1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 3c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 41da177e4SLinus Torvalds * 5c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 81da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 9e1f8e874SFrancois Cami * Andrew Morton 101da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 111da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1289ada679SChristoph Lameter * 13cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 14c54fce6eSTejun Heo * 15c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 16c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 17c54fce6eSTejun Heo * 18c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 19c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 20b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 21b11895c4SLibin * normal work items and the other for high priority ones) and some extra 22b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 23b11895c4SLibin * number of these backing pools is dynamic. 24c54fce6eSTejun Heo * 259a261491SBenjamin Peterson * Please read Documentation/core-api/workqueue.rst for details. 261da177e4SLinus Torvalds */ 271da177e4SLinus Torvalds 289984de1aSPaul Gortmaker #include <linux/export.h> 291da177e4SLinus Torvalds #include <linux/kernel.h> 301da177e4SLinus Torvalds #include <linux/sched.h> 311da177e4SLinus Torvalds #include <linux/init.h> 321da177e4SLinus Torvalds #include <linux/signal.h> 331da177e4SLinus Torvalds #include <linux/completion.h> 341da177e4SLinus Torvalds #include <linux/workqueue.h> 351da177e4SLinus Torvalds #include <linux/slab.h> 361da177e4SLinus Torvalds #include <linux/cpu.h> 371da177e4SLinus Torvalds #include <linux/notifier.h> 381da177e4SLinus Torvalds #include <linux/kthread.h> 391fa44ecaSJames Bottomley #include <linux/hardirq.h> 4046934023SChristoph Lameter #include <linux/mempolicy.h> 41341a5958SRafael J. Wysocki #include <linux/freezer.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51c98a9805STal Shorer #include <linux/sched/isolation.h> 52cd2440d6SPetr Mladek #include <linux/sched/debug.h> 5362635ea8SSergey Senozhatsky #include <linux/nmi.h> 54940d71c6SSergey Senozhatsky #include <linux/kvm_para.h> 55e22bee78STejun Heo 56ea138446STejun Heo #include "workqueue_internal.h" 571da177e4SLinus Torvalds 58c8e55f36STejun Heo enum { 59bc2ae0f5STejun Heo /* 6024647570STejun Heo * worker_pool flags 61bc2ae0f5STejun Heo * 6224647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 63bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 64bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 65bc2ae0f5STejun Heo * is in effect. 66bc2ae0f5STejun Heo * 67bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 68bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6924647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 70bc2ae0f5STejun Heo * 71bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 721258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 734736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 74bc2ae0f5STejun Heo */ 75692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7624647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 77db7bccf4STejun Heo 78c8e55f36STejun Heo /* worker flags */ 79c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 80c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 81e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 82fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 83f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 84a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 85e22bee78STejun Heo 86a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 87a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 88db7bccf4STejun Heo 89e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 904ce62e9eSTejun Heo 9129c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 92c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 93db7bccf4STejun Heo 94e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 95e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 96e22bee78STejun Heo 973233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 983233cdbdSTejun Heo /* call for help after 10ms 993233cdbdSTejun Heo (min two ticks) */ 100e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 101e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1021da177e4SLinus Torvalds 1031da177e4SLinus Torvalds /* 104e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1058698a745SDongsheng Yang * all cpus. Give MIN_NICE. 106e22bee78STejun Heo */ 1078698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1088698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 109ecf6881fSTejun Heo 110ecf6881fSTejun Heo WQ_NAME_LEN = 24, 111c8e55f36STejun Heo }; 112c8e55f36STejun Heo 1131da177e4SLinus Torvalds /* 1144690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1154690c4abSTejun Heo * 116e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 117e41e704bSTejun Heo * everyone else. 1184690c4abSTejun Heo * 119e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 120e22bee78STejun Heo * only be modified and accessed from the local cpu. 121e22bee78STejun Heo * 122d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1234690c4abSTejun Heo * 124d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 125d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 126d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 127d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 128e22bee78STejun Heo * 1291258fae7STejun Heo * A: wq_pool_attach_mutex protected. 130822d8405STejun Heo * 13168e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13276af4d93STejun Heo * 13324acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1345bcab335STejun Heo * 1355b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1365b95e1afSLai Jiangshan * 1375b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 13824acfb71SThomas Gleixner * RCU for reads. 1395b95e1afSLai Jiangshan * 1403c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1413c25a55dSLai Jiangshan * 14224acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1432e109a28STejun Heo * 1442e109a28STejun Heo * MD: wq_mayday_lock protected. 145cd2440d6SPetr Mladek * 146cd2440d6SPetr Mladek * WD: Used internally by the watchdog. 1474690c4abSTejun Heo */ 1484690c4abSTejun Heo 1492eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 150c34056a3STejun Heo 151bd7bdd43STejun Heo struct worker_pool { 152a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 153d84ff051STejun Heo int cpu; /* I: the associated cpu */ 154f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1559daf9e67STejun Heo int id; /* I: pool ID */ 15611ebea50STejun Heo unsigned int flags; /* X: flags */ 157bd7bdd43STejun Heo 15882607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 159cd2440d6SPetr Mladek bool cpu_stall; /* WD: stalled cpu bound pool */ 16082607adcSTejun Heo 161bc35f7efSLai Jiangshan /* 162bc35f7efSLai Jiangshan * The counter is incremented in a process context on the associated CPU 163bc35f7efSLai Jiangshan * w/ preemption disabled, and decremented or reset in the same context 164bc35f7efSLai Jiangshan * but w/ pool->lock held. The readers grab pool->lock and are 165bc35f7efSLai Jiangshan * guaranteed to see if the counter reached zero. 166bc35f7efSLai Jiangshan */ 167bc35f7efSLai Jiangshan int nr_running; 16884f91c62SLai Jiangshan 169bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 170ea1abd61SLai Jiangshan 1715826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1725826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 173bd7bdd43STejun Heo 1742c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 175bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 1763f959aa3SValentin Schneider struct work_struct idle_cull_work; /* L: worker idle cleanup */ 1773f959aa3SValentin Schneider 178bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 179bd7bdd43STejun Heo 180c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 181c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 182c9e7cf27STejun Heo /* L: hash of busy workers */ 183c9e7cf27STejun Heo 1842607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 18592f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 186e02b9312SValentin Schneider struct list_head dying_workers; /* A: workers about to die */ 18760f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 188e19e397aSTejun Heo 1897cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 190e19e397aSTejun Heo 1917a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 19268e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 19368e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1947a4e344cSTejun Heo 195e19e397aSTejun Heo /* 19624acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 19729c91e99STejun Heo * from get_work_pool(). 19829c91e99STejun Heo */ 19929c91e99STejun Heo struct rcu_head rcu; 20084f91c62SLai Jiangshan }; 2018b03ae3cSTejun Heo 2028b03ae3cSTejun Heo /* 203725e8ec5STejun Heo * Per-pool_workqueue statistics. These can be monitored using 204725e8ec5STejun Heo * tools/workqueue/wq_monitor.py. 205725e8ec5STejun Heo */ 206725e8ec5STejun Heo enum pool_workqueue_stats { 207725e8ec5STejun Heo PWQ_STAT_STARTED, /* work items started execution */ 208725e8ec5STejun Heo PWQ_STAT_COMPLETED, /* work items completed execution */ 209725e8ec5STejun Heo PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 210725e8ec5STejun Heo PWQ_STAT_MAYDAY, /* maydays to rescuer */ 211725e8ec5STejun Heo PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 212725e8ec5STejun Heo 213725e8ec5STejun Heo PWQ_NR_STATS, 214725e8ec5STejun Heo }; 215725e8ec5STejun Heo 216725e8ec5STejun Heo /* 217112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 218112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 219112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 220112202d9STejun Heo * number of flag bits. 2211da177e4SLinus Torvalds */ 222112202d9STejun Heo struct pool_workqueue { 223bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2244690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 22573f53c4aSTejun Heo int work_color; /* L: current color */ 22673f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2278864b4e5STejun Heo int refcnt; /* L: reference count */ 22873f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 22973f53c4aSTejun Heo /* L: nr of in_flight works */ 230018f3a13SLai Jiangshan 231018f3a13SLai Jiangshan /* 232018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 233018f3a13SLai Jiangshan * 234018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 235018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 236018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 237018f3a13SLai Jiangshan * 238018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 239018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 240018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 241018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 242018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 243018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 244018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 245018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 246018f3a13SLai Jiangshan */ 2471e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 248a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 249f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2503c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2512e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2528864b4e5STejun Heo 253725e8ec5STejun Heo u64 stats[PWQ_NR_STATS]; 254725e8ec5STejun Heo 2558864b4e5STejun Heo /* 2568864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2578864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 25824acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 259b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2608864b4e5STejun Heo */ 2618864b4e5STejun Heo struct work_struct unbound_release_work; 2628864b4e5STejun Heo struct rcu_head rcu; 263e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2641da177e4SLinus Torvalds 2651da177e4SLinus Torvalds /* 26673f53c4aSTejun Heo * Structure used to wait for workqueue flush. 26773f53c4aSTejun Heo */ 26873f53c4aSTejun Heo struct wq_flusher { 2693c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2703c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 27173f53c4aSTejun Heo struct completion done; /* flush completion */ 27273f53c4aSTejun Heo }; 2731da177e4SLinus Torvalds 274226223abSTejun Heo struct wq_device; 275226223abSTejun Heo 27673f53c4aSTejun Heo /* 277c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 278c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2791da177e4SLinus Torvalds */ 2801da177e4SLinus Torvalds struct workqueue_struct { 2813c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 282e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 28373f53c4aSTejun Heo 2843c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2853c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2863c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 287112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2883c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2893c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2903c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 29173f53c4aSTejun Heo 2922e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 29330ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 294e22bee78STejun Heo 29587fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 296a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 297226223abSTejun Heo 2985b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2995b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 3006029a918STejun Heo 301226223abSTejun Heo #ifdef CONFIG_SYSFS 302226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 303226223abSTejun Heo #endif 3044e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 305669de8bdSBart Van Assche char *lock_name; 306669de8bdSBart Van Assche struct lock_class_key key; 3074e6045f1SJohannes Berg struct lockdep_map lockdep_map; 3084e6045f1SJohannes Berg #endif 309ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 3102728fd2fSTejun Heo 311e2dca7adSTejun Heo /* 31224acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 31324acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 314e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 315e2dca7adSTejun Heo */ 316e2dca7adSTejun Heo struct rcu_head rcu; 317e2dca7adSTejun Heo 3182728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 3192728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 3202728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 3215b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 3221da177e4SLinus Torvalds }; 3231da177e4SLinus Torvalds 324e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 325e904e6c2STejun Heo 326bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 327bce90380STejun Heo /* possible CPUs of each node */ 328bce90380STejun Heo 329d55262c4STejun Heo static bool wq_disable_numa; 330d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 331d55262c4STejun Heo 332cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 333552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 334cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 335cee22a15SViresh Kumar 336863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3373347fa09STejun Heo 338bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 339bce90380STejun Heo 3404c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3414c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3424c16bd32STejun Heo 34368e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3441258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 345a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 346d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 347d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3485bcab335STejun Heo 349e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 35068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3517d19c5ceSTejun Heo 35299c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 353ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 354ef557180SMike Galbraith 355ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 356ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 357b05a7928SFrederic Weisbecker 358f303fccbSTejun Heo /* 359f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 360f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 361f303fccbSTejun Heo * to uncover usages which depend on it. 362f303fccbSTejun Heo */ 363f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 364f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 365f303fccbSTejun Heo #else 366f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 367f303fccbSTejun Heo #endif 368f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 369f303fccbSTejun Heo 3707d19c5ceSTejun Heo /* the per-cpu worker pools */ 37125528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3727d19c5ceSTejun Heo 37368e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3747d19c5ceSTejun Heo 37568e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 37629c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 37729c91e99STejun Heo 378c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 37929c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 38029c91e99STejun Heo 3818a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3828a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3838a2b7538STejun Heo 384d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 385ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 386044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3871aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 388044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 389d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 390044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 391f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 392044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 39324d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3940668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3950668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3960668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3970668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 398d320c038STejun Heo 3997d19c5ceSTejun Heo static int worker_thread(void *__worker); 4006ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 401c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 40255df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4037d19c5ceSTejun Heo 40497bd2347STejun Heo #define CREATE_TRACE_POINTS 40597bd2347STejun Heo #include <trace/events/workqueue.h> 40697bd2347STejun Heo 40768e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 40824acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 409f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 41024acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4115bcab335STejun Heo 4125b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 41324acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 414f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 415f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 41624acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4175b95e1afSLai Jiangshan 418f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 419f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 420f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4217a62c2c8STejun Heo (pool)++) 4224ce62e9eSTejun Heo 42349e3cf44STejun Heo /** 42417116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 42517116969STejun Heo * @pool: iteration cursor 426611c92a0STejun Heo * @pi: integer used for iteration 427fa1b54e6STejun Heo * 42824acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 42968e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 43068e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 431fa1b54e6STejun Heo * 432fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 433fa1b54e6STejun Heo * ignored. 43417116969STejun Heo */ 435611c92a0STejun Heo #define for_each_pool(pool, pi) \ 436611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 43768e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 438fa1b54e6STejun Heo else 43917116969STejun Heo 44017116969STejun Heo /** 441822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 442822d8405STejun Heo * @worker: iteration cursor 443822d8405STejun Heo * @pool: worker_pool to iterate workers of 444822d8405STejun Heo * 4451258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 446822d8405STejun Heo * 447822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 448822d8405STejun Heo * ignored. 449822d8405STejun Heo */ 450da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 451da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4521258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 453822d8405STejun Heo else 454822d8405STejun Heo 455822d8405STejun Heo /** 45649e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 45749e3cf44STejun Heo * @pwq: iteration cursor 45849e3cf44STejun Heo * @wq: the target workqueue 45976af4d93STejun Heo * 46024acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 461794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 462794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 46376af4d93STejun Heo * 46476af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 46576af4d93STejun Heo * ignored. 46649e3cf44STejun Heo */ 46749e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 46849e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4695a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 470f3421797STejun Heo 471dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 472dc186ad7SThomas Gleixner 473f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 474dc186ad7SThomas Gleixner 47599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 47699777288SStanislaw Gruszka { 47799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 47899777288SStanislaw Gruszka } 47999777288SStanislaw Gruszka 480b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 481b9fdac7fSDu, Changbin { 482b9fdac7fSDu, Changbin struct work_struct *work = addr; 483b9fdac7fSDu, Changbin 484b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 485b9fdac7fSDu, Changbin } 486b9fdac7fSDu, Changbin 487dc186ad7SThomas Gleixner /* 488dc186ad7SThomas Gleixner * fixup_init is called when: 489dc186ad7SThomas Gleixner * - an active object is initialized 490dc186ad7SThomas Gleixner */ 49102a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 492dc186ad7SThomas Gleixner { 493dc186ad7SThomas Gleixner struct work_struct *work = addr; 494dc186ad7SThomas Gleixner 495dc186ad7SThomas Gleixner switch (state) { 496dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 497dc186ad7SThomas Gleixner cancel_work_sync(work); 498dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 49902a982a6SDu, Changbin return true; 500dc186ad7SThomas Gleixner default: 50102a982a6SDu, Changbin return false; 502dc186ad7SThomas Gleixner } 503dc186ad7SThomas Gleixner } 504dc186ad7SThomas Gleixner 505dc186ad7SThomas Gleixner /* 506dc186ad7SThomas Gleixner * fixup_free is called when: 507dc186ad7SThomas Gleixner * - an active object is freed 508dc186ad7SThomas Gleixner */ 50902a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 510dc186ad7SThomas Gleixner { 511dc186ad7SThomas Gleixner struct work_struct *work = addr; 512dc186ad7SThomas Gleixner 513dc186ad7SThomas Gleixner switch (state) { 514dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 515dc186ad7SThomas Gleixner cancel_work_sync(work); 516dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 51702a982a6SDu, Changbin return true; 518dc186ad7SThomas Gleixner default: 51902a982a6SDu, Changbin return false; 520dc186ad7SThomas Gleixner } 521dc186ad7SThomas Gleixner } 522dc186ad7SThomas Gleixner 523f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 524dc186ad7SThomas Gleixner .name = "work_struct", 52599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 526b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 527dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 528dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 529dc186ad7SThomas Gleixner }; 530dc186ad7SThomas Gleixner 531dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 532dc186ad7SThomas Gleixner { 533dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 534dc186ad7SThomas Gleixner } 535dc186ad7SThomas Gleixner 536dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 537dc186ad7SThomas Gleixner { 538dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 539dc186ad7SThomas Gleixner } 540dc186ad7SThomas Gleixner 541dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 542dc186ad7SThomas Gleixner { 543dc186ad7SThomas Gleixner if (onstack) 544dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 545dc186ad7SThomas Gleixner else 546dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 547dc186ad7SThomas Gleixner } 548dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 549dc186ad7SThomas Gleixner 550dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 551dc186ad7SThomas Gleixner { 552dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 553dc186ad7SThomas Gleixner } 554dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 555dc186ad7SThomas Gleixner 556ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 557ea2e64f2SThomas Gleixner { 558ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 559ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 560ea2e64f2SThomas Gleixner } 561ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 562ea2e64f2SThomas Gleixner 563dc186ad7SThomas Gleixner #else 564dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 565dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 566dc186ad7SThomas Gleixner #endif 567dc186ad7SThomas Gleixner 5684e8b22bdSLi Bin /** 56967dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5704e8b22bdSLi Bin * @pool: the pool pointer of interest 5714e8b22bdSLi Bin * 5724e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5734e8b22bdSLi Bin * successfully, -errno on failure. 5744e8b22bdSLi Bin */ 5759daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5769daf9e67STejun Heo { 5779daf9e67STejun Heo int ret; 5789daf9e67STejun Heo 57968e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5805bcab335STejun Heo 5814e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5824e8b22bdSLi Bin GFP_KERNEL); 583229641a6STejun Heo if (ret >= 0) { 584e68035fbSTejun Heo pool->id = ret; 585229641a6STejun Heo return 0; 586229641a6STejun Heo } 5879daf9e67STejun Heo return ret; 5889daf9e67STejun Heo } 5899daf9e67STejun Heo 59076af4d93STejun Heo /** 591df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 592df2d5ae4STejun Heo * @wq: the target workqueue 593df2d5ae4STejun Heo * @node: the node ID 594df2d5ae4STejun Heo * 59524acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5965b95e1afSLai Jiangshan * read locked. 597df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 598df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 599d185af30SYacine Belkadi * 600d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 601df2d5ae4STejun Heo */ 602df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 603df2d5ae4STejun Heo int node) 604df2d5ae4STejun Heo { 6055b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 606d6e022f1STejun Heo 607d6e022f1STejun Heo /* 608d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 609d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 610d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 611d6e022f1STejun Heo * happens, this workaround can be removed. 612d6e022f1STejun Heo */ 613d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 614d6e022f1STejun Heo return wq->dfl_pwq; 615d6e022f1STejun Heo 616df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 617df2d5ae4STejun Heo } 618df2d5ae4STejun Heo 61973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 62073f53c4aSTejun Heo { 62173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 62273f53c4aSTejun Heo } 62373f53c4aSTejun Heo 624c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 62573f53c4aSTejun Heo { 626c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 62773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 62873f53c4aSTejun Heo } 62973f53c4aSTejun Heo 63073f53c4aSTejun Heo static int work_next_color(int color) 63173f53c4aSTejun Heo { 63273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 633a848e3b6SOleg Nesterov } 634a848e3b6SOleg Nesterov 6354594bf15SDavid Howells /* 636112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 637112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6387c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6397a22ad75STejun Heo * 640112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 641112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 642bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 643bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6447a22ad75STejun Heo * 645112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6467c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 647112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6487c3eed5cSTejun Heo * available only while the work item is queued. 649bbb68dfaSTejun Heo * 650bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 651bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 652bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 653bbb68dfaSTejun Heo * try to steal the PENDING bit. 6544594bf15SDavid Howells */ 6557a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6567a22ad75STejun Heo unsigned long flags) 6577a22ad75STejun Heo { 6586183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6597a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6607a22ad75STejun Heo } 6617a22ad75STejun Heo 662112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6634690c4abSTejun Heo unsigned long extra_flags) 664365970a1SDavid Howells { 665112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 666112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 667365970a1SDavid Howells } 668365970a1SDavid Howells 6694468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6704468a00fSLai Jiangshan int pool_id) 6714468a00fSLai Jiangshan { 6724468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6734468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6744468a00fSLai Jiangshan } 6754468a00fSLai Jiangshan 6767c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6777c3eed5cSTejun Heo int pool_id) 6784d707b9fSOleg Nesterov { 67923657bb1STejun Heo /* 68023657bb1STejun Heo * The following wmb is paired with the implied mb in 68123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 68223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 68323657bb1STejun Heo * owner. 68423657bb1STejun Heo */ 68523657bb1STejun Heo smp_wmb(); 6867c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 687346c09f8SRoman Pen /* 688346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 689346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 690346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6918bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 692346c09f8SRoman Pen * the same @work. E.g. consider this case: 693346c09f8SRoman Pen * 694346c09f8SRoman Pen * CPU#0 CPU#1 695346c09f8SRoman Pen * ---------------------------- -------------------------------- 696346c09f8SRoman Pen * 697346c09f8SRoman Pen * 1 STORE event_indicated 698346c09f8SRoman Pen * 2 queue_work_on() { 699346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 700346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 701346c09f8SRoman Pen * 5 set_work_data() # clear bit 702346c09f8SRoman Pen * 6 smp_mb() 703346c09f8SRoman Pen * 7 work->current_func() { 704346c09f8SRoman Pen * 8 LOAD event_indicated 705346c09f8SRoman Pen * } 706346c09f8SRoman Pen * 707346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 708346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 709346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 710346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 711346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 712346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 713346c09f8SRoman Pen * before actual STORE. 714346c09f8SRoman Pen */ 715346c09f8SRoman Pen smp_mb(); 7164d707b9fSOleg Nesterov } 7174d707b9fSOleg Nesterov 7187a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 719365970a1SDavid Howells { 7207c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7217c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7227a22ad75STejun Heo } 7237a22ad75STejun Heo 724112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7257a22ad75STejun Heo { 726e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7277a22ad75STejun Heo 728112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 729e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 730e120153dSTejun Heo else 731e120153dSTejun Heo return NULL; 7327a22ad75STejun Heo } 7337a22ad75STejun Heo 7347c3eed5cSTejun Heo /** 7357c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7367c3eed5cSTejun Heo * @work: the work item of interest 7377c3eed5cSTejun Heo * 73868e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 73924acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 74024acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 741fa1b54e6STejun Heo * 742fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 743fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 744fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 745fa1b54e6STejun Heo * returned pool is and stays online. 746d185af30SYacine Belkadi * 747d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7487c3eed5cSTejun Heo */ 7497c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7507a22ad75STejun Heo { 751e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7527c3eed5cSTejun Heo int pool_id; 7537a22ad75STejun Heo 75468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 755fa1b54e6STejun Heo 756112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 757112202d9STejun Heo return ((struct pool_workqueue *) 7587c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7597a22ad75STejun Heo 7607c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7617c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7627a22ad75STejun Heo return NULL; 7637a22ad75STejun Heo 764fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7657c3eed5cSTejun Heo } 7667c3eed5cSTejun Heo 7677c3eed5cSTejun Heo /** 7687c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7697c3eed5cSTejun Heo * @work: the work item of interest 7707c3eed5cSTejun Heo * 771d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7727c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7737c3eed5cSTejun Heo */ 7747c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7757c3eed5cSTejun Heo { 77654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7777c3eed5cSTejun Heo 778112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 779112202d9STejun Heo return ((struct pool_workqueue *) 78054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 78154d5b7d0SLai Jiangshan 78254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7837c3eed5cSTejun Heo } 7847c3eed5cSTejun Heo 785bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 786bbb68dfaSTejun Heo { 7877c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 788bbb68dfaSTejun Heo 7897c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7907c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 791bbb68dfaSTejun Heo } 792bbb68dfaSTejun Heo 793bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 794bbb68dfaSTejun Heo { 795bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 796bbb68dfaSTejun Heo 797112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 798bbb68dfaSTejun Heo } 799bbb68dfaSTejun Heo 800e22bee78STejun Heo /* 8013270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8023270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 803d565ed63STejun Heo * they're being called with pool->lock held. 804e22bee78STejun Heo */ 805e22bee78STejun Heo 80663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 807649027d7STejun Heo { 808bc35f7efSLai Jiangshan return !pool->nr_running; 809649027d7STejun Heo } 810649027d7STejun Heo 811e22bee78STejun Heo /* 812e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 813e22bee78STejun Heo * running workers. 