker return ret ?: count; 52056ba94429SFrederic Weisbecker } 52066ba94429SFrederic Weisbecker 52076ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 52086ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 52096ba94429SFrederic Weisbecker { 52106ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52116ba94429SFrederic Weisbecker int written; 52126ba94429SFrederic Weisbecker 52136ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 52146ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 52156ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 52166ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 52176ba94429SFrederic Weisbecker return written; 52186ba94429SFrederic Weisbecker } 52196ba94429SFrederic Weisbecker 52206ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 52216ba94429SFrederic Weisbecker struct device_attribute *attr, 52226ba94429SFrederic Weisbecker const char *buf, size_t count) 52236ba94429SFrederic Weisbecker { 52246ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52256ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5226d4d3e257SLai Jiangshan int ret = -ENOMEM; 5227d4d3e257SLai Jiangshan 5228d4d3e257SLai Jiangshan apply_wqattrs_lock(); 52296ba94429SFrederic Weisbecker 52306ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 52316ba94429SFrederic Weisbecker if (!attrs) 5232d4d3e257SLai Jiangshan goto out_unlock; 52336ba94429SFrederic Weisbecker 52346ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 52356ba94429SFrederic Weisbecker if (!ret) 5236d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 52376ba94429SFrederic Weisbecker 5238d4d3e257SLai Jiangshan out_unlock: 5239d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52406ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52416ba94429SFrederic Weisbecker return ret ?: count; 52426ba94429SFrederic Weisbecker } 52436ba94429SFrederic Weisbecker 52446ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 52456ba94429SFrederic Weisbecker char *buf) 52466ba94429SFrederic Weisbecker { 52476ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52486ba94429SFrederic Weisbecker int written; 52496ba94429SFrederic Weisbecker 52506ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 52516ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 52526ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 52536ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 52546ba94429SFrederic Weisbecker 52556ba94429SFrederic Weisbecker return written; 52566ba94429SFrederic Weisbecker } 52576ba94429SFrederic Weisbecker 52586ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 52596ba94429SFrederic Weisbecker const char *buf, size_t count) 52606ba94429SFrederic Weisbecker { 52616ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 52626ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5263d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5264d4d3e257SLai Jiangshan 5265d4d3e257SLai Jiangshan apply_wqattrs_lock(); 52666ba94429SFrederic Weisbecker 52676ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 52686ba94429SFrederic Weisbecker if (!attrs) 5269d4d3e257SLai Jiangshan goto out_unlock; 52706ba94429SFrederic Weisbecker 52716ba94429SFrederic Weisbecker ret = -EINVAL; 52726ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 52736ba94429SFrederic Weisbecker attrs->no_numa = !v; 5274d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 52756ba94429SFrederic Weisbecker } 52766ba94429SFrederic Weisbecker 5277d4d3e257SLai Jiangshan out_unlock: 5278d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 52796ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 52806ba94429SFrederic Weisbecker return ret ?: count; 52816ba94429SFrederic Weisbecker } 52826ba94429SFrederic Weisbecker 52836ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 52846ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 52856ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 52866ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 52876ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 52886ba94429SFrederic Weisbecker __ATTR_NULL, 52896ba94429SFrederic Weisbecker }; 52906ba94429SFrederic Weisbecker 52916ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 52926ba94429SFrederic Weisbecker .name = "workqueue", 52936ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 52946ba94429SFrederic Weisbecker }; 52956ba94429SFrederic Weisbecker 5296b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5297b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5298b05a7928SFrederic Weisbecker { 5299b05a7928SFrederic Weisbecker int written; 5300b05a7928SFrederic Weisbecker 5301042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5302b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5303b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5304042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5305b05a7928SFrederic Weisbecker 5306b05a7928SFrederic Weisbecker return written; 5307b05a7928SFrederic Weisbecker } 5308b05a7928SFrederic Weisbecker 5309042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5310042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5311042f7df1SLai