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 * 129bdf8b9bfSTejun Heo * K: Only modified by worker while holding pool->lock. Can be safely read by 130bdf8b9bfSTejun Heo * self, while holding pool->lock or from IRQ context if %current is the 131bdf8b9bfSTejun Heo * kworker. 132bdf8b9bfSTejun Heo * 133bdf8b9bfSTejun Heo * S: Only modified by worker self. 134bdf8b9bfSTejun Heo * 1351258fae7STejun Heo * A: wq_pool_attach_mutex protected. 136822d8405STejun Heo * 13768e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13876af4d93STejun Heo * 13924acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1405bcab335STejun Heo * 1415b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1425b95e1afSLai Jiangshan * 1435b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 14424acfb71SThomas Gleixner * RCU for reads. 1455b95e1afSLai Jiangshan * 1463c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1473c25a55dSLai Jiangshan * 14824acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1492e109a28STejun Heo * 1502e109a28STejun Heo * MD: wq_mayday_lock protected. 151cd2440d6SPetr Mladek * 152cd2440d6SPetr Mladek * WD: Used internally by the watchdog. 1534690c4abSTejun Heo */ 1544690c4abSTejun Heo 1552eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 156c34056a3STejun Heo 157bd7bdd43STejun Heo struct worker_pool { 158a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 159d84ff051STejun Heo int cpu; /* I: the associated cpu */ 160f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1619daf9e67STejun Heo int id; /* I: pool ID */ 16211ebea50STejun Heo unsigned int flags; /* X: flags */ 163bd7bdd43STejun Heo 16482607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 165cd2440d6SPetr Mladek bool cpu_stall; /* WD: stalled cpu bound pool */ 16682607adcSTejun Heo 167bc35f7efSLai Jiangshan /* 168bc35f7efSLai Jiangshan * The counter is incremented in a process context on the associated CPU 169bc35f7efSLai Jiangshan * w/ preemption disabled, and decremented or reset in the same context 170bc35f7efSLai Jiangshan * but w/ pool->lock held. The readers grab pool->lock and are 171bc35f7efSLai Jiangshan * guaranteed to see if the counter reached zero. 172bc35f7efSLai Jiangshan */ 173bc35f7efSLai Jiangshan int nr_running; 17484f91c62SLai Jiangshan 175bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 176ea1abd61SLai Jiangshan 1775826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1785826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 179bd7bdd43STejun Heo 1802c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 181bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 1823f959aa3SValentin Schneider struct work_struct idle_cull_work; /* L: worker idle cleanup */ 1833f959aa3SValentin Schneider 184bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 185bd7bdd43STejun Heo 186c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 187c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 188c9e7cf27STejun Heo /* L: hash of busy workers */ 189c9e7cf27STejun Heo 1902607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 19192f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 192e02b9312SValentin Schneider struct list_head dying_workers; /* A: workers about to die */ 19360f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 194e19e397aSTejun Heo 1957cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 196e19e397aSTejun Heo 1977a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 19868e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 19968e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 2007a4e344cSTejun Heo 201e19e397aSTejun Heo /* 20224acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 20329c91e99STejun Heo * from get_work_pool(). 20429c91e99STejun Heo */ 20529c91e99STejun Heo struct rcu_head rcu; 20684f91c62SLai Jiangshan }; 2078b03ae3cSTejun Heo 2088b03ae3cSTejun Heo /* 209725e8ec5STejun Heo * Per-pool_workqueue statistics. These can be monitored using 210725e8ec5STejun Heo * tools/workqueue/wq_monitor.py. 211725e8ec5STejun Heo */ 212725e8ec5STejun Heo enum pool_workqueue_stats { 213725e8ec5STejun Heo PWQ_STAT_STARTED, /* work items started execution */ 214725e8ec5STejun Heo PWQ_STAT_COMPLETED, /* work items completed execution */ 2158a1dd1e5STejun Heo PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 216616db877STejun Heo PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 217725e8ec5STejun Heo PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 218725e8ec5STejun Heo PWQ_STAT_MAYDAY, /* maydays to rescuer */ 219725e8ec5STejun Heo PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 220725e8ec5STejun Heo 221725e8ec5STejun Heo PWQ_NR_STATS, 222725e8ec5STejun Heo }; 223725e8ec5STejun Heo 224725e8ec5STejun Heo /* 225112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 226112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 227112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 228112202d9STejun Heo * number of flag bits. 2291da177e4SLinus Torvalds */ 230112202d9STejun Heo struct pool_workqueue { 231bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2324690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 23373f53c4aSTejun Heo int work_color; /* L: current color */ 23473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2358864b4e5STejun Heo int refcnt; /* L: reference count */ 23673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 23773f53c4aSTejun Heo /* L: nr of in_flight works */ 238018f3a13SLai Jiangshan 239018f3a13SLai Jiangshan /* 240018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 241018f3a13SLai Jiangshan * 242018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 243018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 244018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 245018f3a13SLai Jiangshan * 246018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 247018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 248018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 249018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 250018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 251018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 252018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 253018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 254018f3a13SLai Jiangshan */ 2551e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 256a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 257f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2583c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2592e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2608864b4e5STejun Heo 261725e8ec5STejun Heo u64 stats[PWQ_NR_STATS]; 262725e8ec5STejun Heo 2638864b4e5STejun Heo /* 2648864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2658864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 26624acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 267b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2688864b4e5STejun Heo */ 2698864b4e5STejun Heo struct work_struct unbound_release_work; 2708864b4e5STejun Heo struct rcu_head rcu; 271e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2721da177e4SLinus Torvalds 2731da177e4SLinus Torvalds /* 27473f53c4aSTejun Heo * Structure used to wait for workqueue flush. 27573f53c4aSTejun Heo */ 27673f53c4aSTejun Heo struct wq_flusher { 2773c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2783c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 27973f53c4aSTejun Heo struct completion done; /* flush completion */ 28073f53c4aSTejun Heo }; 2811da177e4SLinus Torvalds 282226223abSTejun Heo struct wq_device; 283226223abSTejun Heo 28473f53c4aSTejun Heo /* 285c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 286c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2871da177e4SLinus Torvalds */ 2881da177e4SLinus Torvalds struct workqueue_struct { 2893c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 290e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 29173f53c4aSTejun Heo 2923c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2933c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2943c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 295112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2963c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2973c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2983c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 29973f53c4aSTejun Heo 3002e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 30130ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 302e22bee78STejun Heo 30387fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 304a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 305226223abSTejun Heo 3065b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 3075b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 3086029a918STejun Heo 309226223abSTejun Heo #ifdef CONFIG_SYSFS 310226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 311226223abSTejun Heo #endif 3124e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 313669de8bdSBart Van Assche char *lock_name; 314669de8bdSBart Van Assche struct lock_class_key key; 3154e6045f1SJohannes Berg struct lockdep_map lockdep_map; 3164e6045f1SJohannes Berg #endif 317ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 3182728fd2fSTejun Heo 319e2dca7adSTejun Heo /* 32024acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 32124acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 322e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 323e2dca7adSTejun Heo */ 324e2dca7adSTejun Heo struct rcu_head rcu; 325e2dca7adSTejun Heo 3262728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 3272728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 3282728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 3295b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 3301da177e4SLinus Torvalds }; 3311da177e4SLinus Torvalds 332e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 333e904e6c2STejun Heo 334bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 335bce90380STejun Heo /* possible CPUs of each node */ 336bce90380STejun Heo 337616db877STejun Heo /* 338616db877STejun Heo * Per-cpu work items which run for longer than the following threshold are 339616db877STejun Heo * automatically considered CPU intensive and excluded from concurrency 340616db877STejun Heo * management to prevent them from noticeably delaying other per-cpu work items. 341616db877STejun Heo */ 342616db877STejun Heo static unsigned long wq_cpu_intensive_thresh_us = 10000; 343616db877STejun Heo module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 344616db877STejun Heo 345d55262c4STejun Heo static bool wq_disable_numa; 346d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 347d55262c4STejun Heo 348cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 349552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 350cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 351cee22a15SViresh Kumar 352863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3533347fa09STejun Heo 354bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 355bce90380STejun Heo 3564c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3574c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3584c16bd32STejun Heo 35968e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3601258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 361a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 362d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 363d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3645bcab335STejun Heo 365e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 36668e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3677d19c5ceSTejun Heo 36899c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 369ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 370ef557180SMike Galbraith 371ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 372ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 373b05a7928SFrederic Weisbecker 374f303fccbSTejun Heo /* 375f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 376f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 377f303fccbSTejun Heo * to uncover usages which depend on it. 378f303fccbSTejun Heo */ 379f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 380f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 381f303fccbSTejun Heo #else 382f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 383f303fccbSTejun Heo #endif 384f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 385f303fccbSTejun Heo 3867d19c5ceSTejun Heo /* the per-cpu worker pools */ 38725528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3887d19c5ceSTejun Heo 38968e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3907d19c5ceSTejun Heo 39168e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 39229c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 39329c91e99STejun Heo 394c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 39529c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 39629c91e99STejun Heo 3978a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3988a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3998a2b7538STejun Heo 400d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 401ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 402044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 4031aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 404044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 405d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 406044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 407f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 408044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 40924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 4100668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 4110668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 4120668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 4130668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 414d320c038STejun Heo 4157d19c5ceSTejun Heo static int worker_thread(void *__worker); 4166ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 417c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 41855df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4197d19c5ceSTejun Heo 42097bd2347STejun Heo #define CREATE_TRACE_POINTS 42197bd2347STejun Heo #include <trace/events/workqueue.h> 42297bd2347STejun Heo 42368e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 42424acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 425f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 42624acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4275bcab335STejun Heo 4285b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 42924acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 430f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 431f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 43224acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4335b95e1afSLai Jiangshan 434f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 435f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 436f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4377a62c2c8STejun Heo (pool)++) 4384ce62e9eSTejun Heo 43949e3cf44STejun Heo /** 44017116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 44117116969STejun Heo * @pool: iteration cursor 442611c92a0STejun Heo * @pi: integer used for iteration 443fa1b54e6STejun Heo * 44424acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 44568e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 44668e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 447fa1b54e6STejun Heo * 448fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 449fa1b54e6STejun Heo * ignored. 45017116969STejun Heo */ 451611c92a0STejun Heo #define for_each_pool(pool, pi) \ 452611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 45368e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 454fa1b54e6STejun Heo else 45517116969STejun Heo 45617116969STejun Heo /** 457822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 458822d8405STejun Heo * @worker: iteration cursor 459822d8405STejun Heo * @pool: worker_pool to iterate workers of 460822d8405STejun Heo * 4611258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 462822d8405STejun Heo * 463822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 464822d8405STejun Heo * ignored. 465822d8405STejun Heo */ 466da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 467da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4681258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 469822d8405STejun Heo else 470822d8405STejun Heo 471822d8405STejun Heo /** 47249e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 47349e3cf44STejun Heo * @pwq: iteration cursor 47449e3cf44STejun Heo * @wq: the target workqueue 47576af4d93STejun Heo * 47624acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 477794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 478794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 47976af4d93STejun Heo * 48076af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 48176af4d93STejun Heo * ignored. 48249e3cf44STejun Heo */ 48349e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 48449e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4855a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 486f3421797STejun Heo 487dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 488dc186ad7SThomas Gleixner 489f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 490dc186ad7SThomas Gleixner 49199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 49299777288SStanislaw Gruszka { 49399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 49499777288SStanislaw Gruszka } 49599777288SStanislaw Gruszka 496b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 497b9fdac7fSDu, Changbin { 498b9fdac7fSDu, Changbin struct work_struct *work = addr; 499b9fdac7fSDu, Changbin 500b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 501b9fdac7fSDu, Changbin } 502b9fdac7fSDu, Changbin 503dc186ad7SThomas Gleixner /* 504dc186ad7SThomas Gleixner * fixup_init is called when: 505dc186ad7SThomas Gleixner * - an active object is initialized 506dc186ad7SThomas Gleixner */ 50702a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 508dc186ad7SThomas Gleixner { 509dc186ad7SThomas Gleixner struct work_struct *work = addr; 510dc186ad7SThomas Gleixner 511dc186ad7SThomas Gleixner switch (state) { 512dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 513dc186ad7SThomas Gleixner cancel_work_sync(work); 514dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 51502a982a6SDu, Changbin return true; 516dc186ad7SThomas Gleixner default: 51702a982a6SDu, Changbin return false; 518dc186ad7SThomas Gleixner } 519dc186ad7SThomas Gleixner } 520dc186ad7SThomas Gleixner 521dc186ad7SThomas Gleixner /* 522dc186ad7SThomas Gleixner * fixup_free is called when: 523dc186ad7SThomas Gleixner * - an active object is freed 524dc186ad7SThomas Gleixner */ 52502a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 526dc186ad7SThomas Gleixner { 527dc186ad7SThomas Gleixner struct work_struct *work = addr; 528dc186ad7SThomas Gleixner 529dc186ad7SThomas Gleixner switch (state) { 530dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 531dc186ad7SThomas Gleixner cancel_work_sync(work); 532dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 53302a982a6SDu, Changbin return true; 534dc186ad7SThomas Gleixner default: 53502a982a6SDu, Changbin return false; 536dc186ad7SThomas Gleixner } 537dc186ad7SThomas Gleixner } 538dc186ad7SThomas Gleixner 539f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 540dc186ad7SThomas Gleixner .name = "work_struct", 54199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 542b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 543dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 544dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 545dc186ad7SThomas Gleixner }; 546dc186ad7SThomas Gleixner 547dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 548dc186ad7SThomas Gleixner { 549dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 550dc186ad7SThomas Gleixner } 551dc186ad7SThomas Gleixner 552dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 553dc186ad7SThomas Gleixner { 554dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 555dc186ad7SThomas Gleixner } 556dc186ad7SThomas Gleixner 557dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 558dc186ad7SThomas Gleixner { 559dc186ad7SThomas Gleixner if (onstack) 560dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 561dc186ad7SThomas Gleixner else 562dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 563dc186ad7SThomas Gleixner } 564dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 565dc186ad7SThomas Gleixner 566dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 567dc186ad7SThomas Gleixner { 568dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 569dc186ad7SThomas Gleixner } 570dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 571dc186ad7SThomas Gleixner 572ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 573ea2e64f2SThomas Gleixner { 574ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 575ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 576ea2e64f2SThomas Gleixner } 577ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 578ea2e64f2SThomas Gleixner 579dc186ad7SThomas Gleixner #else 580dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 581dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 582dc186ad7SThomas Gleixner #endif 583dc186ad7SThomas Gleixner 5844e8b22bdSLi Bin /** 58567dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5864e8b22bdSLi Bin * @pool: the pool pointer of interest 5874e8b22bdSLi Bin * 5884e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5894e8b22bdSLi Bin * successfully, -errno on failure. 5904e8b22bdSLi Bin */ 5919daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5929daf9e67STejun Heo { 5939daf9e67STejun Heo int ret; 5949daf9e67STejun Heo 59568e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5965bcab335STejun Heo 5974e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5984e8b22bdSLi Bin GFP_KERNEL); 599229641a6STejun Heo if (ret >= 0) { 600e68035fbSTejun Heo pool->id = ret; 601229641a6STejun Heo return 0; 602229641a6STejun Heo } 6039daf9e67STejun Heo return ret; 6049daf9e67STejun Heo } 6059daf9e67STejun Heo 60676af4d93STejun Heo /** 607df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 608df2d5ae4STejun Heo * @wq: the target workqueue 609df2d5ae4STejun Heo * @node: the node ID 610df2d5ae4STejun Heo * 61124acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 6125b95e1afSLai Jiangshan * read locked. 613df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 614df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 615d185af30SYacine Belkadi * 616d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 617df2d5ae4STejun Heo */ 618df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 619df2d5ae4STejun Heo int node) 620df2d5ae4STejun Heo { 6215b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 622d6e022f1STejun Heo 623d6e022f1STejun Heo /* 624d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 625d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 626d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 627d6e022f1STejun Heo * happens, this workaround can be removed. 628d6e022f1STejun Heo */ 629d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 630d6e022f1STejun Heo return wq->dfl_pwq; 631d6e022f1STejun Heo 632df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 633df2d5ae4STejun Heo } 634df2d5ae4STejun Heo 63573f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 63673f53c4aSTejun Heo { 63773f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 63873f53c4aSTejun Heo } 63973f53c4aSTejun Heo 640c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 64173f53c4aSTejun Heo { 642c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 64373f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 64473f53c4aSTejun Heo } 64573f53c4aSTejun Heo 64673f53c4aSTejun Heo static int work_next_color(int color) 64773f53c4aSTejun Heo { 64873f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 649a848e3b6SOleg Nesterov } 650a848e3b6SOleg Nesterov 6514594bf15SDavid Howells /* 652112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 653112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6547c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6557a22ad75STejun Heo * 656112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 657112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 658bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 659bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6607a22ad75STejun Heo * 661112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6627c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 663112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6647c3eed5cSTejun Heo * available only while the work item is queued. 665bbb68dfaSTejun Heo * 666bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 667bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 668bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 669bbb68dfaSTejun Heo * try to steal the PENDING bit. 6704594bf15SDavid Howells */ 6717a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6727a22ad75STejun Heo unsigned long flags) 6737a22ad75STejun Heo { 6746183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6757a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6767a22ad75STejun Heo } 6777a22ad75STejun Heo 678112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6794690c4abSTejun Heo unsigned long extra_flags) 680365970a1SDavid Howells { 681112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 682112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 683365970a1SDavid Howells } 684365970a1SDavid Howells 6854468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6864468a00fSLai Jiangshan int pool_id) 6874468a00fSLai Jiangshan { 6884468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6894468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6904468a00fSLai Jiangshan } 6914468a00fSLai Jiangshan 6927c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6937c3eed5cSTejun Heo int pool_id) 6944d707b9fSOleg Nesterov { 69523657bb1STejun Heo /* 69623657bb1STejun Heo * The following wmb is paired with the implied mb in 69723657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 69823657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 69923657bb1STejun Heo * owner. 70023657bb1STejun Heo */ 70123657bb1STejun Heo smp_wmb(); 7027c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 703346c09f8SRoman Pen /* 704346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 705346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 706346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 7078bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 708346c09f8SRoman Pen * the same @work. E.g. consider this case: 709346c09f8SRoman Pen * 710346c09f8SRoman Pen * CPU#0 CPU#1 711346c09f8SRoman Pen * ---------------------------- -------------------------------- 712346c09f8SRoman Pen * 713346c09f8SRoman Pen * 1 STORE event_indicated 714346c09f8SRoman Pen * 2 queue_work_on() { 715346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 716346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 717346c09f8SRoman Pen * 5 set_work_data() # clear bit 718346c09f8SRoman Pen * 6 smp_mb() 719346c09f8SRoman Pen * 7 work->current_func() { 720346c09f8SRoman Pen * 8 LOAD event_indicated 721346c09f8SRoman Pen * } 722346c09f8SRoman Pen * 723346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 724346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 725346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 726346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 727346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 728346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 729346c09f8SRoman Pen * before actual STORE. 730346c09f8SRoman Pen */ 731346c09f8SRoman Pen smp_mb(); 7324d707b9fSOleg Nesterov } 7334d707b9fSOleg Nesterov 7347a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 735365970a1SDavid Howells { 7367c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7377c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7387a22ad75STejun Heo } 7397a22ad75STejun Heo 740afa4bb77SLinus Torvalds static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 741afa4bb77SLinus Torvalds { 742afa4bb77SLinus Torvalds return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 743afa4bb77SLinus Torvalds } 744afa4bb77SLinus Torvalds 745112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7467a22ad75STejun Heo { 747e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7487a22ad75STejun Heo 749112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 750afa4bb77SLinus Torvalds return work_struct_pwq(data); 751e120153dSTejun Heo else 752e120153dSTejun Heo return NULL; 7537a22ad75STejun Heo } 7547a22ad75STejun Heo 7557c3eed5cSTejun Heo /** 7567c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7577c3eed5cSTejun Heo * @work: the work item of interest 7587c3eed5cSTejun Heo * 75968e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 76024acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 76124acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 762fa1b54e6STejun Heo * 763fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 764fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 765fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 766fa1b54e6STejun Heo * returned pool is and stays online. 767d185af30SYacine Belkadi * 768d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7697c3eed5cSTejun Heo */ 7707c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7717a22ad75STejun Heo { 772e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7737c3eed5cSTejun Heo int pool_id; 7747a22ad75STejun Heo 77568e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 776fa1b54e6STejun Heo 777112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 778afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool; 7797a22ad75STejun Heo 7807c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7817c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7827a22ad75STejun Heo return NULL; 7837a22ad75STejun Heo 784fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7857c3eed5cSTejun Heo } 7867c3eed5cSTejun Heo 7877c3eed5cSTejun Heo /** 7887c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7897c3eed5cSTejun Heo * @work: the work item of interest 7907c3eed5cSTejun Heo * 791d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7927c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7937c3eed5cSTejun Heo */ 7947c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7957c3eed5cSTejun Heo { 79654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7977c3eed5cSTejun Heo 798112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 799afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool->id; 80054d5b7d0SLai Jiangshan 80154d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 8027c3eed5cSTejun Heo } 8037c3eed5cSTejun Heo 804bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 805bbb68dfaSTejun Heo { 8067c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 807bbb68dfaSTejun Heo 8087c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 8097c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 810bbb68dfaSTejun Heo } 811bbb68dfaSTejun Heo 812bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 813bbb68dfaSTejun Heo { 814bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 815bbb68dfaSTejun Heo 816112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 817bbb68dfaSTejun Heo } 818bbb68dfaSTejun Heo 819e22bee78STejun Heo /* 8203270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8213270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 822d565ed63STejun Heo * they're being called with pool->lock held. 823e22bee78STejun Heo */ 824e22bee78STejun Heo 82563d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 826649027d7STejun Heo { 827bc35f7efSLai Jiangshan return !pool->nr_running; 828649027d7STejun Heo } 829649027d7STejun Heo 830e22bee78STejun Heo /* 831e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 832e22bee78STejun Heo * running workers. 833974271c4STejun Heo * 834974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 835706026c2STejun Heo * function will always return %true for unbound pools as long as the 836974271c4STejun Heo * worklist isn't empty. 837e22bee78STejun Heo */ 83863d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 839e22bee78STejun Heo { 84063d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 841e22bee78STejun Heo } 842e22bee78STejun Heo 843e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 84463d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 845e22bee78STejun Heo { 84663d95a91STejun Heo return pool->nr_idle; 847e22bee78STejun Heo } 848e22bee78STejun Heo 849e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 85063d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 851e22bee78STejun Heo { 852bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 853e22bee78STejun Heo } 854e22bee78STejun Heo 855e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 85663d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 857e22bee78STejun Heo { 85863d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 859e22bee78STejun Heo } 860e22bee78STejun Heo 861e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 86263d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 863e22bee78STejun Heo { 864692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 86563d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 86663d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 867e22bee78STejun Heo 868e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 869e22bee78STejun Heo } 870e22bee78STejun Heo 871e22bee78STejun Heo /* 872e22bee78STejun Heo * Wake up functions. 873e22bee78STejun Heo */ 874e22bee78STejun Heo 8752c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8761037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8777e11629dSTejun Heo { 87863d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8797e11629dSTejun Heo return NULL; 8807e11629dSTejun Heo 88163d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8827e11629dSTejun Heo } 8837e11629dSTejun Heo 8847e11629dSTejun Heo /** 8857e11629dSTejun Heo * wake_up_worker - wake up an idle worker 88663d95a91STejun Heo * @pool: worker pool to wake worker from 8877e11629dSTejun Heo * 88863d95a91STejun Heo * Wake up the first idle worker of @pool. 8897e11629dSTejun Heo * 8907e11629dSTejun Heo * CONTEXT: 891a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8927e11629dSTejun Heo */ 89363d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8947e11629dSTejun Heo { 8951037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8967e11629dSTejun Heo 8977e11629dSTejun Heo if (likely(worker)) 8987e11629dSTejun Heo wake_up_process(worker->task); 8997e11629dSTejun Heo } 9007e11629dSTejun Heo 9014690c4abSTejun Heo /** 902e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 903cb444766STejun Heo * @worker: self 904d302f017STejun Heo * @flags: flags to set 905d302f017STejun Heo * 906228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 907d302f017STejun Heo * 908cb444766STejun Heo * CONTEXT: 909a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 910d302f017STejun Heo */ 911228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 912d302f017STejun Heo { 913bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 914e22bee78STejun Heo 915cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 916cb444766STejun Heo 917228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 918e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 919e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 920bc35f7efSLai Jiangshan pool->nr_running--; 921e22bee78STejun Heo } 922e22bee78STejun Heo 923d302f017STejun Heo worker->flags |= flags; 924d302f017STejun Heo } 925d302f017STejun Heo 926d302f017STejun Heo /** 927e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 928cb444766STejun Heo * @worker: self 929d302f017STejun Heo * @flags: flags to clear 930d302f017STejun Heo * 931e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 932d302f017STejun Heo * 933cb444766STejun Heo * CONTEXT: 934a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 935d302f017STejun Heo */ 936d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 937d302f017STejun Heo { 93863d95a91STejun Heo struct worker_pool *pool = worker->pool; 939e22bee78STejun Heo unsigned int oflags = worker->flags; 940e22bee78STejun Heo 941cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 942cb444766STejun Heo 943d302f017STejun Heo worker->flags &= ~flags; 944e22bee78STejun Heo 94542c025f3STejun Heo /* 94642c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 94742c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 94842c025f3STejun Heo * of multiple flags, not a single flag. 94942c025f3STejun Heo */ 950e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 951e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 952bc35f7efSLai Jiangshan pool->nr_running++; 953d302f017STejun Heo } 954d302f017STejun Heo 95563638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 95663638450STejun Heo 95763638450STejun Heo /* 95863638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 95963638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 96063638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 96163638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 96263638450STejun Heo * should be using an unbound workqueue instead. 