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 740112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7417a22ad75STejun Heo { 742e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7437a22ad75STejun Heo 744112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 745e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 746e120153dSTejun Heo else 747e120153dSTejun Heo return NULL; 7487a22ad75STejun Heo } 7497a22ad75STejun Heo 7507c3eed5cSTejun Heo /** 7517c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7527c3eed5cSTejun Heo * @work: the work item of interest 7537c3eed5cSTejun Heo * 75468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 75524acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 75624acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 757fa1b54e6STejun Heo * 758fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 759fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 760fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 761fa1b54e6STejun Heo * returned pool is and stays online. 762d185af30SYacine Belkadi * 763d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7647c3eed5cSTejun Heo */ 7657c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7667a22ad75STejun Heo { 767e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7687c3eed5cSTejun Heo int pool_id; 7697a22ad75STejun Heo 77068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 771fa1b54e6STejun Heo 772112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 773112202d9STejun Heo return ((struct pool_workqueue *) 7747c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7757a22ad75STejun Heo 7767c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7777c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7787a22ad75STejun Heo return NULL; 7797a22ad75STejun Heo 780fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7817c3eed5cSTejun Heo } 7827c3eed5cSTejun Heo 7837c3eed5cSTejun Heo /** 7847c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7857c3eed5cSTejun Heo * @work: the work item of interest 7867c3eed5cSTejun Heo * 787d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7887c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7897c3eed5cSTejun Heo */ 7907c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7917c3eed5cSTejun Heo { 79254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7937c3eed5cSTejun Heo 794112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 795112202d9STejun Heo return ((struct pool_workqueue *) 79654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 79754d5b7d0SLai Jiangshan 79854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7997c3eed5cSTejun Heo } 8007c3eed5cSTejun Heo 801bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 802bbb68dfaSTejun Heo { 8037c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 804bbb68dfaSTejun Heo 8057c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 8067c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 807bbb68dfaSTejun Heo } 808bbb68dfaSTejun Heo 809bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 810bbb68dfaSTejun Heo { 811bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 812bbb68dfaSTejun Heo 813112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 814bbb68dfaSTejun Heo } 815bbb68dfaSTejun Heo 816e22bee78STejun Heo /* 8173270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8183270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 819d565ed63STejun Heo * they're being called with pool->lock held. 820e22bee78STejun Heo */ 821e22bee78STejun Heo 82263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 823649027d7STejun Heo { 824bc35f7efSLai Jiangshan return !pool->nr_running; 825649027d7STejun Heo } 826649027d7STejun Heo 827e22bee78STejun Heo /* 828e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 829e22bee78STejun Heo * running workers. 830974271c4STejun Heo * 831974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 832706026c2STejun Heo * function will always return %true for unbound pools as long as the 833974271c4STejun Heo * worklist isn't empty. 834e22bee78STejun Heo */ 83563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 836e22bee78STejun Heo { 83763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 838e22bee78STejun Heo } 839e22bee78STejun Heo 840e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 84163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 842e22bee78STejun Heo { 84363d95a91STejun Heo return pool->nr_idle; 844e22bee78STejun Heo } 845e22bee78STejun Heo 846e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 84763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 848e22bee78STejun Heo { 849bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 850e22bee78STejun Heo } 851e22bee78STejun Heo 852e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 85363d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 854e22bee78STejun Heo { 85563d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 856e22bee78STejun Heo } 857e22bee78STejun Heo 858e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 85963d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 860e22bee78STejun Heo { 861692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 86263d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 86363d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 864e22bee78STejun Heo 865e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 866e22bee78STejun Heo } 867e22bee78STejun Heo 868e22bee78STejun Heo /* 869e22bee78STejun Heo * Wake up functions. 870e22bee78STejun Heo */ 871e22bee78STejun Heo 8722c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8731037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8747e11629dSTejun Heo { 87563d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8767e11629dSTejun Heo return NULL; 8777e11629dSTejun Heo 87863d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8797e11629dSTejun Heo } 8807e11629dSTejun Heo 8817e11629dSTejun Heo /** 8827e11629dSTejun Heo * wake_up_worker - wake up an idle worker 88363d95a91STejun Heo * @pool: worker pool to wake worker from 8847e11629dSTejun Heo * 88563d95a91STejun Heo * Wake up the first idle worker of @pool. 8867e11629dSTejun Heo * 8877e11629dSTejun Heo * CONTEXT: 888a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8897e11629dSTejun Heo */ 89063d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8917e11629dSTejun Heo { 8921037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8937e11629dSTejun Heo 8947e11629dSTejun Heo if (likely(worker)) 8957e11629dSTejun Heo wake_up_process(worker->task); 8967e11629dSTejun Heo } 8977e11629dSTejun Heo 8984690c4abSTejun Heo /** 899c54d5046STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 900c54d5046STejun Heo * @worker: self 901c54d5046STejun Heo * @flags: flags to set 902c54d5046STejun Heo * 903c54d5046STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. 904c54d5046STejun Heo * 905c54d5046STejun Heo * CONTEXT: 906c54d5046STejun Heo * raw_spin_lock_irq(pool->lock) 907c54d5046STejun Heo */ 908c54d5046STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags) 909c54d5046STejun Heo { 910c54d5046STejun Heo struct worker_pool *pool = worker->pool; 911c54d5046STejun Heo 912c54d5046STejun Heo WARN_ON_ONCE(worker->task != current); 913c54d5046STejun Heo 914c54d5046STejun Heo /* If transitioning into NOT_RUNNING, adjust nr_running. */ 915c54d5046STejun Heo if ((flags & WORKER_NOT_RUNNING) && 916c54d5046STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 917c54d5046STejun Heo pool->nr_running--; 918c54d5046STejun Heo } 919c54d5046STejun Heo 920c54d5046STejun Heo worker->flags |= flags; 921c54d5046STejun Heo } 922c54d5046STejun Heo 923c54d5046STejun Heo /** 924c54d5046STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 925c54d5046STejun Heo * @worker: self 926c54d5046STejun Heo * @flags: flags to clear 927c54d5046STejun Heo * 928c54d5046STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 929c54d5046STejun Heo * 930c54d5046STejun Heo * CONTEXT: 931c54d5046STejun Heo * raw_spin_lock_irq(pool->lock) 932c54d5046STejun Heo */ 933c54d5046STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 934c54d5046STejun Heo { 935c54d5046STejun Heo struct worker_pool *pool = worker->pool; 936c54d5046STejun Heo unsigned int oflags = worker->flags; 937c54d5046STejun Heo 938c54d5046STejun Heo WARN_ON_ONCE(worker->task != current); 939c54d5046STejun Heo 940c54d5046STejun Heo worker->flags &= ~flags; 941c54d5046STejun Heo 942c54d5046STejun Heo /* 943c54d5046STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 944c54d5046STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 945c54d5046STejun Heo * of multiple flags, not a single flag. 946c54d5046STejun Heo */ 947c54d5046STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 948c54d5046STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 949c54d5046STejun Heo pool->nr_running++; 950c54d5046STejun Heo } 951c54d5046STejun Heo 95263638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 95363638450STejun Heo 95463638450STejun Heo /* 95563638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 95663638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 95763638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 95863638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 95963638450STejun Heo * should be using an unbound workqueue instead. 96063638450STejun Heo * 96163638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 96263638450STejun Heo * and report them so that they can be examined and converted to use unbound 96363638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 96463638450STejun Heo * function is tracked and reported with exponential backoff. 96563638450STejun Heo */ 96663638450STejun Heo #define WCI_MAX_ENTS 128 96763638450STejun Heo 96863638450STejun Heo struct wci_ent { 96963638450STejun Heo work_func_t func; 97063638450STejun Heo atomic64_t cnt; 97163638450STejun Heo struct hlist_node hash_node; 97263638450STejun Heo }; 97363638450STejun Heo 97463638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 97563638450STejun Heo static int wci_nr_ents; 97663638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 97763638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 97863638450STejun Heo 97963638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 98063638450STejun Heo { 98163638450STejun Heo struct wci_ent *ent; 98263638450STejun Heo 98363638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 98463638450STejun Heo (unsigned long)func) { 98563638450STejun Heo if (ent->func == func) 98663638450STejun Heo return ent; 98763638450STejun Heo } 98863638450STejun Heo return NULL; 98963638450STejun Heo } 99063638450STejun Heo 99163638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 99263638450STejun Heo { 99363638450STejun Heo struct wci_ent *ent; 99463638450STejun Heo 99563638450STejun Heo restart: 99663638450STejun Heo ent = wci_find_ent(func); 99763638450STejun Heo if (ent) { 99863638450STejun Heo u64 cnt; 99963638450STejun Heo 100063638450STejun Heo /* 100163638450STejun Heo * Start reporting from the fourth time and back off 100263638450STejun Heo * exponentially. 100363638450STejun Heo */ 100463638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 100563638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 100663638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 100763638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 100863638450STejun Heo atomic64_read(&ent->cnt)); 100963638450STejun Heo return; 101063638450STejun Heo } 101163638450STejun Heo 101263638450STejun Heo /* 101363638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 101463638450STejun Heo * is exhausted, something went really wrong and we probably made enough 101563638450STejun Heo * noise already. 101663638450STejun Heo */ 101763638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 101863638450STejun Heo return; 101963638450STejun Heo 102063638450STejun Heo raw_spin_lock(&wci_lock); 102163638450STejun Heo 102263638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 102363638450STejun Heo raw_spin_unlock(&wci_lock); 102463638450STejun Heo return; 102563638450STejun Heo } 102663638450STejun Heo 102763638450STejun Heo if (wci_find_ent(func)) { 102863638450STejun Heo raw_spin_unlock(&wci_lock); 102963638450STejun Heo goto restart; 103063638450STejun Heo } 103163638450STejun Heo 103263638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 103363638450STejun Heo ent->func = func; 103463638450STejun Heo atomic64_set(&ent->cnt, 1); 103563638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 103663638450STejun Heo 103763638450STejun Heo raw_spin_unlock(&wci_lock); 103863638450STejun Heo } 103963638450STejun Heo 104063638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 104163638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 104263638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 104363638450STejun Heo 1044c54d5046STejun Heo /** 10456d25be57SThomas Gleixner * wq_worker_running - a worker is running again 1046e22bee78STejun Heo * @task: task waking up 1047e22bee78STejun Heo * 10486d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 1049e22bee78STejun Heo */ 10506d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 1051e22bee78STejun Heo { 1052e22bee78STejun Heo struct worker *worker = kthread_data(task); 1053e22bee78STejun Heo 1054*c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 10556d25be57SThomas Gleixner return; 105607edfeceSFrederic Weisbecker 105707edfeceSFrederic Weisbecker /* 105807edfeceSFrederic Weisbecker * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 105907edfeceSFrederic Weisbecker * and the nr_running increment below, we may ruin the nr_running reset 106007edfeceSFrederic Weisbecker * and leave with an unexpected pool->nr_running == 1 on the newly unbound 106107edfeceSFrederic Weisbecker * pool. Protect against such race. 106207edfeceSFrederic Weisbecker */ 106307edfeceSFrederic Weisbecker preempt_disable(); 10646d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 1065bc35f7efSLai Jiangshan worker->pool->nr_running++; 106607edfeceSFrederic Weisbecker preempt_enable(); 1067616db877STejun Heo 1068616db877STejun Heo /* 1069616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1070616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1071616db877STejun Heo */ 1072616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1073616db877STejun Heo 1074*c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 107536576000SJoonsoo Kim } 1076e22bee78STejun Heo 1077e22bee78STejun Heo /** 1078e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 1079e22bee78STejun Heo * @task: task going to sleep 1080e22bee78STejun Heo * 10816d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 1082ccf45156SLai Jiangshan * going to sleep. 1083e22bee78STejun Heo */ 10846d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 1085e22bee78STejun Heo { 1086cc5bff38SLai Jiangshan struct worker *worker = kthread_data(task); 1087111c225aSTejun Heo struct worker_pool *pool; 1088e22bee78STejun Heo 1089111c225aSTejun Heo /* 1090111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 1091111c225aSTejun Heo * workers, also reach here, let's not access anything before 1092111c225aSTejun Heo * checking NOT_RUNNING. 1093111c225aSTejun Heo */ 10942d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 10956d25be57SThomas Gleixner return; 1096e22bee78STejun Heo 1097111c225aSTejun Heo pool = worker->pool; 1098111c225aSTejun Heo 109962849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 1100*c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 11016d25be57SThomas Gleixner return; 11026d25be57SThomas Gleixner 1103*c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 1104a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1105e22bee78STejun Heo 1106e22bee78STejun Heo /* 110745c753f5SFrederic Weisbecker * Recheck in case unbind_workers() preempted us. We don't 110845c753f5SFrederic Weisbecker * want to decrement nr_running after the worker is unbound 110945c753f5SFrederic Weisbecker * and nr_running has been reset. 111045c753f5SFrederic Weisbecker */ 111145c753f5SFrederic Weisbecker if (worker->flags & WORKER_NOT_RUNNING) { 111245c753f5SFrederic Weisbecker raw_spin_unlock_irq(&pool->lock); 111345c753f5SFrederic Weisbecker return; 111445c753f5SFrederic Weisbecker } 111545c753f5SFrederic Weisbecker 1116bc35f7efSLai Jiangshan pool->nr_running--; 1117725e8ec5STejun Heo if (need_more_worker(pool)) { 1118725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1119cc5bff38SLai Jiangshan wake_up_worker(pool); 1120725e8ec5STejun Heo } 1121a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1122e22bee78STejun Heo } 1123e22bee78STejun Heo 1124e22bee78STejun Heo /** 1125616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1126616db877STejun Heo * @task: task currently running 1127616db877STejun Heo * 1128616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1129616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1130616db877STejun Heo */ 1131616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1132616db877STejun Heo { 1133616db877STejun Heo struct worker *worker = kthread_data(task); 1134616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1135616db877STejun Heo struct worker_pool *pool = worker->pool; 1136616db877STejun Heo 1137616db877STejun Heo if (!pwq) 1138616db877STejun Heo return; 1139616db877STejun Heo 11408a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 11418a1dd1e5STejun Heo 1142616db877STejun Heo /* 1143616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1144616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1145616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1146*c8f6219bSZqiang * 1147*c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1148*c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1149*c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1150*c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1151*c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1152*c8f6219bSZqiang * We probably want to make this prettier in the future. 1153616db877STejun Heo */ 1154*c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1155616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1156616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1157616db877STejun Heo return; 1158616db877STejun Heo 1159616db877STejun Heo raw_spin_lock(&pool->lock); 1160616db877STejun Heo 1161616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 116263638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1163616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1164616db877STejun Heo 1165616db877STejun Heo if (need_more_worker(pool)) { 1166616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1167616db877STejun Heo wake_up_worker(pool); 1168616db877STejun Heo } 1169616db877STejun Heo 1170616db877STejun Heo raw_spin_unlock(&pool->lock); 1171616db877STejun Heo } 1172616db877STejun Heo 1173616db877STejun Heo /** 11741b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 11758194fe94SBart Van Assche * @task: Task to retrieve last work function of. 11761b69ac6bSJohannes Weiner * 11771b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 11781b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 11791b69ac6bSJohannes Weiner * 11801b69ac6bSJohannes Weiner * CONTEXT: 1181a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 11821b69ac6bSJohannes Weiner * 11834b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 11844b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 11854b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 11864b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 11874b047002SJohannes Weiner * 11884b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 11894b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 11904b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 11914b047002SJohannes Weiner * is guaranteed to not be processing any works. 11924b047002SJohannes Weiner * 11931b69ac6bSJohannes Weiner * Return: 11941b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 11951b69ac6bSJohannes Weiner * hasn't executed any work yet. 11961b69ac6bSJohannes Weiner */ 11971b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 11981b69ac6bSJohannes Weiner { 11991b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 12001b69ac6bSJohannes Weiner 12011b69ac6bSJohannes Weiner return worker->last_func; 12021b69ac6bSJohannes Weiner } 12031b69ac6bSJohannes Weiner 12041b69ac6bSJohannes Weiner /** 12058cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1206c9e7cf27STejun Heo * @pool: pool of interest 12078cca0eeaSTejun Heo * @work: work to find worker for 12088cca0eeaSTejun Heo * 1209c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1210c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1211a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1212a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1213a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1214a2c1c57bSTejun Heo * being executed. 1215a2c1c57bSTejun Heo * 1216a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1217a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1218a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1219a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1220a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1221a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1222a2c1c57bSTejun Heo * 1223c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1224c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1225c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1226c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1227c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1228c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 12298cca0eeaSTejun Heo * 12308cca0eeaSTejun Heo * CONTEXT: 1231a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 12328cca0eeaSTejun Heo * 1233d185af30SYacine Belkadi * Return: 1234d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 12358cca0eeaSTejun Heo * otherwise. 12368cca0eeaSTejun Heo */ 1237c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 12388cca0eeaSTejun Heo struct work_struct *work) 12398cca0eeaSTejun Heo { 124042f8570fSSasha Levin struct worker *worker; 124142f8570fSSasha Levin 1242b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1243a2c1c57bSTejun Heo (unsigned long)work) 1244a2c1c57bSTejun Heo if (worker->current_work == work && 1245a2c1c57bSTejun Heo worker->current_func == work->func) 124642f8570fSSasha Levin return worker; 124742f8570fSSasha Levin 124842f8570fSSasha Levin return NULL; 12498cca0eeaSTejun Heo } 12508cca0eeaSTejun Heo 12518cca0eeaSTejun Heo /** 1252bf4ede01STejun Heo * move_linked_works - move linked works to a list 1253bf4ede01STejun Heo * @work: start of series of works to be scheduled 1254bf4ede01STejun Heo * @head: target list to append @work to 1255402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1256bf4ede01STejun Heo * 1257bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1258bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1259bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1260bf4ede01STejun Heo * 1261bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1262bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1263bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1264bf4ede01STejun Heo * 1265bf4ede01STejun Heo * CONTEXT: 1266a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1267bf4ede01STejun Heo */ 1268bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1269bf4ede01STejun Heo struct work_struct **nextp) 1270bf4ede01STejun Heo { 1271bf4ede01STejun Heo struct work_struct *n; 1272bf4ede01STejun Heo 1273bf4ede01STejun Heo /* 1274bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1275bf4ede01STejun Heo * use NULL for list head. 1276bf4ede01STejun Heo */ 1277bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1278bf4ede01STejun Heo list_move_tail(&work->entry, head); 1279bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1280bf4ede01STejun Heo break; 1281bf4ede01STejun Heo } 1282bf4ede01STejun Heo 1283bf4ede01STejun Heo /* 1284bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1285bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1286bf4ede01STejun Heo * needs to be updated. 1287bf4ede01STejun Heo */ 1288bf4ede01STejun Heo if (nextp) 1289bf4ede01STejun Heo *nextp = n; 1290bf4ede01STejun Heo } 1291bf4ede01STejun Heo 12928864b4e5STejun Heo /** 12938864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 12948864b4e5STejun Heo * @pwq: pool_workqueue to get 12958864b4e5STejun Heo * 12968864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 12978864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 12988864b4e5STejun Heo */ 12998864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 13008864b4e5STejun Heo { 13018864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13028864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 13038864b4e5STejun Heo pwq->refcnt++; 13048864b4e5STejun Heo } 13058864b4e5STejun Heo 13068864b4e5STejun Heo /** 13078864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 13088864b4e5STejun Heo * @pwq: pool_workqueue to put 13098864b4e5STejun Heo * 13108864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 13118864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 13128864b4e5STejun Heo */ 13138864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 13148864b4e5STejun Heo { 13158864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13168864b4e5STejun Heo if (likely(--pwq->refcnt)) 13178864b4e5STejun Heo return; 13188864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 13198864b4e5STejun Heo return; 13208864b4e5STejun Heo /* 13218864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 13228864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 13238864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 13248864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 13258864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 13268864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 13278864b4e5STejun Heo */ 13288864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 13298864b4e5STejun Heo } 13308864b4e5STejun Heo 1331dce90d47STejun Heo /** 1332dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1333dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1334dce90d47STejun Heo * 1335dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1336dce90d47STejun Heo */ 1337dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1338dce90d47STejun Heo { 1339dce90d47STejun Heo if (pwq) { 1340dce90d47STejun Heo /* 134124acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1342dce90d47STejun Heo * following lock operations are safe. 1343dce90d47STejun Heo */ 1344a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1345dce90d47STejun Heo put_pwq(pwq); 1346a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1347dce90d47STejun Heo } 1348dce90d47STejun Heo } 1349dce90d47STejun Heo 1350f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1351bf4ede01STejun Heo { 1352112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1353bf4ede01STejun Heo 1354bf4ede01STejun Heo trace_workqueue_activate_work(work); 135582607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 135682607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1357112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1358f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1359112202d9STejun Heo pwq->nr_active++; 1360bf4ede01STejun Heo } 1361bf4ede01STejun Heo 1362f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 13633aa62497SLai Jiangshan { 1364f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 13653aa62497SLai Jiangshan struct work_struct, entry); 13663aa62497SLai Jiangshan 1367f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13683aa62497SLai Jiangshan } 13693aa62497SLai Jiangshan 1370bf4ede01STejun Heo /** 1371112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1372112202d9STejun Heo * @pwq: pwq of interest 1373c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1374bf4ede01STejun Heo * 1375bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1376112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1377bf4ede01STejun Heo * 1378bf4ede01STejun Heo * CONTEXT: 1379a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1380bf4ede01STejun Heo */ 1381c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1382bf4ede01STejun Heo { 1383c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1384c4560c2cSLai Jiangshan 1385018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1386112202d9STejun Heo pwq->nr_active--; 1387f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1388f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1389112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1390f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1391bf4ede01STejun Heo } 1392018f3a13SLai Jiangshan } 1393018f3a13SLai Jiangshan 1394018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1395bf4ede01STejun Heo 1396bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1397112202d9STejun Heo if (likely(pwq->flush_color != color)) 13988864b4e5STejun Heo goto out_put; 1399bf4ede01STejun Heo 1400bf4ede01STejun Heo /* are there still in-flight works? */ 1401112202d9STejun Heo if (pwq->nr_in_flight[color]) 14028864b4e5STejun Heo goto out_put; 1403bf4ede01STejun Heo 1404112202d9STejun Heo /* this pwq is done, clear flush_color */ 1405112202d9STejun Heo pwq->flush_color = -1; 1406bf4ede01STejun Heo 1407bf4ede01STejun Heo /* 1408112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1409bf4ede01STejun Heo * will handle the rest. 1410bf4ede01STejun Heo */ 1411112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1412112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 14138864b4e5STejun Heo out_put: 14148864b4e5STejun Heo put_pwq(pwq); 1415bf4ede01STejun Heo } 1416bf4ede01STejun Heo 141736e227d2STejun Heo /** 1418bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 141936e227d2STejun Heo * @work: work item to steal 142036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1421bbb68dfaSTejun Heo * @flags: place to store irq state 142236e227d2STejun Heo * 142336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1424d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 142536e227d2STejun Heo * 1426d185af30SYacine Belkadi * Return: 14273eb6b31bSMauro Carvalho Chehab * 14283eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 142936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 143036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 143136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1432bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1433bbb68dfaSTejun Heo * for arbitrarily long 14343eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 143536e227d2STejun Heo * 1436d185af30SYacine Belkadi * Note: 1437bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1438e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1439e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1440e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1441bbb68dfaSTejun Heo * 1442bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1443bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1444bbb68dfaSTejun Heo * 1445e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1446bf4ede01STejun Heo */ 1447bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1448bbb68dfaSTejun Heo unsigned long *flags) 1449bf4ede01STejun Heo { 1450d565ed63STejun Heo struct worker_pool *pool; 1451112202d9STejun Heo struct pool_workqueue *pwq; 1452bf4ede01STejun Heo 1453bbb68dfaSTejun Heo local_irq_save(*flags); 1454bbb68dfaSTejun Heo 145536e227d2STejun Heo /* try to steal the timer if it exists */ 145636e227d2STejun Heo if (is_dwork) { 145736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 145836e227d2STejun Heo 1459e0aecdd8STejun Heo /* 1460e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1461e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1462e0aecdd8STejun Heo * running on the local CPU. 1463e0aecdd8STejun Heo */ 146436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 146536e227d2STejun Heo return 1; 146636e227d2STejun Heo } 146736e227d2STejun Heo 146836e227d2STejun Heo /* try to claim PENDING the normal way */ 1469bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1470bf4ede01STejun Heo return 0; 1471bf4ede01STejun Heo 147224acfb71SThomas Gleixner rcu_read_lock(); 1473bf4ede01STejun Heo /* 1474bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1475bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1476bf4ede01STejun Heo */ 1477d565ed63STejun Heo pool = get_work_pool(work); 1478d565ed63STejun Heo if (!pool) 1479bbb68dfaSTejun Heo goto fail; 1480bf4ede01STejun Heo 1481a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1482bf4ede01STejun Heo /* 1483112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1484112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1485112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1486112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1487112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 14880b3dae68SLai Jiangshan * item is currently queued on that pool. 1489bf4ede01STejun Heo */ 1490112202d9STejun Heo pwq = get_work_pwq(work); 1491112202d9STejun Heo if (pwq && pwq->pool == pool) { 1492bf4ede01STejun Heo debug_work_deactivate(work); 14933aa62497SLai Jiangshan 14943aa62497SLai Jiangshan /* 1495018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1496018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1497018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1498018f3a13SLai Jiangshan * 1499f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1500d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1501f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 150216062836STejun Heo * management later on and cause stall. Make sure the work 150316062836STejun Heo * item is activated before grabbing. 