Jiangshan { 5312042f7df1SLai Jiangshan cpumask_var_t cpumask; 5313042f7df1SLai Jiangshan int ret; 5314042f7df1SLai Jiangshan 5315042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5316042f7df1SLai Jiangshan return -ENOMEM; 5317042f7df1SLai Jiangshan 5318042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5319042f7df1SLai Jiangshan if (!ret) 5320042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5321042f7df1SLai Jiangshan 5322042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5323042f7df1SLai Jiangshan return ret ? ret : count; 5324042f7df1SLai Jiangshan } 5325042f7df1SLai Jiangshan 5326b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5327042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5328042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5329b05a7928SFrederic Weisbecker 53306ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 53316ba94429SFrederic Weisbecker { 5332b05a7928SFrederic Weisbecker int err; 5333b05a7928SFrederic Weisbecker 5334b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5335b05a7928SFrederic Weisbecker if (err) 5336b05a7928SFrederic Weisbecker return err; 5337b05a7928SFrederic Weisbecker 5338b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 53396ba94429SFrederic Weisbecker } 53406ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 53416ba94429SFrederic Weisbecker 53426ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 53436ba94429SFrederic Weisbecker { 53446ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 53456ba94429SFrederic Weisbecker 53466ba94429SFrederic Weisbecker kfree(wq_dev); 53476ba94429SFrederic Weisbecker } 53486ba94429SFrederic Weisbecker 53496ba94429SFrederic Weisbecker /** 53506ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 53516ba94429SFrederic Weisbecker * @wq: the workqueue to register 53526ba94429SFrederic Weisbecker * 53536ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 53546ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 53556ba94429SFrederic Weisbecker * which is the preferred method. 53566ba94429SFrederic Weisbecker * 53576ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 53586ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 53596ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 53606ba94429SFrederic Weisbecker * attributes. 53616ba94429SFrederic Weisbecker * 53626ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 53636ba94429SFrederic Weisbecker */ 53646ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 53656ba94429SFrederic Weisbecker { 53666ba94429SFrederic Weisbecker struct wq_device *wq_dev; 53676ba94429SFrederic Weisbecker int ret; 53686ba94429SFrederic Weisbecker 53696ba94429SFrederic Weisbecker /* 5370402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 53716ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 53726ba94429SFrederic Weisbecker * workqueues. 53736ba94429SFrederic Weisbecker */ 53740a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 53756ba94429SFrederic Weisbecker return -EINVAL; 53766ba94429SFrederic Weisbecker 53776ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 53786ba94429SFrederic Weisbecker if (!wq_dev) 53796ba94429SFrederic Weisbecker return -ENOMEM; 53806ba94429SFrederic Weisbecker 53816ba94429SFrederic Weisbecker wq_dev->wq = wq; 53826ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 53836ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 538423217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 53856ba94429SFrederic Weisbecker 53866ba94429SFrederic Weisbecker /* 53876ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 53886ba94429SFrederic Weisbecker * everything is ready. 53896ba94429SFrederic Weisbecker */ 53906ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 53916ba94429SFrederic Weisbecker 53926ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 53936ba94429SFrederic Weisbecker if (ret) { 53946ba94429SFrederic Weisbecker kfree(wq_dev); 53956ba94429SFrederic Weisbecker wq->wq_dev = NULL; 53966ba94429SFrederic Weisbecker return ret; 53976ba94429SFrederic Weisbecker } 53986ba94429SFrederic Weisbecker 53996ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 54006ba94429SFrederic Weisbecker struct device_attribute *attr; 54016ba94429SFrederic Weisbecker 54026ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 54036ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 54046ba94429SFrederic Weisbecker if (ret) { 54056ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 54066ba94429SFrederic Weisbecker wq->wq_dev = NULL; 54076ba94429SFrederic Weisbecker return ret; 54086ba94429SFrederic Weisbecker } 54096ba94429SFrederic Weisbecker } 54106ba94429SFrederic Weisbecker } 54116ba94429SFrederic Weisbecker 54126ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 54136ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 54146ba94429SFrederic Weisbecker return 0; 54156ba94429SFrederic Weisbecker } 54166ba94429SFrederic Weisbecker 54176ba94429SFrederic Weisbecker /** 54186ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 54196ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 54206ba94429SFrederic