96363638450STejun Heo * 96463638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 96563638450STejun Heo * and report them so that they can be examined and converted to use unbound 96663638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 96763638450STejun Heo * function is tracked and reported with exponential backoff. 96863638450STejun Heo */ 96963638450STejun Heo #define WCI_MAX_ENTS 128 97063638450STejun Heo 97163638450STejun Heo struct wci_ent { 97263638450STejun Heo work_func_t func; 97363638450STejun Heo atomic64_t cnt; 97463638450STejun Heo struct hlist_node hash_node; 97563638450STejun Heo }; 97663638450STejun Heo 97763638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 97863638450STejun Heo static int wci_nr_ents; 97963638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 98063638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 98163638450STejun Heo 98263638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 98363638450STejun Heo { 98463638450STejun Heo struct wci_ent *ent; 98563638450STejun Heo 98663638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 98763638450STejun Heo (unsigned long)func) { 98863638450STejun Heo if (ent->func == func) 98963638450STejun Heo return ent; 99063638450STejun Heo } 99163638450STejun Heo return NULL; 99263638450STejun Heo } 99363638450STejun Heo 99463638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 99563638450STejun Heo { 99663638450STejun Heo struct wci_ent *ent; 99763638450STejun Heo 99863638450STejun Heo restart: 99963638450STejun Heo ent = wci_find_ent(func); 100063638450STejun Heo if (ent) { 100163638450STejun Heo u64 cnt; 100263638450STejun Heo 100363638450STejun Heo /* 100463638450STejun Heo * Start reporting from the fourth time and back off 100563638450STejun Heo * exponentially. 100663638450STejun Heo */ 100763638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 100863638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 100963638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 101063638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 101163638450STejun Heo atomic64_read(&ent->cnt)); 101263638450STejun Heo return; 101363638450STejun Heo } 101463638450STejun Heo 101563638450STejun Heo /* 101663638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 101763638450STejun Heo * is exhausted, something went really wrong and we probably made enough 101863638450STejun Heo * noise already. 101963638450STejun Heo */ 102063638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 102163638450STejun Heo return; 102263638450STejun Heo 102363638450STejun Heo raw_spin_lock(&wci_lock); 102463638450STejun Heo 102563638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 102663638450STejun Heo raw_spin_unlock(&wci_lock); 102763638450STejun Heo return; 102863638450STejun Heo } 102963638450STejun Heo 103063638450STejun Heo if (wci_find_ent(func)) { 103163638450STejun Heo raw_spin_unlock(&wci_lock); 103263638450STejun Heo goto restart; 103363638450STejun Heo } 103463638450STejun Heo 103563638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 103663638450STejun Heo ent->func = func; 103763638450STejun Heo atomic64_set(&ent->cnt, 1); 103863638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 103963638450STejun Heo 104063638450STejun Heo raw_spin_unlock(&wci_lock); 104163638450STejun Heo } 104263638450STejun Heo 104363638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 104463638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 104563638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 104663638450STejun Heo 1047c54d5046STejun Heo /** 10481da177e4SLinus Torvalds * wq_worker_running - a worker is running again 10491da177e4SLinus Torvalds * @task: task waking up 10501da177e4SLinus Torvalds * 10511da177e4SLinus Torvalds * This function is called when a worker returns from schedule() 10521da177e4SLinus Torvalds */ 10531da177e4SLinus Torvalds void wq_worker_running(struct task_struct *task) 10541da177e4SLinus Torvalds { 10551da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 10561da177e4SLinus Torvalds 1057c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 10581da177e4SLinus Torvalds return; 10591da177e4SLinus Torvalds 10601da177e4SLinus Torvalds /* 10611da177e4SLinus Torvalds * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 10621da177e4SLinus Torvalds * and the nr_running increment below, we may ruin the nr_running reset 10631da177e4SLinus Torvalds * and leave with an unexpected pool->nr_running == 1 on the newly unbound 10641da177e4SLinus Torvalds * pool. Protect against such race. 10651da177e4SLinus Torvalds */ 10661da177e4SLinus Torvalds preempt_disable(); 10671da177e4SLinus Torvalds if (!(worker->flags & WORKER_NOT_RUNNING)) 10681da177e4SLinus Torvalds worker->pool->nr_running++; 10691da177e4SLinus Torvalds preempt_enable(); 1070616db877STejun Heo 1071616db877STejun Heo /* 1072616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1073616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1074616db877STejun Heo */ 1075616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1076616db877STejun Heo 1077c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 10781da177e4SLinus Torvalds } 10791da177e4SLinus Torvalds 10801da177e4SLinus Torvalds /** 10811da177e4SLinus Torvalds * wq_worker_sleeping - a worker is going to sleep 10821da177e4SLinus Torvalds * @task: task going to sleep 10831da177e4SLinus Torvalds * 10841da177e4SLinus Torvalds * This function is called from schedule() when a busy worker is 10851da177e4SLinus Torvalds * going to sleep. 10861da177e4SLinus Torvalds */ 10871da177e4SLinus Torvalds void wq_worker_sleeping(struct task_struct *task) 10881da177e4SLinus Torvalds { 10891da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 10901da177e4SLinus Torvalds struct worker_pool *pool; 10911da177e4SLinus Torvalds 10921da177e4SLinus Torvalds /* 10931da177e4SLinus Torvalds * Rescuers, which may not have all the fields set up like normal 10941da177e4SLinus Torvalds * workers, also reach here, let's not access anything before 10951da177e4SLinus Torvalds * checking NOT_RUNNING. 10961da177e4SLinus Torvalds */ 10971da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) 10981da177e4SLinus Torvalds return; 10991da177e4SLinus Torvalds 11001da177e4SLinus Torvalds pool = worker->pool; 11011da177e4SLinus Torvalds 11021da177e4SLinus Torvalds /* Return if preempted before wq_worker_running() was reached */ 1103c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 11041da177e4SLinus Torvalds return; 11051da177e4SLinus Torvalds 1106c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 11071da177e4SLinus Torvalds raw_spin_lock_irq(&pool->lock); 11081da177e4SLinus Torvalds 11091da177e4SLinus Torvalds /* 11101da177e4SLinus Torvalds * Recheck in case unbind_workers() preempted us. We don't 11111da177e4SLinus Torvalds * want to decrement nr_running after the worker is unbound 11121da177e4SLinus Torvalds * and nr_running has been reset. 11131da177e4SLinus Torvalds */ 11141da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) { 11151da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 11161da177e4SLinus Torvalds return; 11171da177e4SLinus Torvalds } 11181da177e4SLinus Torvalds 11191da177e4SLinus Torvalds pool->nr_running--; 1120725e8ec5STejun Heo if (need_more_worker(pool)) { 1121725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 11221da177e4SLinus Torvalds wake_up_worker(pool); 1123725e8ec5STejun Heo } 11241da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 11251da177e4SLinus Torvalds } 11261da177e4SLinus Torvalds 11271da177e4SLinus Torvalds /** 1128616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1129616db877STejun Heo * @task: task currently running 1130616db877STejun Heo * 1131616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1132616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1133616db877STejun Heo */ 1134616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1135616db877STejun Heo { 1136616db877STejun Heo struct worker *worker = kthread_data(task); 1137616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1138616db877STejun Heo struct worker_pool *pool = worker->pool; 1139616db877STejun Heo 1140616db877STejun Heo if (!pwq) 1141616db877STejun Heo return; 1142616db877STejun Heo 11438a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 11448a1dd1e5STejun Heo 114518c8ae81SZqiang if (!wq_cpu_intensive_thresh_us) 114618c8ae81SZqiang return; 114718c8ae81SZqiang 1148616db877STejun Heo /* 1149616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1150616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1151616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1152c8f6219bSZqiang * 1153c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1154c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1155c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1156c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1157c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1158c8f6219bSZqiang * We probably want to make this prettier in the future. 1159616db877STejun Heo */ 1160c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1161616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1162616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1163616db877STejun Heo return; 1164616db877STejun Heo 1165616db877STejun Heo raw_spin_lock(&pool->lock); 1166616db877STejun Heo 1167616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 116863638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1169616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1170616db877STejun Heo 1171616db877STejun Heo if (need_more_worker(pool)) { 1172616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1173616db877STejun Heo wake_up_worker(pool); 1174616db877STejun Heo } 1175616db877STejun Heo 1176616db877STejun Heo raw_spin_unlock(&pool->lock); 1177616db877STejun Heo } 1178616db877STejun Heo 1179616db877STejun Heo /** 11801da177e4SLinus Torvalds * wq_worker_last_func - retrieve worker's last work function 11811da177e4SLinus Torvalds * @task: Task to retrieve last work function of. 11821da177e4SLinus Torvalds * 11831da177e4SLinus Torvalds * Determine the last function a worker executed. This is called from 11841da177e4SLinus Torvalds * the scheduler to get a worker's last known identity. 11851da177e4SLinus Torvalds * 11861da177e4SLinus Torvalds * CONTEXT: 11879b41ea72SAndrew Morton * raw_spin_lock_irq(rq->lock) 11881da177e4SLinus Torvalds * 11891da177e4SLinus Torvalds * This function is called during schedule() when a kworker is going 1190f756d5e2SNathan Lynch * to sleep. It's used by psi to identify aggregation workers during 1191f756d5e2SNathan Lynch * dequeuing, to allow periodic aggregation to shut-off when that 11921da177e4SLinus Torvalds * worker is the last task in the system or cgroup to go to sleep. 11931da177e4SLinus Torvalds * 11941da177e4SLinus Torvalds * As this function doesn't involve any workqueue-related locking, it 11951da177e4SLinus Torvalds * only returns stable values when called from inside the scheduler's 11961da177e4SLinus Torvalds * queuing and dequeuing paths, when @task, which must be a kworker, 11971da177e4SLinus Torvalds * is guaranteed to not be processing any works. 1198365970a1SDavid Howells * 1199365970a1SDavid Howells * Return: 1200365970a1SDavid Howells * The last work function %current executed as a worker, NULL if it 1201365970a1SDavid Howells * hasn't executed any work yet. 1202365970a1SDavid Howells */ 1203365970a1SDavid Howells work_func_t wq_worker_last_func(struct task_struct *task) 1204365970a1SDavid Howells { 1205365970a1SDavid Howells struct worker *worker = kthread_data(task); 1206365970a1SDavid Howells 1207365970a1SDavid Howells return worker->last_func; 1208365970a1SDavid Howells } 1209365970a1SDavid Howells 1210d302f017STejun Heo /** 12118cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1212c9e7cf27STejun Heo * @pool: pool of interest 12138cca0eeaSTejun Heo * @work: work to find worker for 12148cca0eeaSTejun Heo * 1215c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1216c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1217a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1218a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1219a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1220a2c1c57bSTejun Heo * being executed. 1221a2c1c57bSTejun Heo * 1222a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1223a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1224a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1225a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1226a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1227a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1228a2c1c57bSTejun Heo * 1229c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1230c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1231c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1232c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1233c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1234c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 12358cca0eeaSTejun Heo * 12368cca0eeaSTejun Heo * CONTEXT: 1237a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 12388cca0eeaSTejun Heo * 1239d185af30SYacine Belkadi * Return: 1240d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 12418cca0eeaSTejun Heo * otherwise. 12428cca0eeaSTejun Heo */ 1243c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 12448cca0eeaSTejun Heo struct work_struct *work) 12458cca0eeaSTejun Heo { 124642f8570fSSasha Levin struct worker *worker; 124742f8570fSSasha Levin 1248b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1249a2c1c57bSTejun Heo (unsigned long)work) 1250a2c1c57bSTejun Heo if (worker->current_work == work && 1251a2c1c57bSTejun Heo worker->current_func == work->func) 125242f8570fSSasha Levin return worker; 125342f8570fSSasha Levin 125442f8570fSSasha Levin return NULL; 12558cca0eeaSTejun Heo } 12568cca0eeaSTejun Heo 12578cca0eeaSTejun Heo /** 1258bf4ede01STejun Heo * move_linked_works - move linked works to a list 1259bf4ede01STejun Heo * @work: start of series of works to be scheduled 1260bf4ede01STejun Heo * @head: target list to append @work to 1261402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1262bf4ede01STejun Heo * 1263bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1264bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1265bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1266bf4ede01STejun Heo * 1267bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1268bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1269bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1270bf4ede01STejun Heo * 1271bf4ede01STejun Heo * CONTEXT: 1272a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1273bf4ede01STejun Heo */ 1274bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1275bf4ede01STejun Heo struct work_struct **nextp) 1276bf4ede01STejun Heo { 1277bf4ede01STejun Heo struct work_struct *n; 1278bf4ede01STejun Heo 1279bf4ede01STejun Heo /* 1280bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1281bf4ede01STejun Heo * use NULL for list head. 1282bf4ede01STejun Heo */ 1283bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1284bf4ede01STejun Heo list_move_tail(&work->entry, head); 1285bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1286bf4ede01STejun Heo break; 1287bf4ede01STejun Heo } 1288bf4ede01STejun Heo 1289bf4ede01STejun Heo /* 1290bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1291bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1292bf4ede01STejun Heo * needs to be updated. 1293bf4ede01STejun Heo */ 1294bf4ede01STejun Heo if (nextp) 1295bf4ede01STejun Heo *nextp = n; 1296bf4ede01STejun Heo } 1297bf4ede01STejun Heo 12988864b4e5STejun Heo /** 12998864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 13008864b4e5STejun Heo * @pwq: pool_workqueue to get 13018864b4e5STejun Heo * 13028864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 13038864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 13048864b4e5STejun Heo */ 13058864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 13068864b4e5STejun Heo { 13078864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13088864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 13098864b4e5STejun Heo pwq->refcnt++; 13108864b4e5STejun Heo } 13118864b4e5STejun Heo 13128864b4e5STejun Heo /** 13138864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 13148864b4e5STejun Heo * @pwq: pool_workqueue to put 13158864b4e5STejun Heo * 13168864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 13178864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 13188864b4e5STejun Heo */ 13198864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 13208864b4e5STejun Heo { 13218864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13228864b4e5STejun Heo if (likely(--pwq->refcnt)) 13238864b4e5STejun Heo return; 13248864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 13258864b4e5STejun Heo return; 13268864b4e5STejun Heo /* 13278864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 13288864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 13298864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 13308864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 13318864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 13328864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 13338864b4e5STejun Heo */ 13348864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 13358864b4e5STejun Heo } 13368864b4e5STejun Heo 1337dce90d47STejun Heo /** 1338dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1339dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1340dce90d47STejun Heo * 1341dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1342dce90d47STejun Heo */ 1343dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1344dce90d47STejun Heo { 1345dce90d47STejun Heo if (pwq) { 1346dce90d47STejun Heo /* 134724acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1348dce90d47STejun Heo * following lock operations are safe. 1349dce90d47STejun Heo */ 1350a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1351dce90d47STejun Heo put_pwq(pwq); 1352a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1353dce90d47STejun Heo } 1354dce90d47STejun Heo } 1355dce90d47STejun Heo 1356f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1357bf4ede01STejun Heo { 1358112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1359bf4ede01STejun Heo 1360bf4ede01STejun Heo trace_workqueue_activate_work(work); 136182607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 136282607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1363112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1364f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1365112202d9STejun Heo pwq->nr_active++; 1366bf4ede01STejun Heo } 1367bf4ede01STejun Heo 1368f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 13693aa62497SLai Jiangshan { 1370f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 13713aa62497SLai Jiangshan struct work_struct, entry); 13723aa62497SLai Jiangshan 1373f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13743aa62497SLai Jiangshan } 13753aa62497SLai Jiangshan 1376bf4ede01STejun Heo /** 1377112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1378112202d9STejun Heo * @pwq: pwq of interest 1379c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1380bf4ede01STejun Heo * 1381bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1382112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1383bf4ede01STejun Heo * 1384bf4ede01STejun Heo * CONTEXT: 1385a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1386bf4ede01STejun Heo */ 1387c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1388bf4ede01STejun Heo { 1389c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1390c4560c2cSLai Jiangshan 1391018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1392112202d9STejun Heo pwq->nr_active--; 1393f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1394f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1395112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1396f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1397bf4ede01STejun Heo } 1398018f3a13SLai Jiangshan } 1399018f3a13SLai Jiangshan 1400018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1401bf4ede01STejun Heo 1402bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1403112202d9STejun Heo if (likely(pwq->flush_color != color)) 14048864b4e5STejun Heo goto out_put; 1405bf4ede01STejun Heo 1406bf4ede01STejun Heo /* are there still in-flight works? */ 1407112202d9STejun Heo if (pwq->nr_in_flight[color]) 14088864b4e5STejun Heo goto out_put; 1409bf4ede01STejun Heo 1410112202d9STejun Heo /* this pwq is done, clear flush_color */ 1411112202d9STejun Heo pwq->flush_color = -1; 1412bf4ede01STejun Heo 1413bf4ede01STejun Heo /* 1414112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1415bf4ede01STejun Heo * will handle the rest. 1416bf4ede01STejun Heo */ 1417112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1418112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 14198864b4e5STejun Heo out_put: 14208864b4e5STejun Heo put_pwq(pwq); 1421bf4ede01STejun Heo } 1422bf4ede01STejun Heo 142336e227d2STejun Heo /** 1424bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 142536e227d2STejun Heo * @work: work item to steal 142636e227d2STejun Heo * @is_dwork: @work is a delayed_work 1427bbb68dfaSTejun Heo * @flags: place to store irq state 142836e227d2STejun Heo * 142936e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1430d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 143136e227d2STejun Heo * 1432d185af30SYacine Belkadi * Return: 14333eb6b31bSMauro Carvalho Chehab * 14343eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 143536e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 143636e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 143736e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1438bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1439bbb68dfaSTejun Heo * for arbitrarily long 14403eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 144136e227d2STejun Heo * 1442d185af30SYacine Belkadi * Note: 1443bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1444e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1445e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1446e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1447bbb68dfaSTejun Heo * 1448bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1449bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1450bbb68dfaSTejun Heo * 1451e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1452bf4ede01STejun Heo */ 1453bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1454bbb68dfaSTejun Heo unsigned long *flags) 1455bf4ede01STejun Heo { 1456d565ed63STejun Heo struct worker_pool *pool; 1457112202d9STejun Heo struct pool_workqueue *pwq; 1458bf4ede01STejun Heo 1459bbb68dfaSTejun Heo local_irq_save(*flags); 1460bbb68dfaSTejun Heo 146136e227d2STejun Heo /* try to steal the timer if it exists */ 146236e227d2STejun Heo if (is_dwork) { 146336e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 146436e227d2STejun Heo 1465e0aecdd8STejun Heo /* 1466e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1467e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1468e0aecdd8STejun Heo * running on the local CPU. 1469e0aecdd8STejun Heo */ 147036e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 147136e227d2STejun Heo return 1; 147236e227d2STejun Heo } 147336e227d2STejun Heo 147436e227d2STejun Heo /* try to claim PENDING the normal way */ 1475bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1476bf4ede01STejun Heo return 0; 1477bf4ede01STejun Heo 147824acfb71SThomas Gleixner rcu_read_lock(); 1479bf4ede01STejun Heo /* 1480bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1481bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1482bf4ede01STejun Heo */ 1483d565ed63STejun Heo pool = get_work_pool(work); 1484d565ed63STejun Heo if (!pool) 1485bbb68dfaSTejun Heo goto fail; 1486bf4ede01STejun Heo 1487a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1488bf4ede01STejun Heo /* 1489112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1490112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1491112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1492112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1493112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 14940b3dae68SLai Jiangshan * item is currently queued on that pool. 1495bf4ede01STejun Heo */ 1496112202d9STejun Heo pwq = get_work_pwq(work); 1497112202d9STejun Heo if (pwq && pwq->pool == pool) { 1498bf4ede01STejun Heo debug_work_deactivate(work); 14993aa62497SLai Jiangshan 15003aa62497SLai Jiangshan /* 1501018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1502018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1503018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1504018f3a13SLai Jiangshan * 1505f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1506d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1507f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 150816062836STejun Heo * management later on and cause stall. Make sure the work 150916062836STejun Heo * item is activated before grabbing. 15103aa62497SLai Jiangshan */ 1511f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1512f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 15133aa62497SLai Jiangshan 1514bf4ede01STejun Heo list_del_init(&work->entry); 1515c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 151636e227d2STejun Heo 1517112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 15184468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 15194468a00fSLai Jiangshan 1520a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 152124acfb71SThomas Gleixner rcu_read_unlock(); 152236e227d2STejun Heo return 1; 1523bf4ede01STejun Heo } 1524a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1525bbb68dfaSTejun Heo fail: 152624acfb71SThomas Gleixner rcu_read_unlock(); 1527bbb68dfaSTejun Heo local_irq_restore(*flags); 1528bbb68dfaSTejun Heo if (work_is_canceling(work)) 1529bbb68dfaSTejun Heo return -ENOENT; 1530bbb68dfaSTejun Heo cpu_relax(); 153136e227d2STejun Heo return -EAGAIN; 1532bf4ede01STejun Heo } 1533bf4ede01STejun Heo 1534bf4ede01STejun Heo /** 1535706026c2STejun Heo * insert_work - insert a work into a pool 1536112202d9STejun Heo * @pwq: pwq @work belongs to 15374690c4abSTejun Heo * @work: work to insert 15384690c4abSTejun Heo * @head: insertion point 15394690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 15404690c4abSTejun Heo * 1541112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1542706026c2STejun Heo * work_struct flags. 15434690c4abSTejun Heo * 15444690c4abSTejun Heo * CONTEXT: 1545a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1546365970a1SDavid Howells */ 1547112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1548112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1549b89deed3SOleg Nesterov { 1550112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1551e1d8aa9fSFrederic Weisbecker 1552e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1553f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1554e89a85d6SWalter Wu 15554690c4abSTejun Heo /* we own @work, set data and link */ 1556112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 15571a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 15588864b4e5STejun Heo get_pwq(pwq); 1559e22bee78STejun Heo 156063d95a91STejun Heo if (__need_more_worker(pool)) 156163d95a91STejun Heo wake_up_worker(pool); 1562b89deed3SOleg Nesterov } 1563b89deed3SOleg Nesterov 1564c8efcc25STejun Heo /* 1565c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 15668d03ecfeSTejun Heo * same workqueue. 1567c8efcc25STejun Heo */ 1568c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1569c8efcc25STejun Heo { 1570c8efcc25STejun Heo struct worker *worker; 1571c8efcc25STejun Heo 15728d03ecfeSTejun Heo worker = current_wq_worker(); 1573c8efcc25STejun Heo /* 1574bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 15758d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1576c8efcc25STejun Heo */ 1577112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1578c8efcc25STejun Heo } 1579c8efcc25STejun Heo 1580ef557180SMike Galbraith /* 1581ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1582ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1583ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1584ef557180SMike Galbraith */ 1585ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1586ef557180SMike Galbraith { 1587ef557180SMike Galbraith int new_cpu; 1588ef557180SMike Galbraith 1589f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1590ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1591ef557180SMike Galbraith return cpu; 1592a8ec5880SAmmar Faizi } else { 1593a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1594f303fccbSTejun Heo } 1595f303fccbSTejun Heo 1596ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1597ef557180SMike Galbraith return cpu; 1598ef557180SMike Galbraith 1599ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1600ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1601ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1602ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1603ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1604ef557180SMike Galbraith return cpu; 1605ef557180SMike Galbraith } 1606ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1607ef557180SMike Galbraith 1608ef557180SMike Galbraith return new_cpu; 1609ef557180SMike Galbraith } 1610ef557180SMike Galbraith 1611d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 16121da177e4SLinus Torvalds struct work_struct *work) 16131da177e4SLinus Torvalds { 1614112202d9STejun Heo struct pool_workqueue *pwq; 1615c9178087STejun Heo struct worker_pool *last_pool; 16161e19ffc6STejun Heo struct list_head *worklist; 16178a2e8e5dSTejun Heo unsigned int work_flags; 1618b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 16198930cabaSTejun Heo 16208930cabaSTejun Heo /* 16218930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 16228930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 16238930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 16248930cabaSTejun Heo * happen with IRQ disabled. 16258930cabaSTejun Heo */ 16268e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 16271da177e4SLinus Torvalds 16281e19ffc6STejun Heo 162933e3f0a3SRichard Clark /* 163033e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 163133e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 163233e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 163333e3f0a3SRichard Clark */ 163433e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 163533e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1636e41e704bSTejun Heo return; 163724acfb71SThomas Gleixner rcu_read_lock(); 16389e8cd2f5STejun Heo retry: 1639aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1640aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1641df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1642ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1643df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1644aa202f1fSHillf Danton } else { 1645aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1646aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1647aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1648aa202f1fSHillf Danton } 1649f3421797STejun Heo 165018aa9effSTejun Heo /* 1651c9178087STejun Heo * If @work was previously on a different pool, it might still be 1652c9178087STejun Heo * running there, in which case the work needs to be queued on that 1653c9178087STejun Heo * pool to guarantee non-reentrancy. 165418aa9effSTejun Heo */ 1655c9e7cf27STejun Heo last_pool = get_work_pool(work); 1656112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 165718aa9effSTejun Heo struct worker *worker; 165818aa9effSTejun Heo 1659a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 166018aa9effSTejun Heo 1661c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 166218aa9effSTejun Heo 1663112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1664c9178087STejun Heo pwq = worker->current_pwq; 16658594fadeSLai Jiangshan } else { 166618aa9effSTejun Heo /* meh... not running there, queue here */ 1667a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1668a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 166918aa9effSTejun Heo } 16708930cabaSTejun Heo } else { 1671a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 16728930cabaSTejun Heo } 1673502ca9d8STejun Heo 16749e8cd2f5STejun Heo /* 16759e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 16769e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 16779e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1678df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1679df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 16809e8cd2f5STejun Heo * make forward-progress. 16819e8cd2f5STejun Heo */ 16829e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 16839e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1684a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 16859e8cd2f5STejun Heo cpu_relax(); 16869e8cd2f5STejun Heo goto retry; 16879e8cd2f5STejun Heo } 16889e8cd2f5STejun Heo /* oops */ 16899e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 16909e8cd2f5STejun Heo wq->name, cpu); 16919e8cd2f5STejun Heo } 16929e8cd2f5STejun Heo 1693112202d9STejun Heo /* pwq determined, queue */ 1694112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1695502ca9d8STejun Heo 169624acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 169724acfb71SThomas Gleixner goto out; 16981e19ffc6STejun Heo 1699112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1700112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 17011e19ffc6STejun Heo 1702112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1703cdadf009STejun Heo trace_workqueue_activate_work(work); 1704112202d9STejun Heo pwq->nr_active++; 1705112202d9STejun Heo worklist = &pwq->pool->worklist; 170682607adcSTejun Heo if (list_empty(worklist)) 170782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 17088a2e8e5dSTejun Heo } else { 1709f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1710f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 17118a2e8e5dSTejun Heo } 17121e19ffc6STejun Heo 17130687c66bSZqiang debug_work_activate(work); 1714112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 17151e19ffc6STejun Heo 171624acfb71SThomas Gleixner out: 1717a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 171824acfb71SThomas Gleixner rcu_read_unlock(); 17191da177e4SLinus Torvalds } 17201da177e4SLinus Torvalds 17210fcb78c2SRolf Eike Beer /** 1722c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1723c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1724c1a220e7SZhang Rui * @wq: workqueue to use 1725c1a220e7SZhang Rui * @work: work to queue 1726c1a220e7SZhang Rui * 1727c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1728443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1729443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1730854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1731854f5cc5SPaul E. McKenney * online will get a splat. 