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> 55aa6fde93STejun Heo #include <linux/delay.h> 56e22bee78STejun Heo 57ea138446STejun Heo #include "workqueue_internal.h" 581da177e4SLinus Torvalds 59c8e55f36STejun Heo enum { 60bc2ae0f5STejun Heo /* 6124647570STejun Heo * worker_pool flags 62bc2ae0f5STejun Heo * 6324647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 64bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 65bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 66bc2ae0f5STejun Heo * is in effect. 67bc2ae0f5STejun Heo * 68bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 69bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 7024647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 71bc2ae0f5STejun Heo * 72bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 731258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 744736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 75bc2ae0f5STejun Heo */ 76692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7724647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 78db7bccf4STejun Heo 79c8e55f36STejun Heo /* worker flags */ 80c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 81c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 82e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 83fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 84f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 85a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 86e22bee78STejun Heo 87a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 88a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 89db7bccf4STejun Heo 90e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 914ce62e9eSTejun Heo 9229c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 93c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 94db7bccf4STejun Heo 95e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 96e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 97e22bee78STejun Heo 983233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 993233cdbdSTejun Heo /* call for help after 10ms 1003233cdbdSTejun Heo (min two ticks) */ 101e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 102e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds /* 105e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1068698a745SDongsheng Yang * all cpus. Give MIN_NICE. 107e22bee78STejun Heo */ 1088698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1098698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 110ecf6881fSTejun Heo 111ecf6881fSTejun Heo WQ_NAME_LEN = 24, 112c8e55f36STejun Heo }; 113c8e55f36STejun Heo 1141da177e4SLinus Torvalds /* 1154690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1164690c4abSTejun Heo * 117e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 118e41e704bSTejun Heo * everyone else. 1194690c4abSTejun Heo * 120e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 121e22bee78STejun Heo * only be modified and accessed from the local cpu. 122e22bee78STejun Heo * 123d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1244690c4abSTejun Heo * 125bdf8b9bfSTejun Heo * K: Only modified by worker while holding pool->lock. Can be safely read by 126bdf8b9bfSTejun Heo * self, while holding pool->lock or from IRQ context if %current is the 127bdf8b9bfSTejun Heo * kworker. 128bdf8b9bfSTejun Heo * 129bdf8b9bfSTejun Heo * S: Only modified by worker self. 130bdf8b9bfSTejun Heo * 1311258fae7STejun Heo * A: wq_pool_attach_mutex protected. 132822d8405STejun Heo * 13368e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13476af4d93STejun Heo * 13524acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1365bcab335STejun Heo * 1375b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1385b95e1afSLai Jiangshan * 1395b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 14024acfb71SThomas Gleixner * RCU for reads. 1415b95e1afSLai Jiangshan * 1423c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1433c25a55dSLai Jiangshan * 14424acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1452e109a28STejun Heo * 1462e109a28STejun Heo * MD: wq_mayday_lock protected. 147cd2440d6SPetr Mladek * 148cd2440d6SPetr Mladek * WD: Used internally by the watchdog. 1494690c4abSTejun Heo */ 1504690c4abSTejun Heo 1512eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 152c34056a3STejun Heo 153bd7bdd43STejun Heo struct worker_pool { 154a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 155d84ff051STejun Heo int cpu; /* I: the associated cpu */ 156f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1579daf9e67STejun Heo int id; /* I: pool ID */ 158bc8b50c2STejun Heo unsigned int flags; /* L: flags */ 159bd7bdd43STejun Heo 16082607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 161cd2440d6SPetr Mladek bool cpu_stall; /* WD: stalled cpu bound pool */ 16282607adcSTejun Heo 163bc35f7efSLai Jiangshan /* 164bc35f7efSLai Jiangshan * The counter is incremented in a process context on the associated CPU 165bc35f7efSLai Jiangshan * w/ preemption disabled, and decremented or reset in the same context 166bc35f7efSLai Jiangshan * but w/ pool->lock held. The readers grab pool->lock and are 167bc35f7efSLai Jiangshan * guaranteed to see if the counter reached zero. 168bc35f7efSLai Jiangshan */ 169bc35f7efSLai Jiangshan int nr_running; 17084f91c62SLai Jiangshan 171bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 172ea1abd61SLai Jiangshan 1735826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1745826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 175bd7bdd43STejun Heo 1762c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 177bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 1783f959aa3SValentin Schneider struct work_struct idle_cull_work; /* L: worker idle cleanup */ 1793f959aa3SValentin Schneider 180bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 181bd7bdd43STejun Heo 182c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 183c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 184c9e7cf27STejun Heo /* L: hash of busy workers */ 185c9e7cf27STejun Heo 1862607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 18792f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 188e02b9312SValentin Schneider struct list_head dying_workers; /* A: workers about to die */ 18960f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 190e19e397aSTejun Heo 1917cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 192e19e397aSTejun Heo 1937a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 19468e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 19568e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1967a4e344cSTejun Heo 197e19e397aSTejun Heo /* 19824acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 19929c91e99STejun Heo * from get_work_pool(). 20029c91e99STejun Heo */ 20129c91e99STejun Heo struct rcu_head rcu; 20284f91c62SLai Jiangshan }; 2038b03ae3cSTejun Heo 2048b03ae3cSTejun Heo /* 205725e8ec5STejun Heo * Per-pool_workqueue statistics. These can be monitored using 206725e8ec5STejun Heo * tools/workqueue/wq_monitor.py. 207725e8ec5STejun Heo */ 208725e8ec5STejun Heo enum pool_workqueue_stats { 209725e8ec5STejun Heo PWQ_STAT_STARTED, /* work items started execution */ 210725e8ec5STejun Heo PWQ_STAT_COMPLETED, /* work items completed execution */ 2118a1dd1e5STejun Heo PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 212616db877STejun Heo PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 213725e8ec5STejun Heo PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 214725e8ec5STejun Heo PWQ_STAT_MAYDAY, /* maydays to rescuer */ 215725e8ec5STejun Heo PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 216725e8ec5STejun Heo 217725e8ec5STejun Heo PWQ_NR_STATS, 218725e8ec5STejun Heo }; 219725e8ec5STejun Heo 220725e8ec5STejun Heo /* 221112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 222112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 223112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 224112202d9STejun Heo * number of flag bits. 2251da177e4SLinus Torvalds */ 226112202d9STejun Heo struct pool_workqueue { 227bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2284690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 22973f53c4aSTejun Heo int work_color; /* L: current color */ 23073f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2318864b4e5STejun Heo int refcnt; /* L: reference count */ 23273f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 23373f53c4aSTejun Heo /* L: nr of in_flight works */ 234018f3a13SLai Jiangshan 235018f3a13SLai Jiangshan /* 236018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 237018f3a13SLai Jiangshan * 238018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 239018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 240018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 241018f3a13SLai Jiangshan * 242018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 243018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 244018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 245018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 246018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 247018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 248018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 249018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 250018f3a13SLai Jiangshan */ 2511e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 252a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 253f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2543c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2552e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2568864b4e5STejun Heo 257725e8ec5STejun Heo u64 stats[PWQ_NR_STATS]; 258725e8ec5STejun Heo 2598864b4e5STejun Heo /* 260967b494eSTejun Heo * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 261687a9aa5STejun Heo * and pwq_release_workfn() for details. pool_workqueue itself is also 262687a9aa5STejun Heo * RCU protected so that the first pwq can be determined without 263967b494eSTejun Heo * grabbing wq->mutex. 2648864b4e5STejun Heo */ 265687a9aa5STejun Heo struct kthread_work release_work; 2668864b4e5STejun Heo struct rcu_head rcu; 267e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2681da177e4SLinus Torvalds 2691da177e4SLinus Torvalds /* 27073f53c4aSTejun Heo * Structure used to wait for workqueue flush. 27173f53c4aSTejun Heo */ 27273f53c4aSTejun Heo struct wq_flusher { 2733c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2743c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 27573f53c4aSTejun Heo struct completion done; /* flush completion */ 27673f53c4aSTejun Heo }; 2771da177e4SLinus Torvalds 278226223abSTejun Heo struct wq_device; 279226223abSTejun Heo 28073f53c4aSTejun Heo /* 281c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 282c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2831da177e4SLinus Torvalds */ 2841da177e4SLinus Torvalds struct workqueue_struct { 2853c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 286e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 28773f53c4aSTejun Heo 2883c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2893c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2903c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 291112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2923c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2933c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2943c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 29573f53c4aSTejun Heo 2962e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 29730ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 298e22bee78STejun Heo 29987fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 300a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 301226223abSTejun Heo 3025b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 3035b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 3046029a918STejun Heo 305226223abSTejun Heo #ifdef CONFIG_SYSFS 306226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 307226223abSTejun Heo #endif 3084e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 309669de8bdSBart Van Assche char *lock_name; 310669de8bdSBart Van Assche struct lock_class_key key; 3114e6045f1SJohannes Berg struct lockdep_map lockdep_map; 3124e6045f1SJohannes Berg #endif 313ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 3142728fd2fSTejun Heo 315e2dca7adSTejun Heo /* 31624acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 31724acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 318e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 319e2dca7adSTejun Heo */ 320e2dca7adSTejun Heo struct rcu_head rcu; 321e2dca7adSTejun Heo 3222728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 3232728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 324687a9aa5STejun Heo struct pool_workqueue __percpu **cpu_pwq; /* I: per-cpu pwqs */ 3255b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 3261da177e4SLinus Torvalds }; 3271da177e4SLinus Torvalds 328e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 329e904e6c2STejun Heo 330bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 331bce90380STejun Heo /* possible CPUs of each node */ 332bce90380STejun Heo 333616db877STejun Heo /* 334616db877STejun Heo * Per-cpu work items which run for longer than the following threshold are 335616db877STejun Heo * automatically considered CPU intensive and excluded from concurrency 336616db877STejun Heo * management to prevent them from noticeably delaying other per-cpu work items. 337aa6fde93STejun Heo * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 338aa6fde93STejun Heo * The actual value is initialized in wq_cpu_intensive_thresh_init(). 339616db877STejun Heo */ 340aa6fde93STejun Heo static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 341616db877STejun Heo module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 342616db877STejun Heo 343cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 344552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 345cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 346cee22a15SViresh Kumar 347863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3483347fa09STejun Heo 349bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 350bce90380STejun Heo 3514c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3524c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3534c16bd32STejun Heo 35468e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3551258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 356a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 357d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 358d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3595bcab335STejun Heo 360e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 36168e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3627d19c5ceSTejun Heo 36399c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 364ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 365ef557180SMike Galbraith 366ace3c549Stiozhang /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 367ace3c549Stiozhang static struct cpumask wq_cmdline_cpumask __initdata; 368ace3c549Stiozhang 369ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 370ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 371b05a7928SFrederic Weisbecker 372f303fccbSTejun Heo /* 373f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 374f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 375f303fccbSTejun Heo * to uncover usages which depend on it. 376f303fccbSTejun Heo */ 377f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 378f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 379f303fccbSTejun Heo #else 380f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 381f303fccbSTejun Heo #endif 382f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 383f303fccbSTejun Heo 3847d19c5ceSTejun Heo /* the per-cpu worker pools */ 38525528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3867d19c5ceSTejun Heo 38768e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3887d19c5ceSTejun Heo 38968e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 39029c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 39129c91e99STejun Heo 392c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 39329c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 39429c91e99STejun Heo 3958a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3968a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3978a2b7538STejun Heo 398967b494eSTejun Heo /* 399967b494eSTejun Heo * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 400967b494eSTejun Heo * process context while holding a pool lock. Bounce to a dedicated kthread 401967b494eSTejun Heo * worker to avoid A-A deadlocks. 402967b494eSTejun Heo */ 403967b494eSTejun Heo static struct kthread_worker *pwq_release_worker; 404967b494eSTejun Heo 405d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 406ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 407044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 4081aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 409044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 410d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 411044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 412f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 413044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 41424d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 4150668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 4160668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 4170668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 4180668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 419d320c038STejun Heo 4207d19c5ceSTejun Heo static int worker_thread(void *__worker); 4216ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 422c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 42355df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4247d19c5ceSTejun Heo 42597bd2347STejun Heo #define CREATE_TRACE_POINTS 42697bd2347STejun Heo #include <trace/events/workqueue.h> 42797bd2347STejun Heo 42868e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 42924acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 430f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 43124acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4325bcab335STejun Heo 4335b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 43424acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 435f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 436f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 43724acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4385b95e1afSLai Jiangshan 439f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 440f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 441f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4427a62c2c8STejun Heo (pool)++) 4434ce62e9eSTejun Heo 44449e3cf44STejun Heo /** 44517116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 44617116969STejun Heo * @pool: iteration cursor 447611c92a0STejun Heo * @pi: integer used for iteration 448fa1b54e6STejun Heo * 44924acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 45068e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 45168e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 452fa1b54e6STejun Heo * 453fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 454fa1b54e6STejun Heo * ignored. 45517116969STejun Heo */ 456611c92a0STejun Heo #define for_each_pool(pool, pi) \ 457611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 45868e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 459fa1b54e6STejun Heo else 46017116969STejun Heo 46117116969STejun Heo /** 462822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 463822d8405STejun Heo * @worker: iteration cursor 464822d8405STejun Heo * @pool: worker_pool to iterate workers of 465822d8405STejun Heo * 4661258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 467822d8405STejun Heo * 468822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 469822d8405STejun Heo * ignored. 470822d8405STejun Heo */ 471da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 472da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4731258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 474822d8405STejun Heo else 475822d8405STejun Heo 476822d8405STejun Heo /** 47749e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 47849e3cf44STejun Heo * @pwq: iteration cursor 47949e3cf44STejun Heo * @wq: the target workqueue 48076af4d93STejun Heo * 48124acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 482794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 483794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 48476af4d93STejun Heo * 48576af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 48676af4d93STejun Heo * ignored. 48749e3cf44STejun Heo */ 48849e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 48949e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4905a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 491f3421797STejun Heo 492dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 493dc186ad7SThomas Gleixner 494f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 495dc186ad7SThomas Gleixner 49699777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 49799777288SStanislaw Gruszka { 49899777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 49999777288SStanislaw Gruszka } 50099777288SStanislaw Gruszka 501b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 502b9fdac7fSDu, Changbin { 503b9fdac7fSDu, Changbin struct work_struct *work = addr; 504b9fdac7fSDu, Changbin 505b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 506b9fdac7fSDu, Changbin } 507b9fdac7fSDu, Changbin 508dc186ad7SThomas Gleixner /* 509dc186ad7SThomas Gleixner * fixup_init is called when: 510dc186ad7SThomas Gleixner * - an active object is initialized 511dc186ad7SThomas Gleixner */ 51202a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 513dc186ad7SThomas Gleixner { 514dc186ad7SThomas Gleixner struct work_struct *work = addr; 515dc186ad7SThomas Gleixner 516dc186ad7SThomas Gleixner switch (state) { 517dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 518dc186ad7SThomas Gleixner cancel_work_sync(work); 519dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 52002a982a6SDu, Changbin return true; 521dc186ad7SThomas Gleixner default: 52202a982a6SDu, Changbin return false; 523dc186ad7SThomas Gleixner } 524dc186ad7SThomas Gleixner } 525dc186ad7SThomas Gleixner 526dc186ad7SThomas Gleixner /* 527dc186ad7SThomas Gleixner * fixup_free is called when: 528dc186ad7SThomas Gleixner * - an active object is freed 529dc186ad7SThomas Gleixner */ 53002a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 531dc186ad7SThomas Gleixner { 532dc186ad7SThomas Gleixner struct work_struct *work = addr; 533dc186ad7SThomas Gleixner 534dc186ad7SThomas Gleixner switch (state) { 535dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 536dc186ad7SThomas Gleixner cancel_work_sync(work); 537dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 53802a982a6SDu, Changbin return true; 539dc186ad7SThomas Gleixner default: 54002a982a6SDu, Changbin return false; 541dc186ad7SThomas Gleixner } 542dc186ad7SThomas Gleixner } 543dc186ad7SThomas Gleixner 544f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 545dc186ad7SThomas Gleixner .name = "work_struct", 54699777288SStanislaw Gruszka .debug_hint = work_debug_hint, 547b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 548dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 549dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 550dc186ad7SThomas Gleixner }; 551dc186ad7SThomas Gleixner 552dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 553dc186ad7SThomas Gleixner { 554dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 555dc186ad7SThomas Gleixner } 556dc186ad7SThomas Gleixner 557dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 558dc186ad7SThomas Gleixner { 559dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 560dc186ad7SThomas Gleixner } 561dc186ad7SThomas Gleixner 562dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 563dc186ad7SThomas Gleixner { 564dc186ad7SThomas Gleixner if (onstack) 565dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 566dc186ad7SThomas Gleixner else 567dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 568dc186ad7SThomas Gleixner } 569dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 570dc186ad7SThomas Gleixner 571dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 572dc186ad7SThomas Gleixner { 573dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 574dc186ad7SThomas Gleixner } 575dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 576dc186ad7SThomas Gleixner 577ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 578ea2e64f2SThomas Gleixner { 579ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 580ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 581ea2e64f2SThomas Gleixner } 582ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 583ea2e64f2SThomas Gleixner 584dc186ad7SThomas Gleixner #else 585dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 586dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 587dc186ad7SThomas Gleixner #endif 588dc186ad7SThomas Gleixner 5894e8b22bdSLi Bin /** 59067dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5914e8b22bdSLi Bin * @pool: the pool pointer of interest 5924e8b22bdSLi Bin * 5934e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5944e8b22bdSLi Bin * successfully, -errno on failure. 5954e8b22bdSLi Bin */ 5969daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5979daf9e67STejun Heo { 5989daf9e67STejun Heo int ret; 5999daf9e67STejun Heo 60068e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 6015bcab335STejun Heo 6024e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 6034e8b22bdSLi Bin GFP_KERNEL); 604229641a6STejun Heo if (ret >= 0) { 605e68035fbSTejun Heo pool->id = ret; 606229641a6STejun Heo return 0; 607229641a6STejun Heo } 6089daf9e67STejun Heo return ret; 6099daf9e67STejun Heo } 6109daf9e67STejun Heo 61176af4d93STejun Heo /** 612df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 613df2d5ae4STejun Heo * @wq: the target workqueue 614df2d5ae4STejun Heo * @node: the node ID 615df2d5ae4STejun Heo * 61624acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 6175b95e1afSLai Jiangshan * read locked. 618df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 619df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 620d185af30SYacine Belkadi * 621d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 622df2d5ae4STejun Heo */ 623df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 624df2d5ae4STejun Heo int node) 625df2d5ae4STejun Heo { 6265b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 627d6e022f1STejun Heo 628d6e022f1STejun Heo /* 629d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 630d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 631d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 632d6e022f1STejun Heo * happens, this workaround can be removed. 633d6e022f1STejun Heo */ 634d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 635d6e022f1STejun Heo return wq->dfl_pwq; 636d6e022f1STejun Heo 637df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 638df2d5ae4STejun Heo } 639df2d5ae4STejun Heo 64073f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 64173f53c4aSTejun Heo { 64273f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 64373f53c4aSTejun Heo } 64473f53c4aSTejun Heo 645c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 64673f53c4aSTejun Heo { 647c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 64873f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 64973f53c4aSTejun Heo } 65073f53c4aSTejun Heo 65173f53c4aSTejun Heo static int work_next_color(int color) 65273f53c4aSTejun Heo { 65373f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 654a848e3b6SOleg Nesterov } 655a848e3b6SOleg Nesterov 6564594bf15SDavid Howells /* 657112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 658112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6597c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6607a22ad75STejun Heo * 661112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 662112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 663bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 664bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6657a22ad75STejun Heo * 666112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6677c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 668112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6697c3eed5cSTejun Heo * available only while the work item is queued. 670bbb68dfaSTejun Heo * 671bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 672bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 673bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 674bbb68dfaSTejun Heo * try to steal the PENDING bit. 6754594bf15SDavid Howells */ 6767a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6777a22ad75STejun Heo unsigned long flags) 6787a22ad75STejun Heo { 6796183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6807a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6817a22ad75STejun Heo } 6827a22ad75STejun Heo 683112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6844690c4abSTejun Heo unsigned long extra_flags) 685365970a1SDavid Howells { 686112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 687112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 688365970a1SDavid Howells } 689365970a1SDavid Howells 6904468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6914468a00fSLai Jiangshan int pool_id) 6924468a00fSLai Jiangshan { 6934468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6944468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6954468a00fSLai Jiangshan } 6964468a00fSLai Jiangshan 6977c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6987c3eed5cSTejun Heo int pool_id) 6994d707b9fSOleg Nesterov { 70023657bb1STejun Heo /* 70123657bb1STejun Heo * The following wmb is paired with the implied mb in 70223657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 70323657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 70423657bb1STejun Heo * owner. 70523657bb1STejun Heo */ 70623657bb1STejun Heo smp_wmb(); 7077c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 708346c09f8SRoman Pen /* 709346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 710346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 711346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 7128bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 713346c09f8SRoman Pen * the same @work. E.g. consider this case: 714346c09f8SRoman Pen * 715346c09f8SRoman Pen * CPU#0 CPU#1 716346c09f8SRoman Pen * ---------------------------- -------------------------------- 717346c09f8SRoman Pen * 718346c09f8SRoman Pen * 1 STORE event_indicated 719346c09f8SRoman Pen * 2 queue_work_on() { 720346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 721346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 722346c09f8SRoman Pen * 5 set_work_data() # clear bit 723346c09f8SRoman Pen * 6 smp_mb() 724346c09f8SRoman Pen * 7 work->current_func() { 725346c09f8SRoman Pen * 8 LOAD event_indicated 726346c09f8SRoman Pen * } 727346c09f8SRoman Pen * 728346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 729346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 730346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 731346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 732346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 733346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 734346c09f8SRoman Pen * before actual STORE. 735346c09f8SRoman Pen */ 736346c09f8SRoman Pen smp_mb(); 7374d707b9fSOleg Nesterov } 7384d707b9fSOleg Nesterov 7397a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 740365970a1SDavid Howells { 7417c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7427c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7437a22ad75STejun Heo } 7447a22ad75STejun Heo 745afa4bb77SLinus Torvalds static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 746afa4bb77SLinus Torvalds { 747afa4bb77SLinus Torvalds return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 748afa4bb77SLinus Torvalds } 749afa4bb77SLinus Torvalds 750112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7517a22ad75STejun Heo { 752e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7537a22ad75STejun Heo 754112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 755afa4bb77SLinus Torvalds return work_struct_pwq(data); 756e120153dSTejun Heo else 757e120153dSTejun Heo return NULL; 7587a22ad75STejun Heo } 7597a22ad75STejun Heo 7607c3eed5cSTejun Heo /** 7617c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7627c3eed5cSTejun Heo * @work: the work item of interest 7637c3eed5cSTejun Heo * 76468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 76524acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 76624acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 767fa1b54e6STejun Heo * 768fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 769fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 770fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 771fa1b54e6STejun Heo * returned pool is and stays online. 772d185af30SYacine Belkadi * 773d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7747c3eed5cSTejun Heo */ 7757c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7767a22ad75STejun Heo { 777e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7787c3eed5cSTejun Heo int pool_id; 7797a22ad75STejun Heo 78068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 781fa1b54e6STejun Heo 782112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 783afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool; 7847a22ad75STejun Heo 7857c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7867c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7877a22ad75STejun Heo return NULL; 7887a22ad75STejun Heo 789fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7907c3eed5cSTejun Heo } 7917c3eed5cSTejun Heo 7927c3eed5cSTejun Heo /** 7937c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7947c3eed5cSTejun Heo * @work: the work item of interest 7957c3eed5cSTejun Heo * 796d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7977c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7987c3eed5cSTejun Heo */ 7997c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 8007c3eed5cSTejun Heo { 80154d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 8027c3eed5cSTejun Heo 803112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 804afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool->id; 80554d5b7d0SLai Jiangshan 80654d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 8077c3eed5cSTejun Heo } 8087c3eed5cSTejun Heo 809bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 810bbb68dfaSTejun Heo { 8117c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 812bbb68dfaSTejun Heo 8137c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 8147c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 815bbb68dfaSTejun Heo } 816bbb68dfaSTejun Heo 817bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 818bbb68dfaSTejun Heo { 819bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 820bbb68dfaSTejun Heo 821112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 822bbb68dfaSTejun Heo } 823bbb68dfaSTejun Heo 824e22bee78STejun Heo /* 8253270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8263270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 827d565ed63STejun Heo * they're being called with pool->lock held. 828e22bee78STejun Heo */ 829e22bee78STejun Heo 83063d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 831649027d7STejun Heo { 832bc35f7efSLai Jiangshan return !pool->nr_running; 833649027d7STejun Heo } 834649027d7STejun Heo 835e22bee78STejun Heo /* 836e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 837e22bee78STejun Heo * running workers. 838974271c4STejun Heo * 839974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 840706026c2STejun Heo * function will always return %true for unbound pools as long as the 841974271c4STejun Heo * worklist isn't empty. 