Weisbecker * 54216ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 54226ba94429SFrederic Weisbecker */ 54236ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 54246ba94429SFrederic Weisbecker { 54256ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 54266ba94429SFrederic Weisbecker 54276ba94429SFrederic Weisbecker if (!wq->wq_dev) 54286ba94429SFrederic Weisbecker return; 54296ba94429SFrederic Weisbecker 54306ba94429SFrederic Weisbecker wq->wq_dev = NULL; 54316ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 54326ba94429SFrederic Weisbecker } 54336ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 54346ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 54356ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 54366ba94429SFrederic Weisbecker 543782607adcSTejun Heo /* 543882607adcSTejun Heo * Workqueue watchdog. 543982607adcSTejun Heo * 544082607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 544182607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 544282607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 544382607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 544482607adcSTejun Heo * largely opaque. 544582607adcSTejun Heo * 544682607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 544782607adcSTejun Heo * state if some pools failed to make forward progress for a while where 544882607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 544982607adcSTejun Heo * 545082607adcSTejun Heo * This mechanism is controlled through the kernel parameter 545182607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 545282607adcSTejun Heo * corresponding sysfs parameter file. 545382607adcSTejun Heo */ 545482607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 545582607adcSTejun Heo 545682607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 54575cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 545882607adcSTejun Heo 545982607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 546082607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 546182607adcSTejun Heo 546282607adcSTejun Heo static void wq_watchdog_reset_touched(void) 546382607adcSTejun Heo { 546482607adcSTejun Heo int cpu; 546582607adcSTejun Heo 546682607adcSTejun Heo wq_watchdog_touched = jiffies; 546782607adcSTejun Heo for_each_possible_cpu(cpu) 546882607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 546982607adcSTejun Heo } 547082607adcSTejun Heo 54715cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 547282607adcSTejun Heo { 547382607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 547482607adcSTejun Heo bool lockup_detected = false; 547582607adcSTejun Heo struct worker_pool *pool; 547682607adcSTejun Heo int pi; 547782607adcSTejun Heo 547882607adcSTejun Heo if (!thresh) 547982607adcSTejun Heo return; 548082607adcSTejun Heo 548182607adcSTejun Heo rcu_read_lock(); 548282607adcSTejun Heo 548382607adcSTejun Heo for_each_pool(pool, pi) { 548482607adcSTejun Heo unsigned long pool_ts, touched, ts; 548582607adcSTejun Heo 548682607adcSTejun Heo if (list_empty(&pool->worklist)) 548782607adcSTejun Heo continue; 548882607adcSTejun Heo 548982607adcSTejun Heo /* get the latest of pool and touched timestamps */ 549082607adcSTejun Heo pool_ts = READ_ONCE(pool->watchdog_ts); 549182607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 549282607adcSTejun Heo 549382607adcSTejun Heo if (time_after(pool_ts, touched)) 549482607adcSTejun Heo ts = pool_ts; 549582607adcSTejun Heo else 549682607adcSTejun Heo ts = touched; 549782607adcSTejun Heo 549882607adcSTejun Heo if (pool->cpu >= 0) { 549982607adcSTejun Heo unsigned long cpu_touched = 550082607adcSTejun Heo READ_ONCE(per_cpu(wq_watchdog_touched_cpu, 550182607adcSTejun Heo pool->cpu)); 550282607adcSTejun Heo if (time_after(cpu_touched, ts)) 550382607adcSTejun Heo ts = cpu_touched; 550482607adcSTejun Heo } 550582607adcSTejun Heo 550682607adcSTejun Heo /* did we stall? */ 550782607adcSTejun Heo if (time_after(jiffies, ts + thresh)) { 550882607adcSTejun Heo lockup_detected = true; 550982607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 551082607adcSTejun Heo pr_cont_pool_info(pool); 551182607adcSTejun Heo pr_cont(" stuck for %us!\n", 551282607adcSTejun Heo jiffies_to_msecs(jiffies - pool_ts) / 1000); 551382607adcSTejun Heo } 551482607adcSTejun Heo } 551582607adcSTejun Heo 551682607adcSTejun Heo rcu_read_unlock(); 551782607adcSTejun Heo 551882607adcSTejun Heo if (lockup_detected) 551982607adcSTejun Heo show_workqueue_state(); 552082607adcSTejun Heo 552182607adcSTejun Heo wq_watchdog_reset_touched(); 552282607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 552382607adcSTejun Heo } 552482607adcSTejun Heo 552582607adcSTejun Heo void wq_watchdog_touch(int cpu) 552682607adcSTejun Heo { 552782607adcSTejun Heo if (cpu >= 0) 552882607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 552982607adcSTejun Heo else 553082607adcSTejun Heo wq_watchdog_touched = jiffies; 553182607adcSTejun Heo } 553282607adcSTejun Heo 553382607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 553482607adcSTejun Heo { 553582607adcSTejun Heo wq_watchdog_thresh = 0; 