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 */ 324636b927eSTejun Heo struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */ 3251da177e4SLinus Torvalds }; 3261da177e4SLinus Torvalds 327e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 328e904e6c2STejun Heo 329fef59c9cSTejun Heo static cpumask_var_t *wq_pod_cpus; /* possible CPUs of each node */ 330bce90380STejun Heo 331616db877STejun Heo /* 332616db877STejun Heo * Per-cpu work items which run for longer than the following threshold are 333616db877STejun Heo * automatically considered CPU intensive and excluded from concurrency 334616db877STejun Heo * management to prevent them from noticeably delaying other per-cpu work items. 335aa6fde93STejun Heo * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 336aa6fde93STejun Heo * The actual value is initialized in wq_cpu_intensive_thresh_init(). 337616db877STejun Heo */ 338aa6fde93STejun Heo static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 339616db877STejun Heo module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 340616db877STejun Heo 341cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 342552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 343cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 344cee22a15SViresh Kumar 345863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3463347fa09STejun Heo 347fef59c9cSTejun Heo static bool wq_pod_enabled; /* unbound CPU pod affinity enabled */ 348bce90380STejun Heo 349fef59c9cSTejun Heo /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 350fef59c9cSTejun Heo static struct workqueue_attrs *wq_update_pod_attrs_buf; 3514c16bd32STejun Heo 35268e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3531258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 354a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 355d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 356d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3575bcab335STejun Heo 358e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 35968e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3607d19c5ceSTejun Heo 36199c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 362ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 363ef557180SMike Galbraith 364ace3c549Stiozhang /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 365ace3c549Stiozhang static struct cpumask wq_cmdline_cpumask __initdata; 366ace3c549Stiozhang 367ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 368ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 369b05a7928SFrederic Weisbecker 370f303fccbSTejun Heo /* 371f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 372f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 373f303fccbSTejun Heo * to uncover usages which depend on it. 374f303fccbSTejun Heo */ 375f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 376f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 377f303fccbSTejun Heo #else 378f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 379f303fccbSTejun Heo #endif 380f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 381f303fccbSTejun Heo 3827d19c5ceSTejun Heo /* the per-cpu worker pools */ 38325528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3847d19c5ceSTejun Heo 38568e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3867d19c5ceSTejun Heo 38768e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 38829c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 38929c91e99STejun Heo 390c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 39129c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 39229c91e99STejun Heo 3938a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3948a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3958a2b7538STejun Heo 396967b494eSTejun Heo /* 397967b494eSTejun Heo * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 398967b494eSTejun Heo * process context while holding a pool lock. Bounce to a dedicated kthread 399967b494eSTejun Heo * worker to avoid A-A deadlocks. 400967b494eSTejun Heo */ 401967b494eSTejun Heo static struct kthread_worker *pwq_release_worker; 402967b494eSTejun Heo 403d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 404ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 405044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 4061aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 407044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 408d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 409044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 410f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 411044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 41224d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 4130668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 4140668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 4150668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 4160668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 417d320c038STejun Heo 4187d19c5ceSTejun Heo static int worker_thread(void *__worker); 4196ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 420c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 42155df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4227d19c5ceSTejun Heo 42397bd2347STejun Heo #define CREATE_TRACE_POINTS 42497bd2347STejun Heo #include <trace/events/workqueue.h> 42597bd2347STejun Heo 42668e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 42724acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 428f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 42924acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4305bcab335STejun Heo 4315b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 43224acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 433f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 434f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 43524acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4365b95e1afSLai Jiangshan 437f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 438f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 439f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4407a62c2c8STejun Heo (pool)++) 4414ce62e9eSTejun Heo 44249e3cf44STejun Heo /** 44317116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 44417116969STejun Heo * @pool: iteration cursor 445611c92a0STejun Heo * @pi: integer used for iteration 446fa1b54e6STejun Heo * 44724acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 44868e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 44968e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 450fa1b54e6STejun Heo * 451fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 452fa1b54e6STejun Heo * ignored. 45317116969STejun Heo */ 454611c92a0STejun Heo #define for_each_pool(pool, pi) \ 455611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 45668e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 457fa1b54e6STejun Heo else 45817116969STejun Heo 45917116969STejun Heo /** 460822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 461822d8405STejun Heo * @worker: iteration cursor 462822d8405STejun Heo * @pool: worker_pool to iterate workers of 463822d8405STejun Heo * 4641258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 465822d8405STejun Heo * 466822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 467822d8405STejun Heo * ignored. 468822d8405STejun Heo */ 469da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 470da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4711258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 472822d8405STejun Heo else 473822d8405STejun Heo 474822d8405STejun Heo /** 47549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 47649e3cf44STejun Heo * @pwq: iteration cursor 47749e3cf44STejun Heo * @wq: the target workqueue 47876af4d93STejun Heo * 47924acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 480794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 481794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 48276af4d93STejun Heo * 48376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 48476af4d93STejun Heo * ignored. 48549e3cf44STejun Heo */ 48649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 48749e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4885a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 489f3421797STejun Heo 490dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 491dc186ad7SThomas Gleixner 492f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 493dc186ad7SThomas Gleixner 49499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 49599777288SStanislaw Gruszka { 49699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 49799777288SStanislaw Gruszka } 49899777288SStanislaw Gruszka 499b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 500b9fdac7fSDu, Changbin { 501b9fdac7fSDu, Changbin struct work_struct *work = addr; 502b9fdac7fSDu, Changbin 503b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 504b9fdac7fSDu, Changbin } 505b9fdac7fSDu, Changbin 506dc186ad7SThomas Gleixner /* 507dc186ad7SThomas Gleixner * fixup_init is called when: 508dc186ad7SThomas Gleixner * - an active object is initialized 509dc186ad7SThomas Gleixner */ 51002a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 511dc186ad7SThomas Gleixner { 512dc186ad7SThomas Gleixner struct work_struct *work = addr; 513dc186ad7SThomas Gleixner 514dc186ad7SThomas Gleixner switch (state) { 515dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 516dc186ad7SThomas Gleixner cancel_work_sync(work); 517dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 51802a982a6SDu, Changbin return true; 519dc186ad7SThomas Gleixner default: 52002a982a6SDu, Changbin return false; 521dc186ad7SThomas Gleixner } 522dc186ad7SThomas Gleixner } 523dc186ad7SThomas Gleixner 524dc186ad7SThomas Gleixner /* 525dc186ad7SThomas Gleixner * fixup_free is called when: 526dc186ad7SThomas Gleixner * - an active object is freed 527dc186ad7SThomas Gleixner */ 52802a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 529dc186ad7SThomas Gleixner { 530dc186ad7SThomas Gleixner struct work_struct *work = addr; 531dc186ad7SThomas Gleixner 532dc186ad7SThomas Gleixner switch (state) { 533dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 534dc186ad7SThomas Gleixner cancel_work_sync(work); 535dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 53602a982a6SDu, Changbin return true; 537dc186ad7SThomas Gleixner default: 53802a982a6SDu, Changbin return false; 539dc186ad7SThomas Gleixner } 540dc186ad7SThomas Gleixner } 541dc186ad7SThomas Gleixner 542f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 543dc186ad7SThomas Gleixner .name = "work_struct", 54499777288SStanislaw Gruszka .debug_hint = work_debug_hint, 545b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 546dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 547dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 548dc186ad7SThomas Gleixner }; 549dc186ad7SThomas Gleixner 550dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 551dc186ad7SThomas Gleixner { 552dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 553dc186ad7SThomas Gleixner } 554dc186ad7SThomas Gleixner 555dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 556dc186ad7SThomas Gleixner { 557dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 558dc186ad7SThomas Gleixner } 559dc186ad7SThomas Gleixner 560dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 561dc186ad7SThomas Gleixner { 562dc186ad7SThomas Gleixner if (onstack) 563dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 564dc186ad7SThomas Gleixner else 565dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 566dc186ad7SThomas Gleixner } 567dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 568dc186ad7SThomas Gleixner 569dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 570dc186ad7SThomas Gleixner { 571dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 572dc186ad7SThomas Gleixner } 573dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 574dc186ad7SThomas Gleixner 575ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 576ea2e64f2SThomas Gleixner { 577ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 578ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 579ea2e64f2SThomas Gleixner } 580ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 581ea2e64f2SThomas Gleixner 582dc186ad7SThomas Gleixner #else 583dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 584dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 585dc186ad7SThomas Gleixner #endif 586dc186ad7SThomas Gleixner 5874e8b22bdSLi Bin /** 58867dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5894e8b22bdSLi Bin * @pool: the pool pointer of interest 5904e8b22bdSLi Bin * 5914e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5924e8b22bdSLi Bin * successfully, -errno on failure. 5934e8b22bdSLi Bin */ 5949daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5959daf9e67STejun Heo { 5969daf9e67STejun Heo int ret; 5979daf9e67STejun Heo 59868e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5995bcab335STejun Heo 6004e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 6014e8b22bdSLi Bin GFP_KERNEL); 602229641a6STejun Heo if (ret >= 0) { 603e68035fbSTejun Heo pool->id = ret; 604229641a6STejun Heo return 0; 605229641a6STejun Heo } 6069daf9e67STejun Heo return ret; 6079daf9e67STejun Heo } 6089daf9e67STejun Heo 60973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 61073f53c4aSTejun Heo { 61173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 61273f53c4aSTejun Heo } 61373f53c4aSTejun Heo 614c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 61573f53c4aSTejun Heo { 616c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 61773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 61873f53c4aSTejun Heo } 61973f53c4aSTejun Heo 62073f53c4aSTejun Heo static int work_next_color(int color) 62173f53c4aSTejun Heo { 62273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 623a848e3b6SOleg Nesterov } 624a848e3b6SOleg Nesterov 6254594bf15SDavid Howells /* 626112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 627112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6287c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6297a22ad75STejun Heo * 630112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 631112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 632bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 633bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6347a22ad75STejun Heo * 635112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6367c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 637112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6387c3eed5cSTejun Heo * available only while the work item is queued. 639bbb68dfaSTejun Heo * 640bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 641bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 642bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 643bbb68dfaSTejun Heo * try to steal the PENDING bit. 6444594bf15SDavid Howells */ 6457a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6467a22ad75STejun Heo unsigned long flags) 6477a22ad75STejun Heo { 6486183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6497a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6507a22ad75STejun Heo } 6517a22ad75STejun Heo 652112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6534690c4abSTejun Heo unsigned long extra_flags) 654365970a1SDavid Howells { 655112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 656112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 657365970a1SDavid Howells } 658365970a1SDavid Howells 6594468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6604468a00fSLai Jiangshan int pool_id) 6614468a00fSLai Jiangshan { 6624468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6634468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6644468a00fSLai Jiangshan } 6654468a00fSLai Jiangshan 6667c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6677c3eed5cSTejun Heo int pool_id) 6684d707b9fSOleg Nesterov { 66923657bb1STejun Heo /* 67023657bb1STejun Heo * The following wmb is paired with the implied mb in 67123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 67223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 67323657bb1STejun Heo * owner. 67423657bb1STejun Heo */ 67523657bb1STejun Heo smp_wmb(); 6767c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 677346c09f8SRoman Pen /* 678346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 679346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 680346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6818bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 682346c09f8SRoman Pen * the same @work. E.g. consider this case: 683346c09f8SRoman Pen * 684346c09f8SRoman Pen * CPU#0 CPU#1 685346c09f8SRoman Pen * ---------------------------- -------------------------------- 686346c09f8SRoman Pen * 687346c09f8SRoman Pen * 1 STORE event_indicated 688346c09f8SRoman Pen * 2 queue_work_on() { 689346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 690346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 691346c09f8SRoman Pen * 5 set_work_data() # clear bit 692346c09f8SRoman Pen * 6 smp_mb() 693346c09f8SRoman Pen * 7 work->current_func() { 694346c09f8SRoman Pen * 8 LOAD event_indicated 695346c09f8SRoman Pen * } 696346c09f8SRoman Pen * 697346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 698346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 699346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 700346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 701346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 702346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 703346c09f8SRoman Pen * before actual STORE. 704346c09f8SRoman Pen */ 705346c09f8SRoman Pen smp_mb(); 7064d707b9fSOleg Nesterov } 7074d707b9fSOleg Nesterov 7087a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 709365970a1SDavid Howells { 7107c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7117c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7127a22ad75STejun Heo } 7137a22ad75STejun Heo 714afa4bb77SLinus Torvalds static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 715afa4bb77SLinus Torvalds { 716afa4bb77SLinus Torvalds return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 717afa4bb77SLinus Torvalds } 718afa4bb77SLinus Torvalds 719112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7207a22ad75STejun Heo { 721e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7227a22ad75STejun Heo 723112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 724afa4bb77SLinus Torvalds return work_struct_pwq(data); 725e120153dSTejun Heo else 726e120153dSTejun Heo return NULL; 7277a22ad75STejun Heo } 7287a22ad75STejun Heo 7297c3eed5cSTejun Heo /** 7307c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7317c3eed5cSTejun Heo * @work: the work item of interest 7327c3eed5cSTejun Heo * 73368e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 73424acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 73524acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 736fa1b54e6STejun Heo * 737fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 738fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 739fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 740fa1b54e6STejun Heo * returned pool is and stays online. 741d185af30SYacine Belkadi * 742d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7437c3eed5cSTejun Heo */ 7447c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7457a22ad75STejun Heo { 746e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7477c3eed5cSTejun Heo int pool_id; 7487a22ad75STejun Heo 74968e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 750fa1b54e6STejun Heo 751112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 752afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool; 7537a22ad75STejun Heo 7547c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7557c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7567a22ad75STejun Heo return NULL; 7577a22ad75STejun Heo 758fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7597c3eed5cSTejun Heo } 7607c3eed5cSTejun Heo 7617c3eed5cSTejun Heo /** 7627c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7637c3eed5cSTejun Heo * @work: the work item of interest 7647c3eed5cSTejun Heo * 765d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7667c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7677c3eed5cSTejun Heo */ 7687c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7697c3eed5cSTejun Heo { 77054d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7717c3eed5cSTejun Heo 772112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 773afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool->id; 77454d5b7d0SLai Jiangshan 77554d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7767c3eed5cSTejun Heo } 7777c3eed5cSTejun Heo 778bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 779bbb68dfaSTejun Heo { 7807c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 781bbb68dfaSTejun Heo 7827c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7837c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 784bbb68dfaSTejun Heo } 785bbb68dfaSTejun Heo 786bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 787bbb68dfaSTejun Heo { 788bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 789bbb68dfaSTejun Heo 790112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 791bbb68dfaSTejun Heo } 792bbb68dfaSTejun Heo 793e22bee78STejun Heo /* 7943270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7953270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 796d565ed63STejun Heo * they're being called with pool->lock held. 797e22bee78STejun Heo */ 798e22bee78STejun Heo 79963d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 800649027d7STejun Heo { 801bc35f7efSLai Jiangshan return !pool->nr_running; 802649027d7STejun Heo } 803649027d7STejun Heo 804e22bee78STejun Heo /* 805e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 806e22bee78STejun Heo * running workers. 