1732d185af30SYacine Belkadi * 1733d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1734c1a220e7SZhang Rui */ 1735d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1736d4283e93STejun Heo struct work_struct *work) 1737c1a220e7SZhang Rui { 1738d4283e93STejun Heo bool ret = false; 17398930cabaSTejun Heo unsigned long flags; 17408930cabaSTejun Heo 17418930cabaSTejun Heo local_irq_save(flags); 1742c1a220e7SZhang Rui 174322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17444690c4abSTejun Heo __queue_work(cpu, wq, work); 1745d4283e93STejun Heo ret = true; 1746c1a220e7SZhang Rui } 17478930cabaSTejun Heo 17488930cabaSTejun Heo local_irq_restore(flags); 1749c1a220e7SZhang Rui return ret; 1750c1a220e7SZhang Rui } 1751ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1752c1a220e7SZhang Rui 17538204e0c1SAlexander Duyck /** 17548204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 17558204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 17568204e0c1SAlexander Duyck * 17578204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 17588204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 17598204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 17608204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 17618204e0c1SAlexander Duyck */ 17628204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 17638204e0c1SAlexander Duyck { 17648204e0c1SAlexander Duyck int cpu; 17658204e0c1SAlexander Duyck 17668204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 17678204e0c1SAlexander Duyck if (!wq_numa_enabled) 17688204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17698204e0c1SAlexander Duyck 17708204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 17718204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 17728204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17738204e0c1SAlexander Duyck 17748204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 17758204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 17768204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 17778204e0c1SAlexander Duyck return cpu; 17788204e0c1SAlexander Duyck 17798204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 17808204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 17818204e0c1SAlexander Duyck 17828204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 17838204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 17848204e0c1SAlexander Duyck } 17858204e0c1SAlexander Duyck 17868204e0c1SAlexander Duyck /** 17878204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 17888204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 17898204e0c1SAlexander Duyck * @wq: workqueue to use 17908204e0c1SAlexander Duyck * @work: work to queue 17918204e0c1SAlexander Duyck * 17928204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 17938204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 17948204e0c1SAlexander Duyck * NUMA node. 17958204e0c1SAlexander Duyck * 17968204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 17978204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 17988204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 17998204e0c1SAlexander Duyck * 18008204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 18018204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 18028204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 18038204e0c1SAlexander Duyck * 18048204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 18058204e0c1SAlexander Duyck */ 18068204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18078204e0c1SAlexander Duyck struct work_struct *work) 18088204e0c1SAlexander Duyck { 18098204e0c1SAlexander Duyck unsigned long flags; 18108204e0c1SAlexander Duyck bool ret = false; 18118204e0c1SAlexander Duyck 18128204e0c1SAlexander Duyck /* 18138204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 18148204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 18158204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 18168204e0c1SAlexander Duyck * 18178204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 18188204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 18198204e0c1SAlexander Duyck * some round robin type logic. 18208204e0c1SAlexander Duyck */ 18218204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 18228204e0c1SAlexander Duyck 18238204e0c1SAlexander Duyck local_irq_save(flags); 18248204e0c1SAlexander Duyck 18258204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18268204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 18278204e0c1SAlexander Duyck 18288204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 18298204e0c1SAlexander Duyck ret = true; 18308204e0c1SAlexander Duyck } 18318204e0c1SAlexander Duyck 18328204e0c1SAlexander Duyck local_irq_restore(flags); 18338204e0c1SAlexander Duyck return ret; 18348204e0c1SAlexander Duyck } 18358204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 18368204e0c1SAlexander Duyck 18378c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 18381da177e4SLinus Torvalds { 18398c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 18401da177e4SLinus Torvalds 1841e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 184260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 18431da177e4SLinus Torvalds } 18441438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 18451da177e4SLinus Torvalds 18467beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 184752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18481da177e4SLinus Torvalds { 18497beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 18507beb2edfSTejun Heo struct work_struct *work = &dwork->work; 18511da177e4SLinus Torvalds 1852637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 18534b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1854fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1855fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 18567beb2edfSTejun Heo 18578852aac2STejun Heo /* 18588852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 18598852aac2STejun Heo * both optimization and correctness. The earliest @timer can 18608852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 18618852aac2STejun Heo * on that there's no such delay when @delay is 0. 18628852aac2STejun Heo */ 18638852aac2STejun Heo if (!delay) { 18648852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 18658852aac2STejun Heo return; 18668852aac2STejun Heo } 18678852aac2STejun Heo 186860c057bcSLai Jiangshan dwork->wq = wq; 18691265057fSTejun Heo dwork->cpu = cpu; 18707beb2edfSTejun Heo timer->expires = jiffies + delay; 18717beb2edfSTejun Heo 1872041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 18737beb2edfSTejun Heo add_timer_on(timer, cpu); 1874041bd12eSTejun Heo else 1875041bd12eSTejun Heo add_timer(timer); 18767beb2edfSTejun Heo } 18771da177e4SLinus Torvalds 18780fcb78c2SRolf Eike Beer /** 18790fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 18800fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 18810fcb78c2SRolf Eike Beer * @wq: workqueue to use 1882af9997e4SRandy Dunlap * @dwork: work to queue 18830fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 18840fcb78c2SRolf Eike Beer * 1885d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1886715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1887715f1300STejun Heo * execution. 18880fcb78c2SRolf Eike Beer */ 1889d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 189052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18917a6bc1cdSVenkatesh Pallipadi { 189252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1893d4283e93STejun Heo bool ret = false; 18948930cabaSTejun Heo unsigned long flags; 18958930cabaSTejun Heo 18968930cabaSTejun Heo /* read the comment in __queue_work() */ 18978930cabaSTejun Heo local_irq_save(flags); 18987a6bc1cdSVenkatesh Pallipadi 189922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 19007beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1901d4283e93STejun Heo ret = true; 19027a6bc1cdSVenkatesh Pallipadi } 19038930cabaSTejun Heo 19048930cabaSTejun Heo local_irq_restore(flags); 19057a6bc1cdSVenkatesh Pallipadi return ret; 19067a6bc1cdSVenkatesh Pallipadi } 1907ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19081da177e4SLinus Torvalds 1909c8e55f36STejun Heo /** 19108376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19118376fe22STejun Heo * @cpu: CPU number to execute work on 19128376fe22STejun Heo * @wq: workqueue to use 19138376fe22STejun Heo * @dwork: work to queue 19148376fe22STejun Heo * @delay: number of jiffies to wait before queueing 19158376fe22STejun Heo * 19168376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 19178376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 19188376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 19198376fe22STejun Heo * current state. 19208376fe22STejun Heo * 1921d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 19228376fe22STejun Heo * pending and its timer was modified. 19238376fe22STejun Heo * 1924e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 19258376fe22STejun Heo * See try_to_grab_pending() for details. 19268376fe22STejun Heo */ 19278376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 19288376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 19298376fe22STejun Heo { 19308376fe22STejun Heo unsigned long flags; 19318376fe22STejun Heo int ret; 19328376fe22STejun Heo 19338376fe22STejun Heo do { 19348376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 19358376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 19368376fe22STejun Heo 19378376fe22STejun Heo if (likely(ret >= 0)) { 19388376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 19398376fe22STejun Heo local_irq_restore(flags); 19408376fe22STejun Heo } 19418376fe22STejun Heo 19428376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 19438376fe22STejun Heo return ret; 19448376fe22STejun Heo } 19458376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 19468376fe22STejun Heo 194705f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 194805f0fe6bSTejun Heo { 194905f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 195005f0fe6bSTejun Heo 195105f0fe6bSTejun Heo /* read the comment in __queue_work() */ 195205f0fe6bSTejun Heo local_irq_disable(); 195305f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 195405f0fe6bSTejun Heo local_irq_enable(); 195505f0fe6bSTejun Heo } 195605f0fe6bSTejun Heo 195705f0fe6bSTejun Heo /** 195805f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 195905f0fe6bSTejun Heo * @wq: workqueue to use 196005f0fe6bSTejun Heo * @rwork: work to queue 196105f0fe6bSTejun Heo * 196205f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 196305f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1964bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 196505f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 196605f0fe6bSTejun Heo */ 196705f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 196805f0fe6bSTejun Heo { 196905f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 197005f0fe6bSTejun Heo 197105f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 197205f0fe6bSTejun Heo rwork->wq = wq; 1973a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 197405f0fe6bSTejun Heo return true; 197505f0fe6bSTejun Heo } 197605f0fe6bSTejun Heo 197705f0fe6bSTejun Heo return false; 197805f0fe6bSTejun Heo } 197905f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 198005f0fe6bSTejun Heo 19818376fe22STejun Heo /** 1982c8e55f36STejun Heo * worker_enter_idle - enter idle state 1983c8e55f36STejun Heo * @worker: worker which is entering idle state 1984c8e55f36STejun Heo * 1985c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1986c8e55f36STejun Heo * necessary. 1987c8e55f36STejun Heo * 1988c8e55f36STejun Heo * LOCKING: 1989a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1990c8e55f36STejun Heo */ 1991c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 19921da177e4SLinus Torvalds { 1993bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1994c8e55f36STejun Heo 19956183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 19966183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 19976183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 19986183c009STejun Heo return; 1999c8e55f36STejun Heo 2000051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 2001cb444766STejun Heo worker->flags |= WORKER_IDLE; 2002bd7bdd43STejun Heo pool->nr_idle++; 2003e22bee78STejun Heo worker->last_active = jiffies; 2004c8e55f36STejun Heo 2005c8e55f36STejun Heo /* idle_list is LIFO */ 2006bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 2007db7bccf4STejun Heo 200863d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 2009628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 2010cb444766STejun Heo 2011989442d7SLai Jiangshan /* Sanity check nr_running. */ 2012bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 2013c8e55f36STejun Heo } 2014c8e55f36STejun Heo 2015c8e55f36STejun Heo /** 2016c8e55f36STejun Heo * worker_leave_idle - leave idle state 2017c8e55f36STejun Heo * @worker: worker which is leaving idle state 2018c8e55f36STejun Heo * 2019c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 2020c8e55f36STejun Heo * 2021c8e55f36STejun Heo * LOCKING: 2022a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2023c8e55f36STejun Heo */ 2024c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 2025c8e55f36STejun Heo { 2026bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2027c8e55f36STejun Heo 20286183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 20296183c009STejun Heo return; 2030d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 2031bd7bdd43STejun Heo pool->nr_idle--; 2032c8e55f36STejun Heo list_del_init(&worker->entry); 2033c8e55f36STejun Heo } 2034c8e55f36STejun Heo 2035f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 2036c34056a3STejun Heo { 2037c34056a3STejun Heo struct worker *worker; 2038c34056a3STejun Heo 2039f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2040c8e55f36STejun Heo if (worker) { 2041c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 2042affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 2043da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2044e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2045e22bee78STejun Heo worker->flags = WORKER_PREP; 2046c8e55f36STejun Heo } 2047c34056a3STejun Heo return worker; 2048c34056a3STejun Heo } 2049c34056a3STejun Heo 2050c34056a3STejun Heo /** 20514736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20524736cbf7SLai Jiangshan * @worker: worker to be attached 20534736cbf7SLai Jiangshan * @pool: the target pool 20544736cbf7SLai Jiangshan * 20554736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20564736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20574736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20584736cbf7SLai Jiangshan */ 20594736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 20604736cbf7SLai Jiangshan struct worker_pool *pool) 20614736cbf7SLai Jiangshan { 20621258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 20634736cbf7SLai Jiangshan 20644736cbf7SLai Jiangshan /* 20651258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 20661258fae7STejun Heo * stable across this function. See the comments above the flag 20671258fae7STejun Heo * definition for details. 20684736cbf7SLai Jiangshan */ 20694736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 20704736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 20715c25b5ffSPeter Zijlstra else 20725c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 20734736cbf7SLai Jiangshan 2074640f17c8SPeter Zijlstra if (worker->rescue_wq) 2075640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 2076640f17c8SPeter Zijlstra 20774736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2078a2d812a2STejun Heo worker->pool = pool; 20794736cbf7SLai Jiangshan 20801258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 20814736cbf7SLai Jiangshan } 20824736cbf7SLai Jiangshan 20834736cbf7SLai Jiangshan /** 208460f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 208560f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 208660f5a4bcSLai Jiangshan * 20874736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 20884736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 20894736cbf7SLai Jiangshan * other reference to the pool. 209060f5a4bcSLai Jiangshan */ 2091a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 209260f5a4bcSLai Jiangshan { 2093a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 209460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 209560f5a4bcSLai Jiangshan 20961258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2097a2d812a2STejun Heo 20985c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2099da028469SLai Jiangshan list_del(&worker->node); 2100a2d812a2STejun Heo worker->pool = NULL; 2101a2d812a2STejun Heo 2102e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 210360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 21041258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 210560f5a4bcSLai Jiangshan 2106b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2107b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2108b62c0751SLai Jiangshan 210960f5a4bcSLai Jiangshan if (detach_completion) 211060f5a4bcSLai Jiangshan complete(detach_completion); 211160f5a4bcSLai Jiangshan } 211260f5a4bcSLai Jiangshan 211360f5a4bcSLai Jiangshan /** 2114c34056a3STejun Heo * create_worker - create a new workqueue worker 211563d95a91STejun Heo * @pool: pool the new worker will belong to 2116c34056a3STejun Heo * 2117051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2118c34056a3STejun Heo * 2119c34056a3STejun Heo * CONTEXT: 2120c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2121c34056a3STejun Heo * 2122d185af30SYacine Belkadi * Return: 2123c34056a3STejun Heo * Pointer to the newly created worker. 2124c34056a3STejun Heo */ 2125bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2126c34056a3STejun Heo { 2127e441b56fSZhen Lei struct worker *worker; 2128e441b56fSZhen Lei int id; 2129e3c916a4STejun Heo char id_buf[16]; 2130c34056a3STejun Heo 21317cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2132e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 21333f0ea0b8SPetr Mladek if (id < 0) { 21343f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 21353f0ea0b8SPetr Mladek ERR_PTR(id)); 2136e441b56fSZhen Lei return NULL; 21373f0ea0b8SPetr Mladek } 2138c34056a3STejun Heo 2139f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 21403f0ea0b8SPetr Mladek if (!worker) { 21413f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2142c34056a3STejun Heo goto fail; 21433f0ea0b8SPetr Mladek } 2144c34056a3STejun Heo 2145c34056a3STejun Heo worker->id = id; 2146c34056a3STejun Heo 214729c91e99STejun Heo if (pool->cpu >= 0) 2148e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2149e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2150f3421797STejun Heo else 2151e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2152e3c916a4STejun Heo 2153f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2154e3c916a4STejun Heo "kworker/%s", id_buf); 21553f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 215660f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 215760f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 215860f54038SPetr Mladek id_buf); 215960f54038SPetr Mladek } else { 21603f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 21613f0ea0b8SPetr Mladek worker->task); 216260f54038SPetr Mladek } 2163c34056a3STejun Heo goto fail; 21643f0ea0b8SPetr Mladek } 2165c34056a3STejun Heo 216691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 216725834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 216891151228SOleg Nesterov 2169da028469SLai Jiangshan /* successful, attach the worker to the pool */ 21704736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2171822d8405STejun Heo 2172051e1850SLai Jiangshan /* start the newly created worker */ 2173a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2174051e1850SLai Jiangshan worker->pool->nr_workers++; 2175051e1850SLai Jiangshan worker_enter_idle(worker); 2176051e1850SLai Jiangshan wake_up_process(worker->task); 2177a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2178051e1850SLai Jiangshan 2179c34056a3STejun Heo return worker; 2180822d8405STejun Heo 2181c34056a3STejun Heo fail: 2182e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2183c34056a3STejun Heo kfree(worker); 2184c34056a3STejun Heo return NULL; 2185c34056a3STejun Heo } 2186c34056a3STejun Heo 2187793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2188793777bcSValentin Schneider { 2189793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2190793777bcSValentin Schneider 2191793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2192793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2193793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2194793777bcSValentin Schneider else 2195793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2196793777bcSValentin Schneider } 2197793777bcSValentin Schneider 2198e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2199e02b9312SValentin Schneider { 2200e02b9312SValentin Schneider struct worker *worker, *tmp; 2201e02b9312SValentin Schneider 2202e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2203e02b9312SValentin Schneider list_del_init(&worker->entry); 2204e02b9312SValentin Schneider unbind_worker(worker); 2205e02b9312SValentin Schneider /* 2206e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2207e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2208e02b9312SValentin Schneider * wouldn't have gotten here. 2209c34056a3STejun Heo * 2210e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2211e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2212e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2213e02b9312SValentin Schneider * outside of pool->lock. 2214e02b9312SValentin Schneider */ 2215e02b9312SValentin Schneider wake_up_process(worker->task); 2216e02b9312SValentin Schneider } 2217e02b9312SValentin Schneider } 2218e02b9312SValentin Schneider 2219e02b9312SValentin Schneider /** 2220e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2221e02b9312SValentin Schneider * @worker: worker to be destroyed 2222e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2223e02b9312SValentin Schneider * 2224e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2225e02b9312SValentin Schneider * should be idle. 2226c8e55f36STejun Heo * 2227c8e55f36STejun Heo * CONTEXT: 2228a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2229c34056a3STejun Heo */ 2230e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2231c34056a3STejun Heo { 2232bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2233c34056a3STejun Heo 2234cd549687STejun Heo lockdep_assert_held(&pool->lock); 2235e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2236cd549687STejun Heo 2237c34056a3STejun Heo /* sanity check frenzy */ 22386183c009STejun Heo if (WARN_ON(worker->current_work) || 223973eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 224073eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 22416183c009STejun Heo return; 2242c34056a3STejun Heo 2243bd7bdd43STejun Heo pool->nr_workers--; 2244bd7bdd43STejun Heo pool->nr_idle--; 2245c8e55f36STejun Heo 2246cb444766STejun Heo worker->flags |= WORKER_DIE; 2247e02b9312SValentin Schneider 2248e02b9312SValentin Schneider list_move(&worker->entry, list); 2249e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2250c34056a3STejun Heo } 2251c34056a3STejun Heo 22523f959aa3SValentin Schneider /** 22533f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 22543f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 22553f959aa3SValentin Schneider * 22563f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 22573f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 22583f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 22593f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 22603f959aa3SValentin Schneider * it expire and re-evaluate things from there. 22613f959aa3SValentin Schneider */ 226232a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2263e22bee78STejun Heo { 226432a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 22653f959aa3SValentin Schneider bool do_cull = false; 22663f959aa3SValentin Schneider 22673f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 22683f959aa3SValentin Schneider return; 22693f959aa3SValentin Schneider 22703f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 22713f959aa3SValentin Schneider 22723f959aa3SValentin Schneider if (too_many_workers(pool)) { 22733f959aa3SValentin Schneider struct worker *worker; 22743f959aa3SValentin Schneider unsigned long expires; 22753f959aa3SValentin Schneider 22763f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 22773f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 22783f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 22793f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 22803f959aa3SValentin Schneider 22813f959aa3SValentin Schneider if (!do_cull) 22823f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 22833f959aa3SValentin Schneider } 22843f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 22853f959aa3SValentin Schneider 22863f959aa3SValentin Schneider if (do_cull) 22873f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 22883f959aa3SValentin Schneider } 22893f959aa3SValentin Schneider 22903f959aa3SValentin Schneider /** 22913f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 22923f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 22933f959aa3SValentin Schneider * 22943f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 22953f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2296e02b9312SValentin Schneider * 2297e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2298e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2299e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 23003f959aa3SValentin Schneider */ 23013f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 23023f959aa3SValentin Schneider { 23033f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2304e02b9312SValentin Schneider struct list_head cull_list; 2305e22bee78STejun Heo 2306e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 2307e02b9312SValentin Schneider /* 2308e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2309e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2310e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2311e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2312e02b9312SValentin Schneider */ 2313e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2314a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2315e22bee78STejun Heo 23163347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2317e22bee78STejun Heo struct worker *worker; 2318e22bee78STejun Heo unsigned long expires; 2319e22bee78STejun Heo 232063d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2321e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2322e22bee78STejun Heo 23233347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 232463d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 23253347fc9fSLai Jiangshan break; 2326e22bee78STejun Heo } 23273347fc9fSLai Jiangshan 2328e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2329e22bee78STejun Heo } 2330e22bee78STejun Heo 2331a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2332e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2333e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2334e22bee78STejun Heo } 2335e22bee78STejun Heo 2336493a1724STejun Heo static void send_mayday(struct work_struct *work) 2337e22bee78STejun Heo { 2338112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2339112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2340493a1724STejun Heo 23412e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2342e22bee78STejun Heo 2343493008a8STejun Heo if (!wq->rescuer) 2344493a1724STejun Heo return; 2345e22bee78STejun Heo 2346e22bee78STejun Heo /* mayday mayday mayday */ 2347493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 234877668c8bSLai Jiangshan /* 234977668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 235077668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 235177668c8bSLai Jiangshan * rescuer is done with it. 235277668c8bSLai Jiangshan */ 235377668c8bSLai Jiangshan get_pwq(pwq); 2354493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2355e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2356725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2357493a1724STejun Heo } 2358e22bee78STejun Heo } 2359e22bee78STejun Heo 236032a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2361e22bee78STejun Heo { 236232a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2363e22bee78STejun Heo struct work_struct *work; 2364e22bee78STejun Heo 2365a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2366a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2367e22bee78STejun Heo 236863d95a91STejun Heo if (need_to_create_worker(pool)) { 2369e22bee78STejun Heo /* 2370e22bee78STejun Heo * We've been trying to create a new worker but 2371e22bee78STejun Heo * haven't been successful. We might be hitting an 2372e22bee78STejun Heo * allocation deadlock. Send distress signals to 2373e22bee78STejun Heo * rescuers. 2374e22bee78STejun Heo */ 237563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2376e22bee78STejun Heo send_mayday(work); 2377e22bee78STejun Heo } 2378e22bee78STejun Heo 2379a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2380a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2381e22bee78STejun Heo 238263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2383e22bee78STejun Heo } 2384e22bee78STejun Heo 2385e22bee78STejun Heo /** 2386e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 238763d95a91STejun Heo * @pool: pool to create a new worker for 2388e22bee78STejun Heo * 238963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2390e22bee78STejun Heo * have at least one idle worker on return from this function. If 2391e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 239263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2393e22bee78STejun Heo * possible allocation deadlock. 2394e22bee78STejun Heo * 2395c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2396c5aa87bbSTejun Heo * may_start_working() %true. 2397e22bee78STejun Heo * 2398e22bee78STejun Heo * LOCKING: 2399a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2400e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2401e22bee78STejun Heo * manager. 2402e22bee78STejun Heo */ 240329187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2404d565ed63STejun Heo __releases(&pool->lock) 2405d565ed63STejun Heo __acquires(&pool->lock) 2406e22bee78STejun Heo { 2407e22bee78STejun Heo restart: 2408a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24099f9c2364STejun Heo 2410e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 241163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2412e22bee78STejun Heo 2413e22bee78STejun Heo while (true) { 2414051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2415e22bee78STejun Heo break; 2416e22bee78STejun Heo 2417e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 24189f9c2364STejun Heo 241963d95a91STejun Heo if (!need_to_create_worker(pool)) 2420e22bee78STejun Heo break; 2421e22bee78STejun Heo } 2422e22bee78STejun Heo 242363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2424a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2425051e1850SLai Jiangshan /* 2426051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2427051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2428051e1850SLai Jiangshan * already become busy. 2429051e1850SLai Jiangshan */ 243063d95a91STejun Heo if (need_to_create_worker(pool)) 2431e22bee78STejun Heo goto restart; 2432e22bee78STejun Heo } 2433e22bee78STejun Heo 2434e22bee78STejun Heo /** 2435e22bee78STejun Heo * manage_workers - manage worker pool 2436e22bee78STejun Heo * @worker: self 2437e22bee78STejun Heo * 2438706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2439e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2440706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2441e22bee78STejun Heo * 2442e22bee78STejun Heo * The caller can safely start processing works on false return. On 2443e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2444e22bee78STejun Heo * and may_start_working() is true. 2445e22bee78STejun Heo * 2446e22bee78STejun Heo * CONTEXT: 2447a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2448e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2449e22bee78STejun Heo * 2450d185af30SYacine Belkadi * Return: 245129187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 245229187a9eSTejun Heo * start processing works, %true if management function was performed and 245329187a9eSTejun Heo * the conditions that the caller verified before calling the function may 245429187a9eSTejun Heo * no longer be true. 2455e22bee78STejun Heo */ 2456e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2457e22bee78STejun Heo { 245863d95a91STejun Heo struct worker_pool *pool = worker->pool; 2459e22bee78STejun Heo 2460692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 246129187a9eSTejun Heo return false; 2462692b4825STejun Heo 2463692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 24642607d7a6STejun Heo pool->manager = worker; 2465e22bee78STejun Heo 246629187a9eSTejun Heo maybe_create_worker(pool); 2467e22bee78STejun Heo 24682607d7a6STejun Heo pool->manager = NULL; 2469692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2470d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 247129187a9eSTejun Heo return true; 2472e22bee78STejun Heo } 2473e22bee78STejun Heo 2474a62428c0STejun Heo /** 2475a62428c0STejun Heo * process_one_work - process single work 2476c34056a3STejun Heo * @worker: self 2477a62428c0STejun Heo * @work: work to process 2478a62428c0STejun Heo * 2479a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2480a62428c0STejun Heo * process a single work including synchronization against and 2481a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2482a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2483a62428c0STejun Heo * call this function to process a work. 2484a62428c0STejun Heo * 2485a62428c0STejun Heo * CONTEXT: 2486a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2487a62428c0STejun Heo */ 2488c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2489d565ed63STejun Heo __releases(&pool->lock) 2490d565ed63STejun Heo __acquires(&pool->lock) 24911da177e4SLinus Torvalds { 2492112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2493bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2494c4560c2cSLai Jiangshan unsigned long work_data; 24957e11629dSTejun Heo struct worker *collision; 24964e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 24974e6045f1SJohannes Berg /* 2498a62428c0STejun Heo * It is permissible to free the struct work_struct from 2499a62428c0STejun Heo * inside the function that is called from it, this we need to 2500a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2501a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2502a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 25034e6045f1SJohannes Berg */ 25044d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 25054d82a1deSPeter Zijlstra 25064d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 25074e6045f1SJohannes Berg #endif 2508807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 250985327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2510ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 251125511a47STejun Heo 25127e11629dSTejun Heo /* 25137e11629dSTejun Heo * A single work shouldn't be executed concurrently by 25147e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 25157e11629dSTejun Heo * already processing the work. If so, defer the work to the 25167e11629dSTejun Heo * currently executing one. 25177e11629dSTejun Heo */ 2518c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 25197e11629dSTejun Heo if (unlikely(collision)) { 25207e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 25217e11629dSTejun Heo return; 25227e11629dSTejun Heo } 25231da177e4SLinus Torvalds 25248930cabaSTejun Heo /* claim and dequeue */ 25251da177e4SLinus Torvalds debug_work_deactivate(work); 2526c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2527c34056a3STejun Heo worker->current_work = work; 2528a2c1c57bSTejun Heo worker->current_func = work->func; 2529112202d9STejun Heo worker->current_pwq = pwq; 2530616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2531c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2532d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 25337a22ad75STejun Heo 25348bf89593STejun Heo /* 25358bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 25368bf89593STejun Heo * overridden through set_worker_desc(). 