553682607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 553782607adcSTejun Heo 553882607adcSTejun Heo if (thresh) { 553982607adcSTejun Heo wq_watchdog_thresh = thresh; 554082607adcSTejun Heo wq_watchdog_reset_touched(); 554182607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 554282607adcSTejun Heo } 554382607adcSTejun Heo } 554482607adcSTejun Heo 554582607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 554682607adcSTejun Heo const struct kernel_param *kp) 554782607adcSTejun Heo { 554882607adcSTejun Heo unsigned long thresh; 554982607adcSTejun Heo int ret; 555082607adcSTejun Heo 555182607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 555282607adcSTejun Heo if (ret) 555382607adcSTejun Heo return ret; 555482607adcSTejun Heo 555582607adcSTejun Heo if (system_wq) 555682607adcSTejun Heo wq_watchdog_set_thresh(thresh); 555782607adcSTejun Heo else 555882607adcSTejun Heo wq_watchdog_thresh = thresh; 555982607adcSTejun Heo 556082607adcSTejun Heo return 0; 556182607adcSTejun Heo } 556282607adcSTejun Heo 556382607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 556482607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 556582607adcSTejun Heo .get = param_get_ulong, 556682607adcSTejun Heo }; 556782607adcSTejun Heo 556882607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 556982607adcSTejun Heo 0644); 557082607adcSTejun Heo 557182607adcSTejun Heo static void wq_watchdog_init(void) 557282607adcSTejun Heo { 55735cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 557482607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 557582607adcSTejun Heo } 557682607adcSTejun Heo 557782607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 557882607adcSTejun Heo 557982607adcSTejun Heo static inline void wq_watchdog_init(void) { } 558082607adcSTejun Heo 558182607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 558282607adcSTejun Heo 5583bce90380STejun Heo static void __init wq_numa_init(void) 5584bce90380STejun Heo { 5585bce90380STejun Heo cpumask_var_t *tbl; 5586bce90380STejun Heo int node, cpu; 5587bce90380STejun Heo 5588bce90380STejun Heo if (num_possible_nodes() <= 1) 5589bce90380STejun Heo return; 5590bce90380STejun Heo 5591d55262c4STejun Heo if (wq_disable_numa) { 5592d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5593d55262c4STejun Heo return; 5594d55262c4STejun Heo } 5595d55262c4STejun Heo 55964c16bd32STejun Heo wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL); 55974c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 55984c16bd32STejun Heo 5599bce90380STejun Heo /* 5600bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5601bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5602bce90380STejun Heo * fully initialized by now. 5603bce90380STejun Heo */ 5604ddcb57e2SLai Jiangshan tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL); 5605bce90380STejun Heo BUG_ON(!tbl); 5606bce90380STejun Heo 5607bce90380STejun Heo for_each_node(node) 56085a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 56091be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5610bce90380STejun Heo 5611bce90380STejun Heo for_each_possible_cpu(cpu) { 5612bce90380STejun Heo node = cpu_to_node(cpu); 5613bce90380STejun Heo if (WARN_ON(node == NUMA_NO_NODE)) { 5614bce90380STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5615bce90380STejun Heo /* happens iff arch is bonkers, let's just proceed */ 5616bce90380STejun Heo return; 5617bce90380STejun Heo } 5618bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 5619bce90380STejun Heo } 5620bce90380STejun Heo 5621bce90380STejun Heo wq_numa_possible_cpumask = tbl; 5622bce90380STejun Heo wq_numa_enabled = true; 5623bce90380STejun Heo } 5624bce90380STejun Heo 56253347fa09STejun Heo /** 56263347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 56273347fa09STejun Heo * 56283347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 56293347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 56303347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 56313347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 56323347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 56333347fa09STejun Heo * created and scheduled right before early initcalls. 56343347fa09STejun Heo */ 56353347fa09STejun Heo int __init workqueue_init_early(void) 56361da177e4SLinus Torvalds { 56377a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 56387a4e344cSTejun Heo int i, cpu; 5639c34056a3STejun Heo 5640e904e6c2STejun Heo WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 5641e904e6c2STejun Heo 5642b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 5643c98a9805STal Shorer cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_FLAG_DOMAIN)); 5644b05a7928SFrederic Weisbecker 5645e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 5646e904e6c2STejun Heo 5647706026c2STejun Heo /* initialize CPU pools */ 564829c91e99STejun Heo for_each_possible_cpu(cpu) { 56494ce62e9eSTejun Heo struct worker_pool *pool; 56508b03ae3cSTejun Heo 56517a4e344cSTejun Heo i = 0; 5652f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 56537a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 