842e22bee78STejun Heo */ 84363d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 844e22bee78STejun Heo { 84563d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 846e22bee78STejun Heo } 847e22bee78STejun Heo 848e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 84963d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 850e22bee78STejun Heo { 85163d95a91STejun Heo return pool->nr_idle; 852e22bee78STejun Heo } 853e22bee78STejun Heo 854e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 85563d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 856e22bee78STejun Heo { 857bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 858e22bee78STejun Heo } 859e22bee78STejun Heo 860e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 86163d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 862e22bee78STejun Heo { 86363d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 864e22bee78STejun Heo } 865e22bee78STejun Heo 866e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 86763d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 868e22bee78STejun Heo { 869692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 87063d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 87163d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 872e22bee78STejun Heo 873e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 874e22bee78STejun Heo } 875e22bee78STejun Heo 8764690c4abSTejun Heo /** 877e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 878cb444766STejun Heo * @worker: self 879d302f017STejun Heo * @flags: flags to set 880d302f017STejun Heo * 881228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 882d302f017STejun Heo */ 883228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 884d302f017STejun Heo { 885bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 886e22bee78STejun Heo 887bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 888cb444766STejun Heo 889228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 890e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 891e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 892bc35f7efSLai Jiangshan pool->nr_running--; 893e22bee78STejun Heo } 894e22bee78STejun Heo 895d302f017STejun Heo worker->flags |= flags; 896d302f017STejun Heo } 897d302f017STejun Heo 898d302f017STejun Heo /** 899e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 900cb444766STejun Heo * @worker: self 901d302f017STejun Heo * @flags: flags to clear 902d302f017STejun Heo * 903e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 904d302f017STejun Heo */ 905d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 906d302f017STejun Heo { 90763d95a91STejun Heo struct worker_pool *pool = worker->pool; 908e22bee78STejun Heo unsigned int oflags = worker->flags; 909e22bee78STejun Heo 910bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 911cb444766STejun Heo 912d302f017STejun Heo worker->flags &= ~flags; 913e22bee78STejun Heo 91442c025f3STejun Heo /* 91542c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 91642c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 91742c025f3STejun Heo * of multiple flags, not a single flag. 91842c025f3STejun Heo */ 919e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 920e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 921bc35f7efSLai Jiangshan pool->nr_running++; 922d302f017STejun Heo } 923d302f017STejun Heo 924797e8345STejun Heo /* Return the first idle worker. Called with pool->lock held. */ 925797e8345STejun Heo static struct worker *first_idle_worker(struct worker_pool *pool) 926797e8345STejun Heo { 927797e8345STejun Heo if (unlikely(list_empty(&pool->idle_list))) 928797e8345STejun Heo return NULL; 929797e8345STejun Heo 930797e8345STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 931797e8345STejun Heo } 932797e8345STejun Heo 933797e8345STejun Heo /** 934797e8345STejun Heo * worker_enter_idle - enter idle state 935797e8345STejun Heo * @worker: worker which is entering idle state 936797e8345STejun Heo * 937797e8345STejun Heo * @worker is entering idle state. Update stats and idle timer if 938797e8345STejun Heo * necessary. 939797e8345STejun Heo * 940797e8345STejun Heo * LOCKING: 941797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 942797e8345STejun Heo */ 943797e8345STejun Heo static void worker_enter_idle(struct worker *worker) 944797e8345STejun Heo { 945797e8345STejun Heo struct worker_pool *pool = worker->pool; 946797e8345STejun Heo 947797e8345STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 948797e8345STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 949797e8345STejun Heo (worker->hentry.next || worker->hentry.pprev))) 950797e8345STejun Heo return; 951797e8345STejun Heo 952797e8345STejun Heo /* can't use worker_set_flags(), also called from create_worker() */ 953797e8345STejun Heo worker->flags |= WORKER_IDLE; 954797e8345STejun Heo pool->nr_idle++; 955797e8345STejun Heo worker->last_active = jiffies; 956797e8345STejun Heo 957797e8345STejun Heo /* idle_list is LIFO */ 958797e8345STejun Heo list_add(&worker->entry, &pool->idle_list); 959797e8345STejun Heo 960797e8345STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 961797e8345STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 962797e8345STejun Heo 963797e8345STejun Heo /* Sanity check nr_running. */ 964797e8345STejun Heo WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 965797e8345STejun Heo } 966797e8345STejun Heo 967797e8345STejun Heo /** 968797e8345STejun Heo * worker_leave_idle - leave idle state 969797e8345STejun Heo * @worker: worker which is leaving idle state 970797e8345STejun Heo * 971797e8345STejun Heo * @worker is leaving idle state. Update stats. 972797e8345STejun Heo * 973797e8345STejun Heo * LOCKING: 974797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 975797e8345STejun Heo */ 976797e8345STejun Heo static void worker_leave_idle(struct worker *worker) 977797e8345STejun Heo { 978797e8345STejun Heo struct worker_pool *pool = worker->pool; 979797e8345STejun Heo 980797e8345STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 981797e8345STejun Heo return; 982797e8345STejun Heo worker_clr_flags(worker, WORKER_IDLE); 983797e8345STejun Heo pool->nr_idle--; 984797e8345STejun Heo list_del_init(&worker->entry); 985797e8345STejun Heo } 986797e8345STejun Heo 987797e8345STejun Heo /** 988797e8345STejun Heo * find_worker_executing_work - find worker which is executing a work 989797e8345STejun Heo * @pool: pool of interest 990797e8345STejun Heo * @work: work to find worker for 991797e8345STejun Heo * 992797e8345STejun Heo * Find a worker which is executing @work on @pool by searching 993797e8345STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 994797e8345STejun Heo * to match, its current execution should match the address of @work and 995797e8345STejun Heo * its work function. This is to avoid unwanted dependency between 996797e8345STejun Heo * unrelated work executions through a work item being recycled while still 997797e8345STejun Heo * being executed. 998797e8345STejun Heo * 999797e8345STejun Heo * This is a bit tricky. A work item may be freed once its execution 1000797e8345STejun Heo * starts and nothing prevents the freed area from being recycled for 1001797e8345STejun Heo * another work item. If the same work item address ends up being reused 1002797e8345STejun Heo * before the original execution finishes, workqueue will identify the 1003797e8345STejun Heo * recycled work item as currently executing and make it wait until the 1004797e8345STejun Heo * current execution finishes, introducing an unwanted dependency. 1005797e8345STejun Heo * 1006797e8345STejun Heo * This function checks the work item address and work function to avoid 1007797e8345STejun Heo * false positives. Note that this isn't complete as one may construct a 1008797e8345STejun Heo * work function which can introduce dependency onto itself through a 1009797e8345STejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1010797e8345STejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1011797e8345STejun Heo * actually occurs, it should be easy to locate the culprit work function. 1012797e8345STejun Heo * 1013797e8345STejun Heo * CONTEXT: 1014797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1015797e8345STejun Heo * 1016797e8345STejun Heo * Return: 1017797e8345STejun Heo * Pointer to worker which is executing @work if found, %NULL 1018797e8345STejun Heo * otherwise. 1019797e8345STejun Heo */ 1020797e8345STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 1021797e8345STejun Heo struct work_struct *work) 1022797e8345STejun Heo { 1023797e8345STejun Heo struct worker *worker; 1024797e8345STejun Heo 1025797e8345STejun Heo hash_for_each_possible(pool->busy_hash, worker, hentry, 1026797e8345STejun Heo (unsigned long)work) 1027797e8345STejun Heo if (worker->current_work == work && 1028797e8345STejun Heo worker->current_func == work->func) 1029797e8345STejun Heo return worker; 1030797e8345STejun Heo 1031797e8345STejun Heo return NULL; 1032797e8345STejun Heo } 1033797e8345STejun Heo 1034797e8345STejun Heo /** 1035797e8345STejun Heo * move_linked_works - move linked works to a list 1036797e8345STejun Heo * @work: start of series of works to be scheduled 1037797e8345STejun Heo * @head: target list to append @work to 1038797e8345STejun Heo * @nextp: out parameter for nested worklist walking 1039797e8345STejun Heo * 1040797e8345STejun Heo * Schedule linked works starting from @work to @head. Work series to 1041797e8345STejun Heo * be scheduled starts at @work and includes any consecutive work with 1042797e8345STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1043797e8345STejun Heo * 1044797e8345STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1045797e8345STejun Heo * the last scheduled work. This allows move_linked_works() to be 1046797e8345STejun Heo * nested inside outer list_for_each_entry_safe(). 1047797e8345STejun Heo * 1048797e8345STejun Heo * CONTEXT: 1049797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1050797e8345STejun Heo */ 1051797e8345STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1052797e8345STejun Heo struct work_struct **nextp) 1053797e8345STejun Heo { 1054797e8345STejun Heo struct work_struct *n; 1055797e8345STejun Heo 1056797e8345STejun Heo /* 1057797e8345STejun Heo * Linked worklist will always end before the end of the list, 1058797e8345STejun Heo * use NULL for list head. 1059797e8345STejun Heo */ 1060797e8345STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1061797e8345STejun Heo list_move_tail(&work->entry, head); 1062797e8345STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1063797e8345STejun Heo break; 1064797e8345STejun Heo } 1065797e8345STejun Heo 1066797e8345STejun Heo /* 1067797e8345STejun Heo * If we're already inside safe list traversal and have moved 1068797e8345STejun Heo * multiple works to the scheduled queue, the next position 1069797e8345STejun Heo * needs to be updated. 1070797e8345STejun Heo */ 1071797e8345STejun Heo if (nextp) 1072797e8345STejun Heo *nextp = n; 1073797e8345STejun Heo } 1074797e8345STejun Heo 1075797e8345STejun Heo /** 1076797e8345STejun Heo * wake_up_worker - wake up an idle worker 1077797e8345STejun Heo * @pool: worker pool to wake worker from 1078797e8345STejun Heo * 1079797e8345STejun Heo * Wake up the first idle worker of @pool. 1080797e8345STejun Heo * 1081797e8345STejun Heo * CONTEXT: 1082797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1083797e8345STejun Heo */ 1084797e8345STejun Heo static void wake_up_worker(struct worker_pool *pool) 1085797e8345STejun Heo { 1086797e8345STejun Heo struct worker *worker = first_idle_worker(pool); 1087797e8345STejun Heo 1088797e8345STejun Heo if (likely(worker)) 1089797e8345STejun Heo wake_up_process(worker->task); 1090797e8345STejun Heo } 1091797e8345STejun Heo 109263638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 109363638450STejun Heo 109463638450STejun Heo /* 109563638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 109663638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 109763638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 109863638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 109963638450STejun Heo * should be using an unbound workqueue instead. 110063638450STejun Heo * 110163638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 110263638450STejun Heo * and report them so that they can be examined and converted to use unbound 110363638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 110463638450STejun Heo * function is tracked and reported with exponential backoff. 110563638450STejun Heo */ 110663638450STejun Heo #define WCI_MAX_ENTS 128 110763638450STejun Heo 110863638450STejun Heo struct wci_ent { 110963638450STejun Heo work_func_t func; 111063638450STejun Heo atomic64_t cnt; 111163638450STejun Heo struct hlist_node hash_node; 111263638450STejun Heo }; 111363638450STejun Heo 111463638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 111563638450STejun Heo static int wci_nr_ents; 111663638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 111763638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 111863638450STejun Heo 111963638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 112063638450STejun Heo { 112163638450STejun Heo struct wci_ent *ent; 112263638450STejun Heo 112363638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 112463638450STejun Heo (unsigned long)func) { 112563638450STejun Heo if (ent->func == func) 112663638450STejun Heo return ent; 112763638450STejun Heo } 112863638450STejun Heo return NULL; 112963638450STejun Heo } 113063638450STejun Heo 113163638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 113263638450STejun Heo { 113363638450STejun Heo struct wci_ent *ent; 113463638450STejun Heo 113563638450STejun Heo restart: 113663638450STejun Heo ent = wci_find_ent(func); 113763638450STejun Heo if (ent) { 113863638450STejun Heo u64 cnt; 113963638450STejun Heo 114063638450STejun Heo /* 114163638450STejun Heo * Start reporting from the fourth time and back off 114263638450STejun Heo * exponentially. 114363638450STejun Heo */ 114463638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 114563638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 114663638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 114763638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 114863638450STejun Heo atomic64_read(&ent->cnt)); 114963638450STejun Heo return; 115063638450STejun Heo } 115163638450STejun Heo 115263638450STejun Heo /* 115363638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 115463638450STejun Heo * is exhausted, something went really wrong and we probably made enough 115563638450STejun Heo * noise already. 115663638450STejun Heo */ 115763638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 115863638450STejun Heo return; 115963638450STejun Heo 116063638450STejun Heo raw_spin_lock(&wci_lock); 116163638450STejun Heo 116263638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 116363638450STejun Heo raw_spin_unlock(&wci_lock); 116463638450STejun Heo return; 116563638450STejun Heo } 116663638450STejun Heo 116763638450STejun Heo if (wci_find_ent(func)) { 116863638450STejun Heo raw_spin_unlock(&wci_lock); 116963638450STejun Heo goto restart; 117063638450STejun Heo } 117163638450STejun Heo 117263638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 117363638450STejun Heo ent->func = func; 117463638450STejun Heo atomic64_set(&ent->cnt, 1); 117563638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 117663638450STejun Heo 117763638450STejun Heo raw_spin_unlock(&wci_lock); 117863638450STejun Heo } 117963638450STejun Heo 118063638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 118163638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 118263638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 118363638450STejun Heo 1184c54d5046STejun Heo /** 11851da177e4SLinus Torvalds * wq_worker_running - a worker is running again 11861da177e4SLinus Torvalds * @task: task waking up 11871da177e4SLinus Torvalds * 11881da177e4SLinus Torvalds * This function is called when a worker returns from schedule() 11891da177e4SLinus Torvalds */ 11901da177e4SLinus Torvalds void wq_worker_running(struct task_struct *task) 11911da177e4SLinus Torvalds { 11921da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 11931da177e4SLinus Torvalds 1194c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 11951da177e4SLinus Torvalds return; 11961da177e4SLinus Torvalds 11971da177e4SLinus Torvalds /* 11981da177e4SLinus Torvalds * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 11991da177e4SLinus Torvalds * and the nr_running increment below, we may ruin the nr_running reset 12001da177e4SLinus Torvalds * and leave with an unexpected pool->nr_running == 1 on the newly unbound 12011da177e4SLinus Torvalds * pool. Protect against such race. 12021da177e4SLinus Torvalds */ 12031da177e4SLinus Torvalds preempt_disable(); 12041da177e4SLinus Torvalds if (!(worker->flags & WORKER_NOT_RUNNING)) 12051da177e4SLinus Torvalds worker->pool->nr_running++; 12061da177e4SLinus Torvalds preempt_enable(); 1207616db877STejun Heo 1208616db877STejun Heo /* 1209616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1210616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1211616db877STejun Heo */ 1212616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1213616db877STejun Heo 1214c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 12151da177e4SLinus Torvalds } 12161da177e4SLinus Torvalds 12171da177e4SLinus Torvalds /** 12181da177e4SLinus Torvalds * wq_worker_sleeping - a worker is going to sleep 12191da177e4SLinus Torvalds * @task: task going to sleep 12201da177e4SLinus Torvalds * 12211da177e4SLinus Torvalds * This function is called from schedule() when a busy worker is 12221da177e4SLinus Torvalds * going to sleep. 12231da177e4SLinus Torvalds */ 12241da177e4SLinus Torvalds void wq_worker_sleeping(struct task_struct *task) 12251da177e4SLinus Torvalds { 12261da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 12271da177e4SLinus Torvalds struct worker_pool *pool; 12281da177e4SLinus Torvalds 12291da177e4SLinus Torvalds /* 12301da177e4SLinus Torvalds * Rescuers, which may not have all the fields set up like normal 12311da177e4SLinus Torvalds * workers, also reach here, let's not access anything before 12321da177e4SLinus Torvalds * checking NOT_RUNNING. 12331da177e4SLinus Torvalds */ 12341da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) 12351da177e4SLinus Torvalds return; 12361da177e4SLinus Torvalds 12371da177e4SLinus Torvalds pool = worker->pool; 12381da177e4SLinus Torvalds 12391da177e4SLinus Torvalds /* Return if preempted before wq_worker_running() was reached */ 1240c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 12411da177e4SLinus Torvalds return; 12421da177e4SLinus Torvalds 1243c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 12441da177e4SLinus Torvalds raw_spin_lock_irq(&pool->lock); 12451da177e4SLinus Torvalds 12461da177e4SLinus Torvalds /* 12471da177e4SLinus Torvalds * Recheck in case unbind_workers() preempted us. We don't 12481da177e4SLinus Torvalds * want to decrement nr_running after the worker is unbound 12491da177e4SLinus Torvalds * and nr_running has been reset. 12501da177e4SLinus Torvalds */ 12511da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) { 12521da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 12531da177e4SLinus Torvalds return; 12541da177e4SLinus Torvalds } 12551da177e4SLinus Torvalds 12561da177e4SLinus Torvalds pool->nr_running--; 1257725e8ec5STejun Heo if (need_more_worker(pool)) { 1258725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 12591da177e4SLinus Torvalds wake_up_worker(pool); 1260725e8ec5STejun Heo } 12611da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 12621da177e4SLinus Torvalds } 12631da177e4SLinus Torvalds 12641da177e4SLinus Torvalds /** 1265616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1266616db877STejun Heo * @task: task currently running 1267616db877STejun Heo * 1268616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1269616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1270616db877STejun Heo */ 1271616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1272616db877STejun Heo { 1273616db877STejun Heo struct worker *worker = kthread_data(task); 1274616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1275616db877STejun Heo struct worker_pool *pool = worker->pool; 1276616db877STejun Heo 1277616db877STejun Heo if (!pwq) 1278616db877STejun Heo return; 1279616db877STejun Heo 12808a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 12818a1dd1e5STejun Heo 128218c8ae81SZqiang if (!wq_cpu_intensive_thresh_us) 128318c8ae81SZqiang return; 128418c8ae81SZqiang 1285616db877STejun Heo /* 1286616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1287616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1288616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1289c8f6219bSZqiang * 1290c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1291c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1292c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1293c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1294c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1295c8f6219bSZqiang * We probably want to make this prettier in the future. 1296616db877STejun Heo */ 1297c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1298616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1299616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1300616db877STejun Heo return; 1301616db877STejun Heo 1302616db877STejun Heo raw_spin_lock(&pool->lock); 1303616db877STejun Heo 1304616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 130563638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1306616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1307616db877STejun Heo 1308616db877STejun Heo if (need_more_worker(pool)) { 1309616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1310616db877STejun Heo wake_up_worker(pool); 1311616db877STejun Heo } 1312616db877STejun Heo 1313616db877STejun Heo raw_spin_unlock(&pool->lock); 1314616db877STejun Heo } 1315616db877STejun Heo 1316616db877STejun Heo /** 13171da177e4SLinus Torvalds * wq_worker_last_func - retrieve worker's last work function 13181da177e4SLinus Torvalds * @task: Task to retrieve last work function of. 13191da177e4SLinus Torvalds * 13201da177e4SLinus Torvalds * Determine the last function a worker executed. This is called from 13211da177e4SLinus Torvalds * the scheduler to get a worker's last known identity. 13221da177e4SLinus Torvalds * 13231da177e4SLinus Torvalds * CONTEXT: 13249b41ea72SAndrew Morton * raw_spin_lock_irq(rq->lock) 13251da177e4SLinus Torvalds * 13261da177e4SLinus Torvalds * This function is called during schedule() when a kworker is going 1327f756d5e2SNathan Lynch * to sleep. It's used by psi to identify aggregation workers during 1328f756d5e2SNathan Lynch * dequeuing, to allow periodic aggregation to shut-off when that 13291da177e4SLinus Torvalds * worker is the last task in the system or cgroup to go to sleep. 13301da177e4SLinus Torvalds * 13311da177e4SLinus Torvalds * As this function doesn't involve any workqueue-related locking, it 13321da177e4SLinus Torvalds * only returns stable values when called from inside the scheduler's 13331da177e4SLinus Torvalds * queuing and dequeuing paths, when @task, which must be a kworker, 13341da177e4SLinus Torvalds * is guaranteed to not be processing any works. 1335365970a1SDavid Howells * 1336365970a1SDavid Howells * Return: 1337365970a1SDavid Howells * The last work function %current executed as a worker, NULL if it 1338365970a1SDavid Howells * hasn't executed any work yet. 1339365970a1SDavid Howells */ 1340365970a1SDavid Howells work_func_t wq_worker_last_func(struct task_struct *task) 1341365970a1SDavid Howells { 1342365970a1SDavid Howells struct worker *worker = kthread_data(task); 1343365970a1SDavid Howells 1344365970a1SDavid Howells return worker->last_func; 1345365970a1SDavid Howells } 1346365970a1SDavid Howells 1347d302f017STejun Heo /** 13488864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 13498864b4e5STejun Heo * @pwq: pool_workqueue to get 13508864b4e5STejun Heo * 13518864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 13528864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 13538864b4e5STejun Heo */ 13548864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 13558864b4e5STejun Heo { 13568864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13578864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 13588864b4e5STejun Heo pwq->refcnt++; 13598864b4e5STejun Heo } 13608864b4e5STejun Heo 13618864b4e5STejun Heo /** 13628864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 13638864b4e5STejun Heo * @pwq: pool_workqueue to put 13648864b4e5STejun Heo * 13658864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 13668864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 13678864b4e5STejun Heo */ 13688864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 13698864b4e5STejun Heo { 13708864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13718864b4e5STejun Heo if (likely(--pwq->refcnt)) 13728864b4e5STejun Heo return; 13738864b4e5STejun Heo /* 1374967b494eSTejun Heo * @pwq can't be released under pool->lock, bounce to a dedicated 1375967b494eSTejun Heo * kthread_worker to avoid A-A deadlocks. 13768864b4e5STejun Heo */ 1377687a9aa5STejun Heo kthread_queue_work(pwq_release_worker, &pwq->release_work); 13788864b4e5STejun Heo } 13798864b4e5STejun Heo 1380dce90d47STejun Heo /** 1381dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1382dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1383dce90d47STejun Heo * 1384dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1385dce90d47STejun Heo */ 1386dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1387dce90d47STejun Heo { 1388dce90d47STejun Heo if (pwq) { 1389dce90d47STejun Heo /* 139024acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1391dce90d47STejun Heo * following lock operations are safe. 1392dce90d47STejun Heo */ 1393a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1394dce90d47STejun Heo put_pwq(pwq); 1395a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1396dce90d47STejun Heo } 1397dce90d47STejun Heo } 1398dce90d47STejun Heo 1399f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1400bf4ede01STejun Heo { 1401112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1402bf4ede01STejun Heo 1403bf4ede01STejun Heo trace_workqueue_activate_work(work); 140482607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 140582607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1406112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1407f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1408112202d9STejun Heo pwq->nr_active++; 1409bf4ede01STejun Heo } 1410bf4ede01STejun Heo 1411f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 14123aa62497SLai Jiangshan { 1413f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 14143aa62497SLai Jiangshan struct work_struct, entry); 14153aa62497SLai Jiangshan 1416f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 14173aa62497SLai Jiangshan } 14183aa62497SLai Jiangshan 1419bf4ede01STejun Heo /** 1420112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1421112202d9STejun Heo * @pwq: pwq of interest 1422c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1423bf4ede01STejun Heo * 1424bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1425112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1426bf4ede01STejun Heo * 1427bf4ede01STejun Heo * CONTEXT: 1428a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1429bf4ede01STejun Heo */ 1430c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1431bf4ede01STejun Heo { 1432c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1433c4560c2cSLai Jiangshan 1434018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1435112202d9STejun Heo pwq->nr_active--; 1436f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1437f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1438112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1439f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1440bf4ede01STejun Heo } 1441018f3a13SLai Jiangshan } 1442018f3a13SLai Jiangshan 1443018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1444bf4ede01STejun Heo 1445bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1446112202d9STejun Heo if (likely(pwq->flush_color != color)) 14478864b4e5STejun Heo goto out_put; 1448bf4ede01STejun Heo 1449bf4ede01STejun Heo /* are there still in-flight works? */ 1450112202d9STejun Heo if (pwq->nr_in_flight[color]) 14518864b4e5STejun Heo goto out_put; 1452bf4ede01STejun Heo 1453112202d9STejun Heo /* this pwq is done, clear flush_color */ 1454112202d9STejun Heo pwq->flush_color = -1; 1455bf4ede01STejun Heo 1456bf4ede01STejun Heo /* 1457112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1458bf4ede01STejun Heo * will handle the rest. 1459bf4ede01STejun Heo */ 1460112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1461112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 14628864b4e5STejun Heo out_put: 14638864b4e5STejun Heo put_pwq(pwq); 1464bf4ede01STejun Heo } 1465bf4ede01STejun Heo 146636e227d2STejun Heo /** 1467bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 146836e227d2STejun Heo * @work: work item to steal 146936e227d2STejun Heo * @is_dwork: @work is a delayed_work 1470bbb68dfaSTejun Heo * @flags: place to store irq state 147136e227d2STejun Heo * 147236e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1473d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 147436e227d2STejun Heo * 1475d185af30SYacine Belkadi * Return: 14763eb6b31bSMauro Carvalho Chehab * 14773eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 147836e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 147936e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 148036e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1481bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1482bbb68dfaSTejun Heo * for arbitrarily long 14833eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 148436e227d2STejun Heo * 1485d185af30SYacine Belkadi * Note: 1486bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1487e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1488e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1489e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1490bbb68dfaSTejun Heo * 1491bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1492bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1493bbb68dfaSTejun Heo * 1494e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1495bf4ede01STejun Heo */ 1496bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1497bbb68dfaSTejun Heo unsigned long *flags) 1498bf4ede01STejun Heo { 1499d565ed63STejun Heo struct worker_pool *pool; 1500112202d9STejun Heo struct pool_workqueue *pwq; 1501bf4ede01STejun Heo 1502bbb68dfaSTejun Heo local_irq_save(*flags); 1503bbb68dfaSTejun Heo 150436e227d2STejun Heo /* try to steal the timer if it exists */ 150536e227d2STejun Heo if (is_dwork) { 150636e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 150736e227d2STejun Heo 1508e0aecdd8STejun Heo /* 1509e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1510e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1511e0aecdd8STejun Heo * running on the local CPU. 1512e0aecdd8STejun Heo */ 151336e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 151436e227d2STejun Heo return 1; 151536e227d2STejun Heo } 151636e227d2STejun Heo 151736e227d2STejun Heo /* try to claim PENDING the normal way */ 1518bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1519bf4ede01STejun Heo return 0; 1520bf4ede01STejun Heo 152124acfb71SThomas Gleixner rcu_read_lock(); 1522bf4ede01STejun Heo /* 1523bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1524bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1525bf4ede01STejun Heo */ 1526d565ed63STejun Heo pool = get_work_pool(work); 1527d565ed63STejun Heo if (!pool) 1528bbb68dfaSTejun Heo goto fail; 1529bf4ede01STejun Heo 1530a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1531bf4ede01STejun Heo /* 1532112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1533112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1534112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1535112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1536112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 15370b3dae68SLai Jiangshan * item is currently queued on that pool. 1538bf4ede01STejun Heo */ 1539112202d9STejun Heo pwq = get_work_pwq(work); 1540112202d9STejun Heo if (pwq && pwq->pool == pool) { 1541bf4ede01STejun Heo debug_work_deactivate(work); 15423aa62497SLai Jiangshan 15433aa62497SLai Jiangshan /* 1544018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1545018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1546018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1547018f3a13SLai Jiangshan * 1548f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1549d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1550f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 155116062836STejun Heo * management later on and cause stall. Make sure the work 155216062836STejun Heo * item is activated before grabbing. 15533aa62497SLai Jiangshan */ 1554f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1555f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 15563aa62497SLai Jiangshan 1557bf4ede01STejun Heo list_del_init(&work->entry); 1558c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 155936e227d2STejun Heo 1560112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 15614468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 15624468a00fSLai Jiangshan 1563a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 156424acfb71SThomas Gleixner rcu_read_unlock(); 156536e227d2STejun Heo return 1; 1566bf4ede01STejun Heo } 1567a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1568bbb68dfaSTejun Heo fail: 156924acfb71SThomas Gleixner rcu_read_unlock(); 1570bbb68dfaSTejun Heo local_irq_restore(*flags); 1571bbb68dfaSTejun Heo if (work_is_canceling(work)) 1572bbb68dfaSTejun Heo return -ENOENT; 1573bbb68dfaSTejun Heo cpu_relax(); 157436e227d2STejun Heo return -EAGAIN; 1575bf4ede01STejun Heo } 1576bf4ede01STejun Heo 1577bf4ede01STejun Heo /** 1578706026c2STejun Heo * insert_work - insert a work into a pool 1579112202d9STejun Heo * @pwq: pwq @work belongs to 15804690c4abSTejun Heo * @work: work to insert 15814690c4abSTejun Heo * @head: insertion point 15824690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 15834690c4abSTejun Heo * 1584112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1585706026c2STejun Heo * work_struct flags. 15864690c4abSTejun Heo * 15874690c4abSTejun Heo * CONTEXT: 1588a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1589365970a1SDavid Howells */ 1590112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1591112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1592b89deed3SOleg Nesterov { 1593fe089f87STejun Heo debug_work_activate(work); 1594e1d8aa9fSFrederic Weisbecker 1595e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1596f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1597e89a85d6SWalter Wu 15984690c4abSTejun Heo /* we own @work, set data and link */ 1599112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 16001a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 16018864b4e5STejun Heo get_pwq(pwq); 1602b89deed3SOleg Nesterov } 1603b89deed3SOleg Nesterov 1604c8efcc25STejun Heo /* 1605c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 16068d03ecfeSTejun Heo * same workqueue. 1607c8efcc25STejun Heo */ 1608c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1609c8efcc25STejun Heo { 1610c8efcc25STejun Heo struct worker *worker; 1611c8efcc25STejun Heo 16128d03ecfeSTejun Heo worker = current_wq_worker(); 1613c8efcc25STejun Heo /* 1614bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 16158d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1616c8efcc25STejun Heo */ 1617112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1618c8efcc25STejun Heo } 1619c8efcc25STejun Heo 1620ef557180SMike Galbraith /* 1621ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1622ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1623ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1624ef557180SMike Galbraith */ 1625ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1626ef557180SMike Galbraith { 1627ef557180SMike Galbraith int new_cpu; 1628ef557180SMike Galbraith 1629f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1630ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1631ef557180SMike Galbraith return cpu; 1632a8ec5880SAmmar Faizi } else { 1633a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1634f303fccbSTejun Heo } 1635f303fccbSTejun Heo 1636ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1637ef557180SMike Galbraith return cpu; 1638ef557180SMike Galbraith 1639ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1640ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1641ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1642ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1643ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1644ef557180SMike Galbraith return cpu; 1645ef557180SMike Galbraith } 1646ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1647ef557180SMike Galbraith 1648ef557180SMike Galbraith return new_cpu; 1649ef557180SMike Galbraith } 1650ef557180SMike Galbraith 1651d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 16521da177e4SLinus Torvalds struct work_struct *work) 16531da177e4SLinus Torvalds { 1654112202d9STejun Heo struct pool_workqueue *pwq; 1655fe089f87STejun Heo struct worker_pool *last_pool, *pool; 16568a2e8e5dSTejun Heo unsigned int work_flags; 1657b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 16588930cabaSTejun Heo 16598930cabaSTejun Heo /* 16608930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 16618930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 16628930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 16638930cabaSTejun Heo * happen with IRQ disabled. 