807974271c4STejun Heo * 808974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 809706026c2STejun Heo * function will always return %true for unbound pools as long as the 810974271c4STejun Heo * worklist isn't empty. 811e22bee78STejun Heo */ 81263d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 813e22bee78STejun Heo { 81463d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 815e22bee78STejun Heo } 816e22bee78STejun Heo 817e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 81863d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 819e22bee78STejun Heo { 82063d95a91STejun Heo return pool->nr_idle; 821e22bee78STejun Heo } 822e22bee78STejun Heo 823e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 82463d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 825e22bee78STejun Heo { 826bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 827e22bee78STejun Heo } 828e22bee78STejun Heo 829e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 83063d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 831e22bee78STejun Heo { 83263d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 833e22bee78STejun Heo } 834e22bee78STejun Heo 835e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 83663d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 837e22bee78STejun Heo { 838692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 83963d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 84063d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 841e22bee78STejun Heo 842e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 843e22bee78STejun Heo } 844e22bee78STejun Heo 8454690c4abSTejun Heo /** 846e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 847cb444766STejun Heo * @worker: self 848d302f017STejun Heo * @flags: flags to set 849d302f017STejun Heo * 850228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 851d302f017STejun Heo */ 852228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 853d302f017STejun Heo { 854bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 855e22bee78STejun Heo 856bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 857cb444766STejun Heo 858228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 859e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 860e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 861bc35f7efSLai Jiangshan pool->nr_running--; 862e22bee78STejun Heo } 863e22bee78STejun Heo 864d302f017STejun Heo worker->flags |= flags; 865d302f017STejun Heo } 866d302f017STejun Heo 867d302f017STejun Heo /** 868e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 869cb444766STejun Heo * @worker: self 870d302f017STejun Heo * @flags: flags to clear 871d302f017STejun Heo * 872e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 873d302f017STejun Heo */ 874d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 875d302f017STejun Heo { 87663d95a91STejun Heo struct worker_pool *pool = worker->pool; 877e22bee78STejun Heo unsigned int oflags = worker->flags; 878e22bee78STejun Heo 879bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 880cb444766STejun Heo 881d302f017STejun Heo worker->flags &= ~flags; 882e22bee78STejun Heo 88342c025f3STejun Heo /* 88442c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 88542c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 88642c025f3STejun Heo * of multiple flags, not a single flag. 88742c025f3STejun Heo */ 888e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 889e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 890bc35f7efSLai Jiangshan pool->nr_running++; 891d302f017STejun Heo } 892d302f017STejun Heo 893797e8345STejun Heo /* Return the first idle worker. Called with pool->lock held. */ 894797e8345STejun Heo static struct worker *first_idle_worker(struct worker_pool *pool) 895797e8345STejun Heo { 896797e8345STejun Heo if (unlikely(list_empty(&pool->idle_list))) 897797e8345STejun Heo return NULL; 898797e8345STejun Heo 899797e8345STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 900797e8345STejun Heo } 901797e8345STejun Heo 902797e8345STejun Heo /** 903797e8345STejun Heo * worker_enter_idle - enter idle state 904797e8345STejun Heo * @worker: worker which is entering idle state 905797e8345STejun Heo * 906797e8345STejun Heo * @worker is entering idle state. Update stats and idle timer if 907797e8345STejun Heo * necessary. 908797e8345STejun Heo * 909797e8345STejun Heo * LOCKING: 910797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 911797e8345STejun Heo */ 912797e8345STejun Heo static void worker_enter_idle(struct worker *worker) 913797e8345STejun Heo { 914797e8345STejun Heo struct worker_pool *pool = worker->pool; 915797e8345STejun Heo 916797e8345STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 917797e8345STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 918797e8345STejun Heo (worker->hentry.next || worker->hentry.pprev))) 919797e8345STejun Heo return; 920797e8345STejun Heo 921797e8345STejun Heo /* can't use worker_set_flags(), also called from create_worker() */ 922797e8345STejun Heo worker->flags |= WORKER_IDLE; 923797e8345STejun Heo pool->nr_idle++; 924797e8345STejun Heo worker->last_active = jiffies; 925797e8345STejun Heo 926797e8345STejun Heo /* idle_list is LIFO */ 927797e8345STejun Heo list_add(&worker->entry, &pool->idle_list); 928797e8345STejun Heo 929797e8345STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 930797e8345STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 931797e8345STejun Heo 932797e8345STejun Heo /* Sanity check nr_running. */ 933797e8345STejun Heo WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 934797e8345STejun Heo } 935797e8345STejun Heo 936797e8345STejun Heo /** 937797e8345STejun Heo * worker_leave_idle - leave idle state 938797e8345STejun Heo * @worker: worker which is leaving idle state 939797e8345STejun Heo * 940797e8345STejun Heo * @worker is leaving idle state. Update stats. 941797e8345STejun Heo * 942797e8345STejun Heo * LOCKING: 943797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 944797e8345STejun Heo */ 945797e8345STejun Heo static void worker_leave_idle(struct worker *worker) 946797e8345STejun Heo { 947797e8345STejun Heo struct worker_pool *pool = worker->pool; 948797e8345STejun Heo 949797e8345STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 950797e8345STejun Heo return; 951797e8345STejun Heo worker_clr_flags(worker, WORKER_IDLE); 952797e8345STejun Heo pool->nr_idle--; 953797e8345STejun Heo list_del_init(&worker->entry); 954797e8345STejun Heo } 955797e8345STejun Heo 956797e8345STejun Heo /** 957797e8345STejun Heo * find_worker_executing_work - find worker which is executing a work 958797e8345STejun Heo * @pool: pool of interest 959797e8345STejun Heo * @work: work to find worker for 960797e8345STejun Heo * 961797e8345STejun Heo * Find a worker which is executing @work on @pool by searching 962797e8345STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 963797e8345STejun Heo * to match, its current execution should match the address of @work and 964797e8345STejun Heo * its work function. This is to avoid unwanted dependency between 965797e8345STejun Heo * unrelated work executions through a work item being recycled while still 966797e8345STejun Heo * being executed. 967797e8345STejun Heo * 968797e8345STejun Heo * This is a bit tricky. A work item may be freed once its execution 969797e8345STejun Heo * starts and nothing prevents the freed area from being recycled for 970797e8345STejun Heo * another work item. If the same work item address ends up being reused 971797e8345STejun Heo * before the original execution finishes, workqueue will identify the 972797e8345STejun Heo * recycled work item as currently executing and make it wait until the 973797e8345STejun Heo * current execution finishes, introducing an unwanted dependency. 974797e8345STejun Heo * 975797e8345STejun Heo * This function checks the work item address and work function to avoid 976797e8345STejun Heo * false positives. Note that this isn't complete as one may construct a 977797e8345STejun Heo * work function which can introduce dependency onto itself through a 978797e8345STejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 979797e8345STejun Heo * foot that badly, there's only so much we can do, and if such deadlock 980797e8345STejun Heo * actually occurs, it should be easy to locate the culprit work function. 981797e8345STejun Heo * 982797e8345STejun Heo * CONTEXT: 983797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 984797e8345STejun Heo * 985797e8345STejun Heo * Return: 986797e8345STejun Heo * Pointer to worker which is executing @work if found, %NULL 987797e8345STejun Heo * otherwise. 988797e8345STejun Heo */ 989797e8345STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 990797e8345STejun Heo struct work_struct *work) 991797e8345STejun Heo { 992797e8345STejun Heo struct worker *worker; 993797e8345STejun Heo 994797e8345STejun Heo hash_for_each_possible(pool->busy_hash, worker, hentry, 995797e8345STejun Heo (unsigned long)work) 996797e8345STejun Heo if (worker->current_work == work && 997797e8345STejun Heo worker->current_func == work->func) 998797e8345STejun Heo return worker; 999797e8345STejun Heo 1000797e8345STejun Heo return NULL; 1001797e8345STejun Heo } 1002797e8345STejun Heo 1003797e8345STejun Heo /** 1004797e8345STejun Heo * move_linked_works - move linked works to a list 1005797e8345STejun Heo * @work: start of series of works to be scheduled 1006797e8345STejun Heo * @head: target list to append @work to 1007797e8345STejun Heo * @nextp: out parameter for nested worklist walking 1008797e8345STejun Heo * 1009797e8345STejun Heo * Schedule linked works starting from @work to @head. Work series to 1010797e8345STejun Heo * be scheduled starts at @work and includes any consecutive work with 1011797e8345STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1012797e8345STejun Heo * 1013797e8345STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1014797e8345STejun Heo * the last scheduled work. This allows move_linked_works() to be 1015797e8345STejun Heo * nested inside outer list_for_each_entry_safe(). 1016797e8345STejun Heo * 1017797e8345STejun Heo * CONTEXT: 1018797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1019797e8345STejun Heo */ 1020797e8345STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1021797e8345STejun Heo struct work_struct **nextp) 1022797e8345STejun Heo { 1023797e8345STejun Heo struct work_struct *n; 1024797e8345STejun Heo 1025797e8345STejun Heo /* 1026797e8345STejun Heo * Linked worklist will always end before the end of the list, 1027797e8345STejun Heo * use NULL for list head. 1028797e8345STejun Heo */ 1029797e8345STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1030797e8345STejun Heo list_move_tail(&work->entry, head); 1031797e8345STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1032797e8345STejun Heo break; 1033797e8345STejun Heo } 1034797e8345STejun Heo 1035797e8345STejun Heo /* 1036797e8345STejun Heo * If we're already inside safe list traversal and have moved 1037797e8345STejun Heo * multiple works to the scheduled queue, the next position 1038797e8345STejun Heo * needs to be updated. 1039797e8345STejun Heo */ 1040797e8345STejun Heo if (nextp) 1041797e8345STejun Heo *nextp = n; 1042797e8345STejun Heo } 1043797e8345STejun Heo 1044797e8345STejun Heo /** 1045797e8345STejun Heo * wake_up_worker - wake up an idle worker 1046797e8345STejun Heo * @pool: worker pool to wake worker from 1047797e8345STejun Heo * 1048797e8345STejun Heo * Wake up the first idle worker of @pool. 1049797e8345STejun Heo * 1050797e8345STejun Heo * CONTEXT: 1051797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1052797e8345STejun Heo */ 1053797e8345STejun Heo static void wake_up_worker(struct worker_pool *pool) 1054797e8345STejun Heo { 1055797e8345STejun Heo struct worker *worker = first_idle_worker(pool); 1056797e8345STejun Heo 1057797e8345STejun Heo if (likely(worker)) 1058797e8345STejun Heo wake_up_process(worker->task); 1059797e8345STejun Heo } 1060797e8345STejun Heo 106163638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 106263638450STejun Heo 106363638450STejun Heo /* 106463638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 106563638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 106663638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 106763638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 106863638450STejun Heo * should be using an unbound workqueue instead. 106963638450STejun Heo * 107063638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 107163638450STejun Heo * and report them so that they can be examined and converted to use unbound 107263638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 107363638450STejun Heo * function is tracked and reported with exponential backoff. 107463638450STejun Heo */ 107563638450STejun Heo #define WCI_MAX_ENTS 128 107663638450STejun Heo 107763638450STejun Heo struct wci_ent { 107863638450STejun Heo work_func_t func; 107963638450STejun Heo atomic64_t cnt; 108063638450STejun Heo struct hlist_node hash_node; 108163638450STejun Heo }; 108263638450STejun Heo 108363638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 108463638450STejun Heo static int wci_nr_ents; 108563638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 108663638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 108763638450STejun Heo 108863638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 108963638450STejun Heo { 109063638450STejun Heo struct wci_ent *ent; 109163638450STejun Heo 109263638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 109363638450STejun Heo (unsigned long)func) { 109463638450STejun Heo if (ent->func == func) 109563638450STejun Heo return ent; 109663638450STejun Heo } 109763638450STejun Heo return NULL; 109863638450STejun Heo } 109963638450STejun Heo 110063638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 110163638450STejun Heo { 110263638450STejun Heo struct wci_ent *ent; 110363638450STejun Heo 110463638450STejun Heo restart: 110563638450STejun Heo ent = wci_find_ent(func); 110663638450STejun Heo if (ent) { 110763638450STejun Heo u64 cnt; 110863638450STejun Heo 110963638450STejun Heo /* 111063638450STejun Heo * Start reporting from the fourth time and back off 111163638450STejun Heo * exponentially. 111263638450STejun Heo */ 111363638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 111463638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 111563638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 111663638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 111763638450STejun Heo atomic64_read(&ent->cnt)); 111863638450STejun Heo return; 111963638450STejun Heo } 112063638450STejun Heo 112163638450STejun Heo /* 112263638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 112363638450STejun Heo * is exhausted, something went really wrong and we probably made enough 112463638450STejun Heo * noise already. 112563638450STejun Heo */ 112663638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 112763638450STejun Heo return; 112863638450STejun Heo 112963638450STejun Heo raw_spin_lock(&wci_lock); 113063638450STejun Heo 113163638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 113263638450STejun Heo raw_spin_unlock(&wci_lock); 113363638450STejun Heo return; 113463638450STejun Heo } 113563638450STejun Heo 113663638450STejun Heo if (wci_find_ent(func)) { 113763638450STejun Heo raw_spin_unlock(&wci_lock); 113863638450STejun Heo goto restart; 113963638450STejun Heo } 114063638450STejun Heo 114163638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 114263638450STejun Heo ent->func = func; 114363638450STejun Heo atomic64_set(&ent->cnt, 1); 114463638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 114563638450STejun Heo 114663638450STejun Heo raw_spin_unlock(&wci_lock); 114763638450STejun Heo } 114863638450STejun Heo 114963638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 115063638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 115163638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 115263638450STejun Heo 1153c54d5046STejun Heo /** 11541da177e4SLinus Torvalds * wq_worker_running - a worker is running again 11551da177e4SLinus Torvalds * @task: task waking up 11561da177e4SLinus Torvalds * 11571da177e4SLinus Torvalds * This function is called when a worker returns from schedule() 11581da177e4SLinus Torvalds */ 11591da177e4SLinus Torvalds void wq_worker_running(struct task_struct *task) 11601da177e4SLinus Torvalds { 11611da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 11621da177e4SLinus Torvalds 1163c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 11641da177e4SLinus Torvalds return; 11651da177e4SLinus Torvalds 11661da177e4SLinus Torvalds /* 11671da177e4SLinus Torvalds * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 11681da177e4SLinus Torvalds * and the nr_running increment below, we may ruin the nr_running reset 11691da177e4SLinus Torvalds * and leave with an unexpected pool->nr_running == 1 on the newly unbound 11701da177e4SLinus Torvalds * pool. Protect against such race. 11711da177e4SLinus Torvalds */ 11721da177e4SLinus Torvalds preempt_disable(); 11731da177e4SLinus Torvalds if (!(worker->flags & WORKER_NOT_RUNNING)) 11741da177e4SLinus Torvalds worker->pool->nr_running++; 11751da177e4SLinus Torvalds preempt_enable(); 1176616db877STejun Heo 1177616db877STejun Heo /* 1178616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1179616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1180616db877STejun Heo */ 1181616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1182616db877STejun Heo 1183c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 11841da177e4SLinus Torvalds } 11851da177e4SLinus Torvalds 11861da177e4SLinus Torvalds /** 11871da177e4SLinus Torvalds * wq_worker_sleeping - a worker is going to sleep 11881da177e4SLinus Torvalds * @task: task going to sleep 11891da177e4SLinus Torvalds * 11901da177e4SLinus Torvalds * This function is called from schedule() when a busy worker is 11911da177e4SLinus Torvalds * going to sleep. 11921da177e4SLinus Torvalds */ 11931da177e4SLinus Torvalds void wq_worker_sleeping(struct task_struct *task) 11941da177e4SLinus Torvalds { 11951da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 11961da177e4SLinus Torvalds struct worker_pool *pool; 11971da177e4SLinus Torvalds 11981da177e4SLinus Torvalds /* 11991da177e4SLinus Torvalds * Rescuers, which may not have all the fields set up like normal 12001da177e4SLinus Torvalds * workers, also reach here, let's not access anything before 12011da177e4SLinus Torvalds * checking NOT_RUNNING. 12021da177e4SLinus Torvalds */ 12031da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) 12041da177e4SLinus Torvalds return; 12051da177e4SLinus Torvalds 12061da177e4SLinus Torvalds pool = worker->pool; 12071da177e4SLinus Torvalds 12081da177e4SLinus Torvalds /* Return if preempted before wq_worker_running() was reached */ 1209c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 12101da177e4SLinus Torvalds return; 12111da177e4SLinus Torvalds 1212c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 12131da177e4SLinus Torvalds raw_spin_lock_irq(&pool->lock); 12141da177e4SLinus Torvalds 12151da177e4SLinus Torvalds /* 12161da177e4SLinus Torvalds * Recheck in case unbind_workers() preempted us. We don't 12171da177e4SLinus Torvalds * want to decrement nr_running after the worker is unbound 12181da177e4SLinus Torvalds * and nr_running has been reset. 