15043aa62497SLai Jiangshan */ 1505f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1506f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 15073aa62497SLai Jiangshan 1508bf4ede01STejun Heo list_del_init(&work->entry); 1509c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 151036e227d2STejun Heo 1511112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 15124468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 15134468a00fSLai Jiangshan 1514a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 151524acfb71SThomas Gleixner rcu_read_unlock(); 151636e227d2STejun Heo return 1; 1517bf4ede01STejun Heo } 1518a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1519bbb68dfaSTejun Heo fail: 152024acfb71SThomas Gleixner rcu_read_unlock(); 1521bbb68dfaSTejun Heo local_irq_restore(*flags); 1522bbb68dfaSTejun Heo if (work_is_canceling(work)) 1523bbb68dfaSTejun Heo return -ENOENT; 1524bbb68dfaSTejun Heo cpu_relax(); 152536e227d2STejun Heo return -EAGAIN; 1526bf4ede01STejun Heo } 1527bf4ede01STejun Heo 1528bf4ede01STejun Heo /** 1529706026c2STejun Heo * insert_work - insert a work into a pool 1530112202d9STejun Heo * @pwq: pwq @work belongs to 15314690c4abSTejun Heo * @work: work to insert 15324690c4abSTejun Heo * @head: insertion point 15334690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 15344690c4abSTejun Heo * 1535112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1536706026c2STejun Heo * work_struct flags. 15374690c4abSTejun Heo * 15384690c4abSTejun Heo * CONTEXT: 1539a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1540365970a1SDavid Howells */ 1541112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1542112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1543b89deed3SOleg Nesterov { 1544112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1545e1d8aa9fSFrederic Weisbecker 1546e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1547f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1548e89a85d6SWalter Wu 15494690c4abSTejun Heo /* we own @work, set data and link */ 1550112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 15511a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 15528864b4e5STejun Heo get_pwq(pwq); 1553e22bee78STejun Heo 155463d95a91STejun Heo if (__need_more_worker(pool)) 155563d95a91STejun Heo wake_up_worker(pool); 1556b89deed3SOleg Nesterov } 1557b89deed3SOleg Nesterov 1558c8efcc25STejun Heo /* 1559c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 15608d03ecfeSTejun Heo * same workqueue. 1561c8efcc25STejun Heo */ 1562c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1563c8efcc25STejun Heo { 1564c8efcc25STejun Heo struct worker *worker; 1565c8efcc25STejun Heo 15668d03ecfeSTejun Heo worker = current_wq_worker(); 1567c8efcc25STejun Heo /* 1568bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 15698d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1570c8efcc25STejun Heo */ 1571112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1572c8efcc25STejun Heo } 1573c8efcc25STejun Heo 1574ef557180SMike Galbraith /* 1575ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1576ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1577ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1578ef557180SMike Galbraith */ 1579ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1580ef557180SMike Galbraith { 1581ef557180SMike Galbraith int new_cpu; 1582ef557180SMike Galbraith 1583f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1584ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1585ef557180SMike Galbraith return cpu; 1586a8ec5880SAmmar Faizi } else { 1587a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1588f303fccbSTejun Heo } 1589f303fccbSTejun Heo 1590ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1591ef557180SMike Galbraith return cpu; 1592ef557180SMike Galbraith 1593ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1594ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1595ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1596ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1597ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1598ef557180SMike Galbraith return cpu; 1599ef557180SMike Galbraith } 1600ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1601ef557180SMike Galbraith 1602ef557180SMike Galbraith return new_cpu; 1603ef557180SMike Galbraith } 1604ef557180SMike Galbraith 1605d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 16061da177e4SLinus Torvalds struct work_struct *work) 16071da177e4SLinus Torvalds { 1608112202d9STejun Heo struct pool_workqueue *pwq; 1609c9178087STejun Heo struct worker_pool *last_pool; 16101e19ffc6STejun Heo struct list_head *worklist; 16118a2e8e5dSTejun Heo unsigned int work_flags; 1612b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 16138930cabaSTejun Heo 16148930cabaSTejun Heo /* 16158930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 16168930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 16178930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 16188930cabaSTejun Heo * happen with IRQ disabled. 16198930cabaSTejun Heo */ 16208e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 16211da177e4SLinus Torvalds 16221e19ffc6STejun Heo 162333e3f0a3SRichard Clark /* 162433e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 162533e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 162633e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 162733e3f0a3SRichard Clark */ 162833e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 162933e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1630e41e704bSTejun Heo return; 163124acfb71SThomas Gleixner rcu_read_lock(); 16329e8cd2f5STejun Heo retry: 1633aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1634aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1635df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1636ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1637df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1638aa202f1fSHillf Danton } else { 1639aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1640aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1641aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1642aa202f1fSHillf Danton } 1643f3421797STejun Heo 164418aa9effSTejun Heo /* 1645c9178087STejun Heo * If @work was previously on a different pool, it might still be 1646c9178087STejun Heo * running there, in which case the work needs to be queued on that 1647c9178087STejun Heo * pool to guarantee non-reentrancy. 164818aa9effSTejun Heo */ 1649c9e7cf27STejun Heo last_pool = get_work_pool(work); 1650112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 165118aa9effSTejun Heo struct worker *worker; 165218aa9effSTejun Heo 1653a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 165418aa9effSTejun Heo 1655c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 165618aa9effSTejun Heo 1657112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1658c9178087STejun Heo pwq = worker->current_pwq; 16598594fadeSLai Jiangshan } else { 166018aa9effSTejun Heo /* meh... not running there, queue here */ 1661a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1662a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 166318aa9effSTejun Heo } 16648930cabaSTejun Heo } else { 1665a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 16668930cabaSTejun Heo } 1667502ca9d8STejun Heo 16689e8cd2f5STejun Heo /* 16699e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 16709e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 16719e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1672df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1673df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 16749e8cd2f5STejun Heo * make forward-progress. 16759e8cd2f5STejun Heo */ 16769e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 16779e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1678a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 16799e8cd2f5STejun Heo cpu_relax(); 16809e8cd2f5STejun Heo goto retry; 16819e8cd2f5STejun Heo } 16829e8cd2f5STejun Heo /* oops */ 16839e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 16849e8cd2f5STejun Heo wq->name, cpu); 16859e8cd2f5STejun Heo } 16869e8cd2f5STejun Heo 1687112202d9STejun Heo /* pwq determined, queue */ 1688112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1689502ca9d8STejun Heo 169024acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 169124acfb71SThomas Gleixner goto out; 16921e19ffc6STejun Heo 1693112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1694112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 16951e19ffc6STejun Heo 1696112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1697cdadf009STejun Heo trace_workqueue_activate_work(work); 1698112202d9STejun Heo pwq->nr_active++; 1699112202d9STejun Heo worklist = &pwq->pool->worklist; 170082607adcSTejun Heo if (list_empty(worklist)) 170182607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 17028a2e8e5dSTejun Heo } else { 1703f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1704f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 17058a2e8e5dSTejun Heo } 17061e19ffc6STejun Heo 17070687c66bSZqiang debug_work_activate(work); 1708112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 17091e19ffc6STejun Heo 171024acfb71SThomas Gleixner out: 1711a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 171224acfb71SThomas Gleixner rcu_read_unlock(); 17131da177e4SLinus Torvalds } 17141da177e4SLinus Torvalds 17150fcb78c2SRolf Eike Beer /** 1716c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1717c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1718c1a220e7SZhang Rui * @wq: workqueue to use 1719c1a220e7SZhang Rui * @work: work to queue 1720c1a220e7SZhang Rui * 1721c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1722443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1723443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1724854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1725854f5cc5SPaul E. McKenney * online will get a splat. 1726d185af30SYacine Belkadi * 1727d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1728c1a220e7SZhang Rui */ 1729d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1730d4283e93STejun Heo struct work_struct *work) 1731c1a220e7SZhang Rui { 1732d4283e93STejun Heo bool ret = false; 17338930cabaSTejun Heo unsigned long flags; 17348930cabaSTejun Heo 17358930cabaSTejun Heo local_irq_save(flags); 1736c1a220e7SZhang Rui 173722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17384690c4abSTejun Heo __queue_work(cpu, wq, work); 1739d4283e93STejun Heo ret = true; 1740c1a220e7SZhang Rui } 17418930cabaSTejun Heo 17428930cabaSTejun Heo local_irq_restore(flags); 1743c1a220e7SZhang Rui return ret; 1744c1a220e7SZhang Rui } 1745ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1746c1a220e7SZhang Rui 17478204e0c1SAlexander Duyck /** 17488204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 17498204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 17508204e0c1SAlexander Duyck * 17518204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 17528204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 17538204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 17548204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 17558204e0c1SAlexander Duyck */ 17568204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 17578204e0c1SAlexander Duyck { 17588204e0c1SAlexander Duyck int cpu; 17598204e0c1SAlexander Duyck 17608204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 17618204e0c1SAlexander Duyck if (!wq_numa_enabled) 17628204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17638204e0c1SAlexander Duyck 17648204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 17658204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 17668204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17678204e0c1SAlexander Duyck 17688204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 17698204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 17708204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 17718204e0c1SAlexander Duyck return cpu; 17728204e0c1SAlexander Duyck 17738204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 17748204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 17758204e0c1SAlexander Duyck 17768204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 17778204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 17788204e0c1SAlexander Duyck } 17798204e0c1SAlexander Duyck 17808204e0c1SAlexander Duyck /** 17818204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 17828204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 17838204e0c1SAlexander Duyck * @wq: workqueue to use 17848204e0c1SAlexander Duyck * @work: work to queue 17858204e0c1SAlexander Duyck * 17868204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 17878204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 17888204e0c1SAlexander Duyck * NUMA node. 17898204e0c1SAlexander Duyck * 17908204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 17918204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 17928204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 17938204e0c1SAlexander Duyck * 17948204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 17958204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 17968204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 17978204e0c1SAlexander Duyck * 17988204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 17998204e0c1SAlexander Duyck */ 18008204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18018204e0c1SAlexander Duyck struct work_struct *work) 18028204e0c1SAlexander Duyck { 18038204e0c1SAlexander Duyck unsigned long flags; 18048204e0c1SAlexander Duyck bool ret = false; 18058204e0c1SAlexander Duyck 18068204e0c1SAlexander Duyck /* 18078204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 18088204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 18098204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 18108204e0c1SAlexander Duyck * 18118204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 18128204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 18138204e0c1SAlexander Duyck * some round robin type logic. 18148204e0c1SAlexander Duyck */ 18158204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 18168204e0c1SAlexander Duyck 18178204e0c1SAlexander Duyck local_irq_save(flags); 18188204e0c1SAlexander Duyck 18198204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18208204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 18218204e0c1SAlexander Duyck 18228204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 18238204e0c1SAlexander Duyck ret = true; 18248204e0c1SAlexander Duyck } 18258204e0c1SAlexander Duyck 18268204e0c1SAlexander Duyck local_irq_restore(flags); 18278204e0c1SAlexander Duyck return ret; 18288204e0c1SAlexander Duyck } 18298204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 18308204e0c1SAlexander Duyck 18318c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 18321da177e4SLinus Torvalds { 18338c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 18341da177e4SLinus Torvalds 1835e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 183660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 18371da177e4SLinus Torvalds } 18381438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 18391da177e4SLinus Torvalds 18407beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 184152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18421da177e4SLinus Torvalds { 18437beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 18447beb2edfSTejun Heo struct work_struct *work = &dwork->work; 18451da177e4SLinus Torvalds 1846637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 18474b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1848fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1849fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 18507beb2edfSTejun Heo 18518852aac2STejun Heo /* 18528852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 18538852aac2STejun Heo * both optimization and correctness. The earliest @timer can 18548852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 18558852aac2STejun Heo * on that there's no such delay when @delay is 0. 18568852aac2STejun Heo */ 18578852aac2STejun Heo if (!delay) { 18588852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 18598852aac2STejun Heo return; 18608852aac2STejun Heo } 18618852aac2STejun Heo 186260c057bcSLai Jiangshan dwork->wq = wq; 18631265057fSTejun Heo dwork->cpu = cpu; 18647beb2edfSTejun Heo timer->expires = jiffies + delay; 18657beb2edfSTejun Heo 1866041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 18677beb2edfSTejun Heo add_timer_on(timer, cpu); 1868041bd12eSTejun Heo else 1869041bd12eSTejun Heo add_timer(timer); 18707beb2edfSTejun Heo } 18711da177e4SLinus Torvalds 18720fcb78c2SRolf Eike Beer /** 18730fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 18740fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 18750fcb78c2SRolf Eike Beer * @wq: workqueue to use 1876af9997e4SRandy Dunlap * @dwork: work to queue 18770fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 18780fcb78c2SRolf Eike Beer * 1879d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1880715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1881715f1300STejun Heo * execution. 18820fcb78c2SRolf Eike Beer */ 1883d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 188452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18857a6bc1cdSVenkatesh Pallipadi { 188652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1887d4283e93STejun Heo bool ret = false; 18888930cabaSTejun Heo unsigned long flags; 18898930cabaSTejun Heo 18908930cabaSTejun Heo /* read the comment in __queue_work() */ 18918930cabaSTejun Heo local_irq_save(flags); 18927a6bc1cdSVenkatesh Pallipadi 189322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18947beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1895d4283e93STejun Heo ret = true; 18967a6bc1cdSVenkatesh Pallipadi } 18978930cabaSTejun Heo 18988930cabaSTejun Heo local_irq_restore(flags); 18997a6bc1cdSVenkatesh Pallipadi return ret; 19007a6bc1cdSVenkatesh Pallipadi } 1901ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19021da177e4SLinus Torvalds 1903c8e55f36STejun Heo /** 19048376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19058376fe22STejun Heo * @cpu: CPU number to execute work on 19068376fe22STejun Heo * @wq: workqueue to use 19078376fe22STejun Heo * @dwork: work to queue 19088376fe22STejun Heo * @delay: number of jiffies to wait before queueing 19098376fe22STejun Heo * 19108376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 19118376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 19128376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 19138376fe22STejun Heo * current state. 19148376fe22STejun Heo * 1915d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 19168376fe22STejun Heo * pending and its timer was modified. 19178376fe22STejun Heo * 1918e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 19198376fe22STejun Heo * See try_to_grab_pending() for details. 19208376fe22STejun Heo */ 19218376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 19228376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 19238376fe22STejun Heo { 19248376fe22STejun Heo unsigned long flags; 19258376fe22STejun Heo int ret; 19268376fe22STejun Heo 19278376fe22STejun Heo do { 19288376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 19298376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 19308376fe22STejun Heo 19318376fe22STejun Heo if (likely(ret >= 0)) { 19328376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 19338376fe22STejun Heo local_irq_restore(flags); 19348376fe22STejun Heo } 19358376fe22STejun Heo 19368376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 19378376fe22STejun Heo return ret; 19388376fe22STejun Heo } 19398376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 19408376fe22STejun Heo 194105f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 194205f0fe6bSTejun Heo { 194305f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 194405f0fe6bSTejun Heo 194505f0fe6bSTejun Heo /* read the comment in __queue_work() */ 194605f0fe6bSTejun Heo local_irq_disable(); 194705f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 194805f0fe6bSTejun Heo local_irq_enable(); 194905f0fe6bSTejun Heo } 195005f0fe6bSTejun Heo 195105f0fe6bSTejun Heo /** 195205f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 195305f0fe6bSTejun Heo * @wq: workqueue to use 195405f0fe6bSTejun Heo * @rwork: work to queue 195505f0fe6bSTejun Heo * 195605f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 195705f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1958bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 195905f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 196005f0fe6bSTejun Heo */ 196105f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 196205f0fe6bSTejun Heo { 196305f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 196405f0fe6bSTejun Heo 196505f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 196605f0fe6bSTejun Heo rwork->wq = wq; 1967a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 196805f0fe6bSTejun Heo return true; 196905f0fe6bSTejun Heo } 197005f0fe6bSTejun Heo 197105f0fe6bSTejun Heo return false; 197205f0fe6bSTejun Heo } 197305f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 197405f0fe6bSTejun Heo 19758376fe22STejun Heo /** 1976c8e55f36STejun Heo * worker_enter_idle - enter idle state 1977c8e55f36STejun Heo * @worker: worker which is entering idle state 1978c8e55f36STejun Heo * 1979c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1980c8e55f36STejun Heo * necessary. 1981c8e55f36STejun Heo * 1982c8e55f36STejun Heo * LOCKING: 1983a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1984c8e55f36STejun Heo */ 1985c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 19861da177e4SLinus Torvalds { 1987bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1988c8e55f36STejun Heo 19896183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 19906183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 19916183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 19926183c009STejun Heo return; 1993c8e55f36STejun Heo 1994051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1995cb444766STejun Heo worker->flags |= WORKER_IDLE; 1996bd7bdd43STejun Heo pool->nr_idle++; 1997e22bee78STejun Heo worker->last_active = jiffies; 1998c8e55f36STejun Heo 1999c8e55f36STejun Heo /* idle_list is LIFO */ 2000bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 2001db7bccf4STejun Heo 200263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 2003628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 2004cb444766STejun Heo 2005989442d7SLai Jiangshan /* Sanity check nr_running. */ 2006bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 2007c8e55f36STejun Heo } 2008c8e55f36STejun Heo 2009c8e55f36STejun Heo /** 2010c8e55f36STejun Heo * worker_leave_idle - leave idle state 2011c8e55f36STejun Heo * @worker: worker which is leaving idle state 2012c8e55f36STejun Heo * 2013c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 2014c8e55f36STejun Heo * 2015c8e55f36STejun Heo * LOCKING: 2016a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2017c8e55f36STejun Heo */ 2018c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 2019c8e55f36STejun Heo { 2020bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2021c8e55f36STejun Heo 20226183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 20236183c009STejun Heo return; 2024d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 2025bd7bdd43STejun Heo pool->nr_idle--; 2026c8e55f36STejun Heo list_del_init(&worker->entry); 2027c8e55f36STejun Heo } 2028c8e55f36STejun Heo 2029f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 2030c34056a3STejun Heo { 2031c34056a3STejun Heo struct worker *worker; 2032c34056a3STejun Heo 2033f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2034c8e55f36STejun Heo if (worker) { 2035c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 2036affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 2037da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2038e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2039e22bee78STejun Heo worker->flags = WORKER_PREP; 2040c8e55f36STejun Heo } 2041c34056a3STejun Heo return worker; 2042c34056a3STejun Heo } 2043c34056a3STejun Heo 2044c34056a3STejun Heo /** 20454736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20464736cbf7SLai Jiangshan * @worker: worker to be attached 20474736cbf7SLai Jiangshan * @pool: the target pool 20484736cbf7SLai Jiangshan * 20494736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20504736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20514736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20524736cbf7SLai Jiangshan */ 20534736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 20544736cbf7SLai Jiangshan struct worker_pool *pool) 20554736cbf7SLai Jiangshan { 20561258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 20574736cbf7SLai Jiangshan 20584736cbf7SLai Jiangshan /* 20591258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 20601258fae7STejun Heo * stable across this function. See the comments above the flag 20611258fae7STejun Heo * definition for details. 20624736cbf7SLai Jiangshan */ 20634736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 20644736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 20655c25b5ffSPeter Zijlstra else 20665c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 20674736cbf7SLai Jiangshan 2068640f17c8SPeter Zijlstra if (worker->rescue_wq) 2069640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 2070640f17c8SPeter Zijlstra 20714736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2072a2d812a2STejun Heo worker->pool = pool; 20734736cbf7SLai Jiangshan 20741258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 20754736cbf7SLai Jiangshan } 20764736cbf7SLai Jiangshan 20774736cbf7SLai Jiangshan /** 207860f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 207960f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 208060f5a4bcSLai Jiangshan * 20814736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 20824736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 20834736cbf7SLai Jiangshan * other reference to the pool. 208460f5a4bcSLai Jiangshan */ 2085a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 208660f5a4bcSLai Jiangshan { 2087a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 208860f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 208960f5a4bcSLai Jiangshan 20901258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2091a2d812a2STejun Heo 20925c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2093da028469SLai Jiangshan list_del(&worker->node); 2094a2d812a2STejun Heo worker->pool = NULL; 2095a2d812a2STejun Heo 2096e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 209760f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 20981258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 209960f5a4bcSLai Jiangshan 2100b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2101b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2102b62c0751SLai Jiangshan 210360f5a4bcSLai Jiangshan if (detach_completion) 210460f5a4bcSLai Jiangshan complete(detach_completion); 210560f5a4bcSLai Jiangshan } 210660f5a4bcSLai Jiangshan 210760f5a4bcSLai Jiangshan /** 2108c34056a3STejun Heo * create_worker - create a new workqueue worker 210963d95a91STejun Heo * @pool: pool the new worker will belong to 2110c34056a3STejun Heo * 2111051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2112c34056a3STejun Heo * 2113c34056a3STejun Heo * CONTEXT: 2114c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2115c34056a3STejun Heo * 2116d185af30SYacine Belkadi * Return: 2117c34056a3STejun Heo * Pointer to the newly created worker. 2118c34056a3STejun Heo */ 2119bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2120c34056a3STejun Heo { 2121e441b56fSZhen Lei struct worker *worker; 2122e441b56fSZhen Lei int id; 2123e3c916a4STejun Heo char id_buf[16]; 2124c34056a3STejun Heo 21257cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2126e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 21273f0ea0b8SPetr Mladek if (id < 0) { 21283f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 21293f0ea0b8SPetr Mladek ERR_PTR(id)); 2130e441b56fSZhen Lei return NULL; 21313f0ea0b8SPetr Mladek } 2132c34056a3STejun Heo 2133f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 21343f0ea0b8SPetr Mladek if (!worker) { 21353f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2136c34056a3STejun Heo goto fail; 21373f0ea0b8SPetr Mladek } 2138c34056a3STejun Heo 2139c34056a3STejun Heo worker->id = id; 2140c34056a3STejun Heo 214129c91e99STejun Heo if (pool->cpu >= 0) 2142e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2143e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2144f3421797STejun Heo else 2145e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2146e3c916a4STejun Heo 2147f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2148e3c916a4STejun Heo "kworker/%s", id_buf); 21493f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 215060f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 215160f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 215260f54038SPetr Mladek id_buf); 215360f54038SPetr Mladek } else { 21543f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 21553f0ea0b8SPetr Mladek worker->task); 215660f54038SPetr Mladek } 2157c34056a3STejun Heo goto fail; 21583f0ea0b8SPetr Mladek } 2159c34056a3STejun Heo 216091151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 216125834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 216291151228SOleg Nesterov 2163da028469SLai Jiangshan /* successful, attach the worker to the pool */ 21644736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2165822d8405STejun Heo 2166051e1850SLai Jiangshan /* start the newly created worker */ 2167a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2168051e1850SLai Jiangshan worker->pool->nr_workers++; 2169051e1850SLai Jiangshan worker_enter_idle(worker); 2170051e1850SLai Jiangshan wake_up_process(worker->task); 2171a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2172051e1850SLai Jiangshan 2173c34056a3STejun Heo return worker; 2174822d8405STejun Heo 2175c34056a3STejun Heo fail: 2176e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2177c34056a3STejun Heo kfree(worker); 2178c34056a3STejun Heo return NULL; 2179c34056a3STejun Heo } 2180c34056a3STejun Heo 2181793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2182793777bcSValentin Schneider { 2183793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2184793777bcSValentin Schneider 2185793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2186793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2187793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2188793777bcSValentin Schneider else 2189793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2190793777bcSValentin Schneider } 2191793777bcSValentin Schneider 2192e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2193e02b9312SValentin Schneider { 2194e02b9312SValentin Schneider struct worker *worker, *tmp; 2195e02b9312SValentin Schneider 2196e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2197e02b9312SValentin Schneider list_del_init(&worker->entry); 2198e02b9312SValentin Schneider unbind_worker(worker); 2199e02b9312SValentin Schneider /* 2200e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2201e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2202e02b9312SValentin Schneider * wouldn't have gotten here. 2203c34056a3STejun Heo * 2204e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2205e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2206e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2207e02b9312SValentin Schneider * outside of pool->lock. 2208e02b9312SValentin Schneider */ 2209e02b9312SValentin Schneider wake_up_process(worker->task); 2210e02b9312SValentin Schneider } 2211e02b9312SValentin Schneider } 2212e02b9312SValentin Schneider 2213e02b9312SValentin Schneider /** 2214e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2215e02b9312SValentin Schneider * @worker: worker to be destroyed 2216e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2217e02b9312SValentin Schneider * 2218e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2219e02b9312SValentin Schneider * should be idle. 2220c8e55f36STejun Heo * 2221c8e55f36STejun Heo * CONTEXT: 2222a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2223c34056a3STejun Heo */ 2224e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2225c34056a3STejun Heo { 2226bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2227c34056a3STejun Heo 2228cd549687STejun Heo lockdep_assert_held(&pool->lock); 2229e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2230cd549687STejun Heo 2231c34056a3STejun Heo /* sanity check frenzy */ 22326183c009STejun Heo if (WARN_ON(worker->current_work) || 223373eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 223473eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 22356183c009STejun Heo return; 2236c34056a3STejun Heo 2237bd7bdd43STejun Heo pool->nr_workers--; 2238bd7bdd43STejun Heo pool->nr_idle--; 2239c8e55f36STejun Heo 2240cb444766STejun Heo worker->flags |= WORKER_DIE; 2241e02b9312SValentin Schneider 2242e02b9312SValentin Schneider list_move(&worker->entry, list); 2243e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2244c34056a3STejun Heo } 2245c34056a3STejun Heo 22463f959aa3SValentin Schneider /** 22473f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 22483f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 22493f959aa3SValentin Schneider * 22503f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 22513f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 22523f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 22533f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 22543f959aa3SValentin Schneider * it expire and re-evaluate things from there. 