16648930cabaSTejun Heo */ 16658e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 16661da177e4SLinus Torvalds 16671e19ffc6STejun Heo 166833e3f0a3SRichard Clark /* 166933e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 167033e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 167133e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 167233e3f0a3SRichard Clark */ 167333e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 167433e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1675e41e704bSTejun Heo return; 167624acfb71SThomas Gleixner rcu_read_lock(); 16779e8cd2f5STejun Heo retry: 1678aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1679aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1680df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1681ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1682df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1683aa202f1fSHillf Danton } else { 1684aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1685aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1686687a9aa5STejun Heo pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 1687aa202f1fSHillf Danton } 1688f3421797STejun Heo 1689fe089f87STejun Heo pool = pwq->pool; 1690fe089f87STejun Heo 169118aa9effSTejun Heo /* 1692c9178087STejun Heo * If @work was previously on a different pool, it might still be 1693c9178087STejun Heo * running there, in which case the work needs to be queued on that 1694c9178087STejun Heo * pool to guarantee non-reentrancy. 169518aa9effSTejun Heo */ 1696c9e7cf27STejun Heo last_pool = get_work_pool(work); 1697fe089f87STejun Heo if (last_pool && last_pool != pool) { 169818aa9effSTejun Heo struct worker *worker; 169918aa9effSTejun Heo 1700a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 170118aa9effSTejun Heo 1702c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 170318aa9effSTejun Heo 1704112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1705c9178087STejun Heo pwq = worker->current_pwq; 1706fe089f87STejun Heo pool = pwq->pool; 1707fe089f87STejun Heo WARN_ON_ONCE(pool != last_pool); 17088594fadeSLai Jiangshan } else { 170918aa9effSTejun Heo /* meh... not running there, queue here */ 1710a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1711fe089f87STejun Heo raw_spin_lock(&pool->lock); 171218aa9effSTejun Heo } 17138930cabaSTejun Heo } else { 1714fe089f87STejun Heo raw_spin_lock(&pool->lock); 17158930cabaSTejun Heo } 1716502ca9d8STejun Heo 17179e8cd2f5STejun Heo /* 17189e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 17199e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 17209e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1721df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1722df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 17239e8cd2f5STejun Heo * make forward-progress. 17249e8cd2f5STejun Heo */ 17259e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 17269e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1727fe089f87STejun Heo raw_spin_unlock(&pool->lock); 17289e8cd2f5STejun Heo cpu_relax(); 17299e8cd2f5STejun Heo goto retry; 17309e8cd2f5STejun Heo } 17319e8cd2f5STejun Heo /* oops */ 17329e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 17339e8cd2f5STejun Heo wq->name, cpu); 17349e8cd2f5STejun Heo } 17359e8cd2f5STejun Heo 1736112202d9STejun Heo /* pwq determined, queue */ 1737112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1738502ca9d8STejun Heo 173924acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 174024acfb71SThomas Gleixner goto out; 17411e19ffc6STejun Heo 1742112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1743112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 17441e19ffc6STejun Heo 1745112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1746fe089f87STejun Heo if (list_empty(&pool->worklist)) 1747fe089f87STejun Heo pool->watchdog_ts = jiffies; 1748fe089f87STejun Heo 1749cdadf009STejun Heo trace_workqueue_activate_work(work); 1750112202d9STejun Heo pwq->nr_active++; 1751fe089f87STejun Heo insert_work(pwq, work, &pool->worklist, work_flags); 1752fe089f87STejun Heo 1753fe089f87STejun Heo if (__need_more_worker(pool)) 1754fe089f87STejun Heo wake_up_worker(pool); 17558a2e8e5dSTejun Heo } else { 1756f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1757fe089f87STejun Heo insert_work(pwq, work, &pwq->inactive_works, work_flags); 17588a2e8e5dSTejun Heo } 17591e19ffc6STejun Heo 176024acfb71SThomas Gleixner out: 1761fe089f87STejun Heo raw_spin_unlock(&pool->lock); 176224acfb71SThomas Gleixner rcu_read_unlock(); 17631da177e4SLinus Torvalds } 17641da177e4SLinus Torvalds 17650fcb78c2SRolf Eike Beer /** 1766c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1767c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1768c1a220e7SZhang Rui * @wq: workqueue to use 1769c1a220e7SZhang Rui * @work: work to queue 1770c1a220e7SZhang Rui * 1771c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1772443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1773443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1774854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1775854f5cc5SPaul E. McKenney * online will get a splat. 1776d185af30SYacine Belkadi * 1777d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1778c1a220e7SZhang Rui */ 1779d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1780d4283e93STejun Heo struct work_struct *work) 1781c1a220e7SZhang Rui { 1782d4283e93STejun Heo bool ret = false; 17838930cabaSTejun Heo unsigned long flags; 17848930cabaSTejun Heo 17858930cabaSTejun Heo local_irq_save(flags); 1786c1a220e7SZhang Rui 178722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17884690c4abSTejun Heo __queue_work(cpu, wq, work); 1789d4283e93STejun Heo ret = true; 1790c1a220e7SZhang Rui } 17918930cabaSTejun Heo 17928930cabaSTejun Heo local_irq_restore(flags); 1793c1a220e7SZhang Rui return ret; 1794c1a220e7SZhang Rui } 1795ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1796c1a220e7SZhang Rui 17978204e0c1SAlexander Duyck /** 17988204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 17998204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 18008204e0c1SAlexander Duyck * 18018204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 18028204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 18038204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 18048204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 18058204e0c1SAlexander Duyck */ 18068204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 18078204e0c1SAlexander Duyck { 18088204e0c1SAlexander Duyck int cpu; 18098204e0c1SAlexander Duyck 18108204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 18118204e0c1SAlexander Duyck if (!wq_numa_enabled) 18128204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 18138204e0c1SAlexander Duyck 18148204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 18158204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 18168204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 18178204e0c1SAlexander Duyck 18188204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 18198204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 18208204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 18218204e0c1SAlexander Duyck return cpu; 18228204e0c1SAlexander Duyck 18238204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 18248204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 18258204e0c1SAlexander Duyck 18268204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 18278204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 18288204e0c1SAlexander Duyck } 18298204e0c1SAlexander Duyck 18308204e0c1SAlexander Duyck /** 18318204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 18328204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 18338204e0c1SAlexander Duyck * @wq: workqueue to use 18348204e0c1SAlexander Duyck * @work: work to queue 18358204e0c1SAlexander Duyck * 18368204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 18378204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 18388204e0c1SAlexander Duyck * NUMA node. 18398204e0c1SAlexander Duyck * 18408204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 18418204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 18428204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 18438204e0c1SAlexander Duyck * 18448204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 18458204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 18468204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 18478204e0c1SAlexander Duyck * 18488204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 18498204e0c1SAlexander Duyck */ 18508204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18518204e0c1SAlexander Duyck struct work_struct *work) 18528204e0c1SAlexander Duyck { 18538204e0c1SAlexander Duyck unsigned long flags; 18548204e0c1SAlexander Duyck bool ret = false; 18558204e0c1SAlexander Duyck 18568204e0c1SAlexander Duyck /* 18578204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 18588204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 18598204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 18608204e0c1SAlexander Duyck * 18618204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 18628204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 18638204e0c1SAlexander Duyck * some round robin type logic. 18648204e0c1SAlexander Duyck */ 18658204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 18668204e0c1SAlexander Duyck 18678204e0c1SAlexander Duyck local_irq_save(flags); 18688204e0c1SAlexander Duyck 18698204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18708204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 18718204e0c1SAlexander Duyck 18728204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 18738204e0c1SAlexander Duyck ret = true; 18748204e0c1SAlexander Duyck } 18758204e0c1SAlexander Duyck 18768204e0c1SAlexander Duyck local_irq_restore(flags); 18778204e0c1SAlexander Duyck return ret; 18788204e0c1SAlexander Duyck } 18798204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 18808204e0c1SAlexander Duyck 18818c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 18821da177e4SLinus Torvalds { 18838c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 18841da177e4SLinus Torvalds 1885e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 188660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 18871da177e4SLinus Torvalds } 18881438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 18891da177e4SLinus Torvalds 18907beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 189152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18921da177e4SLinus Torvalds { 18937beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 18947beb2edfSTejun Heo struct work_struct *work = &dwork->work; 18951da177e4SLinus Torvalds 1896637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 18974b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1898fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1899fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 19007beb2edfSTejun Heo 19018852aac2STejun Heo /* 19028852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 19038852aac2STejun Heo * both optimization and correctness. The earliest @timer can 19048852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 19058852aac2STejun Heo * on that there's no such delay when @delay is 0. 19068852aac2STejun Heo */ 19078852aac2STejun Heo if (!delay) { 19088852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 19098852aac2STejun Heo return; 19108852aac2STejun Heo } 19118852aac2STejun Heo 191260c057bcSLai Jiangshan dwork->wq = wq; 19131265057fSTejun Heo dwork->cpu = cpu; 19147beb2edfSTejun Heo timer->expires = jiffies + delay; 19157beb2edfSTejun Heo 1916041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 19177beb2edfSTejun Heo add_timer_on(timer, cpu); 1918041bd12eSTejun Heo else 1919041bd12eSTejun Heo add_timer(timer); 19207beb2edfSTejun Heo } 19211da177e4SLinus Torvalds 19220fcb78c2SRolf Eike Beer /** 19230fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 19240fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 19250fcb78c2SRolf Eike Beer * @wq: workqueue to use 1926af9997e4SRandy Dunlap * @dwork: work to queue 19270fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 19280fcb78c2SRolf Eike Beer * 1929d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1930715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1931715f1300STejun Heo * execution. 19320fcb78c2SRolf Eike Beer */ 1933d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 193452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 19357a6bc1cdSVenkatesh Pallipadi { 193652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1937d4283e93STejun Heo bool ret = false; 19388930cabaSTejun Heo unsigned long flags; 19398930cabaSTejun Heo 19408930cabaSTejun Heo /* read the comment in __queue_work() */ 19418930cabaSTejun Heo local_irq_save(flags); 19427a6bc1cdSVenkatesh Pallipadi 194322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 19447beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1945d4283e93STejun Heo ret = true; 19467a6bc1cdSVenkatesh Pallipadi } 19478930cabaSTejun Heo 19488930cabaSTejun Heo local_irq_restore(flags); 19497a6bc1cdSVenkatesh Pallipadi return ret; 19507a6bc1cdSVenkatesh Pallipadi } 1951ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19521da177e4SLinus Torvalds 1953c8e55f36STejun Heo /** 19548376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19558376fe22STejun Heo * @cpu: CPU number to execute work on 19568376fe22STejun Heo * @wq: workqueue to use 19578376fe22STejun Heo * @dwork: work to queue 19588376fe22STejun Heo * @delay: number of jiffies to wait before queueing 19598376fe22STejun Heo * 19608376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 19618376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 19628376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 19638376fe22STejun Heo * current state. 19648376fe22STejun Heo * 1965d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 19668376fe22STejun Heo * pending and its timer was modified. 19678376fe22STejun Heo * 1968e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 19698376fe22STejun Heo * See try_to_grab_pending() for details. 19708376fe22STejun Heo */ 19718376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 19728376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 19738376fe22STejun Heo { 19748376fe22STejun Heo unsigned long flags; 19758376fe22STejun Heo int ret; 19768376fe22STejun Heo 19778376fe22STejun Heo do { 19788376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 19798376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 19808376fe22STejun Heo 19818376fe22STejun Heo if (likely(ret >= 0)) { 19828376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 19838376fe22STejun Heo local_irq_restore(flags); 19848376fe22STejun Heo } 19858376fe22STejun Heo 19868376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 19878376fe22STejun Heo return ret; 19888376fe22STejun Heo } 19898376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 19908376fe22STejun Heo 199105f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 199205f0fe6bSTejun Heo { 199305f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 199405f0fe6bSTejun Heo 199505f0fe6bSTejun Heo /* read the comment in __queue_work() */ 199605f0fe6bSTejun Heo local_irq_disable(); 199705f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 199805f0fe6bSTejun Heo local_irq_enable(); 199905f0fe6bSTejun Heo } 200005f0fe6bSTejun Heo 200105f0fe6bSTejun Heo /** 200205f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 200305f0fe6bSTejun Heo * @wq: workqueue to use 200405f0fe6bSTejun Heo * @rwork: work to queue 200505f0fe6bSTejun Heo * 200605f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 200705f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 2008bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 200905f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 201005f0fe6bSTejun Heo */ 201105f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 201205f0fe6bSTejun Heo { 201305f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 201405f0fe6bSTejun Heo 201505f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 201605f0fe6bSTejun Heo rwork->wq = wq; 2017a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 201805f0fe6bSTejun Heo return true; 201905f0fe6bSTejun Heo } 202005f0fe6bSTejun Heo 202105f0fe6bSTejun Heo return false; 202205f0fe6bSTejun Heo } 202305f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 202405f0fe6bSTejun Heo 2025f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 2026c34056a3STejun Heo { 2027c34056a3STejun Heo struct worker *worker; 2028c34056a3STejun Heo 2029f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2030c8e55f36STejun Heo if (worker) { 2031c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 2032affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 2033da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2034e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2035e22bee78STejun Heo worker->flags = WORKER_PREP; 2036c8e55f36STejun Heo } 2037c34056a3STejun Heo return worker; 2038c34056a3STejun Heo } 2039c34056a3STejun Heo 2040c34056a3STejun Heo /** 20414736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20424736cbf7SLai Jiangshan * @worker: worker to be attached 20434736cbf7SLai Jiangshan * @pool: the target pool 20444736cbf7SLai Jiangshan * 20454736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20464736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20474736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20484736cbf7SLai Jiangshan */ 20494736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 20504736cbf7SLai Jiangshan struct worker_pool *pool) 20514736cbf7SLai Jiangshan { 20521258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 20534736cbf7SLai Jiangshan 20544736cbf7SLai Jiangshan /* 20551258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 20561258fae7STejun Heo * stable across this function. See the comments above the flag 20571258fae7STejun Heo * definition for details. 20584736cbf7SLai Jiangshan */ 20594736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 20604736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 20615c25b5ffSPeter Zijlstra else 20625c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 20634736cbf7SLai Jiangshan 2064640f17c8SPeter Zijlstra if (worker->rescue_wq) 2065640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 2066640f17c8SPeter Zijlstra 20674736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2068a2d812a2STejun Heo worker->pool = pool; 20694736cbf7SLai Jiangshan 20701258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 20714736cbf7SLai Jiangshan } 20724736cbf7SLai Jiangshan 20734736cbf7SLai Jiangshan /** 207460f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 207560f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 207660f5a4bcSLai Jiangshan * 20774736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 20784736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 20794736cbf7SLai Jiangshan * other reference to the pool. 208060f5a4bcSLai Jiangshan */ 2081a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 208260f5a4bcSLai Jiangshan { 2083a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 208460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 208560f5a4bcSLai Jiangshan 20861258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2087a2d812a2STejun Heo 20885c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2089da028469SLai Jiangshan list_del(&worker->node); 2090a2d812a2STejun Heo worker->pool = NULL; 2091a2d812a2STejun Heo 2092e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 209360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 20941258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 209560f5a4bcSLai Jiangshan 2096b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2097b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2098b62c0751SLai Jiangshan 209960f5a4bcSLai Jiangshan if (detach_completion) 210060f5a4bcSLai Jiangshan complete(detach_completion); 210160f5a4bcSLai Jiangshan } 210260f5a4bcSLai Jiangshan 210360f5a4bcSLai Jiangshan /** 2104c34056a3STejun Heo * create_worker - create a new workqueue worker 210563d95a91STejun Heo * @pool: pool the new worker will belong to 2106c34056a3STejun Heo * 2107051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2108c34056a3STejun Heo * 2109c34056a3STejun Heo * CONTEXT: 2110c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2111c34056a3STejun Heo * 2112d185af30SYacine Belkadi * Return: 2113c34056a3STejun Heo * Pointer to the newly created worker. 2114c34056a3STejun Heo */ 2115bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2116c34056a3STejun Heo { 2117e441b56fSZhen Lei struct worker *worker; 2118e441b56fSZhen Lei int id; 2119e3c916a4STejun Heo char id_buf[16]; 2120c34056a3STejun Heo 21217cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2122e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 21233f0ea0b8SPetr Mladek if (id < 0) { 21243f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 21253f0ea0b8SPetr Mladek ERR_PTR(id)); 2126e441b56fSZhen Lei return NULL; 21273f0ea0b8SPetr Mladek } 2128c34056a3STejun Heo 2129f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 21303f0ea0b8SPetr Mladek if (!worker) { 21313f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2132c34056a3STejun Heo goto fail; 21333f0ea0b8SPetr Mladek } 2134c34056a3STejun Heo 2135c34056a3STejun Heo worker->id = id; 2136c34056a3STejun Heo 213729c91e99STejun Heo if (pool->cpu >= 0) 2138e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2139e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2140f3421797STejun Heo else 2141e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2142e3c916a4STejun Heo 2143f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2144e3c916a4STejun Heo "kworker/%s", id_buf); 21453f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 214660f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 214760f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 214860f54038SPetr Mladek id_buf); 214960f54038SPetr Mladek } else { 21503f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 21513f0ea0b8SPetr Mladek worker->task); 215260f54038SPetr Mladek } 2153c34056a3STejun Heo goto fail; 21543f0ea0b8SPetr Mladek } 2155c34056a3STejun Heo 215691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 215725834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 215891151228SOleg Nesterov 2159da028469SLai Jiangshan /* successful, attach the worker to the pool */ 21604736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2161822d8405STejun Heo 2162051e1850SLai Jiangshan /* start the newly created worker */ 2163a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2164051e1850SLai Jiangshan worker->pool->nr_workers++; 2165051e1850SLai Jiangshan worker_enter_idle(worker); 2166051e1850SLai Jiangshan wake_up_process(worker->task); 2167a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2168051e1850SLai Jiangshan 2169c34056a3STejun Heo return worker; 2170822d8405STejun Heo 2171c34056a3STejun Heo fail: 2172e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2173c34056a3STejun Heo kfree(worker); 2174c34056a3STejun Heo return NULL; 2175c34056a3STejun Heo } 2176c34056a3STejun Heo 2177793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2178793777bcSValentin Schneider { 2179793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2180793777bcSValentin Schneider 2181793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2182793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2183793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2184793777bcSValentin Schneider else 2185793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2186793777bcSValentin Schneider } 2187793777bcSValentin Schneider 2188e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2189e02b9312SValentin Schneider { 2190e02b9312SValentin Schneider struct worker *worker, *tmp; 2191e02b9312SValentin Schneider 2192e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2193e02b9312SValentin Schneider list_del_init(&worker->entry); 2194e02b9312SValentin Schneider unbind_worker(worker); 2195e02b9312SValentin Schneider /* 2196e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2197e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2198e02b9312SValentin Schneider * wouldn't have gotten here. 2199c34056a3STejun Heo * 2200e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2201e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2202e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2203e02b9312SValentin Schneider * outside of pool->lock. 2204e02b9312SValentin Schneider */ 2205e02b9312SValentin Schneider wake_up_process(worker->task); 2206e02b9312SValentin Schneider } 2207e02b9312SValentin Schneider } 2208e02b9312SValentin Schneider 2209e02b9312SValentin Schneider /** 2210e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2211e02b9312SValentin Schneider * @worker: worker to be destroyed 2212e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2213e02b9312SValentin Schneider * 2214e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2215e02b9312SValentin Schneider * should be idle. 2216c8e55f36STejun Heo * 2217c8e55f36STejun Heo * CONTEXT: 2218a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2219c34056a3STejun Heo */ 2220e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2221c34056a3STejun Heo { 2222bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2223c34056a3STejun Heo 2224cd549687STejun Heo lockdep_assert_held(&pool->lock); 2225e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2226cd549687STejun Heo 2227c34056a3STejun Heo /* sanity check frenzy */ 22286183c009STejun Heo if (WARN_ON(worker->current_work) || 222973eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 223073eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 22316183c009STejun Heo return; 2232c34056a3STejun Heo 2233bd7bdd43STejun Heo pool->nr_workers--; 2234bd7bdd43STejun Heo pool->nr_idle--; 2235c8e55f36STejun Heo 2236cb444766STejun Heo worker->flags |= WORKER_DIE; 2237e02b9312SValentin Schneider 2238e02b9312SValentin Schneider list_move(&worker->entry, list); 2239e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2240c34056a3STejun Heo } 2241c34056a3STejun Heo 22423f959aa3SValentin Schneider /** 22433f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 22443f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 22453f959aa3SValentin Schneider * 22463f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 22473f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 22483f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 22493f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 22503f959aa3SValentin Schneider * it expire and re-evaluate things from there. 22513f959aa3SValentin Schneider */ 225232a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2253e22bee78STejun Heo { 225432a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 22553f959aa3SValentin Schneider bool do_cull = false; 22563f959aa3SValentin Schneider 22573f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 22583f959aa3SValentin Schneider return; 22593f959aa3SValentin Schneider 22603f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 22613f959aa3SValentin Schneider 22623f959aa3SValentin Schneider if (too_many_workers(pool)) { 22633f959aa3SValentin Schneider struct worker *worker; 22643f959aa3SValentin Schneider unsigned long expires; 22653f959aa3SValentin Schneider 22663f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 22673f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 22683f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 22693f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 22703f959aa3SValentin Schneider 22713f959aa3SValentin Schneider if (!do_cull) 22723f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 22733f959aa3SValentin Schneider } 22743f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 22753f959aa3SValentin Schneider 22763f959aa3SValentin Schneider if (do_cull) 22773f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 22783f959aa3SValentin Schneider } 22793f959aa3SValentin Schneider 22803f959aa3SValentin Schneider /** 22813f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 22823f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 22833f959aa3SValentin Schneider * 22843f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 22853f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2286e02b9312SValentin Schneider * 2287e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2288e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2289e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 22903f959aa3SValentin Schneider */ 22913f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 22923f959aa3SValentin Schneider { 22933f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 22949680540cSYang Yingliang LIST_HEAD(cull_list); 2295e22bee78STejun Heo 2296e02b9312SValentin Schneider /* 2297e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2298e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2299e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2300e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2301e02b9312SValentin Schneider */ 2302e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2303a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2304e22bee78STejun Heo 23053347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2306e22bee78STejun Heo struct worker *worker; 2307e22bee78STejun Heo unsigned long expires; 2308e22bee78STejun Heo 230963d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2310e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2311e22bee78STejun Heo 23123347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 231363d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 23143347fc9fSLai Jiangshan break; 2315e22bee78STejun Heo } 23163347fc9fSLai Jiangshan 2317e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2318e22bee78STejun Heo } 2319e22bee78STejun Heo 2320a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2321e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2322e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2323e22bee78STejun Heo } 2324e22bee78STejun Heo 2325493a1724STejun Heo static void send_mayday(struct work_struct *work) 2326e22bee78STejun Heo { 2327112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2328112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2329493a1724STejun Heo 23302e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2331e22bee78STejun Heo 2332493008a8STejun Heo if (!wq->rescuer) 2333493a1724STejun Heo return; 2334e22bee78STejun Heo 2335e22bee78STejun Heo /* mayday mayday mayday */ 2336493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 233777668c8bSLai Jiangshan /* 233877668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 233977668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 234077668c8bSLai Jiangshan * rescuer is done with it. 234177668c8bSLai Jiangshan */ 234277668c8bSLai Jiangshan get_pwq(pwq); 2343493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2344e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2345725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2346493a1724STejun Heo } 2347e22bee78STejun Heo } 2348e22bee78STejun Heo 234932a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2350e22bee78STejun Heo { 235132a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2352e22bee78STejun Heo struct work_struct *work; 2353e22bee78STejun Heo 2354a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2355a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2356e22bee78STejun Heo 235763d95a91STejun Heo if (need_to_create_worker(pool)) { 2358e22bee78STejun Heo /* 2359e22bee78STejun Heo * We've been trying to create a new worker but 2360e22bee78STejun Heo * haven't been successful. We might be hitting an 2361e22bee78STejun Heo * allocation deadlock. Send distress signals to 2362e22bee78STejun Heo * rescuers. 2363e22bee78STejun Heo */ 236463d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2365e22bee78STejun Heo send_mayday(work); 2366e22bee78STejun Heo } 2367e22bee78STejun Heo 2368a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2369a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2370e22bee78STejun Heo 237163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2372e22bee78STejun Heo } 2373e22bee78STejun Heo 2374e22bee78STejun Heo /** 2375e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 237663d95a91STejun Heo * @pool: pool to create a new worker for 2377e22bee78STejun Heo * 237863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2379e22bee78STejun Heo * have at least one idle worker on return from this function. If 2380e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 238163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2382e22bee78STejun Heo * possible allocation deadlock. 2383e22bee78STejun Heo * 2384c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2385c5aa87bbSTejun Heo * may_start_working() %true. 2386e22bee78STejun Heo * 2387e22bee78STejun Heo * LOCKING: 2388a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2389e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2390e22bee78STejun Heo * manager. 