12191da177e4SLinus Torvalds */ 12201da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) { 12211da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 12221da177e4SLinus Torvalds return; 12231da177e4SLinus Torvalds } 12241da177e4SLinus Torvalds 12251da177e4SLinus Torvalds pool->nr_running--; 1226725e8ec5STejun Heo if (need_more_worker(pool)) { 1227725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 12281da177e4SLinus Torvalds wake_up_worker(pool); 1229725e8ec5STejun Heo } 12301da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 12311da177e4SLinus Torvalds } 12321da177e4SLinus Torvalds 12331da177e4SLinus Torvalds /** 1234616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1235616db877STejun Heo * @task: task currently running 1236616db877STejun Heo * 1237616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1238616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1239616db877STejun Heo */ 1240616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1241616db877STejun Heo { 1242616db877STejun Heo struct worker *worker = kthread_data(task); 1243616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1244616db877STejun Heo struct worker_pool *pool = worker->pool; 1245616db877STejun Heo 1246616db877STejun Heo if (!pwq) 1247616db877STejun Heo return; 1248616db877STejun Heo 12498a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 12508a1dd1e5STejun Heo 125118c8ae81SZqiang if (!wq_cpu_intensive_thresh_us) 125218c8ae81SZqiang return; 125318c8ae81SZqiang 1254616db877STejun Heo /* 1255616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1256616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1257616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1258c8f6219bSZqiang * 1259c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1260c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1261c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1262c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1263c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1264c8f6219bSZqiang * We probably want to make this prettier in the future. 1265616db877STejun Heo */ 1266c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1267616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1268616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1269616db877STejun Heo return; 1270616db877STejun Heo 1271616db877STejun Heo raw_spin_lock(&pool->lock); 1272616db877STejun Heo 1273616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 127463638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1275616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1276616db877STejun Heo 1277616db877STejun Heo if (need_more_worker(pool)) { 1278616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1279616db877STejun Heo wake_up_worker(pool); 1280616db877STejun Heo } 1281616db877STejun Heo 1282616db877STejun Heo raw_spin_unlock(&pool->lock); 1283616db877STejun Heo } 1284616db877STejun Heo 1285616db877STejun Heo /** 12861da177e4SLinus Torvalds * wq_worker_last_func - retrieve worker's last work function 12871da177e4SLinus Torvalds * @task: Task to retrieve last work function of. 12881da177e4SLinus Torvalds * 12891da177e4SLinus Torvalds * Determine the last function a worker executed. This is called from 12901da177e4SLinus Torvalds * the scheduler to get a worker's last known identity. 12911da177e4SLinus Torvalds * 12921da177e4SLinus Torvalds * CONTEXT: 12939b41ea72SAndrew Morton * raw_spin_lock_irq(rq->lock) 12941da177e4SLinus Torvalds * 12951da177e4SLinus Torvalds * This function is called during schedule() when a kworker is going 1296f756d5e2SNathan Lynch * to sleep. It's used by psi to identify aggregation workers during 1297f756d5e2SNathan Lynch * dequeuing, to allow periodic aggregation to shut-off when that 12981da177e4SLinus Torvalds * worker is the last task in the system or cgroup to go to sleep. 12991da177e4SLinus Torvalds * 13001da177e4SLinus Torvalds * As this function doesn't involve any workqueue-related locking, it 13011da177e4SLinus Torvalds * only returns stable values when called from inside the scheduler's 13021da177e4SLinus Torvalds * queuing and dequeuing paths, when @task, which must be a kworker, 13031da177e4SLinus Torvalds * is guaranteed to not be processing any works. 1304365970a1SDavid Howells * 1305365970a1SDavid Howells * Return: 1306365970a1SDavid Howells * The last work function %current executed as a worker, NULL if it 1307365970a1SDavid Howells * hasn't executed any work yet. 1308365970a1SDavid Howells */ 1309365970a1SDavid Howells work_func_t wq_worker_last_func(struct task_struct *task) 1310365970a1SDavid Howells { 1311365970a1SDavid Howells struct worker *worker = kthread_data(task); 1312365970a1SDavid Howells 1313365970a1SDavid Howells return worker->last_func; 1314365970a1SDavid Howells } 1315365970a1SDavid Howells 1316d302f017STejun Heo /** 13178864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 13188864b4e5STejun Heo * @pwq: pool_workqueue to get 13198864b4e5STejun Heo * 13208864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 13218864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 13228864b4e5STejun Heo */ 13238864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 13248864b4e5STejun Heo { 13258864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13268864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 13278864b4e5STejun Heo pwq->refcnt++; 13288864b4e5STejun Heo } 13298864b4e5STejun Heo 13308864b4e5STejun Heo /** 13318864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 13328864b4e5STejun Heo * @pwq: pool_workqueue to put 13338864b4e5STejun Heo * 13348864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 13358864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 13368864b4e5STejun Heo */ 13378864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 13388864b4e5STejun Heo { 13398864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13408864b4e5STejun Heo if (likely(--pwq->refcnt)) 13418864b4e5STejun Heo return; 13428864b4e5STejun Heo /* 1343967b494eSTejun Heo * @pwq can't be released under pool->lock, bounce to a dedicated 1344967b494eSTejun Heo * kthread_worker to avoid A-A deadlocks. 13458864b4e5STejun Heo */ 1346687a9aa5STejun Heo kthread_queue_work(pwq_release_worker, &pwq->release_work); 13478864b4e5STejun Heo } 13488864b4e5STejun Heo 1349dce90d47STejun Heo /** 1350dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1351dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1352dce90d47STejun Heo * 1353dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1354dce90d47STejun Heo */ 1355dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1356dce90d47STejun Heo { 1357dce90d47STejun Heo if (pwq) { 1358dce90d47STejun Heo /* 135924acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1360dce90d47STejun Heo * following lock operations are safe. 1361dce90d47STejun Heo */ 1362a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1363dce90d47STejun Heo put_pwq(pwq); 1364a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1365dce90d47STejun Heo } 1366dce90d47STejun Heo } 1367dce90d47STejun Heo 1368f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1369bf4ede01STejun Heo { 1370112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1371bf4ede01STejun Heo 1372bf4ede01STejun Heo trace_workqueue_activate_work(work); 137382607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 137482607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1375112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1376f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1377112202d9STejun Heo pwq->nr_active++; 1378bf4ede01STejun Heo } 1379bf4ede01STejun Heo 1380f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 13813aa62497SLai Jiangshan { 1382f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 13833aa62497SLai Jiangshan struct work_struct, entry); 13843aa62497SLai Jiangshan 1385f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13863aa62497SLai Jiangshan } 13873aa62497SLai Jiangshan 1388bf4ede01STejun Heo /** 1389112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1390112202d9STejun Heo * @pwq: pwq of interest 1391c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1392bf4ede01STejun Heo * 1393bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1394112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1395bf4ede01STejun Heo * 1396bf4ede01STejun Heo * CONTEXT: 1397a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1398bf4ede01STejun Heo */ 1399c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1400bf4ede01STejun Heo { 1401c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1402c4560c2cSLai Jiangshan 1403018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1404112202d9STejun Heo pwq->nr_active--; 1405f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1406f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1407112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1408f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1409bf4ede01STejun Heo } 1410018f3a13SLai Jiangshan } 1411018f3a13SLai Jiangshan 1412018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1413bf4ede01STejun Heo 1414bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1415112202d9STejun Heo if (likely(pwq->flush_color != color)) 14168864b4e5STejun Heo goto out_put; 1417bf4ede01STejun Heo 1418bf4ede01STejun Heo /* are there still in-flight works? */ 1419112202d9STejun Heo if (pwq->nr_in_flight[color]) 14208864b4e5STejun Heo goto out_put; 1421bf4ede01STejun Heo 1422112202d9STejun Heo /* this pwq is done, clear flush_color */ 1423112202d9STejun Heo pwq->flush_color = -1; 1424bf4ede01STejun Heo 1425bf4ede01STejun Heo /* 1426112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1427bf4ede01STejun Heo * will handle the rest. 1428bf4ede01STejun Heo */ 1429112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1430112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 14318864b4e5STejun Heo out_put: 14328864b4e5STejun Heo put_pwq(pwq); 1433bf4ede01STejun Heo } 1434bf4ede01STejun Heo 143536e227d2STejun Heo /** 1436bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 143736e227d2STejun Heo * @work: work item to steal 143836e227d2STejun Heo * @is_dwork: @work is a delayed_work 1439bbb68dfaSTejun Heo * @flags: place to store irq state 144036e227d2STejun Heo * 144136e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1442d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 144336e227d2STejun Heo * 1444d185af30SYacine Belkadi * Return: 14453eb6b31bSMauro Carvalho Chehab * 14463eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 144736e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 144836e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 144936e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1450bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1451bbb68dfaSTejun Heo * for arbitrarily long 14523eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 145336e227d2STejun Heo * 1454d185af30SYacine Belkadi * Note: 1455bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1456e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1457e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1458e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1459bbb68dfaSTejun Heo * 1460bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1461bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1462bbb68dfaSTejun Heo * 1463e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1464bf4ede01STejun Heo */ 1465bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1466bbb68dfaSTejun Heo unsigned long *flags) 1467bf4ede01STejun Heo { 1468d565ed63STejun Heo struct worker_pool *pool; 1469112202d9STejun Heo struct pool_workqueue *pwq; 1470bf4ede01STejun Heo 1471bbb68dfaSTejun Heo local_irq_save(*flags); 1472bbb68dfaSTejun Heo 147336e227d2STejun Heo /* try to steal the timer if it exists */ 147436e227d2STejun Heo if (is_dwork) { 147536e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 147636e227d2STejun Heo 1477e0aecdd8STejun Heo /* 1478e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1479e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1480e0aecdd8STejun Heo * running on the local CPU. 1481e0aecdd8STejun Heo */ 148236e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 148336e227d2STejun Heo return 1; 148436e227d2STejun Heo } 148536e227d2STejun Heo 148636e227d2STejun Heo /* try to claim PENDING the normal way */ 1487bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1488bf4ede01STejun Heo return 0; 1489bf4ede01STejun Heo 149024acfb71SThomas Gleixner rcu_read_lock(); 1491bf4ede01STejun Heo /* 1492bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1493bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1494bf4ede01STejun Heo */ 1495d565ed63STejun Heo pool = get_work_pool(work); 1496d565ed63STejun Heo if (!pool) 1497bbb68dfaSTejun Heo goto fail; 1498bf4ede01STejun Heo 1499a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1500bf4ede01STejun Heo /* 1501112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1502112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1503112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1504112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1505112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 15060b3dae68SLai Jiangshan * item is currently queued on that pool. 1507bf4ede01STejun Heo */ 1508112202d9STejun Heo pwq = get_work_pwq(work); 1509112202d9STejun Heo if (pwq && pwq->pool == pool) { 1510bf4ede01STejun Heo debug_work_deactivate(work); 15113aa62497SLai Jiangshan 15123aa62497SLai Jiangshan /* 1513018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1514018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1515018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1516018f3a13SLai Jiangshan * 1517f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1518d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1519f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 152016062836STejun Heo * management later on and cause stall. Make sure the work 152116062836STejun Heo * item is activated before grabbing. 15223aa62497SLai Jiangshan */ 1523f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1524f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 15253aa62497SLai Jiangshan 1526bf4ede01STejun Heo list_del_init(&work->entry); 1527c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 152836e227d2STejun Heo 1529112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 15304468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 15314468a00fSLai Jiangshan 1532a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 153324acfb71SThomas Gleixner rcu_read_unlock(); 153436e227d2STejun Heo return 1; 1535bf4ede01STejun Heo } 1536a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1537bbb68dfaSTejun Heo fail: 153824acfb71SThomas Gleixner rcu_read_unlock(); 1539bbb68dfaSTejun Heo local_irq_restore(*flags); 1540bbb68dfaSTejun Heo if (work_is_canceling(work)) 1541bbb68dfaSTejun Heo return -ENOENT; 1542bbb68dfaSTejun Heo cpu_relax(); 154336e227d2STejun Heo return -EAGAIN; 1544bf4ede01STejun Heo } 1545bf4ede01STejun Heo 1546bf4ede01STejun Heo /** 1547706026c2STejun Heo * insert_work - insert a work into a pool 1548112202d9STejun Heo * @pwq: pwq @work belongs to 15494690c4abSTejun Heo * @work: work to insert 15504690c4abSTejun Heo * @head: insertion point 15514690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 15524690c4abSTejun Heo * 1553112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1554706026c2STejun Heo * work_struct flags. 15554690c4abSTejun Heo * 15564690c4abSTejun Heo * CONTEXT: 1557a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1558365970a1SDavid Howells */ 1559112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1560112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1561b89deed3SOleg Nesterov { 1562fe089f87STejun Heo debug_work_activate(work); 1563e1d8aa9fSFrederic Weisbecker 1564e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1565f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1566e89a85d6SWalter Wu 15674690c4abSTejun Heo /* we own @work, set data and link */ 1568112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 15691a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 15708864b4e5STejun Heo get_pwq(pwq); 1571b89deed3SOleg Nesterov } 1572b89deed3SOleg Nesterov 1573c8efcc25STejun Heo /* 1574c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 15758d03ecfeSTejun Heo * same workqueue. 1576c8efcc25STejun Heo */ 1577c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1578c8efcc25STejun Heo { 1579c8efcc25STejun Heo struct worker *worker; 1580c8efcc25STejun Heo 15818d03ecfeSTejun Heo worker = current_wq_worker(); 1582c8efcc25STejun Heo /* 1583bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 15848d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1585c8efcc25STejun Heo */ 1586112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1587c8efcc25STejun Heo } 1588c8efcc25STejun Heo 1589ef557180SMike Galbraith /* 1590ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1591ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1592ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1593ef557180SMike Galbraith */ 1594ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1595ef557180SMike Galbraith { 1596ef557180SMike Galbraith int new_cpu; 1597ef557180SMike Galbraith 1598f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1599ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1600ef557180SMike Galbraith return cpu; 1601a8ec5880SAmmar Faizi } else { 1602a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1603f303fccbSTejun Heo } 1604f303fccbSTejun Heo 1605ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1606ef557180SMike Galbraith return cpu; 1607ef557180SMike Galbraith 1608ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1609ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1610ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1611ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1612ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1613ef557180SMike Galbraith return cpu; 1614ef557180SMike Galbraith } 1615ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1616ef557180SMike Galbraith 1617ef557180SMike Galbraith return new_cpu; 1618ef557180SMike Galbraith } 1619ef557180SMike Galbraith 1620d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 16211da177e4SLinus Torvalds struct work_struct *work) 16221da177e4SLinus Torvalds { 1623112202d9STejun Heo struct pool_workqueue *pwq; 1624fe089f87STejun Heo struct worker_pool *last_pool, *pool; 16258a2e8e5dSTejun Heo unsigned int work_flags; 1626b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 16278930cabaSTejun Heo 16288930cabaSTejun Heo /* 16298930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 16308930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 16318930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 16328930cabaSTejun Heo * happen with IRQ disabled. 16338930cabaSTejun Heo */ 16348e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 16351da177e4SLinus Torvalds 16361e19ffc6STejun Heo 163733e3f0a3SRichard Clark /* 163833e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 163933e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 164033e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 164133e3f0a3SRichard Clark */ 164233e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 164333e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1644e41e704bSTejun Heo return; 164524acfb71SThomas Gleixner rcu_read_lock(); 16469e8cd2f5STejun Heo retry: 1647aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1648636b927eSTejun Heo if (req_cpu == WORK_CPU_UNBOUND) { 1649636b927eSTejun Heo if (wq->flags & WQ_UNBOUND) 1650ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1651636b927eSTejun Heo else 1652aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1653aa202f1fSHillf Danton } 1654f3421797STejun Heo 1655636b927eSTejun Heo pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 1656fe089f87STejun Heo pool = pwq->pool; 1657fe089f87STejun Heo 165818aa9effSTejun Heo /* 1659c9178087STejun Heo * If @work was previously on a different pool, it might still be 1660c9178087STejun Heo * running there, in which case the work needs to be queued on that 1661c9178087STejun Heo * pool to guarantee non-reentrancy. 166218aa9effSTejun Heo */ 1663c9e7cf27STejun Heo last_pool = get_work_pool(work); 1664fe089f87STejun Heo if (last_pool && last_pool != pool) { 166518aa9effSTejun Heo struct worker *worker; 166618aa9effSTejun Heo 1667a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 166818aa9effSTejun Heo 1669c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 167018aa9effSTejun Heo 1671112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1672c9178087STejun Heo pwq = worker->current_pwq; 1673fe089f87STejun Heo pool = pwq->pool; 1674fe089f87STejun Heo WARN_ON_ONCE(pool != last_pool); 16758594fadeSLai Jiangshan } else { 167618aa9effSTejun Heo /* meh... not running there, queue here */ 1677a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1678fe089f87STejun Heo raw_spin_lock(&pool->lock); 167918aa9effSTejun Heo } 16808930cabaSTejun Heo } else { 1681fe089f87STejun Heo raw_spin_lock(&pool->lock); 16828930cabaSTejun Heo } 1683502ca9d8STejun Heo 16849e8cd2f5STejun Heo /* 1685636b927eSTejun Heo * pwq is determined and locked. For unbound pools, we could have raced 1686636b927eSTejun Heo * with pwq release and it could already be dead. If its refcnt is zero, 1687636b927eSTejun Heo * repeat pwq selection. Note that unbound pwqs never die without 1688636b927eSTejun Heo * another pwq replacing it in cpu_pwq or while work items are executing 1689636b927eSTejun Heo * on it, so the retrying is guaranteed to make forward-progress. 