22553f959aa3SValentin Schneider */ 225632a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2257e22bee78STejun Heo { 225832a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 22593f959aa3SValentin Schneider bool do_cull = false; 22603f959aa3SValentin Schneider 22613f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 22623f959aa3SValentin Schneider return; 22633f959aa3SValentin Schneider 22643f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 22653f959aa3SValentin Schneider 22663f959aa3SValentin Schneider if (too_many_workers(pool)) { 22673f959aa3SValentin Schneider struct worker *worker; 22683f959aa3SValentin Schneider unsigned long expires; 22693f959aa3SValentin Schneider 22703f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 22713f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 22723f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 22733f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 22743f959aa3SValentin Schneider 22753f959aa3SValentin Schneider if (!do_cull) 22763f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 22773f959aa3SValentin Schneider } 22783f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 22793f959aa3SValentin Schneider 22803f959aa3SValentin Schneider if (do_cull) 22813f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 22823f959aa3SValentin Schneider } 22833f959aa3SValentin Schneider 22843f959aa3SValentin Schneider /** 22853f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 22863f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 22873f959aa3SValentin Schneider * 22883f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 22893f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2290e02b9312SValentin Schneider * 2291e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2292e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2293e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 22943f959aa3SValentin Schneider */ 22953f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 22963f959aa3SValentin Schneider { 22973f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2298e02b9312SValentin Schneider struct list_head cull_list; 2299e22bee78STejun Heo 2300e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 2301e02b9312SValentin Schneider /* 2302e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2303e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2304e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2305e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2306e02b9312SValentin Schneider */ 2307e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2308a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2309e22bee78STejun Heo 23103347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2311e22bee78STejun Heo struct worker *worker; 2312e22bee78STejun Heo unsigned long expires; 2313e22bee78STejun Heo 231463d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2315e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2316e22bee78STejun Heo 23173347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 231863d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 23193347fc9fSLai Jiangshan break; 2320e22bee78STejun Heo } 23213347fc9fSLai Jiangshan 2322e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2323e22bee78STejun Heo } 2324e22bee78STejun Heo 2325a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2326e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2327e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2328e22bee78STejun Heo } 2329e22bee78STejun Heo 2330493a1724STejun Heo static void send_mayday(struct work_struct *work) 2331e22bee78STejun Heo { 2332112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2333112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2334493a1724STejun Heo 23352e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2336e22bee78STejun Heo 2337493008a8STejun Heo if (!wq->rescuer) 2338493a1724STejun Heo return; 2339e22bee78STejun Heo 2340e22bee78STejun Heo /* mayday mayday mayday */ 2341493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 234277668c8bSLai Jiangshan /* 234377668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 234477668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 234577668c8bSLai Jiangshan * rescuer is done with it. 234677668c8bSLai Jiangshan */ 234777668c8bSLai Jiangshan get_pwq(pwq); 2348493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2349e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2350725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2351493a1724STejun Heo } 2352e22bee78STejun Heo } 2353e22bee78STejun Heo 235432a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2355e22bee78STejun Heo { 235632a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2357e22bee78STejun Heo struct work_struct *work; 2358e22bee78STejun Heo 2359a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2360a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2361e22bee78STejun Heo 236263d95a91STejun Heo if (need_to_create_worker(pool)) { 2363e22bee78STejun Heo /* 2364e22bee78STejun Heo * We've been trying to create a new worker but 2365e22bee78STejun Heo * haven't been successful. We might be hitting an 2366e22bee78STejun Heo * allocation deadlock. Send distress signals to 2367e22bee78STejun Heo * rescuers. 2368e22bee78STejun Heo */ 236963d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2370e22bee78STejun Heo send_mayday(work); 2371e22bee78STejun Heo } 2372e22bee78STejun Heo 2373a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2374a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2375e22bee78STejun Heo 237663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2377e22bee78STejun Heo } 2378e22bee78STejun Heo 2379e22bee78STejun Heo /** 2380e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 238163d95a91STejun Heo * @pool: pool to create a new worker for 2382e22bee78STejun Heo * 238363d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2384e22bee78STejun Heo * have at least one idle worker on return from this function. If 2385e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 238663d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2387e22bee78STejun Heo * possible allocation deadlock. 2388e22bee78STejun Heo * 2389c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2390c5aa87bbSTejun Heo * may_start_working() %true. 2391e22bee78STejun Heo * 2392e22bee78STejun Heo * LOCKING: 2393a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2394e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2395e22bee78STejun Heo * manager. 2396e22bee78STejun Heo */ 239729187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2398d565ed63STejun Heo __releases(&pool->lock) 2399d565ed63STejun Heo __acquires(&pool->lock) 2400e22bee78STejun Heo { 2401e22bee78STejun Heo restart: 2402a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24039f9c2364STejun Heo 2404e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 240563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2406e22bee78STejun Heo 2407e22bee78STejun Heo while (true) { 2408051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2409e22bee78STejun Heo break; 2410e22bee78STejun Heo 2411e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 24129f9c2364STejun Heo 241363d95a91STejun Heo if (!need_to_create_worker(pool)) 2414e22bee78STejun Heo break; 2415e22bee78STejun Heo } 2416e22bee78STejun Heo 241763d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2418a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2419051e1850SLai Jiangshan /* 2420051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2421051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2422051e1850SLai Jiangshan * already become busy. 2423051e1850SLai Jiangshan */ 242463d95a91STejun Heo if (need_to_create_worker(pool)) 2425e22bee78STejun Heo goto restart; 2426e22bee78STejun Heo } 2427e22bee78STejun Heo 2428e22bee78STejun Heo /** 2429e22bee78STejun Heo * manage_workers - manage worker pool 2430e22bee78STejun Heo * @worker: self 2431e22bee78STejun Heo * 2432706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2433e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2434706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2435e22bee78STejun Heo * 2436e22bee78STejun Heo * The caller can safely start processing works on false return. On 2437e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2438e22bee78STejun Heo * and may_start_working() is true. 2439e22bee78STejun Heo * 2440e22bee78STejun Heo * CONTEXT: 2441a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2442e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2443e22bee78STejun Heo * 2444d185af30SYacine Belkadi * Return: 244529187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 244629187a9eSTejun Heo * start processing works, %true if management function was performed and 244729187a9eSTejun Heo * the conditions that the caller verified before calling the function may 244829187a9eSTejun Heo * no longer be true. 2449e22bee78STejun Heo */ 2450e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2451e22bee78STejun Heo { 245263d95a91STejun Heo struct worker_pool *pool = worker->pool; 2453e22bee78STejun Heo 2454692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 245529187a9eSTejun Heo return false; 2456692b4825STejun Heo 2457692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 24582607d7a6STejun Heo pool->manager = worker; 2459e22bee78STejun Heo 246029187a9eSTejun Heo maybe_create_worker(pool); 2461e22bee78STejun Heo 24622607d7a6STejun Heo pool->manager = NULL; 2463692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2464d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 246529187a9eSTejun Heo return true; 2466e22bee78STejun Heo } 2467e22bee78STejun Heo 2468a62428c0STejun Heo /** 2469a62428c0STejun Heo * process_one_work - process single work 2470c34056a3STejun Heo * @worker: self 2471a62428c0STejun Heo * @work: work to process 2472a62428c0STejun Heo * 2473a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2474a62428c0STejun Heo * process a single work including synchronization against and 2475a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2476a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2477a62428c0STejun Heo * call this function to process a work. 2478a62428c0STejun Heo * 2479a62428c0STejun Heo * CONTEXT: 2480a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2481a62428c0STejun Heo */ 2482c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2483d565ed63STejun Heo __releases(&pool->lock) 2484d565ed63STejun Heo __acquires(&pool->lock) 24851da177e4SLinus Torvalds { 2486112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2487bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2488c4560c2cSLai Jiangshan unsigned long work_data; 24897e11629dSTejun Heo struct worker *collision; 24904e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 24914e6045f1SJohannes Berg /* 2492a62428c0STejun Heo * It is permissible to free the struct work_struct from 2493a62428c0STejun Heo * inside the function that is called from it, this we need to 2494a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2495a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2496a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 24974e6045f1SJohannes Berg */ 24984d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 24994d82a1deSPeter Zijlstra 25004d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 25014e6045f1SJohannes Berg #endif 2502807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 250385327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2504ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 250525511a47STejun Heo 25067e11629dSTejun Heo /* 25077e11629dSTejun Heo * A single work shouldn't be executed concurrently by 25087e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 25097e11629dSTejun Heo * already processing the work. If so, defer the work to the 25107e11629dSTejun Heo * currently executing one. 25117e11629dSTejun Heo */ 2512c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 25137e11629dSTejun Heo if (unlikely(collision)) { 25147e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 25157e11629dSTejun Heo return; 25167e11629dSTejun Heo } 25171da177e4SLinus Torvalds 25188930cabaSTejun Heo /* claim and dequeue */ 25191da177e4SLinus Torvalds debug_work_deactivate(work); 2520c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2521c34056a3STejun Heo worker->current_work = work; 2522a2c1c57bSTejun Heo worker->current_func = work->func; 2523112202d9STejun Heo worker->current_pwq = pwq; 2524616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2525c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2526d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 25277a22ad75STejun Heo 25288bf89593STejun Heo /* 25298bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 25308bf89593STejun Heo * overridden through set_worker_desc(). 25318bf89593STejun Heo */ 25328bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 25338bf89593STejun Heo 2534a62428c0STejun Heo list_del_init(&work->entry); 2535a62428c0STejun Heo 2536649027d7STejun Heo /* 2537228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2538228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2539228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2540228f1d00SLai Jiangshan * execution of the pending work items. 2541fb0e7bebSTejun Heo */ 2542616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2543228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2544fb0e7bebSTejun Heo 2545974271c4STejun Heo /* 2546a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2547a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2548a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2549a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2550228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2551974271c4STejun Heo */ 2552a489a03eSLai Jiangshan if (need_more_worker(pool)) 255363d95a91STejun Heo wake_up_worker(pool); 2554974271c4STejun Heo 25558930cabaSTejun Heo /* 25567c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2557d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 255823657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 255923657bb1STejun Heo * disabled. 25608930cabaSTejun Heo */ 25617c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 25621da177e4SLinus Torvalds 2563a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2564365970a1SDavid Howells 2565a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 25663295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2567e6f3faa7SPeter Zijlstra /* 2568f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2569f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2570e6f3faa7SPeter Zijlstra * 2571e6f3faa7SPeter Zijlstra * However, that would result in: 2572e6f3faa7SPeter Zijlstra * 2573e6f3faa7SPeter Zijlstra * A(W1) 2574e6f3faa7SPeter Zijlstra * WFC(C) 2575e6f3faa7SPeter Zijlstra * A(W1) 2576e6f3faa7SPeter Zijlstra * C(C) 2577e6f3faa7SPeter Zijlstra * 2578e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2579e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2580e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2581f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2582e6f3faa7SPeter Zijlstra * these locks. 2583e6f3faa7SPeter Zijlstra * 2584e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2585e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2586e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2587e6f3faa7SPeter Zijlstra */ 2588f52be570SPeter Zijlstra lockdep_invariant_state(true); 2589725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2590e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2591a2c1c57bSTejun Heo worker->current_func(work); 2592e36c886aSArjan van de Ven /* 2593e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2594e36c886aSArjan van de Ven * point will only record its address. 2595e36c886aSArjan van de Ven */ 25961c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2597725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 25983295f0efSIngo Molnar lock_map_release(&lockdep_map); 2599112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 26001da177e4SLinus Torvalds 2601d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2602044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2603d75f773cSSakari Ailus " last function: %ps\n", 2604a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2605a2c1c57bSTejun Heo worker->current_func); 2606d5abe669SPeter Zijlstra debug_show_held_locks(current); 2607d5abe669SPeter Zijlstra dump_stack(); 2608d5abe669SPeter Zijlstra } 2609d5abe669SPeter Zijlstra 2610b22ce278STejun Heo /* 2611025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2612b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2613b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2614b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2615789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2616789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2617b22ce278STejun Heo */ 2618a7e6425eSPaul E. McKenney cond_resched(); 2619b22ce278STejun Heo 2620a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2621a62428c0STejun Heo 2622616db877STejun Heo /* 2623616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2624616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2625616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2626616db877STejun Heo */ 2627fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2628fb0e7bebSTejun Heo 26291b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 26301b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 26311b69ac6bSJohannes Weiner 2632a62428c0STejun Heo /* we're done with it, release */ 263342f8570fSSasha Levin hash_del(&worker->hentry); 2634c34056a3STejun Heo worker->current_work = NULL; 2635a2c1c57bSTejun Heo worker->current_func = NULL; 2636112202d9STejun Heo worker->current_pwq = NULL; 2637d812796eSLai Jiangshan worker->current_color = INT_MAX; 2638c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26391da177e4SLinus Torvalds } 26401da177e4SLinus Torvalds 2641affee4b2STejun Heo /** 2642affee4b2STejun Heo * process_scheduled_works - process scheduled works 2643affee4b2STejun Heo * @worker: self 2644affee4b2STejun Heo * 2645affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2646affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2647affee4b2STejun Heo * fetches a work from the top and executes it. 2648affee4b2STejun Heo * 2649affee4b2STejun Heo * CONTEXT: 2650a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2651affee4b2STejun Heo * multiple times. 2652affee4b2STejun Heo */ 2653affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26541da177e4SLinus Torvalds { 2655affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2656affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2657a62428c0STejun Heo struct work_struct, entry); 2658c34056a3STejun Heo process_one_work(worker, work); 2659a62428c0STejun Heo } 26601da177e4SLinus Torvalds } 26611da177e4SLinus Torvalds 2662197f6accSTejun Heo static void set_pf_worker(bool val) 2663197f6accSTejun Heo { 2664197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2665197f6accSTejun Heo if (val) 2666197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2667197f6accSTejun Heo else 2668197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2669197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2670197f6accSTejun Heo } 2671197f6accSTejun Heo 26724690c4abSTejun Heo /** 26734690c4abSTejun Heo * worker_thread - the worker thread function 2674c34056a3STejun Heo * @__worker: self 26754690c4abSTejun Heo * 2676c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2677c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2678c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2679c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2680c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2681d185af30SYacine Belkadi * 2682d185af30SYacine Belkadi * Return: 0 26834690c4abSTejun Heo */ 2684c34056a3STejun Heo static int worker_thread(void *__worker) 26851da177e4SLinus Torvalds { 2686c34056a3STejun Heo struct worker *worker = __worker; 2687bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 26881da177e4SLinus Torvalds 2689e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2690197f6accSTejun Heo set_pf_worker(true); 2691c8e55f36STejun Heo woke_up: 2692a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2693affee4b2STejun Heo 2694a9ab775bSTejun Heo /* am I supposed to die? */ 2695a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2696a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2697197f6accSTejun Heo set_pf_worker(false); 269860f5a4bcSLai Jiangshan 269960f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2700e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2701a2d812a2STejun Heo worker_detach_from_pool(worker); 2702e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 270360f5a4bcSLai Jiangshan kfree(worker); 2704c8e55f36STejun Heo return 0; 2705c8e55f36STejun Heo } 2706c8e55f36STejun Heo 2707c8e55f36STejun Heo worker_leave_idle(worker); 2708db7bccf4STejun Heo recheck: 2709e22bee78STejun Heo /* no more worker necessary? */ 271063d95a91STejun Heo if (!need_more_worker(pool)) 2711e22bee78STejun Heo goto sleep; 2712e22bee78STejun Heo 2713e22bee78STejun Heo /* do we need to manage? */ 271463d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2715e22bee78STejun Heo goto recheck; 2716e22bee78STejun Heo 2717c8e55f36STejun Heo /* 2718c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2719c8e55f36STejun Heo * preparing to process a work or actually processing it. 2720c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2721c8e55f36STejun Heo */ 27226183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2723c8e55f36STejun Heo 2724e22bee78STejun Heo /* 2725a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2726a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2727a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2728a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2729a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2730e22bee78STejun Heo */ 2731a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2732e22bee78STejun Heo 2733e22bee78STejun Heo do { 2734affee4b2STejun Heo struct work_struct *work = 2735bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2736affee4b2STejun Heo struct work_struct, entry); 2737affee4b2STejun Heo 273882607adcSTejun Heo pool->watchdog_ts = jiffies; 273982607adcSTejun Heo 2740c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2741affee4b2STejun Heo /* optimization path, not strictly necessary */ 2742affee4b2STejun Heo process_one_work(worker, work); 2743affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2744affee4b2STejun Heo process_scheduled_works(worker); 2745affee4b2STejun Heo } else { 2746c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2747affee4b2STejun Heo process_scheduled_works(worker); 2748affee4b2STejun Heo } 274963d95a91STejun Heo } while (keep_working(pool)); 2750affee4b2STejun Heo 2751228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2752d313dd85STejun Heo sleep: 2753c8e55f36STejun Heo /* 2754d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2755d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2756d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2757d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2758d565ed63STejun Heo * event. 2759c8e55f36STejun Heo */ 2760c8e55f36STejun Heo worker_enter_idle(worker); 2761c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2762a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27631da177e4SLinus Torvalds schedule(); 2764c8e55f36STejun Heo goto woke_up; 27651da177e4SLinus Torvalds } 27661da177e4SLinus Torvalds 2767e22bee78STejun Heo /** 2768e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2769111c225aSTejun Heo * @__rescuer: self 2770e22bee78STejun Heo * 2771e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2772493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2773e22bee78STejun Heo * 2774706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2775e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2776e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2777e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2778e22bee78STejun Heo * the problem rescuer solves. 2779e22bee78STejun Heo * 2780706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2781706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2782e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2783e22bee78STejun Heo * 2784e22bee78STejun Heo * This should happen rarely. 2785d185af30SYacine Belkadi * 2786d185af30SYacine Belkadi * Return: 0 2787e22bee78STejun Heo */ 2788111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2789e22bee78STejun Heo { 2790111c225aSTejun Heo struct worker *rescuer = __rescuer; 2791111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2792e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 27934d595b86SLai Jiangshan bool should_stop; 2794e22bee78STejun Heo 2795e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2796111c225aSTejun Heo 2797111c225aSTejun Heo /* 2798111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2799111c225aSTejun Heo * doesn't participate in concurrency management. 2800111c225aSTejun Heo */ 2801197f6accSTejun Heo set_pf_worker(true); 2802e22bee78STejun Heo repeat: 2803c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 28041da177e4SLinus Torvalds 28054d595b86SLai Jiangshan /* 28064d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 28074d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 28084d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 28094d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 28104d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 28114d595b86SLai Jiangshan * list is always empty on exit. 28124d595b86SLai Jiangshan */ 28134d595b86SLai Jiangshan should_stop = kthread_should_stop(); 28141da177e4SLinus Torvalds 2815493a1724STejun Heo /* see whether any pwq is asking for help */ 2816a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2817493a1724STejun Heo 2818493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2819493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2820493a1724STejun Heo struct pool_workqueue, mayday_node); 2821112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2822e22bee78STejun Heo struct work_struct *work, *n; 282382607adcSTejun Heo bool first = true; 2824e22bee78STejun Heo 2825e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2826493a1724STejun Heo list_del_init(&pwq->mayday_node); 2827493a1724STejun Heo 2828a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2829e22bee78STejun Heo 283051697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 283151697d39SLai Jiangshan 2832a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2833e22bee78STejun Heo 2834e22bee78STejun Heo /* 2835e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2836e22bee78STejun Heo * process'em. 2837e22bee78STejun Heo */ 28380479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 283982607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 284082607adcSTejun Heo if (get_work_pwq(work) == pwq) { 284182607adcSTejun Heo if (first) 284282607adcSTejun Heo pool->watchdog_ts = jiffies; 2843e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2844725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 284582607adcSTejun Heo } 284682607adcSTejun Heo first = false; 284782607adcSTejun Heo } 2848e22bee78STejun Heo 2849008847f6SNeilBrown if (!list_empty(scheduled)) { 2850e22bee78STejun Heo process_scheduled_works(rescuer); 28517576958aSTejun Heo 28527576958aSTejun Heo /* 2853008847f6SNeilBrown * The above execution of rescued work items could 2854008847f6SNeilBrown * have created more to rescue through 2855f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2856008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2857008847f6SNeilBrown * that such back-to-back work items, which may be 2858008847f6SNeilBrown * being used to relieve memory pressure, don't 2859008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2860008847f6SNeilBrown */ 28614f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2862a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2863e66b39afSTejun Heo /* 2864e66b39afSTejun Heo * Queue iff we aren't racing destruction 2865e66b39afSTejun Heo * and somebody else hasn't queued it already. 2866e66b39afSTejun Heo */ 2867e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2868008847f6SNeilBrown get_pwq(pwq); 2869e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2870e66b39afSTejun Heo } 2871a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2872008847f6SNeilBrown } 2873008847f6SNeilBrown } 2874008847f6SNeilBrown 2875008847f6SNeilBrown /* 287677668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 287713b1d625SLai Jiangshan * go away while we're still attached to it. 287877668c8bSLai Jiangshan */ 287977668c8bSLai Jiangshan put_pwq(pwq); 288077668c8bSLai Jiangshan 288177668c8bSLai Jiangshan /* 2882d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 28837576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 28847576958aSTejun Heo * and stalling the execution. 28857576958aSTejun Heo */ 2886d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 288763d95a91STejun Heo wake_up_worker(pool); 28887576958aSTejun Heo 2889a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 289013b1d625SLai Jiangshan 2891a2d812a2STejun Heo worker_detach_from_pool(rescuer); 289213b1d625SLai Jiangshan 2893a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 28941da177e4SLinus Torvalds } 28951da177e4SLinus Torvalds 2896a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2897493a1724STejun Heo 28984d595b86SLai Jiangshan if (should_stop) { 28994d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2900197f6accSTejun Heo set_pf_worker(false); 29014d595b86SLai Jiangshan return 0; 29024d595b86SLai Jiangshan } 29034d595b86SLai Jiangshan 2904111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2905111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2906e22bee78STejun Heo schedule(); 2907e22bee78STejun Heo goto repeat; 29081da177e4SLinus Torvalds } 29091da177e4SLinus Torvalds 2910fca839c0STejun Heo /** 2911fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2912fca839c0STejun Heo * @target_wq: workqueue being flushed 2913fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2914fca839c0STejun Heo * 2915fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2916fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2917fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2918fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2919fca839c0STejun Heo * a deadlock. 