814974271c4STejun Heo * 815974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 816706026c2STejun Heo * function will always return %true for unbound pools as long as the 817974271c4STejun Heo * worklist isn't empty. 818e22bee78STejun Heo */ 81963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 820e22bee78STejun Heo { 82163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 822e22bee78STejun Heo } 823e22bee78STejun Heo 824e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 82563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 826e22bee78STejun Heo { 82763d95a91STejun Heo return pool->nr_idle; 828e22bee78STejun Heo } 829e22bee78STejun Heo 830e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 83163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 832e22bee78STejun Heo { 833bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 834e22bee78STejun Heo } 835e22bee78STejun Heo 836e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 83763d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 838e22bee78STejun Heo { 83963d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 840e22bee78STejun Heo } 841e22bee78STejun Heo 842e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 84363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 844e22bee78STejun Heo { 845692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 84663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 84763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 848e22bee78STejun Heo 849e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 850e22bee78STejun Heo } 851e22bee78STejun Heo 852e22bee78STejun Heo /* 853e22bee78STejun Heo * Wake up functions. 854e22bee78STejun Heo */ 855e22bee78STejun Heo 8562c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8571037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8587e11629dSTejun Heo { 85963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8607e11629dSTejun Heo return NULL; 8617e11629dSTejun Heo 86263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8637e11629dSTejun Heo } 8647e11629dSTejun Heo 8657e11629dSTejun Heo /** 8667e11629dSTejun Heo * wake_up_worker - wake up an idle worker 86763d95a91STejun Heo * @pool: worker pool to wake worker from 8687e11629dSTejun Heo * 86963d95a91STejun Heo * Wake up the first idle worker of @pool. 8707e11629dSTejun Heo * 8717e11629dSTejun Heo * CONTEXT: 872a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8737e11629dSTejun Heo */ 87463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8757e11629dSTejun Heo { 8761037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8777e11629dSTejun Heo 8787e11629dSTejun Heo if (likely(worker)) 8797e11629dSTejun Heo wake_up_process(worker->task); 8807e11629dSTejun Heo } 8817e11629dSTejun Heo 8824690c4abSTejun Heo /** 883*c54d5046STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 884*c54d5046STejun Heo * @worker: self 885*c54d5046STejun Heo * @flags: flags to set 886*c54d5046STejun Heo * 887*c54d5046STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. 888*c54d5046STejun Heo * 889*c54d5046STejun Heo * CONTEXT: 890*c54d5046STejun Heo * raw_spin_lock_irq(pool->lock) 891*c54d5046STejun Heo */ 892*c54d5046STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags) 893*c54d5046STejun Heo { 894*c54d5046STejun Heo struct worker_pool *pool = worker->pool; 895*c54d5046STejun Heo 896*c54d5046STejun Heo WARN_ON_ONCE(worker->task != current); 897*c54d5046STejun Heo 898*c54d5046STejun Heo /* If transitioning into NOT_RUNNING, adjust nr_running. */ 899*c54d5046STejun Heo if ((flags & WORKER_NOT_RUNNING) && 900*c54d5046STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 901*c54d5046STejun Heo pool->nr_running--; 902*c54d5046STejun Heo } 903*c54d5046STejun Heo 904*c54d5046STejun Heo worker->flags |= flags; 905*c54d5046STejun Heo } 906*c54d5046STejun Heo 907*c54d5046STejun Heo /** 908*c54d5046STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 909*c54d5046STejun Heo * @worker: self 910*c54d5046STejun Heo * @flags: flags to clear 911*c54d5046STejun Heo * 912*c54d5046STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 913*c54d5046STejun Heo * 914*c54d5046STejun Heo * CONTEXT: 915*c54d5046STejun Heo * raw_spin_lock_irq(pool->lock) 916*c54d5046STejun Heo */ 917*c54d5046STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 918*c54d5046STejun Heo { 919*c54d5046STejun Heo struct worker_pool *pool = worker->pool; 920*c54d5046STejun Heo unsigned int oflags = worker->flags; 921*c54d5046STejun Heo 922*c54d5046STejun Heo WARN_ON_ONCE(worker->task != current); 923*c54d5046STejun Heo 924*c54d5046STejun Heo worker->flags &= ~flags; 925*c54d5046STejun Heo 926*c54d5046STejun Heo /* 927*c54d5046STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 928*c54d5046STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 929*c54d5046STejun Heo * of multiple flags, not a single flag. 930*c54d5046STejun Heo */ 931*c54d5046STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 932*c54d5046STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 933*c54d5046STejun Heo pool->nr_running++; 934*c54d5046STejun Heo } 935*c54d5046STejun Heo 936*c54d5046STejun Heo /** 9376d25be57SThomas Gleixner * wq_worker_running - a worker is running again 938e22bee78STejun Heo * @task: task waking up 939e22bee78STejun Heo * 9406d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 941e22bee78STejun Heo */ 9426d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 943e22bee78STejun Heo { 944e22bee78STejun Heo struct worker *worker = kthread_data(task); 945e22bee78STejun Heo 9466d25be57SThomas Gleixner if (!worker->sleeping) 9476d25be57SThomas Gleixner return; 94807edfeceSFrederic Weisbecker 94907edfeceSFrederic Weisbecker /* 95007edfeceSFrederic Weisbecker * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 95107edfeceSFrederic Weisbecker * and the nr_running increment below, we may ruin the nr_running reset 95207edfeceSFrederic Weisbecker * and leave with an unexpected pool->nr_running == 1 on the newly unbound 95307edfeceSFrederic Weisbecker * pool. Protect against such race. 95407edfeceSFrederic Weisbecker */ 95507edfeceSFrederic Weisbecker preempt_disable(); 9566d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 957bc35f7efSLai Jiangshan worker->pool->nr_running++; 95807edfeceSFrederic Weisbecker preempt_enable(); 9596d25be57SThomas Gleixner worker->sleeping = 0; 96036576000SJoonsoo Kim } 961e22bee78STejun Heo 962e22bee78STejun Heo /** 963e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 964e22bee78STejun Heo * @task: task going to sleep 965e22bee78STejun Heo * 9666d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 967ccf45156SLai Jiangshan * going to sleep. 968e22bee78STejun Heo */ 9696d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 970e22bee78STejun Heo { 971cc5bff38SLai Jiangshan struct worker *worker = kthread_data(task); 972111c225aSTejun Heo struct worker_pool *pool; 973e22bee78STejun Heo 974111c225aSTejun Heo /* 975111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 976111c225aSTejun Heo * workers, also reach here, let's not access anything before 977111c225aSTejun Heo * checking NOT_RUNNING. 978111c225aSTejun Heo */ 9792d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 9806d25be57SThomas Gleixner return; 981e22bee78STejun Heo 982111c225aSTejun Heo pool = worker->pool; 983111c225aSTejun Heo 98462849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 98562849a96SSebastian Andrzej Siewior if (worker->sleeping) 9866d25be57SThomas Gleixner return; 9876d25be57SThomas Gleixner 9886d25be57SThomas Gleixner worker->sleeping = 1; 989a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 990e22bee78STejun Heo 991e22bee78STejun Heo /* 99245c753f5SFrederic Weisbecker * Recheck in case unbind_workers() preempted us. We don't 99345c753f5SFrederic Weisbecker * want to decrement nr_running after the worker is unbound 99445c753f5SFrederic Weisbecker * and nr_running has been reset. 99545c753f5SFrederic Weisbecker */ 99645c753f5SFrederic Weisbecker if (worker->flags & WORKER_NOT_RUNNING) { 99745c753f5SFrederic Weisbecker raw_spin_unlock_irq(&pool->lock); 99845c753f5SFrederic Weisbecker return; 99945c753f5SFrederic Weisbecker } 100045c753f5SFrederic Weisbecker 1001bc35f7efSLai Jiangshan pool->nr_running--; 1002725e8ec5STejun Heo if (need_more_worker(pool)) { 1003725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1004cc5bff38SLai Jiangshan wake_up_worker(pool); 1005725e8ec5STejun Heo } 1006a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1007e22bee78STejun Heo } 1008e22bee78STejun Heo 1009e22bee78STejun Heo /** 10101b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 10118194fe94SBart Van Assche * @task: Task to retrieve last work function of. 10121b69ac6bSJohannes Weiner * 10131b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 10141b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 10151b69ac6bSJohannes Weiner * 10161b69ac6bSJohannes Weiner * CONTEXT: 1017a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 10181b69ac6bSJohannes Weiner * 10194b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 10204b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 10214b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 10224b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 10234b047002SJohannes Weiner * 10244b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 10254b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 10264b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 10274b047002SJohannes Weiner * is guaranteed to not be processing any works. 10284b047002SJohannes Weiner * 10291b69ac6bSJohannes Weiner * Return: 10301b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 10311b69ac6bSJohannes Weiner * hasn't executed any work yet. 10321b69ac6bSJohannes Weiner */ 10331b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 10341b69ac6bSJohannes Weiner { 10351b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 10361b69ac6bSJohannes Weiner 10371b69ac6bSJohannes Weiner return worker->last_func; 10381b69ac6bSJohannes Weiner } 10391b69ac6bSJohannes Weiner 10401b69ac6bSJohannes Weiner /** 10418cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1042c9e7cf27STejun Heo * @pool: pool of interest 10438cca0eeaSTejun Heo * @work: work to find worker for 10448cca0eeaSTejun Heo * 1045c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1046c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1047a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1048a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1049a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1050a2c1c57bSTejun Heo * being executed. 1051a2c1c57bSTejun Heo * 1052a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1053a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1054a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1055a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1056a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1057a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1058a2c1c57bSTejun Heo * 1059c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1060c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1061c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1062c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1063c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1064c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10658cca0eeaSTejun Heo * 10668cca0eeaSTejun Heo * CONTEXT: 1067a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10688cca0eeaSTejun Heo * 1069d185af30SYacine Belkadi * Return: 1070d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10718cca0eeaSTejun Heo * otherwise. 10728cca0eeaSTejun Heo */ 1073c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10748cca0eeaSTejun Heo struct work_struct *work) 10758cca0eeaSTejun Heo { 107642f8570fSSasha Levin struct worker *worker; 107742f8570fSSasha Levin 1078b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1079a2c1c57bSTejun Heo (unsigned long)work) 1080a2c1c57bSTejun Heo if (worker->current_work == work && 1081a2c1c57bSTejun Heo worker->current_func == work->func) 108242f8570fSSasha Levin return worker; 108342f8570fSSasha Levin 108442f8570fSSasha Levin return NULL; 10858cca0eeaSTejun Heo } 10868cca0eeaSTejun Heo 10878cca0eeaSTejun Heo /** 1088bf4ede01STejun Heo * move_linked_works - move linked works to a list 1089bf4ede01STejun Heo * @work: start of series of works to be scheduled 1090bf4ede01STejun Heo * @head: target list to append @work to 1091402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1092bf4ede01STejun Heo * 1093bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1094bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1095bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1096bf4ede01STejun Heo * 1097bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1098bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1099bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1100bf4ede01STejun Heo * 1101bf4ede01STejun Heo * CONTEXT: 1102a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1103bf4ede01STejun Heo */ 1104bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1105bf4ede01STejun Heo struct work_struct **nextp) 1106bf4ede01STejun Heo { 1107bf4ede01STejun Heo struct work_struct *n; 1108bf4ede01STejun Heo 1109bf4ede01STejun Heo /* 1110bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1111bf4ede01STejun Heo * use NULL for list head. 1112bf4ede01STejun Heo */ 1113bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1114bf4ede01STejun Heo list_move_tail(&work->entry, head); 1115bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1116bf4ede01STejun Heo break; 1117bf4ede01STejun Heo } 1118bf4ede01STejun Heo 1119bf4ede01STejun Heo /* 1120bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1121bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1122bf4ede01STejun Heo * needs to be updated. 1123bf4ede01STejun Heo */ 1124bf4ede01STejun Heo if (nextp) 1125bf4ede01STejun Heo *nextp = n; 1126bf4ede01STejun Heo } 1127bf4ede01STejun Heo 11288864b4e5STejun Heo /** 11298864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 11308864b4e5STejun Heo * @pwq: pool_workqueue to get 11318864b4e5STejun Heo * 11328864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 11338864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 11348864b4e5STejun Heo */ 11358864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 11368864b4e5STejun Heo { 11378864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11388864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 11398864b4e5STejun Heo pwq->refcnt++; 11408864b4e5STejun Heo } 11418864b4e5STejun Heo 11428864b4e5STejun Heo /** 11438864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 11448864b4e5STejun Heo * @pwq: pool_workqueue to put 11458864b4e5STejun Heo * 11468864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11478864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11488864b4e5STejun Heo */ 11498864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11508864b4e5STejun Heo { 11518864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11528864b4e5STejun Heo if (likely(--pwq->refcnt)) 11538864b4e5STejun Heo return; 11548864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11558864b4e5STejun Heo return; 11568864b4e5STejun Heo /* 11578864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11588864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11598864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11608864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11618864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11628864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11638864b4e5STejun Heo */ 11648864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11658864b4e5STejun Heo } 11668864b4e5STejun Heo 1167dce90d47STejun Heo /** 1168dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1169dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1170dce90d47STejun Heo * 1171dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1172dce90d47STejun Heo */ 1173dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1174dce90d47STejun Heo { 1175dce90d47STejun Heo if (pwq) { 1176dce90d47STejun Heo /* 117724acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1178dce90d47STejun Heo * following lock operations are safe. 1179dce90d47STejun Heo */ 1180a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1181dce90d47STejun Heo put_pwq(pwq); 1182a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1183dce90d47STejun Heo } 1184dce90d47STejun Heo } 1185dce90d47STejun Heo 1186f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1187bf4ede01STejun Heo { 1188112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1189bf4ede01STejun Heo 1190bf4ede01STejun Heo trace_workqueue_activate_work(work); 119182607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 119282607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1193112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1194f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1195112202d9STejun Heo pwq->nr_active++; 1196bf4ede01STejun Heo } 1197bf4ede01STejun Heo 1198f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 11993aa62497SLai Jiangshan { 1200f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 12013aa62497SLai Jiangshan struct work_struct, entry); 12023aa62497SLai Jiangshan 1203f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 12043aa62497SLai Jiangshan } 12053aa62497SLai Jiangshan 1206bf4ede01STejun Heo /** 1207112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1208112202d9STejun Heo * @pwq: pwq of interest 1209c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1210bf4ede01STejun Heo * 1211bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1212112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1213bf4ede01STejun Heo * 1214bf4ede01STejun Heo * CONTEXT: 1215a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1216bf4ede01STejun Heo */ 1217c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1218bf4ede01STejun Heo { 1219c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1220c4560c2cSLai Jiangshan 1221018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1222112202d9STejun Heo pwq->nr_active--; 1223f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1224f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1225112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1226f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1227bf4ede01STejun Heo } 1228018f3a13SLai Jiangshan } 1229018f3a13SLai Jiangshan 1230018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1231bf4ede01STejun Heo 1232bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1233112202d9STejun Heo if (likely(pwq->flush_color != color)) 12348864b4e5STejun Heo goto out_put; 1235bf4ede01STejun Heo 1236bf4ede01STejun Heo /* are there still in-flight works? */ 1237112202d9STejun Heo if (pwq->nr_in_flight[color]) 12388864b4e5STejun Heo goto out_put; 1239bf4ede01STejun Heo 1240112202d9STejun Heo /* this pwq is done, clear flush_color */ 1241112202d9STejun Heo pwq->flush_color = -1; 1242bf4ede01STejun Heo 1243bf4ede01STejun Heo /* 1244112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1245bf4ede01STejun Heo * will handle the rest. 1246bf4ede01STejun Heo */ 1247112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1248112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12498864b4e5STejun Heo out_put: 12508864b4e5STejun Heo put_pwq(pwq); 1251bf4ede01STejun Heo } 1252bf4ede01STejun Heo 125336e227d2STejun Heo /** 1254bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 125536e227d2STejun Heo * @work: work item to steal 125636e227d2STejun Heo * @is_dwork: @work is a delayed_work 1257bbb68dfaSTejun Heo * @flags: place to store irq state 125836e227d2STejun Heo * 125936e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1260d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 126136e227d2STejun Heo * 1262d185af30SYacine Belkadi * Return: 12633eb6b31bSMauro Carvalho Chehab * 12643eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 126536e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 126636e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 126736e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1268bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1269bbb68dfaSTejun Heo * for arbitrarily long 12703eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 127136e227d2STejun Heo * 1272d185af30SYacine Belkadi * Note: 1273bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1274e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1275e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1276e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1277bbb68dfaSTejun Heo * 1278bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1279bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1280bbb68dfaSTejun Heo * 1281e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1282bf4ede01STejun Heo */ 1283bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1284bbb68dfaSTejun Heo unsigned long *flags) 1285bf4ede01STejun Heo { 1286d565ed63STejun Heo struct worker_pool *pool; 1287112202d9STejun Heo struct pool_workqueue *pwq; 1288bf4ede01STejun Heo 1289bbb68dfaSTejun Heo local_irq_save(*flags); 1290bbb68dfaSTejun Heo 129136e227d2STejun Heo /* try to steal the timer if it exists */ 129236e227d2STejun Heo if (is_dwork) { 129336e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 129436e227d2STejun Heo 1295e0aecdd8STejun Heo /* 1296e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1297e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1298e0aecdd8STejun Heo * running on the local CPU. 1299e0aecdd8STejun Heo */ 130036e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 130136e227d2STejun Heo return 1; 130236e227d2STejun Heo } 130336e227d2STejun Heo 130436e227d2STejun Heo /* try to claim PENDING the normal way */ 1305bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1306bf4ede01STejun Heo return 0; 1307bf4ede01STejun Heo 130824acfb71SThomas Gleixner rcu_read_lock(); 1309bf4ede01STejun Heo /* 1310bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1311bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1312bf4ede01STejun Heo */ 1313d565ed63STejun Heo pool = get_work_pool(work); 1314d565ed63STejun Heo if (!pool) 1315bbb68dfaSTejun Heo goto fail; 1316bf4ede01STejun Heo 1317a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1318bf4ede01STejun Heo /* 1319112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1320112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1321112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1322112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1323112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 13240b3dae68SLai Jiangshan * item is currently queued on that pool. 1325bf4ede01STejun Heo */ 1326112202d9STejun Heo pwq = get_work_pwq(work); 1327112202d9STejun Heo if (pwq && pwq->pool == pool) { 1328bf4ede01STejun Heo debug_work_deactivate(work); 13293aa62497SLai Jiangshan 13303aa62497SLai Jiangshan /* 1331018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1332018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1333018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1334018f3a13SLai Jiangshan * 1335f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1336d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1337f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 133816062836STejun Heo * management later on and cause stall. Make sure the work 133916062836STejun Heo * item is activated before grabbing. 13403aa62497SLai Jiangshan */ 1341f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1342f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13433aa62497SLai Jiangshan 1344bf4ede01STejun Heo list_del_init(&work->entry); 1345c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 134636e227d2STejun Heo 1347112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 13484468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 13494468a00fSLai Jiangshan 1350a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 135124acfb71SThomas Gleixner rcu_read_unlock(); 135236e227d2STejun Heo return 1; 1353bf4ede01STejun Heo } 1354a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1355bbb68dfaSTejun Heo fail: 135624acfb71SThomas Gleixner rcu_read_unlock(); 1357bbb68dfaSTejun Heo local_irq_restore(*flags); 1358bbb68dfaSTejun Heo if (work_is_canceling(work)) 1359bbb68dfaSTejun Heo return -ENOENT; 1360bbb68dfaSTejun Heo cpu_relax(); 136136e227d2STejun Heo return -EAGAIN; 1362bf4ede01STejun Heo } 1363bf4ede01STejun Heo 1364bf4ede01STejun Heo /** 1365706026c2STejun Heo * insert_work - insert a work into a pool 1366112202d9STejun Heo * @pwq: pwq @work belongs to 13674690c4abSTejun Heo * @work: work to insert 13684690c4abSTejun Heo * @head: insertion point 13694690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13704690c4abSTejun Heo * 1371112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1372706026c2STejun Heo * work_struct flags. 13734690c4abSTejun Heo * 13744690c4abSTejun Heo * CONTEXT: 1375a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1376365970a1SDavid Howells */ 1377112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1378112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1379b89deed3SOleg Nesterov { 1380112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1381e1d8aa9fSFrederic Weisbecker 1382e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1383f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1384e89a85d6SWalter Wu 13854690c4abSTejun Heo /* we own @work, set data and link */ 1386112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13871a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13888864b4e5STejun Heo get_pwq(pwq); 1389e22bee78STejun Heo 139063d95a91STejun Heo if (__need_more_worker(pool)) 139163d95a91STejun Heo wake_up_worker(pool); 1392b89deed3SOleg Nesterov } 1393b89deed3SOleg Nesterov 1394c8efcc25STejun Heo /* 1395c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13968d03ecfeSTejun Heo * same workqueue. 1397c8efcc25STejun Heo */ 1398c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1399c8efcc25STejun Heo { 1400c8efcc25STejun Heo struct worker *worker; 1401c8efcc25STejun Heo 14028d03ecfeSTejun Heo worker = current_wq_worker(); 1403c8efcc25STejun Heo /* 1404bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 14058d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1406c8efcc25STejun Heo */ 1407112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1408c8efcc25STejun Heo } 1409c8efcc25STejun Heo 1410ef557180SMike Galbraith /* 1411ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1412ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1413ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1414ef557180SMike Galbraith */ 1415ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1416ef557180SMike Galbraith { 1417ef557180SMike Galbraith int new_cpu; 1418ef557180SMike Galbraith 1419f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1420ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1421ef557180SMike Galbraith return cpu; 1422a8ec5880SAmmar Faizi } else { 1423a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1424f303fccbSTejun Heo } 1425f303fccbSTejun Heo 1426ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1427ef557180SMike Galbraith return cpu; 1428ef557180SMike Galbraith 1429ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1430ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1431ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1432ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1433ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1434ef557180SMike Galbraith return cpu; 1435ef557180SMike Galbraith } 1436ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1437ef557180SMike Galbraith 1438ef557180SMike Galbraith return new_cpu; 1439ef557180SMike Galbraith } 1440ef557180SMike Galbraith 1441d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 14421da177e4SLinus Torvalds struct work_struct *work) 14431da177e4SLinus Torvalds { 1444112202d9STejun Heo struct pool_workqueue *pwq; 1445c9178087STejun Heo struct worker_pool *last_pool; 14461e19ffc6STejun Heo struct list_head *worklist; 14478a2e8e5dSTejun Heo unsigned int work_flags; 1448b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14498930cabaSTejun Heo 14508930cabaSTejun Heo /* 14518930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14528930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14538930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14548930cabaSTejun Heo * happen with IRQ disabled. 14558930cabaSTejun Heo */ 14568e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14571da177e4SLinus Torvalds 14581e19ffc6STejun Heo 145933e3f0a3SRichard Clark /* 146033e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 146133e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 146233e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 146333e3f0a3SRichard Clark */ 146433e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 146533e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1466e41e704bSTejun Heo return; 146724acfb71SThomas Gleixner rcu_read_lock(); 14689e8cd2f5STejun Heo retry: 1469aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1470aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1471df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1472ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1473df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1474aa202f1fSHillf Danton } else { 1475aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1476aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1477aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1478aa202f1fSHillf Danton } 1479f3421797STejun Heo 148018aa9effSTejun Heo /* 1481c9178087STejun Heo * If @work was previously on a different pool, it might still be 1482c9178087STejun Heo * running there, in which case the work needs to be queued on that 1483c9178087STejun Heo * pool to guarantee non-reentrancy. 148418aa9effSTejun Heo */ 1485c9e7cf27STejun Heo last_pool = get_work_pool(work); 1486112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 148718aa9effSTejun Heo struct worker *worker; 148818aa9effSTejun Heo 1489a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 149018aa9effSTejun Heo 1491c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 149218aa9effSTejun Heo 1493112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1494c9178087STejun Heo pwq = worker->current_pwq; 14958594fadeSLai Jiangshan } else { 149618aa9effSTejun Heo /* meh... not running there, queue here */ 1497a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1498a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 149918aa9effSTejun Heo } 15008930cabaSTejun Heo } else { 1501a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 15028930cabaSTejun Heo } 1503502ca9d8STejun Heo 15049e8cd2f5STejun Heo /* 15059e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 15069e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 15079e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1508df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1509df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 15109e8cd2f5STejun Heo * make forward-progress. 15119e8cd2f5STejun Heo */ 15129e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 15139e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1514a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 15159e8cd2f5STejun Heo cpu_relax(); 15169e8cd2f5STejun Heo goto retry; 15179e8cd2f5STejun Heo } 15189e8cd2f5STejun Heo /* oops */ 15199e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 15209e8cd2f5STejun Heo wq->name, cpu); 15219e8cd2f5STejun Heo } 15229e8cd2f5STejun Heo 1523112202d9STejun Heo /* pwq determined, queue */ 1524112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1525502ca9d8STejun Heo 152624acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 152724acfb71SThomas Gleixner goto out; 15281e19ffc6STejun Heo 1529112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1530112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 15311e19ffc6STejun Heo 1532112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1533cdadf009STejun Heo trace_workqueue_activate_work(work); 1534112202d9STejun Heo pwq->nr_active++; 1535112202d9STejun Heo worklist = &pwq->pool->worklist; 153682607adcSTejun Heo if (list_empty(worklist)) 153782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 15388a2e8e5dSTejun Heo } else { 1539f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1540f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 15418a2e8e5dSTejun Heo } 15421e19ffc6STejun Heo 15430687c66bSZqiang debug_work_activate(work); 1544112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15451e19ffc6STejun Heo 154624acfb71SThomas Gleixner out: 1547a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 154824acfb71SThomas Gleixner rcu_read_unlock(); 15491da177e4SLinus Torvalds } 15501da177e4SLinus Torvalds 15510fcb78c2SRolf Eike Beer /** 1552c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1553c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1554c1a220e7SZhang Rui * @wq: workqueue to use 1555c1a220e7SZhang Rui * @work: work to queue 1556c1a220e7SZhang Rui * 1557c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1558443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1559443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1560854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1561854f5cc5SPaul E. McKenney * online will get a splat. 