16909e8cd2f5STejun Heo */ 16919e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 16929e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1693fe089f87STejun Heo raw_spin_unlock(&pool->lock); 16949e8cd2f5STejun Heo cpu_relax(); 16959e8cd2f5STejun Heo goto retry; 16969e8cd2f5STejun Heo } 16979e8cd2f5STejun Heo /* oops */ 16989e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 16999e8cd2f5STejun Heo wq->name, cpu); 17009e8cd2f5STejun Heo } 17019e8cd2f5STejun Heo 1702112202d9STejun Heo /* pwq determined, queue */ 1703112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1704502ca9d8STejun Heo 170524acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 170624acfb71SThomas Gleixner goto out; 17071e19ffc6STejun Heo 1708112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1709112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 17101e19ffc6STejun Heo 1711112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1712fe089f87STejun Heo if (list_empty(&pool->worklist)) 1713fe089f87STejun Heo pool->watchdog_ts = jiffies; 1714fe089f87STejun Heo 1715cdadf009STejun Heo trace_workqueue_activate_work(work); 1716112202d9STejun Heo pwq->nr_active++; 1717fe089f87STejun Heo insert_work(pwq, work, &pool->worklist, work_flags); 1718fe089f87STejun Heo 1719fe089f87STejun Heo if (__need_more_worker(pool)) 1720fe089f87STejun Heo wake_up_worker(pool); 17218a2e8e5dSTejun Heo } else { 1722f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1723fe089f87STejun Heo insert_work(pwq, work, &pwq->inactive_works, work_flags); 17248a2e8e5dSTejun Heo } 17251e19ffc6STejun Heo 172624acfb71SThomas Gleixner out: 1727fe089f87STejun Heo raw_spin_unlock(&pool->lock); 172824acfb71SThomas Gleixner rcu_read_unlock(); 17291da177e4SLinus Torvalds } 17301da177e4SLinus Torvalds 17310fcb78c2SRolf Eike Beer /** 1732c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1733c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1734c1a220e7SZhang Rui * @wq: workqueue to use 1735c1a220e7SZhang Rui * @work: work to queue 1736c1a220e7SZhang Rui * 1737c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1738443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1739443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1740854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1741854f5cc5SPaul E. McKenney * online will get a splat. 1742d185af30SYacine Belkadi * 1743d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1744c1a220e7SZhang Rui */ 1745d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1746d4283e93STejun Heo struct work_struct *work) 1747c1a220e7SZhang Rui { 1748d4283e93STejun Heo bool ret = false; 17498930cabaSTejun Heo unsigned long flags; 17508930cabaSTejun Heo 17518930cabaSTejun Heo local_irq_save(flags); 1752c1a220e7SZhang Rui 175322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17544690c4abSTejun Heo __queue_work(cpu, wq, work); 1755d4283e93STejun Heo ret = true; 1756c1a220e7SZhang Rui } 17578930cabaSTejun Heo 17588930cabaSTejun Heo local_irq_restore(flags); 1759c1a220e7SZhang Rui return ret; 1760c1a220e7SZhang Rui } 1761ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1762c1a220e7SZhang Rui 17638204e0c1SAlexander Duyck /** 1764fef59c9cSTejun Heo * select_numa_node_cpu - Select a CPU based on NUMA node 17658204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 17668204e0c1SAlexander Duyck * 17678204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 17688204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 17698204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 17708204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 17718204e0c1SAlexander Duyck */ 1772fef59c9cSTejun Heo static int select_numa_node_cpu(int node) 17738204e0c1SAlexander Duyck { 17748204e0c1SAlexander Duyck int cpu; 17758204e0c1SAlexander Duyck 17768204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 1777fef59c9cSTejun Heo if (!wq_pod_enabled) 17788204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17798204e0c1SAlexander Duyck 17808204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 17818204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 17828204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17838204e0c1SAlexander Duyck 17848204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 17858204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 17868204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 17878204e0c1SAlexander Duyck return cpu; 17888204e0c1SAlexander Duyck 17898204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 17908204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 17918204e0c1SAlexander Duyck 17928204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 17938204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 17948204e0c1SAlexander Duyck } 17958204e0c1SAlexander Duyck 17968204e0c1SAlexander Duyck /** 17978204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 17988204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 17998204e0c1SAlexander Duyck * @wq: workqueue to use 18008204e0c1SAlexander Duyck * @work: work to queue 18018204e0c1SAlexander Duyck * 18028204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 18038204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 18048204e0c1SAlexander Duyck * NUMA node. 18058204e0c1SAlexander Duyck * 18068204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 18078204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 18088204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 18098204e0c1SAlexander Duyck * 18108204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 18118204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 18128204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 18138204e0c1SAlexander Duyck * 18148204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 18158204e0c1SAlexander Duyck */ 18168204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18178204e0c1SAlexander Duyck struct work_struct *work) 18188204e0c1SAlexander Duyck { 18198204e0c1SAlexander Duyck unsigned long flags; 18208204e0c1SAlexander Duyck bool ret = false; 18218204e0c1SAlexander Duyck 18228204e0c1SAlexander Duyck /* 18238204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 18248204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 18258204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 18268204e0c1SAlexander Duyck * 18278204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 18288204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 18298204e0c1SAlexander Duyck * some round robin type logic. 18308204e0c1SAlexander Duyck */ 18318204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 18328204e0c1SAlexander Duyck 18338204e0c1SAlexander Duyck local_irq_save(flags); 18348204e0c1SAlexander Duyck 18358204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1836fef59c9cSTejun Heo int cpu = select_numa_node_cpu(node); 18378204e0c1SAlexander Duyck 18388204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 18398204e0c1SAlexander Duyck ret = true; 18408204e0c1SAlexander Duyck } 18418204e0c1SAlexander Duyck 18428204e0c1SAlexander Duyck local_irq_restore(flags); 18438204e0c1SAlexander Duyck return ret; 18448204e0c1SAlexander Duyck } 18458204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 18468204e0c1SAlexander Duyck 18478c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 18481da177e4SLinus Torvalds { 18498c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 18501da177e4SLinus Torvalds 1851e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 185260c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 18531da177e4SLinus Torvalds } 18541438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 18551da177e4SLinus Torvalds 18567beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 185752bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18581da177e4SLinus Torvalds { 18597beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 18607beb2edfSTejun Heo struct work_struct *work = &dwork->work; 18611da177e4SLinus Torvalds 1862637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 18634b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1864fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1865fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 18667beb2edfSTejun Heo 18678852aac2STejun Heo /* 18688852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 18698852aac2STejun Heo * both optimization and correctness. The earliest @timer can 18708852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 18718852aac2STejun Heo * on that there's no such delay when @delay is 0. 18728852aac2STejun Heo */ 18738852aac2STejun Heo if (!delay) { 18748852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 18758852aac2STejun Heo return; 18768852aac2STejun Heo } 18778852aac2STejun Heo 187860c057bcSLai Jiangshan dwork->wq = wq; 18791265057fSTejun Heo dwork->cpu = cpu; 18807beb2edfSTejun Heo timer->expires = jiffies + delay; 18817beb2edfSTejun Heo 1882041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 18837beb2edfSTejun Heo add_timer_on(timer, cpu); 1884041bd12eSTejun Heo else 1885041bd12eSTejun Heo add_timer(timer); 18867beb2edfSTejun Heo } 18871da177e4SLinus Torvalds 18880fcb78c2SRolf Eike Beer /** 18890fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 18900fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 18910fcb78c2SRolf Eike Beer * @wq: workqueue to use 1892af9997e4SRandy Dunlap * @dwork: work to queue 18930fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 18940fcb78c2SRolf Eike Beer * 1895d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1896715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1897715f1300STejun Heo * execution. 18980fcb78c2SRolf Eike Beer */ 1899d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 190052bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 19017a6bc1cdSVenkatesh Pallipadi { 190252bad64dSDavid Howells struct work_struct *work = &dwork->work; 1903d4283e93STejun Heo bool ret = false; 19048930cabaSTejun Heo unsigned long flags; 19058930cabaSTejun Heo 19068930cabaSTejun Heo /* read the comment in __queue_work() */ 19078930cabaSTejun Heo local_irq_save(flags); 19087a6bc1cdSVenkatesh Pallipadi 190922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 19107beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1911d4283e93STejun Heo ret = true; 19127a6bc1cdSVenkatesh Pallipadi } 19138930cabaSTejun Heo 19148930cabaSTejun Heo local_irq_restore(flags); 19157a6bc1cdSVenkatesh Pallipadi return ret; 19167a6bc1cdSVenkatesh Pallipadi } 1917ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19181da177e4SLinus Torvalds 1919c8e55f36STejun Heo /** 19208376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19218376fe22STejun Heo * @cpu: CPU number to execute work on 19228376fe22STejun Heo * @wq: workqueue to use 19238376fe22STejun Heo * @dwork: work to queue 19248376fe22STejun Heo * @delay: number of jiffies to wait before queueing 19258376fe22STejun Heo * 19268376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 19278376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 19288376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 19298376fe22STejun Heo * current state. 19308376fe22STejun Heo * 1931d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 19328376fe22STejun Heo * pending and its timer was modified. 19338376fe22STejun Heo * 1934e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 19358376fe22STejun Heo * See try_to_grab_pending() for details. 19368376fe22STejun Heo */ 19378376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 19388376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 19398376fe22STejun Heo { 19408376fe22STejun Heo unsigned long flags; 19418376fe22STejun Heo int ret; 19428376fe22STejun Heo 19438376fe22STejun Heo do { 19448376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 19458376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 19468376fe22STejun Heo 19478376fe22STejun Heo if (likely(ret >= 0)) { 19488376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 19498376fe22STejun Heo local_irq_restore(flags); 19508376fe22STejun Heo } 19518376fe22STejun Heo 19528376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 19538376fe22STejun Heo return ret; 19548376fe22STejun Heo } 19558376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 19568376fe22STejun Heo 195705f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 195805f0fe6bSTejun Heo { 195905f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 196005f0fe6bSTejun Heo 196105f0fe6bSTejun Heo /* read the comment in __queue_work() */ 196205f0fe6bSTejun Heo local_irq_disable(); 196305f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 196405f0fe6bSTejun Heo local_irq_enable(); 196505f0fe6bSTejun Heo } 196605f0fe6bSTejun Heo 196705f0fe6bSTejun Heo /** 196805f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 196905f0fe6bSTejun Heo * @wq: workqueue to use 197005f0fe6bSTejun Heo * @rwork: work to queue 197105f0fe6bSTejun Heo * 197205f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 197305f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1974bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 197505f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 197605f0fe6bSTejun Heo */ 197705f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 197805f0fe6bSTejun Heo { 197905f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 198005f0fe6bSTejun Heo 198105f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 198205f0fe6bSTejun Heo rwork->wq = wq; 1983a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 198405f0fe6bSTejun Heo return true; 198505f0fe6bSTejun Heo } 198605f0fe6bSTejun Heo 198705f0fe6bSTejun Heo return false; 198805f0fe6bSTejun Heo } 198905f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 199005f0fe6bSTejun Heo 1991f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1992c34056a3STejun Heo { 1993c34056a3STejun Heo struct worker *worker; 1994c34056a3STejun Heo 1995f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1996c8e55f36STejun Heo if (worker) { 1997c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1998affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1999da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2000e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2001e22bee78STejun Heo worker->flags = WORKER_PREP; 2002c8e55f36STejun Heo } 2003c34056a3STejun Heo return worker; 2004c34056a3STejun Heo } 2005c34056a3STejun Heo 2006c34056a3STejun Heo /** 20074736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20084736cbf7SLai Jiangshan * @worker: worker to be attached 20094736cbf7SLai Jiangshan * @pool: the target pool 20104736cbf7SLai Jiangshan * 20114736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20124736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20134736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20144736cbf7SLai Jiangshan */ 20154736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 20164736cbf7SLai Jiangshan struct worker_pool *pool) 20174736cbf7SLai Jiangshan { 20181258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 20194736cbf7SLai Jiangshan 20204736cbf7SLai Jiangshan /* 20211258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 20221258fae7STejun Heo * stable across this function. See the comments above the flag 20231258fae7STejun Heo * definition for details. 20244736cbf7SLai Jiangshan */ 20254736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 20264736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 20275c25b5ffSPeter Zijlstra else 20285c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 20294736cbf7SLai Jiangshan 2030640f17c8SPeter Zijlstra if (worker->rescue_wq) 2031640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 2032640f17c8SPeter Zijlstra 20334736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2034a2d812a2STejun Heo worker->pool = pool; 20354736cbf7SLai Jiangshan 20361258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 20374736cbf7SLai Jiangshan } 20384736cbf7SLai Jiangshan 20394736cbf7SLai Jiangshan /** 204060f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 204160f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 204260f5a4bcSLai Jiangshan * 20434736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 20444736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 20454736cbf7SLai Jiangshan * other reference to the pool. 204660f5a4bcSLai Jiangshan */ 2047a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 204860f5a4bcSLai Jiangshan { 2049a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 205060f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 205160f5a4bcSLai Jiangshan 20521258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2053a2d812a2STejun Heo 20545c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2055da028469SLai Jiangshan list_del(&worker->node); 2056a2d812a2STejun Heo worker->pool = NULL; 2057a2d812a2STejun Heo 2058e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 205960f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 20601258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 206160f5a4bcSLai Jiangshan 2062b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2063b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2064b62c0751SLai Jiangshan 206560f5a4bcSLai Jiangshan if (detach_completion) 206660f5a4bcSLai Jiangshan complete(detach_completion); 206760f5a4bcSLai Jiangshan } 206860f5a4bcSLai Jiangshan 206960f5a4bcSLai Jiangshan /** 2070c34056a3STejun Heo * create_worker - create a new workqueue worker 207163d95a91STejun Heo * @pool: pool the new worker will belong to 2072c34056a3STejun Heo * 2073051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2074c34056a3STejun Heo * 2075c34056a3STejun Heo * CONTEXT: 2076c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2077c34056a3STejun Heo * 2078d185af30SYacine Belkadi * Return: 2079c34056a3STejun Heo * Pointer to the newly created worker. 2080c34056a3STejun Heo */ 2081bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2082c34056a3STejun Heo { 2083e441b56fSZhen Lei struct worker *worker; 2084e441b56fSZhen Lei int id; 2085e3c916a4STejun Heo char id_buf[16]; 2086c34056a3STejun Heo 20877cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2088e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 20893f0ea0b8SPetr Mladek if (id < 0) { 20903f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 20913f0ea0b8SPetr Mladek ERR_PTR(id)); 2092e441b56fSZhen Lei return NULL; 20933f0ea0b8SPetr Mladek } 2094c34056a3STejun Heo 2095f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 20963f0ea0b8SPetr Mladek if (!worker) { 20973f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2098c34056a3STejun Heo goto fail; 20993f0ea0b8SPetr Mladek } 2100c34056a3STejun Heo 2101c34056a3STejun Heo worker->id = id; 2102c34056a3STejun Heo 210329c91e99STejun Heo if (pool->cpu >= 0) 2104e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2105e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2106f3421797STejun Heo else 2107e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2108e3c916a4STejun Heo 2109f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2110e3c916a4STejun Heo "kworker/%s", id_buf); 21113f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 211260f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 211360f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 211460f54038SPetr Mladek id_buf); 211560f54038SPetr Mladek } else { 21163f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 21173f0ea0b8SPetr Mladek worker->task); 211860f54038SPetr Mladek } 2119c34056a3STejun Heo goto fail; 21203f0ea0b8SPetr Mladek } 2121c34056a3STejun Heo 212291151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 212325834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 212491151228SOleg Nesterov 2125da028469SLai Jiangshan /* successful, attach the worker to the pool */ 21264736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2127822d8405STejun Heo 2128051e1850SLai Jiangshan /* start the newly created worker */ 2129a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2130051e1850SLai Jiangshan worker->pool->nr_workers++; 2131051e1850SLai Jiangshan worker_enter_idle(worker); 2132051e1850SLai Jiangshan wake_up_process(worker->task); 2133a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2134051e1850SLai Jiangshan 2135c34056a3STejun Heo return worker; 2136822d8405STejun Heo 2137c34056a3STejun Heo fail: 2138e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2139c34056a3STejun Heo kfree(worker); 2140c34056a3STejun Heo return NULL; 2141c34056a3STejun Heo } 2142c34056a3STejun Heo 2143793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2144793777bcSValentin Schneider { 2145793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2146793777bcSValentin Schneider 2147793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2148793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2149793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2150793777bcSValentin Schneider else 2151793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2152793777bcSValentin Schneider } 2153793777bcSValentin Schneider 2154e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2155e02b9312SValentin Schneider { 2156e02b9312SValentin Schneider struct worker *worker, *tmp; 2157e02b9312SValentin Schneider 2158e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2159e02b9312SValentin Schneider list_del_init(&worker->entry); 2160e02b9312SValentin Schneider unbind_worker(worker); 2161e02b9312SValentin Schneider /* 2162e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2163e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2164e02b9312SValentin Schneider * wouldn't have gotten here. 2165c34056a3STejun Heo * 2166e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2167e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2168e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2169e02b9312SValentin Schneider * outside of pool->lock. 2170e02b9312SValentin Schneider */ 2171e02b9312SValentin Schneider wake_up_process(worker->task); 2172e02b9312SValentin Schneider } 2173e02b9312SValentin Schneider } 2174e02b9312SValentin Schneider 2175e02b9312SValentin Schneider /** 2176e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2177e02b9312SValentin Schneider * @worker: worker to be destroyed 2178e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2179e02b9312SValentin Schneider * 2180e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2181e02b9312SValentin Schneider * should be idle. 2182c8e55f36STejun Heo * 2183c8e55f36STejun Heo * CONTEXT: 2184a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2185c34056a3STejun Heo */ 2186e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2187c34056a3STejun Heo { 2188bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2189c34056a3STejun Heo 2190cd549687STejun Heo lockdep_assert_held(&pool->lock); 2191e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2192cd549687STejun Heo 2193c34056a3STejun Heo /* sanity check frenzy */ 21946183c009STejun Heo if (WARN_ON(worker->current_work) || 219573eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 219673eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 21976183c009STejun Heo return; 2198c34056a3STejun Heo 2199bd7bdd43STejun Heo pool->nr_workers--; 2200bd7bdd43STejun Heo pool->nr_idle--; 2201c8e55f36STejun Heo 2202cb444766STejun Heo worker->flags |= WORKER_DIE; 2203e02b9312SValentin Schneider 2204e02b9312SValentin Schneider list_move(&worker->entry, list); 2205e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2206c34056a3STejun Heo } 2207c34056a3STejun Heo 22083f959aa3SValentin Schneider /** 22093f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 22103f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 22113f959aa3SValentin Schneider * 22123f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 22133f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 22143f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 22153f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 22163f959aa3SValentin Schneider * it expire and re-evaluate things from there. 