2920fca839c0STejun Heo */ 2921fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2922fca839c0STejun Heo struct work_struct *target_work) 2923fca839c0STejun Heo { 2924fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2925fca839c0STejun Heo struct worker *worker; 2926fca839c0STejun Heo 2927fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2928fca839c0STejun Heo return; 2929fca839c0STejun Heo 2930fca839c0STejun Heo worker = current_wq_worker(); 2931fca839c0STejun Heo 2932fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2933d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2934fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 293523d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 293623d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2937d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2938fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2939fca839c0STejun Heo target_wq->name, target_func); 2940fca839c0STejun Heo } 2941fca839c0STejun Heo 2942fc2e4d70SOleg Nesterov struct wq_barrier { 2943fc2e4d70SOleg Nesterov struct work_struct work; 2944fc2e4d70SOleg Nesterov struct completion done; 29452607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2946fc2e4d70SOleg Nesterov }; 2947fc2e4d70SOleg Nesterov 2948fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2949fc2e4d70SOleg Nesterov { 2950fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2951fc2e4d70SOleg Nesterov complete(&barr->done); 2952fc2e4d70SOleg Nesterov } 2953fc2e4d70SOleg Nesterov 29544690c4abSTejun Heo /** 29554690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2956112202d9STejun Heo * @pwq: pwq to insert barrier into 29574690c4abSTejun Heo * @barr: wq_barrier to insert 2958affee4b2STejun Heo * @target: target work to attach @barr to 2959affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29604690c4abSTejun Heo * 2961affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2962affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2963affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2964affee4b2STejun Heo * cpu. 2965affee4b2STejun Heo * 2966affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2967affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2968affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2969affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2970affee4b2STejun Heo * after a work with LINKED flag set. 2971affee4b2STejun Heo * 2972affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2973112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 29744690c4abSTejun Heo * 29754690c4abSTejun Heo * CONTEXT: 2976a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 29774690c4abSTejun Heo */ 2978112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2979affee4b2STejun Heo struct wq_barrier *barr, 2980affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2981fc2e4d70SOleg Nesterov { 2982d812796eSLai Jiangshan unsigned int work_flags = 0; 2983d812796eSLai Jiangshan unsigned int work_color; 2984affee4b2STejun Heo struct list_head *head; 2985affee4b2STejun Heo 2986dc186ad7SThomas Gleixner /* 2987d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2988dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2989dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2990dc186ad7SThomas Gleixner * might deadlock. 2991dc186ad7SThomas Gleixner */ 2992ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 299322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 299452fa5bc5SBoqun Feng 2995fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2996fd1a5b04SByungchul Park 29972607d7a6STejun Heo barr->task = current; 299883c22520SOleg Nesterov 2999018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 3000018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 3001018f3a13SLai Jiangshan 3002affee4b2STejun Heo /* 3003affee4b2STejun Heo * If @target is currently being executed, schedule the 3004affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 3005affee4b2STejun Heo */ 3006d812796eSLai Jiangshan if (worker) { 3007affee4b2STejun Heo head = worker->scheduled.next; 3008d812796eSLai Jiangshan work_color = worker->current_color; 3009d812796eSLai Jiangshan } else { 3010affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 3011affee4b2STejun Heo 3012affee4b2STejun Heo head = target->entry.next; 3013affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 3014d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 3015d812796eSLai Jiangshan work_color = get_work_color(*bits); 3016affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3017affee4b2STejun Heo } 3018affee4b2STejun Heo 3019d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 3020d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 3021d812796eSLai Jiangshan 3022dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 3023d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 3024fc2e4d70SOleg Nesterov } 3025fc2e4d70SOleg Nesterov 302673f53c4aSTejun Heo /** 3027112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 302873f53c4aSTejun Heo * @wq: workqueue being flushed 302973f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 303073f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 303173f53c4aSTejun Heo * 3032112202d9STejun Heo * Prepare pwqs for workqueue flushing. 303373f53c4aSTejun Heo * 3034112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 3035112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 3036112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3037112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 3038112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 303973f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 304073f53c4aSTejun Heo * 304173f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 304273f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 304373f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 304473f53c4aSTejun Heo * is returned. 304573f53c4aSTejun Heo * 3046112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 304773f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 304873f53c4aSTejun Heo * advanced to @work_color. 304973f53c4aSTejun Heo * 305073f53c4aSTejun Heo * CONTEXT: 30513c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 305273f53c4aSTejun Heo * 3053d185af30SYacine Belkadi * Return: 305473f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 305573f53c4aSTejun Heo * otherwise. 305673f53c4aSTejun Heo */ 3057112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 305873f53c4aSTejun Heo int flush_color, int work_color) 30591da177e4SLinus Torvalds { 306073f53c4aSTejun Heo bool wait = false; 306149e3cf44STejun Heo struct pool_workqueue *pwq; 30621da177e4SLinus Torvalds 306373f53c4aSTejun Heo if (flush_color >= 0) { 30646183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3065112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3066dc186ad7SThomas Gleixner } 306714441960SOleg Nesterov 306849e3cf44STejun Heo for_each_pwq(pwq, wq) { 3069112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30701da177e4SLinus Torvalds 3071a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 307273f53c4aSTejun Heo 307373f53c4aSTejun Heo if (flush_color >= 0) { 30746183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 307573f53c4aSTejun Heo 3076112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3077112202d9STejun Heo pwq->flush_color = flush_color; 3078112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 307973f53c4aSTejun Heo wait = true; 30801da177e4SLinus Torvalds } 308173f53c4aSTejun Heo } 308273f53c4aSTejun Heo 308373f53c4aSTejun Heo if (work_color >= 0) { 30846183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3085112202d9STejun Heo pwq->work_color = work_color; 308673f53c4aSTejun Heo } 308773f53c4aSTejun Heo 3088a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 30891da177e4SLinus Torvalds } 30901da177e4SLinus Torvalds 3091112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 309273f53c4aSTejun Heo complete(&wq->first_flusher->done); 309373f53c4aSTejun Heo 309473f53c4aSTejun Heo return wait; 309583c22520SOleg Nesterov } 30961da177e4SLinus Torvalds 30970fcb78c2SRolf Eike Beer /** 3098c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 30990fcb78c2SRolf Eike Beer * @wq: workqueue to flush 31001da177e4SLinus Torvalds * 3101c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3102c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 31031da177e4SLinus Torvalds */ 3104c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 31051da177e4SLinus Torvalds { 310673f53c4aSTejun Heo struct wq_flusher this_flusher = { 310773f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 310873f53c4aSTejun Heo .flush_color = -1, 3109fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 311073f53c4aSTejun Heo }; 311173f53c4aSTejun Heo int next_color; 3112b1f4ec17SOleg Nesterov 31133347fa09STejun Heo if (WARN_ON(!wq_online)) 31143347fa09STejun Heo return; 31153347fa09STejun Heo 311687915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 311787915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 311887915adcSJohannes Berg 31193c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 312073f53c4aSTejun Heo 312173f53c4aSTejun Heo /* 312273f53c4aSTejun Heo * Start-to-wait phase 312373f53c4aSTejun Heo */ 312473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 312573f53c4aSTejun Heo 312673f53c4aSTejun Heo if (next_color != wq->flush_color) { 312773f53c4aSTejun Heo /* 312873f53c4aSTejun Heo * Color space is not full. The current work_color 312973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 313073f53c4aSTejun Heo * by one. 313173f53c4aSTejun Heo */ 31326183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 313373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 313473f53c4aSTejun Heo wq->work_color = next_color; 313573f53c4aSTejun Heo 313673f53c4aSTejun Heo if (!wq->first_flusher) { 313773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 31386183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 313973f53c4aSTejun Heo 314073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 314173f53c4aSTejun Heo 3142112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 314373f53c4aSTejun Heo wq->work_color)) { 314473f53c4aSTejun Heo /* nothing to flush, done */ 314573f53c4aSTejun Heo wq->flush_color = next_color; 314673f53c4aSTejun Heo wq->first_flusher = NULL; 314773f53c4aSTejun Heo goto out_unlock; 314873f53c4aSTejun Heo } 314973f53c4aSTejun Heo } else { 315073f53c4aSTejun Heo /* wait in queue */ 31516183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 315273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3153112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 315473f53c4aSTejun Heo } 315573f53c4aSTejun Heo } else { 315673f53c4aSTejun Heo /* 315773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 315873f53c4aSTejun Heo * The next flush completion will assign us 315973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 316073f53c4aSTejun Heo */ 316173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 316273f53c4aSTejun Heo } 316373f53c4aSTejun Heo 3164fca839c0STejun Heo check_flush_dependency(wq, NULL); 3165fca839c0STejun Heo 31663c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 316773f53c4aSTejun Heo 316873f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 316973f53c4aSTejun Heo 317073f53c4aSTejun Heo /* 317173f53c4aSTejun Heo * Wake-up-and-cascade phase 317273f53c4aSTejun Heo * 317373f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 317473f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 317573f53c4aSTejun Heo */ 317600d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 317773f53c4aSTejun Heo return; 317873f53c4aSTejun Heo 31793c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 318073f53c4aSTejun Heo 31814ce48b37STejun Heo /* we might have raced, check again with mutex held */ 31824ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 31834ce48b37STejun Heo goto out_unlock; 31844ce48b37STejun Heo 318500d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 318673f53c4aSTejun Heo 31876183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 31886183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 318973f53c4aSTejun Heo 319073f53c4aSTejun Heo while (true) { 319173f53c4aSTejun Heo struct wq_flusher *next, *tmp; 319273f53c4aSTejun Heo 319373f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 319473f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 319573f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 319673f53c4aSTejun Heo break; 319773f53c4aSTejun Heo list_del_init(&next->list); 319873f53c4aSTejun Heo complete(&next->done); 319973f53c4aSTejun Heo } 320073f53c4aSTejun Heo 32016183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 320273f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 320373f53c4aSTejun Heo 320473f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 320573f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 320673f53c4aSTejun Heo 320773f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 320873f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 320973f53c4aSTejun Heo /* 321073f53c4aSTejun Heo * Assign the same color to all overflowed 321173f53c4aSTejun Heo * flushers, advance work_color and append to 321273f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 321373f53c4aSTejun Heo * phase for these overflowed flushers. 321473f53c4aSTejun Heo */ 321573f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 321673f53c4aSTejun Heo tmp->flush_color = wq->work_color; 321773f53c4aSTejun Heo 321873f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 321973f53c4aSTejun Heo 322073f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 322173f53c4aSTejun Heo &wq->flusher_queue); 3222112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 322373f53c4aSTejun Heo } 322473f53c4aSTejun Heo 322573f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 32266183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 322773f53c4aSTejun Heo break; 322873f53c4aSTejun Heo } 322973f53c4aSTejun Heo 323073f53c4aSTejun Heo /* 323173f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3232112202d9STejun Heo * the new first flusher and arm pwqs. 323373f53c4aSTejun Heo */ 32346183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 32356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 323673f53c4aSTejun Heo 323773f53c4aSTejun Heo list_del_init(&next->list); 323873f53c4aSTejun Heo wq->first_flusher = next; 323973f53c4aSTejun Heo 3240112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 324173f53c4aSTejun Heo break; 324273f53c4aSTejun Heo 324373f53c4aSTejun Heo /* 324473f53c4aSTejun Heo * Meh... this color is already done, clear first 324573f53c4aSTejun Heo * flusher and repeat cascading. 324673f53c4aSTejun Heo */ 324773f53c4aSTejun Heo wq->first_flusher = NULL; 324873f53c4aSTejun Heo } 324973f53c4aSTejun Heo 325073f53c4aSTejun Heo out_unlock: 32513c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32521da177e4SLinus Torvalds } 3253c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32541da177e4SLinus Torvalds 32559c5a2ba7STejun Heo /** 32569c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32579c5a2ba7STejun Heo * @wq: workqueue to drain 32589c5a2ba7STejun Heo * 32599c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32609c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32619c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3262b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32639c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32649c5a2ba7STejun Heo * takes too long. 32659c5a2ba7STejun Heo */ 32669c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32679c5a2ba7STejun Heo { 32689c5a2ba7STejun Heo unsigned int flush_cnt = 0; 326949e3cf44STejun Heo struct pool_workqueue *pwq; 32709c5a2ba7STejun Heo 32719c5a2ba7STejun Heo /* 32729c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 32739c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3274618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 32759c5a2ba7STejun Heo */ 327687fc741eSLai Jiangshan mutex_lock(&wq->mutex); 32779c5a2ba7STejun Heo if (!wq->nr_drainers++) 3278618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 327987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32809c5a2ba7STejun Heo reflush: 3281c4f135d6STetsuo Handa __flush_workqueue(wq); 32829c5a2ba7STejun Heo 3283b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 328476af4d93STejun Heo 328549e3cf44STejun Heo for_each_pwq(pwq, wq) { 3286fa2563e4SThomas Tuttle bool drained; 32879c5a2ba7STejun Heo 3288a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3289f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3290a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3291fa2563e4SThomas Tuttle 3292fa2563e4SThomas Tuttle if (drained) 32939c5a2ba7STejun Heo continue; 32949c5a2ba7STejun Heo 32959c5a2ba7STejun Heo if (++flush_cnt == 10 || 32969c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3297e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3298e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 329976af4d93STejun Heo 3300b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 33019c5a2ba7STejun Heo goto reflush; 33029c5a2ba7STejun Heo } 33039c5a2ba7STejun Heo 33049c5a2ba7STejun Heo if (!--wq->nr_drainers) 3305618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 330687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 33079c5a2ba7STejun Heo } 33089c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 33099c5a2ba7STejun Heo 3310d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3311d6e89786SJohannes Berg bool from_cancel) 3312baf59022STejun Heo { 3313baf59022STejun Heo struct worker *worker = NULL; 3314c9e7cf27STejun Heo struct worker_pool *pool; 3315112202d9STejun Heo struct pool_workqueue *pwq; 3316baf59022STejun Heo 3317baf59022STejun Heo might_sleep(); 3318baf59022STejun Heo 331924acfb71SThomas Gleixner rcu_read_lock(); 3320fa1b54e6STejun Heo pool = get_work_pool(work); 3321fa1b54e6STejun Heo if (!pool) { 332224acfb71SThomas Gleixner rcu_read_unlock(); 3323fa1b54e6STejun Heo return false; 3324fa1b54e6STejun Heo } 3325fa1b54e6STejun Heo 3326a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 33270b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3328112202d9STejun Heo pwq = get_work_pwq(work); 3329112202d9STejun Heo if (pwq) { 3330112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3331baf59022STejun Heo goto already_gone; 3332606a5020STejun Heo } else { 3333c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3334baf59022STejun Heo if (!worker) 3335baf59022STejun Heo goto already_gone; 3336112202d9STejun Heo pwq = worker->current_pwq; 3337606a5020STejun Heo } 3338baf59022STejun Heo 3339fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3340fca839c0STejun Heo 3341112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3342a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3343baf59022STejun Heo 3344e159489bSTejun Heo /* 3345a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3346a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3347a1d14934SPeter Zijlstra * 3348a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3349a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3350a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3351a1d14934SPeter Zijlstra * forward progress. 3352e159489bSTejun Heo */ 3353d6e89786SJohannes Berg if (!from_cancel && 3354d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3355112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3356112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3357a1d14934SPeter Zijlstra } 335824acfb71SThomas Gleixner rcu_read_unlock(); 3359baf59022STejun Heo return true; 3360baf59022STejun Heo already_gone: 3361a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 336224acfb71SThomas Gleixner rcu_read_unlock(); 3363baf59022STejun Heo return false; 3364baf59022STejun Heo } 3365baf59022STejun Heo 3366d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3367d6e89786SJohannes Berg { 3368d6e89786SJohannes Berg struct wq_barrier barr; 3369d6e89786SJohannes Berg 3370d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3371d6e89786SJohannes Berg return false; 3372d6e89786SJohannes Berg 33734d43d395STetsuo Handa if (WARN_ON(!work->func)) 33744d43d395STetsuo Handa return false; 33754d43d395STetsuo Handa 337687915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 337787915adcSJohannes Berg lock_map_release(&work->lockdep_map); 337887915adcSJohannes Berg 3379d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3380d6e89786SJohannes Berg wait_for_completion(&barr.done); 3381d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3382d6e89786SJohannes Berg return true; 3383d6e89786SJohannes Berg } else { 3384d6e89786SJohannes Berg return false; 3385d6e89786SJohannes Berg } 3386d6e89786SJohannes Berg } 3387d6e89786SJohannes Berg 3388db700897SOleg Nesterov /** 3389401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3390401a8d04STejun Heo * @work: the work to flush 3391db700897SOleg Nesterov * 3392606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3393606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3394401a8d04STejun Heo * 3395d185af30SYacine Belkadi * Return: 3396401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3397401a8d04STejun Heo * %false if it was already idle. 3398db700897SOleg Nesterov */ 3399401a8d04STejun Heo bool flush_work(struct work_struct *work) 3400db700897SOleg Nesterov { 3401d6e89786SJohannes Berg return __flush_work(work, false); 3402606a5020STejun Heo } 3403db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3404db700897SOleg Nesterov 34058603e1b3STejun Heo struct cwt_wait { 3406ac6424b9SIngo Molnar wait_queue_entry_t wait; 34078603e1b3STejun Heo struct work_struct *work; 34088603e1b3STejun Heo }; 34098603e1b3STejun Heo 3410ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 34118603e1b3STejun Heo { 34128603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 34138603e1b3STejun Heo 34148603e1b3STejun Heo if (cwait->work != key) 34158603e1b3STejun Heo return 0; 34168603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 34178603e1b3STejun Heo } 34188603e1b3STejun Heo 341936e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3420401a8d04STejun Heo { 34218603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3422bbb68dfaSTejun Heo unsigned long flags; 34231f1f642eSOleg Nesterov int ret; 34241f1f642eSOleg Nesterov 34251f1f642eSOleg Nesterov do { 3426bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3427bbb68dfaSTejun Heo /* 34288603e1b3STejun Heo * If someone else is already canceling, wait for it to 34298603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 34308603e1b3STejun Heo * because we may get scheduled between @work's completion 34318603e1b3STejun Heo * and the other canceling task resuming and clearing 34328603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 34338603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 34348603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 34358603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 34368603e1b3STejun Heo * we're hogging the CPU. 34378603e1b3STejun Heo * 34388603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 34398603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 34408603e1b3STejun Heo * wake function which matches @work along with exclusive 34418603e1b3STejun Heo * wait and wakeup. 3442bbb68dfaSTejun Heo */ 34438603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 34448603e1b3STejun Heo struct cwt_wait cwait; 34458603e1b3STejun Heo 34468603e1b3STejun Heo init_wait(&cwait.wait); 34478603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34488603e1b3STejun Heo cwait.work = work; 34498603e1b3STejun Heo 34508603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34518603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34528603e1b3STejun Heo if (work_is_canceling(work)) 34538603e1b3STejun Heo schedule(); 34548603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34558603e1b3STejun Heo } 34561f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34571f1f642eSOleg Nesterov 3458bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3459bbb68dfaSTejun Heo mark_work_canceling(work); 3460bbb68dfaSTejun Heo local_irq_restore(flags); 3461bbb68dfaSTejun Heo 34623347fa09STejun Heo /* 34633347fa09STejun Heo * This allows canceling during early boot. We know that @work 34643347fa09STejun Heo * isn't executing. 34653347fa09STejun Heo */ 34663347fa09STejun Heo if (wq_online) 3467d6e89786SJohannes Berg __flush_work(work, true); 34683347fa09STejun Heo 34697a22ad75STejun Heo clear_work_data(work); 34708603e1b3STejun Heo 34718603e1b3STejun Heo /* 34728603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 34738603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 34748603e1b3STejun Heo * visible there. 34758603e1b3STejun Heo */ 34768603e1b3STejun Heo smp_mb(); 34778603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 34788603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 34798603e1b3STejun Heo 34801f1f642eSOleg Nesterov return ret; 34811f1f642eSOleg Nesterov } 34821f1f642eSOleg Nesterov 34836e84d644SOleg Nesterov /** 3484401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3485401a8d04STejun Heo * @work: the work to cancel 34866e84d644SOleg Nesterov * 3487401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3488401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3489401a8d04STejun Heo * another workqueue. On return from this function, @work is 3490401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 34911f1f642eSOleg Nesterov * 3492401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3493401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 34946e84d644SOleg Nesterov * 3495401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 34966e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3497401a8d04STejun Heo * 3498d185af30SYacine Belkadi * Return: 3499401a8d04STejun Heo * %true if @work was pending, %false otherwise. 35006e84d644SOleg Nesterov */ 3501401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 35026e84d644SOleg Nesterov { 350336e227d2STejun Heo return __cancel_work_timer(work, false); 3504b89deed3SOleg Nesterov } 350528e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3506b89deed3SOleg Nesterov 35076e84d644SOleg Nesterov /** 3508401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3509401a8d04STejun Heo * @dwork: the delayed work to flush 35106e84d644SOleg Nesterov * 3511401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3512401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3513401a8d04STejun Heo * considers the last queueing instance of @dwork. 35141f1f642eSOleg Nesterov * 3515d185af30SYacine Belkadi * Return: 3516401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3517401a8d04STejun Heo * %false if it was already idle. 35186e84d644SOleg Nesterov */ 3519401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3520401a8d04STejun Heo { 35218930cabaSTejun Heo local_irq_disable(); 3522401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 352360c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 35248930cabaSTejun Heo local_irq_enable(); 3525401a8d04STejun Heo return flush_work(&dwork->work); 3526401a8d04STejun Heo } 3527401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3528401a8d04STejun Heo 352905f0fe6bSTejun Heo /** 353005f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 353105f0fe6bSTejun Heo * @rwork: the rcu work to flush 353205f0fe6bSTejun Heo * 353305f0fe6bSTejun Heo * Return: 353405f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 353505f0fe6bSTejun Heo * %false if it was already idle. 353605f0fe6bSTejun Heo */ 353705f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 353805f0fe6bSTejun Heo { 353905f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 354005f0fe6bSTejun Heo rcu_barrier(); 354105f0fe6bSTejun Heo flush_work(&rwork->work); 354205f0fe6bSTejun Heo return true; 354305f0fe6bSTejun Heo } else { 354405f0fe6bSTejun Heo return flush_work(&rwork->work); 354505f0fe6bSTejun Heo } 354605f0fe6bSTejun Heo } 354705f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 354805f0fe6bSTejun Heo 3549f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3550f72b8792SJens Axboe { 3551f72b8792SJens Axboe unsigned long flags; 3552f72b8792SJens Axboe int ret; 3553f72b8792SJens Axboe 3554f72b8792SJens Axboe do { 3555f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3556f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3557f72b8792SJens Axboe 3558f72b8792SJens Axboe if (unlikely(ret < 0)) 3559f72b8792SJens Axboe return false; 3560f72b8792SJens Axboe 3561f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3562f72b8792SJens Axboe local_irq_restore(flags); 3563f72b8792SJens Axboe return ret; 3564f72b8792SJens Axboe } 3565f72b8792SJens Axboe 356673b4b532SAndrey Grodzovsky /* 356773b4b532SAndrey Grodzovsky * See cancel_delayed_work() 356873b4b532SAndrey Grodzovsky */ 356973b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 357073b4b532SAndrey Grodzovsky { 357173b4b532SAndrey Grodzovsky return __cancel_work(work, false); 357273b4b532SAndrey Grodzovsky } 357373b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 357473b4b532SAndrey Grodzovsky 3575401a8d04STejun Heo /** 357657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 357757b30ae7STejun Heo * @dwork: delayed_work to cancel 357809383498STejun Heo * 3579d185af30SYacine Belkadi * Kill off a pending delayed_work. 3580d185af30SYacine Belkadi * 3581d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3582d185af30SYacine Belkadi * pending. 3583d185af30SYacine Belkadi * 3584d185af30SYacine Belkadi * Note: 3585d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3586d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3587d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 358809383498STejun Heo * 358957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 359009383498STejun Heo */ 359157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 359209383498STejun Heo { 3593f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 359409383498STejun Heo } 359557b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 359609383498STejun Heo 359709383498STejun Heo /** 3598401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3599401a8d04STejun Heo * @dwork: the delayed work cancel 3600401a8d04STejun Heo * 3601401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3602401a8d04STejun Heo * 3603d185af30SYacine Belkadi * Return: 3604401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3605401a8d04STejun Heo */ 3606401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 36076e84d644SOleg Nesterov { 360836e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 36096e84d644SOleg Nesterov } 3610f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 36111da177e4SLinus Torvalds 36120fcb78c2SRolf Eike Beer /** 361331ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3614b6136773SAndrew Morton * @func: the function to call 3615b6136773SAndrew Morton * 361631ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 361731ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3618b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 361931ddd871STejun Heo * 3620d185af30SYacine Belkadi * Return: 362131ddd871STejun Heo * 0 on success, -errno on failure. 