2391e22bee78STejun Heo */ 239229187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2393d565ed63STejun Heo __releases(&pool->lock) 2394d565ed63STejun Heo __acquires(&pool->lock) 2395e22bee78STejun Heo { 2396e22bee78STejun Heo restart: 2397a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 23989f9c2364STejun Heo 2399e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 240063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2401e22bee78STejun Heo 2402e22bee78STejun Heo while (true) { 2403051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2404e22bee78STejun Heo break; 2405e22bee78STejun Heo 2406e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 24079f9c2364STejun Heo 240863d95a91STejun Heo if (!need_to_create_worker(pool)) 2409e22bee78STejun Heo break; 2410e22bee78STejun Heo } 2411e22bee78STejun Heo 241263d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2413a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2414051e1850SLai Jiangshan /* 2415051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2416051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2417051e1850SLai Jiangshan * already become busy. 2418051e1850SLai Jiangshan */ 241963d95a91STejun Heo if (need_to_create_worker(pool)) 2420e22bee78STejun Heo goto restart; 2421e22bee78STejun Heo } 2422e22bee78STejun Heo 2423e22bee78STejun Heo /** 2424e22bee78STejun Heo * manage_workers - manage worker pool 2425e22bee78STejun Heo * @worker: self 2426e22bee78STejun Heo * 2427706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2428e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2429706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2430e22bee78STejun Heo * 2431e22bee78STejun Heo * The caller can safely start processing works on false return. On 2432e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2433e22bee78STejun Heo * and may_start_working() is true. 2434e22bee78STejun Heo * 2435e22bee78STejun Heo * CONTEXT: 2436a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2437e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2438e22bee78STejun Heo * 2439d185af30SYacine Belkadi * Return: 244029187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 244129187a9eSTejun Heo * start processing works, %true if management function was performed and 244229187a9eSTejun Heo * the conditions that the caller verified before calling the function may 244329187a9eSTejun Heo * no longer be true. 2444e22bee78STejun Heo */ 2445e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2446e22bee78STejun Heo { 244763d95a91STejun Heo struct worker_pool *pool = worker->pool; 2448e22bee78STejun Heo 2449692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 245029187a9eSTejun Heo return false; 2451692b4825STejun Heo 2452692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 24532607d7a6STejun Heo pool->manager = worker; 2454e22bee78STejun Heo 245529187a9eSTejun Heo maybe_create_worker(pool); 2456e22bee78STejun Heo 24572607d7a6STejun Heo pool->manager = NULL; 2458692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2459d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 246029187a9eSTejun Heo return true; 2461e22bee78STejun Heo } 2462e22bee78STejun Heo 2463a62428c0STejun Heo /** 2464a62428c0STejun Heo * process_one_work - process single work 2465c34056a3STejun Heo * @worker: self 2466a62428c0STejun Heo * @work: work to process 2467a62428c0STejun Heo * 2468a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2469a62428c0STejun Heo * process a single work including synchronization against and 2470a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2471a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2472a62428c0STejun Heo * call this function to process a work. 2473a62428c0STejun Heo * 2474a62428c0STejun Heo * CONTEXT: 2475a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2476a62428c0STejun Heo */ 2477c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2478d565ed63STejun Heo __releases(&pool->lock) 2479d565ed63STejun Heo __acquires(&pool->lock) 24801da177e4SLinus Torvalds { 2481112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2482bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2483c4560c2cSLai Jiangshan unsigned long work_data; 24847e11629dSTejun Heo struct worker *collision; 24854e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 24864e6045f1SJohannes Berg /* 2487a62428c0STejun Heo * It is permissible to free the struct work_struct from 2488a62428c0STejun Heo * inside the function that is called from it, this we need to 2489a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2490a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2491a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 24924e6045f1SJohannes Berg */ 24934d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 24944d82a1deSPeter Zijlstra 24954d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 24964e6045f1SJohannes Berg #endif 2497807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 249885327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2499ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 250025511a47STejun Heo 25017e11629dSTejun Heo /* 25027e11629dSTejun Heo * A single work shouldn't be executed concurrently by 25037e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 25047e11629dSTejun Heo * already processing the work. If so, defer the work to the 25057e11629dSTejun Heo * currently executing one. 25067e11629dSTejun Heo */ 2507c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 25087e11629dSTejun Heo if (unlikely(collision)) { 25097e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 25107e11629dSTejun Heo return; 25117e11629dSTejun Heo } 25121da177e4SLinus Torvalds 25138930cabaSTejun Heo /* claim and dequeue */ 25141da177e4SLinus Torvalds debug_work_deactivate(work); 2515c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2516c34056a3STejun Heo worker->current_work = work; 2517a2c1c57bSTejun Heo worker->current_func = work->func; 2518112202d9STejun Heo worker->current_pwq = pwq; 2519616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2520c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2521d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 25227a22ad75STejun Heo 25238bf89593STejun Heo /* 25248bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 25258bf89593STejun Heo * overridden through set_worker_desc(). 25268bf89593STejun Heo */ 25278bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 25288bf89593STejun Heo 2529a62428c0STejun Heo list_del_init(&work->entry); 2530a62428c0STejun Heo 2531649027d7STejun Heo /* 2532228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2533228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2534228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2535228f1d00SLai Jiangshan * execution of the pending work items. 2536fb0e7bebSTejun Heo */ 2537616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2538228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2539fb0e7bebSTejun Heo 2540974271c4STejun Heo /* 2541a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2542a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2543a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2544a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2545228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2546974271c4STejun Heo */ 2547a489a03eSLai Jiangshan if (need_more_worker(pool)) 254863d95a91STejun Heo wake_up_worker(pool); 2549974271c4STejun Heo 25508930cabaSTejun Heo /* 25517c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2552d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 255323657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 255423657bb1STejun Heo * disabled. 25558930cabaSTejun Heo */ 25567c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 25571da177e4SLinus Torvalds 2558a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2559365970a1SDavid Howells 2560a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 25613295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2562e6f3faa7SPeter Zijlstra /* 2563f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2564f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2565e6f3faa7SPeter Zijlstra * 2566e6f3faa7SPeter Zijlstra * However, that would result in: 2567e6f3faa7SPeter Zijlstra * 2568e6f3faa7SPeter Zijlstra * A(W1) 2569e6f3faa7SPeter Zijlstra * WFC(C) 2570e6f3faa7SPeter Zijlstra * A(W1) 2571e6f3faa7SPeter Zijlstra * C(C) 2572e6f3faa7SPeter Zijlstra * 2573e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2574e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2575e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2576f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2577e6f3faa7SPeter Zijlstra * these locks. 2578e6f3faa7SPeter Zijlstra * 2579e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2580e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2581e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2582e6f3faa7SPeter Zijlstra */ 2583f52be570SPeter Zijlstra lockdep_invariant_state(true); 2584725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2585e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2586a2c1c57bSTejun Heo worker->current_func(work); 2587e36c886aSArjan van de Ven /* 2588e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2589e36c886aSArjan van de Ven * point will only record its address. 2590e36c886aSArjan van de Ven */ 25911c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2592725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 25933295f0efSIngo Molnar lock_map_release(&lockdep_map); 2594112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 25951da177e4SLinus Torvalds 2596d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2597044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2598d75f773cSSakari Ailus " last function: %ps\n", 2599a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2600a2c1c57bSTejun Heo worker->current_func); 2601d5abe669SPeter Zijlstra debug_show_held_locks(current); 2602d5abe669SPeter Zijlstra dump_stack(); 2603d5abe669SPeter Zijlstra } 2604d5abe669SPeter Zijlstra 2605b22ce278STejun Heo /* 2606025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2607b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2608b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2609b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2610789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2611789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2612b22ce278STejun Heo */ 2613a7e6425eSPaul E. McKenney cond_resched(); 2614b22ce278STejun Heo 2615a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2616a62428c0STejun Heo 2617616db877STejun Heo /* 2618616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2619616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2620616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2621616db877STejun Heo */ 2622fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2623fb0e7bebSTejun Heo 26241b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 26251b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 26261b69ac6bSJohannes Weiner 2627a62428c0STejun Heo /* we're done with it, release */ 262842f8570fSSasha Levin hash_del(&worker->hentry); 2629c34056a3STejun Heo worker->current_work = NULL; 2630a2c1c57bSTejun Heo worker->current_func = NULL; 2631112202d9STejun Heo worker->current_pwq = NULL; 2632d812796eSLai Jiangshan worker->current_color = INT_MAX; 2633c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26341da177e4SLinus Torvalds } 26351da177e4SLinus Torvalds 2636affee4b2STejun Heo /** 2637affee4b2STejun Heo * process_scheduled_works - process scheduled works 2638affee4b2STejun Heo * @worker: self 2639affee4b2STejun Heo * 2640affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2641affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2642affee4b2STejun Heo * fetches a work from the top and executes it. 2643affee4b2STejun Heo * 2644affee4b2STejun Heo * CONTEXT: 2645a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2646affee4b2STejun Heo * multiple times. 2647affee4b2STejun Heo */ 2648affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26491da177e4SLinus Torvalds { 2650c0ab017dSTejun Heo struct work_struct *work; 2651c0ab017dSTejun Heo bool first = true; 2652c0ab017dSTejun Heo 2653c0ab017dSTejun Heo while ((work = list_first_entry_or_null(&worker->scheduled, 2654c0ab017dSTejun Heo struct work_struct, entry))) { 2655c0ab017dSTejun Heo if (first) { 2656c0ab017dSTejun Heo worker->pool->watchdog_ts = jiffies; 2657c0ab017dSTejun Heo first = false; 2658c0ab017dSTejun Heo } 2659c34056a3STejun Heo process_one_work(worker, work); 2660a62428c0STejun Heo } 26611da177e4SLinus Torvalds } 26621da177e4SLinus Torvalds 2663197f6accSTejun Heo static void set_pf_worker(bool val) 2664197f6accSTejun Heo { 2665197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2666197f6accSTejun Heo if (val) 2667197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2668197f6accSTejun Heo else 2669197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2670197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2671197f6accSTejun Heo } 2672197f6accSTejun Heo 26734690c4abSTejun Heo /** 26744690c4abSTejun Heo * worker_thread - the worker thread function 2675c34056a3STejun Heo * @__worker: self 26764690c4abSTejun Heo * 2677c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2678c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2679c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2680c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2681c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2682d185af30SYacine Belkadi * 2683d185af30SYacine Belkadi * Return: 0 26844690c4abSTejun Heo */ 2685c34056a3STejun Heo static int worker_thread(void *__worker) 26861da177e4SLinus Torvalds { 2687c34056a3STejun Heo struct worker *worker = __worker; 2688bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 26891da177e4SLinus Torvalds 2690e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2691197f6accSTejun Heo set_pf_worker(true); 2692c8e55f36STejun Heo woke_up: 2693a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2694affee4b2STejun Heo 2695a9ab775bSTejun Heo /* am I supposed to die? */ 2696a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2697a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2698197f6accSTejun Heo set_pf_worker(false); 269960f5a4bcSLai Jiangshan 270060f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2701e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2702a2d812a2STejun Heo worker_detach_from_pool(worker); 2703e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 270460f5a4bcSLai Jiangshan kfree(worker); 2705c8e55f36STejun Heo return 0; 2706c8e55f36STejun Heo } 2707c8e55f36STejun Heo 2708c8e55f36STejun Heo worker_leave_idle(worker); 2709db7bccf4STejun Heo recheck: 2710e22bee78STejun Heo /* no more worker necessary? */ 271163d95a91STejun Heo if (!need_more_worker(pool)) 2712e22bee78STejun Heo goto sleep; 2713e22bee78STejun Heo 2714e22bee78STejun Heo /* do we need to manage? */ 271563d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2716e22bee78STejun Heo goto recheck; 2717e22bee78STejun Heo 2718c8e55f36STejun Heo /* 2719c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2720c8e55f36STejun Heo * preparing to process a work or actually processing it. 2721c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2722c8e55f36STejun Heo */ 27236183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2724c8e55f36STejun Heo 2725e22bee78STejun Heo /* 2726a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2727a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2728a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2729a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2730a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2731e22bee78STejun Heo */ 2732a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2733e22bee78STejun Heo 2734e22bee78STejun Heo do { 2735affee4b2STejun Heo struct work_struct *work = 2736bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2737affee4b2STejun Heo struct work_struct, entry); 2738affee4b2STejun Heo 2739c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2740affee4b2STejun Heo process_scheduled_works(worker); 274163d95a91STejun Heo } while (keep_working(pool)); 2742affee4b2STejun Heo 2743228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2744d313dd85STejun Heo sleep: 2745c8e55f36STejun Heo /* 2746d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2747d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2748d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2749d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2750d565ed63STejun Heo * event. 2751c8e55f36STejun Heo */ 2752c8e55f36STejun Heo worker_enter_idle(worker); 2753c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2754a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27551da177e4SLinus Torvalds schedule(); 2756c8e55f36STejun Heo goto woke_up; 27571da177e4SLinus Torvalds } 27581da177e4SLinus Torvalds 2759e22bee78STejun Heo /** 2760e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2761111c225aSTejun Heo * @__rescuer: self 2762e22bee78STejun Heo * 2763e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2764493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2765e22bee78STejun Heo * 2766706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2767e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2768e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2769e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2770e22bee78STejun Heo * the problem rescuer solves. 2771e22bee78STejun Heo * 2772706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2773706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2774e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2775e22bee78STejun Heo * 2776e22bee78STejun Heo * This should happen rarely. 2777d185af30SYacine Belkadi * 2778d185af30SYacine Belkadi * Return: 0 2779e22bee78STejun Heo */ 2780111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2781e22bee78STejun Heo { 2782111c225aSTejun Heo struct worker *rescuer = __rescuer; 2783111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2784e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 27854d595b86SLai Jiangshan bool should_stop; 2786e22bee78STejun Heo 2787e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2788111c225aSTejun Heo 2789111c225aSTejun Heo /* 2790111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2791111c225aSTejun Heo * doesn't participate in concurrency management. 2792111c225aSTejun Heo */ 2793197f6accSTejun Heo set_pf_worker(true); 2794e22bee78STejun Heo repeat: 2795c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 27961da177e4SLinus Torvalds 27974d595b86SLai Jiangshan /* 27984d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 27994d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 28004d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 28014d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 28024d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 28034d595b86SLai Jiangshan * list is always empty on exit. 28044d595b86SLai Jiangshan */ 28054d595b86SLai Jiangshan should_stop = kthread_should_stop(); 28061da177e4SLinus Torvalds 2807493a1724STejun Heo /* see whether any pwq is asking for help */ 2808a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2809493a1724STejun Heo 2810493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2811493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2812493a1724STejun Heo struct pool_workqueue, mayday_node); 2813112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2814e22bee78STejun Heo struct work_struct *work, *n; 2815e22bee78STejun Heo 2816e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2817493a1724STejun Heo list_del_init(&pwq->mayday_node); 2818493a1724STejun Heo 2819a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2820e22bee78STejun Heo 282151697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 282251697d39SLai Jiangshan 2823a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2824e22bee78STejun Heo 2825e22bee78STejun Heo /* 2826e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2827e22bee78STejun Heo * process'em. 2828e22bee78STejun Heo */ 28290479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 283082607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 283182607adcSTejun Heo if (get_work_pwq(work) == pwq) { 2832e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2833725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 283482607adcSTejun Heo } 283582607adcSTejun Heo } 2836e22bee78STejun Heo 2837008847f6SNeilBrown if (!list_empty(scheduled)) { 2838e22bee78STejun Heo process_scheduled_works(rescuer); 28397576958aSTejun Heo 28407576958aSTejun Heo /* 2841008847f6SNeilBrown * The above execution of rescued work items could 2842008847f6SNeilBrown * have created more to rescue through 2843f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2844008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2845008847f6SNeilBrown * that such back-to-back work items, which may be 2846008847f6SNeilBrown * being used to relieve memory pressure, don't 2847008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2848008847f6SNeilBrown */ 28494f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2850a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2851e66b39afSTejun Heo /* 2852e66b39afSTejun Heo * Queue iff we aren't racing destruction 2853e66b39afSTejun Heo * and somebody else hasn't queued it already. 2854e66b39afSTejun Heo */ 2855e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2856008847f6SNeilBrown get_pwq(pwq); 2857e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2858e66b39afSTejun Heo } 2859a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2860008847f6SNeilBrown } 2861008847f6SNeilBrown } 2862008847f6SNeilBrown 2863008847f6SNeilBrown /* 286477668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 286513b1d625SLai Jiangshan * go away while we're still attached to it. 286677668c8bSLai Jiangshan */ 286777668c8bSLai Jiangshan put_pwq(pwq); 286877668c8bSLai Jiangshan 286977668c8bSLai Jiangshan /* 2870d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 28717576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 28727576958aSTejun Heo * and stalling the execution. 28737576958aSTejun Heo */ 2874d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 287563d95a91STejun Heo wake_up_worker(pool); 28767576958aSTejun Heo 2877a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 287813b1d625SLai Jiangshan 2879a2d812a2STejun Heo worker_detach_from_pool(rescuer); 288013b1d625SLai Jiangshan 2881a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 28821da177e4SLinus Torvalds } 28831da177e4SLinus Torvalds 2884a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2885493a1724STejun Heo 28864d595b86SLai Jiangshan if (should_stop) { 28874d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2888197f6accSTejun Heo set_pf_worker(false); 28894d595b86SLai Jiangshan return 0; 28904d595b86SLai Jiangshan } 28914d595b86SLai Jiangshan 2892111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2893111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2894e22bee78STejun Heo schedule(); 2895e22bee78STejun Heo goto repeat; 28961da177e4SLinus Torvalds } 28971da177e4SLinus Torvalds 2898fca839c0STejun Heo /** 2899fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2900fca839c0STejun Heo * @target_wq: workqueue being flushed 2901fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2902fca839c0STejun Heo * 2903fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2904fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2905fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2906fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2907fca839c0STejun Heo * a deadlock. 2908fca839c0STejun Heo */ 2909fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2910fca839c0STejun Heo struct work_struct *target_work) 2911fca839c0STejun Heo { 2912fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2913fca839c0STejun Heo struct worker *worker; 2914fca839c0STejun Heo 2915fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2916fca839c0STejun Heo return; 2917fca839c0STejun Heo 2918fca839c0STejun Heo worker = current_wq_worker(); 2919fca839c0STejun Heo 2920fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2921d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2922fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 292323d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 292423d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2925d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2926fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2927fca839c0STejun Heo target_wq->name, target_func); 2928fca839c0STejun Heo } 2929fca839c0STejun Heo 2930fc2e4d70SOleg Nesterov struct wq_barrier { 2931fc2e4d70SOleg Nesterov struct work_struct work; 2932fc2e4d70SOleg Nesterov struct completion done; 29332607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2934fc2e4d70SOleg Nesterov }; 2935fc2e4d70SOleg Nesterov 2936fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2937fc2e4d70SOleg Nesterov { 2938fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2939fc2e4d70SOleg Nesterov complete(&barr->done); 2940fc2e4d70SOleg Nesterov } 2941fc2e4d70SOleg Nesterov 29424690c4abSTejun Heo /** 29434690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2944112202d9STejun Heo * @pwq: pwq to insert barrier into 29454690c4abSTejun Heo * @barr: wq_barrier to insert 2946affee4b2STejun Heo * @target: target work to attach @barr to 2947affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29484690c4abSTejun Heo * 2949affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2950affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2951affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2952affee4b2STejun Heo * cpu. 2953affee4b2STejun Heo * 2954affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2955affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2956affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2957affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2958affee4b2STejun Heo * after a work with LINKED flag set. 2959affee4b2STejun Heo * 2960affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2961112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 29624690c4abSTejun Heo * 29634690c4abSTejun Heo * CONTEXT: 2964a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 29654690c4abSTejun Heo */ 2966112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2967affee4b2STejun Heo struct wq_barrier *barr, 2968affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2969fc2e4d70SOleg Nesterov { 2970d812796eSLai Jiangshan unsigned int work_flags = 0; 2971d812796eSLai Jiangshan unsigned int work_color; 2972affee4b2STejun Heo struct list_head *head; 2973affee4b2STejun Heo 2974dc186ad7SThomas Gleixner /* 2975d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2976dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2977dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2978dc186ad7SThomas Gleixner * might deadlock. 2979dc186ad7SThomas Gleixner */ 2980ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 298122df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 298252fa5bc5SBoqun Feng 2983fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2984fd1a5b04SByungchul Park 29852607d7a6STejun Heo barr->task = current; 298683c22520SOleg Nesterov 2987018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2988018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2989018f3a13SLai Jiangshan 2990affee4b2STejun Heo /* 2991affee4b2STejun Heo * If @target is currently being executed, schedule the 2992affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2993affee4b2STejun Heo */ 2994d812796eSLai Jiangshan if (worker) { 2995affee4b2STejun Heo head = worker->scheduled.next; 2996d812796eSLai Jiangshan work_color = worker->current_color; 2997d812796eSLai Jiangshan } else { 2998affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2999affee4b2STejun Heo 3000affee4b2STejun Heo head = target->entry.next; 3001affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 3002d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 3003d812796eSLai Jiangshan work_color = get_work_color(*bits); 3004affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3005affee4b2STejun Heo } 3006affee4b2STejun Heo 3007d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 3008d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 3009d812796eSLai Jiangshan 3010d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 3011fc2e4d70SOleg Nesterov } 3012fc2e4d70SOleg Nesterov 301373f53c4aSTejun Heo /** 3014112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 301573f53c4aSTejun Heo * @wq: workqueue being flushed 301673f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 301773f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 301873f53c4aSTejun Heo * 3019112202d9STejun Heo * Prepare pwqs for workqueue flushing. 302073f53c4aSTejun Heo * 3021112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 3022112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 3023112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3024112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 3025112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 302673f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 302773f53c4aSTejun Heo * 302873f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 302973f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 303073f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 303173f53c4aSTejun Heo * is returned. 303273f53c4aSTejun Heo * 3033112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 303473f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 303573f53c4aSTejun Heo * advanced to @work_color. 303673f53c4aSTejun Heo * 303773f53c4aSTejun Heo * CONTEXT: 30383c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 303973f53c4aSTejun Heo * 3040d185af30SYacine Belkadi * Return: 304173f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 304273f53c4aSTejun Heo * otherwise. 304373f53c4aSTejun Heo */ 3044112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 304573f53c4aSTejun Heo int flush_color, int work_color) 30461da177e4SLinus Torvalds { 304773f53c4aSTejun Heo bool wait = false; 304849e3cf44STejun Heo struct pool_workqueue *pwq; 30491da177e4SLinus Torvalds 305073f53c4aSTejun Heo if (flush_color >= 0) { 30516183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3052112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3053dc186ad7SThomas Gleixner } 305414441960SOleg Nesterov 305549e3cf44STejun Heo for_each_pwq(pwq, wq) { 3056112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30571da177e4SLinus Torvalds 3058a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 305973f53c4aSTejun Heo 306073f53c4aSTejun Heo if (flush_color >= 0) { 30616183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 306273f53c4aSTejun Heo 3063112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3064112202d9STejun Heo pwq->flush_color = flush_color; 3065112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 306673f53c4aSTejun Heo wait = true; 30671da177e4SLinus Torvalds } 306873f53c4aSTejun Heo } 306973f53c4aSTejun Heo 307073f53c4aSTejun Heo if (work_color >= 0) { 30716183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3072112202d9STejun Heo pwq->work_color = work_color; 307373f53c4aSTejun Heo } 307473f53c4aSTejun Heo 3075a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 30761da177e4SLinus Torvalds } 30771da177e4SLinus Torvalds 3078112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 307973f53c4aSTejun Heo complete(&wq->first_flusher->done); 308073f53c4aSTejun Heo 308173f53c4aSTejun Heo return wait; 308283c22520SOleg Nesterov } 30831da177e4SLinus Torvalds 30840fcb78c2SRolf Eike Beer /** 3085c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 30860fcb78c2SRolf Eike Beer * @wq: workqueue to flush 30871da177e4SLinus Torvalds * 3088c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3089c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 30901da177e4SLinus Torvalds */ 3091c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 30921da177e4SLinus Torvalds { 309373f53c4aSTejun Heo struct wq_flusher this_flusher = { 309473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 309573f53c4aSTejun Heo .flush_color = -1, 3096fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 309773f53c4aSTejun Heo }; 309873f53c4aSTejun Heo int next_color; 3099b1f4ec17SOleg Nesterov 31003347fa09STejun Heo if (WARN_ON(!wq_online)) 31013347fa09STejun Heo return; 31023347fa09STejun Heo 310387915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 310487915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 310587915adcSJohannes Berg 31063c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 310773f53c4aSTejun Heo 310873f53c4aSTejun Heo /* 310973f53c4aSTejun Heo * Start-to-wait phase 311073f53c4aSTejun Heo */ 311173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 311273f53c4aSTejun Heo 311373f53c4aSTejun Heo if (next_color != wq->flush_color) { 311473f53c4aSTejun Heo /* 311573f53c4aSTejun Heo * Color space is not full. The current work_color 311673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 311773f53c4aSTejun Heo * by one. 311873f53c4aSTejun Heo */ 31196183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 312073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 312173f53c4aSTejun Heo wq->work_color = next_color; 312273f53c4aSTejun Heo 312373f53c4aSTejun Heo if (!wq->first_flusher) { 312473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 31256183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 312673f53c4aSTejun Heo 312773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 312873f53c4aSTejun Heo 3129112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 313073f53c4aSTejun Heo wq->work_color)) { 313173f53c4aSTejun Heo /* nothing to flush, done */ 313273f53c4aSTejun Heo wq->flush_color = next_color; 313373f53c4aSTejun Heo wq->first_flusher = NULL; 313473f53c4aSTejun Heo goto out_unlock; 313573f53c4aSTejun Heo } 313673f53c4aSTejun Heo } else { 313773f53c4aSTejun Heo /* wait in queue */ 31386183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 313973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3140112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 314173f53c4aSTejun Heo } 314273f53c4aSTejun Heo } else { 314373f53c4aSTejun Heo /* 314473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 314573f53c4aSTejun Heo * The next flush completion will assign us 314673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 314773f53c4aSTejun Heo */ 314873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 314973f53c4aSTejun Heo } 315073f53c4aSTejun Heo 3151fca839c0STejun Heo check_flush_dependency(wq, NULL); 3152fca839c0STejun Heo 31533c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 315473f53c4aSTejun Heo 315573f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 315673f53c4aSTejun Heo 315773f53c4aSTejun Heo /* 315873f53c4aSTejun Heo * Wake-up-and-cascade phase 315973f53c4aSTejun Heo * 316073f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 316173f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 316273f53c4aSTejun Heo */ 316300d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 316473f53c4aSTejun Heo return; 316573f53c4aSTejun Heo 31663c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 316773f53c4aSTejun Heo 31684ce48b37STejun Heo /* we might have raced, check again with mutex held */ 31694ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 31704ce48b37STejun Heo goto out_unlock; 31714ce48b37STejun Heo 317200d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 317373f53c4aSTejun Heo 31746183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 31756183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 317673f53c4aSTejun Heo 317773f53c4aSTejun Heo while (true) { 317873f53c4aSTejun Heo struct wq_flusher *next, *tmp; 317973f53c4aSTejun Heo 318073f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 318173f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 318273f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 318373f53c4aSTejun Heo break; 318473f53c4aSTejun Heo list_del_init(&next->list); 318573f53c4aSTejun Heo complete(&next->done); 318673f53c4aSTejun Heo } 318773f53c4aSTejun Heo 31886183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 318973f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 319073f53c4aSTejun Heo 319173f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 319273f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 319373f53c4aSTejun Heo 319473f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 319573f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 319673f53c4aSTejun Heo /* 319773f53c4aSTejun Heo * Assign the same color to all overflowed 319873f53c4aSTejun Heo * flushers, advance work_color and append to 319973f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 320073f53c4aSTejun Heo * phase for these overflowed flushers. 320173f53c4aSTejun Heo */ 320273f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 320373f53c4aSTejun Heo tmp->flush_color = wq->work_color; 320473f53c4aSTejun Heo 320573f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 320673f53c4aSTejun Heo 320773f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 320873f53c4aSTejun Heo &wq->flusher_queue); 3209112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 321073f53c4aSTejun Heo } 321173f53c4aSTejun Heo 321273f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 32136183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 321473f53c4aSTejun Heo break; 321573f53c4aSTejun Heo } 321673f53c4aSTejun Heo 321773f53c4aSTejun Heo /* 321873f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3219112202d9STejun Heo * the new first flusher and arm pwqs. 322073f53c4aSTejun Heo */ 32216183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 32226183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 322373f53c4aSTejun Heo 322473f53c4aSTejun Heo list_del_init(&next->list); 322573f53c4aSTejun Heo wq->first_flusher = next; 322673f53c4aSTejun Heo 3227112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 322873f53c4aSTejun Heo break; 322973f53c4aSTejun Heo 323073f53c4aSTejun Heo /* 323173f53c4aSTejun Heo * Meh... this color is already done, clear first 323273f53c4aSTejun Heo * flusher and repeat cascading. 323373f53c4aSTejun Heo */ 323473f53c4aSTejun Heo wq->first_flusher = NULL; 323573f53c4aSTejun Heo } 323673f53c4aSTejun Heo 323773f53c4aSTejun Heo out_unlock: 32383c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32391da177e4SLinus Torvalds } 3240c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32411da177e4SLinus Torvalds 32429c5a2ba7STejun Heo /** 32439c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32449c5a2ba7STejun Heo * @wq: workqueue to drain 32459c5a2ba7STejun Heo * 32469c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32479c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32489c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3249b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32509c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32519c5a2ba7STejun Heo * takes too long. 32529c5a2ba7STejun Heo */ 32539c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32549c5a2ba7STejun Heo { 32559c5a2ba7STejun Heo unsigned int flush_cnt = 0; 325649e3cf44STejun Heo struct pool_workqueue *pwq; 32579c5a2ba7STejun Heo 32589c5a2ba7STejun Heo /* 32599c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 32609c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3261618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 32629c5a2ba7STejun Heo */ 326387fc741eSLai Jiangshan mutex_lock(&wq->mutex); 32649c5a2ba7STejun Heo if (!wq->nr_drainers++) 3265618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 326687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32679c5a2ba7STejun Heo reflush: 3268c4f135d6STetsuo Handa __flush_workqueue(wq); 32699c5a2ba7STejun Heo 3270b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 327176af4d93STejun Heo 327249e3cf44STejun Heo for_each_pwq(pwq, wq) { 3273fa2563e4SThomas Tuttle bool drained; 32749c5a2ba7STejun Heo 3275a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3276f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3277a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3278fa2563e4SThomas Tuttle 3279fa2563e4SThomas Tuttle if (drained) 32809c5a2ba7STejun Heo continue; 32819c5a2ba7STejun Heo 32829c5a2ba7STejun Heo if (++flush_cnt == 10 || 32839c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3284e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3285e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 328676af4d93STejun Heo 3287b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 32889c5a2ba7STejun Heo goto reflush; 32899c5a2ba7STejun Heo } 32909c5a2ba7STejun Heo 32919c5a2ba7STejun Heo if (!