22173f959aa3SValentin Schneider */ 221832a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2219e22bee78STejun Heo { 222032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 22213f959aa3SValentin Schneider bool do_cull = false; 22223f959aa3SValentin Schneider 22233f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 22243f959aa3SValentin Schneider return; 22253f959aa3SValentin Schneider 22263f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 22273f959aa3SValentin Schneider 22283f959aa3SValentin Schneider if (too_many_workers(pool)) { 22293f959aa3SValentin Schneider struct worker *worker; 22303f959aa3SValentin Schneider unsigned long expires; 22313f959aa3SValentin Schneider 22323f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 22333f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 22343f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 22353f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 22363f959aa3SValentin Schneider 22373f959aa3SValentin Schneider if (!do_cull) 22383f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 22393f959aa3SValentin Schneider } 22403f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 22413f959aa3SValentin Schneider 22423f959aa3SValentin Schneider if (do_cull) 22433f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 22443f959aa3SValentin Schneider } 22453f959aa3SValentin Schneider 22463f959aa3SValentin Schneider /** 22473f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 22483f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 22493f959aa3SValentin Schneider * 22503f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 22513f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2252e02b9312SValentin Schneider * 2253e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2254e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2255e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 22563f959aa3SValentin Schneider */ 22573f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 22583f959aa3SValentin Schneider { 22593f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 22609680540cSYang Yingliang LIST_HEAD(cull_list); 2261e22bee78STejun Heo 2262e02b9312SValentin Schneider /* 2263e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2264e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2265e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2266e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2267e02b9312SValentin Schneider */ 2268e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2269a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2270e22bee78STejun Heo 22713347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2272e22bee78STejun Heo struct worker *worker; 2273e22bee78STejun Heo unsigned long expires; 2274e22bee78STejun Heo 227563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2276e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2277e22bee78STejun Heo 22783347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 227963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 22803347fc9fSLai Jiangshan break; 2281e22bee78STejun Heo } 22823347fc9fSLai Jiangshan 2283e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2284e22bee78STejun Heo } 2285e22bee78STejun Heo 2286a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2287e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2288e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2289e22bee78STejun Heo } 2290e22bee78STejun Heo 2291493a1724STejun Heo static void send_mayday(struct work_struct *work) 2292e22bee78STejun Heo { 2293112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2294112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2295493a1724STejun Heo 22962e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2297e22bee78STejun Heo 2298493008a8STejun Heo if (!wq->rescuer) 2299493a1724STejun Heo return; 2300e22bee78STejun Heo 2301e22bee78STejun Heo /* mayday mayday mayday */ 2302493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 230377668c8bSLai Jiangshan /* 230477668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 230577668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 230677668c8bSLai Jiangshan * rescuer is done with it. 230777668c8bSLai Jiangshan */ 230877668c8bSLai Jiangshan get_pwq(pwq); 2309493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2310e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2311725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2312493a1724STejun Heo } 2313e22bee78STejun Heo } 2314e22bee78STejun Heo 231532a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2316e22bee78STejun Heo { 231732a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2318e22bee78STejun Heo struct work_struct *work; 2319e22bee78STejun Heo 2320a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2321a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2322e22bee78STejun Heo 232363d95a91STejun Heo if (need_to_create_worker(pool)) { 2324e22bee78STejun Heo /* 2325e22bee78STejun Heo * We've been trying to create a new worker but 2326e22bee78STejun Heo * haven't been successful. We might be hitting an 2327e22bee78STejun Heo * allocation deadlock. Send distress signals to 2328e22bee78STejun Heo * rescuers. 2329e22bee78STejun Heo */ 233063d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2331e22bee78STejun Heo send_mayday(work); 2332e22bee78STejun Heo } 2333e22bee78STejun Heo 2334a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2335a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2336e22bee78STejun Heo 233763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2338e22bee78STejun Heo } 2339e22bee78STejun Heo 2340e22bee78STejun Heo /** 2341e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 234263d95a91STejun Heo * @pool: pool to create a new worker for 2343e22bee78STejun Heo * 234463d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2345e22bee78STejun Heo * have at least one idle worker on return from this function. If 2346e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 234763d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2348e22bee78STejun Heo * possible allocation deadlock. 2349e22bee78STejun Heo * 2350c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2351c5aa87bbSTejun Heo * may_start_working() %true. 2352e22bee78STejun Heo * 2353e22bee78STejun Heo * LOCKING: 2354a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2355e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2356e22bee78STejun Heo * manager. 2357e22bee78STejun Heo */ 235829187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2359d565ed63STejun Heo __releases(&pool->lock) 2360d565ed63STejun Heo __acquires(&pool->lock) 2361e22bee78STejun Heo { 2362e22bee78STejun Heo restart: 2363a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 23649f9c2364STejun Heo 2365e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 236663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2367e22bee78STejun Heo 2368e22bee78STejun Heo while (true) { 2369051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2370e22bee78STejun Heo break; 2371e22bee78STejun Heo 2372e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 23739f9c2364STejun Heo 237463d95a91STejun Heo if (!need_to_create_worker(pool)) 2375e22bee78STejun Heo break; 2376e22bee78STejun Heo } 2377e22bee78STejun Heo 237863d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2379a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2380051e1850SLai Jiangshan /* 2381051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2382051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2383051e1850SLai Jiangshan * already become busy. 2384051e1850SLai Jiangshan */ 238563d95a91STejun Heo if (need_to_create_worker(pool)) 2386e22bee78STejun Heo goto restart; 2387e22bee78STejun Heo } 2388e22bee78STejun Heo 2389e22bee78STejun Heo /** 2390e22bee78STejun Heo * manage_workers - manage worker pool 2391e22bee78STejun Heo * @worker: self 2392e22bee78STejun Heo * 2393706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2394e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2395706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2396e22bee78STejun Heo * 2397e22bee78STejun Heo * The caller can safely start processing works on false return. On 2398e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2399e22bee78STejun Heo * and may_start_working() is true. 2400e22bee78STejun Heo * 2401e22bee78STejun Heo * CONTEXT: 2402a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2403e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2404e22bee78STejun Heo * 2405d185af30SYacine Belkadi * Return: 240629187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 240729187a9eSTejun Heo * start processing works, %true if management function was performed and 240829187a9eSTejun Heo * the conditions that the caller verified before calling the function may 240929187a9eSTejun Heo * no longer be true. 2410e22bee78STejun Heo */ 2411e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2412e22bee78STejun Heo { 241363d95a91STejun Heo struct worker_pool *pool = worker->pool; 2414e22bee78STejun Heo 2415692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 241629187a9eSTejun Heo return false; 2417692b4825STejun Heo 2418692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 24192607d7a6STejun Heo pool->manager = worker; 2420e22bee78STejun Heo 242129187a9eSTejun Heo maybe_create_worker(pool); 2422e22bee78STejun Heo 24232607d7a6STejun Heo pool->manager = NULL; 2424692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2425d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 242629187a9eSTejun Heo return true; 2427e22bee78STejun Heo } 2428e22bee78STejun Heo 2429a62428c0STejun Heo /** 2430a62428c0STejun Heo * process_one_work - process single work 2431c34056a3STejun Heo * @worker: self 2432a62428c0STejun Heo * @work: work to process 2433a62428c0STejun Heo * 2434a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2435a62428c0STejun Heo * process a single work including synchronization against and 2436a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2437a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2438a62428c0STejun Heo * call this function to process a work. 2439a62428c0STejun Heo * 2440a62428c0STejun Heo * CONTEXT: 2441a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2442a62428c0STejun Heo */ 2443c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2444d565ed63STejun Heo __releases(&pool->lock) 2445d565ed63STejun Heo __acquires(&pool->lock) 24461da177e4SLinus Torvalds { 2447112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2448bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2449c4560c2cSLai Jiangshan unsigned long work_data; 24507e11629dSTejun Heo struct worker *collision; 24514e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 24524e6045f1SJohannes Berg /* 2453a62428c0STejun Heo * It is permissible to free the struct work_struct from 2454a62428c0STejun Heo * inside the function that is called from it, this we need to 2455a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2456a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2457a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 24584e6045f1SJohannes Berg */ 24594d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 24604d82a1deSPeter Zijlstra 24614d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 24624e6045f1SJohannes Berg #endif 2463807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 246485327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2465ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 246625511a47STejun Heo 24677e11629dSTejun Heo /* 24687e11629dSTejun Heo * A single work shouldn't be executed concurrently by 24697e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 24707e11629dSTejun Heo * already processing the work. If so, defer the work to the 24717e11629dSTejun Heo * currently executing one. 24727e11629dSTejun Heo */ 2473c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 24747e11629dSTejun Heo if (unlikely(collision)) { 24757e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 24767e11629dSTejun Heo return; 24777e11629dSTejun Heo } 24781da177e4SLinus Torvalds 24798930cabaSTejun Heo /* claim and dequeue */ 24801da177e4SLinus Torvalds debug_work_deactivate(work); 2481c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2482c34056a3STejun Heo worker->current_work = work; 2483a2c1c57bSTejun Heo worker->current_func = work->func; 2484112202d9STejun Heo worker->current_pwq = pwq; 2485616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2486c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2487d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 24887a22ad75STejun Heo 24898bf89593STejun Heo /* 24908bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 24918bf89593STejun Heo * overridden through set_worker_desc(). 24928bf89593STejun Heo */ 24938bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 24948bf89593STejun Heo 2495a62428c0STejun Heo list_del_init(&work->entry); 2496a62428c0STejun Heo 2497649027d7STejun Heo /* 2498228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2499228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2500228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2501228f1d00SLai Jiangshan * execution of the pending work items. 2502fb0e7bebSTejun Heo */ 2503616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2504228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2505fb0e7bebSTejun Heo 2506974271c4STejun Heo /* 2507a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2508a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2509a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2510a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2511228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2512974271c4STejun Heo */ 2513a489a03eSLai Jiangshan if (need_more_worker(pool)) 251463d95a91STejun Heo wake_up_worker(pool); 2515974271c4STejun Heo 25168930cabaSTejun Heo /* 25177c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2518d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 251923657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 252023657bb1STejun Heo * disabled. 25218930cabaSTejun Heo */ 25227c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 25231da177e4SLinus Torvalds 2524a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2525365970a1SDavid Howells 2526a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 25273295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2528e6f3faa7SPeter Zijlstra /* 2529f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2530f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2531e6f3faa7SPeter Zijlstra * 2532e6f3faa7SPeter Zijlstra * However, that would result in: 2533e6f3faa7SPeter Zijlstra * 2534e6f3faa7SPeter Zijlstra * A(W1) 2535e6f3faa7SPeter Zijlstra * WFC(C) 2536e6f3faa7SPeter Zijlstra * A(W1) 2537e6f3faa7SPeter Zijlstra * C(C) 2538e6f3faa7SPeter Zijlstra * 2539e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2540e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2541e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2542f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2543e6f3faa7SPeter Zijlstra * these locks. 2544e6f3faa7SPeter Zijlstra * 2545e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2546e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2547e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2548e6f3faa7SPeter Zijlstra */ 2549f52be570SPeter Zijlstra lockdep_invariant_state(true); 2550725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2551e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2552a2c1c57bSTejun Heo worker->current_func(work); 2553e36c886aSArjan van de Ven /* 2554e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2555e36c886aSArjan van de Ven * point will only record its address. 2556e36c886aSArjan van de Ven */ 25571c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2558725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 25593295f0efSIngo Molnar lock_map_release(&lockdep_map); 2560112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 25611da177e4SLinus Torvalds 2562d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2563044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2564d75f773cSSakari Ailus " last function: %ps\n", 2565a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2566a2c1c57bSTejun Heo worker->current_func); 2567d5abe669SPeter Zijlstra debug_show_held_locks(current); 2568d5abe669SPeter Zijlstra dump_stack(); 2569d5abe669SPeter Zijlstra } 2570d5abe669SPeter Zijlstra 2571b22ce278STejun Heo /* 2572025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2573b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2574b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2575b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2576789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2577789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2578b22ce278STejun Heo */ 2579a7e6425eSPaul E. McKenney cond_resched(); 2580b22ce278STejun Heo 2581a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2582a62428c0STejun Heo 2583616db877STejun Heo /* 2584616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2585616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2586616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2587616db877STejun Heo */ 2588fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2589fb0e7bebSTejun Heo 25901b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 25911b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 25921b69ac6bSJohannes Weiner 2593a62428c0STejun Heo /* we're done with it, release */ 259442f8570fSSasha Levin hash_del(&worker->hentry); 2595c34056a3STejun Heo worker->current_work = NULL; 2596a2c1c57bSTejun Heo worker->current_func = NULL; 2597112202d9STejun Heo worker->current_pwq = NULL; 2598d812796eSLai Jiangshan worker->current_color = INT_MAX; 2599c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26001da177e4SLinus Torvalds } 26011da177e4SLinus Torvalds 2602affee4b2STejun Heo /** 2603affee4b2STejun Heo * process_scheduled_works - process scheduled works 2604affee4b2STejun Heo * @worker: self 2605affee4b2STejun Heo * 2606affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2607affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2608affee4b2STejun Heo * fetches a work from the top and executes it. 2609affee4b2STejun Heo * 2610affee4b2STejun Heo * CONTEXT: 2611a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2612affee4b2STejun Heo * multiple times. 