3622b6136773SAndrew Morton */ 362365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 362415316ba8SChristoph Lameter { 362515316ba8SChristoph Lameter int cpu; 362638f51568SNamhyung Kim struct work_struct __percpu *works; 362715316ba8SChristoph Lameter 3628b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3629b6136773SAndrew Morton if (!works) 363015316ba8SChristoph Lameter return -ENOMEM; 3631b6136773SAndrew Morton 3632ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 363393981800STejun Heo 363415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 36359bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 36369bfb1839SIngo Molnar 36379bfb1839SIngo Molnar INIT_WORK(work, func); 36388de6d308SOleg Nesterov schedule_work_on(cpu, work); 363915316ba8SChristoph Lameter } 364093981800STejun Heo 364193981800STejun Heo for_each_online_cpu(cpu) 36428616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 364393981800STejun Heo 3644ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3645b6136773SAndrew Morton free_percpu(works); 364615316ba8SChristoph Lameter return 0; 364715316ba8SChristoph Lameter } 364815316ba8SChristoph Lameter 3649eef6a7d5SAlan Stern /** 36501fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36511fa44ecaSJames Bottomley * @fn: the function to execute 36521fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36531fa44ecaSJames Bottomley * be available when the work executes) 36541fa44ecaSJames Bottomley * 36551fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36561fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36571fa44ecaSJames Bottomley * 3658d185af30SYacine Belkadi * Return: 0 - function was executed 36591fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36601fa44ecaSJames Bottomley */ 366165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36621fa44ecaSJames Bottomley { 36631fa44ecaSJames Bottomley if (!in_interrupt()) { 366465f27f38SDavid Howells fn(&ew->work); 36651fa44ecaSJames Bottomley return 0; 36661fa44ecaSJames Bottomley } 36671fa44ecaSJames Bottomley 366865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36691fa44ecaSJames Bottomley schedule_work(&ew->work); 36701fa44ecaSJames Bottomley 36711fa44ecaSJames Bottomley return 1; 36721fa44ecaSJames Bottomley } 36731fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 36741fa44ecaSJames Bottomley 36757a4e344cSTejun Heo /** 36767a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 36777a4e344cSTejun Heo * @attrs: workqueue_attrs to free 36787a4e344cSTejun Heo * 36797a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 36807a4e344cSTejun Heo */ 3681513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 36827a4e344cSTejun Heo { 36837a4e344cSTejun Heo if (attrs) { 36847a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 36857a4e344cSTejun Heo kfree(attrs); 36867a4e344cSTejun Heo } 36877a4e344cSTejun Heo } 36887a4e344cSTejun Heo 36897a4e344cSTejun Heo /** 36907a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 36917a4e344cSTejun Heo * 36927a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3693d185af30SYacine Belkadi * return it. 3694d185af30SYacine Belkadi * 3695d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 36967a4e344cSTejun Heo */ 3697513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 36987a4e344cSTejun Heo { 36997a4e344cSTejun Heo struct workqueue_attrs *attrs; 37007a4e344cSTejun Heo 3701be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 37027a4e344cSTejun Heo if (!attrs) 37037a4e344cSTejun Heo goto fail; 3704be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 37057a4e344cSTejun Heo goto fail; 37067a4e344cSTejun Heo 370713e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 37087a4e344cSTejun Heo return attrs; 37097a4e344cSTejun Heo fail: 37107a4e344cSTejun Heo free_workqueue_attrs(attrs); 37117a4e344cSTejun Heo return NULL; 37127a4e344cSTejun Heo } 37137a4e344cSTejun Heo 371429c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 371529c91e99STejun Heo const struct workqueue_attrs *from) 371629c91e99STejun Heo { 371729c91e99STejun Heo to->nice = from->nice; 371829c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 37192865a8fbSShaohua Li /* 37202865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 37212865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 37222865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 37232865a8fbSShaohua Li */ 37242865a8fbSShaohua Li to->no_numa = from->no_numa; 372529c91e99STejun Heo } 372629c91e99STejun Heo 372729c91e99STejun Heo /* hash value of the content of @attr */ 372829c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 372929c91e99STejun Heo { 373029c91e99STejun Heo u32 hash = 0; 373129c91e99STejun Heo 373229c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 373313e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 373413e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 373529c91e99STejun Heo return hash; 373629c91e99STejun Heo } 373729c91e99STejun Heo 373829c91e99STejun Heo /* content equality test */ 373929c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 374029c91e99STejun Heo const struct workqueue_attrs *b) 374129c91e99STejun Heo { 374229c91e99STejun Heo if (a->nice != b->nice) 374329c91e99STejun Heo return false; 374429c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 374529c91e99STejun Heo return false; 374629c91e99STejun Heo return true; 374729c91e99STejun Heo } 374829c91e99STejun Heo 37497a4e344cSTejun Heo /** 37507a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 37517a4e344cSTejun Heo * @pool: worker_pool to initialize 37527a4e344cSTejun Heo * 3753402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3754d185af30SYacine Belkadi * 3755d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 375629c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 375729c91e99STejun Heo * on @pool safely to release it. 37587a4e344cSTejun Heo */ 37597a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 37604e1a1f9aSTejun Heo { 3761a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 376229c91e99STejun Heo pool->id = -1; 376329c91e99STejun Heo pool->cpu = -1; 3764f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 37654e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 376682607adcSTejun Heo pool->watchdog_ts = jiffies; 37674e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37684e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 37694e1a1f9aSTejun Heo hash_init(pool->busy_hash); 37704e1a1f9aSTejun Heo 377132a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 37723f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 37734e1a1f9aSTejun Heo 377432a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 37754e1a1f9aSTejun Heo 3776da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3777e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 37787a4e344cSTejun Heo 37797cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 378029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 378129c91e99STejun Heo pool->refcnt = 1; 378229c91e99STejun Heo 378329c91e99STejun Heo /* shouldn't fail above this point */ 3784be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 37857a4e344cSTejun Heo if (!pool->attrs) 37867a4e344cSTejun Heo return -ENOMEM; 37877a4e344cSTejun Heo return 0; 37884e1a1f9aSTejun Heo } 37894e1a1f9aSTejun Heo 3790669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3791669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3792669de8bdSBart Van Assche { 3793669de8bdSBart Van Assche char *lock_name; 3794669de8bdSBart Van Assche 3795669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3796669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3797669de8bdSBart Van Assche if (!lock_name) 3798669de8bdSBart Van Assche lock_name = wq->name; 379969a106c0SQian Cai 380069a106c0SQian Cai wq->lock_name = lock_name; 3801669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3802669de8bdSBart Van Assche } 3803669de8bdSBart Van Assche 3804669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3805669de8bdSBart Van Assche { 3806669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3807669de8bdSBart Van Assche } 3808669de8bdSBart Van Assche 3809669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3810669de8bdSBart Van Assche { 3811669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3812669de8bdSBart Van Assche kfree(wq->lock_name); 3813669de8bdSBart Van Assche } 3814669de8bdSBart Van Assche #else 3815669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3816669de8bdSBart Van Assche { 3817669de8bdSBart Van Assche } 3818669de8bdSBart Van Assche 3819669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3820669de8bdSBart Van Assche { 3821669de8bdSBart Van Assche } 3822669de8bdSBart Van Assche 3823669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3824669de8bdSBart Van Assche { 3825669de8bdSBart Van Assche } 3826669de8bdSBart Van Assche #endif 3827669de8bdSBart Van Assche 3828e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3829e2dca7adSTejun Heo { 3830e2dca7adSTejun Heo struct workqueue_struct *wq = 3831e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3832e2dca7adSTejun Heo 3833669de8bdSBart Van Assche wq_free_lockdep(wq); 3834669de8bdSBart Van Assche 3835e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3836e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3837e2dca7adSTejun Heo else 3838e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3839e2dca7adSTejun Heo 3840e2dca7adSTejun Heo kfree(wq); 3841e2dca7adSTejun Heo } 3842e2dca7adSTejun Heo 384329c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 384429c91e99STejun Heo { 384529c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 384629c91e99STejun Heo 38477cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 384829c91e99STejun Heo free_workqueue_attrs(pool->attrs); 384929c91e99STejun Heo kfree(pool); 385029c91e99STejun Heo } 385129c91e99STejun Heo 385229c91e99STejun Heo /** 385329c91e99STejun Heo * put_unbound_pool - put a worker_pool 385429c91e99STejun Heo * @pool: worker_pool to put 385529c91e99STejun Heo * 385624acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3857c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3858c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3859c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3860a892caccSTejun Heo * 3861a892caccSTejun Heo * Should be called with wq_pool_mutex held. 386229c91e99STejun Heo */ 386329c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 386429c91e99STejun Heo { 386560f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 3866e02b9312SValentin Schneider struct list_head cull_list; 386729c91e99STejun Heo struct worker *worker; 386829c91e99STejun Heo 3869e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 3870e02b9312SValentin Schneider 3871a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3872a892caccSTejun Heo 3873a892caccSTejun Heo if (--pool->refcnt) 387429c91e99STejun Heo return; 387529c91e99STejun Heo 387629c91e99STejun Heo /* sanity checks */ 387761d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3878a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 387929c91e99STejun Heo return; 388029c91e99STejun Heo 388129c91e99STejun Heo /* release id and unhash */ 388229c91e99STejun Heo if (pool->id >= 0) 388329c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 388429c91e99STejun Heo hash_del(&pool->hash_node); 388529c91e99STejun Heo 3886c5aa87bbSTejun Heo /* 3887692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3888692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3889692b4825STejun Heo * manager and @pool gets freed with the flag set. 38909ab03be4SValentin Schneider * 38919ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 38929ab03be4SValentin Schneider * only get here with 38939ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 38949ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 38959ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 38969ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 38979ab03be4SValentin Schneider * drops pool->lock 3898c5aa87bbSTejun Heo */ 38999ab03be4SValentin Schneider while (true) { 39009ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 39019ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3902d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3903e02b9312SValentin Schneider 3904e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 39059ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 39069ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3907692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 39089ab03be4SValentin Schneider break; 39099ab03be4SValentin Schneider } 39109ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3911e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 39129ab03be4SValentin Schneider } 3913692b4825STejun Heo 39141037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3915e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 391629c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3917a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 391860f5a4bcSLai Jiangshan 3919e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3920e02b9312SValentin Schneider 3921e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 392260f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 39231258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 392460f5a4bcSLai Jiangshan 392560f5a4bcSLai Jiangshan if (pool->detach_completion) 392660f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 392760f5a4bcSLai Jiangshan 392829c91e99STejun Heo /* shut down the timers */ 392929c91e99STejun Heo del_timer_sync(&pool->idle_timer); 39303f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 393129c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 393229c91e99STejun Heo 393324acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 393425b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 393529c91e99STejun Heo } 393629c91e99STejun Heo 393729c91e99STejun Heo /** 393829c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 393929c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 394029c91e99STejun Heo * 394129c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 394229c91e99STejun Heo * reference count and return it. If there already is a matching 394329c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3944d185af30SYacine Belkadi * create a new one. 3945a892caccSTejun Heo * 3946a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3947d185af30SYacine Belkadi * 3948d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3949d185af30SYacine Belkadi * On failure, %NULL. 395029c91e99STejun Heo */ 395129c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 395229c91e99STejun Heo { 395329c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 395429c91e99STejun Heo struct worker_pool *pool; 3955f3f90ad4STejun Heo int node; 3956e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 395729c91e99STejun Heo 3958a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 395929c91e99STejun Heo 396029c91e99STejun Heo /* do we already have a matching pool? */ 396129c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 396229c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 396329c91e99STejun Heo pool->refcnt++; 39643fb1823cSLai Jiangshan return pool; 396529c91e99STejun Heo } 396629c91e99STejun Heo } 396729c91e99STejun Heo 3968e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3969e2273584SXunlei Pang if (wq_numa_enabled) { 3970e2273584SXunlei Pang for_each_node(node) { 3971e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3972e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3973e2273584SXunlei Pang target_node = node; 3974e2273584SXunlei Pang break; 3975e2273584SXunlei Pang } 3976e2273584SXunlei Pang } 3977e2273584SXunlei Pang } 3978e2273584SXunlei Pang 397929c91e99STejun Heo /* nope, create a new one */ 3980e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 398129c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 398229c91e99STejun Heo goto fail; 398329c91e99STejun Heo 39848864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 398529c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3986e2273584SXunlei Pang pool->node = target_node; 398729c91e99STejun Heo 39882865a8fbSShaohua Li /* 39892865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 39902865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 39912865a8fbSShaohua Li */ 39922865a8fbSShaohua Li pool->attrs->no_numa = false; 39932865a8fbSShaohua Li 399429c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 399529c91e99STejun Heo goto fail; 399629c91e99STejun Heo 399729c91e99STejun Heo /* create and start the initial worker */ 39983347fa09STejun Heo if (wq_online && !create_worker(pool)) 399929c91e99STejun Heo goto fail; 400029c91e99STejun Heo 400129c91e99STejun Heo /* install */ 400229c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 40033fb1823cSLai Jiangshan 400429c91e99STejun Heo return pool; 400529c91e99STejun Heo fail: 400629c91e99STejun Heo if (pool) 400729c91e99STejun Heo put_unbound_pool(pool); 400829c91e99STejun Heo return NULL; 400929c91e99STejun Heo } 401029c91e99STejun Heo 40118864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 40128864b4e5STejun Heo { 40138864b4e5STejun Heo kmem_cache_free(pwq_cache, 40148864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 40158864b4e5STejun Heo } 40168864b4e5STejun Heo 40178864b4e5STejun Heo /* 40188864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 40198864b4e5STejun Heo * and needs to be destroyed. 40208864b4e5STejun Heo */ 40218864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 40228864b4e5STejun Heo { 40238864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 40248864b4e5STejun Heo unbound_release_work); 40258864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 40268864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 4027b42b0bddSYang Yingliang bool is_last = false; 40288864b4e5STejun Heo 4029b42b0bddSYang Yingliang /* 4030b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 4031b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 4032b42b0bddSYang Yingliang */ 4033b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 40348864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 40358864b4e5STejun Heo return; 40368864b4e5STejun Heo 40373c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 40388864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 4039bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 40403c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 4041b42b0bddSYang Yingliang } 40428864b4e5STejun Heo 4043a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 40448864b4e5STejun Heo put_unbound_pool(pool); 4045a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 4046a892caccSTejun Heo 404725b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 40488864b4e5STejun Heo 40498864b4e5STejun Heo /* 40508864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 4051e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 40528864b4e5STejun Heo */ 4053669de8bdSBart Van Assche if (is_last) { 4054669de8bdSBart Van Assche wq_unregister_lockdep(wq); 405525b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 40566029a918STejun Heo } 4057669de8bdSBart Van Assche } 40588864b4e5STejun Heo 40590fbd95aaSTejun Heo /** 4060699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 40610fbd95aaSTejun Heo * @pwq: target pool_workqueue 40620fbd95aaSTejun Heo * 4063699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4064f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4065699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 40660fbd95aaSTejun Heo */ 4067699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 40680fbd95aaSTejun Heo { 4069699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4070699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 40713347fa09STejun Heo unsigned long flags; 4072699ce097STejun Heo 4073699ce097STejun Heo /* for @wq->saved_max_active */ 4074a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4075699ce097STejun Heo 4076699ce097STejun Heo /* fast exit for non-freezable wqs */ 4077699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4078699ce097STejun Heo return; 4079699ce097STejun Heo 40803347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4081a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4082699ce097STejun Heo 408374b414eaSLai Jiangshan /* 408474b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 408574b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 408674b414eaSLai Jiangshan * is updated and visible. 408774b414eaSLai Jiangshan */ 408874b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 408901341fbdSYunfeng Ye bool kick = false; 409001341fbdSYunfeng Ye 4091699ce097STejun Heo pwq->max_active = wq->saved_max_active; 40920fbd95aaSTejun Heo 4093f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 409401341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 4095f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 409601341fbdSYunfeng Ye kick = true; 409701341fbdSYunfeng Ye } 4098951a078aSLai Jiangshan 4099951a078aSLai Jiangshan /* 4100951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 410101341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 410201341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 410301341fbdSYunfeng Ye * let's kick iff work items were activated. 4104951a078aSLai Jiangshan */ 410501341fbdSYunfeng Ye if (kick) 4106951a078aSLai Jiangshan wake_up_worker(pwq->pool); 4107699ce097STejun Heo } else { 4108699ce097STejun Heo pwq->max_active = 0; 4109699ce097STejun Heo } 4110699ce097STejun Heo 4111a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 41120fbd95aaSTejun Heo } 41130fbd95aaSTejun Heo 411467dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4115f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4116f147f29eSTejun Heo struct worker_pool *pool) 4117d2c1d404STejun Heo { 4118d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4119d2c1d404STejun Heo 4120e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4121e50aba9aSTejun Heo 4122d2c1d404STejun Heo pwq->pool = pool; 4123d2c1d404STejun Heo pwq->wq = wq; 4124d2c1d404STejun Heo pwq->flush_color = -1; 41258864b4e5STejun Heo pwq->refcnt = 1; 4126f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 41271befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4128d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 41298864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 4130f147f29eSTejun Heo } 4131d2c1d404STejun Heo 4132f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 41331befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4134f147f29eSTejun Heo { 4135f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4136f147f29eSTejun Heo 4137f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 413875ccf595STejun Heo 41391befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 41401befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 41411befcf30STejun Heo return; 41421befcf30STejun Heo 414329b1cb41SLai Jiangshan /* set the matching work_color */ 414475ccf595STejun Heo pwq->work_color = wq->work_color; 4145983ca25eSTejun Heo 4146983ca25eSTejun Heo /* sync max_active to the current setting */ 4147983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4148983ca25eSTejun Heo 4149983ca25eSTejun Heo /* link in @pwq */ 41509e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4151df2d5ae4STejun Heo } 41526029a918STejun Heo 4153f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4154f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4155f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4156f147f29eSTejun Heo { 4157f147f29eSTejun Heo struct worker_pool *pool; 4158f147f29eSTejun Heo struct pool_workqueue *pwq; 4159f147f29eSTejun Heo 4160f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4161f147f29eSTejun Heo 4162f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4163f147f29eSTejun Heo if (!pool) 4164f147f29eSTejun Heo return NULL; 4165f147f29eSTejun Heo 4166e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4167f147f29eSTejun Heo if (!pwq) { 4168f147f29eSTejun Heo put_unbound_pool(pool); 4169f147f29eSTejun Heo return NULL; 4170f147f29eSTejun Heo } 4171f147f29eSTejun Heo 4172f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4173f147f29eSTejun Heo return pwq; 4174d2c1d404STejun Heo } 4175d2c1d404STejun Heo 41764c16bd32STejun Heo /** 417730186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 4178042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 41794c16bd32STejun Heo * @node: the target NUMA node 41804c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 41814c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 41824c16bd32STejun Heo * 41834c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 41844c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 4185d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 41864c16bd32STejun Heo * 41874c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 41884c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 41894c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 41904c16bd32STejun Heo * @attrs->cpumask. 41914c16bd32STejun Heo * 41924c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 41934c16bd32STejun Heo * stable. 4194d185af30SYacine Belkadi * 4195d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 4196d185af30SYacine Belkadi * %false if equal. 41974c16bd32STejun Heo */ 41984c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 41994c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 42004c16bd32STejun Heo { 4201d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 42024c16bd32STejun Heo goto use_dfl; 42034c16bd32STejun Heo 42044c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 42054c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 42064c16bd32STejun Heo if (cpu_going_down >= 0) 42074c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 42084c16bd32STejun Heo 42094c16bd32STejun Heo if (cpumask_empty(cpumask)) 42104c16bd32STejun Heo goto use_dfl; 42114c16bd32STejun Heo 42124c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 42134c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 42141ad0f0a7SMichael Bringmann 42151ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 42161ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 42171ad0f0a7SMichael Bringmann "possible intersect\n"); 42181ad0f0a7SMichael Bringmann return false; 42191ad0f0a7SMichael Bringmann } 42201ad0f0a7SMichael Bringmann 42214c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 42224c16bd32STejun Heo 42234c16bd32STejun Heo use_dfl: 42244c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 42254c16bd32STejun Heo return false; 42264c16bd32STejun Heo } 42274c16bd32STejun Heo 42281befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 42291befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 42301befcf30STejun Heo int node, 42311befcf30STejun Heo struct pool_workqueue *pwq) 42321befcf30STejun Heo { 42331befcf30STejun Heo struct pool_workqueue *old_pwq; 42341befcf30STejun Heo 42355b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42361befcf30STejun Heo lockdep_assert_held(&wq->mutex); 42371befcf30STejun Heo 42381befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 42391befcf30STejun Heo link_pwq(pwq); 42401befcf30STejun Heo 42411befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42421befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 42431befcf30STejun Heo return old_pwq; 42441befcf30STejun Heo } 42451befcf30STejun Heo 42462d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 42472d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 42482d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 42492d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4250042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 42512d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 42522d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 42532d5f0764SLai Jiangshan }; 42542d5f0764SLai Jiangshan 42552d5f0764SLai Jiangshan /* free the resources after success or abort */ 42562d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 42572d5f0764SLai Jiangshan { 42582d5f0764SLai Jiangshan if (ctx) { 42592d5f0764SLai Jiangshan int node; 42602d5f0764SLai Jiangshan 42612d5f0764SLai Jiangshan for_each_node(node) 42622d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 42632d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 42642d5f0764SLai Jiangshan 42652d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 42662d5f0764SLai Jiangshan 42672d5f0764SLai Jiangshan kfree(ctx); 42682d5f0764SLai Jiangshan } 42692d5f0764SLai Jiangshan } 42702d5f0764SLai Jiangshan 42712d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 42722d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 42732d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 427499c621efSLai Jiangshan const struct workqueue_attrs *attrs, 427599c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 42762d5f0764SLai Jiangshan { 42772d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 42782d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 42792d5f0764SLai Jiangshan int node; 42802d5f0764SLai Jiangshan 42812d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42822d5f0764SLai Jiangshan 4283acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 42842d5f0764SLai Jiangshan 4285be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4286be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 42872d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 42882d5f0764SLai Jiangshan goto out_free; 42892d5f0764SLai Jiangshan 4290042f7df1SLai Jiangshan /* 429199c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 429299c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4293042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4294042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4295042f7df1SLai Jiangshan */ 42962d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 429799c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4298042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 429999c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 43002d5f0764SLai Jiangshan 43012d5f0764SLai Jiangshan /* 43022d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 43032d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 43042d5f0764SLai Jiangshan * pools. 43052d5f0764SLai Jiangshan */ 43062d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 43072d5f0764SLai Jiangshan 43082d5f0764SLai Jiangshan /* 43092d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 43102d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 43112d5f0764SLai Jiangshan * it even if we don't use it immediately. 