--wq->nr_drainers) 3292618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 329387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32949c5a2ba7STejun Heo } 32959c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 32969c5a2ba7STejun Heo 3297d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3298d6e89786SJohannes Berg bool from_cancel) 3299baf59022STejun Heo { 3300baf59022STejun Heo struct worker *worker = NULL; 3301c9e7cf27STejun Heo struct worker_pool *pool; 3302112202d9STejun Heo struct pool_workqueue *pwq; 3303baf59022STejun Heo 3304baf59022STejun Heo might_sleep(); 3305baf59022STejun Heo 330624acfb71SThomas Gleixner rcu_read_lock(); 3307fa1b54e6STejun Heo pool = get_work_pool(work); 3308fa1b54e6STejun Heo if (!pool) { 330924acfb71SThomas Gleixner rcu_read_unlock(); 3310fa1b54e6STejun Heo return false; 3311fa1b54e6STejun Heo } 3312fa1b54e6STejun Heo 3313a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 33140b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3315112202d9STejun Heo pwq = get_work_pwq(work); 3316112202d9STejun Heo if (pwq) { 3317112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3318baf59022STejun Heo goto already_gone; 3319606a5020STejun Heo } else { 3320c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3321baf59022STejun Heo if (!worker) 3322baf59022STejun Heo goto already_gone; 3323112202d9STejun Heo pwq = worker->current_pwq; 3324606a5020STejun Heo } 3325baf59022STejun Heo 3326fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3327fca839c0STejun Heo 3328112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3329a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3330baf59022STejun Heo 3331e159489bSTejun Heo /* 3332a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3333a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3334a1d14934SPeter Zijlstra * 3335a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3336a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3337a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3338a1d14934SPeter Zijlstra * forward progress. 3339e159489bSTejun Heo */ 3340d6e89786SJohannes Berg if (!from_cancel && 3341d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3342112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3343112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3344a1d14934SPeter Zijlstra } 334524acfb71SThomas Gleixner rcu_read_unlock(); 3346baf59022STejun Heo return true; 3347baf59022STejun Heo already_gone: 3348a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 334924acfb71SThomas Gleixner rcu_read_unlock(); 3350baf59022STejun Heo return false; 3351baf59022STejun Heo } 3352baf59022STejun Heo 3353d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3354d6e89786SJohannes Berg { 3355d6e89786SJohannes Berg struct wq_barrier barr; 3356d6e89786SJohannes Berg 3357d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3358d6e89786SJohannes Berg return false; 3359d6e89786SJohannes Berg 33604d43d395STetsuo Handa if (WARN_ON(!work->func)) 33614d43d395STetsuo Handa return false; 33624d43d395STetsuo Handa 336387915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 336487915adcSJohannes Berg lock_map_release(&work->lockdep_map); 336587915adcSJohannes Berg 3366d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3367d6e89786SJohannes Berg wait_for_completion(&barr.done); 3368d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3369d6e89786SJohannes Berg return true; 3370d6e89786SJohannes Berg } else { 3371d6e89786SJohannes Berg return false; 3372d6e89786SJohannes Berg } 3373d6e89786SJohannes Berg } 3374d6e89786SJohannes Berg 3375db700897SOleg Nesterov /** 3376401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3377401a8d04STejun Heo * @work: the work to flush 3378db700897SOleg Nesterov * 3379606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3380606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3381401a8d04STejun Heo * 3382d185af30SYacine Belkadi * Return: 3383401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3384401a8d04STejun Heo * %false if it was already idle. 3385db700897SOleg Nesterov */ 3386401a8d04STejun Heo bool flush_work(struct work_struct *work) 3387db700897SOleg Nesterov { 3388d6e89786SJohannes Berg return __flush_work(work, false); 3389606a5020STejun Heo } 3390db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3391db700897SOleg Nesterov 33928603e1b3STejun Heo struct cwt_wait { 3393ac6424b9SIngo Molnar wait_queue_entry_t wait; 33948603e1b3STejun Heo struct work_struct *work; 33958603e1b3STejun Heo }; 33968603e1b3STejun Heo 3397ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 33988603e1b3STejun Heo { 33998603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 34008603e1b3STejun Heo 34018603e1b3STejun Heo if (cwait->work != key) 34028603e1b3STejun Heo return 0; 34038603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 34048603e1b3STejun Heo } 34058603e1b3STejun Heo 340636e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3407401a8d04STejun Heo { 34088603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3409bbb68dfaSTejun Heo unsigned long flags; 34101f1f642eSOleg Nesterov int ret; 34111f1f642eSOleg Nesterov 34121f1f642eSOleg Nesterov do { 3413bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3414bbb68dfaSTejun Heo /* 34158603e1b3STejun Heo * If someone else is already canceling, wait for it to 34168603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 34178603e1b3STejun Heo * because we may get scheduled between @work's completion 34188603e1b3STejun Heo * and the other canceling task resuming and clearing 34198603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 34208603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 34218603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 34228603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 34238603e1b3STejun Heo * we're hogging the CPU. 34248603e1b3STejun Heo * 34258603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 34268603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 34278603e1b3STejun Heo * wake function which matches @work along with exclusive 34288603e1b3STejun Heo * wait and wakeup. 3429bbb68dfaSTejun Heo */ 34308603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 34318603e1b3STejun Heo struct cwt_wait cwait; 34328603e1b3STejun Heo 34338603e1b3STejun Heo init_wait(&cwait.wait); 34348603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34358603e1b3STejun Heo cwait.work = work; 34368603e1b3STejun Heo 34378603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34388603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34398603e1b3STejun Heo if (work_is_canceling(work)) 34408603e1b3STejun Heo schedule(); 34418603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34428603e1b3STejun Heo } 34431f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34441f1f642eSOleg Nesterov 3445bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3446bbb68dfaSTejun Heo mark_work_canceling(work); 3447bbb68dfaSTejun Heo local_irq_restore(flags); 3448bbb68dfaSTejun Heo 34493347fa09STejun Heo /* 34503347fa09STejun Heo * This allows canceling during early boot. We know that @work 34513347fa09STejun Heo * isn't executing. 34523347fa09STejun Heo */ 34533347fa09STejun Heo if (wq_online) 3454d6e89786SJohannes Berg __flush_work(work, true); 34553347fa09STejun Heo 34567a22ad75STejun Heo clear_work_data(work); 34578603e1b3STejun Heo 34588603e1b3STejun Heo /* 34598603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 34608603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 34618603e1b3STejun Heo * visible there. 34628603e1b3STejun Heo */ 34638603e1b3STejun Heo smp_mb(); 34648603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 34658603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 34668603e1b3STejun Heo 34671f1f642eSOleg Nesterov return ret; 34681f1f642eSOleg Nesterov } 34691f1f642eSOleg Nesterov 34706e84d644SOleg Nesterov /** 3471401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3472401a8d04STejun Heo * @work: the work to cancel 34736e84d644SOleg Nesterov * 3474401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3475401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3476401a8d04STejun Heo * another workqueue. On return from this function, @work is 3477401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 34781f1f642eSOleg Nesterov * 3479401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3480401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 34816e84d644SOleg Nesterov * 3482401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 34836e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3484401a8d04STejun Heo * 3485d185af30SYacine Belkadi * Return: 3486401a8d04STejun Heo * %true if @work was pending, %false otherwise. 34876e84d644SOleg Nesterov */ 3488401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 34896e84d644SOleg Nesterov { 349036e227d2STejun Heo return __cancel_work_timer(work, false); 3491b89deed3SOleg Nesterov } 349228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3493b89deed3SOleg Nesterov 34946e84d644SOleg Nesterov /** 3495401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3496401a8d04STejun Heo * @dwork: the delayed work to flush 34976e84d644SOleg Nesterov * 3498401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3499401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3500401a8d04STejun Heo * considers the last queueing instance of @dwork. 35011f1f642eSOleg Nesterov * 3502d185af30SYacine Belkadi * Return: 3503401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3504401a8d04STejun Heo * %false if it was already idle. 35056e84d644SOleg Nesterov */ 3506401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3507401a8d04STejun Heo { 35088930cabaSTejun Heo local_irq_disable(); 3509401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 351060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 35118930cabaSTejun Heo local_irq_enable(); 3512401a8d04STejun Heo return flush_work(&dwork->work); 3513401a8d04STejun Heo } 3514401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3515401a8d04STejun Heo 351605f0fe6bSTejun Heo /** 351705f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 351805f0fe6bSTejun Heo * @rwork: the rcu work to flush 351905f0fe6bSTejun Heo * 352005f0fe6bSTejun Heo * Return: 352105f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 352205f0fe6bSTejun Heo * %false if it was already idle. 352305f0fe6bSTejun Heo */ 352405f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 352505f0fe6bSTejun Heo { 352605f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 352705f0fe6bSTejun Heo rcu_barrier(); 352805f0fe6bSTejun Heo flush_work(&rwork->work); 352905f0fe6bSTejun Heo return true; 353005f0fe6bSTejun Heo } else { 353105f0fe6bSTejun Heo return flush_work(&rwork->work); 353205f0fe6bSTejun Heo } 353305f0fe6bSTejun Heo } 353405f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 353505f0fe6bSTejun Heo 3536f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3537f72b8792SJens Axboe { 3538f72b8792SJens Axboe unsigned long flags; 3539f72b8792SJens Axboe int ret; 3540f72b8792SJens Axboe 3541f72b8792SJens Axboe do { 3542f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3543f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3544f72b8792SJens Axboe 3545f72b8792SJens Axboe if (unlikely(ret < 0)) 3546f72b8792SJens Axboe return false; 3547f72b8792SJens Axboe 3548f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3549f72b8792SJens Axboe local_irq_restore(flags); 3550f72b8792SJens Axboe return ret; 3551f72b8792SJens Axboe } 3552f72b8792SJens Axboe 355373b4b532SAndrey Grodzovsky /* 355473b4b532SAndrey Grodzovsky * See cancel_delayed_work() 355573b4b532SAndrey Grodzovsky */ 355673b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 355773b4b532SAndrey Grodzovsky { 355873b4b532SAndrey Grodzovsky return __cancel_work(work, false); 355973b4b532SAndrey Grodzovsky } 356073b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 356173b4b532SAndrey Grodzovsky 3562401a8d04STejun Heo /** 356357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 356457b30ae7STejun Heo * @dwork: delayed_work to cancel 356509383498STejun Heo * 3566d185af30SYacine Belkadi * Kill off a pending delayed_work. 3567d185af30SYacine Belkadi * 3568d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3569d185af30SYacine Belkadi * pending. 3570d185af30SYacine Belkadi * 3571d185af30SYacine Belkadi * Note: 3572d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3573d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3574d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 357509383498STejun Heo * 357657b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 357709383498STejun Heo */ 357857b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 357909383498STejun Heo { 3580f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 358109383498STejun Heo } 358257b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 358309383498STejun Heo 358409383498STejun Heo /** 3585401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3586401a8d04STejun Heo * @dwork: the delayed work cancel 3587401a8d04STejun Heo * 3588401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3589401a8d04STejun Heo * 3590d185af30SYacine Belkadi * Return: 3591401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3592401a8d04STejun Heo */ 3593401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 35946e84d644SOleg Nesterov { 359536e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 35966e84d644SOleg Nesterov } 3597f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 35981da177e4SLinus Torvalds 35990fcb78c2SRolf Eike Beer /** 360031ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3601b6136773SAndrew Morton * @func: the function to call 3602b6136773SAndrew Morton * 360331ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 360431ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3605b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 360631ddd871STejun Heo * 3607d185af30SYacine Belkadi * Return: 360831ddd871STejun Heo * 0 on success, -errno on failure. 3609b6136773SAndrew Morton */ 361065f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 361115316ba8SChristoph Lameter { 361215316ba8SChristoph Lameter int cpu; 361338f51568SNamhyung Kim struct work_struct __percpu *works; 361415316ba8SChristoph Lameter 3615b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3616b6136773SAndrew Morton if (!works) 361715316ba8SChristoph Lameter return -ENOMEM; 3618b6136773SAndrew Morton 3619ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 362093981800STejun Heo 362115316ba8SChristoph Lameter for_each_online_cpu(cpu) { 36229bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 36239bfb1839SIngo Molnar 36249bfb1839SIngo Molnar INIT_WORK(work, func); 36258de6d308SOleg Nesterov schedule_work_on(cpu, work); 362615316ba8SChristoph Lameter } 362793981800STejun Heo 362893981800STejun Heo for_each_online_cpu(cpu) 36298616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 363093981800STejun Heo 3631ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3632b6136773SAndrew Morton free_percpu(works); 363315316ba8SChristoph Lameter return 0; 363415316ba8SChristoph Lameter } 363515316ba8SChristoph Lameter 3636eef6a7d5SAlan Stern /** 36371fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36381fa44ecaSJames Bottomley * @fn: the function to execute 36391fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36401fa44ecaSJames Bottomley * be available when the work executes) 36411fa44ecaSJames Bottomley * 36421fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36431fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36441fa44ecaSJames Bottomley * 3645d185af30SYacine Belkadi * Return: 0 - function was executed 36461fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36471fa44ecaSJames Bottomley */ 364865f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36491fa44ecaSJames Bottomley { 36501fa44ecaSJames Bottomley if (!in_interrupt()) { 365165f27f38SDavid Howells fn(&ew->work); 36521fa44ecaSJames Bottomley return 0; 36531fa44ecaSJames Bottomley } 36541fa44ecaSJames Bottomley 365565f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36561fa44ecaSJames Bottomley schedule_work(&ew->work); 36571fa44ecaSJames Bottomley 36581fa44ecaSJames Bottomley return 1; 36591fa44ecaSJames Bottomley } 36601fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 36611fa44ecaSJames Bottomley 36627a4e344cSTejun Heo /** 36637a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 36647a4e344cSTejun Heo * @attrs: workqueue_attrs to free 36657a4e344cSTejun Heo * 36667a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 36677a4e344cSTejun Heo */ 3668513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 36697a4e344cSTejun Heo { 36707a4e344cSTejun Heo if (attrs) { 36717a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 36727a4e344cSTejun Heo kfree(attrs); 36737a4e344cSTejun Heo } 36747a4e344cSTejun Heo } 36757a4e344cSTejun Heo 36767a4e344cSTejun Heo /** 36777a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 36787a4e344cSTejun Heo * 36797a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3680d185af30SYacine Belkadi * return it. 3681d185af30SYacine Belkadi * 3682d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 36837a4e344cSTejun Heo */ 3684513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 36857a4e344cSTejun Heo { 36867a4e344cSTejun Heo struct workqueue_attrs *attrs; 36877a4e344cSTejun Heo 3688be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 36897a4e344cSTejun Heo if (!attrs) 36907a4e344cSTejun Heo goto fail; 3691be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 36927a4e344cSTejun Heo goto fail; 36937a4e344cSTejun Heo 369413e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 36957a4e344cSTejun Heo return attrs; 36967a4e344cSTejun Heo fail: 36977a4e344cSTejun Heo free_workqueue_attrs(attrs); 36987a4e344cSTejun Heo return NULL; 36997a4e344cSTejun Heo } 37007a4e344cSTejun Heo 370129c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 370229c91e99STejun Heo const struct workqueue_attrs *from) 370329c91e99STejun Heo { 370429c91e99STejun Heo to->nice = from->nice; 370529c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 37062865a8fbSShaohua Li /* 37072865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 37082865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 37092865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 37102865a8fbSShaohua Li */ 37112865a8fbSShaohua Li to->no_numa = from->no_numa; 371229c91e99STejun Heo } 371329c91e99STejun Heo 371429c91e99STejun Heo /* hash value of the content of @attr */ 371529c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 371629c91e99STejun Heo { 371729c91e99STejun Heo u32 hash = 0; 371829c91e99STejun Heo 371929c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 372013e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 372113e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 372229c91e99STejun Heo return hash; 372329c91e99STejun Heo } 372429c91e99STejun Heo 372529c91e99STejun Heo /* content equality test */ 372629c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 372729c91e99STejun Heo const struct workqueue_attrs *b) 372829c91e99STejun Heo { 372929c91e99STejun Heo if (a->nice != b->nice) 373029c91e99STejun Heo return false; 373129c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 373229c91e99STejun Heo return false; 373329c91e99STejun Heo return true; 373429c91e99STejun Heo } 373529c91e99STejun Heo 37367a4e344cSTejun Heo /** 37377a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 37387a4e344cSTejun Heo * @pool: worker_pool to initialize 37397a4e344cSTejun Heo * 3740402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3741d185af30SYacine Belkadi * 3742d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 374329c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 374429c91e99STejun Heo * on @pool safely to release it. 37457a4e344cSTejun Heo */ 37467a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 37474e1a1f9aSTejun Heo { 3748a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 374929c91e99STejun Heo pool->id = -1; 375029c91e99STejun Heo pool->cpu = -1; 3751f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 37524e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 375382607adcSTejun Heo pool->watchdog_ts = jiffies; 37544e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37554e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 37564e1a1f9aSTejun Heo hash_init(pool->busy_hash); 37574e1a1f9aSTejun Heo 375832a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 37593f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 37604e1a1f9aSTejun Heo 376132a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 37624e1a1f9aSTejun Heo 3763da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3764e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 37657a4e344cSTejun Heo 37667cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 376729c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 376829c91e99STejun Heo pool->refcnt = 1; 376929c91e99STejun Heo 377029c91e99STejun Heo /* shouldn't fail above this point */ 3771be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 37727a4e344cSTejun Heo if (!pool->attrs) 37737a4e344cSTejun Heo return -ENOMEM; 37747a4e344cSTejun Heo return 0; 37754e1a1f9aSTejun Heo } 37764e1a1f9aSTejun Heo 3777669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3778669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3779669de8bdSBart Van Assche { 3780669de8bdSBart Van Assche char *lock_name; 3781669de8bdSBart Van Assche 3782669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3783669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3784669de8bdSBart Van Assche if (!lock_name) 3785669de8bdSBart Van Assche lock_name = wq->name; 378669a106c0SQian Cai 378769a106c0SQian Cai wq->lock_name = lock_name; 3788669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3789669de8bdSBart Van Assche } 3790669de8bdSBart Van Assche 3791669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3792669de8bdSBart Van Assche { 3793669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3794669de8bdSBart Van Assche } 3795669de8bdSBart Van Assche 3796669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3797669de8bdSBart Van Assche { 3798669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3799669de8bdSBart Van Assche kfree(wq->lock_name); 3800669de8bdSBart Van Assche } 3801669de8bdSBart Van Assche #else 3802669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3803669de8bdSBart Van Assche { 3804669de8bdSBart Van Assche } 3805669de8bdSBart Van Assche 3806669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3807669de8bdSBart Van Assche { 3808669de8bdSBart Van Assche } 3809669de8bdSBart Van Assche 3810669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3811669de8bdSBart Van Assche { 3812669de8bdSBart Van Assche } 3813669de8bdSBart Van Assche #endif 3814669de8bdSBart Van Assche 3815e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3816e2dca7adSTejun Heo { 3817e2dca7adSTejun Heo struct workqueue_struct *wq = 3818e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3819e2dca7adSTejun Heo 3820669de8bdSBart Van Assche wq_free_lockdep(wq); 3821669de8bdSBart Van Assche 3822e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3823ee1ceef7STejun Heo free_percpu(wq->cpu_pwq); 3824e2dca7adSTejun Heo else 3825e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3826e2dca7adSTejun Heo 3827e2dca7adSTejun Heo kfree(wq); 3828e2dca7adSTejun Heo } 3829e2dca7adSTejun Heo 383029c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 383129c91e99STejun Heo { 383229c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 383329c91e99STejun Heo 38347cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 383529c91e99STejun Heo free_workqueue_attrs(pool->attrs); 383629c91e99STejun Heo kfree(pool); 383729c91e99STejun Heo } 383829c91e99STejun Heo 383929c91e99STejun Heo /** 384029c91e99STejun Heo * put_unbound_pool - put a worker_pool 384129c91e99STejun Heo * @pool: worker_pool to put 384229c91e99STejun Heo * 384324acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3844c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3845c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3846c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3847a892caccSTejun Heo * 3848a892caccSTejun Heo * Should be called with wq_pool_mutex held. 384929c91e99STejun Heo */ 385029c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 385129c91e99STejun Heo { 385260f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 385329c91e99STejun Heo struct worker *worker; 38549680540cSYang Yingliang LIST_HEAD(cull_list); 3855e02b9312SValentin Schneider 3856a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3857a892caccSTejun Heo 3858a892caccSTejun Heo if (--pool->refcnt) 385929c91e99STejun Heo return; 386029c91e99STejun Heo 386129c91e99STejun Heo /* sanity checks */ 386261d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3863a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 386429c91e99STejun Heo return; 386529c91e99STejun Heo 386629c91e99STejun Heo /* release id and unhash */ 386729c91e99STejun Heo if (pool->id >= 0) 386829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 386929c91e99STejun Heo hash_del(&pool->hash_node); 387029c91e99STejun Heo 3871c5aa87bbSTejun Heo /* 3872692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3873692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3874692b4825STejun Heo * manager and @pool gets freed with the flag set. 38759ab03be4SValentin Schneider * 38769ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 38779ab03be4SValentin Schneider * only get here with 38789ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 38799ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 38809ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 38819ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 38829ab03be4SValentin Schneider * drops pool->lock 3883c5aa87bbSTejun Heo */ 38849ab03be4SValentin Schneider while (true) { 38859ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 38869ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3887d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3888e02b9312SValentin Schneider 3889e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 38909ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 38919ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3892692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 38939ab03be4SValentin Schneider break; 38949ab03be4SValentin Schneider } 38959ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3896e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 38979ab03be4SValentin Schneider } 3898692b4825STejun Heo 38991037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3900e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 390129c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3902a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 390360f5a4bcSLai Jiangshan 3904e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3905e02b9312SValentin Schneider 3906e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 390760f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 39081258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 390960f5a4bcSLai Jiangshan 391060f5a4bcSLai Jiangshan if (pool->detach_completion) 391160f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 391260f5a4bcSLai Jiangshan 391329c91e99STejun Heo /* shut down the timers */ 391429c91e99STejun Heo del_timer_sync(&pool->idle_timer); 39153f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 391629c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 391729c91e99STejun Heo 391824acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 391925b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 392029c91e99STejun Heo } 392129c91e99STejun Heo 392229c91e99STejun Heo /** 392329c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 392429c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 392529c91e99STejun Heo * 392629c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 392729c91e99STejun Heo * reference count and return it. If there already is a matching 392829c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3929d185af30SYacine Belkadi * create a new one. 3930a892caccSTejun Heo * 3931a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3932d185af30SYacine Belkadi * 3933d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3934d185af30SYacine Belkadi * On failure, %NULL. 393529c91e99STejun Heo */ 393629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 393729c91e99STejun Heo { 393829c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 393929c91e99STejun Heo struct worker_pool *pool; 3940f3f90ad4STejun Heo int node; 3941e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 394229c91e99STejun Heo 3943a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 394429c91e99STejun Heo 394529c91e99STejun Heo /* do we already have a matching pool? */ 394629c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 394729c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 394829c91e99STejun Heo pool->refcnt++; 39493fb1823cSLai Jiangshan return pool; 395029c91e99STejun Heo } 395129c91e99STejun Heo } 395229c91e99STejun Heo 3953e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3954e2273584SXunlei Pang if (wq_numa_enabled) { 3955e2273584SXunlei Pang for_each_node(node) { 3956e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3957e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3958e2273584SXunlei Pang target_node = node; 3959e2273584SXunlei Pang break; 3960e2273584SXunlei Pang } 3961e2273584SXunlei Pang } 3962e2273584SXunlei Pang } 3963e2273584SXunlei Pang 396429c91e99STejun Heo /* nope, create a new one */ 3965e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 396629c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 396729c91e99STejun Heo goto fail; 396829c91e99STejun Heo 396929c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3970e2273584SXunlei Pang pool->node = target_node; 397129c91e99STejun Heo 39722865a8fbSShaohua Li /* 39732865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 39742865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 39752865a8fbSShaohua Li */ 39762865a8fbSShaohua Li pool->attrs->no_numa = false; 39772865a8fbSShaohua Li 397829c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 397929c91e99STejun Heo goto fail; 398029c91e99STejun Heo 398129c91e99STejun Heo /* create and start the initial worker */ 39823347fa09STejun Heo if (wq_online && !create_worker(pool)) 398329c91e99STejun Heo goto fail; 398429c91e99STejun Heo 398529c91e99STejun Heo /* install */ 398629c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 39873fb1823cSLai Jiangshan 398829c91e99STejun Heo return pool; 398929c91e99STejun Heo fail: 399029c91e99STejun Heo if (pool) 399129c91e99STejun Heo put_unbound_pool(pool); 399229c91e99STejun Heo return NULL; 399329c91e99STejun Heo } 399429c91e99STejun Heo 39958864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 39968864b4e5STejun Heo { 39978864b4e5STejun Heo kmem_cache_free(pwq_cache, 39988864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 39998864b4e5STejun Heo } 40008864b4e5STejun Heo 40018864b4e5STejun Heo /* 4002967b494eSTejun Heo * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 4003967b494eSTejun Heo * refcnt and needs to be destroyed. 40048864b4e5STejun Heo */ 4005687a9aa5STejun Heo static void pwq_release_workfn(struct kthread_work *work) 40068864b4e5STejun Heo { 40078864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 4008687a9aa5STejun Heo release_work); 40098864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 40108864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 4011b42b0bddSYang Yingliang bool is_last = false; 40128864b4e5STejun Heo 4013b42b0bddSYang Yingliang /* 4014687a9aa5STejun Heo * When @pwq is not linked, it doesn't hold any reference to the 4015b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 4016b42b0bddSYang Yingliang */ 4017b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 40183c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 40198864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 4020bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 40213c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 4022b42b0bddSYang Yingliang } 40238864b4e5STejun Heo 4024687a9aa5STejun Heo if (wq->flags & WQ_UNBOUND) { 4025a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 40268864b4e5STejun Heo put_unbound_pool(pool); 4027a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 4028687a9aa5STejun Heo } 4029a892caccSTejun Heo 403025b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 40318864b4e5STejun Heo 40328864b4e5STejun Heo /* 40338864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 4034e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 40358864b4e5STejun Heo */ 4036669de8bdSBart Van Assche if (is_last) { 4037669de8bdSBart Van Assche wq_unregister_lockdep(wq); 403825b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 40396029a918STejun Heo } 4040669de8bdSBart Van Assche } 40418864b4e5STejun Heo 40420fbd95aaSTejun Heo /** 4043699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 40440fbd95aaSTejun Heo * @pwq: target pool_workqueue 40450fbd95aaSTejun Heo * 4046699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4047f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4048699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 40490fbd95aaSTejun Heo */ 4050699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 40510fbd95aaSTejun Heo { 4052699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4053699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 40543347fa09STejun Heo unsigned long flags; 4055699ce097STejun Heo 4056699ce097STejun Heo /* for @wq->saved_max_active */ 4057a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4058699ce097STejun Heo 4059699ce097STejun Heo /* fast exit for non-freezable wqs */ 4060699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4061699ce097STejun Heo return; 4062699ce097STejun Heo 40633347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4064a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4065699ce097STejun Heo 406674b414eaSLai Jiangshan /* 406774b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 406874b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 406974b414eaSLai Jiangshan * is updated and visible. 407074b414eaSLai Jiangshan */ 407174b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 407201341fbdSYunfeng Ye bool kick = false; 407301341fbdSYunfeng Ye 4074699ce097STejun Heo pwq->max_active = wq->saved_max_active; 40750fbd95aaSTejun Heo 4076f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 407701341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 4078f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 407901341fbdSYunfeng Ye kick = true; 408001341fbdSYunfeng Ye } 4081951a078aSLai Jiangshan 4082951a078aSLai Jiangshan /* 4083951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 408401341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 408501341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 408601341fbdSYunfeng Ye * let's kick iff work items were activated. 