25378bf89593STejun Heo */ 25388bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 25398bf89593STejun Heo 2540a62428c0STejun Heo list_del_init(&work->entry); 2541a62428c0STejun Heo 2542649027d7STejun Heo /* 2543228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2544228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2545228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2546228f1d00SLai Jiangshan * execution of the pending work items. 2547fb0e7bebSTejun Heo */ 2548616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2549228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2550fb0e7bebSTejun Heo 2551974271c4STejun Heo /* 2552a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2553a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2554a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2555a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2556228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2557974271c4STejun Heo */ 2558a489a03eSLai Jiangshan if (need_more_worker(pool)) 255963d95a91STejun Heo wake_up_worker(pool); 2560974271c4STejun Heo 25618930cabaSTejun Heo /* 25627c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2563d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 256423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 256523657bb1STejun Heo * disabled. 25668930cabaSTejun Heo */ 25677c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 25681da177e4SLinus Torvalds 2569a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2570365970a1SDavid Howells 2571a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 25723295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2573e6f3faa7SPeter Zijlstra /* 2574f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2575f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2576e6f3faa7SPeter Zijlstra * 2577e6f3faa7SPeter Zijlstra * However, that would result in: 2578e6f3faa7SPeter Zijlstra * 2579e6f3faa7SPeter Zijlstra * A(W1) 2580e6f3faa7SPeter Zijlstra * WFC(C) 2581e6f3faa7SPeter Zijlstra * A(W1) 2582e6f3faa7SPeter Zijlstra * C(C) 2583e6f3faa7SPeter Zijlstra * 2584e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2585e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2586e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2587f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2588e6f3faa7SPeter Zijlstra * these locks. 2589e6f3faa7SPeter Zijlstra * 2590e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2591e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2592e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2593e6f3faa7SPeter Zijlstra */ 2594f52be570SPeter Zijlstra lockdep_invariant_state(true); 2595725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2596e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2597a2c1c57bSTejun Heo worker->current_func(work); 2598e36c886aSArjan van de Ven /* 2599e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2600e36c886aSArjan van de Ven * point will only record its address. 2601e36c886aSArjan van de Ven */ 26021c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2603725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 26043295f0efSIngo Molnar lock_map_release(&lockdep_map); 2605112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 26061da177e4SLinus Torvalds 2607d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2608044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2609d75f773cSSakari Ailus " last function: %ps\n", 2610a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2611a2c1c57bSTejun Heo worker->current_func); 2612d5abe669SPeter Zijlstra debug_show_held_locks(current); 2613d5abe669SPeter Zijlstra dump_stack(); 2614d5abe669SPeter Zijlstra } 2615d5abe669SPeter Zijlstra 2616b22ce278STejun Heo /* 2617025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2618b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2619b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2620b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2621789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2622789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2623b22ce278STejun Heo */ 2624a7e6425eSPaul E. McKenney cond_resched(); 2625b22ce278STejun Heo 2626a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2627a62428c0STejun Heo 2628616db877STejun Heo /* 2629616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2630616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2631616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2632616db877STejun Heo */ 2633fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2634fb0e7bebSTejun Heo 26351b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 26361b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 26371b69ac6bSJohannes Weiner 2638a62428c0STejun Heo /* we're done with it, release */ 263942f8570fSSasha Levin hash_del(&worker->hentry); 2640c34056a3STejun Heo worker->current_work = NULL; 2641a2c1c57bSTejun Heo worker->current_func = NULL; 2642112202d9STejun Heo worker->current_pwq = NULL; 2643d812796eSLai Jiangshan worker->current_color = INT_MAX; 2644c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26451da177e4SLinus Torvalds } 26461da177e4SLinus Torvalds 2647affee4b2STejun Heo /** 2648affee4b2STejun Heo * process_scheduled_works - process scheduled works 2649affee4b2STejun Heo * @worker: self 2650affee4b2STejun Heo * 2651affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2652affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2653affee4b2STejun Heo * fetches a work from the top and executes it. 2654affee4b2STejun Heo * 2655affee4b2STejun Heo * CONTEXT: 2656a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2657affee4b2STejun Heo * multiple times. 2658affee4b2STejun Heo */ 2659affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26601da177e4SLinus Torvalds { 2661affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2662affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2663a62428c0STejun Heo struct work_struct, entry); 2664c34056a3STejun Heo process_one_work(worker, work); 2665a62428c0STejun Heo } 26661da177e4SLinus Torvalds } 26671da177e4SLinus Torvalds 2668197f6accSTejun Heo static void set_pf_worker(bool val) 2669197f6accSTejun Heo { 2670197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2671197f6accSTejun Heo if (val) 2672197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2673197f6accSTejun Heo else 2674197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2675197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2676197f6accSTejun Heo } 2677197f6accSTejun Heo 26784690c4abSTejun Heo /** 26794690c4abSTejun Heo * worker_thread - the worker thread function 2680c34056a3STejun Heo * @__worker: self 26814690c4abSTejun Heo * 2682c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2683c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2684c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2685c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2686c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2687d185af30SYacine Belkadi * 2688d185af30SYacine Belkadi * Return: 0 26894690c4abSTejun Heo */ 2690c34056a3STejun Heo static int worker_thread(void *__worker) 26911da177e4SLinus Torvalds { 2692c34056a3STejun Heo struct worker *worker = __worker; 2693bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 26941da177e4SLinus Torvalds 2695e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2696197f6accSTejun Heo set_pf_worker(true); 2697c8e55f36STejun Heo woke_up: 2698a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2699affee4b2STejun Heo 2700a9ab775bSTejun Heo /* am I supposed to die? */ 2701a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2702a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2703197f6accSTejun Heo set_pf_worker(false); 270460f5a4bcSLai Jiangshan 270560f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2706e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2707a2d812a2STejun Heo worker_detach_from_pool(worker); 2708e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 270960f5a4bcSLai Jiangshan kfree(worker); 2710c8e55f36STejun Heo return 0; 2711c8e55f36STejun Heo } 2712c8e55f36STejun Heo 2713c8e55f36STejun Heo worker_leave_idle(worker); 2714db7bccf4STejun Heo recheck: 2715e22bee78STejun Heo /* no more worker necessary? */ 271663d95a91STejun Heo if (!need_more_worker(pool)) 2717e22bee78STejun Heo goto sleep; 2718e22bee78STejun Heo 2719e22bee78STejun Heo /* do we need to manage? */ 272063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2721e22bee78STejun Heo goto recheck; 2722e22bee78STejun Heo 2723c8e55f36STejun Heo /* 2724c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2725c8e55f36STejun Heo * preparing to process a work or actually processing it. 2726c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2727c8e55f36STejun Heo */ 27286183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2729c8e55f36STejun Heo 2730e22bee78STejun Heo /* 2731a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2732a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2733a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2734a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2735a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2736e22bee78STejun Heo */ 2737a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2738e22bee78STejun Heo 2739e22bee78STejun Heo do { 2740affee4b2STejun Heo struct work_struct *work = 2741bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2742affee4b2STejun Heo struct work_struct, entry); 2743affee4b2STejun Heo 274482607adcSTejun Heo pool->watchdog_ts = jiffies; 274582607adcSTejun Heo 2746c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2747affee4b2STejun Heo /* optimization path, not strictly necessary */ 2748affee4b2STejun Heo process_one_work(worker, work); 2749affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2750affee4b2STejun Heo process_scheduled_works(worker); 2751affee4b2STejun Heo } else { 2752c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2753affee4b2STejun Heo process_scheduled_works(worker); 2754affee4b2STejun Heo } 275563d95a91STejun Heo } while (keep_working(pool)); 2756affee4b2STejun Heo 2757228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2758d313dd85STejun Heo sleep: 2759c8e55f36STejun Heo /* 2760d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2761d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2762d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2763d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2764d565ed63STejun Heo * event. 2765c8e55f36STejun Heo */ 2766c8e55f36STejun Heo worker_enter_idle(worker); 2767c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2768a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27691da177e4SLinus Torvalds schedule(); 2770c8e55f36STejun Heo goto woke_up; 27711da177e4SLinus Torvalds } 27721da177e4SLinus Torvalds 2773e22bee78STejun Heo /** 2774e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2775111c225aSTejun Heo * @__rescuer: self 2776e22bee78STejun Heo * 2777e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2778493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2779e22bee78STejun Heo * 2780706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2781e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2782e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2783e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2784e22bee78STejun Heo * the problem rescuer solves. 2785e22bee78STejun Heo * 2786706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2787706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2788e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2789e22bee78STejun Heo * 2790e22bee78STejun Heo * This should happen rarely. 2791d185af30SYacine Belkadi * 2792d185af30SYacine Belkadi * Return: 0 2793e22bee78STejun Heo */ 2794111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2795e22bee78STejun Heo { 2796111c225aSTejun Heo struct worker *rescuer = __rescuer; 2797111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2798e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 27994d595b86SLai Jiangshan bool should_stop; 2800e22bee78STejun Heo 2801e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2802111c225aSTejun Heo 2803111c225aSTejun Heo /* 2804111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2805111c225aSTejun Heo * doesn't participate in concurrency management. 2806111c225aSTejun Heo */ 2807197f6accSTejun Heo set_pf_worker(true); 2808e22bee78STejun Heo repeat: 2809c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 28101da177e4SLinus Torvalds 28114d595b86SLai Jiangshan /* 28124d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 28134d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 28144d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 28154d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 28164d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 28174d595b86SLai Jiangshan * list is always empty on exit. 28184d595b86SLai Jiangshan */ 28194d595b86SLai Jiangshan should_stop = kthread_should_stop(); 28201da177e4SLinus Torvalds 2821493a1724STejun Heo /* see whether any pwq is asking for help */ 2822a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2823493a1724STejun Heo 2824493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2825493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2826493a1724STejun Heo struct pool_workqueue, mayday_node); 2827112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2828e22bee78STejun Heo struct work_struct *work, *n; 282982607adcSTejun Heo bool first = true; 2830e22bee78STejun Heo 2831e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2832493a1724STejun Heo list_del_init(&pwq->mayday_node); 2833493a1724STejun Heo 2834a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2835e22bee78STejun Heo 283651697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 283751697d39SLai Jiangshan 2838a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2839e22bee78STejun Heo 2840e22bee78STejun Heo /* 2841e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2842e22bee78STejun Heo * process'em. 2843e22bee78STejun Heo */ 28440479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 284582607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 284682607adcSTejun Heo if (get_work_pwq(work) == pwq) { 284782607adcSTejun Heo if (first) 284882607adcSTejun Heo pool->watchdog_ts = jiffies; 2849e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2850725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 285182607adcSTejun Heo } 285282607adcSTejun Heo first = false; 285382607adcSTejun Heo } 2854e22bee78STejun Heo 2855008847f6SNeilBrown if (!list_empty(scheduled)) { 2856e22bee78STejun Heo process_scheduled_works(rescuer); 28577576958aSTejun Heo 28587576958aSTejun Heo /* 2859008847f6SNeilBrown * The above execution of rescued work items could 2860008847f6SNeilBrown * have created more to rescue through 2861f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2862008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2863008847f6SNeilBrown * that such back-to-back work items, which may be 2864008847f6SNeilBrown * being used to relieve memory pressure, don't 2865008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2866008847f6SNeilBrown */ 28674f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2868a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2869e66b39afSTejun Heo /* 2870e66b39afSTejun Heo * Queue iff we aren't racing destruction 2871e66b39afSTejun Heo * and somebody else hasn't queued it already. 2872e66b39afSTejun Heo */ 2873e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2874008847f6SNeilBrown get_pwq(pwq); 2875e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2876e66b39afSTejun Heo } 2877a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2878008847f6SNeilBrown } 2879008847f6SNeilBrown } 2880008847f6SNeilBrown 2881008847f6SNeilBrown /* 288277668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 288313b1d625SLai Jiangshan * go away while we're still attached to it. 288477668c8bSLai Jiangshan */ 288577668c8bSLai Jiangshan put_pwq(pwq); 288677668c8bSLai Jiangshan 288777668c8bSLai Jiangshan /* 2888d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 28897576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 28907576958aSTejun Heo * and stalling the execution. 28917576958aSTejun Heo */ 2892d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 289363d95a91STejun Heo wake_up_worker(pool); 28947576958aSTejun Heo 2895a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 289613b1d625SLai Jiangshan 2897a2d812a2STejun Heo worker_detach_from_pool(rescuer); 289813b1d625SLai Jiangshan 2899a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 29001da177e4SLinus Torvalds } 29011da177e4SLinus Torvalds 2902a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2903493a1724STejun Heo 29044d595b86SLai Jiangshan if (should_stop) { 29054d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2906197f6accSTejun Heo set_pf_worker(false); 29074d595b86SLai Jiangshan return 0; 29084d595b86SLai Jiangshan } 29094d595b86SLai Jiangshan 2910111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2911111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2912e22bee78STejun Heo schedule(); 2913e22bee78STejun Heo goto repeat; 29141da177e4SLinus Torvalds } 29151da177e4SLinus Torvalds 2916fca839c0STejun Heo /** 2917fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2918fca839c0STejun Heo * @target_wq: workqueue being flushed 2919fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2920fca839c0STejun Heo * 2921fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2922fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2923fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2924fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2925fca839c0STejun Heo * a deadlock. 2926fca839c0STejun Heo */ 2927fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2928fca839c0STejun Heo struct work_struct *target_work) 2929fca839c0STejun Heo { 2930fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2931fca839c0STejun Heo struct worker *worker; 2932fca839c0STejun Heo 2933fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2934fca839c0STejun Heo return; 2935fca839c0STejun Heo 2936fca839c0STejun Heo worker = current_wq_worker(); 2937fca839c0STejun Heo 2938fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2939d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2940fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 294123d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 294223d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2943d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2944fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2945fca839c0STejun Heo target_wq->name, target_func); 2946fca839c0STejun Heo } 2947fca839c0STejun Heo 2948fc2e4d70SOleg Nesterov struct wq_barrier { 2949fc2e4d70SOleg Nesterov struct work_struct work; 2950fc2e4d70SOleg Nesterov struct completion done; 29512607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2952fc2e4d70SOleg Nesterov }; 2953fc2e4d70SOleg Nesterov 2954fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2955fc2e4d70SOleg Nesterov { 2956fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2957fc2e4d70SOleg Nesterov complete(&barr->done); 2958fc2e4d70SOleg Nesterov } 2959fc2e4d70SOleg Nesterov 29604690c4abSTejun Heo /** 29614690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2962112202d9STejun Heo * @pwq: pwq to insert barrier into 29634690c4abSTejun Heo * @barr: wq_barrier to insert 2964affee4b2STejun Heo * @target: target work to attach @barr to 2965affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29664690c4abSTejun Heo * 2967affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2968affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2969affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2970affee4b2STejun Heo * cpu. 2971affee4b2STejun Heo * 2972affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2973affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2974affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2975affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2976affee4b2STejun Heo * after a work with LINKED flag set. 2977affee4b2STejun Heo * 2978affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2979112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 29804690c4abSTejun Heo * 29814690c4abSTejun Heo * CONTEXT: 2982a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 29834690c4abSTejun Heo */ 2984112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2985affee4b2STejun Heo struct wq_barrier *barr, 2986affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2987fc2e4d70SOleg Nesterov { 2988d812796eSLai Jiangshan unsigned int work_flags = 0; 2989d812796eSLai Jiangshan unsigned int work_color; 2990affee4b2STejun Heo struct list_head *head; 2991affee4b2STejun Heo 2992dc186ad7SThomas Gleixner /* 2993d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2994dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2995dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2996dc186ad7SThomas Gleixner * might deadlock. 2997dc186ad7SThomas Gleixner */ 2998ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 299922df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 300052fa5bc5SBoqun Feng 3001fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 3002fd1a5b04SByungchul Park 30032607d7a6STejun Heo barr->task = current; 300483c22520SOleg Nesterov 3005018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 3006018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 3007018f3a13SLai Jiangshan 3008affee4b2STejun Heo /* 3009affee4b2STejun Heo * If @target is currently being executed, schedule the 3010affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 3011affee4b2STejun Heo */ 3012d812796eSLai Jiangshan if (worker) { 3013affee4b2STejun Heo head = worker->scheduled.next; 3014d812796eSLai Jiangshan work_color = worker->current_color; 3015d812796eSLai Jiangshan } else { 3016affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 3017affee4b2STejun Heo 3018affee4b2STejun Heo head = target->entry.next; 3019affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 3020d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 3021d812796eSLai Jiangshan work_color = get_work_color(*bits); 3022affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3023affee4b2STejun Heo } 3024affee4b2STejun Heo 3025d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 3026d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 3027d812796eSLai Jiangshan 3028dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 3029d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 3030fc2e4d70SOleg Nesterov } 3031fc2e4d70SOleg Nesterov 303273f53c4aSTejun Heo /** 3033112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 303473f53c4aSTejun Heo * @wq: workqueue being flushed 303573f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 303673f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 303773f53c4aSTejun Heo * 3038112202d9STejun Heo * Prepare pwqs for workqueue flushing. 303973f53c4aSTejun Heo * 3040112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 3041112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 3042112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3043112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 3044112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 304573f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 304673f53c4aSTejun Heo * 304773f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 304873f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 304973f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 305073f53c4aSTejun Heo * is returned. 305173f53c4aSTejun Heo * 3052112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 305373f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 305473f53c4aSTejun Heo * advanced to @work_color. 305573f53c4aSTejun Heo * 305673f53c4aSTejun Heo * CONTEXT: 30573c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 305873f53c4aSTejun Heo * 3059d185af30SYacine Belkadi * Return: 306073f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 306173f53c4aSTejun Heo * otherwise. 306273f53c4aSTejun Heo */ 3063112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 306473f53c4aSTejun Heo int flush_color, int work_color) 30651da177e4SLinus Torvalds { 306673f53c4aSTejun Heo bool wait = false; 306749e3cf44STejun Heo struct pool_workqueue *pwq; 30681da177e4SLinus Torvalds 306973f53c4aSTejun Heo if (flush_color >= 0) { 30706183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3071112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3072dc186ad7SThomas Gleixner } 307314441960SOleg Nesterov 307449e3cf44STejun Heo for_each_pwq(pwq, wq) { 3075112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30761da177e4SLinus Torvalds 3077a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 307873f53c4aSTejun Heo 307973f53c4aSTejun Heo if (flush_color >= 0) { 30806183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 308173f53c4aSTejun Heo 3082112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3083112202d9STejun Heo pwq->flush_color = flush_color; 3084112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 308573f53c4aSTejun Heo wait = true; 30861da177e4SLinus Torvalds } 308773f53c4aSTejun Heo } 308873f53c4aSTejun Heo 308973f53c4aSTejun Heo if (work_color >= 0) { 30906183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3091112202d9STejun Heo pwq->work_color = work_color; 309273f53c4aSTejun Heo } 309373f53c4aSTejun Heo 3094a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 30951da177e4SLinus Torvalds } 30961da177e4SLinus Torvalds 3097112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 309873f53c4aSTejun Heo complete(&wq->first_flusher->done); 309973f53c4aSTejun Heo 310073f53c4aSTejun Heo return wait; 310183c22520SOleg Nesterov } 31021da177e4SLinus Torvalds 31030fcb78c2SRolf Eike Beer /** 3104c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 31050fcb78c2SRolf Eike Beer * @wq: workqueue to flush 31061da177e4SLinus Torvalds * 3107c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3108c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 31091da177e4SLinus Torvalds */ 3110c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 31111da177e4SLinus Torvalds { 311273f53c4aSTejun Heo struct wq_flusher this_flusher = { 311373f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 311473f53c4aSTejun Heo .flush_color = -1, 3115fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 311673f53c4aSTejun Heo }; 311773f53c4aSTejun Heo int next_color; 3118b1f4ec17SOleg Nesterov 31193347fa09STejun Heo if (WARN_ON(!wq_online)) 31203347fa09STejun Heo return; 31213347fa09STejun Heo 312287915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 312387915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 312487915adcSJohannes Berg 31253c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 312673f53c4aSTejun Heo 312773f53c4aSTejun Heo /* 312873f53c4aSTejun Heo * Start-to-wait phase 312973f53c4aSTejun Heo */ 313073f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 313173f53c4aSTejun Heo 313273f53c4aSTejun Heo if (next_color != wq->flush_color) { 313373f53c4aSTejun Heo /* 313473f53c4aSTejun Heo * Color space is not full. The current work_color 313573f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 313673f53c4aSTejun Heo * by one. 313773f53c4aSTejun Heo */ 31386183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 313973f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 314073f53c4aSTejun Heo wq->work_color = next_color; 314173f53c4aSTejun Heo 314273f53c4aSTejun Heo if (!wq->first_flusher) { 314373f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 31446183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 314573f53c4aSTejun Heo 314673f53c4aSTejun Heo wq->first_flusher = &this_flusher; 314773f53c4aSTejun Heo 3148112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 314973f53c4aSTejun Heo wq->work_color)) { 315073f53c4aSTejun Heo /* nothing to flush, done */ 315173f53c4aSTejun Heo wq->flush_color = next_color; 315273f53c4aSTejun Heo wq->first_flusher = NULL; 315373f53c4aSTejun Heo goto out_unlock; 315473f53c4aSTejun Heo } 315573f53c4aSTejun Heo } else { 315673f53c4aSTejun Heo /* wait in queue */ 31576183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 315873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3159112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 316073f53c4aSTejun Heo } 316173f53c4aSTejun Heo } else { 316273f53c4aSTejun Heo /* 316373f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 316473f53c4aSTejun Heo * The next flush completion will assign us 316573f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 316673f53c4aSTejun Heo */ 316773f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 316873f53c4aSTejun Heo } 316973f53c4aSTejun Heo 3170fca839c0STejun Heo check_flush_dependency(wq, NULL); 3171fca839c0STejun Heo 31723c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 317373f53c4aSTejun Heo 317473f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 317573f53c4aSTejun Heo 317673f53c4aSTejun Heo /* 317773f53c4aSTejun Heo * Wake-up-and-cascade phase 317873f53c4aSTejun Heo * 317973f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 318073f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 318173f53c4aSTejun Heo */ 318200d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 318373f53c4aSTejun Heo return; 318473f53c4aSTejun Heo 31853c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 318673f53c4aSTejun Heo 31874ce48b37STejun Heo /* we might have raced, check again with mutex held */ 31884ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 31894ce48b37STejun Heo goto out_unlock; 31904ce48b37STejun Heo 319100d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 319273f53c4aSTejun Heo 31936183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 31946183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 319573f53c4aSTejun Heo 319673f53c4aSTejun Heo while (true) { 319773f53c4aSTejun Heo struct wq_flusher *next, *tmp; 319873f53c4aSTejun Heo 319973f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 320073f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 320173f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 320273f53c4aSTejun Heo break; 320373f53c4aSTejun Heo list_del_init(&next->list); 320473f53c4aSTejun Heo complete(&next->done); 320573f53c4aSTejun Heo } 320673f53c4aSTejun Heo 32076183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 320873f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 320973f53c4aSTejun Heo 321073f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 321173f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 321273f53c4aSTejun Heo 321373f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 321473f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 321573f53c4aSTejun Heo /* 321673f53c4aSTejun Heo * Assign the same color to all overflowed 321773f53c4aSTejun Heo * flushers, advance work_color and append to 321873f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 321973f53c4aSTejun Heo * phase for these overflowed flushers. 322073f53c4aSTejun Heo */ 322173f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 322273f53c4aSTejun Heo tmp->flush_color = wq->work_color; 322373f53c4aSTejun Heo 322473f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 322573f53c4aSTejun Heo 322673f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 322773f53c4aSTejun Heo &wq->flusher_queue); 3228112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 322973f53c4aSTejun Heo } 323073f53c4aSTejun Heo 323173f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 32326183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 323373f53c4aSTejun Heo break; 323473f53c4aSTejun Heo } 323573f53c4aSTejun Heo 323673f53c4aSTejun Heo /* 323773f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3238112202d9STejun Heo * the new first flusher and arm pwqs. 323973f53c4aSTejun Heo */ 32406183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 32416183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 324273f53c4aSTejun Heo 324373f53c4aSTejun Heo list_del_init(&next->list); 324473f53c4aSTejun Heo wq->first_flusher = next; 324573f53c4aSTejun Heo 3246112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 324773f53c4aSTejun Heo break; 324873f53c4aSTejun Heo 324973f53c4aSTejun Heo /* 325073f53c4aSTejun Heo * Meh... this color is already done, clear first 325173f53c4aSTejun Heo * flusher and repeat cascading. 325273f53c4aSTejun Heo */ 325373f53c4aSTejun Heo wq->first_flusher = NULL; 325473f53c4aSTejun Heo } 325573f53c4aSTejun Heo 325673f53c4aSTejun Heo out_unlock: 32573c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32581da177e4SLinus Torvalds } 3259c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32601da177e4SLinus Torvalds 32619c5a2ba7STejun Heo /** 32629c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32639c5a2ba7STejun Heo * @wq: workqueue to drain 32649c5a2ba7STejun Heo * 32659c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32669c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32679c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3268b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32699c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32709c5a2ba7STejun Heo * takes too long. 32719c5a2ba7STejun Heo */ 32729c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32739c5a2ba7STejun Heo { 32749c5a2ba7STejun Heo unsigned int flush_cnt = 0; 327549e3cf44STejun Heo struct pool_workqueue *pwq; 32769c5a2ba7STejun Heo 32779c5a2ba7STejun Heo /* 32789c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 32799c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3280618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 32819c5a2ba7STejun Heo */ 328287fc741eSLai Jiangshan mutex_lock(&wq->mutex); 32839c5a2ba7STejun Heo if (!wq->nr_drainers++) 3284618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 328587fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32869c5a2ba7STejun Heo reflush: 3287c4f135d6STetsuo Handa __flush_workqueue(wq); 32889c5a2ba7STejun Heo 3289b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 329076af4d93STejun Heo 329149e3cf44STejun Heo for_each_pwq(pwq, wq) { 3292fa2563e4SThomas Tuttle bool drained; 32939c5a2ba7STejun Heo 3294a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3295f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3296a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3297fa2563e4SThomas Tuttle 3298fa2563e4SThomas Tuttle if (drained) 32999c5a2ba7STejun Heo continue; 33009c5a2ba7STejun Heo 33019c5a2ba7STejun Heo if (++flush_cnt == 10 || 33029c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3303e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3304e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 330576af4d93STejun Heo 3306b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 33079c5a2ba7STejun Heo goto reflush; 33089c5a2ba7STejun Heo } 33099c5a2ba7STejun Heo 33109c5a2ba7STejun Heo if (!