5654ec22ca5eSTejun Heo pool->cpu = cpu; 56557a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 56567a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 5657f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 56587a4e344cSTejun Heo 56599daf9e67STejun Heo /* alloc pool ID */ 566068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 56619daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 566268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 56634ce62e9eSTejun Heo } 56648b03ae3cSTejun Heo } 56658b03ae3cSTejun Heo 56668a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 566729c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 566829c91e99STejun Heo struct workqueue_attrs *attrs; 566929c91e99STejun Heo 567029c91e99STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 567129c91e99STejun Heo attrs->nice = std_nice[i]; 567229c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 56738a2b7538STejun Heo 56748a2b7538STejun Heo /* 56758a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 56768a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 56778a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 56788a2b7538STejun Heo */ 56798a2b7538STejun Heo BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL))); 56808a2b7538STejun Heo attrs->nice = std_nice[i]; 56818a2b7538STejun Heo attrs->no_numa = true; 56828a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 568329c91e99STejun Heo } 568429c91e99STejun Heo 5685d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 56861aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 5687d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 5688f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 5689f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 569024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 569124d51addSTejun Heo WQ_FREEZABLE, 0); 56920668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 56930668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 56940668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 56950668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 56960668106cSViresh Kumar 0); 56971aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 56980668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 56990668106cSViresh Kumar !system_power_efficient_wq || 57000668106cSViresh Kumar !system_freezable_power_efficient_wq); 570182607adcSTejun Heo 57023347fa09STejun Heo return 0; 57033347fa09STejun Heo } 57043347fa09STejun Heo 57053347fa09STejun Heo /** 57063347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 57073347fa09STejun Heo * 57083347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 57093347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 57103347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 57113347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 57123347fa09STejun Heo * with the initial workers and enable future kworker creations. 57133347fa09STejun Heo */ 57143347fa09STejun Heo int __init workqueue_init(void) 57153347fa09STejun Heo { 57162186d9f9STejun Heo struct workqueue_struct *wq; 57173347fa09STejun Heo struct worker_pool *pool; 57183347fa09STejun Heo int cpu, bkt; 57193347fa09STejun Heo 57202186d9f9STejun Heo /* 57212186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 57222186d9f9STejun Heo * CPU to node mapping may not be available that early on some 57232186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 57242186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 57252186d9f9STejun Heo * affinity, fix them up. 572640c17f75STejun Heo * 572740c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 57282186d9f9STejun Heo */ 57292186d9f9STejun Heo wq_numa_init(); 57302186d9f9STejun Heo 57312186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 57322186d9f9STejun Heo 57332186d9f9STejun Heo for_each_possible_cpu(cpu) { 57342186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 57352186d9f9STejun Heo pool->node = cpu_to_node(cpu); 57362186d9f9STejun Heo } 57372186d9f9STejun Heo } 57382186d9f9STejun Heo 573940c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 57402186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 574140c17f75STejun Heo WARN(init_rescuer(wq), 574240c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 574340c17f75STejun Heo wq->name); 574440c17f75STejun Heo } 57452186d9f9STejun Heo 57462186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 57472186d9f9STejun Heo 57483347fa09STejun Heo /* create the initial workers */ 57493347fa09STejun Heo for_each_online_cpu(cpu) { 57503347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 57513347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 57523347fa09STejun Heo BUG_ON(!create_worker(pool)); 57533347fa09STejun Heo } 57543347fa09STejun Heo } 57553347fa09STejun Heo 57563347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 57573347fa09STejun Heo BUG_ON(!create_worker(pool)); 57583347fa09STejun Heo 57593347fa09STejun Heo wq_online = true; 576082607adcSTejun Heo wq_watchdog_init(); 576182607adcSTejun Heo 57626ee0578bSSuresh Siddha return 0; 57631da177e4SLinus Torvalds } 5764