4087951a078aSLai Jiangshan */ 408801341fbdSYunfeng Ye if (kick) 4089951a078aSLai Jiangshan wake_up_worker(pwq->pool); 4090699ce097STejun Heo } else { 4091699ce097STejun Heo pwq->max_active = 0; 4092699ce097STejun Heo } 4093699ce097STejun Heo 4094a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 40950fbd95aaSTejun Heo } 40960fbd95aaSTejun Heo 409767dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4098f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4099f147f29eSTejun Heo struct worker_pool *pool) 4100d2c1d404STejun Heo { 4101d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4102d2c1d404STejun Heo 4103e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4104e50aba9aSTejun Heo 4105d2c1d404STejun Heo pwq->pool = pool; 4106d2c1d404STejun Heo pwq->wq = wq; 4107d2c1d404STejun Heo pwq->flush_color = -1; 41088864b4e5STejun Heo pwq->refcnt = 1; 4109f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 41101befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4111d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 4112687a9aa5STejun Heo kthread_init_work(&pwq->release_work, pwq_release_workfn); 4113f147f29eSTejun Heo } 4114d2c1d404STejun Heo 4115f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 41161befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4117f147f29eSTejun Heo { 4118f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4119f147f29eSTejun Heo 4120f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 412175ccf595STejun Heo 41221befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 41231befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 41241befcf30STejun Heo return; 41251befcf30STejun Heo 412629b1cb41SLai Jiangshan /* set the matching work_color */ 412775ccf595STejun Heo pwq->work_color = wq->work_color; 4128983ca25eSTejun Heo 4129983ca25eSTejun Heo /* sync max_active to the current setting */ 4130983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4131983ca25eSTejun Heo 4132983ca25eSTejun Heo /* link in @pwq */ 41339e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4134df2d5ae4STejun Heo } 41356029a918STejun Heo 4136f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4137f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4138f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4139f147f29eSTejun Heo { 4140f147f29eSTejun Heo struct worker_pool *pool; 4141f147f29eSTejun Heo struct pool_workqueue *pwq; 4142f147f29eSTejun Heo 4143f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4144f147f29eSTejun Heo 4145f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4146f147f29eSTejun Heo if (!pool) 4147f147f29eSTejun Heo return NULL; 4148f147f29eSTejun Heo 4149e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4150f147f29eSTejun Heo if (!pwq) { 4151f147f29eSTejun Heo put_unbound_pool(pool); 4152f147f29eSTejun Heo return NULL; 4153f147f29eSTejun Heo } 4154f147f29eSTejun Heo 4155f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4156f147f29eSTejun Heo return pwq; 4157d2c1d404STejun Heo } 4158d2c1d404STejun Heo 41594c16bd32STejun Heo /** 416030186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 4161042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 41624c16bd32STejun Heo * @node: the target NUMA node 41634c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 41644c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 41654c16bd32STejun Heo * 41664c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 41674c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 4168d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 41694c16bd32STejun Heo * 41704c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 41714c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 41724c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 41734c16bd32STejun Heo * @attrs->cpumask. 41744c16bd32STejun Heo * 41754c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 41764c16bd32STejun Heo * stable. 4177d185af30SYacine Belkadi * 4178d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 4179d185af30SYacine Belkadi * %false if equal. 41804c16bd32STejun Heo */ 41814c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 41824c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 41834c16bd32STejun Heo { 4184d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 41854c16bd32STejun Heo goto use_dfl; 41864c16bd32STejun Heo 41874c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 41884c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 41894c16bd32STejun Heo if (cpu_going_down >= 0) 41904c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 41914c16bd32STejun Heo 41924c16bd32STejun Heo if (cpumask_empty(cpumask)) 41934c16bd32STejun Heo goto use_dfl; 41944c16bd32STejun Heo 41954c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 41964c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 41971ad0f0a7SMichael Bringmann 41981ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 41991ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 42001ad0f0a7SMichael Bringmann "possible intersect\n"); 42011ad0f0a7SMichael Bringmann return false; 42021ad0f0a7SMichael Bringmann } 42031ad0f0a7SMichael Bringmann 42044c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 42054c16bd32STejun Heo 42064c16bd32STejun Heo use_dfl: 42074c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 42084c16bd32STejun Heo return false; 42094c16bd32STejun Heo } 42104c16bd32STejun Heo 42111befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 42121befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 42131befcf30STejun Heo int node, 42141befcf30STejun Heo struct pool_workqueue *pwq) 42151befcf30STejun Heo { 42161befcf30STejun Heo struct pool_workqueue *old_pwq; 42171befcf30STejun Heo 42185b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42191befcf30STejun Heo lockdep_assert_held(&wq->mutex); 42201befcf30STejun Heo 42211befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 42221befcf30STejun Heo link_pwq(pwq); 42231befcf30STejun Heo 42241befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42251befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 42261befcf30STejun Heo return old_pwq; 42271befcf30STejun Heo } 42281befcf30STejun Heo 42292d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 42302d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 42312d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 42322d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4233042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 42342d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 42352d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 42362d5f0764SLai Jiangshan }; 42372d5f0764SLai Jiangshan 42382d5f0764SLai Jiangshan /* free the resources after success or abort */ 42392d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 42402d5f0764SLai Jiangshan { 42412d5f0764SLai Jiangshan if (ctx) { 42422d5f0764SLai Jiangshan int node; 42432d5f0764SLai Jiangshan 42442d5f0764SLai Jiangshan for_each_node(node) 42452d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 42462d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 42472d5f0764SLai Jiangshan 42482d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 42492d5f0764SLai Jiangshan 42502d5f0764SLai Jiangshan kfree(ctx); 42512d5f0764SLai Jiangshan } 42522d5f0764SLai Jiangshan } 42532d5f0764SLai Jiangshan 42542d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 42552d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 42562d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 425799c621efSLai Jiangshan const struct workqueue_attrs *attrs, 425899c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 42592d5f0764SLai Jiangshan { 42602d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 42612d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 42622d5f0764SLai Jiangshan int node; 42632d5f0764SLai Jiangshan 42642d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42652d5f0764SLai Jiangshan 4266acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 42672d5f0764SLai Jiangshan 4268be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4269be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 42702d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 42712d5f0764SLai Jiangshan goto out_free; 42722d5f0764SLai Jiangshan 4273042f7df1SLai Jiangshan /* 427499c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 427599c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4276042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4277042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4278042f7df1SLai Jiangshan */ 42792d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 428099c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4281042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 428299c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 42832d5f0764SLai Jiangshan 42842d5f0764SLai Jiangshan /* 42852d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 42862d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 42872d5f0764SLai Jiangshan * pools. 42882d5f0764SLai Jiangshan */ 42892d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 42902d5f0764SLai Jiangshan 42912d5f0764SLai Jiangshan /* 42922d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 42932d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 42942d5f0764SLai Jiangshan * it even if we don't use it immediately. 42952d5f0764SLai Jiangshan */ 42962d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 42972d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 42982d5f0764SLai Jiangshan goto out_free; 42992d5f0764SLai Jiangshan 43002d5f0764SLai Jiangshan for_each_node(node) { 4301042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 43022d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 43032d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 43042d5f0764SLai Jiangshan goto out_free; 43052d5f0764SLai Jiangshan } else { 43062d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 43072d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 43082d5f0764SLai Jiangshan } 43092d5f0764SLai Jiangshan } 43102d5f0764SLai Jiangshan 4311042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4312042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4313042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 43142d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4315042f7df1SLai Jiangshan 43162d5f0764SLai Jiangshan ctx->wq = wq; 43172d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43182d5f0764SLai Jiangshan return ctx; 43192d5f0764SLai Jiangshan 43202d5f0764SLai Jiangshan out_free: 43212d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43222d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 43232d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 43242d5f0764SLai Jiangshan return NULL; 43252d5f0764SLai Jiangshan } 43262d5f0764SLai Jiangshan 43272d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 43282d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 43292d5f0764SLai Jiangshan { 43302d5f0764SLai Jiangshan int node; 43312d5f0764SLai Jiangshan 43322d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 43332d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 43342d5f0764SLai Jiangshan 43352d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 43362d5f0764SLai Jiangshan 43372d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 43382d5f0764SLai Jiangshan for_each_node(node) 43392d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 43402d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 43412d5f0764SLai Jiangshan 43422d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 43432d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 43442d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 43452d5f0764SLai Jiangshan 43462d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 43472d5f0764SLai Jiangshan } 43482d5f0764SLai Jiangshan 4349a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4350a0111cf6SLai Jiangshan { 4351a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4352ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4353a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4354a0111cf6SLai Jiangshan } 4355a0111cf6SLai Jiangshan 4356a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4357a0111cf6SLai Jiangshan { 4358a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4359ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4360a0111cf6SLai Jiangshan } 4361a0111cf6SLai Jiangshan 4362a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4363a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4364a0111cf6SLai Jiangshan { 4365a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4366a0111cf6SLai Jiangshan 4367a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4368a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4369a0111cf6SLai Jiangshan return -EINVAL; 4370a0111cf6SLai Jiangshan 4371a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 43720a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 43730a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4374a0111cf6SLai Jiangshan return -EINVAL; 4375a0111cf6SLai Jiangshan 43760a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 43770a94efb5STejun Heo } 43780a94efb5STejun Heo 437999c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 43806201171eSwanghaibin if (!ctx) 43816201171eSwanghaibin return -ENOMEM; 4382a0111cf6SLai Jiangshan 4383a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4384a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4385a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4386a0111cf6SLai Jiangshan 43876201171eSwanghaibin return 0; 4388a0111cf6SLai Jiangshan } 4389a0111cf6SLai Jiangshan 43909e8cd2f5STejun Heo /** 43919e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 43929e8cd2f5STejun Heo * @wq: the target workqueue 43939e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 43949e8cd2f5STejun Heo * 43954c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 43964c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 43974c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 43984c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 43994c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 44004c16bd32STejun Heo * back-to-back will stay on its current pwq. 44019e8cd2f5STejun Heo * 4402d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4403d185af30SYacine Belkadi * 4404ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4405509b3204SDaniel Jordan * 4406d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 44079e8cd2f5STejun Heo */ 4408513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 44099e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 44109e8cd2f5STejun Heo { 4411a0111cf6SLai Jiangshan int ret; 44129e8cd2f5STejun Heo 4413509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4414509b3204SDaniel Jordan 4415509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4416a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4417509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 44182d5f0764SLai Jiangshan 44192d5f0764SLai Jiangshan return ret; 44209e8cd2f5STejun Heo } 44219e8cd2f5STejun Heo 44224c16bd32STejun Heo /** 44234c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 44244c16bd32STejun Heo * @wq: the target workqueue 4425*4cbfd3deSTejun Heo * @cpu: the CPU to update pool association for 4426*4cbfd3deSTejun Heo * @hotplug_cpu: the CPU coming up or going down 44274c16bd32STejun Heo * @online: whether @cpu is coming up or going down 44284c16bd32STejun Heo * 44294c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 44304c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 44314c16bd32STejun Heo * @wq accordingly. 44324c16bd32STejun Heo * 44334c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 44344c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 44354c16bd32STejun Heo * correct. 44364c16bd32STejun Heo * 44374c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 44384c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 44394c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 44404c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 44414c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 44424c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 44434c16bd32STejun Heo * CPU_DOWN_PREPARE. 44444c16bd32STejun Heo */ 44454c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 4446*4cbfd3deSTejun Heo int hotplug_cpu, bool online) 44474c16bd32STejun Heo { 44484c16bd32STejun Heo int node = cpu_to_node(cpu); 4449*4cbfd3deSTejun Heo int off_cpu = online ? -1 : hotplug_cpu; 44504c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 44514c16bd32STejun Heo struct workqueue_attrs *target_attrs; 44524c16bd32STejun Heo cpumask_t *cpumask; 44534c16bd32STejun Heo 44544c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 44554c16bd32STejun Heo 4456f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4457f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 44584c16bd32STejun Heo return; 44594c16bd32STejun Heo 44604c16bd32STejun Heo /* 44614c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 44624c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 44634c16bd32STejun Heo * CPU hotplug exclusion. 44644c16bd32STejun Heo */ 44654c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 44664c16bd32STejun Heo cpumask = target_attrs->cpumask; 44674c16bd32STejun Heo 44684c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 44694c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 44704c16bd32STejun Heo 44714c16bd32STejun Heo /* 44724c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4473042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4474042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4475042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 44764c16bd32STejun Heo */ 4477*4cbfd3deSTejun Heo if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, off_cpu, cpumask)) { 44784c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4479f7142ed4SLai Jiangshan return; 44804c16bd32STejun Heo } else { 44814c16bd32STejun Heo goto use_dfl_pwq; 44824c16bd32STejun Heo } 44834c16bd32STejun Heo 44844c16bd32STejun Heo /* create a new pwq */ 44854c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 44864c16bd32STejun Heo if (!pwq) { 44872d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 44884c16bd32STejun Heo wq->name); 448977f300b1SDaeseok Youn goto use_dfl_pwq; 44904c16bd32STejun Heo } 44914c16bd32STejun Heo 4492f7142ed4SLai Jiangshan /* Install the new pwq. */ 44934c16bd32STejun Heo mutex_lock(&wq->mutex); 44944c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 44954c16bd32STejun Heo goto out_unlock; 44964c16bd32STejun Heo 44974c16bd32STejun Heo use_dfl_pwq: 4498f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4499a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 45004c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4501a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 45024c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 45034c16bd32STejun Heo out_unlock: 45044c16bd32STejun Heo mutex_unlock(&wq->mutex); 45054c16bd32STejun Heo put_pwq_unlocked(old_pwq); 45064c16bd32STejun Heo } 45074c16bd32STejun Heo 450830cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 45091da177e4SLinus Torvalds { 451049e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 45118a2b7538STejun Heo int cpu, ret; 4512e1d8aa9fSFrederic Weisbecker 451330cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4514687a9aa5STejun Heo wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 4515ee1ceef7STejun Heo if (!wq->cpu_pwq) 4516687a9aa5STejun Heo goto enomem; 451730cdf249STejun Heo 451830cdf249STejun Heo for_each_possible_cpu(cpu) { 4519687a9aa5STejun Heo struct pool_workqueue **pwq_p = 4520ee1ceef7STejun Heo per_cpu_ptr(wq->cpu_pwq, cpu); 4521687a9aa5STejun Heo struct worker_pool *pool = 4522687a9aa5STejun Heo &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); 452330cdf249STejun Heo 4524687a9aa5STejun Heo *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 4525687a9aa5STejun Heo pool->node); 4526687a9aa5STejun Heo if (!*pwq_p) 4527687a9aa5STejun Heo goto enomem; 4528687a9aa5STejun Heo 4529687a9aa5STejun Heo init_pwq(*pwq_p, wq, pool); 4530f147f29eSTejun Heo 4531f147f29eSTejun Heo mutex_lock(&wq->mutex); 4532687a9aa5STejun Heo link_pwq(*pwq_p); 4533f147f29eSTejun Heo mutex_unlock(&wq->mutex); 453430cdf249STejun Heo } 453530cdf249STejun Heo return 0; 4536509b3204SDaniel Jordan } 4537509b3204SDaniel Jordan 4538ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4539509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 45408a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 45418a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 45428a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 45438a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 45448a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 45459e8cd2f5STejun Heo } else { 4546509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 45479e8cd2f5STejun Heo } 4548ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4549509b3204SDaniel Jordan 4550509b3204SDaniel Jordan return ret; 4551687a9aa5STejun Heo 4552687a9aa5STejun Heo enomem: 4553687a9aa5STejun Heo if (wq->cpu_pwq) { 4554687a9aa5STejun Heo for_each_possible_cpu(cpu) 4555687a9aa5STejun Heo kfree(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4556687a9aa5STejun Heo free_percpu(wq->cpu_pwq); 4557687a9aa5STejun Heo wq->cpu_pwq = NULL; 4558687a9aa5STejun Heo } 4559687a9aa5STejun Heo return -ENOMEM; 45600f900049STejun Heo } 45610f900049STejun Heo 4562f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4563f3421797STejun Heo const char *name) 4564b71ab8c2STejun Heo { 4565f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4566f3421797STejun Heo 4567f3421797STejun Heo if (max_active < 1 || max_active > lim) 4568044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4569f3421797STejun Heo max_active, name, 1, lim); 4570b71ab8c2STejun Heo 4571f3421797STejun Heo return clamp_val(max_active, 1, lim); 4572b71ab8c2STejun Heo } 4573b71ab8c2STejun Heo 4574983c7515STejun Heo /* 4575983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4576983c7515STejun Heo * to guarantee forward progress. 4577983c7515STejun Heo */ 4578983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4579983c7515STejun Heo { 4580983c7515STejun Heo struct worker *rescuer; 4581b92b36eaSDan Carpenter int ret; 4582983c7515STejun Heo 4583983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4584983c7515STejun Heo return 0; 4585983c7515STejun Heo 4586983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 45874c0736a7SPetr Mladek if (!rescuer) { 45884c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 45894c0736a7SPetr Mladek wq->name); 4590983c7515STejun Heo return -ENOMEM; 45914c0736a7SPetr Mladek } 4592983c7515STejun Heo 4593983c7515STejun Heo rescuer->rescue_wq = wq; 4594983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4595f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4596b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 45974c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 45984c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4599983c7515STejun Heo kfree(rescuer); 4600b92b36eaSDan Carpenter return ret; 4601983c7515STejun Heo } 4602983c7515STejun Heo 4603983c7515STejun Heo wq->rescuer = rescuer; 4604983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4605983c7515STejun Heo wake_up_process(rescuer->task); 4606983c7515STejun Heo 4607983c7515STejun Heo return 0; 4608983c7515STejun Heo } 4609983c7515STejun Heo 4610a2775bbcSMathieu Malaterre __printf(1, 4) 4611669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 461297e37d7bSTejun Heo unsigned int flags, 4613669de8bdSBart Van Assche int max_active, ...) 46143af24433SOleg Nesterov { 4615df2d5ae4STejun Heo size_t tbl_size = 0; 4616ecf6881fSTejun Heo va_list args; 46173af24433SOleg Nesterov struct workqueue_struct *wq; 461849e3cf44STejun Heo struct pool_workqueue *pwq; 4619b196be89STejun Heo 46205c0338c6STejun Heo /* 46215c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 46225c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 46235c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 46245c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 46255c0338c6STejun Heo * on NUMA. 46265c0338c6STejun Heo */ 46275c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 46285c0338c6STejun Heo flags |= __WQ_ORDERED; 46295c0338c6STejun Heo 4630cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4631cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4632cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4633cee22a15SViresh Kumar 4634ecf6881fSTejun Heo /* allocate wq and format name */ 4635df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4636ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4637df2d5ae4STejun Heo 4638df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4639b196be89STejun Heo if (!wq) 4640d2c1d404STejun Heo return NULL; 4641b196be89STejun Heo 46426029a918STejun Heo if (flags & WQ_UNBOUND) { 4643be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 46446029a918STejun Heo if (!wq->unbound_attrs) 46456029a918STejun Heo goto err_free_wq; 46466029a918STejun Heo } 46476029a918STejun Heo 4648669de8bdSBart Van Assche va_start(args, max_active); 4649ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4650b196be89STejun Heo va_end(args); 46513af24433SOleg Nesterov 4652d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4653b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 46543af24433SOleg Nesterov 4655b196be89STejun Heo /* init wq */ 465697e37d7bSTejun Heo wq->flags = flags; 4657a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 46583c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4659112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 466030cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 466173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 466273f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4663493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 46643af24433SOleg Nesterov 4665669de8bdSBart Van Assche wq_init_lockdep(wq); 4666cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 46673af24433SOleg Nesterov 466830cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 466982efcab3SBart Van Assche goto err_unreg_lockdep; 46701537663fSTejun Heo 467140c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4672d2c1d404STejun Heo goto err_destroy; 4673e22bee78STejun Heo 4674226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4675226223abSTejun Heo goto err_destroy; 4676226223abSTejun Heo 46776af8bf3dSOleg Nesterov /* 467868e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 467968e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 468068e13a67SLai Jiangshan * list. 46816af8bf3dSOleg Nesterov */ 468268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4683a0a1a5fdSTejun Heo 4684a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 468549e3cf44STejun Heo for_each_pwq(pwq, wq) 4686699ce097STejun Heo pwq_adjust_max_active(pwq); 4687a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4688a0a1a5fdSTejun Heo 4689e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4690a0a1a5fdSTejun Heo 469168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46923af24433SOleg Nesterov 46933af24433SOleg Nesterov return wq; 4694d2c1d404STejun Heo 469582efcab3SBart Van Assche err_unreg_lockdep: 4696009bb421SBart Van Assche wq_unregister_lockdep(wq); 4697009bb421SBart Van Assche wq_free_lockdep(wq); 469882efcab3SBart Van Assche err_free_wq: 46996029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 47004690c4abSTejun Heo kfree(wq); 4701d2c1d404STejun Heo return NULL; 4702d2c1d404STejun Heo err_destroy: 4703d2c1d404STejun Heo destroy_workqueue(wq); 47044690c4abSTejun Heo return NULL; 47051da177e4SLinus Torvalds } 4706669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 47071da177e4SLinus Torvalds 4708c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4709c29eb853STejun Heo { 4710c29eb853STejun Heo int i; 4711c29eb853STejun Heo 4712c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4713c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4714c29eb853STejun Heo return true; 4715c29eb853STejun Heo 4716c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4717c29eb853STejun Heo return true; 4718f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4719c29eb853STejun Heo return true; 4720c29eb853STejun Heo 4721c29eb853STejun Heo return false; 4722c29eb853STejun Heo } 4723c29eb853STejun Heo 47243af24433SOleg Nesterov /** 47253af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 47263af24433SOleg Nesterov * @wq: target workqueue 47273af24433SOleg Nesterov * 47283af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 47293af24433SOleg Nesterov */ 47303af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 47313af24433SOleg Nesterov { 473249e3cf44STejun Heo struct pool_workqueue *pwq; 4733687a9aa5STejun Heo int cpu, node; 47343af24433SOleg Nesterov 4735def98c84STejun Heo /* 4736def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4737def98c84STejun Heo * lead to sysfs name conflicts. 4738def98c84STejun Heo */ 4739def98c84STejun Heo workqueue_sysfs_unregister(wq); 4740def98c84STejun Heo 474133e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 474233e3f0a3SRichard Clark mutex_lock(&wq->mutex); 474333e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 474433e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 474533e3f0a3SRichard Clark 47469c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 47479c5a2ba7STejun Heo drain_workqueue(wq); 4748c8efcc25STejun Heo 4749def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4750def98c84STejun Heo if (wq->rescuer) { 4751def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4752def98c84STejun Heo 4753def98c84STejun Heo /* this prevents new queueing */ 4754a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4755def98c84STejun Heo wq->rescuer = NULL; 4756a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4757def98c84STejun Heo 4758def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4759def98c84STejun Heo kthread_stop(rescuer->task); 47608efe1223STejun Heo kfree(rescuer); 4761def98c84STejun Heo } 4762def98c84STejun Heo 4763c29eb853STejun Heo /* 4764c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4765c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4766c29eb853STejun Heo */ 4767c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4768b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 476949e3cf44STejun Heo for_each_pwq(pwq, wq) { 4770a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4771c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 47721d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4773e66b39afSTejun Heo __func__, wq->name); 4774c29eb853STejun Heo show_pwq(pwq); 4775a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4776b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4777c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 477855df0933SImran Khan show_one_workqueue(wq); 47796183c009STejun Heo return; 478076af4d93STejun Heo } 4781a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 478276af4d93STejun Heo } 4783b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 47846183c009STejun Heo 4785a0a1a5fdSTejun Heo /* 4786a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4787a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4788a0a1a5fdSTejun Heo */ 4789e2dca7adSTejun Heo list_del_rcu(&wq->list); 479068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47913af24433SOleg Nesterov 47928864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4793687a9aa5STejun Heo for_each_possible_cpu(cpu) 4794687a9aa5STejun Heo put_pwq_unlocked(*per_cpu_ptr(wq->cpu_pwq, cpu)); 47958864b4e5STejun Heo } else { 47968864b4e5STejun Heo /* 47978864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 47984c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 47994c16bd32STejun Heo * @wq will be freed when the last pwq is released. 48008864b4e5STejun Heo */ 48014c16bd32STejun Heo for_each_node(node) { 48024c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 48034c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 48044c16bd32STejun Heo put_pwq_unlocked(pwq); 48054c16bd32STejun Heo } 48064c16bd32STejun Heo 48074c16bd32STejun Heo /* 48084c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 48094c16bd32STejun Heo * put. Don't access it afterwards. 48104c16bd32STejun Heo */ 48114c16bd32STejun Heo pwq = wq->dfl_pwq; 48124c16bd32STejun Heo wq->dfl_pwq = NULL; 4813dce90d47STejun Heo put_pwq_unlocked(pwq); 481429c91e99STejun Heo } 48153af24433SOleg Nesterov } 48163af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 48173af24433SOleg Nesterov 4818dcd989cbSTejun Heo /** 4819dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4820dcd989cbSTejun Heo * @wq: target workqueue 4821dcd989cbSTejun Heo * @max_active: new max_active value. 4822dcd989cbSTejun Heo * 4823dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4824dcd989cbSTejun Heo * 4825dcd989cbSTejun Heo * CONTEXT: 4826dcd989cbSTejun Heo * Don't call from IRQ context. 4827dcd989cbSTejun Heo */ 4828dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4829dcd989cbSTejun Heo { 483049e3cf44STejun Heo struct pool_workqueue *pwq; 4831dcd989cbSTejun Heo 48328719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 48330a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 48348719dceaSTejun Heo return; 48358719dceaSTejun Heo 4836f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4837dcd989cbSTejun Heo 4838a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4839dcd989cbSTejun Heo 48400a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4841dcd989cbSTejun Heo wq->saved_max_active = max_active; 4842dcd989cbSTejun Heo 4843699ce097STejun Heo for_each_pwq(pwq, wq) 4844699ce097STejun Heo pwq_adjust_max_active(pwq); 4845dcd989cbSTejun Heo 4846a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4847dcd989cbSTejun Heo } 4848dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4849dcd989cbSTejun Heo 4850dcd989cbSTejun Heo /** 485127d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 485227d4ee03SLukas Wunner * 485327d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 485427d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 485527d4ee03SLukas Wunner * 485627d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 485727d4ee03SLukas Wunner */ 485827d4ee03SLukas Wunner struct work_struct *current_work(void) 485927d4ee03SLukas Wunner { 486027d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 486127d4ee03SLukas Wunner 486227d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 486327d4ee03SLukas Wunner } 486427d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 486527d4ee03SLukas Wunner 486627d4ee03SLukas Wunner /** 4867e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4868e6267616STejun Heo * 4869e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4870e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4871d185af30SYacine Belkadi * 4872d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4873e6267616STejun Heo */ 4874e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4875e6267616STejun Heo { 4876e6267616STejun Heo struct worker *worker = current_wq_worker(); 4877e6267616STejun Heo 48786a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4879e6267616STejun Heo } 4880e6267616STejun Heo 4881e6267616STejun Heo /** 4882dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4883dcd989cbSTejun Heo * @cpu: CPU in question 4884dcd989cbSTejun Heo * @wq: target workqueue 4885dcd989cbSTejun Heo * 4886dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4887dcd989cbSTejun Heo * no synchronization around this function and the test result is 4888dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4889dcd989cbSTejun Heo * 4890d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4891d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4892d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4893d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4894d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4895d3251859STejun Heo * 4896d185af30SYacine Belkadi * Return: 4897dcd989cbSTejun Heo * %true if congested, %false otherwise. 