--wq->nr_drainers) 3311618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 331287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 33139c5a2ba7STejun Heo } 33149c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 33159c5a2ba7STejun Heo 3316d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3317d6e89786SJohannes Berg bool from_cancel) 3318baf59022STejun Heo { 3319baf59022STejun Heo struct worker *worker = NULL; 3320c9e7cf27STejun Heo struct worker_pool *pool; 3321112202d9STejun Heo struct pool_workqueue *pwq; 3322baf59022STejun Heo 3323baf59022STejun Heo might_sleep(); 3324baf59022STejun Heo 332524acfb71SThomas Gleixner rcu_read_lock(); 3326fa1b54e6STejun Heo pool = get_work_pool(work); 3327fa1b54e6STejun Heo if (!pool) { 332824acfb71SThomas Gleixner rcu_read_unlock(); 3329fa1b54e6STejun Heo return false; 3330fa1b54e6STejun Heo } 3331fa1b54e6STejun Heo 3332a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 33330b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3334112202d9STejun Heo pwq = get_work_pwq(work); 3335112202d9STejun Heo if (pwq) { 3336112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3337baf59022STejun Heo goto already_gone; 3338606a5020STejun Heo } else { 3339c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3340baf59022STejun Heo if (!worker) 3341baf59022STejun Heo goto already_gone; 3342112202d9STejun Heo pwq = worker->current_pwq; 3343606a5020STejun Heo } 3344baf59022STejun Heo 3345fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3346fca839c0STejun Heo 3347112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3348a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3349baf59022STejun Heo 3350e159489bSTejun Heo /* 3351a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3352a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3353a1d14934SPeter Zijlstra * 3354a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3355a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3356a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3357a1d14934SPeter Zijlstra * forward progress. 3358e159489bSTejun Heo */ 3359d6e89786SJohannes Berg if (!from_cancel && 3360d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3361112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3362112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3363a1d14934SPeter Zijlstra } 336424acfb71SThomas Gleixner rcu_read_unlock(); 3365baf59022STejun Heo return true; 3366baf59022STejun Heo already_gone: 3367a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 336824acfb71SThomas Gleixner rcu_read_unlock(); 3369baf59022STejun Heo return false; 3370baf59022STejun Heo } 3371baf59022STejun Heo 3372d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3373d6e89786SJohannes Berg { 3374d6e89786SJohannes Berg struct wq_barrier barr; 3375d6e89786SJohannes Berg 3376d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3377d6e89786SJohannes Berg return false; 3378d6e89786SJohannes Berg 33794d43d395STetsuo Handa if (WARN_ON(!work->func)) 33804d43d395STetsuo Handa return false; 33814d43d395STetsuo Handa 338287915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 338387915adcSJohannes Berg lock_map_release(&work->lockdep_map); 338487915adcSJohannes Berg 3385d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3386d6e89786SJohannes Berg wait_for_completion(&barr.done); 3387d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3388d6e89786SJohannes Berg return true; 3389d6e89786SJohannes Berg } else { 3390d6e89786SJohannes Berg return false; 3391d6e89786SJohannes Berg } 3392d6e89786SJohannes Berg } 3393d6e89786SJohannes Berg 3394db700897SOleg Nesterov /** 3395401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3396401a8d04STejun Heo * @work: the work to flush 3397db700897SOleg Nesterov * 3398606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3399606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3400401a8d04STejun Heo * 3401d185af30SYacine Belkadi * Return: 3402401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3403401a8d04STejun Heo * %false if it was already idle. 3404db700897SOleg Nesterov */ 3405401a8d04STejun Heo bool flush_work(struct work_struct *work) 3406db700897SOleg Nesterov { 3407d6e89786SJohannes Berg return __flush_work(work, false); 3408606a5020STejun Heo } 3409db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3410db700897SOleg Nesterov 34118603e1b3STejun Heo struct cwt_wait { 3412ac6424b9SIngo Molnar wait_queue_entry_t wait; 34138603e1b3STejun Heo struct work_struct *work; 34148603e1b3STejun Heo }; 34158603e1b3STejun Heo 3416ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 34178603e1b3STejun Heo { 34188603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 34198603e1b3STejun Heo 34208603e1b3STejun Heo if (cwait->work != key) 34218603e1b3STejun Heo return 0; 34228603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 34238603e1b3STejun Heo } 34248603e1b3STejun Heo 342536e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3426401a8d04STejun Heo { 34278603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3428bbb68dfaSTejun Heo unsigned long flags; 34291f1f642eSOleg Nesterov int ret; 34301f1f642eSOleg Nesterov 34311f1f642eSOleg Nesterov do { 3432bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3433bbb68dfaSTejun Heo /* 34348603e1b3STejun Heo * If someone else is already canceling, wait for it to 34358603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 34368603e1b3STejun Heo * because we may get scheduled between @work's completion 34378603e1b3STejun Heo * and the other canceling task resuming and clearing 34388603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 34398603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 34408603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 34418603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 34428603e1b3STejun Heo * we're hogging the CPU. 34438603e1b3STejun Heo * 34448603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 34458603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 34468603e1b3STejun Heo * wake function which matches @work along with exclusive 34478603e1b3STejun Heo * wait and wakeup. 3448bbb68dfaSTejun Heo */ 34498603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 34508603e1b3STejun Heo struct cwt_wait cwait; 34518603e1b3STejun Heo 34528603e1b3STejun Heo init_wait(&cwait.wait); 34538603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34548603e1b3STejun Heo cwait.work = work; 34558603e1b3STejun Heo 34568603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34578603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34588603e1b3STejun Heo if (work_is_canceling(work)) 34598603e1b3STejun Heo schedule(); 34608603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34618603e1b3STejun Heo } 34621f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34631f1f642eSOleg Nesterov 3464bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3465bbb68dfaSTejun Heo mark_work_canceling(work); 3466bbb68dfaSTejun Heo local_irq_restore(flags); 3467bbb68dfaSTejun Heo 34683347fa09STejun Heo /* 34693347fa09STejun Heo * This allows canceling during early boot. We know that @work 34703347fa09STejun Heo * isn't executing. 34713347fa09STejun Heo */ 34723347fa09STejun Heo if (wq_online) 3473d6e89786SJohannes Berg __flush_work(work, true); 34743347fa09STejun Heo 34757a22ad75STejun Heo clear_work_data(work); 34768603e1b3STejun Heo 34778603e1b3STejun Heo /* 34788603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 34798603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 34808603e1b3STejun Heo * visible there. 34818603e1b3STejun Heo */ 34828603e1b3STejun Heo smp_mb(); 34838603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 34848603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 34858603e1b3STejun Heo 34861f1f642eSOleg Nesterov return ret; 34871f1f642eSOleg Nesterov } 34881f1f642eSOleg Nesterov 34896e84d644SOleg Nesterov /** 3490401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3491401a8d04STejun Heo * @work: the work to cancel 34926e84d644SOleg Nesterov * 3493401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3494401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3495401a8d04STejun Heo * another workqueue. On return from this function, @work is 3496401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 34971f1f642eSOleg Nesterov * 3498401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3499401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 35006e84d644SOleg Nesterov * 3501401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 35026e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3503401a8d04STejun Heo * 3504d185af30SYacine Belkadi * Return: 3505401a8d04STejun Heo * %true if @work was pending, %false otherwise. 35066e84d644SOleg Nesterov */ 3507401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 35086e84d644SOleg Nesterov { 350936e227d2STejun Heo return __cancel_work_timer(work, false); 3510b89deed3SOleg Nesterov } 351128e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3512b89deed3SOleg Nesterov 35136e84d644SOleg Nesterov /** 3514401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3515401a8d04STejun Heo * @dwork: the delayed work to flush 35166e84d644SOleg Nesterov * 3517401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3518401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3519401a8d04STejun Heo * considers the last queueing instance of @dwork. 35201f1f642eSOleg Nesterov * 3521d185af30SYacine Belkadi * Return: 3522401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3523401a8d04STejun Heo * %false if it was already idle. 35246e84d644SOleg Nesterov */ 3525401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3526401a8d04STejun Heo { 35278930cabaSTejun Heo local_irq_disable(); 3528401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 352960c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 35308930cabaSTejun Heo local_irq_enable(); 3531401a8d04STejun Heo return flush_work(&dwork->work); 3532401a8d04STejun Heo } 3533401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3534401a8d04STejun Heo 353505f0fe6bSTejun Heo /** 353605f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 353705f0fe6bSTejun Heo * @rwork: the rcu work to flush 353805f0fe6bSTejun Heo * 353905f0fe6bSTejun Heo * Return: 354005f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 354105f0fe6bSTejun Heo * %false if it was already idle. 354205f0fe6bSTejun Heo */ 354305f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 354405f0fe6bSTejun Heo { 354505f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 354605f0fe6bSTejun Heo rcu_barrier(); 354705f0fe6bSTejun Heo flush_work(&rwork->work); 354805f0fe6bSTejun Heo return true; 354905f0fe6bSTejun Heo } else { 355005f0fe6bSTejun Heo return flush_work(&rwork->work); 355105f0fe6bSTejun Heo } 355205f0fe6bSTejun Heo } 355305f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 355405f0fe6bSTejun Heo 3555f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3556f72b8792SJens Axboe { 3557f72b8792SJens Axboe unsigned long flags; 3558f72b8792SJens Axboe int ret; 3559f72b8792SJens Axboe 3560f72b8792SJens Axboe do { 3561f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3562f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3563f72b8792SJens Axboe 3564f72b8792SJens Axboe if (unlikely(ret < 0)) 3565f72b8792SJens Axboe return false; 3566f72b8792SJens Axboe 3567f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3568f72b8792SJens Axboe local_irq_restore(flags); 3569f72b8792SJens Axboe return ret; 3570f72b8792SJens Axboe } 3571f72b8792SJens Axboe 357273b4b532SAndrey Grodzovsky /* 357373b4b532SAndrey Grodzovsky * See cancel_delayed_work() 357473b4b532SAndrey Grodzovsky */ 357573b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 357673b4b532SAndrey Grodzovsky { 357773b4b532SAndrey Grodzovsky return __cancel_work(work, false); 357873b4b532SAndrey Grodzovsky } 357973b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 358073b4b532SAndrey Grodzovsky 3581401a8d04STejun Heo /** 358257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 358357b30ae7STejun Heo * @dwork: delayed_work to cancel 358409383498STejun Heo * 3585d185af30SYacine Belkadi * Kill off a pending delayed_work. 3586d185af30SYacine Belkadi * 3587d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3588d185af30SYacine Belkadi * pending. 3589d185af30SYacine Belkadi * 3590d185af30SYacine Belkadi * Note: 3591d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3592d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3593d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 359409383498STejun Heo * 359557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 359609383498STejun Heo */ 359757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 359809383498STejun Heo { 3599f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 360009383498STejun Heo } 360157b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 360209383498STejun Heo 360309383498STejun Heo /** 3604401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3605401a8d04STejun Heo * @dwork: the delayed work cancel 3606401a8d04STejun Heo * 3607401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3608401a8d04STejun Heo * 3609d185af30SYacine Belkadi * Return: 3610401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3611401a8d04STejun Heo */ 3612401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 36136e84d644SOleg Nesterov { 361436e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 36156e84d644SOleg Nesterov } 3616f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 36171da177e4SLinus Torvalds 36180fcb78c2SRolf Eike Beer /** 361931ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3620b6136773SAndrew Morton * @func: the function to call 3621b6136773SAndrew Morton * 362231ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 362331ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3624b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 362531ddd871STejun Heo * 3626d185af30SYacine Belkadi * Return: 362731ddd871STejun Heo * 0 on success, -errno on failure. 3628b6136773SAndrew Morton */ 362965f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 363015316ba8SChristoph Lameter { 363115316ba8SChristoph Lameter int cpu; 363238f51568SNamhyung Kim struct work_struct __percpu *works; 363315316ba8SChristoph Lameter 3634b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3635b6136773SAndrew Morton if (!works) 363615316ba8SChristoph Lameter return -ENOMEM; 3637b6136773SAndrew Morton 3638ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 363993981800STejun Heo 364015316ba8SChristoph Lameter for_each_online_cpu(cpu) { 36419bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 36429bfb1839SIngo Molnar 36439bfb1839SIngo Molnar INIT_WORK(work, func); 36448de6d308SOleg Nesterov schedule_work_on(cpu, work); 364515316ba8SChristoph Lameter } 364693981800STejun Heo 364793981800STejun Heo for_each_online_cpu(cpu) 36488616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 364993981800STejun Heo 3650ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3651b6136773SAndrew Morton free_percpu(works); 365215316ba8SChristoph Lameter return 0; 365315316ba8SChristoph Lameter } 365415316ba8SChristoph Lameter 3655eef6a7d5SAlan Stern /** 36561fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36571fa44ecaSJames Bottomley * @fn: the function to execute 36581fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36591fa44ecaSJames Bottomley * be available when the work executes) 36601fa44ecaSJames Bottomley * 36611fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36621fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36631fa44ecaSJames Bottomley * 3664d185af30SYacine Belkadi * Return: 0 - function was executed 36651fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36661fa44ecaSJames Bottomley */ 366765f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36681fa44ecaSJames Bottomley { 36691fa44ecaSJames Bottomley if (!in_interrupt()) { 367065f27f38SDavid Howells fn(&ew->work); 36711fa44ecaSJames Bottomley return 0; 36721fa44ecaSJames Bottomley } 36731fa44ecaSJames Bottomley 367465f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36751fa44ecaSJames Bottomley schedule_work(&ew->work); 36761fa44ecaSJames Bottomley 36771fa44ecaSJames Bottomley return 1; 36781fa44ecaSJames Bottomley } 36791fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 36801fa44ecaSJames Bottomley 36817a4e344cSTejun Heo /** 36827a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 36837a4e344cSTejun Heo * @attrs: workqueue_attrs to free 36847a4e344cSTejun Heo * 36857a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 36867a4e344cSTejun Heo */ 3687513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 36887a4e344cSTejun Heo { 36897a4e344cSTejun Heo if (attrs) { 36907a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 36917a4e344cSTejun Heo kfree(attrs); 36927a4e344cSTejun Heo } 36937a4e344cSTejun Heo } 36947a4e344cSTejun Heo 36957a4e344cSTejun Heo /** 36967a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 36977a4e344cSTejun Heo * 36987a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3699d185af30SYacine Belkadi * return it. 3700d185af30SYacine Belkadi * 3701d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 37027a4e344cSTejun Heo */ 3703513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 37047a4e344cSTejun Heo { 37057a4e344cSTejun Heo struct workqueue_attrs *attrs; 37067a4e344cSTejun Heo 3707be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 37087a4e344cSTejun Heo if (!attrs) 37097a4e344cSTejun Heo goto fail; 3710be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 37117a4e344cSTejun Heo goto fail; 37127a4e344cSTejun Heo 371313e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 37147a4e344cSTejun Heo return attrs; 37157a4e344cSTejun Heo fail: 37167a4e344cSTejun Heo free_workqueue_attrs(attrs); 37177a4e344cSTejun Heo return NULL; 37187a4e344cSTejun Heo } 37197a4e344cSTejun Heo 372029c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 372129c91e99STejun Heo const struct workqueue_attrs *from) 372229c91e99STejun Heo { 372329c91e99STejun Heo to->nice = from->nice; 372429c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 37252865a8fbSShaohua Li /* 37262865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 37272865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 37282865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 37292865a8fbSShaohua Li */ 37302865a8fbSShaohua Li to->no_numa = from->no_numa; 373129c91e99STejun Heo } 373229c91e99STejun Heo 373329c91e99STejun Heo /* hash value of the content of @attr */ 373429c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 373529c91e99STejun Heo { 373629c91e99STejun Heo u32 hash = 0; 373729c91e99STejun Heo 373829c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 373913e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 374013e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 374129c91e99STejun Heo return hash; 374229c91e99STejun Heo } 374329c91e99STejun Heo 374429c91e99STejun Heo /* content equality test */ 374529c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 374629c91e99STejun Heo const struct workqueue_attrs *b) 374729c91e99STejun Heo { 374829c91e99STejun Heo if (a->nice != b->nice) 374929c91e99STejun Heo return false; 375029c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 375129c91e99STejun Heo return false; 375229c91e99STejun Heo return true; 375329c91e99STejun Heo } 375429c91e99STejun Heo 37557a4e344cSTejun Heo /** 37567a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 37577a4e344cSTejun Heo * @pool: worker_pool to initialize 37587a4e344cSTejun Heo * 3759402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3760d185af30SYacine Belkadi * 3761d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 376229c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 376329c91e99STejun Heo * on @pool safely to release it. 37647a4e344cSTejun Heo */ 37657a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 37664e1a1f9aSTejun Heo { 3767a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 376829c91e99STejun Heo pool->id = -1; 376929c91e99STejun Heo pool->cpu = -1; 3770f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 37714e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 377282607adcSTejun Heo pool->watchdog_ts = jiffies; 37734e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37744e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 37754e1a1f9aSTejun Heo hash_init(pool->busy_hash); 37764e1a1f9aSTejun Heo 377732a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 37783f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 37794e1a1f9aSTejun Heo 378032a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 37814e1a1f9aSTejun Heo 3782da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3783e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 37847a4e344cSTejun Heo 37857cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 378629c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 378729c91e99STejun Heo pool->refcnt = 1; 378829c91e99STejun Heo 378929c91e99STejun Heo /* shouldn't fail above this point */ 3790be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 37917a4e344cSTejun Heo if (!pool->attrs) 37927a4e344cSTejun Heo return -ENOMEM; 37937a4e344cSTejun Heo return 0; 37944e1a1f9aSTejun Heo } 37954e1a1f9aSTejun Heo 3796669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3797669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3798669de8bdSBart Van Assche { 3799669de8bdSBart Van Assche char *lock_name; 3800669de8bdSBart Van Assche 3801669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3802669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3803669de8bdSBart Van Assche if (!lock_name) 3804669de8bdSBart Van Assche lock_name = wq->name; 380569a106c0SQian Cai 380669a106c0SQian Cai wq->lock_name = lock_name; 3807669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3808669de8bdSBart Van Assche } 3809669de8bdSBart Van Assche 3810669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3811669de8bdSBart Van Assche { 3812669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3813669de8bdSBart Van Assche } 3814669de8bdSBart Van Assche 3815669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3816669de8bdSBart Van Assche { 3817669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3818669de8bdSBart Van Assche kfree(wq->lock_name); 3819669de8bdSBart Van Assche } 3820669de8bdSBart Van Assche #else 3821669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3822669de8bdSBart Van Assche { 3823669de8bdSBart Van Assche } 3824669de8bdSBart Van Assche 3825669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3826669de8bdSBart Van Assche { 3827669de8bdSBart Van Assche } 3828669de8bdSBart Van Assche 3829669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3830669de8bdSBart Van Assche { 3831669de8bdSBart Van Assche } 3832669de8bdSBart Van Assche #endif 3833669de8bdSBart Van Assche 3834e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3835e2dca7adSTejun Heo { 3836e2dca7adSTejun Heo struct workqueue_struct *wq = 3837e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3838e2dca7adSTejun Heo 3839669de8bdSBart Van Assche wq_free_lockdep(wq); 3840669de8bdSBart Van Assche 3841e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3842e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3843e2dca7adSTejun Heo else 3844e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3845e2dca7adSTejun Heo 3846e2dca7adSTejun Heo kfree(wq); 3847e2dca7adSTejun Heo } 3848e2dca7adSTejun Heo 384929c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 385029c91e99STejun Heo { 385129c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 385229c91e99STejun Heo 38537cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 385429c91e99STejun Heo free_workqueue_attrs(pool->attrs); 385529c91e99STejun Heo kfree(pool); 385629c91e99STejun Heo } 385729c91e99STejun Heo 385829c91e99STejun Heo /** 385929c91e99STejun Heo * put_unbound_pool - put a worker_pool 386029c91e99STejun Heo * @pool: worker_pool to put 386129c91e99STejun Heo * 386224acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3863c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3864c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3865c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3866a892caccSTejun Heo * 3867a892caccSTejun Heo * Should be called with wq_pool_mutex held. 386829c91e99STejun Heo */ 386929c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 387029c91e99STejun Heo { 387160f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 3872e02b9312SValentin Schneider struct list_head cull_list; 387329c91e99STejun Heo struct worker *worker; 387429c91e99STejun Heo 3875e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 3876e02b9312SValentin Schneider 3877a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3878a892caccSTejun Heo 3879a892caccSTejun Heo if (--pool->refcnt) 388029c91e99STejun Heo return; 388129c91e99STejun Heo 388229c91e99STejun Heo /* sanity checks */ 388361d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3884a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 388529c91e99STejun Heo return; 388629c91e99STejun Heo 388729c91e99STejun Heo /* release id and unhash */ 388829c91e99STejun Heo if (pool->id >= 0) 388929c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 389029c91e99STejun Heo hash_del(&pool->hash_node); 389129c91e99STejun Heo 3892c5aa87bbSTejun Heo /* 3893692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3894692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3895692b4825STejun Heo * manager and @pool gets freed with the flag set. 38969ab03be4SValentin Schneider * 38979ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 38989ab03be4SValentin Schneider * only get here with 38999ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 39009ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 39019ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 39029ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 39039ab03be4SValentin Schneider * drops pool->lock 3904c5aa87bbSTejun Heo */ 39059ab03be4SValentin Schneider while (true) { 39069ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 39079ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3908d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3909e02b9312SValentin Schneider 3910e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 39119ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 39129ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3913692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 39149ab03be4SValentin Schneider break; 39159ab03be4SValentin Schneider } 39169ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3917e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 39189ab03be4SValentin Schneider } 3919692b4825STejun Heo 39201037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3921e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 392229c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3923a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 392460f5a4bcSLai Jiangshan 3925e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3926e02b9312SValentin Schneider 3927e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 392860f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 39291258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 393060f5a4bcSLai Jiangshan 393160f5a4bcSLai Jiangshan if (pool->detach_completion) 393260f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 393360f5a4bcSLai Jiangshan 393429c91e99STejun Heo /* shut down the timers */ 393529c91e99STejun Heo del_timer_sync(&pool->idle_timer); 39363f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 393729c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 393829c91e99STejun Heo 393924acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 394025b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 394129c91e99STejun Heo } 394229c91e99STejun Heo 394329c91e99STejun Heo /** 394429c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 394529c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 394629c91e99STejun Heo * 394729c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 394829c91e99STejun Heo * reference count and return it. If there already is a matching 394929c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3950d185af30SYacine Belkadi * create a new one. 3951a892caccSTejun Heo * 3952a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3953d185af30SYacine Belkadi * 3954d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3955d185af30SYacine Belkadi * On failure, %NULL. 395629c91e99STejun Heo */ 395729c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 395829c91e99STejun Heo { 395929c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 396029c91e99STejun Heo struct worker_pool *pool; 3961f3f90ad4STejun Heo int node; 3962e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 396329c91e99STejun Heo 3964a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 396529c91e99STejun Heo 396629c91e99STejun Heo /* do we already have a matching pool? */ 396729c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 396829c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 396929c91e99STejun Heo pool->refcnt++; 39703fb1823cSLai Jiangshan return pool; 397129c91e99STejun Heo } 397229c91e99STejun Heo } 397329c91e99STejun Heo 3974e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3975e2273584SXunlei Pang if (wq_numa_enabled) { 3976e2273584SXunlei Pang for_each_node(node) { 3977e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3978e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3979e2273584SXunlei Pang target_node = node; 3980e2273584SXunlei Pang break; 3981e2273584SXunlei Pang } 3982e2273584SXunlei Pang } 3983e2273584SXunlei Pang } 3984e2273584SXunlei Pang 398529c91e99STejun Heo /* nope, create a new one */ 3986e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 398729c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 398829c91e99STejun Heo goto fail; 398929c91e99STejun Heo 39908864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 399129c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3992e2273584SXunlei Pang pool->node = target_node; 399329c91e99STejun Heo 39942865a8fbSShaohua Li /* 39952865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 39962865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 39972865a8fbSShaohua Li */ 39982865a8fbSShaohua Li pool->attrs->no_numa = false; 39992865a8fbSShaohua Li 400029c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 400129c91e99STejun Heo goto fail; 400229c91e99STejun Heo 400329c91e99STejun Heo /* create and start the initial worker */ 40043347fa09STejun Heo if (wq_online && !create_worker(pool)) 400529c91e99STejun Heo goto fail; 400629c91e99STejun Heo 400729c91e99STejun Heo /* install */ 400829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 40093fb1823cSLai Jiangshan 401029c91e99STejun Heo return pool; 401129c91e99STejun Heo fail: 401229c91e99STejun Heo if (pool) 401329c91e99STejun Heo put_unbound_pool(pool); 401429c91e99STejun Heo return NULL; 401529c91e99STejun Heo } 401629c91e99STejun Heo 40178864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 40188864b4e5STejun Heo { 40198864b4e5STejun Heo kmem_cache_free(pwq_cache, 40208864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 40218864b4e5STejun Heo } 40228864b4e5STejun Heo 40238864b4e5STejun Heo /* 40248864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 40258864b4e5STejun Heo * and needs to be destroyed. 40268864b4e5STejun Heo */ 40278864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 40288864b4e5STejun Heo { 40298864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 40308864b4e5STejun Heo unbound_release_work); 40318864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 40328864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 4033b42b0bddSYang Yingliang bool is_last = false; 40348864b4e5STejun Heo 4035b42b0bddSYang Yingliang /* 4036b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 4037b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 4038b42b0bddSYang Yingliang */ 4039b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 40408864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 40418864b4e5STejun Heo return; 40428864b4e5STejun Heo 40433c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 40448864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 4045bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 40463c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 4047b42b0bddSYang Yingliang } 40488864b4e5STejun Heo 4049a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 40508864b4e5STejun Heo put_unbound_pool(pool); 4051a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 4052a892caccSTejun Heo 405325b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 40548864b4e5STejun Heo 40558864b4e5STejun Heo /* 40568864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 4057e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 40588864b4e5STejun Heo */ 4059669de8bdSBart Van Assche if (is_last) { 4060669de8bdSBart Van Assche wq_unregister_lockdep(wq); 406125b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 40626029a918STejun Heo } 4063669de8bdSBart Van Assche } 40648864b4e5STejun Heo 40650fbd95aaSTejun Heo /** 4066699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 40670fbd95aaSTejun Heo * @pwq: target pool_workqueue 40680fbd95aaSTejun Heo * 4069699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4070f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4071699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 40720fbd95aaSTejun Heo */ 4073699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 40740fbd95aaSTejun Heo { 4075699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4076699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 40773347fa09STejun Heo unsigned long flags; 4078699ce097STejun Heo 4079699ce097STejun Heo /* for @wq->saved_max_active */ 4080a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4081699ce097STejun Heo 4082699ce097STejun Heo /* fast exit for non-freezable wqs */ 4083699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4084699ce097STejun Heo return; 4085699ce097STejun Heo 40863347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4087a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4088699ce097STejun Heo 408974b414eaSLai Jiangshan /* 409074b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 409174b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 409274b414eaSLai Jiangshan * is updated and visible. 409374b414eaSLai Jiangshan */ 409474b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 409501341fbdSYunfeng Ye bool kick = false; 409601341fbdSYunfeng Ye 4097699ce097STejun Heo pwq->max_active = wq->saved_max_active; 40980fbd95aaSTejun Heo 4099f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 410001341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 4101f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 410201341fbdSYunfeng Ye kick = true; 410301341fbdSYunfeng Ye } 4104951a078aSLai Jiangshan 4105951a078aSLai Jiangshan /* 4106951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 410701341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 410801341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 410901341fbdSYunfeng Ye * let's kick iff work items were activated. 