2613affee4b2STejun Heo */ 2614affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26151da177e4SLinus Torvalds { 2616c0ab017dSTejun Heo struct work_struct *work; 2617c0ab017dSTejun Heo bool first = true; 2618c0ab017dSTejun Heo 2619c0ab017dSTejun Heo while ((work = list_first_entry_or_null(&worker->scheduled, 2620c0ab017dSTejun Heo struct work_struct, entry))) { 2621c0ab017dSTejun Heo if (first) { 2622c0ab017dSTejun Heo worker->pool->watchdog_ts = jiffies; 2623c0ab017dSTejun Heo first = false; 2624c0ab017dSTejun Heo } 2625c34056a3STejun Heo process_one_work(worker, work); 2626a62428c0STejun Heo } 26271da177e4SLinus Torvalds } 26281da177e4SLinus Torvalds 2629197f6accSTejun Heo static void set_pf_worker(bool val) 2630197f6accSTejun Heo { 2631197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2632197f6accSTejun Heo if (val) 2633197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2634197f6accSTejun Heo else 2635197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2636197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2637197f6accSTejun Heo } 2638197f6accSTejun Heo 26394690c4abSTejun Heo /** 26404690c4abSTejun Heo * worker_thread - the worker thread function 2641c34056a3STejun Heo * @__worker: self 26424690c4abSTejun Heo * 2643c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2644c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2645c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2646c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2647c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2648d185af30SYacine Belkadi * 2649d185af30SYacine Belkadi * Return: 0 26504690c4abSTejun Heo */ 2651c34056a3STejun Heo static int worker_thread(void *__worker) 26521da177e4SLinus Torvalds { 2653c34056a3STejun Heo struct worker *worker = __worker; 2654bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 26551da177e4SLinus Torvalds 2656e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2657197f6accSTejun Heo set_pf_worker(true); 2658c8e55f36STejun Heo woke_up: 2659a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2660affee4b2STejun Heo 2661a9ab775bSTejun Heo /* am I supposed to die? */ 2662a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2663a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2664197f6accSTejun Heo set_pf_worker(false); 266560f5a4bcSLai Jiangshan 266660f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2667e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2668a2d812a2STejun Heo worker_detach_from_pool(worker); 2669e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 267060f5a4bcSLai Jiangshan kfree(worker); 2671c8e55f36STejun Heo return 0; 2672c8e55f36STejun Heo } 2673c8e55f36STejun Heo 2674c8e55f36STejun Heo worker_leave_idle(worker); 2675db7bccf4STejun Heo recheck: 2676e22bee78STejun Heo /* no more worker necessary? */ 267763d95a91STejun Heo if (!need_more_worker(pool)) 2678e22bee78STejun Heo goto sleep; 2679e22bee78STejun Heo 2680e22bee78STejun Heo /* do we need to manage? */ 268163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2682e22bee78STejun Heo goto recheck; 2683e22bee78STejun Heo 2684c8e55f36STejun Heo /* 2685c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2686c8e55f36STejun Heo * preparing to process a work or actually processing it. 2687c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2688c8e55f36STejun Heo */ 26896183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2690c8e55f36STejun Heo 2691e22bee78STejun Heo /* 2692a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2693a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2694a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2695a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2696a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2697e22bee78STejun Heo */ 2698a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2699e22bee78STejun Heo 2700e22bee78STejun Heo do { 2701affee4b2STejun Heo struct work_struct *work = 2702bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2703affee4b2STejun Heo struct work_struct, entry); 2704affee4b2STejun Heo 2705c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2706affee4b2STejun Heo process_scheduled_works(worker); 270763d95a91STejun Heo } while (keep_working(pool)); 2708affee4b2STejun Heo 2709228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2710d313dd85STejun Heo sleep: 2711c8e55f36STejun Heo /* 2712d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2713d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2714d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2715d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2716d565ed63STejun Heo * event. 2717c8e55f36STejun Heo */ 2718c8e55f36STejun Heo worker_enter_idle(worker); 2719c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2720a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27211da177e4SLinus Torvalds schedule(); 2722c8e55f36STejun Heo goto woke_up; 27231da177e4SLinus Torvalds } 27241da177e4SLinus Torvalds 2725e22bee78STejun Heo /** 2726e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2727111c225aSTejun Heo * @__rescuer: self 2728e22bee78STejun Heo * 2729e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2730493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2731e22bee78STejun Heo * 2732706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2733e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2734e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2735e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2736e22bee78STejun Heo * the problem rescuer solves. 2737e22bee78STejun Heo * 2738706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2739706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2740e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2741e22bee78STejun Heo * 2742e22bee78STejun Heo * This should happen rarely. 2743d185af30SYacine Belkadi * 2744d185af30SYacine Belkadi * Return: 0 2745e22bee78STejun Heo */ 2746111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2747e22bee78STejun Heo { 2748111c225aSTejun Heo struct worker *rescuer = __rescuer; 2749111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2750e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 27514d595b86SLai Jiangshan bool should_stop; 2752e22bee78STejun Heo 2753e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2754111c225aSTejun Heo 2755111c225aSTejun Heo /* 2756111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2757111c225aSTejun Heo * doesn't participate in concurrency management. 2758111c225aSTejun Heo */ 2759197f6accSTejun Heo set_pf_worker(true); 2760e22bee78STejun Heo repeat: 2761c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 27621da177e4SLinus Torvalds 27634d595b86SLai Jiangshan /* 27644d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 27654d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 27664d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 27674d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 27684d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 27694d595b86SLai Jiangshan * list is always empty on exit. 27704d595b86SLai Jiangshan */ 27714d595b86SLai Jiangshan should_stop = kthread_should_stop(); 27721da177e4SLinus Torvalds 2773493a1724STejun Heo /* see whether any pwq is asking for help */ 2774a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2775493a1724STejun Heo 2776493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2777493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2778493a1724STejun Heo struct pool_workqueue, mayday_node); 2779112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2780e22bee78STejun Heo struct work_struct *work, *n; 2781e22bee78STejun Heo 2782e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2783493a1724STejun Heo list_del_init(&pwq->mayday_node); 2784493a1724STejun Heo 2785a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2786e22bee78STejun Heo 278751697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 278851697d39SLai Jiangshan 2789a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2790e22bee78STejun Heo 2791e22bee78STejun Heo /* 2792e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2793e22bee78STejun Heo * process'em. 2794e22bee78STejun Heo */ 27950479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 279682607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 279782607adcSTejun Heo if (get_work_pwq(work) == pwq) { 2798e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2799725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 280082607adcSTejun Heo } 280182607adcSTejun Heo } 2802e22bee78STejun Heo 2803008847f6SNeilBrown if (!list_empty(scheduled)) { 2804e22bee78STejun Heo process_scheduled_works(rescuer); 28057576958aSTejun Heo 28067576958aSTejun Heo /* 2807008847f6SNeilBrown * The above execution of rescued work items could 2808008847f6SNeilBrown * have created more to rescue through 2809f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2810008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2811008847f6SNeilBrown * that such back-to-back work items, which may be 2812008847f6SNeilBrown * being used to relieve memory pressure, don't 2813008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2814008847f6SNeilBrown */ 28154f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2816a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2817e66b39afSTejun Heo /* 2818e66b39afSTejun Heo * Queue iff we aren't racing destruction 2819e66b39afSTejun Heo * and somebody else hasn't queued it already. 2820e66b39afSTejun Heo */ 2821e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2822008847f6SNeilBrown get_pwq(pwq); 2823e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2824e66b39afSTejun Heo } 2825a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2826008847f6SNeilBrown } 2827008847f6SNeilBrown } 2828008847f6SNeilBrown 2829008847f6SNeilBrown /* 283077668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 283113b1d625SLai Jiangshan * go away while we're still attached to it. 283277668c8bSLai Jiangshan */ 283377668c8bSLai Jiangshan put_pwq(pwq); 283477668c8bSLai Jiangshan 283577668c8bSLai Jiangshan /* 2836d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 28377576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 28387576958aSTejun Heo * and stalling the execution. 28397576958aSTejun Heo */ 2840d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 284163d95a91STejun Heo wake_up_worker(pool); 28427576958aSTejun Heo 2843a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 284413b1d625SLai Jiangshan 2845a2d812a2STejun Heo worker_detach_from_pool(rescuer); 284613b1d625SLai Jiangshan 2847a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 28481da177e4SLinus Torvalds } 28491da177e4SLinus Torvalds 2850a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2851493a1724STejun Heo 28524d595b86SLai Jiangshan if (should_stop) { 28534d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2854197f6accSTejun Heo set_pf_worker(false); 28554d595b86SLai Jiangshan return 0; 28564d595b86SLai Jiangshan } 28574d595b86SLai Jiangshan 2858111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2859111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2860e22bee78STejun Heo schedule(); 2861e22bee78STejun Heo goto repeat; 28621da177e4SLinus Torvalds } 28631da177e4SLinus Torvalds 2864fca839c0STejun Heo /** 2865fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2866fca839c0STejun Heo * @target_wq: workqueue being flushed 2867fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2868fca839c0STejun Heo * 2869fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2870fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2871fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2872fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2873fca839c0STejun Heo * a deadlock. 2874fca839c0STejun Heo */ 2875fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2876fca839c0STejun Heo struct work_struct *target_work) 2877fca839c0STejun Heo { 2878fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2879fca839c0STejun Heo struct worker *worker; 2880fca839c0STejun Heo 2881fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2882fca839c0STejun Heo return; 2883fca839c0STejun Heo 2884fca839c0STejun Heo worker = current_wq_worker(); 2885fca839c0STejun Heo 2886fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2887d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2888fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 288923d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 289023d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2891d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2892fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2893fca839c0STejun Heo target_wq->name, target_func); 2894fca839c0STejun Heo } 2895fca839c0STejun Heo 2896fc2e4d70SOleg Nesterov struct wq_barrier { 2897fc2e4d70SOleg Nesterov struct work_struct work; 2898fc2e4d70SOleg Nesterov struct completion done; 28992607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2900fc2e4d70SOleg Nesterov }; 2901fc2e4d70SOleg Nesterov 2902fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2903fc2e4d70SOleg Nesterov { 2904fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2905fc2e4d70SOleg Nesterov complete(&barr->done); 2906fc2e4d70SOleg Nesterov } 2907fc2e4d70SOleg Nesterov 29084690c4abSTejun Heo /** 29094690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2910112202d9STejun Heo * @pwq: pwq to insert barrier into 29114690c4abSTejun Heo * @barr: wq_barrier to insert 2912affee4b2STejun Heo * @target: target work to attach @barr to 2913affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29144690c4abSTejun Heo * 2915affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2916affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2917affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2918affee4b2STejun Heo * cpu. 2919affee4b2STejun Heo * 2920affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2921affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2922affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2923affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2924affee4b2STejun Heo * after a work with LINKED flag set. 2925affee4b2STejun Heo * 2926affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2927112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 29284690c4abSTejun Heo * 29294690c4abSTejun Heo * CONTEXT: 2930a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 29314690c4abSTejun Heo */ 2932112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2933affee4b2STejun Heo struct wq_barrier *barr, 2934affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2935fc2e4d70SOleg Nesterov { 2936d812796eSLai Jiangshan unsigned int work_flags = 0; 2937d812796eSLai Jiangshan unsigned int work_color; 2938affee4b2STejun Heo struct list_head *head; 2939affee4b2STejun Heo 2940dc186ad7SThomas Gleixner /* 2941d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2942dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2943dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2944dc186ad7SThomas Gleixner * might deadlock. 2945dc186ad7SThomas Gleixner */ 2946ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 294722df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 294852fa5bc5SBoqun Feng 2949fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2950fd1a5b04SByungchul Park 29512607d7a6STejun Heo barr->task = current; 295283c22520SOleg Nesterov 2953018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2954018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2955018f3a13SLai Jiangshan 2956affee4b2STejun Heo /* 2957affee4b2STejun Heo * If @target is currently being executed, schedule the 2958affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2959affee4b2STejun Heo */ 2960d812796eSLai Jiangshan if (worker) { 2961affee4b2STejun Heo head = worker->scheduled.next; 2962d812796eSLai Jiangshan work_color = worker->current_color; 2963d812796eSLai Jiangshan } else { 2964affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2965affee4b2STejun Heo 2966affee4b2STejun Heo head = target->entry.next; 2967affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2968d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 2969d812796eSLai Jiangshan work_color = get_work_color(*bits); 2970affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2971affee4b2STejun Heo } 2972affee4b2STejun Heo 2973d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 2974d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 2975d812796eSLai Jiangshan 2976d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 2977fc2e4d70SOleg Nesterov } 2978fc2e4d70SOleg Nesterov 297973f53c4aSTejun Heo /** 2980112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 298173f53c4aSTejun Heo * @wq: workqueue being flushed 298273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 298373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 298473f53c4aSTejun Heo * 2985112202d9STejun Heo * Prepare pwqs for workqueue flushing. 298673f53c4aSTejun Heo * 2987112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2988112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2989112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2990112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2991112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 299273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 299373f53c4aSTejun Heo * 299473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 299573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 299673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 299773f53c4aSTejun Heo * is returned. 299873f53c4aSTejun Heo * 2999112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 300073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 300173f53c4aSTejun Heo * advanced to @work_color. 300273f53c4aSTejun Heo * 300373f53c4aSTejun Heo * CONTEXT: 30043c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 300573f53c4aSTejun Heo * 3006d185af30SYacine Belkadi * Return: 300773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 300873f53c4aSTejun Heo * otherwise. 300973f53c4aSTejun Heo */ 3010112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 301173f53c4aSTejun Heo int flush_color, int work_color) 30121da177e4SLinus Torvalds { 301373f53c4aSTejun Heo bool wait = false; 301449e3cf44STejun Heo struct pool_workqueue *pwq; 30151da177e4SLinus Torvalds 301673f53c4aSTejun Heo if (flush_color >= 0) { 30176183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3018112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3019dc186ad7SThomas Gleixner } 302014441960SOleg Nesterov 302149e3cf44STejun Heo for_each_pwq(pwq, wq) { 3022112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30231da177e4SLinus Torvalds 3024a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 302573f53c4aSTejun Heo 302673f53c4aSTejun Heo if (flush_color >= 0) { 30276183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 302873f53c4aSTejun Heo 3029112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3030112202d9STejun Heo pwq->flush_color = flush_color; 3031112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 303273f53c4aSTejun Heo wait = true; 30331da177e4SLinus Torvalds } 303473f53c4aSTejun Heo } 303573f53c4aSTejun Heo 303673f53c4aSTejun Heo if (work_color >= 0) { 30376183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3038112202d9STejun Heo pwq->work_color = work_color; 303973f53c4aSTejun Heo } 304073f53c4aSTejun Heo 3041a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 30421da177e4SLinus Torvalds } 30431da177e4SLinus Torvalds 3044112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 304573f53c4aSTejun Heo complete(&wq->first_flusher->done); 304673f53c4aSTejun Heo 304773f53c4aSTejun Heo return wait; 304883c22520SOleg Nesterov } 30491da177e4SLinus Torvalds 30500fcb78c2SRolf Eike Beer /** 3051c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 30520fcb78c2SRolf Eike Beer * @wq: workqueue to flush 30531da177e4SLinus Torvalds * 3054c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3055c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 30561da177e4SLinus Torvalds */ 3057c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 30581da177e4SLinus Torvalds { 305973f53c4aSTejun Heo struct wq_flusher this_flusher = { 306073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 306173f53c4aSTejun Heo .flush_color = -1, 3062fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 306373f53c4aSTejun Heo }; 306473f53c4aSTejun Heo int next_color; 3065b1f4ec17SOleg Nesterov 30663347fa09STejun Heo if (WARN_ON(!wq_online)) 30673347fa09STejun Heo return; 30683347fa09STejun Heo 306987915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 307087915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 307187915adcSJohannes Berg 30723c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 307373f53c4aSTejun Heo 307473f53c4aSTejun Heo /* 307573f53c4aSTejun Heo * Start-to-wait phase 307673f53c4aSTejun Heo */ 307773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 307873f53c4aSTejun Heo 307973f53c4aSTejun Heo if (next_color != wq->flush_color) { 308073f53c4aSTejun Heo /* 308173f53c4aSTejun Heo * Color space is not full. The current work_color 308273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 308373f53c4aSTejun Heo * by one. 308473f53c4aSTejun Heo */ 30856183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 308673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 308773f53c4aSTejun Heo wq->work_color = next_color; 