1562d185af30SYacine Belkadi * 1563d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1564c1a220e7SZhang Rui */ 1565d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1566d4283e93STejun Heo struct work_struct *work) 1567c1a220e7SZhang Rui { 1568d4283e93STejun Heo bool ret = false; 15698930cabaSTejun Heo unsigned long flags; 15708930cabaSTejun Heo 15718930cabaSTejun Heo local_irq_save(flags); 1572c1a220e7SZhang Rui 157322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15744690c4abSTejun Heo __queue_work(cpu, wq, work); 1575d4283e93STejun Heo ret = true; 1576c1a220e7SZhang Rui } 15778930cabaSTejun Heo 15788930cabaSTejun Heo local_irq_restore(flags); 1579c1a220e7SZhang Rui return ret; 1580c1a220e7SZhang Rui } 1581ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1582c1a220e7SZhang Rui 15838204e0c1SAlexander Duyck /** 15848204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15858204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15868204e0c1SAlexander Duyck * 15878204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15888204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15898204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15908204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15918204e0c1SAlexander Duyck */ 15928204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15938204e0c1SAlexander Duyck { 15948204e0c1SAlexander Duyck int cpu; 15958204e0c1SAlexander Duyck 15968204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15978204e0c1SAlexander Duyck if (!wq_numa_enabled) 15988204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15998204e0c1SAlexander Duyck 16008204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 16018204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 16028204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 16038204e0c1SAlexander Duyck 16048204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 16058204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 16068204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 16078204e0c1SAlexander Duyck return cpu; 16088204e0c1SAlexander Duyck 16098204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 16108204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 16118204e0c1SAlexander Duyck 16128204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 16138204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 16148204e0c1SAlexander Duyck } 16158204e0c1SAlexander Duyck 16168204e0c1SAlexander Duyck /** 16178204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 16188204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 16198204e0c1SAlexander Duyck * @wq: workqueue to use 16208204e0c1SAlexander Duyck * @work: work to queue 16218204e0c1SAlexander Duyck * 16228204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 16238204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 16248204e0c1SAlexander Duyck * NUMA node. 16258204e0c1SAlexander Duyck * 16268204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 16278204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 16288204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 16298204e0c1SAlexander Duyck * 16308204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 16318204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 16328204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 16338204e0c1SAlexander Duyck * 16348204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 16358204e0c1SAlexander Duyck */ 16368204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 16378204e0c1SAlexander Duyck struct work_struct *work) 16388204e0c1SAlexander Duyck { 16398204e0c1SAlexander Duyck unsigned long flags; 16408204e0c1SAlexander Duyck bool ret = false; 16418204e0c1SAlexander Duyck 16428204e0c1SAlexander Duyck /* 16438204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 16448204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 16458204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 16468204e0c1SAlexander Duyck * 16478204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16488204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16498204e0c1SAlexander Duyck * some round robin type logic. 16508204e0c1SAlexander Duyck */ 16518204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16528204e0c1SAlexander Duyck 16538204e0c1SAlexander Duyck local_irq_save(flags); 16548204e0c1SAlexander Duyck 16558204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16568204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16578204e0c1SAlexander Duyck 16588204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16598204e0c1SAlexander Duyck ret = true; 16608204e0c1SAlexander Duyck } 16618204e0c1SAlexander Duyck 16628204e0c1SAlexander Duyck local_irq_restore(flags); 16638204e0c1SAlexander Duyck return ret; 16648204e0c1SAlexander Duyck } 16658204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16668204e0c1SAlexander Duyck 16678c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16681da177e4SLinus Torvalds { 16698c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16701da177e4SLinus Torvalds 1671e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 167260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16731da177e4SLinus Torvalds } 16741438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16751da177e4SLinus Torvalds 16767beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 167752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16781da177e4SLinus Torvalds { 16797beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16807beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16811da177e4SLinus Torvalds 1682637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 16834b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1684fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1685fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16867beb2edfSTejun Heo 16878852aac2STejun Heo /* 16888852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16898852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16908852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16918852aac2STejun Heo * on that there's no such delay when @delay is 0. 16928852aac2STejun Heo */ 16938852aac2STejun Heo if (!delay) { 16948852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16958852aac2STejun Heo return; 16968852aac2STejun Heo } 16978852aac2STejun Heo 169860c057bcSLai Jiangshan dwork->wq = wq; 16991265057fSTejun Heo dwork->cpu = cpu; 17007beb2edfSTejun Heo timer->expires = jiffies + delay; 17017beb2edfSTejun Heo 1702041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 17037beb2edfSTejun Heo add_timer_on(timer, cpu); 1704041bd12eSTejun Heo else 1705041bd12eSTejun Heo add_timer(timer); 17067beb2edfSTejun Heo } 17071da177e4SLinus Torvalds 17080fcb78c2SRolf Eike Beer /** 17090fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 17100fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 17110fcb78c2SRolf Eike Beer * @wq: workqueue to use 1712af9997e4SRandy Dunlap * @dwork: work to queue 17130fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 17140fcb78c2SRolf Eike Beer * 1715d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1716715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1717715f1300STejun Heo * execution. 17180fcb78c2SRolf Eike Beer */ 1719d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 172052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 17217a6bc1cdSVenkatesh Pallipadi { 172252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1723d4283e93STejun Heo bool ret = false; 17248930cabaSTejun Heo unsigned long flags; 17258930cabaSTejun Heo 17268930cabaSTejun Heo /* read the comment in __queue_work() */ 17278930cabaSTejun Heo local_irq_save(flags); 17287a6bc1cdSVenkatesh Pallipadi 172922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17307beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1731d4283e93STejun Heo ret = true; 17327a6bc1cdSVenkatesh Pallipadi } 17338930cabaSTejun Heo 17348930cabaSTejun Heo local_irq_restore(flags); 17357a6bc1cdSVenkatesh Pallipadi return ret; 17367a6bc1cdSVenkatesh Pallipadi } 1737ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 17381da177e4SLinus Torvalds 1739c8e55f36STejun Heo /** 17408376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 17418376fe22STejun Heo * @cpu: CPU number to execute work on 17428376fe22STejun Heo * @wq: workqueue to use 17438376fe22STejun Heo * @dwork: work to queue 17448376fe22STejun Heo * @delay: number of jiffies to wait before queueing 17458376fe22STejun Heo * 17468376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17478376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17488376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17498376fe22STejun Heo * current state. 17508376fe22STejun Heo * 1751d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17528376fe22STejun Heo * pending and its timer was modified. 17538376fe22STejun Heo * 1754e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17558376fe22STejun Heo * See try_to_grab_pending() for details. 17568376fe22STejun Heo */ 17578376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17588376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17598376fe22STejun Heo { 17608376fe22STejun Heo unsigned long flags; 17618376fe22STejun Heo int ret; 17628376fe22STejun Heo 17638376fe22STejun Heo do { 17648376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17658376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17668376fe22STejun Heo 17678376fe22STejun Heo if (likely(ret >= 0)) { 17688376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17698376fe22STejun Heo local_irq_restore(flags); 17708376fe22STejun Heo } 17718376fe22STejun Heo 17728376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17738376fe22STejun Heo return ret; 17748376fe22STejun Heo } 17758376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17768376fe22STejun Heo 177705f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 177805f0fe6bSTejun Heo { 177905f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 178005f0fe6bSTejun Heo 178105f0fe6bSTejun Heo /* read the comment in __queue_work() */ 178205f0fe6bSTejun Heo local_irq_disable(); 178305f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 178405f0fe6bSTejun Heo local_irq_enable(); 178505f0fe6bSTejun Heo } 178605f0fe6bSTejun Heo 178705f0fe6bSTejun Heo /** 178805f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 178905f0fe6bSTejun Heo * @wq: workqueue to use 179005f0fe6bSTejun Heo * @rwork: work to queue 179105f0fe6bSTejun Heo * 179205f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 179305f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1794bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 179505f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 179605f0fe6bSTejun Heo */ 179705f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 179805f0fe6bSTejun Heo { 179905f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 180005f0fe6bSTejun Heo 180105f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 180205f0fe6bSTejun Heo rwork->wq = wq; 1803a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 180405f0fe6bSTejun Heo return true; 180505f0fe6bSTejun Heo } 180605f0fe6bSTejun Heo 180705f0fe6bSTejun Heo return false; 180805f0fe6bSTejun Heo } 180905f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 181005f0fe6bSTejun Heo 18118376fe22STejun Heo /** 1812c8e55f36STejun Heo * worker_enter_idle - enter idle state 1813c8e55f36STejun Heo * @worker: worker which is entering idle state 1814c8e55f36STejun Heo * 1815c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1816c8e55f36STejun Heo * necessary. 1817c8e55f36STejun Heo * 1818c8e55f36STejun Heo * LOCKING: 1819a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1820c8e55f36STejun Heo */ 1821c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 18221da177e4SLinus Torvalds { 1823bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1824c8e55f36STejun Heo 18256183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 18266183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 18276183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 18286183c009STejun Heo return; 1829c8e55f36STejun Heo 1830051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1831cb444766STejun Heo worker->flags |= WORKER_IDLE; 1832bd7bdd43STejun Heo pool->nr_idle++; 1833e22bee78STejun Heo worker->last_active = jiffies; 1834c8e55f36STejun Heo 1835c8e55f36STejun Heo /* idle_list is LIFO */ 1836bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1837db7bccf4STejun Heo 183863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1839628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1840cb444766STejun Heo 1841989442d7SLai Jiangshan /* Sanity check nr_running. */ 1842bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 1843c8e55f36STejun Heo } 1844c8e55f36STejun Heo 1845c8e55f36STejun Heo /** 1846c8e55f36STejun Heo * worker_leave_idle - leave idle state 1847c8e55f36STejun Heo * @worker: worker which is leaving idle state 1848c8e55f36STejun Heo * 1849c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1850c8e55f36STejun Heo * 1851c8e55f36STejun Heo * LOCKING: 1852a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1853c8e55f36STejun Heo */ 1854c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1855c8e55f36STejun Heo { 1856bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1857c8e55f36STejun Heo 18586183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18596183c009STejun Heo return; 1860d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1861bd7bdd43STejun Heo pool->nr_idle--; 1862c8e55f36STejun Heo list_del_init(&worker->entry); 1863c8e55f36STejun Heo } 1864c8e55f36STejun Heo 1865f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1866c34056a3STejun Heo { 1867c34056a3STejun Heo struct worker *worker; 1868c34056a3STejun Heo 1869f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1870c8e55f36STejun Heo if (worker) { 1871c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1872affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1873da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1874e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1875e22bee78STejun Heo worker->flags = WORKER_PREP; 1876c8e55f36STejun Heo } 1877c34056a3STejun Heo return worker; 1878c34056a3STejun Heo } 1879c34056a3STejun Heo 1880c34056a3STejun Heo /** 18814736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18824736cbf7SLai Jiangshan * @worker: worker to be attached 18834736cbf7SLai Jiangshan * @pool: the target pool 18844736cbf7SLai Jiangshan * 18854736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18864736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18874736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18884736cbf7SLai Jiangshan */ 18894736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18904736cbf7SLai Jiangshan struct worker_pool *pool) 18914736cbf7SLai Jiangshan { 18921258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18934736cbf7SLai Jiangshan 18944736cbf7SLai Jiangshan /* 18951258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18961258fae7STejun Heo * stable across this function. See the comments above the flag 18971258fae7STejun Heo * definition for details. 18984736cbf7SLai Jiangshan */ 18994736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 19004736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 19015c25b5ffSPeter Zijlstra else 19025c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 19034736cbf7SLai Jiangshan 1904640f17c8SPeter Zijlstra if (worker->rescue_wq) 1905640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1906640f17c8SPeter Zijlstra 19074736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1908a2d812a2STejun Heo worker->pool = pool; 19094736cbf7SLai Jiangshan 19101258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 19114736cbf7SLai Jiangshan } 19124736cbf7SLai Jiangshan 19134736cbf7SLai Jiangshan /** 191460f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 191560f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 191660f5a4bcSLai Jiangshan * 19174736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 19184736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 19194736cbf7SLai Jiangshan * other reference to the pool. 192060f5a4bcSLai Jiangshan */ 1921a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 192260f5a4bcSLai Jiangshan { 1923a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 192460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 192560f5a4bcSLai Jiangshan 19261258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1927a2d812a2STejun Heo 19285c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 1929da028469SLai Jiangshan list_del(&worker->node); 1930a2d812a2STejun Heo worker->pool = NULL; 1931a2d812a2STejun Heo 1932e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 193360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 19341258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 193560f5a4bcSLai Jiangshan 1936b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1937b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1938b62c0751SLai Jiangshan 193960f5a4bcSLai Jiangshan if (detach_completion) 194060f5a4bcSLai Jiangshan complete(detach_completion); 194160f5a4bcSLai Jiangshan } 194260f5a4bcSLai Jiangshan 194360f5a4bcSLai Jiangshan /** 1944c34056a3STejun Heo * create_worker - create a new workqueue worker 194563d95a91STejun Heo * @pool: pool the new worker will belong to 1946c34056a3STejun Heo * 1947051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1948c34056a3STejun Heo * 1949c34056a3STejun Heo * CONTEXT: 1950c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1951c34056a3STejun Heo * 1952d185af30SYacine Belkadi * Return: 1953c34056a3STejun Heo * Pointer to the newly created worker. 1954c34056a3STejun Heo */ 1955bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1956c34056a3STejun Heo { 1957e441b56fSZhen Lei struct worker *worker; 1958e441b56fSZhen Lei int id; 1959e3c916a4STejun Heo char id_buf[16]; 1960c34056a3STejun Heo 19617cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 1962e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 19633f0ea0b8SPetr Mladek if (id < 0) { 19643f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 19653f0ea0b8SPetr Mladek ERR_PTR(id)); 1966e441b56fSZhen Lei return NULL; 19673f0ea0b8SPetr Mladek } 1968c34056a3STejun Heo 1969f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 19703f0ea0b8SPetr Mladek if (!worker) { 19713f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 1972c34056a3STejun Heo goto fail; 19733f0ea0b8SPetr Mladek } 1974c34056a3STejun Heo 1975c34056a3STejun Heo worker->id = id; 1976c34056a3STejun Heo 197729c91e99STejun Heo if (pool->cpu >= 0) 1978e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1979e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1980f3421797STejun Heo else 1981e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1982e3c916a4STejun Heo 1983f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1984e3c916a4STejun Heo "kworker/%s", id_buf); 19853f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 198660f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 198760f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 198860f54038SPetr Mladek id_buf); 198960f54038SPetr Mladek } else { 19903f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 19913f0ea0b8SPetr Mladek worker->task); 199260f54038SPetr Mladek } 1993c34056a3STejun Heo goto fail; 19943f0ea0b8SPetr Mladek } 1995c34056a3STejun Heo 199691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 199725834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 199891151228SOleg Nesterov 1999da028469SLai Jiangshan /* successful, attach the worker to the pool */ 20004736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2001822d8405STejun Heo 2002051e1850SLai Jiangshan /* start the newly created worker */ 2003a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2004051e1850SLai Jiangshan worker->pool->nr_workers++; 2005051e1850SLai Jiangshan worker_enter_idle(worker); 2006051e1850SLai Jiangshan wake_up_process(worker->task); 2007a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2008051e1850SLai Jiangshan 2009c34056a3STejun Heo return worker; 2010822d8405STejun Heo 2011c34056a3STejun Heo fail: 2012e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2013c34056a3STejun Heo kfree(worker); 2014c34056a3STejun Heo return NULL; 2015c34056a3STejun Heo } 2016c34056a3STejun Heo 2017793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2018793777bcSValentin Schneider { 2019793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2020793777bcSValentin Schneider 2021793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2022793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2023793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2024793777bcSValentin Schneider else 2025793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2026793777bcSValentin Schneider } 2027793777bcSValentin Schneider 2028e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2029e02b9312SValentin Schneider { 2030e02b9312SValentin Schneider struct worker *worker, *tmp; 2031e02b9312SValentin Schneider 2032e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2033e02b9312SValentin Schneider list_del_init(&worker->entry); 2034e02b9312SValentin Schneider unbind_worker(worker); 2035e02b9312SValentin Schneider /* 2036e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2037e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2038e02b9312SValentin Schneider * wouldn't have gotten here. 2039c34056a3STejun Heo * 2040e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2041e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2042e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2043e02b9312SValentin Schneider * outside of pool->lock. 2044e02b9312SValentin Schneider */ 2045e02b9312SValentin Schneider wake_up_process(worker->task); 2046e02b9312SValentin Schneider } 2047e02b9312SValentin Schneider } 2048e02b9312SValentin Schneider 2049e02b9312SValentin Schneider /** 2050e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2051e02b9312SValentin Schneider * @worker: worker to be destroyed 2052e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2053e02b9312SValentin Schneider * 2054e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2055e02b9312SValentin Schneider * should be idle. 2056c8e55f36STejun Heo * 2057c8e55f36STejun Heo * CONTEXT: 2058a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2059c34056a3STejun Heo */ 2060e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2061c34056a3STejun Heo { 2062bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2063c34056a3STejun Heo 2064cd549687STejun Heo lockdep_assert_held(&pool->lock); 2065e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2066cd549687STejun Heo 2067c34056a3STejun Heo /* sanity check frenzy */ 20686183c009STejun Heo if (WARN_ON(worker->current_work) || 206973eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 207073eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 20716183c009STejun Heo return; 2072c34056a3STejun Heo 2073bd7bdd43STejun Heo pool->nr_workers--; 2074bd7bdd43STejun Heo pool->nr_idle--; 2075c8e55f36STejun Heo 2076cb444766STejun Heo worker->flags |= WORKER_DIE; 2077e02b9312SValentin Schneider 2078e02b9312SValentin Schneider list_move(&worker->entry, list); 2079e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2080c34056a3STejun Heo } 2081c34056a3STejun Heo 20823f959aa3SValentin Schneider /** 20833f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 20843f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 20853f959aa3SValentin Schneider * 20863f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 20873f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 20883f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 20893f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 20903f959aa3SValentin Schneider * it expire and re-evaluate things from there. 20913f959aa3SValentin Schneider */ 209232a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2093e22bee78STejun Heo { 209432a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 20953f959aa3SValentin Schneider bool do_cull = false; 20963f959aa3SValentin Schneider 20973f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 20983f959aa3SValentin Schneider return; 20993f959aa3SValentin Schneider 21003f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 21013f959aa3SValentin Schneider 21023f959aa3SValentin Schneider if (too_many_workers(pool)) { 21033f959aa3SValentin Schneider struct worker *worker; 21043f959aa3SValentin Schneider unsigned long expires; 21053f959aa3SValentin Schneider 21063f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 21073f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 21083f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 21093f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 21103f959aa3SValentin Schneider 21113f959aa3SValentin Schneider if (!do_cull) 21123f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 21133f959aa3SValentin Schneider } 21143f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 21153f959aa3SValentin Schneider 21163f959aa3SValentin Schneider if (do_cull) 21173f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 21183f959aa3SValentin Schneider } 21193f959aa3SValentin Schneider 21203f959aa3SValentin Schneider /** 21213f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 21223f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 21233f959aa3SValentin Schneider * 21243f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 21253f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2126e02b9312SValentin Schneider * 2127e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2128e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2129e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 21303f959aa3SValentin Schneider */ 21313f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 21323f959aa3SValentin Schneider { 21333f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2134e02b9312SValentin Schneider struct list_head cull_list; 2135e22bee78STejun Heo 2136e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 2137e02b9312SValentin Schneider /* 2138e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2139e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2140e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2141e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2142e02b9312SValentin Schneider */ 2143e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2144a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2145e22bee78STejun Heo 21463347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2147e22bee78STejun Heo struct worker *worker; 2148e22bee78STejun Heo unsigned long expires; 2149e22bee78STejun Heo 215063d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2151e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2152e22bee78STejun Heo 21533347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 215463d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 21553347fc9fSLai Jiangshan break; 2156e22bee78STejun Heo } 21573347fc9fSLai Jiangshan 2158e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2159e22bee78STejun Heo } 2160e22bee78STejun Heo 2161a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2162e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2163e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2164e22bee78STejun Heo } 2165e22bee78STejun Heo 2166493a1724STejun Heo static void send_mayday(struct work_struct *work) 2167e22bee78STejun Heo { 2168112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2169112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2170493a1724STejun Heo 21712e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2172e22bee78STejun Heo 2173493008a8STejun Heo if (!wq->rescuer) 2174493a1724STejun Heo return; 2175e22bee78STejun Heo 2176e22bee78STejun Heo /* mayday mayday mayday */ 2177493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 217877668c8bSLai Jiangshan /* 217977668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 218077668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 218177668c8bSLai Jiangshan * rescuer is done with it. 218277668c8bSLai Jiangshan */ 218377668c8bSLai Jiangshan get_pwq(pwq); 2184493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2185e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2186725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2187493a1724STejun Heo } 2188e22bee78STejun Heo } 2189e22bee78STejun Heo 219032a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2191e22bee78STejun Heo { 219232a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2193e22bee78STejun Heo struct work_struct *work; 2194e22bee78STejun Heo 2195a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2196a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2197e22bee78STejun Heo 219863d95a91STejun Heo if (need_to_create_worker(pool)) { 2199e22bee78STejun Heo /* 2200e22bee78STejun Heo * We've been trying to create a new worker but 2201e22bee78STejun Heo * haven't been successful. We might be hitting an 2202e22bee78STejun Heo * allocation deadlock. Send distress signals to 2203e22bee78STejun Heo * rescuers. 2204e22bee78STejun Heo */ 220563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2206e22bee78STejun Heo send_mayday(work); 2207e22bee78STejun Heo } 2208e22bee78STejun Heo 2209a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2210a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2211e22bee78STejun Heo 221263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2213e22bee78STejun Heo } 2214e22bee78STejun Heo 2215e22bee78STejun Heo /** 2216e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 221763d95a91STejun Heo * @pool: pool to create a new worker for 2218e22bee78STejun Heo * 221963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2220e22bee78STejun Heo * have at least one idle worker on return from this function. If 2221e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 222263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2223e22bee78STejun Heo * possible allocation deadlock. 2224e22bee78STejun Heo * 2225c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2226c5aa87bbSTejun Heo * may_start_working() %true. 2227e22bee78STejun Heo * 2228e22bee78STejun Heo * LOCKING: 2229a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2230e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2231e22bee78STejun Heo * manager. 2232e22bee78STejun Heo */ 223329187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2234d565ed63STejun Heo __releases(&pool->lock) 2235d565ed63STejun Heo __acquires(&pool->lock) 2236e22bee78STejun Heo { 2237e22bee78STejun Heo restart: 2238a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 22399f9c2364STejun Heo 2240e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 224163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2242e22bee78STejun Heo 2243e22bee78STejun Heo while (true) { 2244051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2245e22bee78STejun Heo break; 2246e22bee78STejun Heo 2247e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 22489f9c2364STejun Heo 224963d95a91STejun Heo if (!need_to_create_worker(pool)) 2250e22bee78STejun Heo break; 2251e22bee78STejun Heo } 2252e22bee78STejun Heo 225363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2254a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2255051e1850SLai Jiangshan /* 2256051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2257051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2258051e1850SLai Jiangshan * already become busy. 2259051e1850SLai Jiangshan */ 226063d95a91STejun Heo if (need_to_create_worker(pool)) 2261e22bee78STejun Heo goto restart; 2262e22bee78STejun Heo } 2263e22bee78STejun Heo 2264e22bee78STejun Heo /** 2265e22bee78STejun Heo * manage_workers - manage worker pool 2266e22bee78STejun Heo * @worker: self 2267e22bee78STejun Heo * 2268706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2269e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2270706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2271e22bee78STejun Heo * 2272e22bee78STejun Heo * The caller can safely start processing works on false return. On 2273e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2274e22bee78STejun Heo * and may_start_working() is true. 2275e22bee78STejun Heo * 2276e22bee78STejun Heo * CONTEXT: 2277a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2278e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2279e22bee78STejun Heo * 2280d185af30SYacine Belkadi * Return: 228129187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 228229187a9eSTejun Heo * start processing works, %true if management function was performed and 228329187a9eSTejun Heo * the conditions that the caller verified before calling the function may 228429187a9eSTejun Heo * no longer be true. 2285e22bee78STejun Heo */ 2286e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2287e22bee78STejun Heo { 228863d95a91STejun Heo struct worker_pool *pool = worker->pool; 2289e22bee78STejun Heo 2290692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 229129187a9eSTejun Heo return false; 2292692b4825STejun Heo 2293692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 22942607d7a6STejun Heo pool->manager = worker; 2295e22bee78STejun Heo 229629187a9eSTejun Heo maybe_create_worker(pool); 2297e22bee78STejun Heo 22982607d7a6STejun Heo pool->manager = NULL; 2299692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2300d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 230129187a9eSTejun Heo return true; 2302e22bee78STejun Heo } 2303e22bee78STejun Heo 2304a62428c0STejun Heo /** 2305a62428c0STejun Heo * process_one_work - process single work 2306c34056a3STejun Heo * @worker: self 2307a62428c0STejun Heo * @work: work to process 2308a62428c0STejun Heo * 2309a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2310a62428c0STejun Heo * process a single work including synchronization against and 2311a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2312a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2313a62428c0STejun Heo * call this function to process a work. 2314a62428c0STejun Heo * 2315a62428c0STejun Heo * CONTEXT: 2316a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2317a62428c0STejun Heo */ 2318c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2319d565ed63STejun Heo __releases(&pool->lock) 2320d565ed63STejun Heo __acquires(&pool->lock) 23211da177e4SLinus Torvalds { 2322112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2323bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2324112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2325c4560c2cSLai Jiangshan unsigned long work_data; 23267e11629dSTejun Heo struct worker *collision; 23274e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 23284e6045f1SJohannes Berg /* 2329a62428c0STejun Heo * It is permissible to free the struct work_struct from 2330a62428c0STejun Heo * inside the function that is called from it, this we need to 2331a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2332a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2333a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 23344e6045f1SJohannes Berg */ 23354d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 23364d82a1deSPeter Zijlstra 23374d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 23384e6045f1SJohannes Berg #endif 2339807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 234085327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2341ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 234225511a47STejun Heo 23437e11629dSTejun Heo /* 23447e11629dSTejun Heo * A single work shouldn't be executed concurrently by 23457e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 23467e11629dSTejun Heo * already processing the work. If so, defer the work to the 23477e11629dSTejun Heo * currently executing one. 23487e11629dSTejun Heo */ 2349c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 23507e11629dSTejun Heo if (unlikely(collision)) { 23517e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 23527e11629dSTejun Heo return; 23537e11629dSTejun Heo } 23541da177e4SLinus Torvalds 23558930cabaSTejun Heo /* claim and dequeue */ 23561da177e4SLinus Torvalds debug_work_deactivate(work); 2357c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2358c34056a3STejun Heo worker->current_work = work; 2359a2c1c57bSTejun Heo worker->current_func = work->func; 2360112202d9STejun Heo worker->current_pwq = pwq; 2361c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2362d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 23637a22ad75STejun Heo 23648bf89593STejun Heo /* 23658bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 23668bf89593STejun Heo * overridden through set_worker_desc(). 23678bf89593STejun Heo */ 23688bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 23698bf89593STejun Heo 2370a62428c0STejun Heo list_del_init(&work->entry); 2371a62428c0STejun Heo 2372649027d7STejun Heo /* 2373228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2374228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2375228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2376228f1d00SLai Jiangshan * execution of the pending work items. 2377fb0e7bebSTejun Heo */ 2378fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2379228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2380fb0e7bebSTejun Heo 2381974271c4STejun Heo /* 2382a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2383a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2384a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2385a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2386228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2387974271c4STejun Heo */ 2388a489a03eSLai Jiangshan if (need_more_worker(pool)) 238963d95a91STejun Heo wake_up_worker(pool); 2390974271c4STejun Heo 23918930cabaSTejun Heo /* 23927c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2393d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 239423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 239523657bb1STejun Heo * disabled. 