4110951a078aSLai Jiangshan */ 411101341fbdSYunfeng Ye if (kick) 4112951a078aSLai Jiangshan wake_up_worker(pwq->pool); 4113699ce097STejun Heo } else { 4114699ce097STejun Heo pwq->max_active = 0; 4115699ce097STejun Heo } 4116699ce097STejun Heo 4117a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 41180fbd95aaSTejun Heo } 41190fbd95aaSTejun Heo 412067dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4121f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4122f147f29eSTejun Heo struct worker_pool *pool) 4123d2c1d404STejun Heo { 4124d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4125d2c1d404STejun Heo 4126e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4127e50aba9aSTejun Heo 4128d2c1d404STejun Heo pwq->pool = pool; 4129d2c1d404STejun Heo pwq->wq = wq; 4130d2c1d404STejun Heo pwq->flush_color = -1; 41318864b4e5STejun Heo pwq->refcnt = 1; 4132f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 41331befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4134d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 41358864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 4136f147f29eSTejun Heo } 4137d2c1d404STejun Heo 4138f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 41391befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4140f147f29eSTejun Heo { 4141f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4142f147f29eSTejun Heo 4143f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 414475ccf595STejun Heo 41451befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 41461befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 41471befcf30STejun Heo return; 41481befcf30STejun Heo 414929b1cb41SLai Jiangshan /* set the matching work_color */ 415075ccf595STejun Heo pwq->work_color = wq->work_color; 4151983ca25eSTejun Heo 4152983ca25eSTejun Heo /* sync max_active to the current setting */ 4153983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4154983ca25eSTejun Heo 4155983ca25eSTejun Heo /* link in @pwq */ 41569e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4157df2d5ae4STejun Heo } 41586029a918STejun Heo 4159f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4160f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4161f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4162f147f29eSTejun Heo { 4163f147f29eSTejun Heo struct worker_pool *pool; 4164f147f29eSTejun Heo struct pool_workqueue *pwq; 4165f147f29eSTejun Heo 4166f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4167f147f29eSTejun Heo 4168f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4169f147f29eSTejun Heo if (!pool) 4170f147f29eSTejun Heo return NULL; 4171f147f29eSTejun Heo 4172e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4173f147f29eSTejun Heo if (!pwq) { 4174f147f29eSTejun Heo put_unbound_pool(pool); 4175f147f29eSTejun Heo return NULL; 4176f147f29eSTejun Heo } 4177f147f29eSTejun Heo 4178f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4179f147f29eSTejun Heo return pwq; 4180d2c1d404STejun Heo } 4181d2c1d404STejun Heo 41824c16bd32STejun Heo /** 418330186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 4184042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 41854c16bd32STejun Heo * @node: the target NUMA node 41864c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 41874c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 41884c16bd32STejun Heo * 41894c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 41904c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 4191d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 41924c16bd32STejun Heo * 41934c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 41944c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 41954c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 41964c16bd32STejun Heo * @attrs->cpumask. 41974c16bd32STejun Heo * 41984c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 41994c16bd32STejun Heo * stable. 4200d185af30SYacine Belkadi * 4201d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 4202d185af30SYacine Belkadi * %false if equal. 42034c16bd32STejun Heo */ 42044c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 42054c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 42064c16bd32STejun Heo { 4207d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 42084c16bd32STejun Heo goto use_dfl; 42094c16bd32STejun Heo 42104c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 42114c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 42124c16bd32STejun Heo if (cpu_going_down >= 0) 42134c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 42144c16bd32STejun Heo 42154c16bd32STejun Heo if (cpumask_empty(cpumask)) 42164c16bd32STejun Heo goto use_dfl; 42174c16bd32STejun Heo 42184c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 42194c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 42201ad0f0a7SMichael Bringmann 42211ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 42221ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 42231ad0f0a7SMichael Bringmann "possible intersect\n"); 42241ad0f0a7SMichael Bringmann return false; 42251ad0f0a7SMichael Bringmann } 42261ad0f0a7SMichael Bringmann 42274c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 42284c16bd32STejun Heo 42294c16bd32STejun Heo use_dfl: 42304c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 42314c16bd32STejun Heo return false; 42324c16bd32STejun Heo } 42334c16bd32STejun Heo 42341befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 42351befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 42361befcf30STejun Heo int node, 42371befcf30STejun Heo struct pool_workqueue *pwq) 42381befcf30STejun Heo { 42391befcf30STejun Heo struct pool_workqueue *old_pwq; 42401befcf30STejun Heo 42415b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42421befcf30STejun Heo lockdep_assert_held(&wq->mutex); 42431befcf30STejun Heo 42441befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 42451befcf30STejun Heo link_pwq(pwq); 42461befcf30STejun Heo 42471befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42481befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 42491befcf30STejun Heo return old_pwq; 42501befcf30STejun Heo } 42511befcf30STejun Heo 42522d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 42532d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 42542d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 42552d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4256042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 42572d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 42582d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 42592d5f0764SLai Jiangshan }; 42602d5f0764SLai Jiangshan 42612d5f0764SLai Jiangshan /* free the resources after success or abort */ 42622d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 42632d5f0764SLai Jiangshan { 42642d5f0764SLai Jiangshan if (ctx) { 42652d5f0764SLai Jiangshan int node; 42662d5f0764SLai Jiangshan 42672d5f0764SLai Jiangshan for_each_node(node) 42682d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 42692d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 42702d5f0764SLai Jiangshan 42712d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 42722d5f0764SLai Jiangshan 42732d5f0764SLai Jiangshan kfree(ctx); 42742d5f0764SLai Jiangshan } 42752d5f0764SLai Jiangshan } 42762d5f0764SLai Jiangshan 42772d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 42782d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 42792d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 428099c621efSLai Jiangshan const struct workqueue_attrs *attrs, 428199c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 42822d5f0764SLai Jiangshan { 42832d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 42842d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 42852d5f0764SLai Jiangshan int node; 42862d5f0764SLai Jiangshan 42872d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42882d5f0764SLai Jiangshan 4289acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 42902d5f0764SLai Jiangshan 4291be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4292be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 42932d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 42942d5f0764SLai Jiangshan goto out_free; 42952d5f0764SLai Jiangshan 4296042f7df1SLai Jiangshan /* 429799c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 429899c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4299042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4300042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4301042f7df1SLai Jiangshan */ 43022d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 430399c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4304042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 430599c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 43062d5f0764SLai Jiangshan 43072d5f0764SLai Jiangshan /* 43082d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 43092d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 43102d5f0764SLai Jiangshan * pools. 43112d5f0764SLai Jiangshan */ 43122d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 43132d5f0764SLai Jiangshan 43142d5f0764SLai Jiangshan /* 43152d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 43162d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 43172d5f0764SLai Jiangshan * it even if we don't use it immediately. 43182d5f0764SLai Jiangshan */ 43192d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 43202d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 43212d5f0764SLai Jiangshan goto out_free; 43222d5f0764SLai Jiangshan 43232d5f0764SLai Jiangshan for_each_node(node) { 4324042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 43252d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 43262d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 43272d5f0764SLai Jiangshan goto out_free; 43282d5f0764SLai Jiangshan } else { 43292d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 43302d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 43312d5f0764SLai Jiangshan } 43322d5f0764SLai Jiangshan } 43332d5f0764SLai Jiangshan 4334042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4335042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4336042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 43372d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4338042f7df1SLai Jiangshan 43392d5f0764SLai Jiangshan ctx->wq = wq; 43402d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43412d5f0764SLai Jiangshan return ctx; 43422d5f0764SLai Jiangshan 43432d5f0764SLai Jiangshan out_free: 43442d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43452d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 43462d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 43472d5f0764SLai Jiangshan return NULL; 43482d5f0764SLai Jiangshan } 43492d5f0764SLai Jiangshan 43502d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 43512d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 43522d5f0764SLai Jiangshan { 43532d5f0764SLai Jiangshan int node; 43542d5f0764SLai Jiangshan 43552d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 43562d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 43572d5f0764SLai Jiangshan 43582d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 43592d5f0764SLai Jiangshan 43602d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 43612d5f0764SLai Jiangshan for_each_node(node) 43622d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 43632d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 43642d5f0764SLai Jiangshan 43652d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 43662d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 43672d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 43682d5f0764SLai Jiangshan 43692d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 43702d5f0764SLai Jiangshan } 43712d5f0764SLai Jiangshan 4372a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4373a0111cf6SLai Jiangshan { 4374a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4375ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4376a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4377a0111cf6SLai Jiangshan } 4378a0111cf6SLai Jiangshan 4379a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4380a0111cf6SLai Jiangshan { 4381a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4382ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4383a0111cf6SLai Jiangshan } 4384a0111cf6SLai Jiangshan 4385a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4386a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4387a0111cf6SLai Jiangshan { 4388a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4389a0111cf6SLai Jiangshan 4390a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4391a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4392a0111cf6SLai Jiangshan return -EINVAL; 4393a0111cf6SLai Jiangshan 4394a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 43950a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 43960a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4397a0111cf6SLai Jiangshan return -EINVAL; 4398a0111cf6SLai Jiangshan 43990a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 44000a94efb5STejun Heo } 44010a94efb5STejun Heo 440299c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 44036201171eSwanghaibin if (!ctx) 44046201171eSwanghaibin return -ENOMEM; 4405a0111cf6SLai Jiangshan 4406a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4407a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4408a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4409a0111cf6SLai Jiangshan 44106201171eSwanghaibin return 0; 4411a0111cf6SLai Jiangshan } 4412a0111cf6SLai Jiangshan 44139e8cd2f5STejun Heo /** 44149e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 44159e8cd2f5STejun Heo * @wq: the target workqueue 44169e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 44179e8cd2f5STejun Heo * 44184c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 44194c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 44204c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 44214c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 44224c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 44234c16bd32STejun Heo * back-to-back will stay on its current pwq. 44249e8cd2f5STejun Heo * 4425d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4426d185af30SYacine Belkadi * 4427ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4428509b3204SDaniel Jordan * 4429d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 44309e8cd2f5STejun Heo */ 4431513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 44329e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 44339e8cd2f5STejun Heo { 4434a0111cf6SLai Jiangshan int ret; 44359e8cd2f5STejun Heo 4436509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4437509b3204SDaniel Jordan 4438509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4439a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4440509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 44412d5f0764SLai Jiangshan 44422d5f0764SLai Jiangshan return ret; 44439e8cd2f5STejun Heo } 44449e8cd2f5STejun Heo 44454c16bd32STejun Heo /** 44464c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 44474c16bd32STejun Heo * @wq: the target workqueue 44484c16bd32STejun Heo * @cpu: the CPU coming up or going down 44494c16bd32STejun Heo * @online: whether @cpu is coming up or going down 44504c16bd32STejun Heo * 44514c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 44524c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 44534c16bd32STejun Heo * @wq accordingly. 44544c16bd32STejun Heo * 44554c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 44564c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 44574c16bd32STejun Heo * correct. 44584c16bd32STejun Heo * 44594c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 44604c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 44614c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 44624c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 44634c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 44644c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 44654c16bd32STejun Heo * CPU_DOWN_PREPARE. 44664c16bd32STejun Heo */ 44674c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 44684c16bd32STejun Heo bool online) 44694c16bd32STejun Heo { 44704c16bd32STejun Heo int node = cpu_to_node(cpu); 44714c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 44724c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 44734c16bd32STejun Heo struct workqueue_attrs *target_attrs; 44744c16bd32STejun Heo cpumask_t *cpumask; 44754c16bd32STejun Heo 44764c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 44774c16bd32STejun Heo 4478f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4479f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 44804c16bd32STejun Heo return; 44814c16bd32STejun Heo 44824c16bd32STejun Heo /* 44834c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 44844c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 44854c16bd32STejun Heo * CPU hotplug exclusion. 44864c16bd32STejun Heo */ 44874c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 44884c16bd32STejun Heo cpumask = target_attrs->cpumask; 44894c16bd32STejun Heo 44904c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 44914c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 44924c16bd32STejun Heo 44934c16bd32STejun Heo /* 44944c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4495042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4496042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4497042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 44984c16bd32STejun Heo */ 4499042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 45004c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4501f7142ed4SLai Jiangshan return; 45024c16bd32STejun Heo } else { 45034c16bd32STejun Heo goto use_dfl_pwq; 45044c16bd32STejun Heo } 45054c16bd32STejun Heo 45064c16bd32STejun Heo /* create a new pwq */ 45074c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 45084c16bd32STejun Heo if (!pwq) { 45092d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 45104c16bd32STejun Heo wq->name); 451177f300b1SDaeseok Youn goto use_dfl_pwq; 45124c16bd32STejun Heo } 45134c16bd32STejun Heo 4514f7142ed4SLai Jiangshan /* Install the new pwq. */ 45154c16bd32STejun Heo mutex_lock(&wq->mutex); 45164c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 45174c16bd32STejun Heo goto out_unlock; 45184c16bd32STejun Heo 45194c16bd32STejun Heo use_dfl_pwq: 4520f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4521a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 45224c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4523a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 45244c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 45254c16bd32STejun Heo out_unlock: 45264c16bd32STejun Heo mutex_unlock(&wq->mutex); 45274c16bd32STejun Heo put_pwq_unlocked(old_pwq); 45284c16bd32STejun Heo } 45294c16bd32STejun Heo 453030cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 45311da177e4SLinus Torvalds { 453249e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 45338a2b7538STejun Heo int cpu, ret; 4534e1d8aa9fSFrederic Weisbecker 453530cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4536420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4537420c0ddbSTejun Heo if (!wq->cpu_pwqs) 453830cdf249STejun Heo return -ENOMEM; 453930cdf249STejun Heo 454030cdf249STejun Heo for_each_possible_cpu(cpu) { 45417fb98ea7STejun Heo struct pool_workqueue *pwq = 45427fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 45437a62c2c8STejun Heo struct worker_pool *cpu_pools = 4544f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 454530cdf249STejun Heo 4546f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4547f147f29eSTejun Heo 4548f147f29eSTejun Heo mutex_lock(&wq->mutex); 45491befcf30STejun Heo link_pwq(pwq); 4550f147f29eSTejun Heo mutex_unlock(&wq->mutex); 455130cdf249STejun Heo } 455230cdf249STejun Heo return 0; 4553509b3204SDaniel Jordan } 4554509b3204SDaniel Jordan 4555ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4556509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 45578a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 45588a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 45598a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 45608a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 45618a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 45629e8cd2f5STejun Heo } else { 4563509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 45649e8cd2f5STejun Heo } 4565ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4566509b3204SDaniel Jordan 4567509b3204SDaniel Jordan return ret; 45680f900049STejun Heo } 45690f900049STejun Heo 4570f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4571f3421797STejun Heo const char *name) 4572b71ab8c2STejun Heo { 4573f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4574f3421797STejun Heo 4575f3421797STejun Heo if (max_active < 1 || max_active > lim) 4576044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4577f3421797STejun Heo max_active, name, 1, lim); 4578b71ab8c2STejun Heo 4579f3421797STejun Heo return clamp_val(max_active, 1, lim); 4580b71ab8c2STejun Heo } 4581b71ab8c2STejun Heo 4582983c7515STejun Heo /* 4583983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4584983c7515STejun Heo * to guarantee forward progress. 4585983c7515STejun Heo */ 4586983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4587983c7515STejun Heo { 4588983c7515STejun Heo struct worker *rescuer; 4589b92b36eaSDan Carpenter int ret; 4590983c7515STejun Heo 4591983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4592983c7515STejun Heo return 0; 4593983c7515STejun Heo 4594983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 45954c0736a7SPetr Mladek if (!rescuer) { 45964c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 45974c0736a7SPetr Mladek wq->name); 4598983c7515STejun Heo return -ENOMEM; 45994c0736a7SPetr Mladek } 4600983c7515STejun Heo 4601983c7515STejun Heo rescuer->rescue_wq = wq; 4602983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4603f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4604b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 46054c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 46064c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4607983c7515STejun Heo kfree(rescuer); 4608b92b36eaSDan Carpenter return ret; 4609983c7515STejun Heo } 4610983c7515STejun Heo 4611983c7515STejun Heo wq->rescuer = rescuer; 4612983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4613983c7515STejun Heo wake_up_process(rescuer->task); 4614983c7515STejun Heo 4615983c7515STejun Heo return 0; 4616983c7515STejun Heo } 4617983c7515STejun Heo 4618a2775bbcSMathieu Malaterre __printf(1, 4) 4619669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 462097e37d7bSTejun Heo unsigned int flags, 4621669de8bdSBart Van Assche int max_active, ...) 46223af24433SOleg Nesterov { 4623df2d5ae4STejun Heo size_t tbl_size = 0; 4624ecf6881fSTejun Heo va_list args; 46253af24433SOleg Nesterov struct workqueue_struct *wq; 462649e3cf44STejun Heo struct pool_workqueue *pwq; 4627b196be89STejun Heo 46285c0338c6STejun Heo /* 46295c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 46305c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 46315c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 46325c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 46335c0338c6STejun Heo * on NUMA. 46345c0338c6STejun Heo */ 46355c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 46365c0338c6STejun Heo flags |= __WQ_ORDERED; 46375c0338c6STejun Heo 4638cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4639cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4640cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4641cee22a15SViresh Kumar 4642ecf6881fSTejun Heo /* allocate wq and format name */ 4643df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4644ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4645df2d5ae4STejun Heo 4646df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4647b196be89STejun Heo if (!wq) 4648d2c1d404STejun Heo return NULL; 4649b196be89STejun Heo 46506029a918STejun Heo if (flags & WQ_UNBOUND) { 4651be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 46526029a918STejun Heo if (!wq->unbound_attrs) 46536029a918STejun Heo goto err_free_wq; 46546029a918STejun Heo } 46556029a918STejun Heo 4656669de8bdSBart Van Assche va_start(args, max_active); 4657ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4658b196be89STejun Heo va_end(args); 46593af24433SOleg Nesterov 4660d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4661b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 46623af24433SOleg Nesterov 4663b196be89STejun Heo /* init wq */ 466497e37d7bSTejun Heo wq->flags = flags; 4665a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 46663c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4667112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 466830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 466973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 467073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4671493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 46723af24433SOleg Nesterov 4673669de8bdSBart Van Assche wq_init_lockdep(wq); 4674cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 46753af24433SOleg Nesterov 467630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 467782efcab3SBart Van Assche goto err_unreg_lockdep; 46781537663fSTejun Heo 467940c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4680d2c1d404STejun Heo goto err_destroy; 4681e22bee78STejun Heo 4682226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4683226223abSTejun Heo goto err_destroy; 4684226223abSTejun Heo 46856af8bf3dSOleg Nesterov /* 468668e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 468768e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 468868e13a67SLai Jiangshan * list. 46896af8bf3dSOleg Nesterov */ 469068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4691a0a1a5fdSTejun Heo 4692a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 469349e3cf44STejun Heo for_each_pwq(pwq, wq) 4694699ce097STejun Heo pwq_adjust_max_active(pwq); 4695a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4696a0a1a5fdSTejun Heo 4697e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4698a0a1a5fdSTejun Heo 469968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47003af24433SOleg Nesterov 47013af24433SOleg Nesterov return wq; 4702d2c1d404STejun Heo 470382efcab3SBart Van Assche err_unreg_lockdep: 4704009bb421SBart Van Assche wq_unregister_lockdep(wq); 4705009bb421SBart Van Assche wq_free_lockdep(wq); 470682efcab3SBart Van Assche err_free_wq: 47076029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 47084690c4abSTejun Heo kfree(wq); 4709d2c1d404STejun Heo return NULL; 4710d2c1d404STejun Heo err_destroy: 4711d2c1d404STejun Heo destroy_workqueue(wq); 47124690c4abSTejun Heo return NULL; 47131da177e4SLinus Torvalds } 4714669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 47151da177e4SLinus Torvalds 4716c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4717c29eb853STejun Heo { 4718c29eb853STejun Heo int i; 4719c29eb853STejun Heo 4720c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4721c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4722c29eb853STejun Heo return true; 4723c29eb853STejun Heo 4724c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4725c29eb853STejun Heo return true; 4726f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4727c29eb853STejun Heo return true; 4728c29eb853STejun Heo 4729c29eb853STejun Heo return false; 4730c29eb853STejun Heo } 4731c29eb853STejun Heo 47323af24433SOleg Nesterov /** 47333af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 47343af24433SOleg Nesterov * @wq: target workqueue 47353af24433SOleg Nesterov * 47363af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 47373af24433SOleg Nesterov */ 47383af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 47393af24433SOleg Nesterov { 474049e3cf44STejun Heo struct pool_workqueue *pwq; 47414c16bd32STejun Heo int node; 47423af24433SOleg Nesterov 4743def98c84STejun Heo /* 4744def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4745def98c84STejun Heo * lead to sysfs name conflicts. 4746def98c84STejun Heo */ 4747def98c84STejun Heo workqueue_sysfs_unregister(wq); 4748def98c84STejun Heo 474933e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 475033e3f0a3SRichard Clark mutex_lock(&wq->mutex); 475133e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 475233e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 475333e3f0a3SRichard Clark 47549c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 47559c5a2ba7STejun Heo drain_workqueue(wq); 4756c8efcc25STejun Heo 4757def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4758def98c84STejun Heo if (wq->rescuer) { 4759def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4760def98c84STejun Heo 4761def98c84STejun Heo /* this prevents new queueing */ 4762a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4763def98c84STejun Heo wq->rescuer = NULL; 4764a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4765def98c84STejun Heo 4766def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4767def98c84STejun Heo kthread_stop(rescuer->task); 47688efe1223STejun Heo kfree(rescuer); 4769def98c84STejun Heo } 4770def98c84STejun Heo 4771c29eb853STejun Heo /* 4772c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4773c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4774c29eb853STejun Heo */ 4775c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4776b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 477749e3cf44STejun Heo for_each_pwq(pwq, wq) { 4778a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4779c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 47801d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4781e66b39afSTejun Heo __func__, wq->name); 4782c29eb853STejun Heo show_pwq(pwq); 4783a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4784b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4785c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 478655df0933SImran Khan show_one_workqueue(wq); 47876183c009STejun Heo return; 478876af4d93STejun Heo } 4789a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 479076af4d93STejun Heo } 4791b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 47926183c009STejun Heo 4793a0a1a5fdSTejun Heo /* 4794a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4795a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4796a0a1a5fdSTejun Heo */ 4797e2dca7adSTejun Heo list_del_rcu(&wq->list); 479868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47993af24433SOleg Nesterov 48008864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4801669de8bdSBart Van Assche wq_unregister_lockdep(wq); 480229c91e99STejun Heo /* 48038864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4804e2dca7adSTejun Heo * schedule RCU free. 480529c91e99STejun Heo */ 480625b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 48078864b4e5STejun Heo } else { 48088864b4e5STejun Heo /* 48098864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 48104c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 48114c16bd32STejun Heo * @wq will be freed when the last pwq is released. 48128864b4e5STejun Heo */ 48134c16bd32STejun Heo for_each_node(node) { 48144c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 48154c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 48164c16bd32STejun Heo put_pwq_unlocked(pwq); 48174c16bd32STejun Heo } 48184c16bd32STejun Heo 48194c16bd32STejun Heo /* 48204c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 48214c16bd32STejun Heo * put. Don't access it afterwards. 48224c16bd32STejun Heo */ 48234c16bd32STejun Heo pwq = wq->dfl_pwq; 48244c16bd32STejun Heo wq->dfl_pwq = NULL; 4825dce90d47STejun Heo put_pwq_unlocked(pwq); 482629c91e99STejun Heo } 48273af24433SOleg Nesterov } 48283af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 48293af24433SOleg Nesterov 4830dcd989cbSTejun Heo /** 4831dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4832dcd989cbSTejun Heo * @wq: target workqueue 4833dcd989cbSTejun Heo * @max_active: new max_active value. 4834dcd989cbSTejun Heo * 4835dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4836dcd989cbSTejun Heo * 4837dcd989cbSTejun Heo * CONTEXT: 4838dcd989cbSTejun Heo * Don't call from IRQ context. 4839dcd989cbSTejun Heo */ 4840dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4841dcd989cbSTejun Heo { 484249e3cf44STejun Heo struct pool_workqueue *pwq; 4843dcd989cbSTejun Heo 48448719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 48450a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 48468719dceaSTejun Heo return; 48478719dceaSTejun Heo 4848f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4849dcd989cbSTejun Heo 4850a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4851dcd989cbSTejun Heo 48520a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4853dcd989cbSTejun Heo wq->saved_max_active = max_active; 4854dcd989cbSTejun Heo 4855699ce097STejun Heo for_each_pwq(pwq, wq) 4856699ce097STejun Heo pwq_adjust_max_active(pwq); 4857dcd989cbSTejun Heo 4858a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4859dcd989cbSTejun Heo } 4860dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4861dcd989cbSTejun Heo 4862dcd989cbSTejun Heo /** 486327d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 486427d4ee03SLukas Wunner * 486527d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 486627d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 486727d4ee03SLukas Wunner * 486827d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 486927d4ee03SLukas Wunner */ 487027d4ee03SLukas Wunner struct work_struct *current_work(void) 487127d4ee03SLukas Wunner { 487227d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 487327d4ee03SLukas Wunner 487427d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 487527d4ee03SLukas Wunner } 487627d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 487727d4ee03SLukas Wunner 487827d4ee03SLukas Wunner /** 4879e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4880e6267616STejun Heo * 4881e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4882e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4883d185af30SYacine Belkadi * 4884d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4885e6267616STejun Heo */ 4886e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4887e6267616STejun Heo { 4888e6267616STejun Heo struct worker *worker = current_wq_worker(); 4889e6267616STejun Heo 48906a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4891e6267616STejun Heo } 4892e6267616STejun Heo 4893e6267616STejun Heo /** 4894dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4895dcd989cbSTejun Heo * @cpu: CPU in question 4896dcd989cbSTejun Heo * @wq: target workqueue 4897dcd989cbSTejun Heo * 4898dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4899dcd989cbSTejun Heo * no synchronization around this function and the test result is 4900dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4901dcd989cbSTejun Heo * 4902d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4903d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4904d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4905d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4906d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4907d3251859STejun Heo * 4908d185af30SYacine Belkadi * Return: 4909dcd989cbSTejun Heo * %true if congested, %false otherwise. 