43122d5f0764SLai Jiangshan */ 43132d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 43142d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 43152d5f0764SLai Jiangshan goto out_free; 43162d5f0764SLai Jiangshan 43172d5f0764SLai Jiangshan for_each_node(node) { 4318042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 43192d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 43202d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 43212d5f0764SLai Jiangshan goto out_free; 43222d5f0764SLai Jiangshan } else { 43232d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 43242d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 43252d5f0764SLai Jiangshan } 43262d5f0764SLai Jiangshan } 43272d5f0764SLai Jiangshan 4328042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4329042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4330042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 43312d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4332042f7df1SLai Jiangshan 43332d5f0764SLai Jiangshan ctx->wq = wq; 43342d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43352d5f0764SLai Jiangshan return ctx; 43362d5f0764SLai Jiangshan 43372d5f0764SLai Jiangshan out_free: 43382d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43392d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 43402d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 43412d5f0764SLai Jiangshan return NULL; 43422d5f0764SLai Jiangshan } 43432d5f0764SLai Jiangshan 43442d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 43452d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 43462d5f0764SLai Jiangshan { 43472d5f0764SLai Jiangshan int node; 43482d5f0764SLai Jiangshan 43492d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 43502d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 43512d5f0764SLai Jiangshan 43522d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 43532d5f0764SLai Jiangshan 43542d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 43552d5f0764SLai Jiangshan for_each_node(node) 43562d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 43572d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 43582d5f0764SLai Jiangshan 43592d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 43602d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 43612d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 43622d5f0764SLai Jiangshan 43632d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 43642d5f0764SLai Jiangshan } 43652d5f0764SLai Jiangshan 4366a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4367a0111cf6SLai Jiangshan { 4368a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4369ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4370a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4371a0111cf6SLai Jiangshan } 4372a0111cf6SLai Jiangshan 4373a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4374a0111cf6SLai Jiangshan { 4375a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4376ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4377a0111cf6SLai Jiangshan } 4378a0111cf6SLai Jiangshan 4379a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4380a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4381a0111cf6SLai Jiangshan { 4382a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4383a0111cf6SLai Jiangshan 4384a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4385a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4386a0111cf6SLai Jiangshan return -EINVAL; 4387a0111cf6SLai Jiangshan 4388a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 43890a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 43900a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4391a0111cf6SLai Jiangshan return -EINVAL; 4392a0111cf6SLai Jiangshan 43930a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 43940a94efb5STejun Heo } 43950a94efb5STejun Heo 439699c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 43976201171eSwanghaibin if (!ctx) 43986201171eSwanghaibin return -ENOMEM; 4399a0111cf6SLai Jiangshan 4400a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4401a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4402a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4403a0111cf6SLai Jiangshan 44046201171eSwanghaibin return 0; 4405a0111cf6SLai Jiangshan } 4406a0111cf6SLai Jiangshan 44079e8cd2f5STejun Heo /** 44089e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 44099e8cd2f5STejun Heo * @wq: the target workqueue 44109e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 44119e8cd2f5STejun Heo * 44124c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 44134c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 44144c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 44154c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 44164c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 44174c16bd32STejun Heo * back-to-back will stay on its current pwq. 44189e8cd2f5STejun Heo * 4419d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4420d185af30SYacine Belkadi * 4421ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4422509b3204SDaniel Jordan * 4423d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 44249e8cd2f5STejun Heo */ 4425513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 44269e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 44279e8cd2f5STejun Heo { 4428a0111cf6SLai Jiangshan int ret; 44299e8cd2f5STejun Heo 4430509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4431509b3204SDaniel Jordan 4432509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4433a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4434509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 44352d5f0764SLai Jiangshan 44362d5f0764SLai Jiangshan return ret; 44379e8cd2f5STejun Heo } 44389e8cd2f5STejun Heo 44394c16bd32STejun Heo /** 44404c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 44414c16bd32STejun Heo * @wq: the target workqueue 44424c16bd32STejun Heo * @cpu: the CPU coming up or going down 44434c16bd32STejun Heo * @online: whether @cpu is coming up or going down 44444c16bd32STejun Heo * 44454c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 44464c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 44474c16bd32STejun Heo * @wq accordingly. 44484c16bd32STejun Heo * 44494c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 44504c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 44514c16bd32STejun Heo * correct. 44524c16bd32STejun Heo * 44534c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 44544c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 44554c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 44564c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 44574c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 44584c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 44594c16bd32STejun Heo * CPU_DOWN_PREPARE. 44604c16bd32STejun Heo */ 44614c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 44624c16bd32STejun Heo bool online) 44634c16bd32STejun Heo { 44644c16bd32STejun Heo int node = cpu_to_node(cpu); 44654c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 44664c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 44674c16bd32STejun Heo struct workqueue_attrs *target_attrs; 44684c16bd32STejun Heo cpumask_t *cpumask; 44694c16bd32STejun Heo 44704c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 44714c16bd32STejun Heo 4472f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4473f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 44744c16bd32STejun Heo return; 44754c16bd32STejun Heo 44764c16bd32STejun Heo /* 44774c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 44784c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 44794c16bd32STejun Heo * CPU hotplug exclusion. 44804c16bd32STejun Heo */ 44814c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 44824c16bd32STejun Heo cpumask = target_attrs->cpumask; 44834c16bd32STejun Heo 44844c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 44854c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 44864c16bd32STejun Heo 44874c16bd32STejun Heo /* 44884c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4489042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4490042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4491042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 44924c16bd32STejun Heo */ 4493042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 44944c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4495f7142ed4SLai Jiangshan return; 44964c16bd32STejun Heo } else { 44974c16bd32STejun Heo goto use_dfl_pwq; 44984c16bd32STejun Heo } 44994c16bd32STejun Heo 45004c16bd32STejun Heo /* create a new pwq */ 45014c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 45024c16bd32STejun Heo if (!pwq) { 45032d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 45044c16bd32STejun Heo wq->name); 450577f300b1SDaeseok Youn goto use_dfl_pwq; 45064c16bd32STejun Heo } 45074c16bd32STejun Heo 4508f7142ed4SLai Jiangshan /* Install the new pwq. */ 45094c16bd32STejun Heo mutex_lock(&wq->mutex); 45104c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 45114c16bd32STejun Heo goto out_unlock; 45124c16bd32STejun Heo 45134c16bd32STejun Heo use_dfl_pwq: 4514f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4515a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 45164c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4517a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 45184c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 45194c16bd32STejun Heo out_unlock: 45204c16bd32STejun Heo mutex_unlock(&wq->mutex); 45214c16bd32STejun Heo put_pwq_unlocked(old_pwq); 45224c16bd32STejun Heo } 45234c16bd32STejun Heo 452430cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 45251da177e4SLinus Torvalds { 452649e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 45278a2b7538STejun Heo int cpu, ret; 4528e1d8aa9fSFrederic Weisbecker 452930cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4530420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4531420c0ddbSTejun Heo if (!wq->cpu_pwqs) 453230cdf249STejun Heo return -ENOMEM; 453330cdf249STejun Heo 453430cdf249STejun Heo for_each_possible_cpu(cpu) { 45357fb98ea7STejun Heo struct pool_workqueue *pwq = 45367fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 45377a62c2c8STejun Heo struct worker_pool *cpu_pools = 4538f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 453930cdf249STejun Heo 4540f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4541f147f29eSTejun Heo 4542f147f29eSTejun Heo mutex_lock(&wq->mutex); 45431befcf30STejun Heo link_pwq(pwq); 4544f147f29eSTejun Heo mutex_unlock(&wq->mutex); 454530cdf249STejun Heo } 454630cdf249STejun Heo return 0; 4547509b3204SDaniel Jordan } 4548509b3204SDaniel Jordan 4549ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4550509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 45518a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 45528a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 45538a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 45548a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 45558a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 45569e8cd2f5STejun Heo } else { 4557509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 45589e8cd2f5STejun Heo } 4559ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4560509b3204SDaniel Jordan 4561509b3204SDaniel Jordan return ret; 45620f900049STejun Heo } 45630f900049STejun Heo 4564f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4565f3421797STejun Heo const char *name) 4566b71ab8c2STejun Heo { 4567f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4568f3421797STejun Heo 4569f3421797STejun Heo if (max_active < 1 || max_active > lim) 4570044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4571f3421797STejun Heo max_active, name, 1, lim); 4572b71ab8c2STejun Heo 4573f3421797STejun Heo return clamp_val(max_active, 1, lim); 4574b71ab8c2STejun Heo } 4575b71ab8c2STejun Heo 4576983c7515STejun Heo /* 4577983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4578983c7515STejun Heo * to guarantee forward progress. 4579983c7515STejun Heo */ 4580983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4581983c7515STejun Heo { 4582983c7515STejun Heo struct worker *rescuer; 4583b92b36eaSDan Carpenter int ret; 4584983c7515STejun Heo 4585983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4586983c7515STejun Heo return 0; 4587983c7515STejun Heo 4588983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 45894c0736a7SPetr Mladek if (!rescuer) { 45904c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 45914c0736a7SPetr Mladek wq->name); 4592983c7515STejun Heo return -ENOMEM; 45934c0736a7SPetr Mladek } 4594983c7515STejun Heo 4595983c7515STejun Heo rescuer->rescue_wq = wq; 4596983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4597f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4598b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 45994c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 46004c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4601983c7515STejun Heo kfree(rescuer); 4602b92b36eaSDan Carpenter return ret; 4603983c7515STejun Heo } 4604983c7515STejun Heo 4605983c7515STejun Heo wq->rescuer = rescuer; 4606983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4607983c7515STejun Heo wake_up_process(rescuer->task); 4608983c7515STejun Heo 4609983c7515STejun Heo return 0; 4610983c7515STejun Heo } 4611983c7515STejun Heo 4612a2775bbcSMathieu Malaterre __printf(1, 4) 4613669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 461497e37d7bSTejun Heo unsigned int flags, 4615669de8bdSBart Van Assche int max_active, ...) 46163af24433SOleg Nesterov { 4617df2d5ae4STejun Heo size_t tbl_size = 0; 4618ecf6881fSTejun Heo va_list args; 46193af24433SOleg Nesterov struct workqueue_struct *wq; 462049e3cf44STejun Heo struct pool_workqueue *pwq; 4621b196be89STejun Heo 46225c0338c6STejun Heo /* 46235c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 46245c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 46255c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 46265c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 46275c0338c6STejun Heo * on NUMA. 46285c0338c6STejun Heo */ 46295c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 46305c0338c6STejun Heo flags |= __WQ_ORDERED; 46315c0338c6STejun Heo 4632cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4633cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4634cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4635cee22a15SViresh Kumar 4636ecf6881fSTejun Heo /* allocate wq and format name */ 4637df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4638ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4639df2d5ae4STejun Heo 4640df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4641b196be89STejun Heo if (!wq) 4642d2c1d404STejun Heo return NULL; 4643b196be89STejun Heo 46446029a918STejun Heo if (flags & WQ_UNBOUND) { 4645be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 46466029a918STejun Heo if (!wq->unbound_attrs) 46476029a918STejun Heo goto err_free_wq; 46486029a918STejun Heo } 46496029a918STejun Heo 4650669de8bdSBart Van Assche va_start(args, max_active); 4651ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4652b196be89STejun Heo va_end(args); 46533af24433SOleg Nesterov 4654d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4655b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 46563af24433SOleg Nesterov 4657b196be89STejun Heo /* init wq */ 465897e37d7bSTejun Heo wq->flags = flags; 4659a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 46603c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4661112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 466230cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 466373f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 466473f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4665493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 46663af24433SOleg Nesterov 4667669de8bdSBart Van Assche wq_init_lockdep(wq); 4668cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 46693af24433SOleg Nesterov 467030cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 467182efcab3SBart Van Assche goto err_unreg_lockdep; 46721537663fSTejun Heo 467340c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4674d2c1d404STejun Heo goto err_destroy; 4675e22bee78STejun Heo 4676226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4677226223abSTejun Heo goto err_destroy; 4678226223abSTejun Heo 46796af8bf3dSOleg Nesterov /* 468068e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 468168e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 468268e13a67SLai Jiangshan * list. 46836af8bf3dSOleg Nesterov */ 468468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4685a0a1a5fdSTejun Heo 4686a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 468749e3cf44STejun Heo for_each_pwq(pwq, wq) 4688699ce097STejun Heo pwq_adjust_max_active(pwq); 4689a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4690a0a1a5fdSTejun Heo 4691e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4692a0a1a5fdSTejun Heo 469368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46943af24433SOleg Nesterov 46953af24433SOleg Nesterov return wq; 4696d2c1d404STejun Heo 469782efcab3SBart Van Assche err_unreg_lockdep: 4698009bb421SBart Van Assche wq_unregister_lockdep(wq); 4699009bb421SBart Van Assche wq_free_lockdep(wq); 470082efcab3SBart Van Assche err_free_wq: 47016029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 47024690c4abSTejun Heo kfree(wq); 4703d2c1d404STejun Heo return NULL; 4704d2c1d404STejun Heo err_destroy: 4705d2c1d404STejun Heo destroy_workqueue(wq); 47064690c4abSTejun Heo return NULL; 47071da177e4SLinus Torvalds } 4708669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 47091da177e4SLinus Torvalds 4710c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4711c29eb853STejun Heo { 4712c29eb853STejun Heo int i; 4713c29eb853STejun Heo 4714c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4715c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4716c29eb853STejun Heo return true; 4717c29eb853STejun Heo 4718c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4719c29eb853STejun Heo return true; 4720f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4721c29eb853STejun Heo return true; 4722c29eb853STejun Heo 4723c29eb853STejun Heo return false; 4724c29eb853STejun Heo } 4725c29eb853STejun Heo 47263af24433SOleg Nesterov /** 47273af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 47283af24433SOleg Nesterov * @wq: target workqueue 47293af24433SOleg Nesterov * 47303af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 47313af24433SOleg Nesterov */ 47323af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 47333af24433SOleg Nesterov { 473449e3cf44STejun Heo struct pool_workqueue *pwq; 47354c16bd32STejun Heo int node; 47363af24433SOleg Nesterov 4737def98c84STejun Heo /* 4738def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4739def98c84STejun Heo * lead to sysfs name conflicts. 4740def98c84STejun Heo */ 4741def98c84STejun Heo workqueue_sysfs_unregister(wq); 4742def98c84STejun Heo 474333e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 474433e3f0a3SRichard Clark mutex_lock(&wq->mutex); 474533e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 474633e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 474733e3f0a3SRichard Clark 47489c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 47499c5a2ba7STejun Heo drain_workqueue(wq); 4750c8efcc25STejun Heo 4751def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4752def98c84STejun Heo if (wq->rescuer) { 4753def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4754def98c84STejun Heo 4755def98c84STejun Heo /* this prevents new queueing */ 4756a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4757def98c84STejun Heo wq->rescuer = NULL; 4758a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4759def98c84STejun Heo 4760def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4761def98c84STejun Heo kthread_stop(rescuer->task); 47628efe1223STejun Heo kfree(rescuer); 4763def98c84STejun Heo } 4764def98c84STejun Heo 4765c29eb853STejun Heo /* 4766c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4767c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4768c29eb853STejun Heo */ 4769c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4770b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 477149e3cf44STejun Heo for_each_pwq(pwq, wq) { 4772a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4773c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 47741d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4775e66b39afSTejun Heo __func__, wq->name); 4776c29eb853STejun Heo show_pwq(pwq); 4777a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4778b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4779c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 478055df0933SImran Khan show_one_workqueue(wq); 47816183c009STejun Heo return; 478276af4d93STejun Heo } 4783a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 478476af4d93STejun Heo } 4785b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 47866183c009STejun Heo 4787a0a1a5fdSTejun Heo /* 4788a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4789a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4790a0a1a5fdSTejun Heo */ 4791e2dca7adSTejun Heo list_del_rcu(&wq->list); 479268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47933af24433SOleg Nesterov 47948864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4795669de8bdSBart Van Assche wq_unregister_lockdep(wq); 479629c91e99STejun Heo /* 47978864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4798e2dca7adSTejun Heo * schedule RCU free. 479929c91e99STejun Heo */ 480025b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 48018864b4e5STejun Heo } else { 48028864b4e5STejun Heo /* 48038864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 48044c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 48054c16bd32STejun Heo * @wq will be freed when the last pwq is released. 48068864b4e5STejun Heo */ 48074c16bd32STejun Heo for_each_node(node) { 48084c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 48094c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 48104c16bd32STejun Heo put_pwq_unlocked(pwq); 48114c16bd32STejun Heo } 48124c16bd32STejun Heo 48134c16bd32STejun Heo /* 48144c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 48154c16bd32STejun Heo * put. Don't access it afterwards. 48164c16bd32STejun Heo */ 48174c16bd32STejun Heo pwq = wq->dfl_pwq; 48184c16bd32STejun Heo wq->dfl_pwq = NULL; 4819dce90d47STejun Heo put_pwq_unlocked(pwq); 482029c91e99STejun Heo } 48213af24433SOleg Nesterov } 48223af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 48233af24433SOleg Nesterov 4824dcd989cbSTejun Heo /** 4825dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4826dcd989cbSTejun Heo * @wq: target workqueue 4827dcd989cbSTejun Heo * @max_active: new max_active value. 4828dcd989cbSTejun Heo * 4829dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4830dcd989cbSTejun Heo * 4831dcd989cbSTejun Heo * CONTEXT: 4832dcd989cbSTejun Heo * Don't call from IRQ context. 4833dcd989cbSTejun Heo */ 4834dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4835dcd989cbSTejun Heo { 483649e3cf44STejun Heo struct pool_workqueue *pwq; 4837dcd989cbSTejun Heo 48388719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 48390a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 48408719dceaSTejun Heo return; 48418719dceaSTejun Heo 4842f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4843dcd989cbSTejun Heo 4844a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4845dcd989cbSTejun Heo 48460a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4847dcd989cbSTejun Heo wq->saved_max_active = max_active; 4848dcd989cbSTejun Heo 4849699ce097STejun Heo for_each_pwq(pwq, wq) 4850699ce097STejun Heo pwq_adjust_max_active(pwq); 4851dcd989cbSTejun Heo 4852a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4853dcd989cbSTejun Heo } 4854dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4855dcd989cbSTejun Heo 4856dcd989cbSTejun Heo /** 485727d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 485827d4ee03SLukas Wunner * 485927d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 486027d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 486127d4ee03SLukas Wunner * 486227d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 486327d4ee03SLukas Wunner */ 486427d4ee03SLukas Wunner struct work_struct *current_work(void) 486527d4ee03SLukas Wunner { 486627d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 486727d4ee03SLukas Wunner 486827d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 486927d4ee03SLukas Wunner } 487027d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 487127d4ee03SLukas Wunner 487227d4ee03SLukas Wunner /** 4873e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4874e6267616STejun Heo * 4875e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4876e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4877d185af30SYacine Belkadi * 4878d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4879e6267616STejun Heo */ 4880e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4881e6267616STejun Heo { 4882e6267616STejun Heo struct worker *worker = current_wq_worker(); 4883e6267616STejun Heo 48846a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4885e6267616STejun Heo } 4886e6267616STejun Heo 4887e6267616STejun Heo /** 4888dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4889dcd989cbSTejun Heo * @cpu: CPU in question 4890dcd989cbSTejun Heo * @wq: target workqueue 4891dcd989cbSTejun Heo * 4892dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4893dcd989cbSTejun Heo * no synchronization around this function and the test result is 4894dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4895dcd989cbSTejun Heo * 4896d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4897d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4898d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4899d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4900d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4901d3251859STejun Heo * 4902d185af30SYacine Belkadi * Return: 4903dcd989cbSTejun Heo * %true if congested, %false otherwise. 4904dcd989cbSTejun Heo */ 4905d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4906dcd989cbSTejun Heo { 49077fb98ea7STejun Heo struct pool_workqueue *pwq; 490876af4d93STejun Heo bool ret; 490976af4d93STejun Heo 491024acfb71SThomas Gleixner rcu_read_lock(); 491124acfb71SThomas Gleixner preempt_disable(); 49127fb98ea7STejun Heo 4913d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4914d3251859STejun Heo cpu = smp_processor_id(); 4915d3251859STejun Heo 49167fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 49177fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 49187fb98ea7STejun Heo else 4919df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4920dcd989cbSTejun Heo 4921f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 492224acfb71SThomas Gleixner preempt_enable(); 492324acfb71SThomas Gleixner rcu_read_unlock(); 492476af4d93STejun Heo 492576af4d93STejun Heo return ret; 4926dcd989cbSTejun Heo } 4927dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4928dcd989cbSTejun Heo 4929dcd989cbSTejun Heo /** 4930dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4931dcd989cbSTejun Heo * @work: the work to be tested 4932dcd989cbSTejun Heo * 4933dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4934dcd989cbSTejun Heo * synchronization around this function and the test result is 4935dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4936dcd989cbSTejun Heo * 4937d185af30SYacine Belkadi * Return: 4938dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4939dcd989cbSTejun Heo */ 4940dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4941dcd989cbSTejun Heo { 4942fa1b54e6STejun Heo struct worker_pool *pool; 4943dcd989cbSTejun Heo unsigned long flags; 4944dcd989cbSTejun Heo unsigned int ret = 0; 4945dcd989cbSTejun Heo 4946dcd989cbSTejun Heo if (work_pending(work)) 4947dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4948038366c5SLai Jiangshan 494924acfb71SThomas Gleixner rcu_read_lock(); 4950fa1b54e6STejun Heo pool = get_work_pool(work); 4951038366c5SLai Jiangshan if (pool) { 4952a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4953c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4954dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4955a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4956038366c5SLai Jiangshan } 495724acfb71SThomas Gleixner rcu_read_unlock(); 4958dcd989cbSTejun Heo 4959dcd989cbSTejun Heo return ret; 4960dcd989cbSTejun Heo } 4961dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4962dcd989cbSTejun Heo 49633d1cb205STejun Heo /** 49643d1cb205STejun Heo * set_worker_desc - set description for the current work item 49653d1cb205STejun Heo * @fmt: printf-style format string 49663d1cb205STejun Heo * @...: arguments for the format string 49673d1cb205STejun Heo * 49683d1cb205STejun Heo * This function can be called by a running work function to describe what 49693d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 49703d1cb205STejun Heo * information will be printed out together to help debugging. The 49713d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 49723d1cb205STejun Heo */ 49733d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 49743d1cb205STejun Heo { 49753d1cb205STejun Heo struct worker *worker = current_wq_worker(); 49763d1cb205STejun Heo va_list args; 49773d1cb205STejun Heo 49783d1cb205STejun Heo if (worker) { 49793d1cb205STejun Heo va_start(args, fmt); 49803d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 49813d1cb205STejun Heo va_end(args); 49823d1cb205STejun Heo } 49833d1cb205STejun Heo } 49845c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 49853d1cb205STejun Heo 49863d1cb205STejun Heo /** 49873d1cb205STejun Heo * print_worker_info - print out worker information and description 49883d1cb205STejun Heo * @log_lvl: the log level to use when printing 49893d1cb205STejun Heo * @task: target task 49903d1cb205STejun Heo * 49913d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 49923d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 49933d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 49943d1cb205STejun Heo * 49953d1cb205STejun Heo * This function can be safely called on any task as long as the 49963d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 49973d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 49983d1cb205STejun Heo */ 49993d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 50003d1cb205STejun Heo { 50013d1cb205STejun Heo work_func_t *fn = NULL; 50023d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 50033d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 50043d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 50053d1cb205STejun Heo struct workqueue_struct *wq = NULL; 50063d1cb205STejun Heo struct worker *worker; 50073d1cb205STejun Heo 50083d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 50093d1cb205STejun Heo return; 50103d1cb205STejun Heo 50113d1cb205STejun Heo /* 50123d1cb205STejun Heo * This function is called without any synchronization and @task 50133d1cb205STejun Heo * could be in any state. Be careful with dereferences. 50143d1cb205STejun Heo */ 5015e700591aSPetr Mladek worker = kthread_probe_data(task); 50163d1cb205STejun Heo 50173d1cb205STejun Heo /* 50188bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 50198bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 50203d1cb205STejun Heo */ 5021fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5022fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5023fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5024fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5025fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 50263d1cb205STejun Heo 50273d1cb205STejun Heo if (fn || name[0] || desc[0]) { 5028d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 50298bf89593STejun Heo if (strcmp(name, desc)) 50303d1cb205STejun Heo pr_cont(" (%s)", desc); 50313d1cb205STejun Heo pr_cont("\n"); 50323d1cb205STejun Heo } 50333d1cb205STejun Heo } 50343d1cb205STejun Heo 50353494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 50363494fc30STejun Heo { 50373494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 50383494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 50393494fc30STejun Heo pr_cont(" node=%d", pool->node); 50403494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 50413494fc30STejun Heo } 50423494fc30STejun Heo 5043c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 5044c76feb0dSPaul E. McKenney bool comma; 5045c76feb0dSPaul E. McKenney work_func_t func; 5046c76feb0dSPaul E. McKenney long ctr; 5047c76feb0dSPaul E. McKenney }; 5048c76feb0dSPaul E. McKenney 5049c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5050c76feb0dSPaul E. McKenney { 5051c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 5052c76feb0dSPaul E. McKenney goto out_record; 5053c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 5054c76feb0dSPaul E. McKenney pcwsp->ctr++; 5055c76feb0dSPaul E. McKenney return; 5056c76feb0dSPaul E. McKenney } 5057c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 5058c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5059c76feb0dSPaul E. McKenney else 5060c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5061c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 5062c76feb0dSPaul E. McKenney out_record: 5063c76feb0dSPaul E. McKenney if ((long)func == -1L) 5064c76feb0dSPaul E. McKenney return; 5065c76feb0dSPaul E. McKenney pcwsp->comma = comma; 5066c76feb0dSPaul E. McKenney pcwsp->func = func; 5067c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 5068c76feb0dSPaul E. McKenney } 5069c76feb0dSPaul E. McKenney 5070c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 50713494fc30STejun Heo { 50723494fc30STejun Heo if (work->func == wq_barrier_func) { 50733494fc30STejun Heo struct wq_barrier *barr; 50743494fc30STejun Heo 50753494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 50763494fc30STejun Heo 5077c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 50783494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 50793494fc30STejun Heo task_pid_nr(barr->task)); 50803494fc30STejun Heo } else { 5081c76feb0dSPaul E. McKenney if (!comma) 5082c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5083c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 50843494fc30STejun Heo } 50853494fc30STejun Heo } 50863494fc30STejun Heo 50873494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 50883494fc30STejun Heo { 5089c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 50903494fc30STejun Heo struct worker_pool *pool = pwq->pool; 50913494fc30STejun Heo struct work_struct *work; 50923494fc30STejun Heo struct worker *worker; 50933494fc30STejun Heo bool has_in_flight = false, has_pending = false; 50943494fc30STejun Heo int bkt; 50953494fc30STejun Heo 50963494fc30STejun Heo pr_info(" pwq %d:", pool->id); 50973494fc30STejun Heo pr_cont_pool_info(pool); 50983494fc30STejun Heo 5099e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5100e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 51013494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 51023494fc30STejun Heo 51033494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51043494fc30STejun Heo if (worker->current_pwq == pwq) { 51053494fc30STejun Heo has_in_flight = true; 51063494fc30STejun Heo break; 51073494fc30STejun Heo } 51083494fc30STejun Heo } 51093494fc30STejun Heo if (has_in_flight) { 51103494fc30STejun Heo bool comma = false; 51113494fc30STejun Heo 51123494fc30STejun Heo pr_info(" in-flight:"); 51133494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51143494fc30STejun Heo if (worker->current_pwq != pwq) 51153494fc30STejun Heo continue; 51163494fc30STejun Heo 5117d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 51183494fc30STejun Heo task_pid_nr(worker->task), 511930ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 51203494fc30STejun Heo worker->current_func); 51213494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5122c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5123c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51243494fc30STejun Heo comma = true; 51253494fc30STejun Heo } 51263494fc30STejun Heo pr_cont("\n"); 51273494fc30STejun Heo } 51283494fc30STejun Heo 51293494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51303494fc30STejun Heo if (get_work_pwq(work) == pwq) { 51313494fc30STejun Heo has_pending = true; 51323494fc30STejun Heo break; 51333494fc30STejun Heo } 51343494fc30STejun Heo } 51353494fc30STejun Heo if (has_pending) { 51363494fc30STejun Heo bool comma = false; 51373494fc30STejun Heo 51383494fc30STejun Heo pr_info(" pending:"); 51393494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51403494fc30STejun Heo if (get_work_pwq(work) != pwq) 51413494fc30STejun Heo continue; 51423494fc30STejun Heo 5143c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51443494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51453494fc30STejun Heo } 5146c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51473494fc30STejun Heo pr_cont("\n"); 51483494fc30STejun Heo } 51493494fc30STejun Heo 5150f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 51513494fc30STejun Heo bool comma = false; 51523494fc30STejun Heo 5153f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5154f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5155c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51563494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51573494fc30STejun Heo } 5158c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51593494fc30STejun Heo pr_cont("\n"); 51603494fc30STejun Heo } 51613494fc30STejun Heo } 51623494fc30STejun Heo 51633494fc30STejun Heo /** 516455df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 516555df0933SImran Khan * @wq: workqueue whose state will be printed 51663494fc30STejun Heo */ 516755df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 51683494fc30STejun Heo { 51693494fc30STejun Heo struct pool_workqueue *pwq; 51703494fc30STejun Heo bool idle = true; 517155df0933SImran Khan unsigned long flags; 51723494fc30STejun Heo 51733494fc30STejun Heo for_each_pwq(pwq, wq) { 5174f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 51753494fc30STejun Heo idle = false; 51763494fc30STejun Heo break; 51773494fc30STejun Heo } 51783494fc30STejun Heo } 517955df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 518055df0933SImran Khan return; 51813494fc30STejun Heo 51823494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 51833494fc30STejun Heo 51843494fc30STejun Heo for_each_pwq(pwq, wq) { 5185a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 518657116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 518757116ce1SJohan Hovold /* 518857116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 518957116ce1SJohan Hovold * drivers that queue work while holding locks 519057116ce1SJohan Hovold * also taken in their write paths. 