23968930cabaSTejun Heo */ 23977c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 23981da177e4SLinus Torvalds 2399a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2400365970a1SDavid Howells 2401a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 24023295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2403e6f3faa7SPeter Zijlstra /* 2404f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2405f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2406e6f3faa7SPeter Zijlstra * 2407e6f3faa7SPeter Zijlstra * However, that would result in: 2408e6f3faa7SPeter Zijlstra * 2409e6f3faa7SPeter Zijlstra * A(W1) 2410e6f3faa7SPeter Zijlstra * WFC(C) 2411e6f3faa7SPeter Zijlstra * A(W1) 2412e6f3faa7SPeter Zijlstra * C(C) 2413e6f3faa7SPeter Zijlstra * 2414e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2415e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2416e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2417f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2418e6f3faa7SPeter Zijlstra * these locks. 2419e6f3faa7SPeter Zijlstra * 2420e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2421e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2422e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2423e6f3faa7SPeter Zijlstra */ 2424f52be570SPeter Zijlstra lockdep_invariant_state(true); 2425725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2426e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2427a2c1c57bSTejun Heo worker->current_func(work); 2428e36c886aSArjan van de Ven /* 2429e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2430e36c886aSArjan van de Ven * point will only record its address. 2431e36c886aSArjan van de Ven */ 24321c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2433725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 24343295f0efSIngo Molnar lock_map_release(&lockdep_map); 2435112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 24361da177e4SLinus Torvalds 2437d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2438044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2439d75f773cSSakari Ailus " last function: %ps\n", 2440a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2441a2c1c57bSTejun Heo worker->current_func); 2442d5abe669SPeter Zijlstra debug_show_held_locks(current); 2443d5abe669SPeter Zijlstra dump_stack(); 2444d5abe669SPeter Zijlstra } 2445d5abe669SPeter Zijlstra 2446b22ce278STejun Heo /* 2447025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2448b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2449b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2450b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2451789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2452789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2453b22ce278STejun Heo */ 2454a7e6425eSPaul E. McKenney cond_resched(); 2455b22ce278STejun Heo 2456a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2457a62428c0STejun Heo 2458fb0e7bebSTejun Heo /* clear cpu intensive status */ 2459fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2460fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2461fb0e7bebSTejun Heo 24621b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 24631b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 24641b69ac6bSJohannes Weiner 2465a62428c0STejun Heo /* we're done with it, release */ 246642f8570fSSasha Levin hash_del(&worker->hentry); 2467c34056a3STejun Heo worker->current_work = NULL; 2468a2c1c57bSTejun Heo worker->current_func = NULL; 2469112202d9STejun Heo worker->current_pwq = NULL; 2470d812796eSLai Jiangshan worker->current_color = INT_MAX; 2471c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 24721da177e4SLinus Torvalds } 24731da177e4SLinus Torvalds 2474affee4b2STejun Heo /** 2475affee4b2STejun Heo * process_scheduled_works - process scheduled works 2476affee4b2STejun Heo * @worker: self 2477affee4b2STejun Heo * 2478affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2479affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2480affee4b2STejun Heo * fetches a work from the top and executes it. 2481affee4b2STejun Heo * 2482affee4b2STejun Heo * CONTEXT: 2483a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2484affee4b2STejun Heo * multiple times. 2485affee4b2STejun Heo */ 2486affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 24871da177e4SLinus Torvalds { 2488affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2489affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2490a62428c0STejun Heo struct work_struct, entry); 2491c34056a3STejun Heo process_one_work(worker, work); 2492a62428c0STejun Heo } 24931da177e4SLinus Torvalds } 24941da177e4SLinus Torvalds 2495197f6accSTejun Heo static void set_pf_worker(bool val) 2496197f6accSTejun Heo { 2497197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2498197f6accSTejun Heo if (val) 2499197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2500197f6accSTejun Heo else 2501197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2502197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2503197f6accSTejun Heo } 2504197f6accSTejun Heo 25054690c4abSTejun Heo /** 25064690c4abSTejun Heo * worker_thread - the worker thread function 2507c34056a3STejun Heo * @__worker: self 25084690c4abSTejun Heo * 2509c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2510c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2511c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2512c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2513c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2514d185af30SYacine Belkadi * 2515d185af30SYacine Belkadi * Return: 0 25164690c4abSTejun Heo */ 2517c34056a3STejun Heo static int worker_thread(void *__worker) 25181da177e4SLinus Torvalds { 2519c34056a3STejun Heo struct worker *worker = __worker; 2520bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 25211da177e4SLinus Torvalds 2522e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2523197f6accSTejun Heo set_pf_worker(true); 2524c8e55f36STejun Heo woke_up: 2525a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2526affee4b2STejun Heo 2527a9ab775bSTejun Heo /* am I supposed to die? */ 2528a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2529a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2530197f6accSTejun Heo set_pf_worker(false); 253160f5a4bcSLai Jiangshan 253260f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2533e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2534a2d812a2STejun Heo worker_detach_from_pool(worker); 2535e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 253660f5a4bcSLai Jiangshan kfree(worker); 2537c8e55f36STejun Heo return 0; 2538c8e55f36STejun Heo } 2539c8e55f36STejun Heo 2540c8e55f36STejun Heo worker_leave_idle(worker); 2541db7bccf4STejun Heo recheck: 2542e22bee78STejun Heo /* no more worker necessary? */ 254363d95a91STejun Heo if (!need_more_worker(pool)) 2544e22bee78STejun Heo goto sleep; 2545e22bee78STejun Heo 2546e22bee78STejun Heo /* do we need to manage? */ 254763d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2548e22bee78STejun Heo goto recheck; 2549e22bee78STejun Heo 2550c8e55f36STejun Heo /* 2551c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2552c8e55f36STejun Heo * preparing to process a work or actually processing it. 2553c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2554c8e55f36STejun Heo */ 25556183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2556c8e55f36STejun Heo 2557e22bee78STejun Heo /* 2558a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2559a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2560a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2561a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2562a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2563e22bee78STejun Heo */ 2564a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2565e22bee78STejun Heo 2566e22bee78STejun Heo do { 2567affee4b2STejun Heo struct work_struct *work = 2568bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2569affee4b2STejun Heo struct work_struct, entry); 2570affee4b2STejun Heo 257182607adcSTejun Heo pool->watchdog_ts = jiffies; 257282607adcSTejun Heo 2573c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2574affee4b2STejun Heo /* optimization path, not strictly necessary */ 2575affee4b2STejun Heo process_one_work(worker, work); 2576affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2577affee4b2STejun Heo process_scheduled_works(worker); 2578affee4b2STejun Heo } else { 2579c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2580affee4b2STejun Heo process_scheduled_works(worker); 2581affee4b2STejun Heo } 258263d95a91STejun Heo } while (keep_working(pool)); 2583affee4b2STejun Heo 2584228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2585d313dd85STejun Heo sleep: 2586c8e55f36STejun Heo /* 2587d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2588d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2589d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2590d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2591d565ed63STejun Heo * event. 2592c8e55f36STejun Heo */ 2593c8e55f36STejun Heo worker_enter_idle(worker); 2594c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2595a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 25961da177e4SLinus Torvalds schedule(); 2597c8e55f36STejun Heo goto woke_up; 25981da177e4SLinus Torvalds } 25991da177e4SLinus Torvalds 2600e22bee78STejun Heo /** 2601e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2602111c225aSTejun Heo * @__rescuer: self 2603e22bee78STejun Heo * 2604e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2605493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2606e22bee78STejun Heo * 2607706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2608e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2609e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2610e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2611e22bee78STejun Heo * the problem rescuer solves. 2612e22bee78STejun Heo * 2613706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2614706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2615e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2616e22bee78STejun Heo * 2617e22bee78STejun Heo * This should happen rarely. 2618d185af30SYacine Belkadi * 2619d185af30SYacine Belkadi * Return: 0 2620e22bee78STejun Heo */ 2621111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2622e22bee78STejun Heo { 2623111c225aSTejun Heo struct worker *rescuer = __rescuer; 2624111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2625e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 26264d595b86SLai Jiangshan bool should_stop; 2627e22bee78STejun Heo 2628e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2629111c225aSTejun Heo 2630111c225aSTejun Heo /* 2631111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2632111c225aSTejun Heo * doesn't participate in concurrency management. 2633111c225aSTejun Heo */ 2634197f6accSTejun Heo set_pf_worker(true); 2635e22bee78STejun Heo repeat: 2636c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 26371da177e4SLinus Torvalds 26384d595b86SLai Jiangshan /* 26394d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 26404d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 26414d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 26424d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 26434d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 26444d595b86SLai Jiangshan * list is always empty on exit. 26454d595b86SLai Jiangshan */ 26464d595b86SLai Jiangshan should_stop = kthread_should_stop(); 26471da177e4SLinus Torvalds 2648493a1724STejun Heo /* see whether any pwq is asking for help */ 2649a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2650493a1724STejun Heo 2651493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2652493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2653493a1724STejun Heo struct pool_workqueue, mayday_node); 2654112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2655e22bee78STejun Heo struct work_struct *work, *n; 265682607adcSTejun Heo bool first = true; 2657e22bee78STejun Heo 2658e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2659493a1724STejun Heo list_del_init(&pwq->mayday_node); 2660493a1724STejun Heo 2661a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2662e22bee78STejun Heo 266351697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 266451697d39SLai Jiangshan 2665a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2666e22bee78STejun Heo 2667e22bee78STejun Heo /* 2668e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2669e22bee78STejun Heo * process'em. 2670e22bee78STejun Heo */ 26710479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 267282607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 267382607adcSTejun Heo if (get_work_pwq(work) == pwq) { 267482607adcSTejun Heo if (first) 267582607adcSTejun Heo pool->watchdog_ts = jiffies; 2676e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2677725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 267882607adcSTejun Heo } 267982607adcSTejun Heo first = false; 268082607adcSTejun Heo } 2681e22bee78STejun Heo 2682008847f6SNeilBrown if (!list_empty(scheduled)) { 2683e22bee78STejun Heo process_scheduled_works(rescuer); 26847576958aSTejun Heo 26857576958aSTejun Heo /* 2686008847f6SNeilBrown * The above execution of rescued work items could 2687008847f6SNeilBrown * have created more to rescue through 2688f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2689008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2690008847f6SNeilBrown * that such back-to-back work items, which may be 2691008847f6SNeilBrown * being used to relieve memory pressure, don't 2692008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2693008847f6SNeilBrown */ 26944f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2695a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2696e66b39afSTejun Heo /* 2697e66b39afSTejun Heo * Queue iff we aren't racing destruction 2698e66b39afSTejun Heo * and somebody else hasn't queued it already. 2699e66b39afSTejun Heo */ 2700e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2701008847f6SNeilBrown get_pwq(pwq); 2702e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2703e66b39afSTejun Heo } 2704a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2705008847f6SNeilBrown } 2706008847f6SNeilBrown } 2707008847f6SNeilBrown 2708008847f6SNeilBrown /* 270977668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 271013b1d625SLai Jiangshan * go away while we're still attached to it. 271177668c8bSLai Jiangshan */ 271277668c8bSLai Jiangshan put_pwq(pwq); 271377668c8bSLai Jiangshan 271477668c8bSLai Jiangshan /* 2715d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 27167576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 27177576958aSTejun Heo * and stalling the execution. 27187576958aSTejun Heo */ 2719d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 272063d95a91STejun Heo wake_up_worker(pool); 27217576958aSTejun Heo 2722a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 272313b1d625SLai Jiangshan 2724a2d812a2STejun Heo worker_detach_from_pool(rescuer); 272513b1d625SLai Jiangshan 2726a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 27271da177e4SLinus Torvalds } 27281da177e4SLinus Torvalds 2729a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2730493a1724STejun Heo 27314d595b86SLai Jiangshan if (should_stop) { 27324d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2733197f6accSTejun Heo set_pf_worker(false); 27344d595b86SLai Jiangshan return 0; 27354d595b86SLai Jiangshan } 27364d595b86SLai Jiangshan 2737111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2738111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2739e22bee78STejun Heo schedule(); 2740e22bee78STejun Heo goto repeat; 27411da177e4SLinus Torvalds } 27421da177e4SLinus Torvalds 2743fca839c0STejun Heo /** 2744fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2745fca839c0STejun Heo * @target_wq: workqueue being flushed 2746fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2747fca839c0STejun Heo * 2748fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2749fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2750fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2751fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2752fca839c0STejun Heo * a deadlock. 2753fca839c0STejun Heo */ 2754fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2755fca839c0STejun Heo struct work_struct *target_work) 2756fca839c0STejun Heo { 2757fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2758fca839c0STejun Heo struct worker *worker; 2759fca839c0STejun Heo 2760fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2761fca839c0STejun Heo return; 2762fca839c0STejun Heo 2763fca839c0STejun Heo worker = current_wq_worker(); 2764fca839c0STejun Heo 2765fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2766d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2767fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 276823d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 276923d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2770d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2771fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2772fca839c0STejun Heo target_wq->name, target_func); 2773fca839c0STejun Heo } 2774fca839c0STejun Heo 2775fc2e4d70SOleg Nesterov struct wq_barrier { 2776fc2e4d70SOleg Nesterov struct work_struct work; 2777fc2e4d70SOleg Nesterov struct completion done; 27782607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2779fc2e4d70SOleg Nesterov }; 2780fc2e4d70SOleg Nesterov 2781fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2782fc2e4d70SOleg Nesterov { 2783fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2784fc2e4d70SOleg Nesterov complete(&barr->done); 2785fc2e4d70SOleg Nesterov } 2786fc2e4d70SOleg Nesterov 27874690c4abSTejun Heo /** 27884690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2789112202d9STejun Heo * @pwq: pwq to insert barrier into 27904690c4abSTejun Heo * @barr: wq_barrier to insert 2791affee4b2STejun Heo * @target: target work to attach @barr to 2792affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 27934690c4abSTejun Heo * 2794affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2795affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2796affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2797affee4b2STejun Heo * cpu. 2798affee4b2STejun Heo * 2799affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2800affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2801affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2802affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2803affee4b2STejun Heo * after a work with LINKED flag set. 2804affee4b2STejun Heo * 2805affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2806112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 28074690c4abSTejun Heo * 28084690c4abSTejun Heo * CONTEXT: 2809a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 28104690c4abSTejun Heo */ 2811112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2812affee4b2STejun Heo struct wq_barrier *barr, 2813affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2814fc2e4d70SOleg Nesterov { 2815d812796eSLai Jiangshan unsigned int work_flags = 0; 2816d812796eSLai Jiangshan unsigned int work_color; 2817affee4b2STejun Heo struct list_head *head; 2818affee4b2STejun Heo 2819dc186ad7SThomas Gleixner /* 2820d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2821dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2822dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2823dc186ad7SThomas Gleixner * might deadlock. 2824dc186ad7SThomas Gleixner */ 2825ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 282622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 282752fa5bc5SBoqun Feng 2828fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2829fd1a5b04SByungchul Park 28302607d7a6STejun Heo barr->task = current; 283183c22520SOleg Nesterov 2832018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2833018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2834018f3a13SLai Jiangshan 2835affee4b2STejun Heo /* 2836affee4b2STejun Heo * If @target is currently being executed, schedule the 2837affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2838affee4b2STejun Heo */ 2839d812796eSLai Jiangshan if (worker) { 2840affee4b2STejun Heo head = worker->scheduled.next; 2841d812796eSLai Jiangshan work_color = worker->current_color; 2842d812796eSLai Jiangshan } else { 2843affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2844affee4b2STejun Heo 2845affee4b2STejun Heo head = target->entry.next; 2846affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2847d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 2848d812796eSLai Jiangshan work_color = get_work_color(*bits); 2849affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2850affee4b2STejun Heo } 2851affee4b2STejun Heo 2852d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 2853d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 2854d812796eSLai Jiangshan 2855dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2856d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 2857fc2e4d70SOleg Nesterov } 2858fc2e4d70SOleg Nesterov 285973f53c4aSTejun Heo /** 2860112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 286173f53c4aSTejun Heo * @wq: workqueue being flushed 286273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 286373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 286473f53c4aSTejun Heo * 2865112202d9STejun Heo * Prepare pwqs for workqueue flushing. 286673f53c4aSTejun Heo * 2867112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2868112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2869112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2870112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2871112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 287273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 287373f53c4aSTejun Heo * 287473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 287573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 287673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 287773f53c4aSTejun Heo * is returned. 287873f53c4aSTejun Heo * 2879112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 288073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 288173f53c4aSTejun Heo * advanced to @work_color. 288273f53c4aSTejun Heo * 288373f53c4aSTejun Heo * CONTEXT: 28843c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 288573f53c4aSTejun Heo * 2886d185af30SYacine Belkadi * Return: 288773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 288873f53c4aSTejun Heo * otherwise. 288973f53c4aSTejun Heo */ 2890112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 289173f53c4aSTejun Heo int flush_color, int work_color) 28921da177e4SLinus Torvalds { 289373f53c4aSTejun Heo bool wait = false; 289449e3cf44STejun Heo struct pool_workqueue *pwq; 28951da177e4SLinus Torvalds 289673f53c4aSTejun Heo if (flush_color >= 0) { 28976183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2898112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2899dc186ad7SThomas Gleixner } 290014441960SOleg Nesterov 290149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2902112202d9STejun Heo struct worker_pool *pool = pwq->pool; 29031da177e4SLinus Torvalds 2904a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 290573f53c4aSTejun Heo 290673f53c4aSTejun Heo if (flush_color >= 0) { 29076183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 290873f53c4aSTejun Heo 2909112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2910112202d9STejun Heo pwq->flush_color = flush_color; 2911112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 291273f53c4aSTejun Heo wait = true; 29131da177e4SLinus Torvalds } 291473f53c4aSTejun Heo } 291573f53c4aSTejun Heo 291673f53c4aSTejun Heo if (work_color >= 0) { 29176183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2918112202d9STejun Heo pwq->work_color = work_color; 291973f53c4aSTejun Heo } 292073f53c4aSTejun Heo 2921a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 29221da177e4SLinus Torvalds } 29231da177e4SLinus Torvalds 2924112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 292573f53c4aSTejun Heo complete(&wq->first_flusher->done); 292673f53c4aSTejun Heo 292773f53c4aSTejun Heo return wait; 292883c22520SOleg Nesterov } 29291da177e4SLinus Torvalds 29300fcb78c2SRolf Eike Beer /** 2931c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 29320fcb78c2SRolf Eike Beer * @wq: workqueue to flush 29331da177e4SLinus Torvalds * 2934c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2935c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 29361da177e4SLinus Torvalds */ 2937c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 29381da177e4SLinus Torvalds { 293973f53c4aSTejun Heo struct wq_flusher this_flusher = { 294073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 294173f53c4aSTejun Heo .flush_color = -1, 2942fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 294373f53c4aSTejun Heo }; 294473f53c4aSTejun Heo int next_color; 2945b1f4ec17SOleg Nesterov 29463347fa09STejun Heo if (WARN_ON(!wq_online)) 29473347fa09STejun Heo return; 29483347fa09STejun Heo 294987915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 295087915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 295187915adcSJohannes Berg 29523c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 295373f53c4aSTejun Heo 295473f53c4aSTejun Heo /* 295573f53c4aSTejun Heo * Start-to-wait phase 295673f53c4aSTejun Heo */ 295773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 295873f53c4aSTejun Heo 295973f53c4aSTejun Heo if (next_color != wq->flush_color) { 296073f53c4aSTejun Heo /* 296173f53c4aSTejun Heo * Color space is not full. The current work_color 296273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 296373f53c4aSTejun Heo * by one. 296473f53c4aSTejun Heo */ 29656183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 296673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 296773f53c4aSTejun Heo wq->work_color = next_color; 296873f53c4aSTejun Heo 296973f53c4aSTejun Heo if (!wq->first_flusher) { 297073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 29716183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 297273f53c4aSTejun Heo 297373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 297473f53c4aSTejun Heo 2975112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 297673f53c4aSTejun Heo wq->work_color)) { 297773f53c4aSTejun Heo /* nothing to flush, done */ 297873f53c4aSTejun Heo wq->flush_color = next_color; 297973f53c4aSTejun Heo wq->first_flusher = NULL; 298073f53c4aSTejun Heo goto out_unlock; 298173f53c4aSTejun Heo } 298273f53c4aSTejun Heo } else { 298373f53c4aSTejun Heo /* wait in queue */ 29846183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 298573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2986112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 298773f53c4aSTejun Heo } 298873f53c4aSTejun Heo } else { 298973f53c4aSTejun Heo /* 299073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 299173f53c4aSTejun Heo * The next flush completion will assign us 299273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 299373f53c4aSTejun Heo */ 299473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 299573f53c4aSTejun Heo } 299673f53c4aSTejun Heo 2997fca839c0STejun Heo check_flush_dependency(wq, NULL); 2998fca839c0STejun Heo 29993c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 300073f53c4aSTejun Heo 300173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 300273f53c4aSTejun Heo 300373f53c4aSTejun Heo /* 300473f53c4aSTejun Heo * Wake-up-and-cascade phase 300573f53c4aSTejun Heo * 300673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 300773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 300873f53c4aSTejun Heo */ 300900d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 301073f53c4aSTejun Heo return; 301173f53c4aSTejun Heo 30123c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 301373f53c4aSTejun Heo 30144ce48b37STejun Heo /* we might have raced, check again with mutex held */ 30154ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 30164ce48b37STejun Heo goto out_unlock; 30174ce48b37STejun Heo 301800d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 301973f53c4aSTejun Heo 30206183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 30216183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 302273f53c4aSTejun Heo 302373f53c4aSTejun Heo while (true) { 302473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 302573f53c4aSTejun Heo 302673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 302773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 302873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 302973f53c4aSTejun Heo break; 303073f53c4aSTejun Heo list_del_init(&next->list); 303173f53c4aSTejun Heo complete(&next->done); 303273f53c4aSTejun Heo } 303373f53c4aSTejun Heo 30346183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 303573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 303673f53c4aSTejun Heo 303773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 303873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 303973f53c4aSTejun Heo 304073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 304173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 304273f53c4aSTejun Heo /* 304373f53c4aSTejun Heo * Assign the same color to all overflowed 304473f53c4aSTejun Heo * flushers, advance work_color and append to 304573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 304673f53c4aSTejun Heo * phase for these overflowed flushers. 304773f53c4aSTejun Heo */ 304873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 304973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 305073f53c4aSTejun Heo 305173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 305273f53c4aSTejun Heo 305373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 305473f53c4aSTejun Heo &wq->flusher_queue); 3055112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 305673f53c4aSTejun Heo } 305773f53c4aSTejun Heo 305873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 30596183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 306073f53c4aSTejun Heo break; 306173f53c4aSTejun Heo } 306273f53c4aSTejun Heo 306373f53c4aSTejun Heo /* 306473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3065112202d9STejun Heo * the new first flusher and arm pwqs. 306673f53c4aSTejun Heo */ 30676183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 30686183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 306973f53c4aSTejun Heo 307073f53c4aSTejun Heo list_del_init(&next->list); 307173f53c4aSTejun Heo wq->first_flusher = next; 307273f53c4aSTejun Heo 3073112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 307473f53c4aSTejun Heo break; 307573f53c4aSTejun Heo 307673f53c4aSTejun Heo /* 307773f53c4aSTejun Heo * Meh... this color is already done, clear first 307873f53c4aSTejun Heo * flusher and repeat cascading. 307973f53c4aSTejun Heo */ 308073f53c4aSTejun Heo wq->first_flusher = NULL; 308173f53c4aSTejun Heo } 308273f53c4aSTejun Heo 308373f53c4aSTejun Heo out_unlock: 30843c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 30851da177e4SLinus Torvalds } 3086c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 30871da177e4SLinus Torvalds 30889c5a2ba7STejun Heo /** 30899c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 30909c5a2ba7STejun Heo * @wq: workqueue to drain 30919c5a2ba7STejun Heo * 30929c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 30939c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 30949c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3095b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 30969c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 30979c5a2ba7STejun Heo * takes too long. 30989c5a2ba7STejun Heo */ 30999c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 31009c5a2ba7STejun Heo { 31019c5a2ba7STejun Heo unsigned int flush_cnt = 0; 310249e3cf44STejun Heo struct pool_workqueue *pwq; 31039c5a2ba7STejun Heo 31049c5a2ba7STejun Heo /* 31059c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 31069c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3107618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 31089c5a2ba7STejun Heo */ 310987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 31109c5a2ba7STejun Heo if (!wq->nr_drainers++) 3111618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 311287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 31139c5a2ba7STejun Heo reflush: 3114c4f135d6STetsuo Handa __flush_workqueue(wq); 31159c5a2ba7STejun Heo 3116b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 311776af4d93STejun Heo 311849e3cf44STejun Heo for_each_pwq(pwq, wq) { 3119fa2563e4SThomas Tuttle bool drained; 31209c5a2ba7STejun Heo 3121a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3122f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3123a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3124fa2563e4SThomas Tuttle 3125fa2563e4SThomas Tuttle if (drained) 31269c5a2ba7STejun Heo continue; 31279c5a2ba7STejun Heo 31289c5a2ba7STejun Heo if (++flush_cnt == 10 || 31299c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3130e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3131e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 313276af4d93STejun Heo 3133b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 31349c5a2ba7STejun Heo goto reflush; 31359c5a2ba7STejun Heo } 31369c5a2ba7STejun Heo 31379c5a2ba7STejun Heo if (!--wq->nr_drainers) 3138618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 313987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 31409c5a2ba7STejun Heo } 31419c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 31429c5a2ba7STejun Heo 3143d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3144d6e89786SJohannes Berg bool from_cancel) 3145baf59022STejun Heo { 3146baf59022STejun Heo struct worker *worker = NULL; 3147c9e7cf27STejun Heo struct worker_pool *pool; 3148112202d9STejun Heo struct pool_workqueue *pwq; 3149baf59022STejun Heo 3150baf59022STejun Heo might_sleep(); 3151baf59022STejun Heo 315224acfb71SThomas Gleixner rcu_read_lock(); 3153fa1b54e6STejun Heo pool = get_work_pool(work); 3154fa1b54e6STejun Heo if (!pool) { 315524acfb71SThomas Gleixner rcu_read_unlock(); 3156fa1b54e6STejun Heo return false; 3157fa1b54e6STejun Heo } 3158fa1b54e6STejun Heo 3159a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 31600b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3161112202d9STejun Heo pwq = get_work_pwq(work); 3162112202d9STejun Heo if (pwq) { 3163112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3164baf59022STejun Heo goto already_gone; 3165606a5020STejun Heo } else { 3166c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3167baf59022STejun Heo if (!worker) 3168baf59022STejun Heo goto already_gone; 3169112202d9STejun Heo pwq = worker->current_pwq; 3170606a5020STejun Heo } 3171baf59022STejun Heo 3172fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3173fca839c0STejun Heo 3174112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3175a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3176baf59022STejun Heo 3177e159489bSTejun Heo /* 3178a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3179a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3180a1d14934SPeter Zijlstra * 3181a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3182a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3183a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3184a1d14934SPeter Zijlstra * forward progress. 