4910dcd989cbSTejun Heo */ 4911d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4912dcd989cbSTejun Heo { 49137fb98ea7STejun Heo struct pool_workqueue *pwq; 491476af4d93STejun Heo bool ret; 491576af4d93STejun Heo 491624acfb71SThomas Gleixner rcu_read_lock(); 491724acfb71SThomas Gleixner preempt_disable(); 49187fb98ea7STejun Heo 4919d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4920d3251859STejun Heo cpu = smp_processor_id(); 4921d3251859STejun Heo 49227fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 49237fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 49247fb98ea7STejun Heo else 4925df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4926dcd989cbSTejun Heo 4927f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 492824acfb71SThomas Gleixner preempt_enable(); 492924acfb71SThomas Gleixner rcu_read_unlock(); 493076af4d93STejun Heo 493176af4d93STejun Heo return ret; 4932dcd989cbSTejun Heo } 4933dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4934dcd989cbSTejun Heo 4935dcd989cbSTejun Heo /** 4936dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4937dcd989cbSTejun Heo * @work: the work to be tested 4938dcd989cbSTejun Heo * 4939dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4940dcd989cbSTejun Heo * synchronization around this function and the test result is 4941dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4942dcd989cbSTejun Heo * 4943d185af30SYacine Belkadi * Return: 4944dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4945dcd989cbSTejun Heo */ 4946dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4947dcd989cbSTejun Heo { 4948fa1b54e6STejun Heo struct worker_pool *pool; 4949dcd989cbSTejun Heo unsigned long flags; 4950dcd989cbSTejun Heo unsigned int ret = 0; 4951dcd989cbSTejun Heo 4952dcd989cbSTejun Heo if (work_pending(work)) 4953dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4954038366c5SLai Jiangshan 495524acfb71SThomas Gleixner rcu_read_lock(); 4956fa1b54e6STejun Heo pool = get_work_pool(work); 4957038366c5SLai Jiangshan if (pool) { 4958a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4959c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4960dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4961a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4962038366c5SLai Jiangshan } 496324acfb71SThomas Gleixner rcu_read_unlock(); 4964dcd989cbSTejun Heo 4965dcd989cbSTejun Heo return ret; 4966dcd989cbSTejun Heo } 4967dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4968dcd989cbSTejun Heo 49693d1cb205STejun Heo /** 49703d1cb205STejun Heo * set_worker_desc - set description for the current work item 49713d1cb205STejun Heo * @fmt: printf-style format string 49723d1cb205STejun Heo * @...: arguments for the format string 49733d1cb205STejun Heo * 49743d1cb205STejun Heo * This function can be called by a running work function to describe what 49753d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 49763d1cb205STejun Heo * information will be printed out together to help debugging. The 49773d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 49783d1cb205STejun Heo */ 49793d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 49803d1cb205STejun Heo { 49813d1cb205STejun Heo struct worker *worker = current_wq_worker(); 49823d1cb205STejun Heo va_list args; 49833d1cb205STejun Heo 49843d1cb205STejun Heo if (worker) { 49853d1cb205STejun Heo va_start(args, fmt); 49863d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 49873d1cb205STejun Heo va_end(args); 49883d1cb205STejun Heo } 49893d1cb205STejun Heo } 49905c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 49913d1cb205STejun Heo 49923d1cb205STejun Heo /** 49933d1cb205STejun Heo * print_worker_info - print out worker information and description 49943d1cb205STejun Heo * @log_lvl: the log level to use when printing 49953d1cb205STejun Heo * @task: target task 49963d1cb205STejun Heo * 49973d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 49983d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 49993d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 50003d1cb205STejun Heo * 50013d1cb205STejun Heo * This function can be safely called on any task as long as the 50023d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 50033d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 50043d1cb205STejun Heo */ 50053d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 50063d1cb205STejun Heo { 50073d1cb205STejun Heo work_func_t *fn = NULL; 50083d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 50093d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 50103d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 50113d1cb205STejun Heo struct workqueue_struct *wq = NULL; 50123d1cb205STejun Heo struct worker *worker; 50133d1cb205STejun Heo 50143d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 50153d1cb205STejun Heo return; 50163d1cb205STejun Heo 50173d1cb205STejun Heo /* 50183d1cb205STejun Heo * This function is called without any synchronization and @task 50193d1cb205STejun Heo * could be in any state. Be careful with dereferences. 50203d1cb205STejun Heo */ 5021e700591aSPetr Mladek worker = kthread_probe_data(task); 50223d1cb205STejun Heo 50233d1cb205STejun Heo /* 50248bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 50258bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 50263d1cb205STejun Heo */ 5027fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5028fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5029fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5030fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5031fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 50323d1cb205STejun Heo 50333d1cb205STejun Heo if (fn || name[0] || desc[0]) { 5034d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 50358bf89593STejun Heo if (strcmp(name, desc)) 50363d1cb205STejun Heo pr_cont(" (%s)", desc); 50373d1cb205STejun Heo pr_cont("\n"); 50383d1cb205STejun Heo } 50393d1cb205STejun Heo } 50403d1cb205STejun Heo 50413494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 50423494fc30STejun Heo { 50433494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 50443494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 50453494fc30STejun Heo pr_cont(" node=%d", pool->node); 50463494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 50473494fc30STejun Heo } 50483494fc30STejun Heo 5049c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 5050c76feb0dSPaul E. McKenney bool comma; 5051c76feb0dSPaul E. McKenney work_func_t func; 5052c76feb0dSPaul E. McKenney long ctr; 5053c76feb0dSPaul E. McKenney }; 5054c76feb0dSPaul E. McKenney 5055c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5056c76feb0dSPaul E. McKenney { 5057c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 5058c76feb0dSPaul E. McKenney goto out_record; 5059c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 5060c76feb0dSPaul E. McKenney pcwsp->ctr++; 5061c76feb0dSPaul E. McKenney return; 5062c76feb0dSPaul E. McKenney } 5063c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 5064c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5065c76feb0dSPaul E. McKenney else 5066c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5067c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 5068c76feb0dSPaul E. McKenney out_record: 5069c76feb0dSPaul E. McKenney if ((long)func == -1L) 5070c76feb0dSPaul E. McKenney return; 5071c76feb0dSPaul E. McKenney pcwsp->comma = comma; 5072c76feb0dSPaul E. McKenney pcwsp->func = func; 5073c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 5074c76feb0dSPaul E. McKenney } 5075c76feb0dSPaul E. McKenney 5076c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 50773494fc30STejun Heo { 50783494fc30STejun Heo if (work->func == wq_barrier_func) { 50793494fc30STejun Heo struct wq_barrier *barr; 50803494fc30STejun Heo 50813494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 50823494fc30STejun Heo 5083c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 50843494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 50853494fc30STejun Heo task_pid_nr(barr->task)); 50863494fc30STejun Heo } else { 5087c76feb0dSPaul E. McKenney if (!comma) 5088c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5089c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 50903494fc30STejun Heo } 50913494fc30STejun Heo } 50923494fc30STejun Heo 50933494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 50943494fc30STejun Heo { 5095c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 50963494fc30STejun Heo struct worker_pool *pool = pwq->pool; 50973494fc30STejun Heo struct work_struct *work; 50983494fc30STejun Heo struct worker *worker; 50993494fc30STejun Heo bool has_in_flight = false, has_pending = false; 51003494fc30STejun Heo int bkt; 51013494fc30STejun Heo 51023494fc30STejun Heo pr_info(" pwq %d:", pool->id); 51033494fc30STejun Heo pr_cont_pool_info(pool); 51043494fc30STejun Heo 5105e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5106e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 51073494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 51083494fc30STejun Heo 51093494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51103494fc30STejun Heo if (worker->current_pwq == pwq) { 51113494fc30STejun Heo has_in_flight = true; 51123494fc30STejun Heo break; 51133494fc30STejun Heo } 51143494fc30STejun Heo } 51153494fc30STejun Heo if (has_in_flight) { 51163494fc30STejun Heo bool comma = false; 51173494fc30STejun Heo 51183494fc30STejun Heo pr_info(" in-flight:"); 51193494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51203494fc30STejun Heo if (worker->current_pwq != pwq) 51213494fc30STejun Heo continue; 51223494fc30STejun Heo 5123d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 51243494fc30STejun Heo task_pid_nr(worker->task), 512530ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 51263494fc30STejun Heo worker->current_func); 51273494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5128c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5129c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51303494fc30STejun Heo comma = true; 51313494fc30STejun Heo } 51323494fc30STejun Heo pr_cont("\n"); 51333494fc30STejun Heo } 51343494fc30STejun Heo 51353494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51363494fc30STejun Heo if (get_work_pwq(work) == pwq) { 51373494fc30STejun Heo has_pending = true; 51383494fc30STejun Heo break; 51393494fc30STejun Heo } 51403494fc30STejun Heo } 51413494fc30STejun Heo if (has_pending) { 51423494fc30STejun Heo bool comma = false; 51433494fc30STejun Heo 51443494fc30STejun Heo pr_info(" pending:"); 51453494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51463494fc30STejun Heo if (get_work_pwq(work) != pwq) 51473494fc30STejun Heo continue; 51483494fc30STejun Heo 5149c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51503494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51513494fc30STejun Heo } 5152c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51533494fc30STejun Heo pr_cont("\n"); 51543494fc30STejun Heo } 51553494fc30STejun Heo 5156f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 51573494fc30STejun Heo bool comma = false; 51583494fc30STejun Heo 5159f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5160f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5161c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51623494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51633494fc30STejun Heo } 5164c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51653494fc30STejun Heo pr_cont("\n"); 51663494fc30STejun Heo } 51673494fc30STejun Heo } 51683494fc30STejun Heo 51693494fc30STejun Heo /** 517055df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 517155df0933SImran Khan * @wq: workqueue whose state will be printed 51723494fc30STejun Heo */ 517355df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 51743494fc30STejun Heo { 51753494fc30STejun Heo struct pool_workqueue *pwq; 51763494fc30STejun Heo bool idle = true; 517755df0933SImran Khan unsigned long flags; 51783494fc30STejun Heo 51793494fc30STejun Heo for_each_pwq(pwq, wq) { 5180f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 51813494fc30STejun Heo idle = false; 51823494fc30STejun Heo break; 51833494fc30STejun Heo } 51843494fc30STejun Heo } 518555df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 518655df0933SImran Khan return; 51873494fc30STejun Heo 51883494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 51893494fc30STejun Heo 51903494fc30STejun Heo for_each_pwq(pwq, wq) { 5191a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 519257116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 519357116ce1SJohan Hovold /* 519457116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 519557116ce1SJohan Hovold * drivers that queue work while holding locks 519657116ce1SJohan Hovold * also taken in their write paths. 519757116ce1SJohan Hovold */ 519857116ce1SJohan Hovold printk_deferred_enter(); 51993494fc30STejun Heo show_pwq(pwq); 520057116ce1SJohan Hovold printk_deferred_exit(); 520157116ce1SJohan Hovold } 5202a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 520362635ea8SSergey Senozhatsky /* 520462635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 520555df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 520662635ea8SSergey Senozhatsky * hard lockup. 520762635ea8SSergey Senozhatsky */ 520862635ea8SSergey Senozhatsky touch_nmi_watchdog(); 52093494fc30STejun Heo } 521055df0933SImran Khan 52113494fc30STejun Heo } 52123494fc30STejun Heo 521355df0933SImran Khan /** 521455df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 521555df0933SImran Khan * @pool: worker pool whose state will be printed 521655df0933SImran Khan */ 521755df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 521855df0933SImran Khan { 52193494fc30STejun Heo struct worker *worker; 52203494fc30STejun Heo bool first = true; 522155df0933SImran Khan unsigned long flags; 5222335a42ebSPetr Mladek unsigned long hung = 0; 52233494fc30STejun Heo 5224a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 52253494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 52263494fc30STejun Heo goto next_pool; 5227335a42ebSPetr Mladek 5228335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5229335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5230335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5231335a42ebSPetr Mladek 523257116ce1SJohan Hovold /* 523357116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 523457116ce1SJohan Hovold * queue work while holding locks also taken in their write 523557116ce1SJohan Hovold * paths. 523657116ce1SJohan Hovold */ 523757116ce1SJohan Hovold printk_deferred_enter(); 52383494fc30STejun Heo pr_info("pool %d:", pool->id); 52393494fc30STejun Heo pr_cont_pool_info(pool); 5240335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 52413494fc30STejun Heo if (pool->manager) 52423494fc30STejun Heo pr_cont(" manager: %d", 52433494fc30STejun Heo task_pid_nr(pool->manager->task)); 52443494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 52453494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 52463494fc30STejun Heo task_pid_nr(worker->task)); 52473494fc30STejun Heo first = false; 52483494fc30STejun Heo } 52493494fc30STejun Heo pr_cont("\n"); 525057116ce1SJohan Hovold printk_deferred_exit(); 52513494fc30STejun Heo next_pool: 5252a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 525362635ea8SSergey Senozhatsky /* 525462635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 525555df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 525662635ea8SSergey Senozhatsky * hard lockup. 525762635ea8SSergey Senozhatsky */ 525862635ea8SSergey Senozhatsky touch_nmi_watchdog(); 525955df0933SImran Khan 52603494fc30STejun Heo } 52613494fc30STejun Heo 526255df0933SImran Khan /** 526355df0933SImran Khan * show_all_workqueues - dump workqueue state 526455df0933SImran Khan * 5265704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 526655df0933SImran Khan */ 526755df0933SImran Khan void show_all_workqueues(void) 526855df0933SImran Khan { 526955df0933SImran Khan struct workqueue_struct *wq; 527055df0933SImran Khan struct worker_pool *pool; 527155df0933SImran Khan int pi; 527255df0933SImran Khan 527355df0933SImran Khan rcu_read_lock(); 527455df0933SImran Khan 527555df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 527655df0933SImran Khan 527755df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 527855df0933SImran Khan show_one_workqueue(wq); 527955df0933SImran Khan 528055df0933SImran Khan for_each_pool(pool, pi) 528155df0933SImran Khan show_one_worker_pool(pool); 528255df0933SImran Khan 528324acfb71SThomas Gleixner rcu_read_unlock(); 52843494fc30STejun Heo } 52853494fc30STejun Heo 5286704bc669SJungseung Lee /** 5287704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5288704bc669SJungseung Lee * 5289704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5290704bc669SJungseung Lee * still busy. 5291704bc669SJungseung Lee */ 5292704bc669SJungseung Lee void show_freezable_workqueues(void) 5293704bc669SJungseung Lee { 5294704bc669SJungseung Lee struct workqueue_struct *wq; 5295704bc669SJungseung Lee 5296704bc669SJungseung Lee rcu_read_lock(); 5297704bc669SJungseung Lee 5298704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5299704bc669SJungseung Lee 5300704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5301704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5302704bc669SJungseung Lee continue; 5303704bc669SJungseung Lee show_one_workqueue(wq); 5304704bc669SJungseung Lee } 5305704bc669SJungseung Lee 5306704bc669SJungseung Lee rcu_read_unlock(); 5307704bc669SJungseung Lee } 5308704bc669SJungseung Lee 53096b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 53106b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 53116b59808bSTejun Heo { 53126b59808bSTejun Heo int off; 53136b59808bSTejun Heo 53146b59808bSTejun Heo /* always show the actual comm */ 53156b59808bSTejun Heo off = strscpy(buf, task->comm, size); 53166b59808bSTejun Heo if (off < 0) 53176b59808bSTejun Heo return; 53186b59808bSTejun Heo 5319197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 53206b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 53216b59808bSTejun Heo 5322197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5323197f6accSTejun Heo struct worker *worker = kthread_data(task); 5324197f6accSTejun Heo struct worker_pool *pool = worker->pool; 53256b59808bSTejun Heo 53266b59808bSTejun Heo if (pool) { 5327a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 53286b59808bSTejun Heo /* 5329197f6accSTejun Heo * ->desc tracks information (wq name or 5330197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5331197f6accSTejun Heo * current, prepend '+', otherwise '-'. 53326b59808bSTejun Heo */ 53336b59808bSTejun Heo if (worker->desc[0] != '\0') { 53346b59808bSTejun Heo if (worker->current_work) 53356b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 53366b59808bSTejun Heo worker->desc); 53376b59808bSTejun Heo else 53386b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 53396b59808bSTejun Heo worker->desc); 53406b59808bSTejun Heo } 5341a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 53426b59808bSTejun Heo } 5343197f6accSTejun Heo } 53446b59808bSTejun Heo 53456b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 53466b59808bSTejun Heo } 53476b59808bSTejun Heo 534866448bc2SMathieu Malaterre #ifdef CONFIG_SMP 534966448bc2SMathieu Malaterre 5350db7bccf4STejun Heo /* 5351db7bccf4STejun Heo * CPU hotplug. 5352db7bccf4STejun Heo * 5353e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5354112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5355706026c2STejun Heo * pool which make migrating pending and scheduled works very 5356e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 535794cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5358e22bee78STejun Heo * blocked draining impractical. 5359e22bee78STejun Heo * 536024647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5361628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5362628c78e7STejun Heo * cpu comes back online. 5363db7bccf4STejun Heo */ 5364db7bccf4STejun Heo 5365e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5366db7bccf4STejun Heo { 53674ce62e9eSTejun Heo struct worker_pool *pool; 5368db7bccf4STejun Heo struct worker *worker; 5369db7bccf4STejun Heo 5370f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53711258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5372a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5373e22bee78STejun Heo 5374f2d5a0eeSTejun Heo /* 537592f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 537694cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 537711b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5378b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5379b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5380b4ac9384SLai Jiangshan * is on the same cpu. 5381f2d5a0eeSTejun Heo */ 5382da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5383403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5384db7bccf4STejun Heo 538524647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5386f2d5a0eeSTejun Heo 5387e22bee78STejun Heo /* 5388989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5389989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5390989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5391989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5392989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5393eb283428SLai Jiangshan * are served by workers tied to the pool. 5394e22bee78STejun Heo */ 5395bc35f7efSLai Jiangshan pool->nr_running = 0; 5396eb283428SLai Jiangshan 5397eb283428SLai Jiangshan /* 5398eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5399eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5400eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5401eb283428SLai Jiangshan */ 5402eb283428SLai Jiangshan wake_up_worker(pool); 5403989442d7SLai Jiangshan 5404a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5405989442d7SLai Jiangshan 5406793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5407793777bcSValentin Schneider unbind_worker(worker); 5408989442d7SLai Jiangshan 5409989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5410eb283428SLai Jiangshan } 5411db7bccf4STejun Heo } 5412db7bccf4STejun Heo 5413bd7c089eSTejun Heo /** 5414bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5415bd7c089eSTejun Heo * @pool: pool of interest 5416bd7c089eSTejun Heo * 5417a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5418bd7c089eSTejun Heo */ 5419bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5420bd7c089eSTejun Heo { 5421a9ab775bSTejun Heo struct worker *worker; 5422bd7c089eSTejun Heo 54231258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5424bd7c089eSTejun Heo 5425bd7c089eSTejun Heo /* 5426a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5427a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5428402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5429a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5430a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5431bd7c089eSTejun Heo */ 5432c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5433c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5434c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5435c63a2e52SValentin Schneider pool->attrs->cpumask) < 0); 5436c63a2e52SValentin Schneider } 5437a9ab775bSTejun Heo 5438a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5439f7c17d26SWanpeng Li 54403de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5441a9ab775bSTejun Heo 5442da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5443a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5444a9ab775bSTejun Heo 5445a9ab775bSTejun Heo /* 5446a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5447a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5448a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5449a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5450a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5451a9ab775bSTejun Heo * concurrency management. Note that when or whether 5452a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5453a9ab775bSTejun Heo * 5454c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5455a9ab775bSTejun Heo * tested without holding any lock in 54566d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5457a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5458a9ab775bSTejun Heo * management operations. 5459bd7c089eSTejun Heo */ 5460a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5461a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5462a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5463c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5464bd7c089eSTejun Heo } 5465a9ab775bSTejun Heo 5466a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5467bd7c089eSTejun Heo } 5468bd7c089eSTejun Heo 54697dbc725eSTejun Heo /** 54707dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 54717dbc725eSTejun Heo * @pool: unbound pool of interest 54727dbc725eSTejun Heo * @cpu: the CPU which is coming up 54737dbc725eSTejun Heo * 54747dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 54757dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 54767dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 54777dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 54787dbc725eSTejun Heo */ 54797dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 54807dbc725eSTejun Heo { 54817dbc725eSTejun Heo static cpumask_t cpumask; 54827dbc725eSTejun Heo struct worker *worker; 54837dbc725eSTejun Heo 54841258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 54857dbc725eSTejun Heo 54867dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 54877dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 54887dbc725eSTejun Heo return; 54897dbc725eSTejun Heo 54907dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 54917dbc725eSTejun Heo 54927dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5493da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5494d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 54957dbc725eSTejun Heo } 54967dbc725eSTejun Heo 54977ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 54981da177e4SLinus Torvalds { 54994ce62e9eSTejun Heo struct worker_pool *pool; 55001da177e4SLinus Torvalds 5501f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 55023ce63377STejun Heo if (pool->nr_workers) 55033ce63377STejun Heo continue; 5504051e1850SLai Jiangshan if (!create_worker(pool)) 55057ee681b2SThomas Gleixner return -ENOMEM; 55063af24433SOleg Nesterov } 55077ee681b2SThomas Gleixner return 0; 55087ee681b2SThomas Gleixner } 55091da177e4SLinus Torvalds 55107ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 55117ee681b2SThomas Gleixner { 55127ee681b2SThomas Gleixner struct worker_pool *pool; 55137ee681b2SThomas Gleixner struct workqueue_struct *wq; 55147ee681b2SThomas Gleixner int pi; 55157ee681b2SThomas Gleixner 551668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55177dbc725eSTejun Heo 55187dbc725eSTejun Heo for_each_pool(pool, pi) { 55191258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 552094cf58bbSTejun Heo 5521f05b558dSLai Jiangshan if (pool->cpu == cpu) 552294cf58bbSTejun Heo rebind_workers(pool); 5523f05b558dSLai Jiangshan else if (pool->cpu < 0) 55247dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 552594cf58bbSTejun Heo 55261258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 552794cf58bbSTejun Heo } 55287dbc725eSTejun Heo 55294c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55304c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 55314c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 55324c16bd32STejun Heo 553368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55347ee681b2SThomas Gleixner return 0; 553565758202STejun Heo } 553665758202STejun Heo 55377ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 553865758202STejun Heo { 55394c16bd32STejun Heo struct workqueue_struct *wq; 55408db25e78STejun Heo 55414c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5542e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5543e8b3f8dbSLai Jiangshan return -1; 5544e8b3f8dbSLai Jiangshan 5545e8b3f8dbSLai Jiangshan unbind_workers(cpu); 55464c16bd32STejun Heo 55474c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55484c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 55494c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 55504c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 55514c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 55524c16bd32STejun Heo 55537ee681b2SThomas Gleixner return 0; 555465758202STejun Heo } 555565758202STejun Heo 55562d3854a3SRusty Russell struct work_for_cpu { 5557ed48ece2STejun Heo struct work_struct work; 55582d3854a3SRusty Russell long (*fn)(void *); 55592d3854a3SRusty Russell void *arg; 55602d3854a3SRusty Russell long ret; 55612d3854a3SRusty Russell }; 55622d3854a3SRusty Russell 5563ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 55642d3854a3SRusty Russell { 5565ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5566ed48ece2STejun Heo 55672d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 55682d3854a3SRusty Russell } 55692d3854a3SRusty Russell 55702d3854a3SRusty Russell /** 557122aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 55722d3854a3SRusty Russell * @cpu: the cpu to run on 55732d3854a3SRusty Russell * @fn: the function to run 55742d3854a3SRusty Russell * @arg: the function arg 55752d3854a3SRusty Russell * 557631ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 55776b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5578d185af30SYacine Belkadi * 5579d185af30SYacine Belkadi * Return: The value @fn returns. 55802d3854a3SRusty Russell */ 5581d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 55822d3854a3SRusty Russell { 5583ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 55842d3854a3SRusty Russell 5585ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5586ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 558712997d1aSBjorn Helgaas flush_work(&wfc.work); 5588440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 55892d3854a3SRusty Russell return wfc.ret; 55902d3854a3SRusty Russell } 55912d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 55920e8d6a93SThomas Gleixner 55930e8d6a93SThomas Gleixner /** 55940e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 55950e8d6a93SThomas Gleixner * @cpu: the cpu to run on 55960e8d6a93SThomas Gleixner * @fn: the function to run 55970e8d6a93SThomas Gleixner * @arg: the function argument 55980e8d6a93SThomas Gleixner * 55990e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 56000e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 56010e8d6a93SThomas Gleixner * 56020e8d6a93SThomas Gleixner * Return: The value @fn returns. 56030e8d6a93SThomas Gleixner */ 56040e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 56050e8d6a93SThomas Gleixner { 56060e8d6a93SThomas Gleixner long ret = -ENODEV; 56070e8d6a93SThomas Gleixner 5608ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 56090e8d6a93SThomas Gleixner if (cpu_online(cpu)) 56100e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5611ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 56120e8d6a93SThomas Gleixner return ret; 56130e8d6a93SThomas Gleixner } 56140e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 56152d3854a3SRusty Russell #endif /* CONFIG_SMP */ 56162d3854a3SRusty Russell 5617a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5618e7577c50SRusty Russell 5619a0a1a5fdSTejun Heo /** 5620a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5621a0a1a5fdSTejun Heo * 562258a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5623f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5624706026c2STejun Heo * pool->worklist. 5625a0a1a5fdSTejun Heo * 5626a0a1a5fdSTejun Heo * CONTEXT: 5627a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5628a0a1a5fdSTejun Heo */ 5629a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5630a0a1a5fdSTejun Heo { 563124b8a847STejun Heo struct workqueue_struct *wq; 563224b8a847STejun Heo struct pool_workqueue *pwq; 5633a0a1a5fdSTejun Heo 563468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5635a0a1a5fdSTejun Heo 56366183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5637a0a1a5fdSTejun Heo workqueue_freezing = true; 5638a0a1a5fdSTejun Heo 563924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5640a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5641699ce097STejun Heo for_each_pwq(pwq, wq) 5642699ce097STejun Heo pwq_adjust_max_active(pwq); 5643a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5644a1056305STejun Heo } 56455bcab335STejun Heo 564668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5647a0a1a5fdSTejun Heo } 5648a0a1a5fdSTejun Heo 5649a0a1a5fdSTejun Heo /** 565058a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5651a0a1a5fdSTejun Heo * 5652a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5653a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5654a0a1a5fdSTejun Heo * 5655a0a1a5fdSTejun Heo * CONTEXT: 565668e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5657a0a1a5fdSTejun Heo * 5658d185af30SYacine Belkadi * Return: 565958a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 566058a69cb4STejun Heo * is complete. 5661a0a1a5fdSTejun Heo */ 5662a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5663a0a1a5fdSTejun Heo { 5664a0a1a5fdSTejun Heo bool busy = false; 566524b8a847STejun Heo struct workqueue_struct *wq; 566624b8a847STejun Heo struct pool_workqueue *pwq; 5667a0a1a5fdSTejun Heo 566868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5669a0a1a5fdSTejun Heo 56706183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5671a0a1a5fdSTejun Heo 567224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 567324b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 567424b8a847STejun Heo continue; 5675a0a1a5fdSTejun Heo /* 5676a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5677a0a1a5fdSTejun Heo * to peek without lock. 5678a0a1a5fdSTejun Heo */ 567924acfb71SThomas Gleixner rcu_read_lock(); 568024b8a847STejun Heo for_each_pwq(pwq, wq) { 56816183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5682112202d9STejun Heo if (pwq->nr_active) { 5683a0a1a5fdSTejun Heo busy = true; 568424acfb71SThomas Gleixner rcu_read_unlock(); 5685a0a1a5fdSTejun Heo goto out_unlock; 5686a0a1a5fdSTejun Heo } 5687a0a1a5fdSTejun Heo } 568824acfb71SThomas Gleixner rcu_read_unlock(); 5689a0a1a5fdSTejun Heo } 5690a0a1a5fdSTejun Heo out_unlock: 569168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5692a0a1a5fdSTejun Heo return busy; 5693a0a1a5fdSTejun Heo } 5694a0a1a5fdSTejun Heo 5695a0a1a5fdSTejun Heo /** 5696a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5697a0a1a5fdSTejun Heo * 5698a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5699706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5700a0a1a5fdSTejun Heo * 5701a0a1a5fdSTejun Heo * CONTEXT: 5702a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5703a0a1a5fdSTejun Heo */ 5704a0a1a5fdSTejun Heo void thaw_workqueues(void) 5705a0a1a5fdSTejun Heo { 570624b8a847STejun Heo struct workqueue_struct *wq; 570724b8a847STejun Heo struct pool_workqueue *pwq; 5708a0a1a5fdSTejun Heo 570968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5710a0a1a5fdSTejun Heo 5711a0a1a5fdSTejun Heo if (!workqueue_freezing) 5712a0a1a5fdSTejun Heo goto out_unlock; 5713a0a1a5fdSTejun Heo 571474b414eaSLai Jiangshan workqueue_freezing = false; 571524b8a847STejun Heo 571624b8a847STejun Heo /* restore max_active and repopulate worklist */ 571724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5718a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5719699ce097STejun Heo for_each_pwq(pwq, wq) 5720699ce097STejun Heo pwq_adjust_max_active(pwq); 5721a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 572224b8a847STejun Heo } 572324b8a847STejun Heo 5724a0a1a5fdSTejun Heo out_unlock: 572568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5726a0a1a5fdSTejun Heo } 5727a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5728a0a1a5fdSTejun Heo 572999c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5730042f7df1SLai Jiangshan { 5731042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5732042f7df1SLai Jiangshan int ret = 0; 5733042f7df1SLai Jiangshan struct workqueue_struct *wq; 5734042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5735042f7df1SLai Jiangshan 5736042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5737042f7df1SLai Jiangshan 5738042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5739042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5740042f7df1SLai Jiangshan continue; 5741042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5742042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5743042f7df1SLai Jiangshan continue; 5744042f7df1SLai Jiangshan 574599c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5746042f7df1SLai Jiangshan if (!ctx) { 5747042f7df1SLai Jiangshan ret = -ENOMEM; 5748042f7df1SLai Jiangshan break; 5749042f7df1SLai Jiangshan } 5750042f7df1SLai Jiangshan 5751042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5752042f7df1SLai Jiangshan } 5753042f7df1SLai Jiangshan 5754042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5755042f7df1SLai Jiangshan if (!ret) 5756042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5757042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5758042f7df1SLai Jiangshan } 5759042f7df1SLai Jiangshan 576099c621efSLai Jiangshan if (!ret) { 576199c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 576299c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 576399c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 576499c621efSLai Jiangshan } 5765042f7df1SLai Jiangshan return ret; 5766042f7df1SLai Jiangshan } 5767042f7df1SLai Jiangshan 5768042f7df1SLai Jiangshan /** 5769042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5770042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5771042f7df1SLai Jiangshan * 5772042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5773042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5774042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5775042f7df1SLai Jiangshan * 577667dc8325SCai Huoqing * Return: 0 - Success 5777042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5778042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5779042f7df1SLai Jiangshan */ 5780042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5781042f7df1SLai Jiangshan { 5782042f7df1SLai Jiangshan int ret = -EINVAL; 5783042f7df1SLai Jiangshan 5784c98a9805STal Shorer /* 5785c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5786c98a9805STal Shorer * If the user wishes to include them, we allow that. 5787c98a9805STal Shorer */ 5788042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5789042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5790a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5791d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5792d25302e4SMenglong Dong ret = 0; 5793d25302e4SMenglong Dong goto out_unlock; 5794d25302e4SMenglong Dong } 5795d25302e4SMenglong Dong 579699c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5797042f7df1SLai Jiangshan 5798d25302e4SMenglong Dong out_unlock: 5799a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5800042f7df1SLai Jiangshan } 5801042f7df1SLai Jiangshan 5802042f7df1SLai Jiangshan return ret; 5803042f7df1SLai Jiangshan } 5804042f7df1SLai Jiangshan 58056ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 58066ba94429SFrederic Weisbecker /* 58076ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 58086ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 58096ba94429SFrederic Weisbecker * following attributes. 