308873f53c4aSTejun Heo 308973f53c4aSTejun Heo if (!wq->first_flusher) { 309073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 30916183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 309273f53c4aSTejun Heo 309373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 309473f53c4aSTejun Heo 3095112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 309673f53c4aSTejun Heo wq->work_color)) { 309773f53c4aSTejun Heo /* nothing to flush, done */ 309873f53c4aSTejun Heo wq->flush_color = next_color; 309973f53c4aSTejun Heo wq->first_flusher = NULL; 310073f53c4aSTejun Heo goto out_unlock; 310173f53c4aSTejun Heo } 310273f53c4aSTejun Heo } else { 310373f53c4aSTejun Heo /* wait in queue */ 31046183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 310573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3106112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 310773f53c4aSTejun Heo } 310873f53c4aSTejun Heo } else { 310973f53c4aSTejun Heo /* 311073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 311173f53c4aSTejun Heo * The next flush completion will assign us 311273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 311373f53c4aSTejun Heo */ 311473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 311573f53c4aSTejun Heo } 311673f53c4aSTejun Heo 3117fca839c0STejun Heo check_flush_dependency(wq, NULL); 3118fca839c0STejun Heo 31193c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 312073f53c4aSTejun Heo 312173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 312273f53c4aSTejun Heo 312373f53c4aSTejun Heo /* 312473f53c4aSTejun Heo * Wake-up-and-cascade phase 312573f53c4aSTejun Heo * 312673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 312773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 312873f53c4aSTejun Heo */ 312900d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 313073f53c4aSTejun Heo return; 313173f53c4aSTejun Heo 31323c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 313373f53c4aSTejun Heo 31344ce48b37STejun Heo /* we might have raced, check again with mutex held */ 31354ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 31364ce48b37STejun Heo goto out_unlock; 31374ce48b37STejun Heo 313800d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 313973f53c4aSTejun Heo 31406183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 31416183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 314273f53c4aSTejun Heo 314373f53c4aSTejun Heo while (true) { 314473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 314573f53c4aSTejun Heo 314673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 314773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 314873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 314973f53c4aSTejun Heo break; 315073f53c4aSTejun Heo list_del_init(&next->list); 315173f53c4aSTejun Heo complete(&next->done); 315273f53c4aSTejun Heo } 315373f53c4aSTejun Heo 31546183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 315573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 315673f53c4aSTejun Heo 315773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 315873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 315973f53c4aSTejun Heo 316073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 316173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 316273f53c4aSTejun Heo /* 316373f53c4aSTejun Heo * Assign the same color to all overflowed 316473f53c4aSTejun Heo * flushers, advance work_color and append to 316573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 316673f53c4aSTejun Heo * phase for these overflowed flushers. 316773f53c4aSTejun Heo */ 316873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 316973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 317073f53c4aSTejun Heo 317173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 317273f53c4aSTejun Heo 317373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 317473f53c4aSTejun Heo &wq->flusher_queue); 3175112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 317673f53c4aSTejun Heo } 317773f53c4aSTejun Heo 317873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 31796183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 318073f53c4aSTejun Heo break; 318173f53c4aSTejun Heo } 318273f53c4aSTejun Heo 318373f53c4aSTejun Heo /* 318473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3185112202d9STejun Heo * the new first flusher and arm pwqs. 318673f53c4aSTejun Heo */ 31876183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 31886183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 318973f53c4aSTejun Heo 319073f53c4aSTejun Heo list_del_init(&next->list); 319173f53c4aSTejun Heo wq->first_flusher = next; 319273f53c4aSTejun Heo 3193112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 319473f53c4aSTejun Heo break; 319573f53c4aSTejun Heo 319673f53c4aSTejun Heo /* 319773f53c4aSTejun Heo * Meh... this color is already done, clear first 319873f53c4aSTejun Heo * flusher and repeat cascading. 319973f53c4aSTejun Heo */ 320073f53c4aSTejun Heo wq->first_flusher = NULL; 320173f53c4aSTejun Heo } 320273f53c4aSTejun Heo 320373f53c4aSTejun Heo out_unlock: 32043c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32051da177e4SLinus Torvalds } 3206c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32071da177e4SLinus Torvalds 32089c5a2ba7STejun Heo /** 32099c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32109c5a2ba7STejun Heo * @wq: workqueue to drain 32119c5a2ba7STejun Heo * 32129c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32139c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32149c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3215b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32169c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32179c5a2ba7STejun Heo * takes too long. 32189c5a2ba7STejun Heo */ 32199c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32209c5a2ba7STejun Heo { 32219c5a2ba7STejun Heo unsigned int flush_cnt = 0; 322249e3cf44STejun Heo struct pool_workqueue *pwq; 32239c5a2ba7STejun Heo 32249c5a2ba7STejun Heo /* 32259c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 32269c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3227618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 32289c5a2ba7STejun Heo */ 322987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 32309c5a2ba7STejun Heo if (!wq->nr_drainers++) 3231618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 323287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32339c5a2ba7STejun Heo reflush: 3234c4f135d6STetsuo Handa __flush_workqueue(wq); 32359c5a2ba7STejun Heo 3236b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 323776af4d93STejun Heo 323849e3cf44STejun Heo for_each_pwq(pwq, wq) { 3239fa2563e4SThomas Tuttle bool drained; 32409c5a2ba7STejun Heo 3241a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3242f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3243a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3244fa2563e4SThomas Tuttle 3245fa2563e4SThomas Tuttle if (drained) 32469c5a2ba7STejun Heo continue; 32479c5a2ba7STejun Heo 32489c5a2ba7STejun Heo if (++flush_cnt == 10 || 32499c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3250e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3251e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 325276af4d93STejun Heo 3253b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 32549c5a2ba7STejun Heo goto reflush; 32559c5a2ba7STejun Heo } 32569c5a2ba7STejun Heo 32579c5a2ba7STejun Heo if (!--wq->nr_drainers) 3258618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 325987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32609c5a2ba7STejun Heo } 32619c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 32629c5a2ba7STejun Heo 3263d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3264d6e89786SJohannes Berg bool from_cancel) 3265baf59022STejun Heo { 3266baf59022STejun Heo struct worker *worker = NULL; 3267c9e7cf27STejun Heo struct worker_pool *pool; 3268112202d9STejun Heo struct pool_workqueue *pwq; 3269baf59022STejun Heo 3270baf59022STejun Heo might_sleep(); 3271baf59022STejun Heo 327224acfb71SThomas Gleixner rcu_read_lock(); 3273fa1b54e6STejun Heo pool = get_work_pool(work); 3274fa1b54e6STejun Heo if (!pool) { 327524acfb71SThomas Gleixner rcu_read_unlock(); 3276fa1b54e6STejun Heo return false; 3277fa1b54e6STejun Heo } 3278fa1b54e6STejun Heo 3279a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 32800b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3281112202d9STejun Heo pwq = get_work_pwq(work); 3282112202d9STejun Heo if (pwq) { 3283112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3284baf59022STejun Heo goto already_gone; 3285606a5020STejun Heo } else { 3286c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3287baf59022STejun Heo if (!worker) 3288baf59022STejun Heo goto already_gone; 3289112202d9STejun Heo pwq = worker->current_pwq; 3290606a5020STejun Heo } 3291baf59022STejun Heo 3292fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3293fca839c0STejun Heo 3294112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3295a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3296baf59022STejun Heo 3297e159489bSTejun Heo /* 3298a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3299a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3300a1d14934SPeter Zijlstra * 3301a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3302a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3303a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3304a1d14934SPeter Zijlstra * forward progress. 3305e159489bSTejun Heo */ 3306d6e89786SJohannes Berg if (!from_cancel && 3307d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3308112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3309112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3310a1d14934SPeter Zijlstra } 331124acfb71SThomas Gleixner rcu_read_unlock(); 3312baf59022STejun Heo return true; 3313baf59022STejun Heo already_gone: 3314a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 331524acfb71SThomas Gleixner rcu_read_unlock(); 3316baf59022STejun Heo return false; 3317baf59022STejun Heo } 3318baf59022STejun Heo 3319d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3320d6e89786SJohannes Berg { 3321d6e89786SJohannes Berg struct wq_barrier barr; 3322d6e89786SJohannes Berg 3323d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3324d6e89786SJohannes Berg return false; 3325d6e89786SJohannes Berg 33264d43d395STetsuo Handa if (WARN_ON(!work->func)) 33274d43d395STetsuo Handa return false; 33284d43d395STetsuo Handa 332987915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 333087915adcSJohannes Berg lock_map_release(&work->lockdep_map); 333187915adcSJohannes Berg 3332d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3333d6e89786SJohannes Berg wait_for_completion(&barr.done); 3334d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3335d6e89786SJohannes Berg return true; 3336d6e89786SJohannes Berg } else { 3337d6e89786SJohannes Berg return false; 3338d6e89786SJohannes Berg } 3339d6e89786SJohannes Berg } 3340d6e89786SJohannes Berg 3341db700897SOleg Nesterov /** 3342401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3343401a8d04STejun Heo * @work: the work to flush 3344db700897SOleg Nesterov * 3345606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3346606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3347401a8d04STejun Heo * 3348d185af30SYacine Belkadi * Return: 3349401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3350401a8d04STejun Heo * %false if it was already idle. 3351db700897SOleg Nesterov */ 3352401a8d04STejun Heo bool flush_work(struct work_struct *work) 3353db700897SOleg Nesterov { 3354d6e89786SJohannes Berg return __flush_work(work, false); 3355606a5020STejun Heo } 3356db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3357db700897SOleg Nesterov 33588603e1b3STejun Heo struct cwt_wait { 3359ac6424b9SIngo Molnar wait_queue_entry_t wait; 33608603e1b3STejun Heo struct work_struct *work; 33618603e1b3STejun Heo }; 33628603e1b3STejun Heo 3363ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 33648603e1b3STejun Heo { 33658603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 33668603e1b3STejun Heo 33678603e1b3STejun Heo if (cwait->work != key) 33688603e1b3STejun Heo return 0; 33698603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 33708603e1b3STejun Heo } 33718603e1b3STejun Heo 337236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3373401a8d04STejun Heo { 33748603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3375bbb68dfaSTejun Heo unsigned long flags; 33761f1f642eSOleg Nesterov int ret; 33771f1f642eSOleg Nesterov 33781f1f642eSOleg Nesterov do { 3379bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3380bbb68dfaSTejun Heo /* 33818603e1b3STejun Heo * If someone else is already canceling, wait for it to 33828603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 33838603e1b3STejun Heo * because we may get scheduled between @work's completion 33848603e1b3STejun Heo * and the other canceling task resuming and clearing 33858603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 33868603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 33878603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 33888603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 33898603e1b3STejun Heo * we're hogging the CPU. 33908603e1b3STejun Heo * 33918603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 33928603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 33938603e1b3STejun Heo * wake function which matches @work along with exclusive 33948603e1b3STejun Heo * wait and wakeup. 3395bbb68dfaSTejun Heo */ 33968603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 33978603e1b3STejun Heo struct cwt_wait cwait; 33988603e1b3STejun Heo 33998603e1b3STejun Heo init_wait(&cwait.wait); 34008603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34018603e1b3STejun Heo cwait.work = work; 34028603e1b3STejun Heo 34038603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34048603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34058603e1b3STejun Heo if (work_is_canceling(work)) 34068603e1b3STejun Heo schedule(); 34078603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34088603e1b3STejun Heo } 34091f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34101f1f642eSOleg Nesterov 3411bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3412bbb68dfaSTejun Heo mark_work_canceling(work); 3413bbb68dfaSTejun Heo local_irq_restore(flags); 3414bbb68dfaSTejun Heo 34153347fa09STejun Heo /* 34163347fa09STejun Heo * This allows canceling during early boot. We know that @work 34173347fa09STejun Heo * isn't executing. 34183347fa09STejun Heo */ 34193347fa09STejun Heo if (wq_online) 3420d6e89786SJohannes Berg __flush_work(work, true); 34213347fa09STejun Heo 34227a22ad75STejun Heo clear_work_data(work); 34238603e1b3STejun Heo 34248603e1b3STejun Heo /* 34258603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 34268603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 34278603e1b3STejun Heo * visible there. 34288603e1b3STejun Heo */ 34298603e1b3STejun Heo smp_mb(); 34308603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 34318603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 34328603e1b3STejun Heo 34331f1f642eSOleg Nesterov return ret; 34341f1f642eSOleg Nesterov } 34351f1f642eSOleg Nesterov 34366e84d644SOleg Nesterov /** 3437401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3438401a8d04STejun Heo * @work: the work to cancel 34396e84d644SOleg Nesterov * 3440401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3441401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3442401a8d04STejun Heo * another workqueue. On return from this function, @work is 3443401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 34441f1f642eSOleg Nesterov * 3445401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3446401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 34476e84d644SOleg Nesterov * 3448401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 34496e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3450401a8d04STejun Heo * 3451d185af30SYacine Belkadi * Return: 3452401a8d04STejun Heo * %true if @work was pending, %false otherwise. 34536e84d644SOleg Nesterov */ 3454401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 34556e84d644SOleg Nesterov { 345636e227d2STejun Heo return __cancel_work_timer(work, false); 3457b89deed3SOleg Nesterov } 345828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3459b89deed3SOleg Nesterov 34606e84d644SOleg Nesterov /** 3461401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3462401a8d04STejun Heo * @dwork: the delayed work to flush 34636e84d644SOleg Nesterov * 3464401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3465401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3466401a8d04STejun Heo * considers the last queueing instance of @dwork. 34671f1f642eSOleg Nesterov * 3468d185af30SYacine Belkadi * Return: 3469401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3470401a8d04STejun Heo * %false if it was already idle. 34716e84d644SOleg Nesterov */ 3472401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3473401a8d04STejun Heo { 34748930cabaSTejun Heo local_irq_disable(); 3475401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 347660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 34778930cabaSTejun Heo local_irq_enable(); 3478401a8d04STejun Heo return flush_work(&dwork->work); 3479401a8d04STejun Heo } 3480401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3481401a8d04STejun Heo 348205f0fe6bSTejun Heo /** 348305f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 348405f0fe6bSTejun Heo * @rwork: the rcu work to flush 348505f0fe6bSTejun Heo * 348605f0fe6bSTejun Heo * Return: 348705f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 348805f0fe6bSTejun Heo * %false if it was already idle. 348905f0fe6bSTejun Heo */ 349005f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 349105f0fe6bSTejun Heo { 349205f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 349305f0fe6bSTejun Heo rcu_barrier(); 349405f0fe6bSTejun Heo flush_work(&rwork->work); 349505f0fe6bSTejun Heo return true; 349605f0fe6bSTejun Heo } else { 349705f0fe6bSTejun Heo return flush_work(&rwork->work); 349805f0fe6bSTejun Heo } 349905f0fe6bSTejun Heo } 350005f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 350105f0fe6bSTejun Heo 3502f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3503f72b8792SJens Axboe { 3504f72b8792SJens Axboe unsigned long flags; 3505f72b8792SJens Axboe int ret; 3506f72b8792SJens Axboe 3507f72b8792SJens Axboe do { 3508f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3509f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3510f72b8792SJens Axboe 3511f72b8792SJens Axboe if (unlikely(ret < 0)) 3512f72b8792SJens Axboe return false; 3513f72b8792SJens Axboe 3514f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3515f72b8792SJens Axboe local_irq_restore(flags); 3516f72b8792SJens Axboe return ret; 3517f72b8792SJens Axboe } 3518f72b8792SJens Axboe 351973b4b532SAndrey Grodzovsky /* 352073b4b532SAndrey Grodzovsky * See cancel_delayed_work() 352173b4b532SAndrey Grodzovsky */ 352273b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 352373b4b532SAndrey Grodzovsky { 352473b4b532SAndrey Grodzovsky return __cancel_work(work, false); 352573b4b532SAndrey Grodzovsky } 352673b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 352773b4b532SAndrey Grodzovsky 3528401a8d04STejun Heo /** 352957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 353057b30ae7STejun Heo * @dwork: delayed_work to cancel 353109383498STejun Heo * 3532d185af30SYacine Belkadi * Kill off a pending delayed_work. 3533d185af30SYacine Belkadi * 3534d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3535d185af30SYacine Belkadi * pending. 3536d185af30SYacine Belkadi * 3537d185af30SYacine Belkadi * Note: 3538d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3539d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3540d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 354109383498STejun Heo * 354257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 354309383498STejun Heo */ 354457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 354509383498STejun Heo { 3546f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 354709383498STejun Heo } 354857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 354909383498STejun Heo 355009383498STejun Heo /** 3551401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3552401a8d04STejun Heo * @dwork: the delayed work cancel 3553401a8d04STejun Heo * 3554401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3555401a8d04STejun Heo * 3556d185af30SYacine Belkadi * Return: 3557401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3558401a8d04STejun Heo */ 3559401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 35606e84d644SOleg Nesterov { 356136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 35626e84d644SOleg Nesterov } 3563f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 35641da177e4SLinus Torvalds 35650fcb78c2SRolf Eike Beer /** 356631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3567b6136773SAndrew Morton * @func: the function to call 3568b6136773SAndrew Morton * 356931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 357031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3571b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 357231ddd871STejun Heo * 3573d185af30SYacine Belkadi * Return: 357431ddd871STejun Heo * 0 on success, -errno on failure. 