3185e159489bSTejun Heo */ 3186d6e89786SJohannes Berg if (!from_cancel && 3187d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3188112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3189112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3190a1d14934SPeter Zijlstra } 319124acfb71SThomas Gleixner rcu_read_unlock(); 3192baf59022STejun Heo return true; 3193baf59022STejun Heo already_gone: 3194a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 319524acfb71SThomas Gleixner rcu_read_unlock(); 3196baf59022STejun Heo return false; 3197baf59022STejun Heo } 3198baf59022STejun Heo 3199d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3200d6e89786SJohannes Berg { 3201d6e89786SJohannes Berg struct wq_barrier barr; 3202d6e89786SJohannes Berg 3203d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3204d6e89786SJohannes Berg return false; 3205d6e89786SJohannes Berg 32064d43d395STetsuo Handa if (WARN_ON(!work->func)) 32074d43d395STetsuo Handa return false; 32084d43d395STetsuo Handa 320987915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 321087915adcSJohannes Berg lock_map_release(&work->lockdep_map); 321187915adcSJohannes Berg 3212d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3213d6e89786SJohannes Berg wait_for_completion(&barr.done); 3214d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3215d6e89786SJohannes Berg return true; 3216d6e89786SJohannes Berg } else { 3217d6e89786SJohannes Berg return false; 3218d6e89786SJohannes Berg } 3219d6e89786SJohannes Berg } 3220d6e89786SJohannes Berg 3221db700897SOleg Nesterov /** 3222401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3223401a8d04STejun Heo * @work: the work to flush 3224db700897SOleg Nesterov * 3225606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3226606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3227401a8d04STejun Heo * 3228d185af30SYacine Belkadi * Return: 3229401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3230401a8d04STejun Heo * %false if it was already idle. 3231db700897SOleg Nesterov */ 3232401a8d04STejun Heo bool flush_work(struct work_struct *work) 3233db700897SOleg Nesterov { 3234d6e89786SJohannes Berg return __flush_work(work, false); 3235606a5020STejun Heo } 3236db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3237db700897SOleg Nesterov 32388603e1b3STejun Heo struct cwt_wait { 3239ac6424b9SIngo Molnar wait_queue_entry_t wait; 32408603e1b3STejun Heo struct work_struct *work; 32418603e1b3STejun Heo }; 32428603e1b3STejun Heo 3243ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 32448603e1b3STejun Heo { 32458603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 32468603e1b3STejun Heo 32478603e1b3STejun Heo if (cwait->work != key) 32488603e1b3STejun Heo return 0; 32498603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 32508603e1b3STejun Heo } 32518603e1b3STejun Heo 325236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3253401a8d04STejun Heo { 32548603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3255bbb68dfaSTejun Heo unsigned long flags; 32561f1f642eSOleg Nesterov int ret; 32571f1f642eSOleg Nesterov 32581f1f642eSOleg Nesterov do { 3259bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3260bbb68dfaSTejun Heo /* 32618603e1b3STejun Heo * If someone else is already canceling, wait for it to 32628603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 32638603e1b3STejun Heo * because we may get scheduled between @work's completion 32648603e1b3STejun Heo * and the other canceling task resuming and clearing 32658603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 32668603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 32678603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 32688603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 32698603e1b3STejun Heo * we're hogging the CPU. 32708603e1b3STejun Heo * 32718603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 32728603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 32738603e1b3STejun Heo * wake function which matches @work along with exclusive 32748603e1b3STejun Heo * wait and wakeup. 3275bbb68dfaSTejun Heo */ 32768603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 32778603e1b3STejun Heo struct cwt_wait cwait; 32788603e1b3STejun Heo 32798603e1b3STejun Heo init_wait(&cwait.wait); 32808603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 32818603e1b3STejun Heo cwait.work = work; 32828603e1b3STejun Heo 32838603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 32848603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 32858603e1b3STejun Heo if (work_is_canceling(work)) 32868603e1b3STejun Heo schedule(); 32878603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 32888603e1b3STejun Heo } 32891f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 32901f1f642eSOleg Nesterov 3291bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3292bbb68dfaSTejun Heo mark_work_canceling(work); 3293bbb68dfaSTejun Heo local_irq_restore(flags); 3294bbb68dfaSTejun Heo 32953347fa09STejun Heo /* 32963347fa09STejun Heo * This allows canceling during early boot. We know that @work 32973347fa09STejun Heo * isn't executing. 32983347fa09STejun Heo */ 32993347fa09STejun Heo if (wq_online) 3300d6e89786SJohannes Berg __flush_work(work, true); 33013347fa09STejun Heo 33027a22ad75STejun Heo clear_work_data(work); 33038603e1b3STejun Heo 33048603e1b3STejun Heo /* 33058603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 33068603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 33078603e1b3STejun Heo * visible there. 33088603e1b3STejun Heo */ 33098603e1b3STejun Heo smp_mb(); 33108603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 33118603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 33128603e1b3STejun Heo 33131f1f642eSOleg Nesterov return ret; 33141f1f642eSOleg Nesterov } 33151f1f642eSOleg Nesterov 33166e84d644SOleg Nesterov /** 3317401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3318401a8d04STejun Heo * @work: the work to cancel 33196e84d644SOleg Nesterov * 3320401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3321401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3322401a8d04STejun Heo * another workqueue. On return from this function, @work is 3323401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 33241f1f642eSOleg Nesterov * 3325401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3326401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 33276e84d644SOleg Nesterov * 3328401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 33296e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3330401a8d04STejun Heo * 3331d185af30SYacine Belkadi * Return: 3332401a8d04STejun Heo * %true if @work was pending, %false otherwise. 33336e84d644SOleg Nesterov */ 3334401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 33356e84d644SOleg Nesterov { 333636e227d2STejun Heo return __cancel_work_timer(work, false); 3337b89deed3SOleg Nesterov } 333828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3339b89deed3SOleg Nesterov 33406e84d644SOleg Nesterov /** 3341401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3342401a8d04STejun Heo * @dwork: the delayed work to flush 33436e84d644SOleg Nesterov * 3344401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3345401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3346401a8d04STejun Heo * considers the last queueing instance of @dwork. 33471f1f642eSOleg Nesterov * 3348d185af30SYacine Belkadi * Return: 3349401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3350401a8d04STejun Heo * %false if it was already idle. 33516e84d644SOleg Nesterov */ 3352401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3353401a8d04STejun Heo { 33548930cabaSTejun Heo local_irq_disable(); 3355401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 335660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 33578930cabaSTejun Heo local_irq_enable(); 3358401a8d04STejun Heo return flush_work(&dwork->work); 3359401a8d04STejun Heo } 3360401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3361401a8d04STejun Heo 336205f0fe6bSTejun Heo /** 336305f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 336405f0fe6bSTejun Heo * @rwork: the rcu work to flush 336505f0fe6bSTejun Heo * 336605f0fe6bSTejun Heo * Return: 336705f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 336805f0fe6bSTejun Heo * %false if it was already idle. 336905f0fe6bSTejun Heo */ 337005f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 337105f0fe6bSTejun Heo { 337205f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 337305f0fe6bSTejun Heo rcu_barrier(); 337405f0fe6bSTejun Heo flush_work(&rwork->work); 337505f0fe6bSTejun Heo return true; 337605f0fe6bSTejun Heo } else { 337705f0fe6bSTejun Heo return flush_work(&rwork->work); 337805f0fe6bSTejun Heo } 337905f0fe6bSTejun Heo } 338005f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 338105f0fe6bSTejun Heo 3382f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3383f72b8792SJens Axboe { 3384f72b8792SJens Axboe unsigned long flags; 3385f72b8792SJens Axboe int ret; 3386f72b8792SJens Axboe 3387f72b8792SJens Axboe do { 3388f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3389f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3390f72b8792SJens Axboe 3391f72b8792SJens Axboe if (unlikely(ret < 0)) 3392f72b8792SJens Axboe return false; 3393f72b8792SJens Axboe 3394f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3395f72b8792SJens Axboe local_irq_restore(flags); 3396f72b8792SJens Axboe return ret; 3397f72b8792SJens Axboe } 3398f72b8792SJens Axboe 339973b4b532SAndrey Grodzovsky /* 340073b4b532SAndrey Grodzovsky * See cancel_delayed_work() 340173b4b532SAndrey Grodzovsky */ 340273b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 340373b4b532SAndrey Grodzovsky { 340473b4b532SAndrey Grodzovsky return __cancel_work(work, false); 340573b4b532SAndrey Grodzovsky } 340673b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 340773b4b532SAndrey Grodzovsky 3408401a8d04STejun Heo /** 340957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 341057b30ae7STejun Heo * @dwork: delayed_work to cancel 341109383498STejun Heo * 3412d185af30SYacine Belkadi * Kill off a pending delayed_work. 3413d185af30SYacine Belkadi * 3414d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3415d185af30SYacine Belkadi * pending. 3416d185af30SYacine Belkadi * 3417d185af30SYacine Belkadi * Note: 3418d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3419d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3420d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 342109383498STejun Heo * 342257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 342309383498STejun Heo */ 342457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 342509383498STejun Heo { 3426f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 342709383498STejun Heo } 342857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 342909383498STejun Heo 343009383498STejun Heo /** 3431401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3432401a8d04STejun Heo * @dwork: the delayed work cancel 3433401a8d04STejun Heo * 3434401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3435401a8d04STejun Heo * 3436d185af30SYacine Belkadi * Return: 3437401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3438401a8d04STejun Heo */ 3439401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 34406e84d644SOleg Nesterov { 344136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 34426e84d644SOleg Nesterov } 3443f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 34441da177e4SLinus Torvalds 34450fcb78c2SRolf Eike Beer /** 344631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3447b6136773SAndrew Morton * @func: the function to call 3448b6136773SAndrew Morton * 344931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 345031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3451b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 345231ddd871STejun Heo * 3453d185af30SYacine Belkadi * Return: 345431ddd871STejun Heo * 0 on success, -errno on failure. 3455b6136773SAndrew Morton */ 345665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 345715316ba8SChristoph Lameter { 345815316ba8SChristoph Lameter int cpu; 345938f51568SNamhyung Kim struct work_struct __percpu *works; 346015316ba8SChristoph Lameter 3461b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3462b6136773SAndrew Morton if (!works) 346315316ba8SChristoph Lameter return -ENOMEM; 3464b6136773SAndrew Morton 3465ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 346693981800STejun Heo 346715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 34689bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 34699bfb1839SIngo Molnar 34709bfb1839SIngo Molnar INIT_WORK(work, func); 34718de6d308SOleg Nesterov schedule_work_on(cpu, work); 347215316ba8SChristoph Lameter } 347393981800STejun Heo 347493981800STejun Heo for_each_online_cpu(cpu) 34758616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 347693981800STejun Heo 3477ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3478b6136773SAndrew Morton free_percpu(works); 347915316ba8SChristoph Lameter return 0; 348015316ba8SChristoph Lameter } 348115316ba8SChristoph Lameter 3482eef6a7d5SAlan Stern /** 34831fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 34841fa44ecaSJames Bottomley * @fn: the function to execute 34851fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 34861fa44ecaSJames Bottomley * be available when the work executes) 34871fa44ecaSJames Bottomley * 34881fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 34891fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 34901fa44ecaSJames Bottomley * 3491d185af30SYacine Belkadi * Return: 0 - function was executed 34921fa44ecaSJames Bottomley * 1 - function was scheduled for execution 34931fa44ecaSJames Bottomley */ 349465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 34951fa44ecaSJames Bottomley { 34961fa44ecaSJames Bottomley if (!in_interrupt()) { 349765f27f38SDavid Howells fn(&ew->work); 34981fa44ecaSJames Bottomley return 0; 34991fa44ecaSJames Bottomley } 35001fa44ecaSJames Bottomley 350165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 35021fa44ecaSJames Bottomley schedule_work(&ew->work); 35031fa44ecaSJames Bottomley 35041fa44ecaSJames Bottomley return 1; 35051fa44ecaSJames Bottomley } 35061fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 35071fa44ecaSJames Bottomley 35087a4e344cSTejun Heo /** 35097a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 35107a4e344cSTejun Heo * @attrs: workqueue_attrs to free 35117a4e344cSTejun Heo * 35127a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 35137a4e344cSTejun Heo */ 3514513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 35157a4e344cSTejun Heo { 35167a4e344cSTejun Heo if (attrs) { 35177a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 35187a4e344cSTejun Heo kfree(attrs); 35197a4e344cSTejun Heo } 35207a4e344cSTejun Heo } 35217a4e344cSTejun Heo 35227a4e344cSTejun Heo /** 35237a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 35247a4e344cSTejun Heo * 35257a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3526d185af30SYacine Belkadi * return it. 3527d185af30SYacine Belkadi * 3528d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 35297a4e344cSTejun Heo */ 3530513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 35317a4e344cSTejun Heo { 35327a4e344cSTejun Heo struct workqueue_attrs *attrs; 35337a4e344cSTejun Heo 3534be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 35357a4e344cSTejun Heo if (!attrs) 35367a4e344cSTejun Heo goto fail; 3537be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 35387a4e344cSTejun Heo goto fail; 35397a4e344cSTejun Heo 354013e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 35417a4e344cSTejun Heo return attrs; 35427a4e344cSTejun Heo fail: 35437a4e344cSTejun Heo free_workqueue_attrs(attrs); 35447a4e344cSTejun Heo return NULL; 35457a4e344cSTejun Heo } 35467a4e344cSTejun Heo 354729c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 354829c91e99STejun Heo const struct workqueue_attrs *from) 354929c91e99STejun Heo { 355029c91e99STejun Heo to->nice = from->nice; 355129c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 35522865a8fbSShaohua Li /* 35532865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 35542865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 35552865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 35562865a8fbSShaohua Li */ 35572865a8fbSShaohua Li to->no_numa = from->no_numa; 355829c91e99STejun Heo } 355929c91e99STejun Heo 356029c91e99STejun Heo /* hash value of the content of @attr */ 356129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 356229c91e99STejun Heo { 356329c91e99STejun Heo u32 hash = 0; 356429c91e99STejun Heo 356529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 356613e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 356713e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 356829c91e99STejun Heo return hash; 356929c91e99STejun Heo } 357029c91e99STejun Heo 357129c91e99STejun Heo /* content equality test */ 357229c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 357329c91e99STejun Heo const struct workqueue_attrs *b) 357429c91e99STejun Heo { 357529c91e99STejun Heo if (a->nice != b->nice) 357629c91e99STejun Heo return false; 357729c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 357829c91e99STejun Heo return false; 357929c91e99STejun Heo return true; 358029c91e99STejun Heo } 358129c91e99STejun Heo 35827a4e344cSTejun Heo /** 35837a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 35847a4e344cSTejun Heo * @pool: worker_pool to initialize 35857a4e344cSTejun Heo * 3586402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3587d185af30SYacine Belkadi * 3588d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 358929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 359029c91e99STejun Heo * on @pool safely to release it. 35917a4e344cSTejun Heo */ 35927a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 35934e1a1f9aSTejun Heo { 3594a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 359529c91e99STejun Heo pool->id = -1; 359629c91e99STejun Heo pool->cpu = -1; 3597f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 35984e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 359982607adcSTejun Heo pool->watchdog_ts = jiffies; 36004e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 36014e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 36024e1a1f9aSTejun Heo hash_init(pool->busy_hash); 36034e1a1f9aSTejun Heo 360432a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 36053f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 36064e1a1f9aSTejun Heo 360732a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 36084e1a1f9aSTejun Heo 3609da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3610e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 36117a4e344cSTejun Heo 36127cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 361329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 361429c91e99STejun Heo pool->refcnt = 1; 361529c91e99STejun Heo 361629c91e99STejun Heo /* shouldn't fail above this point */ 3617be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 36187a4e344cSTejun Heo if (!pool->attrs) 36197a4e344cSTejun Heo return -ENOMEM; 36207a4e344cSTejun Heo return 0; 36214e1a1f9aSTejun Heo } 36224e1a1f9aSTejun Heo 3623669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3624669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3625669de8bdSBart Van Assche { 3626669de8bdSBart Van Assche char *lock_name; 3627669de8bdSBart Van Assche 3628669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3629669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3630669de8bdSBart Van Assche if (!lock_name) 3631669de8bdSBart Van Assche lock_name = wq->name; 363269a106c0SQian Cai 363369a106c0SQian Cai wq->lock_name = lock_name; 3634669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3635669de8bdSBart Van Assche } 3636669de8bdSBart Van Assche 3637669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3638669de8bdSBart Van Assche { 3639669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3640669de8bdSBart Van Assche } 3641669de8bdSBart Van Assche 3642669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3643669de8bdSBart Van Assche { 3644669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3645669de8bdSBart Van Assche kfree(wq->lock_name); 3646669de8bdSBart Van Assche } 3647669de8bdSBart Van Assche #else 3648669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3649669de8bdSBart Van Assche { 3650669de8bdSBart Van Assche } 3651669de8bdSBart Van Assche 3652669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3653669de8bdSBart Van Assche { 3654669de8bdSBart Van Assche } 3655669de8bdSBart Van Assche 3656669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3657669de8bdSBart Van Assche { 3658669de8bdSBart Van Assche } 3659669de8bdSBart Van Assche #endif 3660669de8bdSBart Van Assche 3661e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3662e2dca7adSTejun Heo { 3663e2dca7adSTejun Heo struct workqueue_struct *wq = 3664e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3665e2dca7adSTejun Heo 3666669de8bdSBart Van Assche wq_free_lockdep(wq); 3667669de8bdSBart Van Assche 3668e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3669e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3670e2dca7adSTejun Heo else 3671e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3672e2dca7adSTejun Heo 3673e2dca7adSTejun Heo kfree(wq); 3674e2dca7adSTejun Heo } 3675e2dca7adSTejun Heo 367629c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 367729c91e99STejun Heo { 367829c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 367929c91e99STejun Heo 36807cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 368129c91e99STejun Heo free_workqueue_attrs(pool->attrs); 368229c91e99STejun Heo kfree(pool); 368329c91e99STejun Heo } 368429c91e99STejun Heo 368529c91e99STejun Heo /** 368629c91e99STejun Heo * put_unbound_pool - put a worker_pool 368729c91e99STejun Heo * @pool: worker_pool to put 368829c91e99STejun Heo * 368924acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3690c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3691c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3692c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3693a892caccSTejun Heo * 3694a892caccSTejun Heo * Should be called with wq_pool_mutex held. 369529c91e99STejun Heo */ 369629c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 369729c91e99STejun Heo { 369860f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 3699e02b9312SValentin Schneider struct list_head cull_list; 370029c91e99STejun Heo struct worker *worker; 370129c91e99STejun Heo 3702e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 3703e02b9312SValentin Schneider 3704a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3705a892caccSTejun Heo 3706a892caccSTejun Heo if (--pool->refcnt) 370729c91e99STejun Heo return; 370829c91e99STejun Heo 370929c91e99STejun Heo /* sanity checks */ 371061d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3711a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 371229c91e99STejun Heo return; 371329c91e99STejun Heo 371429c91e99STejun Heo /* release id and unhash */ 371529c91e99STejun Heo if (pool->id >= 0) 371629c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 371729c91e99STejun Heo hash_del(&pool->hash_node); 371829c91e99STejun Heo 3719c5aa87bbSTejun Heo /* 3720692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3721692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3722692b4825STejun Heo * manager and @pool gets freed with the flag set. 37239ab03be4SValentin Schneider * 37249ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 37259ab03be4SValentin Schneider * only get here with 37269ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 37279ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 37289ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 37299ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 37309ab03be4SValentin Schneider * drops pool->lock 3731c5aa87bbSTejun Heo */ 37329ab03be4SValentin Schneider while (true) { 37339ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 37349ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3735d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3736e02b9312SValentin Schneider 3737e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 37389ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 37399ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3740692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 37419ab03be4SValentin Schneider break; 37429ab03be4SValentin Schneider } 37439ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3744e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 37459ab03be4SValentin Schneider } 3746692b4825STejun Heo 37471037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3748e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 374929c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3750a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 375160f5a4bcSLai Jiangshan 3752e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3753e02b9312SValentin Schneider 3754e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 375560f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 37561258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 375760f5a4bcSLai Jiangshan 375860f5a4bcSLai Jiangshan if (pool->detach_completion) 375960f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 376060f5a4bcSLai Jiangshan 376129c91e99STejun Heo /* shut down the timers */ 376229c91e99STejun Heo del_timer_sync(&pool->idle_timer); 37633f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 376429c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 376529c91e99STejun Heo 376624acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 376725b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 376829c91e99STejun Heo } 376929c91e99STejun Heo 377029c91e99STejun Heo /** 377129c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 377229c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 377329c91e99STejun Heo * 377429c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 377529c91e99STejun Heo * reference count and return it. If there already is a matching 377629c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3777d185af30SYacine Belkadi * create a new one. 3778a892caccSTejun Heo * 3779a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3780d185af30SYacine Belkadi * 3781d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3782d185af30SYacine Belkadi * On failure, %NULL. 378329c91e99STejun Heo */ 378429c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 378529c91e99STejun Heo { 378629c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 378729c91e99STejun Heo struct worker_pool *pool; 3788f3f90ad4STejun Heo int node; 3789e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 379029c91e99STejun Heo 3791a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 379229c91e99STejun Heo 379329c91e99STejun Heo /* do we already have a matching pool? */ 379429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 379529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 379629c91e99STejun Heo pool->refcnt++; 37973fb1823cSLai Jiangshan return pool; 379829c91e99STejun Heo } 379929c91e99STejun Heo } 380029c91e99STejun Heo 3801e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3802e2273584SXunlei Pang if (wq_numa_enabled) { 3803e2273584SXunlei Pang for_each_node(node) { 3804e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3805e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3806e2273584SXunlei Pang target_node = node; 3807e2273584SXunlei Pang break; 3808e2273584SXunlei Pang } 3809e2273584SXunlei Pang } 3810e2273584SXunlei Pang } 3811e2273584SXunlei Pang 381229c91e99STejun Heo /* nope, create a new one */ 3813e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 381429c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 381529c91e99STejun Heo goto fail; 381629c91e99STejun Heo 38178864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 381829c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3819e2273584SXunlei Pang pool->node = target_node; 382029c91e99STejun Heo 38212865a8fbSShaohua Li /* 38222865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 38232865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 38242865a8fbSShaohua Li */ 38252865a8fbSShaohua Li pool->attrs->no_numa = false; 38262865a8fbSShaohua Li 382729c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 382829c91e99STejun Heo goto fail; 382929c91e99STejun Heo 383029c91e99STejun Heo /* create and start the initial worker */ 38313347fa09STejun Heo if (wq_online && !create_worker(pool)) 383229c91e99STejun Heo goto fail; 383329c91e99STejun Heo 383429c91e99STejun Heo /* install */ 383529c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 38363fb1823cSLai Jiangshan 383729c91e99STejun Heo return pool; 383829c91e99STejun Heo fail: 383929c91e99STejun Heo if (pool) 384029c91e99STejun Heo put_unbound_pool(pool); 384129c91e99STejun Heo return NULL; 384229c91e99STejun Heo } 384329c91e99STejun Heo 38448864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 38458864b4e5STejun Heo { 38468864b4e5STejun Heo kmem_cache_free(pwq_cache, 38478864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 38488864b4e5STejun Heo } 38498864b4e5STejun Heo 38508864b4e5STejun Heo /* 38518864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 38528864b4e5STejun Heo * and needs to be destroyed. 38538864b4e5STejun Heo */ 38548864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 38558864b4e5STejun Heo { 38568864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 38578864b4e5STejun Heo unbound_release_work); 38588864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 38598864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3860b42b0bddSYang Yingliang bool is_last = false; 38618864b4e5STejun Heo 3862b42b0bddSYang Yingliang /* 3863b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3864b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3865b42b0bddSYang Yingliang */ 3866b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 38678864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 38688864b4e5STejun Heo return; 38698864b4e5STejun Heo 38703c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 38718864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3872bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 38733c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3874b42b0bddSYang Yingliang } 38758864b4e5STejun Heo 3876a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 38778864b4e5STejun Heo put_unbound_pool(pool); 3878a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3879a892caccSTejun Heo 388025b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 38818864b4e5STejun Heo 38828864b4e5STejun Heo /* 38838864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3884e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 38858864b4e5STejun Heo */ 3886669de8bdSBart Van Assche if (is_last) { 3887669de8bdSBart Van Assche wq_unregister_lockdep(wq); 388825b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 38896029a918STejun Heo } 3890669de8bdSBart Van Assche } 38918864b4e5STejun Heo 38920fbd95aaSTejun Heo /** 3893699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 38940fbd95aaSTejun Heo * @pwq: target pool_workqueue 38950fbd95aaSTejun Heo * 3896699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3897f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3898699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 38990fbd95aaSTejun Heo */ 3900699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 39010fbd95aaSTejun Heo { 3902699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3903699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 39043347fa09STejun Heo unsigned long flags; 3905699ce097STejun Heo 3906699ce097STejun Heo /* for @wq->saved_max_active */ 3907a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3908699ce097STejun Heo 3909699ce097STejun Heo /* fast exit for non-freezable wqs */ 3910699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3911699ce097STejun Heo return; 3912699ce097STejun Heo 39133347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3914a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3915699ce097STejun Heo 391674b414eaSLai Jiangshan /* 391774b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 391874b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 391974b414eaSLai Jiangshan * is updated and visible. 392074b414eaSLai Jiangshan */ 392174b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 392201341fbdSYunfeng Ye bool kick = false; 392301341fbdSYunfeng Ye 3924699ce097STejun Heo pwq->max_active = wq->saved_max_active; 39250fbd95aaSTejun Heo 3926f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 392701341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3928f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 392901341fbdSYunfeng Ye kick = true; 393001341fbdSYunfeng Ye } 3931951a078aSLai Jiangshan 3932951a078aSLai Jiangshan /* 3933951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 393401341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 393501341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 393601341fbdSYunfeng Ye * let's kick iff work items were activated. 3937951a078aSLai Jiangshan */ 393801341fbdSYunfeng Ye if (kick) 3939951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3940699ce097STejun Heo } else { 3941699ce097STejun Heo pwq->max_active = 0; 3942699ce097STejun Heo } 3943699ce097STejun Heo 3944a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 39450fbd95aaSTejun Heo } 39460fbd95aaSTejun Heo 394767dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3948f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3949f147f29eSTejun Heo struct worker_pool *pool) 3950d2c1d404STejun Heo { 3951d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3952d2c1d404STejun Heo 3953e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3954e50aba9aSTejun Heo 3955d2c1d404STejun Heo pwq->pool = pool; 3956d2c1d404STejun Heo pwq->wq = wq; 3957d2c1d404STejun Heo pwq->flush_color = -1; 39588864b4e5STejun Heo pwq->refcnt = 1; 3959f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 39601befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3961d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 39628864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3963f147f29eSTejun Heo } 3964d2c1d404STejun Heo 3965f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 39661befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3967f147f29eSTejun Heo { 3968f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3969f147f29eSTejun Heo 3970f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 397175ccf595STejun Heo 39721befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 39731befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 39741befcf30STejun Heo return; 39751befcf30STejun Heo 397629b1cb41SLai Jiangshan /* set the matching work_color */ 397775ccf595STejun Heo pwq->work_color = wq->work_color; 3978983ca25eSTejun Heo 3979983ca25eSTejun Heo /* sync max_active to the current setting */ 3980983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3981983ca25eSTejun Heo 3982983ca25eSTejun Heo /* link in @pwq */ 39839e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3984df2d5ae4STejun Heo } 39856029a918STejun Heo 3986f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3987f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3988f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3989f147f29eSTejun Heo { 3990f147f29eSTejun Heo struct worker_pool *pool; 3991f147f29eSTejun Heo struct pool_workqueue *pwq; 3992f147f29eSTejun Heo 3993f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3994f147f29eSTejun Heo 3995f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3996f147f29eSTejun Heo if (!pool) 3997f147f29eSTejun Heo return NULL; 3998f147f29eSTejun Heo 3999e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4000f147f29eSTejun Heo if (!pwq) { 4001f147f29eSTejun Heo put_unbound_pool(pool); 4002f147f29eSTejun Heo return NULL; 4003f147f29eSTejun Heo } 4004f147f29eSTejun Heo 4005f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4006f147f29eSTejun Heo return pwq; 4007d2c1d404STejun Heo } 4008d2c1d404STejun Heo 40094c16bd32STejun Heo /** 401030186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 4011042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 40124c16bd32STejun Heo * @node: the target NUMA node 40134c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 40144c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 40154c16bd32STejun Heo * 40164c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 40174c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 4018d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 40194c16bd32STejun Heo * 40204c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 40214c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 40224c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 40234c16bd32STejun Heo * @attrs->cpumask. 