4898dcd989cbSTejun Heo */ 4899d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4900dcd989cbSTejun Heo { 49017fb98ea7STejun Heo struct pool_workqueue *pwq; 490276af4d93STejun Heo bool ret; 490376af4d93STejun Heo 490424acfb71SThomas Gleixner rcu_read_lock(); 490524acfb71SThomas Gleixner preempt_disable(); 49067fb98ea7STejun Heo 4907d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4908d3251859STejun Heo cpu = smp_processor_id(); 4909d3251859STejun Heo 49107fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 4911687a9aa5STejun Heo pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 49127fb98ea7STejun Heo else 4913df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4914dcd989cbSTejun Heo 4915f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 491624acfb71SThomas Gleixner preempt_enable(); 491724acfb71SThomas Gleixner rcu_read_unlock(); 491876af4d93STejun Heo 491976af4d93STejun Heo return ret; 4920dcd989cbSTejun Heo } 4921dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4922dcd989cbSTejun Heo 4923dcd989cbSTejun Heo /** 4924dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4925dcd989cbSTejun Heo * @work: the work to be tested 4926dcd989cbSTejun Heo * 4927dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4928dcd989cbSTejun Heo * synchronization around this function and the test result is 4929dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4930dcd989cbSTejun Heo * 4931d185af30SYacine Belkadi * Return: 4932dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4933dcd989cbSTejun Heo */ 4934dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4935dcd989cbSTejun Heo { 4936fa1b54e6STejun Heo struct worker_pool *pool; 4937dcd989cbSTejun Heo unsigned long flags; 4938dcd989cbSTejun Heo unsigned int ret = 0; 4939dcd989cbSTejun Heo 4940dcd989cbSTejun Heo if (work_pending(work)) 4941dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4942038366c5SLai Jiangshan 494324acfb71SThomas Gleixner rcu_read_lock(); 4944fa1b54e6STejun Heo pool = get_work_pool(work); 4945038366c5SLai Jiangshan if (pool) { 4946a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4947c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4948dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4949a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4950038366c5SLai Jiangshan } 495124acfb71SThomas Gleixner rcu_read_unlock(); 4952dcd989cbSTejun Heo 4953dcd989cbSTejun Heo return ret; 4954dcd989cbSTejun Heo } 4955dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4956dcd989cbSTejun Heo 49573d1cb205STejun Heo /** 49583d1cb205STejun Heo * set_worker_desc - set description for the current work item 49593d1cb205STejun Heo * @fmt: printf-style format string 49603d1cb205STejun Heo * @...: arguments for the format string 49613d1cb205STejun Heo * 49623d1cb205STejun Heo * This function can be called by a running work function to describe what 49633d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 49643d1cb205STejun Heo * information will be printed out together to help debugging. The 49653d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 49663d1cb205STejun Heo */ 49673d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 49683d1cb205STejun Heo { 49693d1cb205STejun Heo struct worker *worker = current_wq_worker(); 49703d1cb205STejun Heo va_list args; 49713d1cb205STejun Heo 49723d1cb205STejun Heo if (worker) { 49733d1cb205STejun Heo va_start(args, fmt); 49743d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 49753d1cb205STejun Heo va_end(args); 49763d1cb205STejun Heo } 49773d1cb205STejun Heo } 49785c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 49793d1cb205STejun Heo 49803d1cb205STejun Heo /** 49813d1cb205STejun Heo * print_worker_info - print out worker information and description 49823d1cb205STejun Heo * @log_lvl: the log level to use when printing 49833d1cb205STejun Heo * @task: target task 49843d1cb205STejun Heo * 49853d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 49863d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 49873d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 49883d1cb205STejun Heo * 49893d1cb205STejun Heo * This function can be safely called on any task as long as the 49903d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 49913d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 49923d1cb205STejun Heo */ 49933d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 49943d1cb205STejun Heo { 49953d1cb205STejun Heo work_func_t *fn = NULL; 49963d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 49973d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 49983d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 49993d1cb205STejun Heo struct workqueue_struct *wq = NULL; 50003d1cb205STejun Heo struct worker *worker; 50013d1cb205STejun Heo 50023d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 50033d1cb205STejun Heo return; 50043d1cb205STejun Heo 50053d1cb205STejun Heo /* 50063d1cb205STejun Heo * This function is called without any synchronization and @task 50073d1cb205STejun Heo * could be in any state. Be careful with dereferences. 50083d1cb205STejun Heo */ 5009e700591aSPetr Mladek worker = kthread_probe_data(task); 50103d1cb205STejun Heo 50113d1cb205STejun Heo /* 50128bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 50138bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 50143d1cb205STejun Heo */ 5015fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5016fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5017fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5018fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5019fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 50203d1cb205STejun Heo 50213d1cb205STejun Heo if (fn || name[0] || desc[0]) { 5022d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 50238bf89593STejun Heo if (strcmp(name, desc)) 50243d1cb205STejun Heo pr_cont(" (%s)", desc); 50253d1cb205STejun Heo pr_cont("\n"); 50263d1cb205STejun Heo } 50273d1cb205STejun Heo } 50283d1cb205STejun Heo 50293494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 50303494fc30STejun Heo { 50313494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 50323494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 50333494fc30STejun Heo pr_cont(" node=%d", pool->node); 50343494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 50353494fc30STejun Heo } 50363494fc30STejun Heo 5037c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 5038c76feb0dSPaul E. McKenney bool comma; 5039c76feb0dSPaul E. McKenney work_func_t func; 5040c76feb0dSPaul E. McKenney long ctr; 5041c76feb0dSPaul E. McKenney }; 5042c76feb0dSPaul E. McKenney 5043c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5044c76feb0dSPaul E. McKenney { 5045c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 5046c76feb0dSPaul E. McKenney goto out_record; 5047c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 5048c76feb0dSPaul E. McKenney pcwsp->ctr++; 5049c76feb0dSPaul E. McKenney return; 5050c76feb0dSPaul E. McKenney } 5051c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 5052c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5053c76feb0dSPaul E. McKenney else 5054c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5055c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 5056c76feb0dSPaul E. McKenney out_record: 5057c76feb0dSPaul E. McKenney if ((long)func == -1L) 5058c76feb0dSPaul E. McKenney return; 5059c76feb0dSPaul E. McKenney pcwsp->comma = comma; 5060c76feb0dSPaul E. McKenney pcwsp->func = func; 5061c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 5062c76feb0dSPaul E. McKenney } 5063c76feb0dSPaul E. McKenney 5064c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 50653494fc30STejun Heo { 50663494fc30STejun Heo if (work->func == wq_barrier_func) { 50673494fc30STejun Heo struct wq_barrier *barr; 50683494fc30STejun Heo 50693494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 50703494fc30STejun Heo 5071c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 50723494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 50733494fc30STejun Heo task_pid_nr(barr->task)); 50743494fc30STejun Heo } else { 5075c76feb0dSPaul E. McKenney if (!comma) 5076c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5077c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 50783494fc30STejun Heo } 50793494fc30STejun Heo } 50803494fc30STejun Heo 50813494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 50823494fc30STejun Heo { 5083c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 50843494fc30STejun Heo struct worker_pool *pool = pwq->pool; 50853494fc30STejun Heo struct work_struct *work; 50863494fc30STejun Heo struct worker *worker; 50873494fc30STejun Heo bool has_in_flight = false, has_pending = false; 50883494fc30STejun Heo int bkt; 50893494fc30STejun Heo 50903494fc30STejun Heo pr_info(" pwq %d:", pool->id); 50913494fc30STejun Heo pr_cont_pool_info(pool); 50923494fc30STejun Heo 5093e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5094e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 50953494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 50963494fc30STejun Heo 50973494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 50983494fc30STejun Heo if (worker->current_pwq == pwq) { 50993494fc30STejun Heo has_in_flight = true; 51003494fc30STejun Heo break; 51013494fc30STejun Heo } 51023494fc30STejun Heo } 51033494fc30STejun Heo if (has_in_flight) { 51043494fc30STejun Heo bool comma = false; 51053494fc30STejun Heo 51063494fc30STejun Heo pr_info(" in-flight:"); 51073494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51083494fc30STejun Heo if (worker->current_pwq != pwq) 51093494fc30STejun Heo continue; 51103494fc30STejun Heo 5111d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 51123494fc30STejun Heo task_pid_nr(worker->task), 511330ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 51143494fc30STejun Heo worker->current_func); 51153494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5116c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5117c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51183494fc30STejun Heo comma = true; 51193494fc30STejun Heo } 51203494fc30STejun Heo pr_cont("\n"); 51213494fc30STejun Heo } 51223494fc30STejun Heo 51233494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51243494fc30STejun Heo if (get_work_pwq(work) == pwq) { 51253494fc30STejun Heo has_pending = true; 51263494fc30STejun Heo break; 51273494fc30STejun Heo } 51283494fc30STejun Heo } 51293494fc30STejun Heo if (has_pending) { 51303494fc30STejun Heo bool comma = false; 51313494fc30STejun Heo 51323494fc30STejun Heo pr_info(" pending:"); 51333494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51343494fc30STejun Heo if (get_work_pwq(work) != pwq) 51353494fc30STejun Heo continue; 51363494fc30STejun Heo 5137c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51383494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51393494fc30STejun Heo } 5140c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51413494fc30STejun Heo pr_cont("\n"); 51423494fc30STejun Heo } 51433494fc30STejun Heo 5144f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 51453494fc30STejun Heo bool comma = false; 51463494fc30STejun Heo 5147f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5148f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5149c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51503494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51513494fc30STejun Heo } 5152c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51533494fc30STejun Heo pr_cont("\n"); 51543494fc30STejun Heo } 51553494fc30STejun Heo } 51563494fc30STejun Heo 51573494fc30STejun Heo /** 515855df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 515955df0933SImran Khan * @wq: workqueue whose state will be printed 51603494fc30STejun Heo */ 516155df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 51623494fc30STejun Heo { 51633494fc30STejun Heo struct pool_workqueue *pwq; 51643494fc30STejun Heo bool idle = true; 516555df0933SImran Khan unsigned long flags; 51663494fc30STejun Heo 51673494fc30STejun Heo for_each_pwq(pwq, wq) { 5168f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 51693494fc30STejun Heo idle = false; 51703494fc30STejun Heo break; 51713494fc30STejun Heo } 51723494fc30STejun Heo } 517355df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 517455df0933SImran Khan return; 51753494fc30STejun Heo 51763494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 51773494fc30STejun Heo 51783494fc30STejun Heo for_each_pwq(pwq, wq) { 5179a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 518057116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 518157116ce1SJohan Hovold /* 518257116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 518357116ce1SJohan Hovold * drivers that queue work while holding locks 518457116ce1SJohan Hovold * also taken in their write paths. 518557116ce1SJohan Hovold */ 518657116ce1SJohan Hovold printk_deferred_enter(); 51873494fc30STejun Heo show_pwq(pwq); 518857116ce1SJohan Hovold printk_deferred_exit(); 518957116ce1SJohan Hovold } 5190a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 519162635ea8SSergey Senozhatsky /* 519262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 519355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 519462635ea8SSergey Senozhatsky * hard lockup. 519562635ea8SSergey Senozhatsky */ 519662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 51973494fc30STejun Heo } 519855df0933SImran Khan 51993494fc30STejun Heo } 52003494fc30STejun Heo 520155df0933SImran Khan /** 520255df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 520355df0933SImran Khan * @pool: worker pool whose state will be printed 520455df0933SImran Khan */ 520555df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 520655df0933SImran Khan { 52073494fc30STejun Heo struct worker *worker; 52083494fc30STejun Heo bool first = true; 520955df0933SImran Khan unsigned long flags; 5210335a42ebSPetr Mladek unsigned long hung = 0; 52113494fc30STejun Heo 5212a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 52133494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 52143494fc30STejun Heo goto next_pool; 5215335a42ebSPetr Mladek 5216335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5217335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5218335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5219335a42ebSPetr Mladek 522057116ce1SJohan Hovold /* 522157116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 522257116ce1SJohan Hovold * queue work while holding locks also taken in their write 522357116ce1SJohan Hovold * paths. 522457116ce1SJohan Hovold */ 522557116ce1SJohan Hovold printk_deferred_enter(); 52263494fc30STejun Heo pr_info("pool %d:", pool->id); 52273494fc30STejun Heo pr_cont_pool_info(pool); 5228335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 52293494fc30STejun Heo if (pool->manager) 52303494fc30STejun Heo pr_cont(" manager: %d", 52313494fc30STejun Heo task_pid_nr(pool->manager->task)); 52323494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 52333494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 52343494fc30STejun Heo task_pid_nr(worker->task)); 52353494fc30STejun Heo first = false; 52363494fc30STejun Heo } 52373494fc30STejun Heo pr_cont("\n"); 523857116ce1SJohan Hovold printk_deferred_exit(); 52393494fc30STejun Heo next_pool: 5240a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 524162635ea8SSergey Senozhatsky /* 524262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 524355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 524462635ea8SSergey Senozhatsky * hard lockup. 524562635ea8SSergey Senozhatsky */ 524662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 524755df0933SImran Khan 52483494fc30STejun Heo } 52493494fc30STejun Heo 525055df0933SImran Khan /** 525155df0933SImran Khan * show_all_workqueues - dump workqueue state 525255df0933SImran Khan * 5253704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 525455df0933SImran Khan */ 525555df0933SImran Khan void show_all_workqueues(void) 525655df0933SImran Khan { 525755df0933SImran Khan struct workqueue_struct *wq; 525855df0933SImran Khan struct worker_pool *pool; 525955df0933SImran Khan int pi; 526055df0933SImran Khan 526155df0933SImran Khan rcu_read_lock(); 526255df0933SImran Khan 526355df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 526455df0933SImran Khan 526555df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 526655df0933SImran Khan show_one_workqueue(wq); 526755df0933SImran Khan 526855df0933SImran Khan for_each_pool(pool, pi) 526955df0933SImran Khan show_one_worker_pool(pool); 527055df0933SImran Khan 527124acfb71SThomas Gleixner rcu_read_unlock(); 52723494fc30STejun Heo } 52733494fc30STejun Heo 5274704bc669SJungseung Lee /** 5275704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5276704bc669SJungseung Lee * 5277704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5278704bc669SJungseung Lee * still busy. 5279704bc669SJungseung Lee */ 5280704bc669SJungseung Lee void show_freezable_workqueues(void) 5281704bc669SJungseung Lee { 5282704bc669SJungseung Lee struct workqueue_struct *wq; 5283704bc669SJungseung Lee 5284704bc669SJungseung Lee rcu_read_lock(); 5285704bc669SJungseung Lee 5286704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5287704bc669SJungseung Lee 5288704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5289704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5290704bc669SJungseung Lee continue; 5291704bc669SJungseung Lee show_one_workqueue(wq); 5292704bc669SJungseung Lee } 5293704bc669SJungseung Lee 5294704bc669SJungseung Lee rcu_read_unlock(); 5295704bc669SJungseung Lee } 5296704bc669SJungseung Lee 52976b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 52986b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 52996b59808bSTejun Heo { 53006b59808bSTejun Heo int off; 53016b59808bSTejun Heo 53026b59808bSTejun Heo /* always show the actual comm */ 53036b59808bSTejun Heo off = strscpy(buf, task->comm, size); 53046b59808bSTejun Heo if (off < 0) 53056b59808bSTejun Heo return; 53066b59808bSTejun Heo 5307197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 53086b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 53096b59808bSTejun Heo 5310197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5311197f6accSTejun Heo struct worker *worker = kthread_data(task); 5312197f6accSTejun Heo struct worker_pool *pool = worker->pool; 53136b59808bSTejun Heo 53146b59808bSTejun Heo if (pool) { 5315a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 53166b59808bSTejun Heo /* 5317197f6accSTejun Heo * ->desc tracks information (wq name or 5318197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5319197f6accSTejun Heo * current, prepend '+', otherwise '-'. 53206b59808bSTejun Heo */ 53216b59808bSTejun Heo if (worker->desc[0] != '\0') { 53226b59808bSTejun Heo if (worker->current_work) 53236b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 53246b59808bSTejun Heo worker->desc); 53256b59808bSTejun Heo else 53266b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 53276b59808bSTejun Heo worker->desc); 53286b59808bSTejun Heo } 5329a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 53306b59808bSTejun Heo } 5331197f6accSTejun Heo } 53326b59808bSTejun Heo 53336b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 53346b59808bSTejun Heo } 53356b59808bSTejun Heo 533666448bc2SMathieu Malaterre #ifdef CONFIG_SMP 533766448bc2SMathieu Malaterre 5338db7bccf4STejun Heo /* 5339db7bccf4STejun Heo * CPU hotplug. 5340db7bccf4STejun Heo * 5341e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5342112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5343706026c2STejun Heo * pool which make migrating pending and scheduled works very 5344e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 534594cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5346e22bee78STejun Heo * blocked draining impractical. 5347e22bee78STejun Heo * 534824647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5349628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5350628c78e7STejun Heo * cpu comes back online. 5351db7bccf4STejun Heo */ 5352db7bccf4STejun Heo 5353e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5354db7bccf4STejun Heo { 53554ce62e9eSTejun Heo struct worker_pool *pool; 5356db7bccf4STejun Heo struct worker *worker; 5357db7bccf4STejun Heo 5358f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53591258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5360a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5361e22bee78STejun Heo 5362f2d5a0eeSTejun Heo /* 536392f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 536494cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 536511b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5366b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5367b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5368b4ac9384SLai Jiangshan * is on the same cpu. 5369f2d5a0eeSTejun Heo */ 5370da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5371403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5372db7bccf4STejun Heo 537324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5374f2d5a0eeSTejun Heo 5375e22bee78STejun Heo /* 5376989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5377989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5378989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5379989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5380989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5381eb283428SLai Jiangshan * are served by workers tied to the pool. 5382e22bee78STejun Heo */ 5383bc35f7efSLai Jiangshan pool->nr_running = 0; 5384eb283428SLai Jiangshan 5385eb283428SLai Jiangshan /* 5386eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5387eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5388eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5389eb283428SLai Jiangshan */ 5390eb283428SLai Jiangshan wake_up_worker(pool); 5391989442d7SLai Jiangshan 5392a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5393989442d7SLai Jiangshan 5394793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5395793777bcSValentin Schneider unbind_worker(worker); 5396989442d7SLai Jiangshan 5397989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5398eb283428SLai Jiangshan } 5399db7bccf4STejun Heo } 5400db7bccf4STejun Heo 5401bd7c089eSTejun Heo /** 5402bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5403bd7c089eSTejun Heo * @pool: pool of interest 5404bd7c089eSTejun Heo * 5405a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5406bd7c089eSTejun Heo */ 5407bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5408bd7c089eSTejun Heo { 5409a9ab775bSTejun Heo struct worker *worker; 5410bd7c089eSTejun Heo 54111258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5412bd7c089eSTejun Heo 5413bd7c089eSTejun Heo /* 5414a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5415a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5416402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5417a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5418a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5419bd7c089eSTejun Heo */ 5420c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5421c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5422c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5423c63a2e52SValentin Schneider pool->attrs->cpumask) < 0); 5424c63a2e52SValentin Schneider } 5425a9ab775bSTejun Heo 5426a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5427f7c17d26SWanpeng Li 54283de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5429a9ab775bSTejun Heo 5430da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5431a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5432a9ab775bSTejun Heo 5433a9ab775bSTejun Heo /* 5434a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5435a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5436a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5437a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5438a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5439a9ab775bSTejun Heo * concurrency management. Note that when or whether 5440a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5441a9ab775bSTejun Heo * 5442c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5443a9ab775bSTejun Heo * tested without holding any lock in 54446d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5445a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5446a9ab775bSTejun Heo * management operations. 5447bd7c089eSTejun Heo */ 5448a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5449a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5450a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5451c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5452bd7c089eSTejun Heo } 5453a9ab775bSTejun Heo 5454a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5455bd7c089eSTejun Heo } 5456bd7c089eSTejun Heo 54577dbc725eSTejun Heo /** 54587dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 54597dbc725eSTejun Heo * @pool: unbound pool of interest 54607dbc725eSTejun Heo * @cpu: the CPU which is coming up 54617dbc725eSTejun Heo * 54627dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 54637dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 54647dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 54657dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 54667dbc725eSTejun Heo */ 54677dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 54687dbc725eSTejun Heo { 54697dbc725eSTejun Heo static cpumask_t cpumask; 54707dbc725eSTejun Heo struct worker *worker; 54717dbc725eSTejun Heo 54721258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 54737dbc725eSTejun Heo 54747dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 54757dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 54767dbc725eSTejun Heo return; 54777dbc725eSTejun Heo 54787dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 54797dbc725eSTejun Heo 54807dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5481da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5482d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 54837dbc725eSTejun Heo } 54847dbc725eSTejun Heo 54857ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 54861da177e4SLinus Torvalds { 54874ce62e9eSTejun Heo struct worker_pool *pool; 54881da177e4SLinus Torvalds 5489f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54903ce63377STejun Heo if (pool->nr_workers) 54913ce63377STejun Heo continue; 5492051e1850SLai Jiangshan if (!create_worker(pool)) 54937ee681b2SThomas Gleixner return -ENOMEM; 54943af24433SOleg Nesterov } 54957ee681b2SThomas Gleixner return 0; 54967ee681b2SThomas Gleixner } 54971da177e4SLinus Torvalds 54987ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 54997ee681b2SThomas Gleixner { 55007ee681b2SThomas Gleixner struct worker_pool *pool; 55017ee681b2SThomas Gleixner struct workqueue_struct *wq; 55027ee681b2SThomas Gleixner int pi; 55037ee681b2SThomas Gleixner 550468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55057dbc725eSTejun Heo 55067dbc725eSTejun Heo for_each_pool(pool, pi) { 55071258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 550894cf58bbSTejun Heo 5509f05b558dSLai Jiangshan if (pool->cpu == cpu) 551094cf58bbSTejun Heo rebind_workers(pool); 5511f05b558dSLai Jiangshan else if (pool->cpu < 0) 55127dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 551394cf58bbSTejun Heo 55141258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 551594cf58bbSTejun Heo } 55167dbc725eSTejun Heo 55174c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 5518*4cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 5519*4cbfd3deSTejun Heo int tcpu; 5520*4cbfd3deSTejun Heo 5521*4cbfd3deSTejun Heo for_each_possible_cpu(tcpu) { 5522*4cbfd3deSTejun Heo if (cpu_to_node(tcpu) == cpu_to_node(cpu)) { 5523*4cbfd3deSTejun Heo wq_update_unbound_numa(wq, tcpu, cpu, true); 5524*4cbfd3deSTejun Heo } 5525*4cbfd3deSTejun Heo } 5526*4cbfd3deSTejun Heo } 55274c16bd32STejun Heo 552868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55297ee681b2SThomas Gleixner return 0; 553065758202STejun Heo } 553165758202STejun Heo 55327ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 553365758202STejun Heo { 55344c16bd32STejun Heo struct workqueue_struct *wq; 55358db25e78STejun Heo 55364c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5537e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5538e8b3f8dbSLai Jiangshan return -1; 5539e8b3f8dbSLai Jiangshan 5540e8b3f8dbSLai Jiangshan unbind_workers(cpu); 55414c16bd32STejun Heo 55424c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55434c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 5544*4cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 5545*4cbfd3deSTejun Heo int tcpu; 5546*4cbfd3deSTejun Heo 5547*4cbfd3deSTejun Heo for_each_possible_cpu(tcpu) { 5548*4cbfd3deSTejun Heo if (cpu_to_node(tcpu) == cpu_to_node(cpu)) { 5549*4cbfd3deSTejun Heo wq_update_unbound_numa(wq, tcpu, cpu, false); 5550*4cbfd3deSTejun Heo } 5551*4cbfd3deSTejun Heo } 5552*4cbfd3deSTejun Heo } 55534c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 55544c16bd32STejun Heo 55557ee681b2SThomas Gleixner return 0; 555665758202STejun Heo } 555765758202STejun Heo 55582d3854a3SRusty Russell struct work_for_cpu { 5559ed48ece2STejun Heo struct work_struct work; 55602d3854a3SRusty Russell long (*fn)(void *); 55612d3854a3SRusty Russell void *arg; 55622d3854a3SRusty Russell long ret; 55632d3854a3SRusty Russell }; 55642d3854a3SRusty Russell 5565ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 55662d3854a3SRusty Russell { 5567ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5568ed48ece2STejun Heo 55692d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 55702d3854a3SRusty Russell } 55712d3854a3SRusty Russell 55722d3854a3SRusty Russell /** 557322aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 55742d3854a3SRusty Russell * @cpu: the cpu to run on 55752d3854a3SRusty Russell * @fn: the function to run 55762d3854a3SRusty Russell * @arg: the function arg 55772d3854a3SRusty Russell * 557831ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 55796b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5580d185af30SYacine Belkadi * 5581d185af30SYacine Belkadi * Return: The value @fn returns. 55822d3854a3SRusty Russell */ 5583d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 55842d3854a3SRusty Russell { 5585ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 55862d3854a3SRusty Russell 5587ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5588ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 558912997d1aSBjorn Helgaas flush_work(&wfc.work); 5590440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 55912d3854a3SRusty Russell return wfc.ret; 55922d3854a3SRusty Russell } 55932d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 55940e8d6a93SThomas Gleixner 55950e8d6a93SThomas Gleixner /** 55960e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 55970e8d6a93SThomas Gleixner * @cpu: the cpu to run on 55980e8d6a93SThomas Gleixner * @fn: the function to run 55990e8d6a93SThomas Gleixner * @arg: the function argument 56000e8d6a93SThomas Gleixner * 56010e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 56020e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 56030e8d6a93SThomas Gleixner * 56040e8d6a93SThomas Gleixner * Return: The value @fn returns. 56050e8d6a93SThomas Gleixner */ 56060e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 56070e8d6a93SThomas Gleixner { 56080e8d6a93SThomas Gleixner long ret = -ENODEV; 56090e8d6a93SThomas Gleixner 5610ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 56110e8d6a93SThomas Gleixner if (cpu_online(cpu)) 56120e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5613ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 56140e8d6a93SThomas Gleixner return ret; 56150e8d6a93SThomas Gleixner } 56160e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 56172d3854a3SRusty Russell #endif /* CONFIG_SMP */ 56182d3854a3SRusty Russell 5619a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5620e7577c50SRusty Russell 5621a0a1a5fdSTejun Heo /** 5622a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5623a0a1a5fdSTejun Heo * 562458a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5625f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5626706026c2STejun Heo * pool->worklist. 5627a0a1a5fdSTejun Heo * 5628a0a1a5fdSTejun Heo * CONTEXT: 5629a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5630a0a1a5fdSTejun Heo */ 5631a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5632a0a1a5fdSTejun Heo { 563324b8a847STejun Heo struct workqueue_struct *wq; 563424b8a847STejun Heo struct pool_workqueue *pwq; 5635a0a1a5fdSTejun Heo 563668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5637a0a1a5fdSTejun Heo 56386183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5639a0a1a5fdSTejun Heo workqueue_freezing = true; 5640a0a1a5fdSTejun Heo 564124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5642a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5643699ce097STejun Heo for_each_pwq(pwq, wq) 5644699ce097STejun Heo pwq_adjust_max_active(pwq); 5645a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5646a1056305STejun Heo } 56475bcab335STejun Heo 564868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5649a0a1a5fdSTejun Heo } 5650a0a1a5fdSTejun Heo 5651a0a1a5fdSTejun Heo /** 565258a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5653a0a1a5fdSTejun Heo * 5654a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5655a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5656a0a1a5fdSTejun Heo * 5657a0a1a5fdSTejun Heo * CONTEXT: 565868e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5659a0a1a5fdSTejun Heo * 5660d185af30SYacine Belkadi * Return: 566158a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 566258a69cb4STejun Heo * is complete. 5663a0a1a5fdSTejun Heo */ 5664a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5665a0a1a5fdSTejun Heo { 5666a0a1a5fdSTejun Heo bool busy = false; 566724b8a847STejun Heo struct workqueue_struct *wq; 566824b8a847STejun Heo struct pool_workqueue *pwq; 5669a0a1a5fdSTejun Heo 567068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5671a0a1a5fdSTejun Heo 56726183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5673a0a1a5fdSTejun Heo 567424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 567524b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 567624b8a847STejun Heo continue; 5677a0a1a5fdSTejun Heo /* 5678a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5679a0a1a5fdSTejun Heo * to peek without lock. 5680a0a1a5fdSTejun Heo */ 568124acfb71SThomas Gleixner rcu_read_lock(); 568224b8a847STejun Heo for_each_pwq(pwq, wq) { 56836183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5684112202d9STejun Heo if (pwq->nr_active) { 5685a0a1a5fdSTejun Heo busy = true; 568624acfb71SThomas Gleixner rcu_read_unlock(); 5687a0a1a5fdSTejun Heo goto out_unlock; 5688a0a1a5fdSTejun Heo } 5689a0a1a5fdSTejun Heo } 569024acfb71SThomas Gleixner rcu_read_unlock(); 5691a0a1a5fdSTejun Heo } 5692a0a1a5fdSTejun Heo out_unlock: 569368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5694a0a1a5fdSTejun Heo return busy; 5695a0a1a5fdSTejun Heo } 5696a0a1a5fdSTejun Heo 5697a0a1a5fdSTejun Heo /** 5698a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5699a0a1a5fdSTejun Heo * 5700a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5701706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5702a0a1a5fdSTejun Heo * 5703a0a1a5fdSTejun Heo * CONTEXT: 5704a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5705a0a1a5fdSTejun Heo */ 5706a0a1a5fdSTejun Heo void thaw_workqueues(void) 5707a0a1a5fdSTejun Heo { 570824b8a847STejun Heo struct workqueue_struct *wq; 570924b8a847STejun Heo struct pool_workqueue *pwq; 5710a0a1a5fdSTejun Heo 571168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5712a0a1a5fdSTejun Heo 5713a0a1a5fdSTejun Heo if (!workqueue_freezing) 5714a0a1a5fdSTejun Heo goto out_unlock; 5715a0a1a5fdSTejun Heo 571674b414eaSLai Jiangshan workqueue_freezing = false; 571724b8a847STejun Heo 571824b8a847STejun Heo /* restore max_active and repopulate worklist */ 571924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5720a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5721699ce097STejun Heo for_each_pwq(pwq, wq) 5722699ce097STejun Heo pwq_adjust_max_active(pwq); 5723a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 572424b8a847STejun Heo } 572524b8a847STejun Heo 5726a0a1a5fdSTejun Heo out_unlock: 572768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5728a0a1a5fdSTejun Heo } 5729a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5730a0a1a5fdSTejun Heo 573199c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5732042f7df1SLai Jiangshan { 5733042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5734042f7df1SLai Jiangshan int ret = 0; 5735042f7df1SLai Jiangshan struct workqueue_struct *wq; 5736042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5737042f7df1SLai Jiangshan 5738042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5739042f7df1SLai Jiangshan 5740042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5741042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5742042f7df1SLai Jiangshan continue; 5743042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5744042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5745042f7df1SLai Jiangshan continue; 5746042f7df1SLai Jiangshan 574799c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5748042f7df1SLai Jiangshan if (!ctx) { 5749042f7df1SLai Jiangshan ret = -ENOMEM; 5750042f7df1SLai Jiangshan break; 5751042f7df1SLai Jiangshan } 5752042f7df1SLai Jiangshan 5753042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5754042f7df1SLai Jiangshan } 5755042f7df1SLai Jiangshan 5756042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5757042f7df1SLai Jiangshan if (!ret) 5758042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5759042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5760042f7df1SLai Jiangshan } 5761042f7df1SLai Jiangshan 576299c621efSLai Jiangshan if (!ret) { 576399c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 576499c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 576599c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 576699c621efSLai Jiangshan } 5767042f7df1SLai Jiangshan return ret; 5768042f7df1SLai Jiangshan } 5769042f7df1SLai Jiangshan 5770042f7df1SLai Jiangshan /** 5771042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5772042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5773042f7df1SLai Jiangshan * 5774042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5775042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5776042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5777042f7df1SLai Jiangshan * 577867dc8325SCai Huoqing * Return: 0 - Success 5779042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5780042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5781042f7df1SLai Jiangshan */ 5782042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5783042f7df1SLai Jiangshan { 5784042f7df1SLai Jiangshan int ret = -EINVAL; 5785042f7df1SLai Jiangshan 5786c98a9805STal Shorer /* 5787c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5788c98a9805STal Shorer * If the user wishes to include them, we allow that. 5789c98a9805STal Shorer */ 5790042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5791042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5792a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5793d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5794d25302e4SMenglong Dong ret = 0; 5795d25302e4SMenglong Dong goto out_unlock; 5796d25302e4SMenglong Dong } 5797d25302e4SMenglong Dong 579899c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5799042f7df1SLai Jiangshan 5800d25302e4SMenglong Dong out_unlock: 5801a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5802042f7df1SLai Jiangshan } 5803042f7df1SLai Jiangshan 5804042f7df1SLai Jiangshan return ret; 5805042f7df1SLai Jiangshan } 5806042f7df1SLai Jiangshan 58076ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 58086ba94429SFrederic Weisbecker /* 58096ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 58106ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 58116ba94429SFrederic Weisbecker * following attributes. 