58106ba94429SFrederic Weisbecker * 58116ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 58126ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 58136ba94429SFrederic Weisbecker * 58146ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 58156ba94429SFrederic Weisbecker * 58169a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 58176ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 58186ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 58199a19b463SWang Long * numa RW bool : whether enable NUMA affinity 58206ba94429SFrederic Weisbecker */ 58216ba94429SFrederic Weisbecker struct wq_device { 58226ba94429SFrederic Weisbecker struct workqueue_struct *wq; 58236ba94429SFrederic Weisbecker struct device dev; 58246ba94429SFrederic Weisbecker }; 58256ba94429SFrederic Weisbecker 58266ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 58276ba94429SFrederic Weisbecker { 58286ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 58296ba94429SFrederic Weisbecker 58306ba94429SFrederic Weisbecker return wq_dev->wq; 58316ba94429SFrederic Weisbecker } 58326ba94429SFrederic Weisbecker 58336ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 58346ba94429SFrederic Weisbecker char *buf) 58356ba94429SFrederic Weisbecker { 58366ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58376ba94429SFrederic Weisbecker 58386ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 58396ba94429SFrederic Weisbecker } 58406ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 58416ba94429SFrederic Weisbecker 58426ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 58436ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58446ba94429SFrederic Weisbecker { 58456ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58466ba94429SFrederic Weisbecker 58476ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 58486ba94429SFrederic Weisbecker } 58496ba94429SFrederic Weisbecker 58506ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 58516ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 58526ba94429SFrederic Weisbecker size_t count) 58536ba94429SFrederic Weisbecker { 58546ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58556ba94429SFrederic Weisbecker int val; 58566ba94429SFrederic Weisbecker 58576ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 58586ba94429SFrederic Weisbecker return -EINVAL; 58596ba94429SFrederic Weisbecker 58606ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 58616ba94429SFrederic Weisbecker return count; 58626ba94429SFrederic Weisbecker } 58636ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 58646ba94429SFrederic Weisbecker 58656ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 58666ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 58676ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 58686ba94429SFrederic Weisbecker NULL, 58696ba94429SFrederic Weisbecker }; 58706ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 58716ba94429SFrederic Weisbecker 58726ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 58736ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58746ba94429SFrederic Weisbecker { 58756ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58766ba94429SFrederic Weisbecker const char *delim = ""; 58776ba94429SFrederic Weisbecker int node, written = 0; 58786ba94429SFrederic Weisbecker 5879ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 588024acfb71SThomas Gleixner rcu_read_lock(); 58816ba94429SFrederic Weisbecker for_each_node(node) { 58826ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 58836ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 58846ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 58856ba94429SFrederic Weisbecker delim = " "; 58866ba94429SFrederic Weisbecker } 58876ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 588824acfb71SThomas Gleixner rcu_read_unlock(); 5889ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 58906ba94429SFrederic Weisbecker 58916ba94429SFrederic Weisbecker return written; 58926ba94429SFrederic Weisbecker } 58936ba94429SFrederic Weisbecker 58946ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 58956ba94429SFrederic Weisbecker char *buf) 58966ba94429SFrederic Weisbecker { 58976ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58986ba94429SFrederic Weisbecker int written; 58996ba94429SFrederic Weisbecker 59006ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59016ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 59026ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59036ba94429SFrederic Weisbecker 59046ba94429SFrederic Weisbecker return written; 59056ba94429SFrederic Weisbecker } 59066ba94429SFrederic Weisbecker 59076ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 59086ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 59096ba94429SFrederic Weisbecker { 59106ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 59116ba94429SFrederic Weisbecker 5912899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5913899a94feSLai Jiangshan 5914be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 59156ba94429SFrederic Weisbecker if (!attrs) 59166ba94429SFrederic Weisbecker return NULL; 59176ba94429SFrederic Weisbecker 59186ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 59196ba94429SFrederic Weisbecker return attrs; 59206ba94429SFrederic Weisbecker } 59216ba94429SFrederic Weisbecker 59226ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 59236ba94429SFrederic Weisbecker const char *buf, size_t count) 59246ba94429SFrederic Weisbecker { 59256ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59266ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5927d4d3e257SLai Jiangshan int ret = -ENOMEM; 5928d4d3e257SLai Jiangshan 5929d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59306ba94429SFrederic Weisbecker 59316ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59326ba94429SFrederic Weisbecker if (!attrs) 5933d4d3e257SLai Jiangshan goto out_unlock; 59346ba94429SFrederic Weisbecker 59356ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 59366ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5937d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59386ba94429SFrederic Weisbecker else 59396ba94429SFrederic Weisbecker ret = -EINVAL; 59406ba94429SFrederic Weisbecker 5941d4d3e257SLai Jiangshan out_unlock: 5942d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59436ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59446ba94429SFrederic Weisbecker return ret ?: count; 59456ba94429SFrederic Weisbecker } 59466ba94429SFrederic Weisbecker 59476ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 59486ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 59496ba94429SFrederic Weisbecker { 59506ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59516ba94429SFrederic Weisbecker int written; 59526ba94429SFrederic Weisbecker 59536ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59546ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 59556ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 59566ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59576ba94429SFrederic Weisbecker return written; 59586ba94429SFrederic Weisbecker } 59596ba94429SFrederic Weisbecker 59606ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 59616ba94429SFrederic Weisbecker struct device_attribute *attr, 59626ba94429SFrederic Weisbecker const char *buf, size_t count) 59636ba94429SFrederic Weisbecker { 59646ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59656ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5966d4d3e257SLai Jiangshan int ret = -ENOMEM; 5967d4d3e257SLai Jiangshan 5968d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59696ba94429SFrederic Weisbecker 59706ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59716ba94429SFrederic Weisbecker if (!attrs) 5972d4d3e257SLai Jiangshan goto out_unlock; 59736ba94429SFrederic Weisbecker 59746ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 59756ba94429SFrederic Weisbecker if (!ret) 5976d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59776ba94429SFrederic Weisbecker 5978d4d3e257SLai Jiangshan out_unlock: 5979d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59806ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59816ba94429SFrederic Weisbecker return ret ?: count; 59826ba94429SFrederic Weisbecker } 59836ba94429SFrederic Weisbecker 59846ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 59856ba94429SFrederic Weisbecker char *buf) 59866ba94429SFrederic Weisbecker { 59876ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59886ba94429SFrederic Weisbecker int written; 59896ba94429SFrederic Weisbecker 59906ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59916ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 59926ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 59936ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59946ba94429SFrederic Weisbecker 59956ba94429SFrederic Weisbecker return written; 59966ba94429SFrederic Weisbecker } 59976ba94429SFrederic Weisbecker 59986ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 59996ba94429SFrederic Weisbecker const char *buf, size_t count) 60006ba94429SFrederic Weisbecker { 60016ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 60026ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 6003d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 6004d4d3e257SLai Jiangshan 6005d4d3e257SLai Jiangshan apply_wqattrs_lock(); 60066ba94429SFrederic Weisbecker 60076ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 60086ba94429SFrederic Weisbecker if (!attrs) 6009d4d3e257SLai Jiangshan goto out_unlock; 60106ba94429SFrederic Weisbecker 60116ba94429SFrederic Weisbecker ret = -EINVAL; 60126ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 60136ba94429SFrederic Weisbecker attrs->no_numa = !v; 6014d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 60156ba94429SFrederic Weisbecker } 60166ba94429SFrederic Weisbecker 6017d4d3e257SLai Jiangshan out_unlock: 6018d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 60196ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 60206ba94429SFrederic Weisbecker return ret ?: count; 60216ba94429SFrederic Weisbecker } 60226ba94429SFrederic Weisbecker 60236ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 60246ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 60256ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 60266ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 60276ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 60286ba94429SFrederic Weisbecker __ATTR_NULL, 60296ba94429SFrederic Weisbecker }; 60306ba94429SFrederic Weisbecker 60316ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 60326ba94429SFrederic Weisbecker .name = "workqueue", 60336ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 60346ba94429SFrederic Weisbecker }; 60356ba94429SFrederic Weisbecker 6036b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 6037b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 6038b05a7928SFrederic Weisbecker { 6039b05a7928SFrederic Weisbecker int written; 6040b05a7928SFrederic Weisbecker 6041042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 6042b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6043b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 6044042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 6045b05a7928SFrederic Weisbecker 6046b05a7928SFrederic Weisbecker return written; 6047b05a7928SFrederic Weisbecker } 6048b05a7928SFrederic Weisbecker 6049042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 6050042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 6051042f7df1SLai Jiangshan { 6052042f7df1SLai Jiangshan cpumask_var_t cpumask; 6053042f7df1SLai Jiangshan int ret; 6054042f7df1SLai Jiangshan 6055042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6056042f7df1SLai Jiangshan return -ENOMEM; 6057042f7df1SLai Jiangshan 6058042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 6059042f7df1SLai Jiangshan if (!ret) 6060042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 6061042f7df1SLai Jiangshan 6062042f7df1SLai Jiangshan free_cpumask_var(cpumask); 6063042f7df1SLai Jiangshan return ret ? ret : count; 6064042f7df1SLai Jiangshan } 6065042f7df1SLai Jiangshan 6066b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 6067042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6068042f7df1SLai Jiangshan wq_unbound_cpumask_store); 6069b05a7928SFrederic Weisbecker 60706ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 60716ba94429SFrederic Weisbecker { 6072686f6697SGreg Kroah-Hartman struct device *dev_root; 6073b05a7928SFrederic Weisbecker int err; 6074b05a7928SFrederic Weisbecker 6075b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 6076b05a7928SFrederic Weisbecker if (err) 6077b05a7928SFrederic Weisbecker return err; 6078b05a7928SFrederic Weisbecker 6079686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 6080686f6697SGreg Kroah-Hartman if (dev_root) { 6081686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6082686f6697SGreg Kroah-Hartman put_device(dev_root); 6083686f6697SGreg Kroah-Hartman } 6084686f6697SGreg Kroah-Hartman return err; 60856ba94429SFrederic Weisbecker } 60866ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 60876ba94429SFrederic Weisbecker 60886ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 60896ba94429SFrederic Weisbecker { 60906ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 60916ba94429SFrederic Weisbecker 60926ba94429SFrederic Weisbecker kfree(wq_dev); 60936ba94429SFrederic Weisbecker } 60946ba94429SFrederic Weisbecker 60956ba94429SFrederic Weisbecker /** 60966ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 60976ba94429SFrederic Weisbecker * @wq: the workqueue to register 60986ba94429SFrederic Weisbecker * 60996ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 61006ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 61016ba94429SFrederic Weisbecker * which is the preferred method. 61026ba94429SFrederic Weisbecker * 61036ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 61046ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 61056ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 61066ba94429SFrederic Weisbecker * attributes. 61076ba94429SFrederic Weisbecker * 61086ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 61096ba94429SFrederic Weisbecker */ 61106ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 61116ba94429SFrederic Weisbecker { 61126ba94429SFrederic Weisbecker struct wq_device *wq_dev; 61136ba94429SFrederic Weisbecker int ret; 61146ba94429SFrederic Weisbecker 61156ba94429SFrederic Weisbecker /* 6116402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 61176ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 61186ba94429SFrederic Weisbecker * workqueues. 61196ba94429SFrederic Weisbecker */ 61200a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 61216ba94429SFrederic Weisbecker return -EINVAL; 61226ba94429SFrederic Weisbecker 61236ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 61246ba94429SFrederic Weisbecker if (!wq_dev) 61256ba94429SFrederic Weisbecker return -ENOMEM; 61266ba94429SFrederic Weisbecker 61276ba94429SFrederic Weisbecker wq_dev->wq = wq; 61286ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 61296ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 613023217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 61316ba94429SFrederic Weisbecker 61326ba94429SFrederic Weisbecker /* 61336ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 61346ba94429SFrederic Weisbecker * everything is ready. 61356ba94429SFrederic Weisbecker */ 61366ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 61376ba94429SFrederic Weisbecker 61386ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 61396ba94429SFrederic Weisbecker if (ret) { 6140537f4146SArvind Yadav put_device(&wq_dev->dev); 61416ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61426ba94429SFrederic Weisbecker return ret; 61436ba94429SFrederic Weisbecker } 61446ba94429SFrederic Weisbecker 61456ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 61466ba94429SFrederic Weisbecker struct device_attribute *attr; 61476ba94429SFrederic Weisbecker 61486ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 61496ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 61506ba94429SFrederic Weisbecker if (ret) { 61516ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61526ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61536ba94429SFrederic Weisbecker return ret; 61546ba94429SFrederic Weisbecker } 61556ba94429SFrederic Weisbecker } 61566ba94429SFrederic Weisbecker } 61576ba94429SFrederic Weisbecker 61586ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 61596ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 61606ba94429SFrederic Weisbecker return 0; 61616ba94429SFrederic Weisbecker } 61626ba94429SFrederic Weisbecker 61636ba94429SFrederic Weisbecker /** 61646ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 61656ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 61666ba94429SFrederic Weisbecker * 61676ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 61686ba94429SFrederic Weisbecker */ 61696ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 61706ba94429SFrederic Weisbecker { 61716ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 61726ba94429SFrederic Weisbecker 61736ba94429SFrederic Weisbecker if (!wq->wq_dev) 61746ba94429SFrederic Weisbecker return; 61756ba94429SFrederic Weisbecker 61766ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61776ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61786ba94429SFrederic Weisbecker } 61796ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 61806ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 61816ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 61826ba94429SFrederic Weisbecker 618382607adcSTejun Heo /* 618482607adcSTejun Heo * Workqueue watchdog. 618582607adcSTejun Heo * 618682607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 618782607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 618882607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 618982607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 619082607adcSTejun Heo * largely opaque. 619182607adcSTejun Heo * 619282607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 619382607adcSTejun Heo * state if some pools failed to make forward progress for a while where 619482607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 619582607adcSTejun Heo * 619682607adcSTejun Heo * This mechanism is controlled through the kernel parameter 619782607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 619882607adcSTejun Heo * corresponding sysfs parameter file. 619982607adcSTejun Heo */ 620082607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 620182607adcSTejun Heo 620282607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 62035cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 620482607adcSTejun Heo 620582607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 620682607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 620782607adcSTejun Heo 6208cd2440d6SPetr Mladek /* 6209cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6210cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6211cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6212cd2440d6SPetr Mladek * in all other situations. 6213cd2440d6SPetr Mladek */ 6214cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6215cd2440d6SPetr Mladek { 6216cd2440d6SPetr Mladek struct worker *worker; 6217cd2440d6SPetr Mladek unsigned long flags; 6218cd2440d6SPetr Mladek int bkt; 6219cd2440d6SPetr Mladek 6220cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6221cd2440d6SPetr Mladek 6222cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6223cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6224cd2440d6SPetr Mladek /* 6225cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6226cd2440d6SPetr Mladek * drivers that queue work while holding locks 6227cd2440d6SPetr Mladek * also taken in their write paths. 6228cd2440d6SPetr Mladek */ 6229cd2440d6SPetr Mladek printk_deferred_enter(); 6230cd2440d6SPetr Mladek 6231cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6232cd2440d6SPetr Mladek sched_show_task(worker->task); 6233cd2440d6SPetr Mladek 6234cd2440d6SPetr Mladek printk_deferred_exit(); 6235cd2440d6SPetr Mladek } 6236cd2440d6SPetr Mladek } 6237cd2440d6SPetr Mladek 6238cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6239cd2440d6SPetr Mladek } 6240cd2440d6SPetr Mladek 6241cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6242cd2440d6SPetr Mladek { 6243cd2440d6SPetr Mladek struct worker_pool *pool; 6244cd2440d6SPetr Mladek int pi; 6245cd2440d6SPetr Mladek 6246cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6247cd2440d6SPetr Mladek 6248cd2440d6SPetr Mladek rcu_read_lock(); 6249cd2440d6SPetr Mladek 6250cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6251cd2440d6SPetr Mladek if (pool->cpu_stall) 6252cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6253cd2440d6SPetr Mladek 6254cd2440d6SPetr Mladek } 6255cd2440d6SPetr Mladek 6256cd2440d6SPetr Mladek rcu_read_unlock(); 6257cd2440d6SPetr Mladek } 6258cd2440d6SPetr Mladek 625982607adcSTejun Heo static void wq_watchdog_reset_touched(void) 626082607adcSTejun Heo { 626182607adcSTejun Heo int cpu; 626282607adcSTejun Heo 626382607adcSTejun Heo wq_watchdog_touched = jiffies; 626482607adcSTejun Heo for_each_possible_cpu(cpu) 626582607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 626682607adcSTejun Heo } 626782607adcSTejun Heo 62685cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 626982607adcSTejun Heo { 627082607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 627182607adcSTejun Heo bool lockup_detected = false; 6272cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6273940d71c6SSergey Senozhatsky unsigned long now = jiffies; 627482607adcSTejun Heo struct worker_pool *pool; 627582607adcSTejun Heo int pi; 627682607adcSTejun Heo 627782607adcSTejun Heo if (!thresh) 627882607adcSTejun Heo return; 627982607adcSTejun Heo 628082607adcSTejun Heo rcu_read_lock(); 628182607adcSTejun Heo 628282607adcSTejun Heo for_each_pool(pool, pi) { 628382607adcSTejun Heo unsigned long pool_ts, touched, ts; 628482607adcSTejun Heo 6285cd2440d6SPetr Mladek pool->cpu_stall = false; 628682607adcSTejun Heo if (list_empty(&pool->worklist)) 628782607adcSTejun Heo continue; 628882607adcSTejun Heo 6289940d71c6SSergey Senozhatsky /* 6290940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6291940d71c6SSergey Senozhatsky * the watchdog like a stall. 6292940d71c6SSergey Senozhatsky */ 6293940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6294940d71c6SSergey Senozhatsky 629582607adcSTejun Heo /* get the latest of pool and touched timestamps */ 629689e28ce6SWang Qing if (pool->cpu >= 0) 629789e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 629889e28ce6SWang Qing else 629982607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 630089e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 630182607adcSTejun Heo 630282607adcSTejun Heo if (time_after(pool_ts, touched)) 630382607adcSTejun Heo ts = pool_ts; 630482607adcSTejun Heo else 630582607adcSTejun Heo ts = touched; 630682607adcSTejun Heo 630782607adcSTejun Heo /* did we stall? */ 6308940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 630982607adcSTejun Heo lockup_detected = true; 6310cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6311cd2440d6SPetr Mladek pool->cpu_stall = true; 6312cd2440d6SPetr Mladek cpu_pool_stall = true; 6313cd2440d6SPetr Mladek } 631482607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 631582607adcSTejun Heo pr_cont_pool_info(pool); 631682607adcSTejun Heo pr_cont(" stuck for %us!\n", 6317940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 631882607adcSTejun Heo } 6319cd2440d6SPetr Mladek 6320cd2440d6SPetr Mladek 632182607adcSTejun Heo } 632282607adcSTejun Heo 632382607adcSTejun Heo rcu_read_unlock(); 632482607adcSTejun Heo 632582607adcSTejun Heo if (lockup_detected) 632655df0933SImran Khan show_all_workqueues(); 632782607adcSTejun Heo 6328cd2440d6SPetr Mladek if (cpu_pool_stall) 6329cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6330cd2440d6SPetr Mladek 633182607adcSTejun Heo wq_watchdog_reset_touched(); 633282607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 633382607adcSTejun Heo } 633482607adcSTejun Heo 6335cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 633682607adcSTejun Heo { 633782607adcSTejun Heo if (cpu >= 0) 633882607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 633989e28ce6SWang Qing 634082607adcSTejun Heo wq_watchdog_touched = jiffies; 634182607adcSTejun Heo } 634282607adcSTejun Heo 634382607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 634482607adcSTejun Heo { 634582607adcSTejun Heo wq_watchdog_thresh = 0; 634682607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 634782607adcSTejun Heo 634882607adcSTejun Heo if (thresh) { 634982607adcSTejun Heo wq_watchdog_thresh = thresh; 635082607adcSTejun Heo wq_watchdog_reset_touched(); 635182607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 635282607adcSTejun Heo } 635382607adcSTejun Heo } 635482607adcSTejun Heo 635582607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 635682607adcSTejun Heo const struct kernel_param *kp) 635782607adcSTejun Heo { 635882607adcSTejun Heo unsigned long thresh; 635982607adcSTejun Heo int ret; 636082607adcSTejun Heo 636182607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 636282607adcSTejun Heo if (ret) 636382607adcSTejun Heo return ret; 636482607adcSTejun Heo 636582607adcSTejun Heo if (system_wq) 636682607adcSTejun Heo wq_watchdog_set_thresh(thresh); 636782607adcSTejun Heo else 636882607adcSTejun Heo wq_watchdog_thresh = thresh; 636982607adcSTejun Heo 637082607adcSTejun Heo return 0; 637182607adcSTejun Heo } 637282607adcSTejun Heo 637382607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 637482607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 637582607adcSTejun Heo .get = param_get_ulong, 637682607adcSTejun Heo }; 637782607adcSTejun Heo 637882607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 637982607adcSTejun Heo 0644); 638082607adcSTejun Heo 638182607adcSTejun Heo static void wq_watchdog_init(void) 638282607adcSTejun Heo { 63835cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 638482607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 638582607adcSTejun Heo } 638682607adcSTejun Heo 638782607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 638882607adcSTejun Heo 638982607adcSTejun Heo static inline void wq_watchdog_init(void) { } 639082607adcSTejun Heo 639182607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 639282607adcSTejun Heo 6393bce90380STejun Heo static void __init wq_numa_init(void) 6394bce90380STejun Heo { 6395bce90380STejun Heo cpumask_var_t *tbl; 6396bce90380STejun Heo int node, cpu; 6397bce90380STejun Heo 6398bce90380STejun Heo if (num_possible_nodes() <= 1) 6399bce90380STejun Heo return; 6400bce90380STejun Heo 6401d55262c4STejun Heo if (wq_disable_numa) { 6402d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 6403d55262c4STejun Heo return; 6404d55262c4STejun Heo } 6405d55262c4STejun Heo 6406f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6407f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6408f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6409f728c4a9SZhen Lei return; 6410f728c4a9SZhen Lei } 6411f728c4a9SZhen Lei } 6412f728c4a9SZhen Lei 6413be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 64144c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 64154c16bd32STejun Heo 6416bce90380STejun Heo /* 6417bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6418bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6419bce90380STejun Heo * fully initialized by now. 6420bce90380STejun Heo */ 64216396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6422bce90380STejun Heo BUG_ON(!tbl); 6423bce90380STejun Heo 6424bce90380STejun Heo for_each_node(node) 64255a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 64261be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6427bce90380STejun Heo 6428bce90380STejun Heo for_each_possible_cpu(cpu) { 6429bce90380STejun Heo node = cpu_to_node(cpu); 6430bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6431bce90380STejun Heo } 6432bce90380STejun Heo 6433bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6434bce90380STejun Heo wq_numa_enabled = true; 6435bce90380STejun Heo } 6436bce90380STejun Heo 64373347fa09STejun Heo /** 64383347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 64393347fa09STejun Heo * 64403347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 64413347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 64423347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 64433347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 64443347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 64453347fa09STejun Heo * created and scheduled right before early initcalls. 64463347fa09STejun Heo */ 64472333e829SYu Chen void __init workqueue_init_early(void) 64481da177e4SLinus Torvalds { 64497a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 64507a4e344cSTejun Heo int i, cpu; 6451c34056a3STejun Heo 645210cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6453e904e6c2STejun Heo 6454b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 645504d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 645604d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6457b05a7928SFrederic Weisbecker 6458e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6459e904e6c2STejun Heo 6460706026c2STejun Heo /* initialize CPU pools */ 646129c91e99STejun Heo for_each_possible_cpu(cpu) { 64624ce62e9eSTejun Heo struct worker_pool *pool; 64638b03ae3cSTejun Heo 64647a4e344cSTejun Heo i = 0; 6465f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 64667a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6467ec22ca5eSTejun Heo pool->cpu = cpu; 64687a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 64697a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6470f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 64717a4e344cSTejun Heo 64729daf9e67STejun Heo /* alloc pool ID */ 647368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 64749daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 647568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 64764ce62e9eSTejun Heo } 64778b03ae3cSTejun Heo } 64788b03ae3cSTejun Heo 64798a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 648029c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 648129c91e99STejun Heo struct workqueue_attrs *attrs; 648229c91e99STejun Heo 6483be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 648429c91e99STejun Heo attrs->nice = std_nice[i]; 648529c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 64868a2b7538STejun Heo 64878a2b7538STejun Heo /* 64888a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 64898a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 64908a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 64918a2b7538STejun Heo */ 6492be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 64938a2b7538STejun Heo attrs->nice = std_nice[i]; 64948a2b7538STejun Heo attrs->no_numa = true; 64958a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 649629c91e99STejun Heo } 649729c91e99STejun Heo 6498d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 64991aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6500d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6501f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6502f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 650324d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 650424d51addSTejun Heo WQ_FREEZABLE, 0); 65050668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 65060668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 65070668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 65080668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 65090668106cSViresh Kumar 0); 65101aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 65110668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 65120668106cSViresh Kumar !system_power_efficient_wq || 65130668106cSViresh Kumar !system_freezable_power_efficient_wq); 65143347fa09STejun Heo } 65153347fa09STejun Heo 65163347fa09STejun Heo /** 65173347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 65183347fa09STejun Heo * 65193347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 65203347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 65213347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 65223347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 65233347fa09STejun Heo * with the initial workers and enable future kworker creations. 65243347fa09STejun Heo */ 65252333e829SYu Chen void __init workqueue_init(void) 65263347fa09STejun Heo { 65272186d9f9STejun Heo struct workqueue_struct *wq; 65283347fa09STejun Heo struct worker_pool *pool; 65293347fa09STejun Heo int cpu, bkt; 65303347fa09STejun Heo 65312186d9f9STejun Heo /* 65322186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 65332186d9f9STejun Heo * CPU to node mapping may not be available that early on some 65342186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 65352186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 65362186d9f9STejun Heo * affinity, fix them up. 653740c17f75STejun Heo * 653840c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 65392186d9f9STejun Heo */ 65402186d9f9STejun Heo wq_numa_init(); 65412186d9f9STejun Heo 65422186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 65432186d9f9STejun Heo 65442186d9f9STejun Heo for_each_possible_cpu(cpu) { 65452186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65462186d9f9STejun Heo pool->node = cpu_to_node(cpu); 65472186d9f9STejun Heo } 65482186d9f9STejun Heo } 65492186d9f9STejun Heo 655040c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 65512186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 655240c17f75STejun Heo WARN(init_rescuer(wq), 655340c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 655440c17f75STejun Heo wq->name); 655540c17f75STejun Heo } 65562186d9f9STejun Heo 65572186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 65582186d9f9STejun Heo 65593347fa09STejun Heo /* create the initial workers */ 65603347fa09STejun Heo for_each_online_cpu(cpu) { 65613347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65623347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 65633347fa09STejun Heo BUG_ON(!create_worker(pool)); 65643347fa09STejun Heo } 65653347fa09STejun Heo } 65663347fa09STejun Heo 65673347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 65683347fa09STejun Heo BUG_ON(!create_worker(pool)); 65693347fa09STejun Heo 65703347fa09STejun Heo wq_online = true; 657182607adcSTejun Heo wq_watchdog_init(); 65721da177e4SLinus Torvalds } 6573c4f135d6STetsuo Handa 6574*20bdedafSTetsuo Handa void __warn_flushing_systemwide_wq(void) 6575*20bdedafSTetsuo Handa { 6576*20bdedafSTetsuo Handa pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 6577*20bdedafSTetsuo Handa dump_stack(); 6578*20bdedafSTetsuo Handa } 6579c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6580