40244c16bd32STejun Heo * 40254c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 40264c16bd32STejun Heo * stable. 4027d185af30SYacine Belkadi * 4028d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 4029d185af30SYacine Belkadi * %false if equal. 40304c16bd32STejun Heo */ 40314c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 40324c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 40334c16bd32STejun Heo { 4034d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 40354c16bd32STejun Heo goto use_dfl; 40364c16bd32STejun Heo 40374c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 40384c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 40394c16bd32STejun Heo if (cpu_going_down >= 0) 40404c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 40414c16bd32STejun Heo 40424c16bd32STejun Heo if (cpumask_empty(cpumask)) 40434c16bd32STejun Heo goto use_dfl; 40444c16bd32STejun Heo 40454c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 40464c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 40471ad0f0a7SMichael Bringmann 40481ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 40491ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 40501ad0f0a7SMichael Bringmann "possible intersect\n"); 40511ad0f0a7SMichael Bringmann return false; 40521ad0f0a7SMichael Bringmann } 40531ad0f0a7SMichael Bringmann 40544c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 40554c16bd32STejun Heo 40564c16bd32STejun Heo use_dfl: 40574c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 40584c16bd32STejun Heo return false; 40594c16bd32STejun Heo } 40604c16bd32STejun Heo 40611befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 40621befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 40631befcf30STejun Heo int node, 40641befcf30STejun Heo struct pool_workqueue *pwq) 40651befcf30STejun Heo { 40661befcf30STejun Heo struct pool_workqueue *old_pwq; 40671befcf30STejun Heo 40685b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 40691befcf30STejun Heo lockdep_assert_held(&wq->mutex); 40701befcf30STejun Heo 40711befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 40721befcf30STejun Heo link_pwq(pwq); 40731befcf30STejun Heo 40741befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 40751befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 40761befcf30STejun Heo return old_pwq; 40771befcf30STejun Heo } 40781befcf30STejun Heo 40792d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 40802d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 40812d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 40822d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4083042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 40842d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 40852d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 40862d5f0764SLai Jiangshan }; 40872d5f0764SLai Jiangshan 40882d5f0764SLai Jiangshan /* free the resources after success or abort */ 40892d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 40902d5f0764SLai Jiangshan { 40912d5f0764SLai Jiangshan if (ctx) { 40922d5f0764SLai Jiangshan int node; 40932d5f0764SLai Jiangshan 40942d5f0764SLai Jiangshan for_each_node(node) 40952d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 40962d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 40972d5f0764SLai Jiangshan 40982d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 40992d5f0764SLai Jiangshan 41002d5f0764SLai Jiangshan kfree(ctx); 41012d5f0764SLai Jiangshan } 41022d5f0764SLai Jiangshan } 41032d5f0764SLai Jiangshan 41042d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 41052d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 41062d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 410799c621efSLai Jiangshan const struct workqueue_attrs *attrs, 410899c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 41092d5f0764SLai Jiangshan { 41102d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 41112d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 41122d5f0764SLai Jiangshan int node; 41132d5f0764SLai Jiangshan 41142d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 41152d5f0764SLai Jiangshan 4116acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 41172d5f0764SLai Jiangshan 4118be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4119be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 41202d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 41212d5f0764SLai Jiangshan goto out_free; 41222d5f0764SLai Jiangshan 4123042f7df1SLai Jiangshan /* 412499c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 412599c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4126042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4127042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4128042f7df1SLai Jiangshan */ 41292d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 413099c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4131042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 413299c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 41332d5f0764SLai Jiangshan 41342d5f0764SLai Jiangshan /* 41352d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 41362d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 41372d5f0764SLai Jiangshan * pools. 41382d5f0764SLai Jiangshan */ 41392d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 41402d5f0764SLai Jiangshan 41412d5f0764SLai Jiangshan /* 41422d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 41432d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 41442d5f0764SLai Jiangshan * it even if we don't use it immediately. 41452d5f0764SLai Jiangshan */ 41462d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 41472d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 41482d5f0764SLai Jiangshan goto out_free; 41492d5f0764SLai Jiangshan 41502d5f0764SLai Jiangshan for_each_node(node) { 4151042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 41522d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 41532d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 41542d5f0764SLai Jiangshan goto out_free; 41552d5f0764SLai Jiangshan } else { 41562d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 41572d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 41582d5f0764SLai Jiangshan } 41592d5f0764SLai Jiangshan } 41602d5f0764SLai Jiangshan 4161042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4162042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4163042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 41642d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4165042f7df1SLai Jiangshan 41662d5f0764SLai Jiangshan ctx->wq = wq; 41672d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 41682d5f0764SLai Jiangshan return ctx; 41692d5f0764SLai Jiangshan 41702d5f0764SLai Jiangshan out_free: 41712d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 41722d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 41732d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 41742d5f0764SLai Jiangshan return NULL; 41752d5f0764SLai Jiangshan } 41762d5f0764SLai Jiangshan 41772d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 41782d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 41792d5f0764SLai Jiangshan { 41802d5f0764SLai Jiangshan int node; 41812d5f0764SLai Jiangshan 41822d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 41832d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 41842d5f0764SLai Jiangshan 41852d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 41862d5f0764SLai Jiangshan 41872d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 41882d5f0764SLai Jiangshan for_each_node(node) 41892d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 41902d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 41912d5f0764SLai Jiangshan 41922d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 41932d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 41942d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 41952d5f0764SLai Jiangshan 41962d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 41972d5f0764SLai Jiangshan } 41982d5f0764SLai Jiangshan 4199a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4200a0111cf6SLai Jiangshan { 4201a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4202ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4203a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4204a0111cf6SLai Jiangshan } 4205a0111cf6SLai Jiangshan 4206a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4207a0111cf6SLai Jiangshan { 4208a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4209ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4210a0111cf6SLai Jiangshan } 4211a0111cf6SLai Jiangshan 4212a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4213a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4214a0111cf6SLai Jiangshan { 4215a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4216a0111cf6SLai Jiangshan 4217a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4218a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4219a0111cf6SLai Jiangshan return -EINVAL; 4220a0111cf6SLai Jiangshan 4221a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 42220a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 42230a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4224a0111cf6SLai Jiangshan return -EINVAL; 4225a0111cf6SLai Jiangshan 42260a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 42270a94efb5STejun Heo } 42280a94efb5STejun Heo 422999c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 42306201171eSwanghaibin if (!ctx) 42316201171eSwanghaibin return -ENOMEM; 4232a0111cf6SLai Jiangshan 4233a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4234a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4235a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4236a0111cf6SLai Jiangshan 42376201171eSwanghaibin return 0; 4238a0111cf6SLai Jiangshan } 4239a0111cf6SLai Jiangshan 42409e8cd2f5STejun Heo /** 42419e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 42429e8cd2f5STejun Heo * @wq: the target workqueue 42439e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 42449e8cd2f5STejun Heo * 42454c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 42464c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 42474c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 42484c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 42494c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 42504c16bd32STejun Heo * back-to-back will stay on its current pwq. 42519e8cd2f5STejun Heo * 4252d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4253d185af30SYacine Belkadi * 4254ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4255509b3204SDaniel Jordan * 4256d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 42579e8cd2f5STejun Heo */ 4258513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 42599e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 42609e8cd2f5STejun Heo { 4261a0111cf6SLai Jiangshan int ret; 42629e8cd2f5STejun Heo 4263509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4264509b3204SDaniel Jordan 4265509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4266a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4267509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 42682d5f0764SLai Jiangshan 42692d5f0764SLai Jiangshan return ret; 42709e8cd2f5STejun Heo } 42719e8cd2f5STejun Heo 42724c16bd32STejun Heo /** 42734c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 42744c16bd32STejun Heo * @wq: the target workqueue 42754c16bd32STejun Heo * @cpu: the CPU coming up or going down 42764c16bd32STejun Heo * @online: whether @cpu is coming up or going down 42774c16bd32STejun Heo * 42784c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 42794c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 42804c16bd32STejun Heo * @wq accordingly. 42814c16bd32STejun Heo * 42824c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 42834c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 42844c16bd32STejun Heo * correct. 42854c16bd32STejun Heo * 42864c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 42874c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 42884c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 42894c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 42904c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 42914c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 42924c16bd32STejun Heo * CPU_DOWN_PREPARE. 42934c16bd32STejun Heo */ 42944c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 42954c16bd32STejun Heo bool online) 42964c16bd32STejun Heo { 42974c16bd32STejun Heo int node = cpu_to_node(cpu); 42984c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 42994c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 43004c16bd32STejun Heo struct workqueue_attrs *target_attrs; 43014c16bd32STejun Heo cpumask_t *cpumask; 43024c16bd32STejun Heo 43034c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 43044c16bd32STejun Heo 4305f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4306f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 43074c16bd32STejun Heo return; 43084c16bd32STejun Heo 43094c16bd32STejun Heo /* 43104c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 43114c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 43124c16bd32STejun Heo * CPU hotplug exclusion. 43134c16bd32STejun Heo */ 43144c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 43154c16bd32STejun Heo cpumask = target_attrs->cpumask; 43164c16bd32STejun Heo 43174c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 43184c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 43194c16bd32STejun Heo 43204c16bd32STejun Heo /* 43214c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4322042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4323042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4324042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 43254c16bd32STejun Heo */ 4326042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 43274c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4328f7142ed4SLai Jiangshan return; 43294c16bd32STejun Heo } else { 43304c16bd32STejun Heo goto use_dfl_pwq; 43314c16bd32STejun Heo } 43324c16bd32STejun Heo 43334c16bd32STejun Heo /* create a new pwq */ 43344c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 43354c16bd32STejun Heo if (!pwq) { 43362d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 43374c16bd32STejun Heo wq->name); 433877f300b1SDaeseok Youn goto use_dfl_pwq; 43394c16bd32STejun Heo } 43404c16bd32STejun Heo 4341f7142ed4SLai Jiangshan /* Install the new pwq. */ 43424c16bd32STejun Heo mutex_lock(&wq->mutex); 43434c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 43444c16bd32STejun Heo goto out_unlock; 43454c16bd32STejun Heo 43464c16bd32STejun Heo use_dfl_pwq: 4347f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4348a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 43494c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4350a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 43514c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 43524c16bd32STejun Heo out_unlock: 43534c16bd32STejun Heo mutex_unlock(&wq->mutex); 43544c16bd32STejun Heo put_pwq_unlocked(old_pwq); 43554c16bd32STejun Heo } 43564c16bd32STejun Heo 435730cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 43581da177e4SLinus Torvalds { 435949e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 43608a2b7538STejun Heo int cpu, ret; 4361e1d8aa9fSFrederic Weisbecker 436230cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4363420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4364420c0ddbSTejun Heo if (!wq->cpu_pwqs) 436530cdf249STejun Heo return -ENOMEM; 436630cdf249STejun Heo 436730cdf249STejun Heo for_each_possible_cpu(cpu) { 43687fb98ea7STejun Heo struct pool_workqueue *pwq = 43697fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 43707a62c2c8STejun Heo struct worker_pool *cpu_pools = 4371f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 437230cdf249STejun Heo 4373f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4374f147f29eSTejun Heo 4375f147f29eSTejun Heo mutex_lock(&wq->mutex); 43761befcf30STejun Heo link_pwq(pwq); 4377f147f29eSTejun Heo mutex_unlock(&wq->mutex); 437830cdf249STejun Heo } 437930cdf249STejun Heo return 0; 4380509b3204SDaniel Jordan } 4381509b3204SDaniel Jordan 4382ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4383509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 43848a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 43858a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 43868a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 43878a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 43888a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 43899e8cd2f5STejun Heo } else { 4390509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 43919e8cd2f5STejun Heo } 4392ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4393509b3204SDaniel Jordan 4394509b3204SDaniel Jordan return ret; 43950f900049STejun Heo } 43960f900049STejun Heo 4397f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4398f3421797STejun Heo const char *name) 4399b71ab8c2STejun Heo { 4400f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4401f3421797STejun Heo 4402f3421797STejun Heo if (max_active < 1 || max_active > lim) 4403044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4404f3421797STejun Heo max_active, name, 1, lim); 4405b71ab8c2STejun Heo 4406f3421797STejun Heo return clamp_val(max_active, 1, lim); 4407b71ab8c2STejun Heo } 4408b71ab8c2STejun Heo 4409983c7515STejun Heo /* 4410983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4411983c7515STejun Heo * to guarantee forward progress. 4412983c7515STejun Heo */ 4413983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4414983c7515STejun Heo { 4415983c7515STejun Heo struct worker *rescuer; 4416b92b36eaSDan Carpenter int ret; 4417983c7515STejun Heo 4418983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4419983c7515STejun Heo return 0; 4420983c7515STejun Heo 4421983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 44224c0736a7SPetr Mladek if (!rescuer) { 44234c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 44244c0736a7SPetr Mladek wq->name); 4425983c7515STejun Heo return -ENOMEM; 44264c0736a7SPetr Mladek } 4427983c7515STejun Heo 4428983c7515STejun Heo rescuer->rescue_wq = wq; 4429983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4430f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4431b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 44324c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 44334c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4434983c7515STejun Heo kfree(rescuer); 4435b92b36eaSDan Carpenter return ret; 4436983c7515STejun Heo } 4437983c7515STejun Heo 4438983c7515STejun Heo wq->rescuer = rescuer; 4439983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4440983c7515STejun Heo wake_up_process(rescuer->task); 4441983c7515STejun Heo 4442983c7515STejun Heo return 0; 4443983c7515STejun Heo } 4444983c7515STejun Heo 4445a2775bbcSMathieu Malaterre __printf(1, 4) 4446669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 444797e37d7bSTejun Heo unsigned int flags, 4448669de8bdSBart Van Assche int max_active, ...) 44493af24433SOleg Nesterov { 4450df2d5ae4STejun Heo size_t tbl_size = 0; 4451ecf6881fSTejun Heo va_list args; 44523af24433SOleg Nesterov struct workqueue_struct *wq; 445349e3cf44STejun Heo struct pool_workqueue *pwq; 4454b196be89STejun Heo 44555c0338c6STejun Heo /* 44565c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 44575c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 44585c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 44595c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 44605c0338c6STejun Heo * on NUMA. 44615c0338c6STejun Heo */ 44625c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 44635c0338c6STejun Heo flags |= __WQ_ORDERED; 44645c0338c6STejun Heo 4465cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4466cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4467cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4468cee22a15SViresh Kumar 4469ecf6881fSTejun Heo /* allocate wq and format name */ 4470df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4471ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4472df2d5ae4STejun Heo 4473df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4474b196be89STejun Heo if (!wq) 4475d2c1d404STejun Heo return NULL; 4476b196be89STejun Heo 44776029a918STejun Heo if (flags & WQ_UNBOUND) { 4478be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 44796029a918STejun Heo if (!wq->unbound_attrs) 44806029a918STejun Heo goto err_free_wq; 44816029a918STejun Heo } 44826029a918STejun Heo 4483669de8bdSBart Van Assche va_start(args, max_active); 4484ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4485b196be89STejun Heo va_end(args); 44863af24433SOleg Nesterov 4487d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4488b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 44893af24433SOleg Nesterov 4490b196be89STejun Heo /* init wq */ 449197e37d7bSTejun Heo wq->flags = flags; 4492a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 44933c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4494112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 449530cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 449673f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 449773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4498493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 44993af24433SOleg Nesterov 4500669de8bdSBart Van Assche wq_init_lockdep(wq); 4501cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 45023af24433SOleg Nesterov 450330cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 450482efcab3SBart Van Assche goto err_unreg_lockdep; 45051537663fSTejun Heo 450640c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4507d2c1d404STejun Heo goto err_destroy; 4508e22bee78STejun Heo 4509226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4510226223abSTejun Heo goto err_destroy; 4511226223abSTejun Heo 45126af8bf3dSOleg Nesterov /* 451368e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 451468e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 451568e13a67SLai Jiangshan * list. 45166af8bf3dSOleg Nesterov */ 451768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4518a0a1a5fdSTejun Heo 4519a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 452049e3cf44STejun Heo for_each_pwq(pwq, wq) 4521699ce097STejun Heo pwq_adjust_max_active(pwq); 4522a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4523a0a1a5fdSTejun Heo 4524e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4525a0a1a5fdSTejun Heo 452668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45273af24433SOleg Nesterov 45283af24433SOleg Nesterov return wq; 4529d2c1d404STejun Heo 453082efcab3SBart Van Assche err_unreg_lockdep: 4531009bb421SBart Van Assche wq_unregister_lockdep(wq); 4532009bb421SBart Van Assche wq_free_lockdep(wq); 453382efcab3SBart Van Assche err_free_wq: 45346029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 45354690c4abSTejun Heo kfree(wq); 4536d2c1d404STejun Heo return NULL; 4537d2c1d404STejun Heo err_destroy: 4538d2c1d404STejun Heo destroy_workqueue(wq); 45394690c4abSTejun Heo return NULL; 45401da177e4SLinus Torvalds } 4541669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 45421da177e4SLinus Torvalds 4543c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4544c29eb853STejun Heo { 4545c29eb853STejun Heo int i; 4546c29eb853STejun Heo 4547c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4548c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4549c29eb853STejun Heo return true; 4550c29eb853STejun Heo 4551c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4552c29eb853STejun Heo return true; 4553f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4554c29eb853STejun Heo return true; 4555c29eb853STejun Heo 4556c29eb853STejun Heo return false; 4557c29eb853STejun Heo } 4558c29eb853STejun Heo 45593af24433SOleg Nesterov /** 45603af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 45613af24433SOleg Nesterov * @wq: target workqueue 45623af24433SOleg Nesterov * 45633af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 45643af24433SOleg Nesterov */ 45653af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 45663af24433SOleg Nesterov { 456749e3cf44STejun Heo struct pool_workqueue *pwq; 45684c16bd32STejun Heo int node; 45693af24433SOleg Nesterov 4570def98c84STejun Heo /* 4571def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4572def98c84STejun Heo * lead to sysfs name conflicts. 4573def98c84STejun Heo */ 4574def98c84STejun Heo workqueue_sysfs_unregister(wq); 4575def98c84STejun Heo 457633e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 457733e3f0a3SRichard Clark mutex_lock(&wq->mutex); 457833e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 457933e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 458033e3f0a3SRichard Clark 45819c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 45829c5a2ba7STejun Heo drain_workqueue(wq); 4583c8efcc25STejun Heo 4584def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4585def98c84STejun Heo if (wq->rescuer) { 4586def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4587def98c84STejun Heo 4588def98c84STejun Heo /* this prevents new queueing */ 4589a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4590def98c84STejun Heo wq->rescuer = NULL; 4591a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4592def98c84STejun Heo 4593def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4594def98c84STejun Heo kthread_stop(rescuer->task); 45958efe1223STejun Heo kfree(rescuer); 4596def98c84STejun Heo } 4597def98c84STejun Heo 4598c29eb853STejun Heo /* 4599c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4600c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4601c29eb853STejun Heo */ 4602c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4603b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 460449e3cf44STejun Heo for_each_pwq(pwq, wq) { 4605a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4606c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 46071d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4608e66b39afSTejun Heo __func__, wq->name); 4609c29eb853STejun Heo show_pwq(pwq); 4610a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4611b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4612c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 461355df0933SImran Khan show_one_workqueue(wq); 46146183c009STejun Heo return; 461576af4d93STejun Heo } 4616a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 461776af4d93STejun Heo } 4618b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 46196183c009STejun Heo 4620a0a1a5fdSTejun Heo /* 4621a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4622a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4623a0a1a5fdSTejun Heo */ 4624e2dca7adSTejun Heo list_del_rcu(&wq->list); 462568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46263af24433SOleg Nesterov 46278864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4628669de8bdSBart Van Assche wq_unregister_lockdep(wq); 462929c91e99STejun Heo /* 46308864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4631e2dca7adSTejun Heo * schedule RCU free. 463229c91e99STejun Heo */ 463325b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 46348864b4e5STejun Heo } else { 46358864b4e5STejun Heo /* 46368864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 46374c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 46384c16bd32STejun Heo * @wq will be freed when the last pwq is released. 46398864b4e5STejun Heo */ 46404c16bd32STejun Heo for_each_node(node) { 46414c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 46424c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 46434c16bd32STejun Heo put_pwq_unlocked(pwq); 46444c16bd32STejun Heo } 46454c16bd32STejun Heo 46464c16bd32STejun Heo /* 46474c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 46484c16bd32STejun Heo * put. Don't access it afterwards. 46494c16bd32STejun Heo */ 46504c16bd32STejun Heo pwq = wq->dfl_pwq; 46514c16bd32STejun Heo wq->dfl_pwq = NULL; 4652dce90d47STejun Heo put_pwq_unlocked(pwq); 465329c91e99STejun Heo } 46543af24433SOleg Nesterov } 46553af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 46563af24433SOleg Nesterov 4657dcd989cbSTejun Heo /** 4658dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4659dcd989cbSTejun Heo * @wq: target workqueue 4660dcd989cbSTejun Heo * @max_active: new max_active value. 4661dcd989cbSTejun Heo * 4662dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4663dcd989cbSTejun Heo * 4664dcd989cbSTejun Heo * CONTEXT: 4665dcd989cbSTejun Heo * Don't call from IRQ context. 4666dcd989cbSTejun Heo */ 4667dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4668dcd989cbSTejun Heo { 466949e3cf44STejun Heo struct pool_workqueue *pwq; 4670dcd989cbSTejun Heo 46718719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 46720a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 46738719dceaSTejun Heo return; 46748719dceaSTejun Heo 4675f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4676dcd989cbSTejun Heo 4677a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4678dcd989cbSTejun Heo 46790a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4680dcd989cbSTejun Heo wq->saved_max_active = max_active; 4681dcd989cbSTejun Heo 4682699ce097STejun Heo for_each_pwq(pwq, wq) 4683699ce097STejun Heo pwq_adjust_max_active(pwq); 4684dcd989cbSTejun Heo 4685a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4686dcd989cbSTejun Heo } 4687dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4688dcd989cbSTejun Heo 4689dcd989cbSTejun Heo /** 469027d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 469127d4ee03SLukas Wunner * 469227d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 469327d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 469427d4ee03SLukas Wunner * 469527d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 469627d4ee03SLukas Wunner */ 469727d4ee03SLukas Wunner struct work_struct *current_work(void) 469827d4ee03SLukas Wunner { 469927d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 470027d4ee03SLukas Wunner 470127d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 470227d4ee03SLukas Wunner } 470327d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 470427d4ee03SLukas Wunner 470527d4ee03SLukas Wunner /** 4706e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4707e6267616STejun Heo * 4708e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4709e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4710d185af30SYacine Belkadi * 4711d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4712e6267616STejun Heo */ 4713e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4714e6267616STejun Heo { 4715e6267616STejun Heo struct worker *worker = current_wq_worker(); 4716e6267616STejun Heo 47176a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4718e6267616STejun Heo } 4719e6267616STejun Heo 4720e6267616STejun Heo /** 4721dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4722dcd989cbSTejun Heo * @cpu: CPU in question 4723dcd989cbSTejun Heo * @wq: target workqueue 4724dcd989cbSTejun Heo * 4725dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4726dcd989cbSTejun Heo * no synchronization around this function and the test result is 4727dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4728dcd989cbSTejun Heo * 4729d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4730d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4731d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4732d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4733d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4734d3251859STejun Heo * 4735d185af30SYacine Belkadi * Return: 4736dcd989cbSTejun Heo * %true if congested, %false otherwise. 4737dcd989cbSTejun Heo */ 4738d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4739dcd989cbSTejun Heo { 47407fb98ea7STejun Heo struct pool_workqueue *pwq; 474176af4d93STejun Heo bool ret; 474276af4d93STejun Heo 474324acfb71SThomas Gleixner rcu_read_lock(); 474424acfb71SThomas Gleixner preempt_disable(); 47457fb98ea7STejun Heo 4746d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4747d3251859STejun Heo cpu = smp_processor_id(); 4748d3251859STejun Heo 47497fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 47507fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 47517fb98ea7STejun Heo else 4752df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4753dcd989cbSTejun Heo 4754f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 475524acfb71SThomas Gleixner preempt_enable(); 475624acfb71SThomas Gleixner rcu_read_unlock(); 475776af4d93STejun Heo 475876af4d93STejun Heo return ret; 4759dcd989cbSTejun Heo } 4760dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4761dcd989cbSTejun Heo 4762dcd989cbSTejun Heo /** 4763dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4764dcd989cbSTejun Heo * @work: the work to be tested 4765dcd989cbSTejun Heo * 4766dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4767dcd989cbSTejun Heo * synchronization around this function and the test result is 4768dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4769dcd989cbSTejun Heo * 4770d185af30SYacine Belkadi * Return: 4771dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4772dcd989cbSTejun Heo */ 4773dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4774dcd989cbSTejun Heo { 4775fa1b54e6STejun Heo struct worker_pool *pool; 4776dcd989cbSTejun Heo unsigned long flags; 4777dcd989cbSTejun Heo unsigned int ret = 0; 4778dcd989cbSTejun Heo 4779dcd989cbSTejun Heo if (work_pending(work)) 4780dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4781038366c5SLai Jiangshan 478224acfb71SThomas Gleixner rcu_read_lock(); 4783fa1b54e6STejun Heo pool = get_work_pool(work); 4784038366c5SLai Jiangshan if (pool) { 4785a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4786c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4787dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4788a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4789038366c5SLai Jiangshan } 479024acfb71SThomas Gleixner rcu_read_unlock(); 4791dcd989cbSTejun Heo 4792dcd989cbSTejun Heo return ret; 4793dcd989cbSTejun Heo } 4794dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4795dcd989cbSTejun Heo 47963d1cb205STejun Heo /** 47973d1cb205STejun Heo * set_worker_desc - set description for the current work item 47983d1cb205STejun Heo * @fmt: printf-style format string 47993d1cb205STejun Heo * @...: arguments for the format string 48003d1cb205STejun Heo * 48013d1cb205STejun Heo * This function can be called by a running work function to describe what 48023d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 48033d1cb205STejun Heo * information will be printed out together to help debugging. The 48043d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 48053d1cb205STejun Heo */ 48063d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 48073d1cb205STejun Heo { 48083d1cb205STejun Heo struct worker *worker = current_wq_worker(); 48093d1cb205STejun Heo va_list args; 48103d1cb205STejun Heo 48113d1cb205STejun Heo if (worker) { 48123d1cb205STejun Heo va_start(args, fmt); 48133d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 48143d1cb205STejun Heo va_end(args); 48153d1cb205STejun Heo } 48163d1cb205STejun Heo } 48175c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 48183d1cb205STejun Heo 48193d1cb205STejun Heo /** 48203d1cb205STejun Heo * print_worker_info - print out worker information and description 48213d1cb205STejun Heo * @log_lvl: the log level to use when printing 48223d1cb205STejun Heo * @task: target task 48233d1cb205STejun Heo * 48243d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 48253d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 48263d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 48273d1cb205STejun Heo * 48283d1cb205STejun Heo * This function can be safely called on any task as long as the 48293d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 48303d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 48313d1cb205STejun Heo */ 48323d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 48333d1cb205STejun Heo { 48343d1cb205STejun Heo work_func_t *fn = NULL; 48353d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 48363d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 48373d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 48383d1cb205STejun Heo struct workqueue_struct *wq = NULL; 48393d1cb205STejun Heo struct worker *worker; 48403d1cb205STejun Heo 48413d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 48423d1cb205STejun Heo return; 48433d1cb205STejun Heo 48443d1cb205STejun Heo /* 48453d1cb205STejun Heo * This function is called without any synchronization and @task 48463d1cb205STejun Heo * could be in any state. Be careful with dereferences. 