519157116ce1SJohan Hovold */ 519257116ce1SJohan Hovold printk_deferred_enter(); 51933494fc30STejun Heo show_pwq(pwq); 519457116ce1SJohan Hovold printk_deferred_exit(); 519557116ce1SJohan Hovold } 5196a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 519762635ea8SSergey Senozhatsky /* 519862635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 519955df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 520062635ea8SSergey Senozhatsky * hard lockup. 520162635ea8SSergey Senozhatsky */ 520262635ea8SSergey Senozhatsky touch_nmi_watchdog(); 52033494fc30STejun Heo } 520455df0933SImran Khan 52053494fc30STejun Heo } 52063494fc30STejun Heo 520755df0933SImran Khan /** 520855df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 520955df0933SImran Khan * @pool: worker pool whose state will be printed 521055df0933SImran Khan */ 521155df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 521255df0933SImran Khan { 52133494fc30STejun Heo struct worker *worker; 52143494fc30STejun Heo bool first = true; 521555df0933SImran Khan unsigned long flags; 5216335a42ebSPetr Mladek unsigned long hung = 0; 52173494fc30STejun Heo 5218a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 52193494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 52203494fc30STejun Heo goto next_pool; 5221335a42ebSPetr Mladek 5222335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5223335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5224335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5225335a42ebSPetr Mladek 522657116ce1SJohan Hovold /* 522757116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 522857116ce1SJohan Hovold * queue work while holding locks also taken in their write 522957116ce1SJohan Hovold * paths. 523057116ce1SJohan Hovold */ 523157116ce1SJohan Hovold printk_deferred_enter(); 52323494fc30STejun Heo pr_info("pool %d:", pool->id); 52333494fc30STejun Heo pr_cont_pool_info(pool); 5234335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 52353494fc30STejun Heo if (pool->manager) 52363494fc30STejun Heo pr_cont(" manager: %d", 52373494fc30STejun Heo task_pid_nr(pool->manager->task)); 52383494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 52393494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 52403494fc30STejun Heo task_pid_nr(worker->task)); 52413494fc30STejun Heo first = false; 52423494fc30STejun Heo } 52433494fc30STejun Heo pr_cont("\n"); 524457116ce1SJohan Hovold printk_deferred_exit(); 52453494fc30STejun Heo next_pool: 5246a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 524762635ea8SSergey Senozhatsky /* 524862635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 524955df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 525062635ea8SSergey Senozhatsky * hard lockup. 525162635ea8SSergey Senozhatsky */ 525262635ea8SSergey Senozhatsky touch_nmi_watchdog(); 525355df0933SImran Khan 52543494fc30STejun Heo } 52553494fc30STejun Heo 525655df0933SImran Khan /** 525755df0933SImran Khan * show_all_workqueues - dump workqueue state 525855df0933SImran Khan * 5259704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 526055df0933SImran Khan */ 526155df0933SImran Khan void show_all_workqueues(void) 526255df0933SImran Khan { 526355df0933SImran Khan struct workqueue_struct *wq; 526455df0933SImran Khan struct worker_pool *pool; 526555df0933SImran Khan int pi; 526655df0933SImran Khan 526755df0933SImran Khan rcu_read_lock(); 526855df0933SImran Khan 526955df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 527055df0933SImran Khan 527155df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 527255df0933SImran Khan show_one_workqueue(wq); 527355df0933SImran Khan 527455df0933SImran Khan for_each_pool(pool, pi) 527555df0933SImran Khan show_one_worker_pool(pool); 527655df0933SImran Khan 527724acfb71SThomas Gleixner rcu_read_unlock(); 52783494fc30STejun Heo } 52793494fc30STejun Heo 5280704bc669SJungseung Lee /** 5281704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5282704bc669SJungseung Lee * 5283704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5284704bc669SJungseung Lee * still busy. 5285704bc669SJungseung Lee */ 5286704bc669SJungseung Lee void show_freezable_workqueues(void) 5287704bc669SJungseung Lee { 5288704bc669SJungseung Lee struct workqueue_struct *wq; 5289704bc669SJungseung Lee 5290704bc669SJungseung Lee rcu_read_lock(); 5291704bc669SJungseung Lee 5292704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5293704bc669SJungseung Lee 5294704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5295704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5296704bc669SJungseung Lee continue; 5297704bc669SJungseung Lee show_one_workqueue(wq); 5298704bc669SJungseung Lee } 5299704bc669SJungseung Lee 5300704bc669SJungseung Lee rcu_read_unlock(); 5301704bc669SJungseung Lee } 5302704bc669SJungseung Lee 53036b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 53046b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 53056b59808bSTejun Heo { 53066b59808bSTejun Heo int off; 53076b59808bSTejun Heo 53086b59808bSTejun Heo /* always show the actual comm */ 53096b59808bSTejun Heo off = strscpy(buf, task->comm, size); 53106b59808bSTejun Heo if (off < 0) 53116b59808bSTejun Heo return; 53126b59808bSTejun Heo 5313197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 53146b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 53156b59808bSTejun Heo 5316197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5317197f6accSTejun Heo struct worker *worker = kthread_data(task); 5318197f6accSTejun Heo struct worker_pool *pool = worker->pool; 53196b59808bSTejun Heo 53206b59808bSTejun Heo if (pool) { 5321a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 53226b59808bSTejun Heo /* 5323197f6accSTejun Heo * ->desc tracks information (wq name or 5324197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5325197f6accSTejun Heo * current, prepend '+', otherwise '-'. 53266b59808bSTejun Heo */ 53276b59808bSTejun Heo if (worker->desc[0] != '\0') { 53286b59808bSTejun Heo if (worker->current_work) 53296b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 53306b59808bSTejun Heo worker->desc); 53316b59808bSTejun Heo else 53326b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 53336b59808bSTejun Heo worker->desc); 53346b59808bSTejun Heo } 5335a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 53366b59808bSTejun Heo } 5337197f6accSTejun Heo } 53386b59808bSTejun Heo 53396b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 53406b59808bSTejun Heo } 53416b59808bSTejun Heo 534266448bc2SMathieu Malaterre #ifdef CONFIG_SMP 534366448bc2SMathieu Malaterre 5344db7bccf4STejun Heo /* 5345db7bccf4STejun Heo * CPU hotplug. 5346db7bccf4STejun Heo * 5347e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5348112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5349706026c2STejun Heo * pool which make migrating pending and scheduled works very 5350e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 535194cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5352e22bee78STejun Heo * blocked draining impractical. 5353e22bee78STejun Heo * 535424647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5355628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5356628c78e7STejun Heo * cpu comes back online. 5357db7bccf4STejun Heo */ 5358db7bccf4STejun Heo 5359e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5360db7bccf4STejun Heo { 53614ce62e9eSTejun Heo struct worker_pool *pool; 5362db7bccf4STejun Heo struct worker *worker; 5363db7bccf4STejun Heo 5364f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53651258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5366a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5367e22bee78STejun Heo 5368f2d5a0eeSTejun Heo /* 536992f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 537094cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 537111b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5372b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5373b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5374b4ac9384SLai Jiangshan * is on the same cpu. 5375f2d5a0eeSTejun Heo */ 5376da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5377403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5378db7bccf4STejun Heo 537924647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5380f2d5a0eeSTejun Heo 5381e22bee78STejun Heo /* 5382989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5383989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5384989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5385989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5386989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5387eb283428SLai Jiangshan * are served by workers tied to the pool. 5388e22bee78STejun Heo */ 5389bc35f7efSLai Jiangshan pool->nr_running = 0; 5390eb283428SLai Jiangshan 5391eb283428SLai Jiangshan /* 5392eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5393eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5394eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5395eb283428SLai Jiangshan */ 5396eb283428SLai Jiangshan wake_up_worker(pool); 5397989442d7SLai Jiangshan 5398a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5399989442d7SLai Jiangshan 5400793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5401793777bcSValentin Schneider unbind_worker(worker); 5402989442d7SLai Jiangshan 5403989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5404eb283428SLai Jiangshan } 5405db7bccf4STejun Heo } 5406db7bccf4STejun Heo 5407bd7c089eSTejun Heo /** 5408bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5409bd7c089eSTejun Heo * @pool: pool of interest 5410bd7c089eSTejun Heo * 5411a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5412bd7c089eSTejun Heo */ 5413bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5414bd7c089eSTejun Heo { 5415a9ab775bSTejun Heo struct worker *worker; 5416bd7c089eSTejun Heo 54171258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5418bd7c089eSTejun Heo 5419bd7c089eSTejun Heo /* 5420a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5421a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5422402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5423a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5424a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5425bd7c089eSTejun Heo */ 5426c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5427c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5428c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5429c63a2e52SValentin Schneider pool->attrs->cpumask) < 0); 5430c63a2e52SValentin Schneider } 5431a9ab775bSTejun Heo 5432a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5433f7c17d26SWanpeng Li 54343de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5435a9ab775bSTejun Heo 5436da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5437a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5438a9ab775bSTejun Heo 5439a9ab775bSTejun Heo /* 5440a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5441a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5442a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5443a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5444a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5445a9ab775bSTejun Heo * concurrency management. Note that when or whether 5446a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5447a9ab775bSTejun Heo * 5448c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5449a9ab775bSTejun Heo * tested without holding any lock in 54506d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5451a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5452a9ab775bSTejun Heo * management operations. 5453bd7c089eSTejun Heo */ 5454a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5455a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5456a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5457c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5458bd7c089eSTejun Heo } 5459a9ab775bSTejun Heo 5460a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5461bd7c089eSTejun Heo } 5462bd7c089eSTejun Heo 54637dbc725eSTejun Heo /** 54647dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 54657dbc725eSTejun Heo * @pool: unbound pool of interest 54667dbc725eSTejun Heo * @cpu: the CPU which is coming up 54677dbc725eSTejun Heo * 54687dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 54697dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 54707dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 54717dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 54727dbc725eSTejun Heo */ 54737dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 54747dbc725eSTejun Heo { 54757dbc725eSTejun Heo static cpumask_t cpumask; 54767dbc725eSTejun Heo struct worker *worker; 54777dbc725eSTejun Heo 54781258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 54797dbc725eSTejun Heo 54807dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 54817dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 54827dbc725eSTejun Heo return; 54837dbc725eSTejun Heo 54847dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 54857dbc725eSTejun Heo 54867dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5487da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5488d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 54897dbc725eSTejun Heo } 54907dbc725eSTejun Heo 54917ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 54921da177e4SLinus Torvalds { 54934ce62e9eSTejun Heo struct worker_pool *pool; 54941da177e4SLinus Torvalds 5495f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54963ce63377STejun Heo if (pool->nr_workers) 54973ce63377STejun Heo continue; 5498051e1850SLai Jiangshan if (!create_worker(pool)) 54997ee681b2SThomas Gleixner return -ENOMEM; 55003af24433SOleg Nesterov } 55017ee681b2SThomas Gleixner return 0; 55027ee681b2SThomas Gleixner } 55031da177e4SLinus Torvalds 55047ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 55057ee681b2SThomas Gleixner { 55067ee681b2SThomas Gleixner struct worker_pool *pool; 55077ee681b2SThomas Gleixner struct workqueue_struct *wq; 55087ee681b2SThomas Gleixner int pi; 55097ee681b2SThomas Gleixner 551068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55117dbc725eSTejun Heo 55127dbc725eSTejun Heo for_each_pool(pool, pi) { 55131258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 551494cf58bbSTejun Heo 5515f05b558dSLai Jiangshan if (pool->cpu == cpu) 551694cf58bbSTejun Heo rebind_workers(pool); 5517f05b558dSLai Jiangshan else if (pool->cpu < 0) 55187dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 551994cf58bbSTejun Heo 55201258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 552194cf58bbSTejun Heo } 55227dbc725eSTejun Heo 55234c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55244c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 55254c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 55264c16bd32STejun Heo 552768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55287ee681b2SThomas Gleixner return 0; 552965758202STejun Heo } 553065758202STejun Heo 55317ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 553265758202STejun Heo { 55334c16bd32STejun Heo struct workqueue_struct *wq; 55348db25e78STejun Heo 55354c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5536e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5537e8b3f8dbSLai Jiangshan return -1; 5538e8b3f8dbSLai Jiangshan 5539e8b3f8dbSLai Jiangshan unbind_workers(cpu); 55404c16bd32STejun Heo 55414c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55424c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 55434c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 55444c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 55454c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 55464c16bd32STejun Heo 55477ee681b2SThomas Gleixner return 0; 554865758202STejun Heo } 554965758202STejun Heo 55502d3854a3SRusty Russell struct work_for_cpu { 5551ed48ece2STejun Heo struct work_struct work; 55522d3854a3SRusty Russell long (*fn)(void *); 55532d3854a3SRusty Russell void *arg; 55542d3854a3SRusty Russell long ret; 55552d3854a3SRusty Russell }; 55562d3854a3SRusty Russell 5557ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 55582d3854a3SRusty Russell { 5559ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5560ed48ece2STejun Heo 55612d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 55622d3854a3SRusty Russell } 55632d3854a3SRusty Russell 55642d3854a3SRusty Russell /** 556522aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 55662d3854a3SRusty Russell * @cpu: the cpu to run on 55672d3854a3SRusty Russell * @fn: the function to run 55682d3854a3SRusty Russell * @arg: the function arg 55692d3854a3SRusty Russell * 557031ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 55716b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5572d185af30SYacine Belkadi * 5573d185af30SYacine Belkadi * Return: The value @fn returns. 55742d3854a3SRusty Russell */ 5575d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 55762d3854a3SRusty Russell { 5577ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 55782d3854a3SRusty Russell 5579ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5580ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 558112997d1aSBjorn Helgaas flush_work(&wfc.work); 5582440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 55832d3854a3SRusty Russell return wfc.ret; 55842d3854a3SRusty Russell } 55852d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 55860e8d6a93SThomas Gleixner 55870e8d6a93SThomas Gleixner /** 55880e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 55890e8d6a93SThomas Gleixner * @cpu: the cpu to run on 55900e8d6a93SThomas Gleixner * @fn: the function to run 55910e8d6a93SThomas Gleixner * @arg: the function argument 55920e8d6a93SThomas Gleixner * 55930e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 55940e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 55950e8d6a93SThomas Gleixner * 55960e8d6a93SThomas Gleixner * Return: The value @fn returns. 55970e8d6a93SThomas Gleixner */ 55980e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 55990e8d6a93SThomas Gleixner { 56000e8d6a93SThomas Gleixner long ret = -ENODEV; 56010e8d6a93SThomas Gleixner 5602ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 56030e8d6a93SThomas Gleixner if (cpu_online(cpu)) 56040e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5605ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 56060e8d6a93SThomas Gleixner return ret; 56070e8d6a93SThomas Gleixner } 56080e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 56092d3854a3SRusty Russell #endif /* CONFIG_SMP */ 56102d3854a3SRusty Russell 5611a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5612e7577c50SRusty Russell 5613a0a1a5fdSTejun Heo /** 5614a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5615a0a1a5fdSTejun Heo * 561658a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5617f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5618706026c2STejun Heo * pool->worklist. 5619a0a1a5fdSTejun Heo * 5620a0a1a5fdSTejun Heo * CONTEXT: 5621a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5622a0a1a5fdSTejun Heo */ 5623a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5624a0a1a5fdSTejun Heo { 562524b8a847STejun Heo struct workqueue_struct *wq; 562624b8a847STejun Heo struct pool_workqueue *pwq; 5627a0a1a5fdSTejun Heo 562868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5629a0a1a5fdSTejun Heo 56306183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5631a0a1a5fdSTejun Heo workqueue_freezing = true; 5632a0a1a5fdSTejun Heo 563324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5634a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5635699ce097STejun Heo for_each_pwq(pwq, wq) 5636699ce097STejun Heo pwq_adjust_max_active(pwq); 5637a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5638a1056305STejun Heo } 56395bcab335STejun Heo 564068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5641a0a1a5fdSTejun Heo } 5642a0a1a5fdSTejun Heo 5643a0a1a5fdSTejun Heo /** 564458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5645a0a1a5fdSTejun Heo * 5646a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5647a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5648a0a1a5fdSTejun Heo * 5649a0a1a5fdSTejun Heo * CONTEXT: 565068e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5651a0a1a5fdSTejun Heo * 5652d185af30SYacine Belkadi * Return: 565358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 565458a69cb4STejun Heo * is complete. 5655a0a1a5fdSTejun Heo */ 5656a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5657a0a1a5fdSTejun Heo { 5658a0a1a5fdSTejun Heo bool busy = false; 565924b8a847STejun Heo struct workqueue_struct *wq; 566024b8a847STejun Heo struct pool_workqueue *pwq; 5661a0a1a5fdSTejun Heo 566268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5663a0a1a5fdSTejun Heo 56646183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5665a0a1a5fdSTejun Heo 566624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 566724b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 566824b8a847STejun Heo continue; 5669a0a1a5fdSTejun Heo /* 5670a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5671a0a1a5fdSTejun Heo * to peek without lock. 5672a0a1a5fdSTejun Heo */ 567324acfb71SThomas Gleixner rcu_read_lock(); 567424b8a847STejun Heo for_each_pwq(pwq, wq) { 56756183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5676112202d9STejun Heo if (pwq->nr_active) { 5677a0a1a5fdSTejun Heo busy = true; 567824acfb71SThomas Gleixner rcu_read_unlock(); 5679a0a1a5fdSTejun Heo goto out_unlock; 5680a0a1a5fdSTejun Heo } 5681a0a1a5fdSTejun Heo } 568224acfb71SThomas Gleixner rcu_read_unlock(); 5683a0a1a5fdSTejun Heo } 5684a0a1a5fdSTejun Heo out_unlock: 568568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5686a0a1a5fdSTejun Heo return busy; 5687a0a1a5fdSTejun Heo } 5688a0a1a5fdSTejun Heo 5689a0a1a5fdSTejun Heo /** 5690a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5691a0a1a5fdSTejun Heo * 5692a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5693706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5694a0a1a5fdSTejun Heo * 5695a0a1a5fdSTejun Heo * CONTEXT: 5696a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5697a0a1a5fdSTejun Heo */ 5698a0a1a5fdSTejun Heo void thaw_workqueues(void) 5699a0a1a5fdSTejun Heo { 570024b8a847STejun Heo struct workqueue_struct *wq; 570124b8a847STejun Heo struct pool_workqueue *pwq; 5702a0a1a5fdSTejun Heo 570368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5704a0a1a5fdSTejun Heo 5705a0a1a5fdSTejun Heo if (!workqueue_freezing) 5706a0a1a5fdSTejun Heo goto out_unlock; 5707a0a1a5fdSTejun Heo 570874b414eaSLai Jiangshan workqueue_freezing = false; 570924b8a847STejun Heo 571024b8a847STejun Heo /* restore max_active and repopulate worklist */ 571124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5712a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5713699ce097STejun Heo for_each_pwq(pwq, wq) 5714699ce097STejun Heo pwq_adjust_max_active(pwq); 5715a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 571624b8a847STejun Heo } 571724b8a847STejun Heo 5718a0a1a5fdSTejun Heo out_unlock: 571968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5720a0a1a5fdSTejun Heo } 5721a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5722a0a1a5fdSTejun Heo 572399c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5724042f7df1SLai Jiangshan { 5725042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5726042f7df1SLai Jiangshan int ret = 0; 5727042f7df1SLai Jiangshan struct workqueue_struct *wq; 5728042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5729042f7df1SLai Jiangshan 5730042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5731042f7df1SLai Jiangshan 5732042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5733042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5734042f7df1SLai Jiangshan continue; 5735042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5736042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5737042f7df1SLai Jiangshan continue; 5738042f7df1SLai Jiangshan 573999c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5740042f7df1SLai Jiangshan if (!ctx) { 5741042f7df1SLai Jiangshan ret = -ENOMEM; 5742042f7df1SLai Jiangshan break; 5743042f7df1SLai Jiangshan } 5744042f7df1SLai Jiangshan 5745042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5746042f7df1SLai Jiangshan } 5747042f7df1SLai Jiangshan 5748042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5749042f7df1SLai Jiangshan if (!ret) 5750042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5751042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5752042f7df1SLai Jiangshan } 5753042f7df1SLai Jiangshan 575499c621efSLai Jiangshan if (!ret) { 575599c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 575699c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 575799c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 575899c621efSLai Jiangshan } 5759042f7df1SLai Jiangshan return ret; 5760042f7df1SLai Jiangshan } 5761042f7df1SLai Jiangshan 5762042f7df1SLai Jiangshan /** 5763042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5764042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5765042f7df1SLai Jiangshan * 5766042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5767042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5768042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5769042f7df1SLai Jiangshan * 577067dc8325SCai Huoqing * Return: 0 - Success 5771042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5772042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5773042f7df1SLai Jiangshan */ 5774042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5775042f7df1SLai Jiangshan { 5776042f7df1SLai Jiangshan int ret = -EINVAL; 5777042f7df1SLai Jiangshan 5778c98a9805STal Shorer /* 5779c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5780c98a9805STal Shorer * If the user wishes to include them, we allow that. 5781c98a9805STal Shorer */ 5782042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5783042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5784a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5785d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5786d25302e4SMenglong Dong ret = 0; 5787d25302e4SMenglong Dong goto out_unlock; 5788d25302e4SMenglong Dong } 5789d25302e4SMenglong Dong 579099c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5791042f7df1SLai Jiangshan 5792d25302e4SMenglong Dong out_unlock: 5793a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5794042f7df1SLai Jiangshan } 5795042f7df1SLai Jiangshan 5796042f7df1SLai Jiangshan return ret; 5797042f7df1SLai Jiangshan } 5798042f7df1SLai Jiangshan 57996ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 58006ba94429SFrederic Weisbecker /* 58016ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 58026ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 58036ba94429SFrederic Weisbecker * following attributes. 