58126ba94429SFrederic Weisbecker * 58136ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 58146ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 58156ba94429SFrederic Weisbecker * 58166ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 58176ba94429SFrederic Weisbecker * 58186ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 58196ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 58206ba94429SFrederic Weisbecker */ 58216ba94429SFrederic Weisbecker struct wq_device { 58226ba94429SFrederic Weisbecker struct workqueue_struct *wq; 58236ba94429SFrederic Weisbecker struct device dev; 58246ba94429SFrederic Weisbecker }; 58256ba94429SFrederic Weisbecker 58266ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 58276ba94429SFrederic Weisbecker { 58286ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 58296ba94429SFrederic Weisbecker 58306ba94429SFrederic Weisbecker return wq_dev->wq; 58316ba94429SFrederic Weisbecker } 58326ba94429SFrederic Weisbecker 58336ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 58346ba94429SFrederic Weisbecker char *buf) 58356ba94429SFrederic Weisbecker { 58366ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58376ba94429SFrederic Weisbecker 58386ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 58396ba94429SFrederic Weisbecker } 58406ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 58416ba94429SFrederic Weisbecker 58426ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 58436ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58446ba94429SFrederic Weisbecker { 58456ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58466ba94429SFrederic Weisbecker 58476ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 58486ba94429SFrederic Weisbecker } 58496ba94429SFrederic Weisbecker 58506ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 58516ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 58526ba94429SFrederic Weisbecker size_t count) 58536ba94429SFrederic Weisbecker { 58546ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58556ba94429SFrederic Weisbecker int val; 58566ba94429SFrederic Weisbecker 58576ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 58586ba94429SFrederic Weisbecker return -EINVAL; 58596ba94429SFrederic Weisbecker 58606ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 58616ba94429SFrederic Weisbecker return count; 58626ba94429SFrederic Weisbecker } 58636ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 58646ba94429SFrederic Weisbecker 58656ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 58666ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 58676ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 58686ba94429SFrederic Weisbecker NULL, 58696ba94429SFrederic Weisbecker }; 58706ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 58716ba94429SFrederic Weisbecker 58726ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 58736ba94429SFrederic Weisbecker char *buf) 58746ba94429SFrederic Weisbecker { 58756ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58766ba94429SFrederic Weisbecker int written; 58776ba94429SFrederic Weisbecker 58786ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 58796ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 58806ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 58816ba94429SFrederic Weisbecker 58826ba94429SFrederic Weisbecker return written; 58836ba94429SFrederic Weisbecker } 58846ba94429SFrederic Weisbecker 58856ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 58866ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 58876ba94429SFrederic Weisbecker { 58886ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 58896ba94429SFrederic Weisbecker 5890899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5891899a94feSLai Jiangshan 5892be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 58936ba94429SFrederic Weisbecker if (!attrs) 58946ba94429SFrederic Weisbecker return NULL; 58956ba94429SFrederic Weisbecker 58966ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 58976ba94429SFrederic Weisbecker return attrs; 58986ba94429SFrederic Weisbecker } 58996ba94429SFrederic Weisbecker 59006ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 59016ba94429SFrederic Weisbecker const char *buf, size_t count) 59026ba94429SFrederic Weisbecker { 59036ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59046ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5905d4d3e257SLai Jiangshan int ret = -ENOMEM; 5906d4d3e257SLai Jiangshan 5907d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59086ba94429SFrederic Weisbecker 59096ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59106ba94429SFrederic Weisbecker if (!attrs) 5911d4d3e257SLai Jiangshan goto out_unlock; 59126ba94429SFrederic Weisbecker 59136ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 59146ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5915d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59166ba94429SFrederic Weisbecker else 59176ba94429SFrederic Weisbecker ret = -EINVAL; 59186ba94429SFrederic Weisbecker 5919d4d3e257SLai Jiangshan out_unlock: 5920d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59216ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59226ba94429SFrederic Weisbecker return ret ?: count; 59236ba94429SFrederic Weisbecker } 59246ba94429SFrederic Weisbecker 59256ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 59266ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 59276ba94429SFrederic Weisbecker { 59286ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59296ba94429SFrederic Weisbecker int written; 59306ba94429SFrederic Weisbecker 59316ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59326ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 59336ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 59346ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59356ba94429SFrederic Weisbecker return written; 59366ba94429SFrederic Weisbecker } 59376ba94429SFrederic Weisbecker 59386ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 59396ba94429SFrederic Weisbecker struct device_attribute *attr, 59406ba94429SFrederic Weisbecker const char *buf, size_t count) 59416ba94429SFrederic Weisbecker { 59426ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59436ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5944d4d3e257SLai Jiangshan int ret = -ENOMEM; 5945d4d3e257SLai Jiangshan 5946d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59476ba94429SFrederic Weisbecker 59486ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59496ba94429SFrederic Weisbecker if (!attrs) 5950d4d3e257SLai Jiangshan goto out_unlock; 59516ba94429SFrederic Weisbecker 59526ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 59536ba94429SFrederic Weisbecker if (!ret) 5954d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59556ba94429SFrederic Weisbecker 5956d4d3e257SLai Jiangshan out_unlock: 5957d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59586ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59596ba94429SFrederic Weisbecker return ret ?: count; 59606ba94429SFrederic Weisbecker } 59616ba94429SFrederic Weisbecker 59626ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 59636ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 59646ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 59656ba94429SFrederic Weisbecker __ATTR_NULL, 59666ba94429SFrederic Weisbecker }; 59676ba94429SFrederic Weisbecker 59686ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 59696ba94429SFrederic Weisbecker .name = "workqueue", 59706ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 59716ba94429SFrederic Weisbecker }; 59726ba94429SFrederic Weisbecker 5973b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5974b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5975b05a7928SFrederic Weisbecker { 5976b05a7928SFrederic Weisbecker int written; 5977b05a7928SFrederic Weisbecker 5978042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5979b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5980b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5981042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5982b05a7928SFrederic Weisbecker 5983b05a7928SFrederic Weisbecker return written; 5984b05a7928SFrederic Weisbecker } 5985b05a7928SFrederic Weisbecker 5986042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5987042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5988042f7df1SLai Jiangshan { 5989042f7df1SLai Jiangshan cpumask_var_t cpumask; 5990042f7df1SLai Jiangshan int ret; 5991042f7df1SLai Jiangshan 5992042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5993042f7df1SLai Jiangshan return -ENOMEM; 5994042f7df1SLai Jiangshan 5995042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5996042f7df1SLai Jiangshan if (!ret) 5997042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5998042f7df1SLai Jiangshan 5999042f7df1SLai Jiangshan free_cpumask_var(cpumask); 6000042f7df1SLai Jiangshan return ret ? ret : count; 6001042f7df1SLai Jiangshan } 6002042f7df1SLai Jiangshan 6003b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 6004042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6005042f7df1SLai Jiangshan wq_unbound_cpumask_store); 6006b05a7928SFrederic Weisbecker 60076ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 60086ba94429SFrederic Weisbecker { 6009686f6697SGreg Kroah-Hartman struct device *dev_root; 6010b05a7928SFrederic Weisbecker int err; 6011b05a7928SFrederic Weisbecker 6012b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 6013b05a7928SFrederic Weisbecker if (err) 6014b05a7928SFrederic Weisbecker return err; 6015b05a7928SFrederic Weisbecker 6016686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 6017686f6697SGreg Kroah-Hartman if (dev_root) { 6018686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6019686f6697SGreg Kroah-Hartman put_device(dev_root); 6020686f6697SGreg Kroah-Hartman } 6021686f6697SGreg Kroah-Hartman return err; 60226ba94429SFrederic Weisbecker } 60236ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 60246ba94429SFrederic Weisbecker 60256ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 60266ba94429SFrederic Weisbecker { 60276ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 60286ba94429SFrederic Weisbecker 60296ba94429SFrederic Weisbecker kfree(wq_dev); 60306ba94429SFrederic Weisbecker } 60316ba94429SFrederic Weisbecker 60326ba94429SFrederic Weisbecker /** 60336ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 60346ba94429SFrederic Weisbecker * @wq: the workqueue to register 60356ba94429SFrederic Weisbecker * 60366ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 60376ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 60386ba94429SFrederic Weisbecker * which is the preferred method. 60396ba94429SFrederic Weisbecker * 60406ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 60416ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 60426ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 60436ba94429SFrederic Weisbecker * attributes. 60446ba94429SFrederic Weisbecker * 60456ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 60466ba94429SFrederic Weisbecker */ 60476ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 60486ba94429SFrederic Weisbecker { 60496ba94429SFrederic Weisbecker struct wq_device *wq_dev; 60506ba94429SFrederic Weisbecker int ret; 60516ba94429SFrederic Weisbecker 60526ba94429SFrederic Weisbecker /* 6053402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 60546ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 60556ba94429SFrederic Weisbecker * workqueues. 60566ba94429SFrederic Weisbecker */ 60570a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 60586ba94429SFrederic Weisbecker return -EINVAL; 60596ba94429SFrederic Weisbecker 60606ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 60616ba94429SFrederic Weisbecker if (!wq_dev) 60626ba94429SFrederic Weisbecker return -ENOMEM; 60636ba94429SFrederic Weisbecker 60646ba94429SFrederic Weisbecker wq_dev->wq = wq; 60656ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 60666ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 606723217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 60686ba94429SFrederic Weisbecker 60696ba94429SFrederic Weisbecker /* 60706ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 60716ba94429SFrederic Weisbecker * everything is ready. 60726ba94429SFrederic Weisbecker */ 60736ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 60746ba94429SFrederic Weisbecker 60756ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 60766ba94429SFrederic Weisbecker if (ret) { 6077537f4146SArvind Yadav put_device(&wq_dev->dev); 60786ba94429SFrederic Weisbecker wq->wq_dev = NULL; 60796ba94429SFrederic Weisbecker return ret; 60806ba94429SFrederic Weisbecker } 60816ba94429SFrederic Weisbecker 60826ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 60836ba94429SFrederic Weisbecker struct device_attribute *attr; 60846ba94429SFrederic Weisbecker 60856ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 60866ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 60876ba94429SFrederic Weisbecker if (ret) { 60886ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 60896ba94429SFrederic Weisbecker wq->wq_dev = NULL; 60906ba94429SFrederic Weisbecker return ret; 60916ba94429SFrederic Weisbecker } 60926ba94429SFrederic Weisbecker } 60936ba94429SFrederic Weisbecker } 60946ba94429SFrederic Weisbecker 60956ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 60966ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 60976ba94429SFrederic Weisbecker return 0; 60986ba94429SFrederic Weisbecker } 60996ba94429SFrederic Weisbecker 61006ba94429SFrederic Weisbecker /** 61016ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 61026ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 61036ba94429SFrederic Weisbecker * 61046ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 61056ba94429SFrederic Weisbecker */ 61066ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 61076ba94429SFrederic Weisbecker { 61086ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 61096ba94429SFrederic Weisbecker 61106ba94429SFrederic Weisbecker if (!wq->wq_dev) 61116ba94429SFrederic Weisbecker return; 61126ba94429SFrederic Weisbecker 61136ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61146ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61156ba94429SFrederic Weisbecker } 61166ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 61176ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 61186ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 61196ba94429SFrederic Weisbecker 612082607adcSTejun Heo /* 612182607adcSTejun Heo * Workqueue watchdog. 612282607adcSTejun Heo * 612382607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 612482607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 612582607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 612682607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 612782607adcSTejun Heo * largely opaque. 612882607adcSTejun Heo * 612982607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 613082607adcSTejun Heo * state if some pools failed to make forward progress for a while where 613182607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 613282607adcSTejun Heo * 613382607adcSTejun Heo * This mechanism is controlled through the kernel parameter 613482607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 613582607adcSTejun Heo * corresponding sysfs parameter file. 613682607adcSTejun Heo */ 613782607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 613882607adcSTejun Heo 613982607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 61405cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 614182607adcSTejun Heo 614282607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 614382607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 614482607adcSTejun Heo 6145cd2440d6SPetr Mladek /* 6146cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6147cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6148cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6149cd2440d6SPetr Mladek * in all other situations. 6150cd2440d6SPetr Mladek */ 6151cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6152cd2440d6SPetr Mladek { 6153cd2440d6SPetr Mladek struct worker *worker; 6154cd2440d6SPetr Mladek unsigned long flags; 6155cd2440d6SPetr Mladek int bkt; 6156cd2440d6SPetr Mladek 6157cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6158cd2440d6SPetr Mladek 6159cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6160cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6161cd2440d6SPetr Mladek /* 6162cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6163cd2440d6SPetr Mladek * drivers that queue work while holding locks 6164cd2440d6SPetr Mladek * also taken in their write paths. 6165cd2440d6SPetr Mladek */ 6166cd2440d6SPetr Mladek printk_deferred_enter(); 6167cd2440d6SPetr Mladek 6168cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6169cd2440d6SPetr Mladek sched_show_task(worker->task); 6170cd2440d6SPetr Mladek 6171cd2440d6SPetr Mladek printk_deferred_exit(); 6172cd2440d6SPetr Mladek } 6173cd2440d6SPetr Mladek } 6174cd2440d6SPetr Mladek 6175cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6176cd2440d6SPetr Mladek } 6177cd2440d6SPetr Mladek 6178cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6179cd2440d6SPetr Mladek { 6180cd2440d6SPetr Mladek struct worker_pool *pool; 6181cd2440d6SPetr Mladek int pi; 6182cd2440d6SPetr Mladek 6183cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6184cd2440d6SPetr Mladek 6185cd2440d6SPetr Mladek rcu_read_lock(); 6186cd2440d6SPetr Mladek 6187cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6188cd2440d6SPetr Mladek if (pool->cpu_stall) 6189cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6190cd2440d6SPetr Mladek 6191cd2440d6SPetr Mladek } 6192cd2440d6SPetr Mladek 6193cd2440d6SPetr Mladek rcu_read_unlock(); 6194cd2440d6SPetr Mladek } 6195cd2440d6SPetr Mladek 619682607adcSTejun Heo static void wq_watchdog_reset_touched(void) 619782607adcSTejun Heo { 619882607adcSTejun Heo int cpu; 619982607adcSTejun Heo 620082607adcSTejun Heo wq_watchdog_touched = jiffies; 620182607adcSTejun Heo for_each_possible_cpu(cpu) 620282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 620382607adcSTejun Heo } 620482607adcSTejun Heo 62055cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 620682607adcSTejun Heo { 620782607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 620882607adcSTejun Heo bool lockup_detected = false; 6209cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6210940d71c6SSergey Senozhatsky unsigned long now = jiffies; 621182607adcSTejun Heo struct worker_pool *pool; 621282607adcSTejun Heo int pi; 621382607adcSTejun Heo 621482607adcSTejun Heo if (!thresh) 621582607adcSTejun Heo return; 621682607adcSTejun Heo 621782607adcSTejun Heo rcu_read_lock(); 621882607adcSTejun Heo 621982607adcSTejun Heo for_each_pool(pool, pi) { 622082607adcSTejun Heo unsigned long pool_ts, touched, ts; 622182607adcSTejun Heo 6222cd2440d6SPetr Mladek pool->cpu_stall = false; 622382607adcSTejun Heo if (list_empty(&pool->worklist)) 622482607adcSTejun Heo continue; 622582607adcSTejun Heo 6226940d71c6SSergey Senozhatsky /* 6227940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6228940d71c6SSergey Senozhatsky * the watchdog like a stall. 6229940d71c6SSergey Senozhatsky */ 6230940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6231940d71c6SSergey Senozhatsky 623282607adcSTejun Heo /* get the latest of pool and touched timestamps */ 623389e28ce6SWang Qing if (pool->cpu >= 0) 623489e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 623589e28ce6SWang Qing else 623682607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 623789e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 623882607adcSTejun Heo 623982607adcSTejun Heo if (time_after(pool_ts, touched)) 624082607adcSTejun Heo ts = pool_ts; 624182607adcSTejun Heo else 624282607adcSTejun Heo ts = touched; 624382607adcSTejun Heo 624482607adcSTejun Heo /* did we stall? */ 6245940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 624682607adcSTejun Heo lockup_detected = true; 6247cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6248cd2440d6SPetr Mladek pool->cpu_stall = true; 6249cd2440d6SPetr Mladek cpu_pool_stall = true; 6250cd2440d6SPetr Mladek } 625182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 625282607adcSTejun Heo pr_cont_pool_info(pool); 625382607adcSTejun Heo pr_cont(" stuck for %us!\n", 6254940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 625582607adcSTejun Heo } 6256cd2440d6SPetr Mladek 6257cd2440d6SPetr Mladek 625882607adcSTejun Heo } 625982607adcSTejun Heo 626082607adcSTejun Heo rcu_read_unlock(); 626182607adcSTejun Heo 626282607adcSTejun Heo if (lockup_detected) 626355df0933SImran Khan show_all_workqueues(); 626482607adcSTejun Heo 6265cd2440d6SPetr Mladek if (cpu_pool_stall) 6266cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6267cd2440d6SPetr Mladek 626882607adcSTejun Heo wq_watchdog_reset_touched(); 626982607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 627082607adcSTejun Heo } 627182607adcSTejun Heo 6272cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 627382607adcSTejun Heo { 627482607adcSTejun Heo if (cpu >= 0) 627582607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 627689e28ce6SWang Qing 627782607adcSTejun Heo wq_watchdog_touched = jiffies; 627882607adcSTejun Heo } 627982607adcSTejun Heo 628082607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 628182607adcSTejun Heo { 628282607adcSTejun Heo wq_watchdog_thresh = 0; 628382607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 628482607adcSTejun Heo 628582607adcSTejun Heo if (thresh) { 628682607adcSTejun Heo wq_watchdog_thresh = thresh; 628782607adcSTejun Heo wq_watchdog_reset_touched(); 628882607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 628982607adcSTejun Heo } 629082607adcSTejun Heo } 629182607adcSTejun Heo 629282607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 629382607adcSTejun Heo const struct kernel_param *kp) 629482607adcSTejun Heo { 629582607adcSTejun Heo unsigned long thresh; 629682607adcSTejun Heo int ret; 629782607adcSTejun Heo 629882607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 629982607adcSTejun Heo if (ret) 630082607adcSTejun Heo return ret; 630182607adcSTejun Heo 630282607adcSTejun Heo if (system_wq) 630382607adcSTejun Heo wq_watchdog_set_thresh(thresh); 630482607adcSTejun Heo else 630582607adcSTejun Heo wq_watchdog_thresh = thresh; 630682607adcSTejun Heo 630782607adcSTejun Heo return 0; 630882607adcSTejun Heo } 630982607adcSTejun Heo 631082607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 631182607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 631282607adcSTejun Heo .get = param_get_ulong, 631382607adcSTejun Heo }; 631482607adcSTejun Heo 631582607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 631682607adcSTejun Heo 0644); 631782607adcSTejun Heo 631882607adcSTejun Heo static void wq_watchdog_init(void) 631982607adcSTejun Heo { 63205cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 632182607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 632282607adcSTejun Heo } 632382607adcSTejun Heo 632482607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 632582607adcSTejun Heo 632682607adcSTejun Heo static inline void wq_watchdog_init(void) { } 632782607adcSTejun Heo 632882607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 632982607adcSTejun Heo 6330bce90380STejun Heo static void __init wq_numa_init(void) 6331bce90380STejun Heo { 6332bce90380STejun Heo cpumask_var_t *tbl; 6333bce90380STejun Heo int node, cpu; 6334bce90380STejun Heo 6335bce90380STejun Heo if (num_possible_nodes() <= 1) 6336bce90380STejun Heo return; 6337bce90380STejun Heo 6338f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6339f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6340f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6341f728c4a9SZhen Lei return; 6342f728c4a9SZhen Lei } 6343f728c4a9SZhen Lei } 6344f728c4a9SZhen Lei 6345be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 63464c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 63474c16bd32STejun Heo 6348bce90380STejun Heo /* 6349bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6350bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6351bce90380STejun Heo * fully initialized by now. 6352bce90380STejun Heo */ 63536396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6354bce90380STejun Heo BUG_ON(!tbl); 6355bce90380STejun Heo 6356bce90380STejun Heo for_each_node(node) 63575a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 63581be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6359bce90380STejun Heo 6360bce90380STejun Heo for_each_possible_cpu(cpu) { 6361bce90380STejun Heo node = cpu_to_node(cpu); 6362bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6363bce90380STejun Heo } 6364bce90380STejun Heo 6365bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6366bce90380STejun Heo wq_numa_enabled = true; 6367bce90380STejun Heo } 6368bce90380STejun Heo 63693347fa09STejun Heo /** 63703347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 63713347fa09STejun Heo * 63723347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 63733347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 63743347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 63753347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 63763347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 63773347fa09STejun Heo * created and scheduled right before early initcalls. 63783347fa09STejun Heo */ 63792333e829SYu Chen void __init workqueue_init_early(void) 63801da177e4SLinus Torvalds { 63817a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 63827a4e344cSTejun Heo int i, cpu; 6383c34056a3STejun Heo 638410cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6385e904e6c2STejun Heo 6386b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 638704d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 638804d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6389b05a7928SFrederic Weisbecker 6390ace3c549Stiozhang if (!cpumask_empty(&wq_cmdline_cpumask)) 6391ace3c549Stiozhang cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask); 6392ace3c549Stiozhang 6393e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6394e904e6c2STejun Heo 6395706026c2STejun Heo /* initialize CPU pools */ 639629c91e99STejun Heo for_each_possible_cpu(cpu) { 63974ce62e9eSTejun Heo struct worker_pool *pool; 63988b03ae3cSTejun Heo 63997a4e344cSTejun Heo i = 0; 6400f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 64017a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6402ec22ca5eSTejun Heo pool->cpu = cpu; 64037a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 64047a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6405f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 64067a4e344cSTejun Heo 64079daf9e67STejun Heo /* alloc pool ID */ 640868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 64099daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 641068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 64114ce62e9eSTejun Heo } 64128b03ae3cSTejun Heo } 64138b03ae3cSTejun Heo 64148a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 641529c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 641629c91e99STejun Heo struct workqueue_attrs *attrs; 641729c91e99STejun Heo 6418be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 641929c91e99STejun Heo attrs->nice = std_nice[i]; 642029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 64218a2b7538STejun Heo 64228a2b7538STejun Heo /* 64238a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 64248a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 64258a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 64268a2b7538STejun Heo */ 6427be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 64288a2b7538STejun Heo attrs->nice = std_nice[i]; 64298a2b7538STejun Heo attrs->no_numa = true; 64308a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 643129c91e99STejun Heo } 643229c91e99STejun Heo 6433d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 64341aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6435d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6436f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6437f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 643824d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 643924d51addSTejun Heo WQ_FREEZABLE, 0); 64400668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 64410668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 64420668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 64430668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 64440668106cSViresh Kumar 0); 64451aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 64460668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 64470668106cSViresh Kumar !system_power_efficient_wq || 64480668106cSViresh Kumar !system_freezable_power_efficient_wq); 64493347fa09STejun Heo } 64503347fa09STejun Heo 6451aa6fde93STejun Heo static void __init wq_cpu_intensive_thresh_init(void) 6452aa6fde93STejun Heo { 6453aa6fde93STejun Heo unsigned long thresh; 6454aa6fde93STejun Heo unsigned long bogo; 6455aa6fde93STejun Heo 6456aa6fde93STejun Heo /* if the user set it to a specific value, keep it */ 6457aa6fde93STejun Heo if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6458aa6fde93STejun Heo return; 6459aa6fde93STejun Heo 6460967b494eSTejun Heo pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 6461967b494eSTejun Heo BUG_ON(IS_ERR(pwq_release_worker)); 6462967b494eSTejun Heo 6463aa6fde93STejun Heo /* 6464aa6fde93STejun Heo * The default of 10ms is derived from the fact that most modern (as of 6465aa6fde93STejun Heo * 2023) processors can do a lot in 10ms and that it's just below what 6466aa6fde93STejun Heo * most consider human-perceivable. However, the kernel also runs on a 6467aa6fde93STejun Heo * lot slower CPUs including microcontrollers where the threshold is way 6468aa6fde93STejun Heo * too low. 6469aa6fde93STejun Heo * 6470aa6fde93STejun Heo * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6471aa6fde93STejun Heo * This is by no means accurate but it doesn't have to be. The mechanism 6472aa6fde93STejun Heo * is still useful even when the threshold is fully scaled up. Also, as 6473aa6fde93STejun Heo * the reports would usually be applicable to everyone, some machines 6474aa6fde93STejun Heo * operating on longer thresholds won't significantly diminish their 6475aa6fde93STejun Heo * usefulness. 6476aa6fde93STejun Heo */ 6477aa6fde93STejun Heo thresh = 10 * USEC_PER_MSEC; 6478aa6fde93STejun Heo 6479aa6fde93STejun Heo /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6480aa6fde93STejun Heo bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6481aa6fde93STejun Heo if (bogo < 4000) 6482aa6fde93STejun Heo thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6483aa6fde93STejun Heo 6484aa6fde93STejun Heo pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6485aa6fde93STejun Heo loops_per_jiffy, bogo, thresh); 6486aa6fde93STejun Heo 6487aa6fde93STejun Heo wq_cpu_intensive_thresh_us = thresh; 6488aa6fde93STejun Heo } 6489aa6fde93STejun Heo 64903347fa09STejun Heo /** 64913347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 64923347fa09STejun Heo * 64933347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 64943347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 64953347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 64963347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 64973347fa09STejun Heo * with the initial workers and enable future kworker creations. 64983347fa09STejun Heo */ 64992333e829SYu Chen void __init workqueue_init(void) 65003347fa09STejun Heo { 65012186d9f9STejun Heo struct workqueue_struct *wq; 65023347fa09STejun Heo struct worker_pool *pool; 65033347fa09STejun Heo int cpu, bkt; 65043347fa09STejun Heo 6505aa6fde93STejun Heo wq_cpu_intensive_thresh_init(); 6506aa6fde93STejun Heo 65072186d9f9STejun Heo /* 65082186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 65092186d9f9STejun Heo * CPU to node mapping may not be available that early on some 65102186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 65112186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 65122186d9f9STejun Heo * affinity, fix them up. 651340c17f75STejun Heo * 651440c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 65152186d9f9STejun Heo */ 65162186d9f9STejun Heo wq_numa_init(); 65172186d9f9STejun Heo 65182186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 65192186d9f9STejun Heo 65202186d9f9STejun Heo for_each_possible_cpu(cpu) { 65212186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65222186d9f9STejun Heo pool->node = cpu_to_node(cpu); 65232186d9f9STejun Heo } 65242186d9f9STejun Heo } 65252186d9f9STejun Heo 652640c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 6527*4cbfd3deSTejun Heo wq_update_unbound_numa(wq, smp_processor_id(), smp_processor_id(), 6528*4cbfd3deSTejun Heo true); 652940c17f75STejun Heo WARN(init_rescuer(wq), 653040c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 653140c17f75STejun Heo wq->name); 653240c17f75STejun Heo } 65332186d9f9STejun Heo 65342186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 65352186d9f9STejun Heo 65363347fa09STejun Heo /* create the initial workers */ 65373347fa09STejun Heo for_each_online_cpu(cpu) { 65383347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65393347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 65403347fa09STejun Heo BUG_ON(!create_worker(pool)); 65413347fa09STejun Heo } 65423347fa09STejun Heo } 65433347fa09STejun Heo 65443347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 65453347fa09STejun Heo BUG_ON(!create_worker(pool)); 65463347fa09STejun Heo 65473347fa09STejun Heo wq_online = true; 654882607adcSTejun Heo wq_watchdog_init(); 65491da177e4SLinus Torvalds } 6550c4f135d6STetsuo Handa 655120bdedafSTetsuo Handa void __warn_flushing_systemwide_wq(void) 655220bdedafSTetsuo Handa { 655320bdedafSTetsuo Handa pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 655420bdedafSTetsuo Handa dump_stack(); 655520bdedafSTetsuo Handa } 6556c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6557ace3c549Stiozhang 6558ace3c549Stiozhang static int __init workqueue_unbound_cpus_setup(char *str) 6559ace3c549Stiozhang { 6560ace3c549Stiozhang if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6561ace3c549Stiozhang cpumask_clear(&wq_cmdline_cpumask); 6562ace3c549Stiozhang pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6563ace3c549Stiozhang } 6564ace3c549Stiozhang 6565ace3c549Stiozhang return 1; 6566ace3c549Stiozhang } 6567ace3c549Stiozhang __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6568