48473d1cb205STejun Heo */ 4848e700591aSPetr Mladek worker = kthread_probe_data(task); 48493d1cb205STejun Heo 48503d1cb205STejun Heo /* 48518bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 48528bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 48533d1cb205STejun Heo */ 4854fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4855fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4856fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4857fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4858fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 48593d1cb205STejun Heo 48603d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4861d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 48628bf89593STejun Heo if (strcmp(name, desc)) 48633d1cb205STejun Heo pr_cont(" (%s)", desc); 48643d1cb205STejun Heo pr_cont("\n"); 48653d1cb205STejun Heo } 48663d1cb205STejun Heo } 48673d1cb205STejun Heo 48683494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 48693494fc30STejun Heo { 48703494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 48713494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 48723494fc30STejun Heo pr_cont(" node=%d", pool->node); 48733494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 48743494fc30STejun Heo } 48753494fc30STejun Heo 4876c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 4877c76feb0dSPaul E. McKenney bool comma; 4878c76feb0dSPaul E. McKenney work_func_t func; 4879c76feb0dSPaul E. McKenney long ctr; 4880c76feb0dSPaul E. McKenney }; 4881c76feb0dSPaul E. McKenney 4882c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 4883c76feb0dSPaul E. McKenney { 4884c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 4885c76feb0dSPaul E. McKenney goto out_record; 4886c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 4887c76feb0dSPaul E. McKenney pcwsp->ctr++; 4888c76feb0dSPaul E. McKenney return; 4889c76feb0dSPaul E. McKenney } 4890c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 4891c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 4892c76feb0dSPaul E. McKenney else 4893c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 4894c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 4895c76feb0dSPaul E. McKenney out_record: 4896c76feb0dSPaul E. McKenney if ((long)func == -1L) 4897c76feb0dSPaul E. McKenney return; 4898c76feb0dSPaul E. McKenney pcwsp->comma = comma; 4899c76feb0dSPaul E. McKenney pcwsp->func = func; 4900c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 4901c76feb0dSPaul E. McKenney } 4902c76feb0dSPaul E. McKenney 4903c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 49043494fc30STejun Heo { 49053494fc30STejun Heo if (work->func == wq_barrier_func) { 49063494fc30STejun Heo struct wq_barrier *barr; 49073494fc30STejun Heo 49083494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 49093494fc30STejun Heo 4910c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 49113494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 49123494fc30STejun Heo task_pid_nr(barr->task)); 49133494fc30STejun Heo } else { 4914c76feb0dSPaul E. McKenney if (!comma) 4915c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 4916c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 49173494fc30STejun Heo } 49183494fc30STejun Heo } 49193494fc30STejun Heo 49203494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 49213494fc30STejun Heo { 4922c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 49233494fc30STejun Heo struct worker_pool *pool = pwq->pool; 49243494fc30STejun Heo struct work_struct *work; 49253494fc30STejun Heo struct worker *worker; 49263494fc30STejun Heo bool has_in_flight = false, has_pending = false; 49273494fc30STejun Heo int bkt; 49283494fc30STejun Heo 49293494fc30STejun Heo pr_info(" pwq %d:", pool->id); 49303494fc30STejun Heo pr_cont_pool_info(pool); 49313494fc30STejun Heo 4932e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4933e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 49343494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 49353494fc30STejun Heo 49363494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 49373494fc30STejun Heo if (worker->current_pwq == pwq) { 49383494fc30STejun Heo has_in_flight = true; 49393494fc30STejun Heo break; 49403494fc30STejun Heo } 49413494fc30STejun Heo } 49423494fc30STejun Heo if (has_in_flight) { 49433494fc30STejun Heo bool comma = false; 49443494fc30STejun Heo 49453494fc30STejun Heo pr_info(" in-flight:"); 49463494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 49473494fc30STejun Heo if (worker->current_pwq != pwq) 49483494fc30STejun Heo continue; 49493494fc30STejun Heo 4950d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 49513494fc30STejun Heo task_pid_nr(worker->task), 495230ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 49533494fc30STejun Heo worker->current_func); 49543494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 4955c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 4956c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 49573494fc30STejun Heo comma = true; 49583494fc30STejun Heo } 49593494fc30STejun Heo pr_cont("\n"); 49603494fc30STejun Heo } 49613494fc30STejun Heo 49623494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 49633494fc30STejun Heo if (get_work_pwq(work) == pwq) { 49643494fc30STejun Heo has_pending = true; 49653494fc30STejun Heo break; 49663494fc30STejun Heo } 49673494fc30STejun Heo } 49683494fc30STejun Heo if (has_pending) { 49693494fc30STejun Heo bool comma = false; 49703494fc30STejun Heo 49713494fc30STejun Heo pr_info(" pending:"); 49723494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 49733494fc30STejun Heo if (get_work_pwq(work) != pwq) 49743494fc30STejun Heo continue; 49753494fc30STejun Heo 4976c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 49773494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 49783494fc30STejun Heo } 4979c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 49803494fc30STejun Heo pr_cont("\n"); 49813494fc30STejun Heo } 49823494fc30STejun Heo 4983f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 49843494fc30STejun Heo bool comma = false; 49853494fc30STejun Heo 4986f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4987f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 4988c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 49893494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 49903494fc30STejun Heo } 4991c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 49923494fc30STejun Heo pr_cont("\n"); 49933494fc30STejun Heo } 49943494fc30STejun Heo } 49953494fc30STejun Heo 49963494fc30STejun Heo /** 499755df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 499855df0933SImran Khan * @wq: workqueue whose state will be printed 49993494fc30STejun Heo */ 500055df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 50013494fc30STejun Heo { 50023494fc30STejun Heo struct pool_workqueue *pwq; 50033494fc30STejun Heo bool idle = true; 500455df0933SImran Khan unsigned long flags; 50053494fc30STejun Heo 50063494fc30STejun Heo for_each_pwq(pwq, wq) { 5007f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 50083494fc30STejun Heo idle = false; 50093494fc30STejun Heo break; 50103494fc30STejun Heo } 50113494fc30STejun Heo } 501255df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 501355df0933SImran Khan return; 50143494fc30STejun Heo 50153494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 50163494fc30STejun Heo 50173494fc30STejun Heo for_each_pwq(pwq, wq) { 5018a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 501957116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 502057116ce1SJohan Hovold /* 502157116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 502257116ce1SJohan Hovold * drivers that queue work while holding locks 502357116ce1SJohan Hovold * also taken in their write paths. 502457116ce1SJohan Hovold */ 502557116ce1SJohan Hovold printk_deferred_enter(); 50263494fc30STejun Heo show_pwq(pwq); 502757116ce1SJohan Hovold printk_deferred_exit(); 502857116ce1SJohan Hovold } 5029a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 503062635ea8SSergey Senozhatsky /* 503162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 503255df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 503362635ea8SSergey Senozhatsky * hard lockup. 503462635ea8SSergey Senozhatsky */ 503562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 50363494fc30STejun Heo } 503755df0933SImran Khan 50383494fc30STejun Heo } 50393494fc30STejun Heo 504055df0933SImran Khan /** 504155df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 504255df0933SImran Khan * @pool: worker pool whose state will be printed 504355df0933SImran Khan */ 504455df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 504555df0933SImran Khan { 50463494fc30STejun Heo struct worker *worker; 50473494fc30STejun Heo bool first = true; 504855df0933SImran Khan unsigned long flags; 5049335a42ebSPetr Mladek unsigned long hung = 0; 50503494fc30STejun Heo 5051a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 50523494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 50533494fc30STejun Heo goto next_pool; 5054335a42ebSPetr Mladek 5055335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5056335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5057335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5058335a42ebSPetr Mladek 505957116ce1SJohan Hovold /* 506057116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 506157116ce1SJohan Hovold * queue work while holding locks also taken in their write 506257116ce1SJohan Hovold * paths. 506357116ce1SJohan Hovold */ 506457116ce1SJohan Hovold printk_deferred_enter(); 50653494fc30STejun Heo pr_info("pool %d:", pool->id); 50663494fc30STejun Heo pr_cont_pool_info(pool); 5067335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 50683494fc30STejun Heo if (pool->manager) 50693494fc30STejun Heo pr_cont(" manager: %d", 50703494fc30STejun Heo task_pid_nr(pool->manager->task)); 50713494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 50723494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 50733494fc30STejun Heo task_pid_nr(worker->task)); 50743494fc30STejun Heo first = false; 50753494fc30STejun Heo } 50763494fc30STejun Heo pr_cont("\n"); 507757116ce1SJohan Hovold printk_deferred_exit(); 50783494fc30STejun Heo next_pool: 5079a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 508062635ea8SSergey Senozhatsky /* 508162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 508255df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 508362635ea8SSergey Senozhatsky * hard lockup. 508462635ea8SSergey Senozhatsky */ 508562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 508655df0933SImran Khan 50873494fc30STejun Heo } 50883494fc30STejun Heo 508955df0933SImran Khan /** 509055df0933SImran Khan * show_all_workqueues - dump workqueue state 509155df0933SImran Khan * 5092704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 509355df0933SImran Khan */ 509455df0933SImran Khan void show_all_workqueues(void) 509555df0933SImran Khan { 509655df0933SImran Khan struct workqueue_struct *wq; 509755df0933SImran Khan struct worker_pool *pool; 509855df0933SImran Khan int pi; 509955df0933SImran Khan 510055df0933SImran Khan rcu_read_lock(); 510155df0933SImran Khan 510255df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 510355df0933SImran Khan 510455df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 510555df0933SImran Khan show_one_workqueue(wq); 510655df0933SImran Khan 510755df0933SImran Khan for_each_pool(pool, pi) 510855df0933SImran Khan show_one_worker_pool(pool); 510955df0933SImran Khan 511024acfb71SThomas Gleixner rcu_read_unlock(); 51113494fc30STejun Heo } 51123494fc30STejun Heo 5113704bc669SJungseung Lee /** 5114704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5115704bc669SJungseung Lee * 5116704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5117704bc669SJungseung Lee * still busy. 5118704bc669SJungseung Lee */ 5119704bc669SJungseung Lee void show_freezable_workqueues(void) 5120704bc669SJungseung Lee { 5121704bc669SJungseung Lee struct workqueue_struct *wq; 5122704bc669SJungseung Lee 5123704bc669SJungseung Lee rcu_read_lock(); 5124704bc669SJungseung Lee 5125704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5126704bc669SJungseung Lee 5127704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5128704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5129704bc669SJungseung Lee continue; 5130704bc669SJungseung Lee show_one_workqueue(wq); 5131704bc669SJungseung Lee } 5132704bc669SJungseung Lee 5133704bc669SJungseung Lee rcu_read_unlock(); 5134704bc669SJungseung Lee } 5135704bc669SJungseung Lee 51366b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 51376b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 51386b59808bSTejun Heo { 51396b59808bSTejun Heo int off; 51406b59808bSTejun Heo 51416b59808bSTejun Heo /* always show the actual comm */ 51426b59808bSTejun Heo off = strscpy(buf, task->comm, size); 51436b59808bSTejun Heo if (off < 0) 51446b59808bSTejun Heo return; 51456b59808bSTejun Heo 5146197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 51476b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 51486b59808bSTejun Heo 5149197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5150197f6accSTejun Heo struct worker *worker = kthread_data(task); 5151197f6accSTejun Heo struct worker_pool *pool = worker->pool; 51526b59808bSTejun Heo 51536b59808bSTejun Heo if (pool) { 5154a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 51556b59808bSTejun Heo /* 5156197f6accSTejun Heo * ->desc tracks information (wq name or 5157197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5158197f6accSTejun Heo * current, prepend '+', otherwise '-'. 51596b59808bSTejun Heo */ 51606b59808bSTejun Heo if (worker->desc[0] != '\0') { 51616b59808bSTejun Heo if (worker->current_work) 51626b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 51636b59808bSTejun Heo worker->desc); 51646b59808bSTejun Heo else 51656b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 51666b59808bSTejun Heo worker->desc); 51676b59808bSTejun Heo } 5168a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 51696b59808bSTejun Heo } 5170197f6accSTejun Heo } 51716b59808bSTejun Heo 51726b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 51736b59808bSTejun Heo } 51746b59808bSTejun Heo 517566448bc2SMathieu Malaterre #ifdef CONFIG_SMP 517666448bc2SMathieu Malaterre 5177db7bccf4STejun Heo /* 5178db7bccf4STejun Heo * CPU hotplug. 5179db7bccf4STejun Heo * 5180e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5181112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5182706026c2STejun Heo * pool which make migrating pending and scheduled works very 5183e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 518494cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5185e22bee78STejun Heo * blocked draining impractical. 5186e22bee78STejun Heo * 518724647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5188628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5189628c78e7STejun Heo * cpu comes back online. 5190db7bccf4STejun Heo */ 5191db7bccf4STejun Heo 5192e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5193db7bccf4STejun Heo { 51944ce62e9eSTejun Heo struct worker_pool *pool; 5195db7bccf4STejun Heo struct worker *worker; 5196db7bccf4STejun Heo 5197f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51981258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5199a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5200e22bee78STejun Heo 5201f2d5a0eeSTejun Heo /* 520292f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 520394cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 520411b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5205b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5206b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5207b4ac9384SLai Jiangshan * is on the same cpu. 5208f2d5a0eeSTejun Heo */ 5209da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5210403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5211db7bccf4STejun Heo 521224647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5213f2d5a0eeSTejun Heo 5214e22bee78STejun Heo /* 5215989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5216989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5217989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5218989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5219989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5220eb283428SLai Jiangshan * are served by workers tied to the pool. 5221e22bee78STejun Heo */ 5222bc35f7efSLai Jiangshan pool->nr_running = 0; 5223eb283428SLai Jiangshan 5224eb283428SLai Jiangshan /* 5225eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5226eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5227eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5228eb283428SLai Jiangshan */ 5229eb283428SLai Jiangshan wake_up_worker(pool); 5230989442d7SLai Jiangshan 5231a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5232989442d7SLai Jiangshan 5233793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5234793777bcSValentin Schneider unbind_worker(worker); 5235989442d7SLai Jiangshan 5236989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5237eb283428SLai Jiangshan } 5238db7bccf4STejun Heo } 5239db7bccf4STejun Heo 5240bd7c089eSTejun Heo /** 5241bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5242bd7c089eSTejun Heo * @pool: pool of interest 5243bd7c089eSTejun Heo * 5244a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5245bd7c089eSTejun Heo */ 5246bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5247bd7c089eSTejun Heo { 5248a9ab775bSTejun Heo struct worker *worker; 5249bd7c089eSTejun Heo 52501258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5251bd7c089eSTejun Heo 5252bd7c089eSTejun Heo /* 5253a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5254a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5255402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5256a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5257a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5258bd7c089eSTejun Heo */ 5259c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5260c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5261c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5262c63a2e52SValentin Schneider pool->attrs->cpumask) < 0); 5263c63a2e52SValentin Schneider } 5264a9ab775bSTejun Heo 5265a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5266f7c17d26SWanpeng Li 52673de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5268a9ab775bSTejun Heo 5269da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5270a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5271a9ab775bSTejun Heo 5272a9ab775bSTejun Heo /* 5273a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5274a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5275a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5276a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5277a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5278a9ab775bSTejun Heo * concurrency management. Note that when or whether 5279a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5280a9ab775bSTejun Heo * 5281c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5282a9ab775bSTejun Heo * tested without holding any lock in 52836d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5284a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5285a9ab775bSTejun Heo * management operations. 5286bd7c089eSTejun Heo */ 5287a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5288a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5289a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5290c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5291bd7c089eSTejun Heo } 5292a9ab775bSTejun Heo 5293a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5294bd7c089eSTejun Heo } 5295bd7c089eSTejun Heo 52967dbc725eSTejun Heo /** 52977dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 52987dbc725eSTejun Heo * @pool: unbound pool of interest 52997dbc725eSTejun Heo * @cpu: the CPU which is coming up 53007dbc725eSTejun Heo * 53017dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 53027dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 53037dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 53047dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 53057dbc725eSTejun Heo */ 53067dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 53077dbc725eSTejun Heo { 53087dbc725eSTejun Heo static cpumask_t cpumask; 53097dbc725eSTejun Heo struct worker *worker; 53107dbc725eSTejun Heo 53111258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 53127dbc725eSTejun Heo 53137dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 53147dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 53157dbc725eSTejun Heo return; 53167dbc725eSTejun Heo 53177dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 53187dbc725eSTejun Heo 53197dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5320da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5321d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 53227dbc725eSTejun Heo } 53237dbc725eSTejun Heo 53247ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 53251da177e4SLinus Torvalds { 53264ce62e9eSTejun Heo struct worker_pool *pool; 53271da177e4SLinus Torvalds 5328f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53293ce63377STejun Heo if (pool->nr_workers) 53303ce63377STejun Heo continue; 5331051e1850SLai Jiangshan if (!create_worker(pool)) 53327ee681b2SThomas Gleixner return -ENOMEM; 53333af24433SOleg Nesterov } 53347ee681b2SThomas Gleixner return 0; 53357ee681b2SThomas Gleixner } 53361da177e4SLinus Torvalds 53377ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 53387ee681b2SThomas Gleixner { 53397ee681b2SThomas Gleixner struct worker_pool *pool; 53407ee681b2SThomas Gleixner struct workqueue_struct *wq; 53417ee681b2SThomas Gleixner int pi; 53427ee681b2SThomas Gleixner 534368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 53447dbc725eSTejun Heo 53457dbc725eSTejun Heo for_each_pool(pool, pi) { 53461258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 534794cf58bbSTejun Heo 5348f05b558dSLai Jiangshan if (pool->cpu == cpu) 534994cf58bbSTejun Heo rebind_workers(pool); 5350f05b558dSLai Jiangshan else if (pool->cpu < 0) 53517dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 535294cf58bbSTejun Heo 53531258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 535494cf58bbSTejun Heo } 53557dbc725eSTejun Heo 53564c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 53574c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 53584c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 53594c16bd32STejun Heo 536068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 53617ee681b2SThomas Gleixner return 0; 536265758202STejun Heo } 536365758202STejun Heo 53647ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 536565758202STejun Heo { 53664c16bd32STejun Heo struct workqueue_struct *wq; 53678db25e78STejun Heo 53684c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5369e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5370e8b3f8dbSLai Jiangshan return -1; 5371e8b3f8dbSLai Jiangshan 5372e8b3f8dbSLai Jiangshan unbind_workers(cpu); 53734c16bd32STejun Heo 53744c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 53754c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 53764c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 53774c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 53784c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 53794c16bd32STejun Heo 53807ee681b2SThomas Gleixner return 0; 538165758202STejun Heo } 538265758202STejun Heo 53832d3854a3SRusty Russell struct work_for_cpu { 5384ed48ece2STejun Heo struct work_struct work; 53852d3854a3SRusty Russell long (*fn)(void *); 53862d3854a3SRusty Russell void *arg; 53872d3854a3SRusty Russell long ret; 53882d3854a3SRusty Russell }; 53892d3854a3SRusty Russell 5390ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 53912d3854a3SRusty Russell { 5392ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5393ed48ece2STejun Heo 53942d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 53952d3854a3SRusty Russell } 53962d3854a3SRusty Russell 53972d3854a3SRusty Russell /** 539822aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 53992d3854a3SRusty Russell * @cpu: the cpu to run on 54002d3854a3SRusty Russell * @fn: the function to run 54012d3854a3SRusty Russell * @arg: the function arg 54022d3854a3SRusty Russell * 540331ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 54046b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5405d185af30SYacine Belkadi * 5406d185af30SYacine Belkadi * Return: The value @fn returns. 54072d3854a3SRusty Russell */ 5408d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 54092d3854a3SRusty Russell { 5410ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 54112d3854a3SRusty Russell 5412ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5413ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 541412997d1aSBjorn Helgaas flush_work(&wfc.work); 5415440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 54162d3854a3SRusty Russell return wfc.ret; 54172d3854a3SRusty Russell } 54182d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 54190e8d6a93SThomas Gleixner 54200e8d6a93SThomas Gleixner /** 54210e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 54220e8d6a93SThomas Gleixner * @cpu: the cpu to run on 54230e8d6a93SThomas Gleixner * @fn: the function to run 54240e8d6a93SThomas Gleixner * @arg: the function argument 54250e8d6a93SThomas Gleixner * 54260e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 54270e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 54280e8d6a93SThomas Gleixner * 54290e8d6a93SThomas Gleixner * Return: The value @fn returns. 54300e8d6a93SThomas Gleixner */ 54310e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 54320e8d6a93SThomas Gleixner { 54330e8d6a93SThomas Gleixner long ret = -ENODEV; 54340e8d6a93SThomas Gleixner 5435ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 54360e8d6a93SThomas Gleixner if (cpu_online(cpu)) 54370e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5438ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 54390e8d6a93SThomas Gleixner return ret; 54400e8d6a93SThomas Gleixner } 54410e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 54422d3854a3SRusty Russell #endif /* CONFIG_SMP */ 54432d3854a3SRusty Russell 5444a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5445e7577c50SRusty Russell 5446a0a1a5fdSTejun Heo /** 5447a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5448a0a1a5fdSTejun Heo * 544958a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5450f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5451706026c2STejun Heo * pool->worklist. 5452a0a1a5fdSTejun Heo * 5453a0a1a5fdSTejun Heo * CONTEXT: 5454a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5455a0a1a5fdSTejun Heo */ 5456a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5457a0a1a5fdSTejun Heo { 545824b8a847STejun Heo struct workqueue_struct *wq; 545924b8a847STejun Heo struct pool_workqueue *pwq; 5460a0a1a5fdSTejun Heo 546168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5462a0a1a5fdSTejun Heo 54636183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5464a0a1a5fdSTejun Heo workqueue_freezing = true; 5465a0a1a5fdSTejun Heo 546624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5467a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5468699ce097STejun Heo for_each_pwq(pwq, wq) 5469699ce097STejun Heo pwq_adjust_max_active(pwq); 5470a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5471a1056305STejun Heo } 54725bcab335STejun Heo 547368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5474a0a1a5fdSTejun Heo } 5475a0a1a5fdSTejun Heo 5476a0a1a5fdSTejun Heo /** 547758a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5478a0a1a5fdSTejun Heo * 5479a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5480a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5481a0a1a5fdSTejun Heo * 5482a0a1a5fdSTejun Heo * CONTEXT: 548368e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5484a0a1a5fdSTejun Heo * 5485d185af30SYacine Belkadi * Return: 548658a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 548758a69cb4STejun Heo * is complete. 5488a0a1a5fdSTejun Heo */ 5489a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5490a0a1a5fdSTejun Heo { 5491a0a1a5fdSTejun Heo bool busy = false; 549224b8a847STejun Heo struct workqueue_struct *wq; 549324b8a847STejun Heo struct pool_workqueue *pwq; 5494a0a1a5fdSTejun Heo 549568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5496a0a1a5fdSTejun Heo 54976183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5498a0a1a5fdSTejun Heo 549924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 550024b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 550124b8a847STejun Heo continue; 5502a0a1a5fdSTejun Heo /* 5503a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5504a0a1a5fdSTejun Heo * to peek without lock. 5505a0a1a5fdSTejun Heo */ 550624acfb71SThomas Gleixner rcu_read_lock(); 550724b8a847STejun Heo for_each_pwq(pwq, wq) { 55086183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5509112202d9STejun Heo if (pwq->nr_active) { 5510a0a1a5fdSTejun Heo busy = true; 551124acfb71SThomas Gleixner rcu_read_unlock(); 5512a0a1a5fdSTejun Heo goto out_unlock; 5513a0a1a5fdSTejun Heo } 5514a0a1a5fdSTejun Heo } 551524acfb71SThomas Gleixner rcu_read_unlock(); 5516a0a1a5fdSTejun Heo } 5517a0a1a5fdSTejun Heo out_unlock: 551868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5519a0a1a5fdSTejun Heo return busy; 5520a0a1a5fdSTejun Heo } 5521a0a1a5fdSTejun Heo 5522a0a1a5fdSTejun Heo /** 5523a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5524a0a1a5fdSTejun Heo * 5525a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5526706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5527a0a1a5fdSTejun Heo * 5528a0a1a5fdSTejun Heo * CONTEXT: 5529a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5530a0a1a5fdSTejun Heo */ 5531a0a1a5fdSTejun Heo void thaw_workqueues(void) 5532a0a1a5fdSTejun Heo { 553324b8a847STejun Heo struct workqueue_struct *wq; 553424b8a847STejun Heo struct pool_workqueue *pwq; 5535a0a1a5fdSTejun Heo 553668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5537a0a1a5fdSTejun Heo 5538a0a1a5fdSTejun Heo if (!workqueue_freezing) 5539a0a1a5fdSTejun Heo goto out_unlock; 5540a0a1a5fdSTejun Heo 554174b414eaSLai Jiangshan workqueue_freezing = false; 554224b8a847STejun Heo 554324b8a847STejun Heo /* restore max_active and repopulate worklist */ 554424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5545a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5546699ce097STejun Heo for_each_pwq(pwq, wq) 5547699ce097STejun Heo pwq_adjust_max_active(pwq); 5548a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 554924b8a847STejun Heo } 555024b8a847STejun Heo 5551a0a1a5fdSTejun Heo out_unlock: 555268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5553a0a1a5fdSTejun Heo } 5554a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5555a0a1a5fdSTejun Heo 555699c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5557042f7df1SLai Jiangshan { 5558042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5559042f7df1SLai Jiangshan int ret = 0; 5560042f7df1SLai Jiangshan struct workqueue_struct *wq; 5561042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5562042f7df1SLai Jiangshan 5563042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5564042f7df1SLai Jiangshan 5565042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5566042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5567042f7df1SLai Jiangshan continue; 5568042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5569042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5570042f7df1SLai Jiangshan continue; 5571042f7df1SLai Jiangshan 557299c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5573042f7df1SLai Jiangshan if (!ctx) { 5574042f7df1SLai Jiangshan ret = -ENOMEM; 5575042f7df1SLai Jiangshan break; 5576042f7df1SLai Jiangshan } 5577042f7df1SLai Jiangshan 5578042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5579042f7df1SLai Jiangshan } 5580042f7df1SLai Jiangshan 5581042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5582042f7df1SLai Jiangshan if (!ret) 5583042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5584042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5585042f7df1SLai Jiangshan } 5586042f7df1SLai Jiangshan 558799c621efSLai Jiangshan if (!ret) { 558899c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 558999c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 559099c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 559199c621efSLai Jiangshan } 5592042f7df1SLai Jiangshan return ret; 5593042f7df1SLai Jiangshan } 5594042f7df1SLai Jiangshan 5595042f7df1SLai Jiangshan /** 5596042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5597042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5598042f7df1SLai Jiangshan * 5599042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5600042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5601042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5602042f7df1SLai Jiangshan * 560367dc8325SCai Huoqing * Return: 0 - Success 5604042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5605042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5606042f7df1SLai Jiangshan */ 5607042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5608042f7df1SLai Jiangshan { 5609042f7df1SLai Jiangshan int ret = -EINVAL; 5610042f7df1SLai Jiangshan 5611c98a9805STal Shorer /* 5612c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5613c98a9805STal Shorer * If the user wishes to include them, we allow that. 5614c98a9805STal Shorer */ 5615042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5616042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5617a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5618d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5619d25302e4SMenglong Dong ret = 0; 5620d25302e4SMenglong Dong goto out_unlock; 5621d25302e4SMenglong Dong } 5622d25302e4SMenglong Dong 562399c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5624042f7df1SLai Jiangshan 5625d25302e4SMenglong Dong out_unlock: 5626a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5627042f7df1SLai Jiangshan } 5628042f7df1SLai Jiangshan 5629042f7df1SLai Jiangshan return ret; 5630042f7df1SLai Jiangshan } 5631042f7df1SLai Jiangshan 56326ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 56336ba94429SFrederic Weisbecker /* 56346ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 56356ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 56366ba94429SFrederic Weisbecker * following attributes. 