58046ba94429SFrederic Weisbecker * 58056ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 58066ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 58076ba94429SFrederic Weisbecker * 58086ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 58096ba94429SFrederic Weisbecker * 58109a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 58116ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 58126ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 58139a19b463SWang Long * numa RW bool : whether enable NUMA affinity 58146ba94429SFrederic Weisbecker */ 58156ba94429SFrederic Weisbecker struct wq_device { 58166ba94429SFrederic Weisbecker struct workqueue_struct *wq; 58176ba94429SFrederic Weisbecker struct device dev; 58186ba94429SFrederic Weisbecker }; 58196ba94429SFrederic Weisbecker 58206ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 58216ba94429SFrederic Weisbecker { 58226ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 58236ba94429SFrederic Weisbecker 58246ba94429SFrederic Weisbecker return wq_dev->wq; 58256ba94429SFrederic Weisbecker } 58266ba94429SFrederic Weisbecker 58276ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 58286ba94429SFrederic Weisbecker char *buf) 58296ba94429SFrederic Weisbecker { 58306ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58316ba94429SFrederic Weisbecker 58326ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 58336ba94429SFrederic Weisbecker } 58346ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 58356ba94429SFrederic Weisbecker 58366ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 58376ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58386ba94429SFrederic Weisbecker { 58396ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58406ba94429SFrederic Weisbecker 58416ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 58426ba94429SFrederic Weisbecker } 58436ba94429SFrederic Weisbecker 58446ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 58456ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 58466ba94429SFrederic Weisbecker size_t count) 58476ba94429SFrederic Weisbecker { 58486ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58496ba94429SFrederic Weisbecker int val; 58506ba94429SFrederic Weisbecker 58516ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 58526ba94429SFrederic Weisbecker return -EINVAL; 58536ba94429SFrederic Weisbecker 58546ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 58556ba94429SFrederic Weisbecker return count; 58566ba94429SFrederic Weisbecker } 58576ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 58586ba94429SFrederic Weisbecker 58596ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 58606ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 58616ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 58626ba94429SFrederic Weisbecker NULL, 58636ba94429SFrederic Weisbecker }; 58646ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 58656ba94429SFrederic Weisbecker 58666ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 58676ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58686ba94429SFrederic Weisbecker { 58696ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58706ba94429SFrederic Weisbecker const char *delim = ""; 58716ba94429SFrederic Weisbecker int node, written = 0; 58726ba94429SFrederic Weisbecker 5873ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 587424acfb71SThomas Gleixner rcu_read_lock(); 58756ba94429SFrederic Weisbecker for_each_node(node) { 58766ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 58776ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 58786ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 58796ba94429SFrederic Weisbecker delim = " "; 58806ba94429SFrederic Weisbecker } 58816ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 588224acfb71SThomas Gleixner rcu_read_unlock(); 5883ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 58846ba94429SFrederic Weisbecker 58856ba94429SFrederic Weisbecker return written; 58866ba94429SFrederic Weisbecker } 58876ba94429SFrederic Weisbecker 58886ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 58896ba94429SFrederic Weisbecker char *buf) 58906ba94429SFrederic Weisbecker { 58916ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58926ba94429SFrederic Weisbecker int written; 58936ba94429SFrederic Weisbecker 58946ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 58956ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 58966ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 58976ba94429SFrederic Weisbecker 58986ba94429SFrederic Weisbecker return written; 58996ba94429SFrederic Weisbecker } 59006ba94429SFrederic Weisbecker 59016ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 59026ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 59036ba94429SFrederic Weisbecker { 59046ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 59056ba94429SFrederic Weisbecker 5906899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5907899a94feSLai Jiangshan 5908be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 59096ba94429SFrederic Weisbecker if (!attrs) 59106ba94429SFrederic Weisbecker return NULL; 59116ba94429SFrederic Weisbecker 59126ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 59136ba94429SFrederic Weisbecker return attrs; 59146ba94429SFrederic Weisbecker } 59156ba94429SFrederic Weisbecker 59166ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 59176ba94429SFrederic Weisbecker const char *buf, size_t count) 59186ba94429SFrederic Weisbecker { 59196ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59206ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5921d4d3e257SLai Jiangshan int ret = -ENOMEM; 5922d4d3e257SLai Jiangshan 5923d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59246ba94429SFrederic Weisbecker 59256ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59266ba94429SFrederic Weisbecker if (!attrs) 5927d4d3e257SLai Jiangshan goto out_unlock; 59286ba94429SFrederic Weisbecker 59296ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 59306ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5931d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59326ba94429SFrederic Weisbecker else 59336ba94429SFrederic Weisbecker ret = -EINVAL; 59346ba94429SFrederic Weisbecker 5935d4d3e257SLai Jiangshan out_unlock: 5936d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59376ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59386ba94429SFrederic Weisbecker return ret ?: count; 59396ba94429SFrederic Weisbecker } 59406ba94429SFrederic Weisbecker 59416ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 59426ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 59436ba94429SFrederic Weisbecker { 59446ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59456ba94429SFrederic Weisbecker int written; 59466ba94429SFrederic Weisbecker 59476ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59486ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 59496ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 59506ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59516ba94429SFrederic Weisbecker return written; 59526ba94429SFrederic Weisbecker } 59536ba94429SFrederic Weisbecker 59546ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 59556ba94429SFrederic Weisbecker struct device_attribute *attr, 59566ba94429SFrederic Weisbecker const char *buf, size_t count) 59576ba94429SFrederic Weisbecker { 59586ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59596ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5960d4d3e257SLai Jiangshan int ret = -ENOMEM; 5961d4d3e257SLai Jiangshan 5962d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59636ba94429SFrederic Weisbecker 59646ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59656ba94429SFrederic Weisbecker if (!attrs) 5966d4d3e257SLai Jiangshan goto out_unlock; 59676ba94429SFrederic Weisbecker 59686ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 59696ba94429SFrederic Weisbecker if (!ret) 5970d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59716ba94429SFrederic Weisbecker 5972d4d3e257SLai Jiangshan out_unlock: 5973d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59746ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59756ba94429SFrederic Weisbecker return ret ?: count; 59766ba94429SFrederic Weisbecker } 59776ba94429SFrederic Weisbecker 59786ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 59796ba94429SFrederic Weisbecker char *buf) 59806ba94429SFrederic Weisbecker { 59816ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59826ba94429SFrederic Weisbecker int written; 59836ba94429SFrederic Weisbecker 59846ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59856ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 59866ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 59876ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59886ba94429SFrederic Weisbecker 59896ba94429SFrederic Weisbecker return written; 59906ba94429SFrederic Weisbecker } 59916ba94429SFrederic Weisbecker 59926ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 59936ba94429SFrederic Weisbecker const char *buf, size_t count) 59946ba94429SFrederic Weisbecker { 59956ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59966ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5997d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5998d4d3e257SLai Jiangshan 5999d4d3e257SLai Jiangshan apply_wqattrs_lock(); 60006ba94429SFrederic Weisbecker 60016ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 60026ba94429SFrederic Weisbecker if (!attrs) 6003d4d3e257SLai Jiangshan goto out_unlock; 60046ba94429SFrederic Weisbecker 60056ba94429SFrederic Weisbecker ret = -EINVAL; 60066ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 60076ba94429SFrederic Weisbecker attrs->no_numa = !v; 6008d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 60096ba94429SFrederic Weisbecker } 60106ba94429SFrederic Weisbecker 6011d4d3e257SLai Jiangshan out_unlock: 6012d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 60136ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 60146ba94429SFrederic Weisbecker return ret ?: count; 60156ba94429SFrederic Weisbecker } 60166ba94429SFrederic Weisbecker 60176ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 60186ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 60196ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 60206ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 60216ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 60226ba94429SFrederic Weisbecker __ATTR_NULL, 60236ba94429SFrederic Weisbecker }; 60246ba94429SFrederic Weisbecker 60256ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 60266ba94429SFrederic Weisbecker .name = "workqueue", 60276ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 60286ba94429SFrederic Weisbecker }; 60296ba94429SFrederic Weisbecker 6030b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 6031b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 6032b05a7928SFrederic Weisbecker { 6033b05a7928SFrederic Weisbecker int written; 6034b05a7928SFrederic Weisbecker 6035042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 6036b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6037b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 6038042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 6039b05a7928SFrederic Weisbecker 6040b05a7928SFrederic Weisbecker return written; 6041b05a7928SFrederic Weisbecker } 6042b05a7928SFrederic Weisbecker 6043042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 6044042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 6045042f7df1SLai Jiangshan { 6046042f7df1SLai Jiangshan cpumask_var_t cpumask; 6047042f7df1SLai Jiangshan int ret; 6048042f7df1SLai Jiangshan 6049042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6050042f7df1SLai Jiangshan return -ENOMEM; 6051042f7df1SLai Jiangshan 6052042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 6053042f7df1SLai Jiangshan if (!ret) 6054042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 6055042f7df1SLai Jiangshan 6056042f7df1SLai Jiangshan free_cpumask_var(cpumask); 6057042f7df1SLai Jiangshan return ret ? ret : count; 6058042f7df1SLai Jiangshan } 6059042f7df1SLai Jiangshan 6060b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 6061042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6062042f7df1SLai Jiangshan wq_unbound_cpumask_store); 6063b05a7928SFrederic Weisbecker 60646ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 60656ba94429SFrederic Weisbecker { 6066686f6697SGreg Kroah-Hartman struct device *dev_root; 6067b05a7928SFrederic Weisbecker int err; 6068b05a7928SFrederic Weisbecker 6069b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 6070b05a7928SFrederic Weisbecker if (err) 6071b05a7928SFrederic Weisbecker return err; 6072b05a7928SFrederic Weisbecker 6073686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 6074686f6697SGreg Kroah-Hartman if (dev_root) { 6075686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6076686f6697SGreg Kroah-Hartman put_device(dev_root); 6077686f6697SGreg Kroah-Hartman } 6078686f6697SGreg Kroah-Hartman return err; 60796ba94429SFrederic Weisbecker } 60806ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 60816ba94429SFrederic Weisbecker 60826ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 60836ba94429SFrederic Weisbecker { 60846ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 60856ba94429SFrederic Weisbecker 60866ba94429SFrederic Weisbecker kfree(wq_dev); 60876ba94429SFrederic Weisbecker } 60886ba94429SFrederic Weisbecker 60896ba94429SFrederic Weisbecker /** 60906ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 60916ba94429SFrederic Weisbecker * @wq: the workqueue to register 60926ba94429SFrederic Weisbecker * 60936ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 60946ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 60956ba94429SFrederic Weisbecker * which is the preferred method. 60966ba94429SFrederic Weisbecker * 60976ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 60986ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 60996ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 61006ba94429SFrederic Weisbecker * attributes. 61016ba94429SFrederic Weisbecker * 61026ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 61036ba94429SFrederic Weisbecker */ 61046ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 61056ba94429SFrederic Weisbecker { 61066ba94429SFrederic Weisbecker struct wq_device *wq_dev; 61076ba94429SFrederic Weisbecker int ret; 61086ba94429SFrederic Weisbecker 61096ba94429SFrederic Weisbecker /* 6110402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 61116ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 61126ba94429SFrederic Weisbecker * workqueues. 61136ba94429SFrederic Weisbecker */ 61140a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 61156ba94429SFrederic Weisbecker return -EINVAL; 61166ba94429SFrederic Weisbecker 61176ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 61186ba94429SFrederic Weisbecker if (!wq_dev) 61196ba94429SFrederic Weisbecker return -ENOMEM; 61206ba94429SFrederic Weisbecker 61216ba94429SFrederic Weisbecker wq_dev->wq = wq; 61226ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 61236ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 612423217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 61256ba94429SFrederic Weisbecker 61266ba94429SFrederic Weisbecker /* 61276ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 61286ba94429SFrederic Weisbecker * everything is ready. 61296ba94429SFrederic Weisbecker */ 61306ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 61316ba94429SFrederic Weisbecker 61326ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 61336ba94429SFrederic Weisbecker if (ret) { 6134537f4146SArvind Yadav put_device(&wq_dev->dev); 61356ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61366ba94429SFrederic Weisbecker return ret; 61376ba94429SFrederic Weisbecker } 61386ba94429SFrederic Weisbecker 61396ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 61406ba94429SFrederic Weisbecker struct device_attribute *attr; 61416ba94429SFrederic Weisbecker 61426ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 61436ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 61446ba94429SFrederic Weisbecker if (ret) { 61456ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61466ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61476ba94429SFrederic Weisbecker return ret; 61486ba94429SFrederic Weisbecker } 61496ba94429SFrederic Weisbecker } 61506ba94429SFrederic Weisbecker } 61516ba94429SFrederic Weisbecker 61526ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 61536ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 61546ba94429SFrederic Weisbecker return 0; 61556ba94429SFrederic Weisbecker } 61566ba94429SFrederic Weisbecker 61576ba94429SFrederic Weisbecker /** 61586ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 61596ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 61606ba94429SFrederic Weisbecker * 61616ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 61626ba94429SFrederic Weisbecker */ 61636ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 61646ba94429SFrederic Weisbecker { 61656ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 61666ba94429SFrederic Weisbecker 61676ba94429SFrederic Weisbecker if (!wq->wq_dev) 61686ba94429SFrederic Weisbecker return; 61696ba94429SFrederic Weisbecker 61706ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61716ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61726ba94429SFrederic Weisbecker } 61736ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 61746ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 61756ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 61766ba94429SFrederic Weisbecker 617782607adcSTejun Heo /* 617882607adcSTejun Heo * Workqueue watchdog. 617982607adcSTejun Heo * 618082607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 618182607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 618282607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 618382607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 618482607adcSTejun Heo * largely opaque. 618582607adcSTejun Heo * 618682607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 618782607adcSTejun Heo * state if some pools failed to make forward progress for a while where 618882607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 618982607adcSTejun Heo * 619082607adcSTejun Heo * This mechanism is controlled through the kernel parameter 619182607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 619282607adcSTejun Heo * corresponding sysfs parameter file. 619382607adcSTejun Heo */ 619482607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 619582607adcSTejun Heo 619682607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 61975cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 619882607adcSTejun Heo 619982607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 620082607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 620182607adcSTejun Heo 6202cd2440d6SPetr Mladek /* 6203cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6204cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6205cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6206cd2440d6SPetr Mladek * in all other situations. 6207cd2440d6SPetr Mladek */ 6208cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6209cd2440d6SPetr Mladek { 6210cd2440d6SPetr Mladek struct worker *worker; 6211cd2440d6SPetr Mladek unsigned long flags; 6212cd2440d6SPetr Mladek int bkt; 6213cd2440d6SPetr Mladek 6214cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6215cd2440d6SPetr Mladek 6216cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6217cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6218cd2440d6SPetr Mladek /* 6219cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6220cd2440d6SPetr Mladek * drivers that queue work while holding locks 6221cd2440d6SPetr Mladek * also taken in their write paths. 6222cd2440d6SPetr Mladek */ 6223cd2440d6SPetr Mladek printk_deferred_enter(); 6224cd2440d6SPetr Mladek 6225cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6226cd2440d6SPetr Mladek sched_show_task(worker->task); 6227cd2440d6SPetr Mladek 6228cd2440d6SPetr Mladek printk_deferred_exit(); 6229cd2440d6SPetr Mladek } 6230cd2440d6SPetr Mladek } 6231cd2440d6SPetr Mladek 6232cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6233cd2440d6SPetr Mladek } 6234cd2440d6SPetr Mladek 6235cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6236cd2440d6SPetr Mladek { 6237cd2440d6SPetr Mladek struct worker_pool *pool; 6238cd2440d6SPetr Mladek int pi; 6239cd2440d6SPetr Mladek 6240cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6241cd2440d6SPetr Mladek 6242cd2440d6SPetr Mladek rcu_read_lock(); 6243cd2440d6SPetr Mladek 6244cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6245cd2440d6SPetr Mladek if (pool->cpu_stall) 6246cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6247cd2440d6SPetr Mladek 6248cd2440d6SPetr Mladek } 6249cd2440d6SPetr Mladek 6250cd2440d6SPetr Mladek rcu_read_unlock(); 6251cd2440d6SPetr Mladek } 6252cd2440d6SPetr Mladek 625382607adcSTejun Heo static void wq_watchdog_reset_touched(void) 625482607adcSTejun Heo { 625582607adcSTejun Heo int cpu; 625682607adcSTejun Heo 625782607adcSTejun Heo wq_watchdog_touched = jiffies; 625882607adcSTejun Heo for_each_possible_cpu(cpu) 625982607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 626082607adcSTejun Heo } 626182607adcSTejun Heo 62625cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 626382607adcSTejun Heo { 626482607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 626582607adcSTejun Heo bool lockup_detected = false; 6266cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6267940d71c6SSergey Senozhatsky unsigned long now = jiffies; 626882607adcSTejun Heo struct worker_pool *pool; 626982607adcSTejun Heo int pi; 627082607adcSTejun Heo 627182607adcSTejun Heo if (!thresh) 627282607adcSTejun Heo return; 627382607adcSTejun Heo 627482607adcSTejun Heo rcu_read_lock(); 627582607adcSTejun Heo 627682607adcSTejun Heo for_each_pool(pool, pi) { 627782607adcSTejun Heo unsigned long pool_ts, touched, ts; 627882607adcSTejun Heo 6279cd2440d6SPetr Mladek pool->cpu_stall = false; 628082607adcSTejun Heo if (list_empty(&pool->worklist)) 628182607adcSTejun Heo continue; 628282607adcSTejun Heo 6283940d71c6SSergey Senozhatsky /* 6284940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6285940d71c6SSergey Senozhatsky * the watchdog like a stall. 6286940d71c6SSergey Senozhatsky */ 6287940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6288940d71c6SSergey Senozhatsky 628982607adcSTejun Heo /* get the latest of pool and touched timestamps */ 629089e28ce6SWang Qing if (pool->cpu >= 0) 629189e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 629289e28ce6SWang Qing else 629382607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 629489e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 629582607adcSTejun Heo 629682607adcSTejun Heo if (time_after(pool_ts, touched)) 629782607adcSTejun Heo ts = pool_ts; 629882607adcSTejun Heo else 629982607adcSTejun Heo ts = touched; 630082607adcSTejun Heo 630182607adcSTejun Heo /* did we stall? */ 6302940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 630382607adcSTejun Heo lockup_detected = true; 6304cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6305cd2440d6SPetr Mladek pool->cpu_stall = true; 6306cd2440d6SPetr Mladek cpu_pool_stall = true; 6307cd2440d6SPetr Mladek } 630882607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 630982607adcSTejun Heo pr_cont_pool_info(pool); 631082607adcSTejun Heo pr_cont(" stuck for %us!\n", 6311940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 631282607adcSTejun Heo } 6313cd2440d6SPetr Mladek 6314cd2440d6SPetr Mladek 631582607adcSTejun Heo } 631682607adcSTejun Heo 631782607adcSTejun Heo rcu_read_unlock(); 631882607adcSTejun Heo 631982607adcSTejun Heo if (lockup_detected) 632055df0933SImran Khan show_all_workqueues(); 632182607adcSTejun Heo 6322cd2440d6SPetr Mladek if (cpu_pool_stall) 6323cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6324cd2440d6SPetr Mladek 632582607adcSTejun Heo wq_watchdog_reset_touched(); 632682607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 632782607adcSTejun Heo } 632882607adcSTejun Heo 6329cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 633082607adcSTejun Heo { 633182607adcSTejun Heo if (cpu >= 0) 633282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 633389e28ce6SWang Qing 633482607adcSTejun Heo wq_watchdog_touched = jiffies; 633582607adcSTejun Heo } 633682607adcSTejun Heo 633782607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 633882607adcSTejun Heo { 633982607adcSTejun Heo wq_watchdog_thresh = 0; 634082607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 634182607adcSTejun Heo 634282607adcSTejun Heo if (thresh) { 634382607adcSTejun Heo wq_watchdog_thresh = thresh; 634482607adcSTejun Heo wq_watchdog_reset_touched(); 634582607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 634682607adcSTejun Heo } 634782607adcSTejun Heo } 634882607adcSTejun Heo 634982607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 635082607adcSTejun Heo const struct kernel_param *kp) 635182607adcSTejun Heo { 635282607adcSTejun Heo unsigned long thresh; 635382607adcSTejun Heo int ret; 635482607adcSTejun Heo 635582607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 635682607adcSTejun Heo if (ret) 635782607adcSTejun Heo return ret; 635882607adcSTejun Heo 635982607adcSTejun Heo if (system_wq) 636082607adcSTejun Heo wq_watchdog_set_thresh(thresh); 636182607adcSTejun Heo else 636282607adcSTejun Heo wq_watchdog_thresh = thresh; 636382607adcSTejun Heo 636482607adcSTejun Heo return 0; 636582607adcSTejun Heo } 636682607adcSTejun Heo 636782607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 636882607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 636982607adcSTejun Heo .get = param_get_ulong, 637082607adcSTejun Heo }; 637182607adcSTejun Heo 637282607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 637382607adcSTejun Heo 0644); 637482607adcSTejun Heo 637582607adcSTejun Heo static void wq_watchdog_init(void) 637682607adcSTejun Heo { 63775cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 637882607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 637982607adcSTejun Heo } 638082607adcSTejun Heo 638182607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 638282607adcSTejun Heo 638382607adcSTejun Heo static inline void wq_watchdog_init(void) { } 638482607adcSTejun Heo 638582607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 638682607adcSTejun Heo 6387bce90380STejun Heo static void __init wq_numa_init(void) 6388bce90380STejun Heo { 6389bce90380STejun Heo cpumask_var_t *tbl; 6390bce90380STejun Heo int node, cpu; 6391bce90380STejun Heo 6392bce90380STejun Heo if (num_possible_nodes() <= 1) 6393bce90380STejun Heo return; 6394bce90380STejun Heo 6395d55262c4STejun Heo if (wq_disable_numa) { 6396d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 6397d55262c4STejun Heo return; 6398d55262c4STejun Heo } 6399d55262c4STejun Heo 6400f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6401f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6402f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6403f728c4a9SZhen Lei return; 6404f728c4a9SZhen Lei } 6405f728c4a9SZhen Lei } 6406f728c4a9SZhen Lei 6407be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 64084c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 64094c16bd32STejun Heo 6410bce90380STejun Heo /* 6411bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6412bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6413bce90380STejun Heo * fully initialized by now. 6414bce90380STejun Heo */ 64156396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6416bce90380STejun Heo BUG_ON(!tbl); 6417bce90380STejun Heo 6418bce90380STejun Heo for_each_node(node) 64195a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 64201be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6421bce90380STejun Heo 6422bce90380STejun Heo for_each_possible_cpu(cpu) { 6423bce90380STejun Heo node = cpu_to_node(cpu); 6424bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6425bce90380STejun Heo } 6426bce90380STejun Heo 6427bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6428bce90380STejun Heo wq_numa_enabled = true; 6429bce90380STejun Heo } 6430bce90380STejun Heo 64313347fa09STejun Heo /** 64323347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 64333347fa09STejun Heo * 64343347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 64353347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 64363347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 64373347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 64383347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 64393347fa09STejun Heo * created and scheduled right before early initcalls. 64403347fa09STejun Heo */ 64412333e829SYu Chen void __init workqueue_init_early(void) 64421da177e4SLinus Torvalds { 64437a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 64447a4e344cSTejun Heo int i, cpu; 6445c34056a3STejun Heo 644610cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6447e904e6c2STejun Heo 6448b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 644904d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 645004d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6451b05a7928SFrederic Weisbecker 6452e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6453e904e6c2STejun Heo 6454706026c2STejun Heo /* initialize CPU pools */ 645529c91e99STejun Heo for_each_possible_cpu(cpu) { 64564ce62e9eSTejun Heo struct worker_pool *pool; 64578b03ae3cSTejun Heo 64587a4e344cSTejun Heo i = 0; 6459f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 64607a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6461ec22ca5eSTejun Heo pool->cpu = cpu; 64627a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 64637a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6464f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 64657a4e344cSTejun Heo 64669daf9e67STejun Heo /* alloc pool ID */ 646768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 64689daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 646968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 64704ce62e9eSTejun Heo } 64718b03ae3cSTejun Heo } 64728b03ae3cSTejun Heo 64738a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 647429c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 647529c91e99STejun Heo struct workqueue_attrs *attrs; 647629c91e99STejun Heo 6477be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 647829c91e99STejun Heo attrs->nice = std_nice[i]; 647929c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 64808a2b7538STejun Heo 64818a2b7538STejun Heo /* 64828a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 64838a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 64848a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 64858a2b7538STejun Heo */ 6486be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 64878a2b7538STejun Heo attrs->nice = std_nice[i]; 64888a2b7538STejun Heo attrs->no_numa = true; 64898a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 649029c91e99STejun Heo } 649129c91e99STejun Heo 6492d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 64931aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6494d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6495f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6496f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 649724d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 649824d51addSTejun Heo WQ_FREEZABLE, 0); 64990668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 65000668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 65010668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 65020668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 65030668106cSViresh Kumar 0); 65041aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 65050668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 65060668106cSViresh Kumar !system_power_efficient_wq || 65070668106cSViresh Kumar !system_freezable_power_efficient_wq); 65083347fa09STejun Heo } 65093347fa09STejun Heo 65103347fa09STejun Heo /** 65113347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 65123347fa09STejun Heo * 65133347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 65143347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 65153347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 65163347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 65173347fa09STejun Heo * with the initial workers and enable future kworker creations. 65183347fa09STejun Heo */ 65192333e829SYu Chen void __init workqueue_init(void) 65203347fa09STejun Heo { 65212186d9f9STejun Heo struct workqueue_struct *wq; 65223347fa09STejun Heo struct worker_pool *pool; 65233347fa09STejun Heo int cpu, bkt; 65243347fa09STejun Heo 65252186d9f9STejun Heo /* 65262186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 65272186d9f9STejun Heo * CPU to node mapping may not be available that early on some 65282186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 65292186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 65302186d9f9STejun Heo * affinity, fix them up. 653140c17f75STejun Heo * 653240c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 65332186d9f9STejun Heo */ 65342186d9f9STejun Heo wq_numa_init(); 65352186d9f9STejun Heo 65362186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 65372186d9f9STejun Heo 65382186d9f9STejun Heo for_each_possible_cpu(cpu) { 65392186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65402186d9f9STejun Heo pool->node = cpu_to_node(cpu); 65412186d9f9STejun Heo } 65422186d9f9STejun Heo } 65432186d9f9STejun Heo 654440c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 65452186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 654640c17f75STejun Heo WARN(init_rescuer(wq), 654740c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 654840c17f75STejun Heo wq->name); 654940c17f75STejun Heo } 65502186d9f9STejun Heo 65512186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 65522186d9f9STejun Heo 65533347fa09STejun Heo /* create the initial workers */ 65543347fa09STejun Heo for_each_online_cpu(cpu) { 65553347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65563347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 65573347fa09STejun Heo BUG_ON(!create_worker(pool)); 65583347fa09STejun Heo } 65593347fa09STejun Heo } 65603347fa09STejun Heo 65613347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 65623347fa09STejun Heo BUG_ON(!create_worker(pool)); 65633347fa09STejun Heo 65643347fa09STejun Heo wq_online = true; 656582607adcSTejun Heo wq_watchdog_init(); 65661da177e4SLinus Torvalds } 6567c4f135d6STetsuo Handa 6568c4f135d6STetsuo Handa /* 6569c4f135d6STetsuo Handa * Despite the naming, this is a no-op function which is here only for avoiding 6570c4f135d6STetsuo Handa * link error. Since compile-time warning may fail to catch, we will need to 6571c4f135d6STetsuo Handa * emit run-time warning from __flush_workqueue(). 6572c4f135d6STetsuo Handa */ 6573c4f135d6STetsuo Handa void __warn_flushing_systemwide_wq(void) { } 6574c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6575