56376ba94429SFrederic Weisbecker * 56386ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 56396ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 56406ba94429SFrederic Weisbecker * 56416ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 56426ba94429SFrederic Weisbecker * 56439a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 56446ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 56456ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 56469a19b463SWang Long * numa RW bool : whether enable NUMA affinity 56476ba94429SFrederic Weisbecker */ 56486ba94429SFrederic Weisbecker struct wq_device { 56496ba94429SFrederic Weisbecker struct workqueue_struct *wq; 56506ba94429SFrederic Weisbecker struct device dev; 56516ba94429SFrederic Weisbecker }; 56526ba94429SFrederic Weisbecker 56536ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 56546ba94429SFrederic Weisbecker { 56556ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 56566ba94429SFrederic Weisbecker 56576ba94429SFrederic Weisbecker return wq_dev->wq; 56586ba94429SFrederic Weisbecker } 56596ba94429SFrederic Weisbecker 56606ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 56616ba94429SFrederic Weisbecker char *buf) 56626ba94429SFrederic Weisbecker { 56636ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56646ba94429SFrederic Weisbecker 56656ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 56666ba94429SFrederic Weisbecker } 56676ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 56686ba94429SFrederic Weisbecker 56696ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 56706ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 56716ba94429SFrederic Weisbecker { 56726ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56736ba94429SFrederic Weisbecker 56746ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 56756ba94429SFrederic Weisbecker } 56766ba94429SFrederic Weisbecker 56776ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 56786ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 56796ba94429SFrederic Weisbecker size_t count) 56806ba94429SFrederic Weisbecker { 56816ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56826ba94429SFrederic Weisbecker int val; 56836ba94429SFrederic Weisbecker 56846ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 56856ba94429SFrederic Weisbecker return -EINVAL; 56866ba94429SFrederic Weisbecker 56876ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 56886ba94429SFrederic Weisbecker return count; 56896ba94429SFrederic Weisbecker } 56906ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 56916ba94429SFrederic Weisbecker 56926ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 56936ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 56946ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 56956ba94429SFrederic Weisbecker NULL, 56966ba94429SFrederic Weisbecker }; 56976ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 56986ba94429SFrederic Weisbecker 56996ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 57006ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 57016ba94429SFrederic Weisbecker { 57026ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57036ba94429SFrederic Weisbecker const char *delim = ""; 57046ba94429SFrederic Weisbecker int node, written = 0; 57056ba94429SFrederic Weisbecker 5706ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 570724acfb71SThomas Gleixner rcu_read_lock(); 57086ba94429SFrederic Weisbecker for_each_node(node) { 57096ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 57106ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 57116ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 57126ba94429SFrederic Weisbecker delim = " "; 57136ba94429SFrederic Weisbecker } 57146ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 571524acfb71SThomas Gleixner rcu_read_unlock(); 5716ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 57176ba94429SFrederic Weisbecker 57186ba94429SFrederic Weisbecker return written; 57196ba94429SFrederic Weisbecker } 57206ba94429SFrederic Weisbecker 57216ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 57226ba94429SFrederic Weisbecker char *buf) 57236ba94429SFrederic Weisbecker { 57246ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57256ba94429SFrederic Weisbecker int written; 57266ba94429SFrederic Weisbecker 57276ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 57286ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 57296ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 57306ba94429SFrederic Weisbecker 57316ba94429SFrederic Weisbecker return written; 57326ba94429SFrederic Weisbecker } 57336ba94429SFrederic Weisbecker 57346ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 57356ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 57366ba94429SFrederic Weisbecker { 57376ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 57386ba94429SFrederic Weisbecker 5739899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5740899a94feSLai Jiangshan 5741be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 57426ba94429SFrederic Weisbecker if (!attrs) 57436ba94429SFrederic Weisbecker return NULL; 57446ba94429SFrederic Weisbecker 57456ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 57466ba94429SFrederic Weisbecker return attrs; 57476ba94429SFrederic Weisbecker } 57486ba94429SFrederic Weisbecker 57496ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 57506ba94429SFrederic Weisbecker const char *buf, size_t count) 57516ba94429SFrederic Weisbecker { 57526ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57536ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5754d4d3e257SLai Jiangshan int ret = -ENOMEM; 5755d4d3e257SLai Jiangshan 5756d4d3e257SLai Jiangshan apply_wqattrs_lock(); 57576ba94429SFrederic Weisbecker 57586ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 57596ba94429SFrederic Weisbecker if (!attrs) 5760d4d3e257SLai Jiangshan goto out_unlock; 57616ba94429SFrederic Weisbecker 57626ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 57636ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5764d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 57656ba94429SFrederic Weisbecker else 57666ba94429SFrederic Weisbecker ret = -EINVAL; 57676ba94429SFrederic Weisbecker 5768d4d3e257SLai Jiangshan out_unlock: 5769d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 57706ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 57716ba94429SFrederic Weisbecker return ret ?: count; 57726ba94429SFrederic Weisbecker } 57736ba94429SFrederic Weisbecker 57746ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 57756ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 57766ba94429SFrederic Weisbecker { 57776ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57786ba94429SFrederic Weisbecker int written; 57796ba94429SFrederic Weisbecker 57806ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 57816ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 57826ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 57836ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 57846ba94429SFrederic Weisbecker return written; 57856ba94429SFrederic Weisbecker } 57866ba94429SFrederic Weisbecker 57876ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 57886ba94429SFrederic Weisbecker struct device_attribute *attr, 57896ba94429SFrederic Weisbecker const char *buf, size_t count) 57906ba94429SFrederic Weisbecker { 57916ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57926ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5793d4d3e257SLai Jiangshan int ret = -ENOMEM; 5794d4d3e257SLai Jiangshan 5795d4d3e257SLai Jiangshan apply_wqattrs_lock(); 57966ba94429SFrederic Weisbecker 57976ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 57986ba94429SFrederic Weisbecker if (!attrs) 5799d4d3e257SLai Jiangshan goto out_unlock; 58006ba94429SFrederic Weisbecker 58016ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 58026ba94429SFrederic Weisbecker if (!ret) 5803d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 58046ba94429SFrederic Weisbecker 5805d4d3e257SLai Jiangshan out_unlock: 5806d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 58076ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 58086ba94429SFrederic Weisbecker return ret ?: count; 58096ba94429SFrederic Weisbecker } 58106ba94429SFrederic Weisbecker 58116ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 58126ba94429SFrederic Weisbecker char *buf) 58136ba94429SFrederic Weisbecker { 58146ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58156ba94429SFrederic Weisbecker int written; 58166ba94429SFrederic Weisbecker 58176ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 58186ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 58196ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 58206ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 58216ba94429SFrederic Weisbecker 58226ba94429SFrederic Weisbecker return written; 58236ba94429SFrederic Weisbecker } 58246ba94429SFrederic Weisbecker 58256ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 58266ba94429SFrederic Weisbecker const char *buf, size_t count) 58276ba94429SFrederic Weisbecker { 58286ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58296ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5830d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5831d4d3e257SLai Jiangshan 5832d4d3e257SLai Jiangshan apply_wqattrs_lock(); 58336ba94429SFrederic Weisbecker 58346ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 58356ba94429SFrederic Weisbecker if (!attrs) 5836d4d3e257SLai Jiangshan goto out_unlock; 58376ba94429SFrederic Weisbecker 58386ba94429SFrederic Weisbecker ret = -EINVAL; 58396ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 58406ba94429SFrederic Weisbecker attrs->no_numa = !v; 5841d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 58426ba94429SFrederic Weisbecker } 58436ba94429SFrederic Weisbecker 5844d4d3e257SLai Jiangshan out_unlock: 5845d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 58466ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 58476ba94429SFrederic Weisbecker return ret ?: count; 58486ba94429SFrederic Weisbecker } 58496ba94429SFrederic Weisbecker 58506ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 58516ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 58526ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 58536ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 58546ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 58556ba94429SFrederic Weisbecker __ATTR_NULL, 58566ba94429SFrederic Weisbecker }; 58576ba94429SFrederic Weisbecker 58586ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 58596ba94429SFrederic Weisbecker .name = "workqueue", 58606ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 58616ba94429SFrederic Weisbecker }; 58626ba94429SFrederic Weisbecker 5863b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5864b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5865b05a7928SFrederic Weisbecker { 5866b05a7928SFrederic Weisbecker int written; 5867b05a7928SFrederic Weisbecker 5868042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5869b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5870b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5871042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5872b05a7928SFrederic Weisbecker 5873b05a7928SFrederic Weisbecker return written; 5874b05a7928SFrederic Weisbecker } 5875b05a7928SFrederic Weisbecker 5876042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5877042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5878042f7df1SLai Jiangshan { 5879042f7df1SLai Jiangshan cpumask_var_t cpumask; 5880042f7df1SLai Jiangshan int ret; 5881042f7df1SLai Jiangshan 5882042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5883042f7df1SLai Jiangshan return -ENOMEM; 5884042f7df1SLai Jiangshan 5885042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5886042f7df1SLai Jiangshan if (!ret) 5887042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5888042f7df1SLai Jiangshan 5889042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5890042f7df1SLai Jiangshan return ret ? ret : count; 5891042f7df1SLai Jiangshan } 5892042f7df1SLai Jiangshan 5893b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5894042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5895042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5896b05a7928SFrederic Weisbecker 58976ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 58986ba94429SFrederic Weisbecker { 5899686f6697SGreg Kroah-Hartman struct device *dev_root; 5900b05a7928SFrederic Weisbecker int err; 5901b05a7928SFrederic Weisbecker 5902b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5903b05a7928SFrederic Weisbecker if (err) 5904b05a7928SFrederic Weisbecker return err; 5905b05a7928SFrederic Weisbecker 5906686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 5907686f6697SGreg Kroah-Hartman if (dev_root) { 5908686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 5909686f6697SGreg Kroah-Hartman put_device(dev_root); 5910686f6697SGreg Kroah-Hartman } 5911686f6697SGreg Kroah-Hartman return err; 59126ba94429SFrederic Weisbecker } 59136ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 59146ba94429SFrederic Weisbecker 59156ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 59166ba94429SFrederic Weisbecker { 59176ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 59186ba94429SFrederic Weisbecker 59196ba94429SFrederic Weisbecker kfree(wq_dev); 59206ba94429SFrederic Weisbecker } 59216ba94429SFrederic Weisbecker 59226ba94429SFrederic Weisbecker /** 59236ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 59246ba94429SFrederic Weisbecker * @wq: the workqueue to register 59256ba94429SFrederic Weisbecker * 59266ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 59276ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 59286ba94429SFrederic Weisbecker * which is the preferred method. 59296ba94429SFrederic Weisbecker * 59306ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 59316ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 59326ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 59336ba94429SFrederic Weisbecker * attributes. 59346ba94429SFrederic Weisbecker * 59356ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 59366ba94429SFrederic Weisbecker */ 59376ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 59386ba94429SFrederic Weisbecker { 59396ba94429SFrederic Weisbecker struct wq_device *wq_dev; 59406ba94429SFrederic Weisbecker int ret; 59416ba94429SFrederic Weisbecker 59426ba94429SFrederic Weisbecker /* 5943402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 59446ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 59456ba94429SFrederic Weisbecker * workqueues. 59466ba94429SFrederic Weisbecker */ 59470a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 59486ba94429SFrederic Weisbecker return -EINVAL; 59496ba94429SFrederic Weisbecker 59506ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 59516ba94429SFrederic Weisbecker if (!wq_dev) 59526ba94429SFrederic Weisbecker return -ENOMEM; 59536ba94429SFrederic Weisbecker 59546ba94429SFrederic Weisbecker wq_dev->wq = wq; 59556ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 59566ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 595723217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 59586ba94429SFrederic Weisbecker 59596ba94429SFrederic Weisbecker /* 59606ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 59616ba94429SFrederic Weisbecker * everything is ready. 59626ba94429SFrederic Weisbecker */ 59636ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 59646ba94429SFrederic Weisbecker 59656ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 59666ba94429SFrederic Weisbecker if (ret) { 5967537f4146SArvind Yadav put_device(&wq_dev->dev); 59686ba94429SFrederic Weisbecker wq->wq_dev = NULL; 59696ba94429SFrederic Weisbecker return ret; 59706ba94429SFrederic Weisbecker } 59716ba94429SFrederic Weisbecker 59726ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 59736ba94429SFrederic Weisbecker struct device_attribute *attr; 59746ba94429SFrederic Weisbecker 59756ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 59766ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 59776ba94429SFrederic Weisbecker if (ret) { 59786ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 59796ba94429SFrederic Weisbecker wq->wq_dev = NULL; 59806ba94429SFrederic Weisbecker return ret; 59816ba94429SFrederic Weisbecker } 59826ba94429SFrederic Weisbecker } 59836ba94429SFrederic Weisbecker } 59846ba94429SFrederic Weisbecker 59856ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 59866ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 59876ba94429SFrederic Weisbecker return 0; 59886ba94429SFrederic Weisbecker } 59896ba94429SFrederic Weisbecker 59906ba94429SFrederic Weisbecker /** 59916ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 59926ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 59936ba94429SFrederic Weisbecker * 59946ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 59956ba94429SFrederic Weisbecker */ 59966ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 59976ba94429SFrederic Weisbecker { 59986ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 59996ba94429SFrederic Weisbecker 60006ba94429SFrederic Weisbecker if (!wq->wq_dev) 60016ba94429SFrederic Weisbecker return; 60026ba94429SFrederic Weisbecker 60036ba94429SFrederic Weisbecker wq->wq_dev = NULL; 60046ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 60056ba94429SFrederic Weisbecker } 60066ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 60076ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 60086ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 60096ba94429SFrederic Weisbecker 601082607adcSTejun Heo /* 601182607adcSTejun Heo * Workqueue watchdog. 601282607adcSTejun Heo * 601382607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 601482607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 601582607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 601682607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 601782607adcSTejun Heo * largely opaque. 601882607adcSTejun Heo * 601982607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 602082607adcSTejun Heo * state if some pools failed to make forward progress for a while where 602182607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 602282607adcSTejun Heo * 602382607adcSTejun Heo * This mechanism is controlled through the kernel parameter 602482607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 602582607adcSTejun Heo * corresponding sysfs parameter file. 602682607adcSTejun Heo */ 602782607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 602882607adcSTejun Heo 602982607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 60305cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 603182607adcSTejun Heo 603282607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 603382607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 603482607adcSTejun Heo 6035cd2440d6SPetr Mladek /* 6036cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6037cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6038cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6039cd2440d6SPetr Mladek * in all other situations. 6040cd2440d6SPetr Mladek */ 6041cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6042cd2440d6SPetr Mladek { 6043cd2440d6SPetr Mladek struct worker *worker; 6044cd2440d6SPetr Mladek unsigned long flags; 6045cd2440d6SPetr Mladek int bkt; 6046cd2440d6SPetr Mladek 6047cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6048cd2440d6SPetr Mladek 6049cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6050cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6051cd2440d6SPetr Mladek /* 6052cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6053cd2440d6SPetr Mladek * drivers that queue work while holding locks 6054cd2440d6SPetr Mladek * also taken in their write paths. 6055cd2440d6SPetr Mladek */ 6056cd2440d6SPetr Mladek printk_deferred_enter(); 6057cd2440d6SPetr Mladek 6058cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6059cd2440d6SPetr Mladek sched_show_task(worker->task); 6060cd2440d6SPetr Mladek 6061cd2440d6SPetr Mladek printk_deferred_exit(); 6062cd2440d6SPetr Mladek } 6063cd2440d6SPetr Mladek } 6064cd2440d6SPetr Mladek 6065cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6066cd2440d6SPetr Mladek } 6067cd2440d6SPetr Mladek 6068cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6069cd2440d6SPetr Mladek { 6070cd2440d6SPetr Mladek struct worker_pool *pool; 6071cd2440d6SPetr Mladek int pi; 6072cd2440d6SPetr Mladek 6073cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6074cd2440d6SPetr Mladek 6075cd2440d6SPetr Mladek rcu_read_lock(); 6076cd2440d6SPetr Mladek 6077cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6078cd2440d6SPetr Mladek if (pool->cpu_stall) 6079cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6080cd2440d6SPetr Mladek 6081cd2440d6SPetr Mladek } 6082cd2440d6SPetr Mladek 6083cd2440d6SPetr Mladek rcu_read_unlock(); 6084cd2440d6SPetr Mladek } 6085cd2440d6SPetr Mladek 608682607adcSTejun Heo static void wq_watchdog_reset_touched(void) 608782607adcSTejun Heo { 608882607adcSTejun Heo int cpu; 608982607adcSTejun Heo 609082607adcSTejun Heo wq_watchdog_touched = jiffies; 609182607adcSTejun Heo for_each_possible_cpu(cpu) 609282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 609382607adcSTejun Heo } 609482607adcSTejun Heo 60955cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 609682607adcSTejun Heo { 609782607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 609882607adcSTejun Heo bool lockup_detected = false; 6099cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6100940d71c6SSergey Senozhatsky unsigned long now = jiffies; 610182607adcSTejun Heo struct worker_pool *pool; 610282607adcSTejun Heo int pi; 610382607adcSTejun Heo 610482607adcSTejun Heo if (!thresh) 610582607adcSTejun Heo return; 610682607adcSTejun Heo 610782607adcSTejun Heo rcu_read_lock(); 610882607adcSTejun Heo 610982607adcSTejun Heo for_each_pool(pool, pi) { 611082607adcSTejun Heo unsigned long pool_ts, touched, ts; 611182607adcSTejun Heo 6112cd2440d6SPetr Mladek pool->cpu_stall = false; 611382607adcSTejun Heo if (list_empty(&pool->worklist)) 611482607adcSTejun Heo continue; 611582607adcSTejun Heo 6116940d71c6SSergey Senozhatsky /* 6117940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6118940d71c6SSergey Senozhatsky * the watchdog like a stall. 6119940d71c6SSergey Senozhatsky */ 6120940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6121940d71c6SSergey Senozhatsky 612282607adcSTejun Heo /* get the latest of pool and touched timestamps */ 612389e28ce6SWang Qing if (pool->cpu >= 0) 612489e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 612589e28ce6SWang Qing else 612682607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 612789e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 612882607adcSTejun Heo 612982607adcSTejun Heo if (time_after(pool_ts, touched)) 613082607adcSTejun Heo ts = pool_ts; 613182607adcSTejun Heo else 613282607adcSTejun Heo ts = touched; 613382607adcSTejun Heo 613482607adcSTejun Heo /* did we stall? */ 6135940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 613682607adcSTejun Heo lockup_detected = true; 6137cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6138cd2440d6SPetr Mladek pool->cpu_stall = true; 6139cd2440d6SPetr Mladek cpu_pool_stall = true; 6140cd2440d6SPetr Mladek } 614182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 614282607adcSTejun Heo pr_cont_pool_info(pool); 614382607adcSTejun Heo pr_cont(" stuck for %us!\n", 6144940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 614582607adcSTejun Heo } 6146cd2440d6SPetr Mladek 6147cd2440d6SPetr Mladek 614882607adcSTejun Heo } 614982607adcSTejun Heo 615082607adcSTejun Heo rcu_read_unlock(); 615182607adcSTejun Heo 615282607adcSTejun Heo if (lockup_detected) 615355df0933SImran Khan show_all_workqueues(); 615482607adcSTejun Heo 6155cd2440d6SPetr Mladek if (cpu_pool_stall) 6156cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6157cd2440d6SPetr Mladek 615882607adcSTejun Heo wq_watchdog_reset_touched(); 615982607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 616082607adcSTejun Heo } 616182607adcSTejun Heo 6162cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 616382607adcSTejun Heo { 616482607adcSTejun Heo if (cpu >= 0) 616582607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 616689e28ce6SWang Qing 616782607adcSTejun Heo wq_watchdog_touched = jiffies; 616882607adcSTejun Heo } 616982607adcSTejun Heo 617082607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 617182607adcSTejun Heo { 617282607adcSTejun Heo wq_watchdog_thresh = 0; 617382607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 617482607adcSTejun Heo 617582607adcSTejun Heo if (thresh) { 617682607adcSTejun Heo wq_watchdog_thresh = thresh; 617782607adcSTejun Heo wq_watchdog_reset_touched(); 617882607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 617982607adcSTejun Heo } 618082607adcSTejun Heo } 618182607adcSTejun Heo 618282607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 618382607adcSTejun Heo const struct kernel_param *kp) 618482607adcSTejun Heo { 618582607adcSTejun Heo unsigned long thresh; 618682607adcSTejun Heo int ret; 618782607adcSTejun Heo 618882607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 618982607adcSTejun Heo if (ret) 619082607adcSTejun Heo return ret; 619182607adcSTejun Heo 619282607adcSTejun Heo if (system_wq) 619382607adcSTejun Heo wq_watchdog_set_thresh(thresh); 619482607adcSTejun Heo else 619582607adcSTejun Heo wq_watchdog_thresh = thresh; 619682607adcSTejun Heo 619782607adcSTejun Heo return 0; 619882607adcSTejun Heo } 619982607adcSTejun Heo 620082607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 620182607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 620282607adcSTejun Heo .get = param_get_ulong, 620382607adcSTejun Heo }; 620482607adcSTejun Heo 620582607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 620682607adcSTejun Heo 0644); 620782607adcSTejun Heo 620882607adcSTejun Heo static void wq_watchdog_init(void) 620982607adcSTejun Heo { 62105cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 621182607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 621282607adcSTejun Heo } 621382607adcSTejun Heo 621482607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 621582607adcSTejun Heo 621682607adcSTejun Heo static inline void wq_watchdog_init(void) { } 621782607adcSTejun Heo 621882607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 621982607adcSTejun Heo 6220bce90380STejun Heo static void __init wq_numa_init(void) 6221bce90380STejun Heo { 6222bce90380STejun Heo cpumask_var_t *tbl; 6223bce90380STejun Heo int node, cpu; 6224bce90380STejun Heo 6225bce90380STejun Heo if (num_possible_nodes() <= 1) 6226bce90380STejun Heo return; 6227bce90380STejun Heo 6228d55262c4STejun Heo if (wq_disable_numa) { 6229d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 6230d55262c4STejun Heo return; 6231d55262c4STejun Heo } 6232d55262c4STejun Heo 6233f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6234f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6235f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6236f728c4a9SZhen Lei return; 6237f728c4a9SZhen Lei } 6238f728c4a9SZhen Lei } 6239f728c4a9SZhen Lei 6240be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 62414c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 62424c16bd32STejun Heo 6243bce90380STejun Heo /* 6244bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6245bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6246bce90380STejun Heo * fully initialized by now. 6247bce90380STejun Heo */ 62486396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6249bce90380STejun Heo BUG_ON(!tbl); 6250bce90380STejun Heo 6251bce90380STejun Heo for_each_node(node) 62525a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 62531be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6254bce90380STejun Heo 6255bce90380STejun Heo for_each_possible_cpu(cpu) { 6256bce90380STejun Heo node = cpu_to_node(cpu); 6257bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6258bce90380STejun Heo } 6259bce90380STejun Heo 6260bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6261bce90380STejun Heo wq_numa_enabled = true; 6262bce90380STejun Heo } 6263bce90380STejun Heo 62643347fa09STejun Heo /** 62653347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 62663347fa09STejun Heo * 62673347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 62683347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 62693347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 62703347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 62713347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 62723347fa09STejun Heo * created and scheduled right before early initcalls. 62733347fa09STejun Heo */ 62742333e829SYu Chen void __init workqueue_init_early(void) 62751da177e4SLinus Torvalds { 62767a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 62777a4e344cSTejun Heo int i, cpu; 6278c34056a3STejun Heo 627910cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6280e904e6c2STejun Heo 6281b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 628204d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 628304d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6284b05a7928SFrederic Weisbecker 6285e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6286e904e6c2STejun Heo 6287706026c2STejun Heo /* initialize CPU pools */ 628829c91e99STejun Heo for_each_possible_cpu(cpu) { 62894ce62e9eSTejun Heo struct worker_pool *pool; 62908b03ae3cSTejun Heo 62917a4e344cSTejun Heo i = 0; 6292f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 62937a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6294ec22ca5eSTejun Heo pool->cpu = cpu; 62957a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 62967a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6297f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 62987a4e344cSTejun Heo 62999daf9e67STejun Heo /* alloc pool ID */ 630068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 63019daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 630268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 63034ce62e9eSTejun Heo } 63048b03ae3cSTejun Heo } 63058b03ae3cSTejun Heo 63068a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 630729c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 630829c91e99STejun Heo struct workqueue_attrs *attrs; 630929c91e99STejun Heo 6310be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 631129c91e99STejun Heo attrs->nice = std_nice[i]; 631229c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 63138a2b7538STejun Heo 63148a2b7538STejun Heo /* 63158a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 63168a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 63178a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 63188a2b7538STejun Heo */ 6319be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 63208a2b7538STejun Heo attrs->nice = std_nice[i]; 63218a2b7538STejun Heo attrs->no_numa = true; 63228a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 632329c91e99STejun Heo } 632429c91e99STejun Heo 6325d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 63261aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6327d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6328f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6329f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 633024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 633124d51addSTejun Heo WQ_FREEZABLE, 0); 63320668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 63330668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 63340668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 63350668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 63360668106cSViresh Kumar 0); 63371aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 63380668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 63390668106cSViresh Kumar !system_power_efficient_wq || 63400668106cSViresh Kumar !system_freezable_power_efficient_wq); 63413347fa09STejun Heo } 63423347fa09STejun Heo 63433347fa09STejun Heo /** 63443347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 63453347fa09STejun Heo * 63463347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 63473347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 63483347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 63493347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 63503347fa09STejun Heo * with the initial workers and enable future kworker creations. 63513347fa09STejun Heo */ 63522333e829SYu Chen void __init workqueue_init(void) 63533347fa09STejun Heo { 63542186d9f9STejun Heo struct workqueue_struct *wq; 63553347fa09STejun Heo struct worker_pool *pool; 63563347fa09STejun Heo int cpu, bkt; 63573347fa09STejun Heo 63582186d9f9STejun Heo /* 63592186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 63602186d9f9STejun Heo * CPU to node mapping may not be available that early on some 63612186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 63622186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 63632186d9f9STejun Heo * affinity, fix them up. 636440c17f75STejun Heo * 636540c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 63662186d9f9STejun Heo */ 63672186d9f9STejun Heo wq_numa_init(); 63682186d9f9STejun Heo 63692186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 63702186d9f9STejun Heo 63712186d9f9STejun Heo for_each_possible_cpu(cpu) { 63722186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 63732186d9f9STejun Heo pool->node = cpu_to_node(cpu); 63742186d9f9STejun Heo } 63752186d9f9STejun Heo } 63762186d9f9STejun Heo 637740c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 63782186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 637940c17f75STejun Heo WARN(init_rescuer(wq), 638040c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 638140c17f75STejun Heo wq->name); 638240c17f75STejun Heo } 63832186d9f9STejun Heo 63842186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 63852186d9f9STejun Heo 63863347fa09STejun Heo /* create the initial workers */ 63873347fa09STejun Heo for_each_online_cpu(cpu) { 63883347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 63893347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 63903347fa09STejun Heo BUG_ON(!create_worker(pool)); 63913347fa09STejun Heo } 63923347fa09STejun Heo } 63933347fa09STejun Heo 63943347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 63953347fa09STejun Heo BUG_ON(!create_worker(pool)); 63963347fa09STejun Heo 63973347fa09STejun Heo wq_online = true; 639882607adcSTejun Heo wq_watchdog_init(); 63991da177e4SLinus Torvalds } 6400c4f135d6STetsuo Handa 6401c4f135d6STetsuo Handa /* 6402c4f135d6STetsuo Handa * Despite the naming, this is a no-op function which is here only for avoiding 6403c4f135d6STetsuo Handa * link error. Since compile-time warning may fail to catch, we will need to 6404c4f135d6STetsuo Handa * emit run-time warning from __flush_workqueue(). 6405c4f135d6STetsuo Handa */ 6406c4f135d6STetsuo Handa void __warn_flushing_systemwide_wq(void) { } 6407c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6408