3575b6136773SAndrew Morton */ 357665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 357715316ba8SChristoph Lameter { 357815316ba8SChristoph Lameter int cpu; 357938f51568SNamhyung Kim struct work_struct __percpu *works; 358015316ba8SChristoph Lameter 3581b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3582b6136773SAndrew Morton if (!works) 358315316ba8SChristoph Lameter return -ENOMEM; 3584b6136773SAndrew Morton 3585ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 358693981800STejun Heo 358715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 35889bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 35899bfb1839SIngo Molnar 35909bfb1839SIngo Molnar INIT_WORK(work, func); 35918de6d308SOleg Nesterov schedule_work_on(cpu, work); 359215316ba8SChristoph Lameter } 359393981800STejun Heo 359493981800STejun Heo for_each_online_cpu(cpu) 35958616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 359693981800STejun Heo 3597ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3598b6136773SAndrew Morton free_percpu(works); 359915316ba8SChristoph Lameter return 0; 360015316ba8SChristoph Lameter } 360115316ba8SChristoph Lameter 3602eef6a7d5SAlan Stern /** 36031fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36041fa44ecaSJames Bottomley * @fn: the function to execute 36051fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36061fa44ecaSJames Bottomley * be available when the work executes) 36071fa44ecaSJames Bottomley * 36081fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36091fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36101fa44ecaSJames Bottomley * 3611d185af30SYacine Belkadi * Return: 0 - function was executed 36121fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36131fa44ecaSJames Bottomley */ 361465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36151fa44ecaSJames Bottomley { 36161fa44ecaSJames Bottomley if (!in_interrupt()) { 361765f27f38SDavid Howells fn(&ew->work); 36181fa44ecaSJames Bottomley return 0; 36191fa44ecaSJames Bottomley } 36201fa44ecaSJames Bottomley 362165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36221fa44ecaSJames Bottomley schedule_work(&ew->work); 36231fa44ecaSJames Bottomley 36241fa44ecaSJames Bottomley return 1; 36251fa44ecaSJames Bottomley } 36261fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 36271fa44ecaSJames Bottomley 36287a4e344cSTejun Heo /** 36297a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 36307a4e344cSTejun Heo * @attrs: workqueue_attrs to free 36317a4e344cSTejun Heo * 36327a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 36337a4e344cSTejun Heo */ 3634513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 36357a4e344cSTejun Heo { 36367a4e344cSTejun Heo if (attrs) { 36377a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 36387a4e344cSTejun Heo kfree(attrs); 36397a4e344cSTejun Heo } 36407a4e344cSTejun Heo } 36417a4e344cSTejun Heo 36427a4e344cSTejun Heo /** 36437a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 36447a4e344cSTejun Heo * 36457a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3646d185af30SYacine Belkadi * return it. 3647d185af30SYacine Belkadi * 3648d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 36497a4e344cSTejun Heo */ 3650513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 36517a4e344cSTejun Heo { 36527a4e344cSTejun Heo struct workqueue_attrs *attrs; 36537a4e344cSTejun Heo 3654be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 36557a4e344cSTejun Heo if (!attrs) 36567a4e344cSTejun Heo goto fail; 3657be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 36587a4e344cSTejun Heo goto fail; 36597a4e344cSTejun Heo 366013e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 36617a4e344cSTejun Heo return attrs; 36627a4e344cSTejun Heo fail: 36637a4e344cSTejun Heo free_workqueue_attrs(attrs); 36647a4e344cSTejun Heo return NULL; 36657a4e344cSTejun Heo } 36667a4e344cSTejun Heo 366729c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 366829c91e99STejun Heo const struct workqueue_attrs *from) 366929c91e99STejun Heo { 367029c91e99STejun Heo to->nice = from->nice; 367129c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 36722865a8fbSShaohua Li /* 36732865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 3674af73f5c9STejun Heo * ->ordered as it is used for both pool and wq attrs. Instead, 3675af73f5c9STejun Heo * get_unbound_pool() explicitly clears ->ordered after copying. 36762865a8fbSShaohua Li */ 3677af73f5c9STejun Heo to->ordered = from->ordered; 367829c91e99STejun Heo } 367929c91e99STejun Heo 368029c91e99STejun Heo /* hash value of the content of @attr */ 368129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 368229c91e99STejun Heo { 368329c91e99STejun Heo u32 hash = 0; 368429c91e99STejun Heo 368529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 368613e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 368713e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 368829c91e99STejun Heo return hash; 368929c91e99STejun Heo } 369029c91e99STejun Heo 369129c91e99STejun Heo /* content equality test */ 369229c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 369329c91e99STejun Heo const struct workqueue_attrs *b) 369429c91e99STejun Heo { 369529c91e99STejun Heo if (a->nice != b->nice) 369629c91e99STejun Heo return false; 369729c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 369829c91e99STejun Heo return false; 369929c91e99STejun Heo return true; 370029c91e99STejun Heo } 370129c91e99STejun Heo 37027a4e344cSTejun Heo /** 37037a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 37047a4e344cSTejun Heo * @pool: worker_pool to initialize 37057a4e344cSTejun Heo * 3706402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3707d185af30SYacine Belkadi * 3708d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 370929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 371029c91e99STejun Heo * on @pool safely to release it. 37117a4e344cSTejun Heo */ 37127a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 37134e1a1f9aSTejun Heo { 3714a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 371529c91e99STejun Heo pool->id = -1; 371629c91e99STejun Heo pool->cpu = -1; 3717f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 37184e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 371982607adcSTejun Heo pool->watchdog_ts = jiffies; 37204e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37214e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 37224e1a1f9aSTejun Heo hash_init(pool->busy_hash); 37234e1a1f9aSTejun Heo 372432a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 37253f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 37264e1a1f9aSTejun Heo 372732a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 37284e1a1f9aSTejun Heo 3729da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3730e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 37317a4e344cSTejun Heo 37327cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 373329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 373429c91e99STejun Heo pool->refcnt = 1; 373529c91e99STejun Heo 373629c91e99STejun Heo /* shouldn't fail above this point */ 3737be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 37387a4e344cSTejun Heo if (!pool->attrs) 37397a4e344cSTejun Heo return -ENOMEM; 37407a4e344cSTejun Heo return 0; 37414e1a1f9aSTejun Heo } 37424e1a1f9aSTejun Heo 3743669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3744669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3745669de8bdSBart Van Assche { 3746669de8bdSBart Van Assche char *lock_name; 3747669de8bdSBart Van Assche 3748669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3749669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3750669de8bdSBart Van Assche if (!lock_name) 3751669de8bdSBart Van Assche lock_name = wq->name; 375269a106c0SQian Cai 375369a106c0SQian Cai wq->lock_name = lock_name; 3754669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3755669de8bdSBart Van Assche } 3756669de8bdSBart Van Assche 3757669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3758669de8bdSBart Van Assche { 3759669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3760669de8bdSBart Van Assche } 3761669de8bdSBart Van Assche 3762669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3763669de8bdSBart Van Assche { 3764669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3765669de8bdSBart Van Assche kfree(wq->lock_name); 3766669de8bdSBart Van Assche } 3767669de8bdSBart Van Assche #else 3768669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3769669de8bdSBart Van Assche { 3770669de8bdSBart Van Assche } 3771669de8bdSBart Van Assche 3772669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3773669de8bdSBart Van Assche { 3774669de8bdSBart Van Assche } 3775669de8bdSBart Van Assche 3776669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3777669de8bdSBart Van Assche { 3778669de8bdSBart Van Assche } 3779669de8bdSBart Van Assche #endif 3780669de8bdSBart Van Assche 3781e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3782e2dca7adSTejun Heo { 3783e2dca7adSTejun Heo struct workqueue_struct *wq = 3784e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3785e2dca7adSTejun Heo 3786669de8bdSBart Van Assche wq_free_lockdep(wq); 3787ee1ceef7STejun Heo free_percpu(wq->cpu_pwq); 3788e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3789e2dca7adSTejun Heo kfree(wq); 3790e2dca7adSTejun Heo } 3791e2dca7adSTejun Heo 379229c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 379329c91e99STejun Heo { 379429c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 379529c91e99STejun Heo 37967cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 379729c91e99STejun Heo free_workqueue_attrs(pool->attrs); 379829c91e99STejun Heo kfree(pool); 379929c91e99STejun Heo } 380029c91e99STejun Heo 380129c91e99STejun Heo /** 380229c91e99STejun Heo * put_unbound_pool - put a worker_pool 380329c91e99STejun Heo * @pool: worker_pool to put 380429c91e99STejun Heo * 380524acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3806c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3807c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3808c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3809a892caccSTejun Heo * 3810a892caccSTejun Heo * Should be called with wq_pool_mutex held. 381129c91e99STejun Heo */ 381229c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 381329c91e99STejun Heo { 381460f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 381529c91e99STejun Heo struct worker *worker; 38169680540cSYang Yingliang LIST_HEAD(cull_list); 3817e02b9312SValentin Schneider 3818a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3819a892caccSTejun Heo 3820a892caccSTejun Heo if (--pool->refcnt) 382129c91e99STejun Heo return; 382229c91e99STejun Heo 382329c91e99STejun Heo /* sanity checks */ 382461d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3825a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 382629c91e99STejun Heo return; 382729c91e99STejun Heo 382829c91e99STejun Heo /* release id and unhash */ 382929c91e99STejun Heo if (pool->id >= 0) 383029c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 383129c91e99STejun Heo hash_del(&pool->hash_node); 383229c91e99STejun Heo 3833c5aa87bbSTejun Heo /* 3834692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3835692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3836692b4825STejun Heo * manager and @pool gets freed with the flag set. 38379ab03be4SValentin Schneider * 38389ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 38399ab03be4SValentin Schneider * only get here with 38409ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 38419ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 38429ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 38439ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 38449ab03be4SValentin Schneider * drops pool->lock 3845c5aa87bbSTejun Heo */ 38469ab03be4SValentin Schneider while (true) { 38479ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 38489ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3849d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3850e02b9312SValentin Schneider 3851e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 38529ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 38539ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3854692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 38559ab03be4SValentin Schneider break; 38569ab03be4SValentin Schneider } 38579ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3858e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 38599ab03be4SValentin Schneider } 3860692b4825STejun Heo 38611037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3862e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 386329c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3864a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 386560f5a4bcSLai Jiangshan 3866e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3867e02b9312SValentin Schneider 3868e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 386960f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 38701258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 387160f5a4bcSLai Jiangshan 387260f5a4bcSLai Jiangshan if (pool->detach_completion) 387360f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 387460f5a4bcSLai Jiangshan 387529c91e99STejun Heo /* shut down the timers */ 387629c91e99STejun Heo del_timer_sync(&pool->idle_timer); 38773f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 387829c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 387929c91e99STejun Heo 388024acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 388125b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 388229c91e99STejun Heo } 388329c91e99STejun Heo 388429c91e99STejun Heo /** 388529c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 388629c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 388729c91e99STejun Heo * 388829c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 388929c91e99STejun Heo * reference count and return it. If there already is a matching 389029c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3891d185af30SYacine Belkadi * create a new one. 3892a892caccSTejun Heo * 3893a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3894d185af30SYacine Belkadi * 3895d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3896d185af30SYacine Belkadi * On failure, %NULL. 389729c91e99STejun Heo */ 389829c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 389929c91e99STejun Heo { 390029c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 390129c91e99STejun Heo struct worker_pool *pool; 3902fef59c9cSTejun Heo int pod; 3903fef59c9cSTejun Heo int target_pod = NUMA_NO_NODE; 390429c91e99STejun Heo 3905a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 390629c91e99STejun Heo 390729c91e99STejun Heo /* do we already have a matching pool? */ 390829c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 390929c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 391029c91e99STejun Heo pool->refcnt++; 39113fb1823cSLai Jiangshan return pool; 391229c91e99STejun Heo } 391329c91e99STejun Heo } 391429c91e99STejun Heo 3915fef59c9cSTejun Heo /* if cpumask is contained inside a pod, we belong to that pod */ 3916fef59c9cSTejun Heo if (wq_pod_enabled) { 3917fef59c9cSTejun Heo for_each_node(pod) { 3918fef59c9cSTejun Heo if (cpumask_subset(attrs->cpumask, wq_pod_cpus[pod])) { 3919fef59c9cSTejun Heo target_pod = pod; 3920e2273584SXunlei Pang break; 3921e2273584SXunlei Pang } 3922e2273584SXunlei Pang } 3923e2273584SXunlei Pang } 3924e2273584SXunlei Pang 392529c91e99STejun Heo /* nope, create a new one */ 3926fef59c9cSTejun Heo pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_pod); 392729c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 392829c91e99STejun Heo goto fail; 392929c91e99STejun Heo 393029c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3931fef59c9cSTejun Heo pool->node = target_pod; 393229c91e99STejun Heo 39332865a8fbSShaohua Li /* 3934af73f5c9STejun Heo * ordered isn't a worker_pool attribute, always clear it. See 39352865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 39362865a8fbSShaohua Li */ 3937af73f5c9STejun Heo pool->attrs->ordered = false; 39382865a8fbSShaohua Li 393929c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 394029c91e99STejun Heo goto fail; 394129c91e99STejun Heo 394229c91e99STejun Heo /* create and start the initial worker */ 39433347fa09STejun Heo if (wq_online && !create_worker(pool)) 394429c91e99STejun Heo goto fail; 394529c91e99STejun Heo 394629c91e99STejun Heo /* install */ 394729c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 39483fb1823cSLai Jiangshan 394929c91e99STejun Heo return pool; 395029c91e99STejun Heo fail: 395129c91e99STejun Heo if (pool) 395229c91e99STejun Heo put_unbound_pool(pool); 395329c91e99STejun Heo return NULL; 395429c91e99STejun Heo } 395529c91e99STejun Heo 39568864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 39578864b4e5STejun Heo { 39588864b4e5STejun Heo kmem_cache_free(pwq_cache, 39598864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 39608864b4e5STejun Heo } 39618864b4e5STejun Heo 39628864b4e5STejun Heo /* 3963967b494eSTejun Heo * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 3964967b494eSTejun Heo * refcnt and needs to be destroyed. 39658864b4e5STejun Heo */ 3966687a9aa5STejun Heo static void pwq_release_workfn(struct kthread_work *work) 39678864b4e5STejun Heo { 39688864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 3969687a9aa5STejun Heo release_work); 39708864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 39718864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3972b42b0bddSYang Yingliang bool is_last = false; 39738864b4e5STejun Heo 3974b42b0bddSYang Yingliang /* 3975687a9aa5STejun Heo * When @pwq is not linked, it doesn't hold any reference to the 3976b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3977b42b0bddSYang Yingliang */ 3978b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 39793c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 39808864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3981bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 39823c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3983b42b0bddSYang Yingliang } 39848864b4e5STejun Heo 3985687a9aa5STejun Heo if (wq->flags & WQ_UNBOUND) { 3986a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 39878864b4e5STejun Heo put_unbound_pool(pool); 3988a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3989687a9aa5STejun Heo } 3990a892caccSTejun Heo 399125b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 39928864b4e5STejun Heo 39938864b4e5STejun Heo /* 39948864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3995e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 39968864b4e5STejun Heo */ 3997669de8bdSBart Van Assche if (is_last) { 3998669de8bdSBart Van Assche wq_unregister_lockdep(wq); 399925b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 40006029a918STejun Heo } 4001669de8bdSBart Van Assche } 40028864b4e5STejun Heo 40030fbd95aaSTejun Heo /** 4004699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 40050fbd95aaSTejun Heo * @pwq: target pool_workqueue 40060fbd95aaSTejun Heo * 4007699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4008f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4009699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 40100fbd95aaSTejun Heo */ 4011699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 40120fbd95aaSTejun Heo { 4013699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4014699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 40153347fa09STejun Heo unsigned long flags; 4016699ce097STejun Heo 4017699ce097STejun Heo /* for @wq->saved_max_active */ 4018a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4019699ce097STejun Heo 4020699ce097STejun Heo /* fast exit for non-freezable wqs */ 4021699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4022699ce097STejun Heo return; 4023699ce097STejun Heo 40243347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4025a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4026699ce097STejun Heo 402774b414eaSLai Jiangshan /* 402874b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 402974b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 403074b414eaSLai Jiangshan * is updated and visible. 403174b414eaSLai Jiangshan */ 403274b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 403301341fbdSYunfeng Ye bool kick = false; 403401341fbdSYunfeng Ye 4035699ce097STejun Heo pwq->max_active = wq->saved_max_active; 40360fbd95aaSTejun Heo 4037f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 403801341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 4039f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 404001341fbdSYunfeng Ye kick = true; 404101341fbdSYunfeng Ye } 4042951a078aSLai Jiangshan 4043951a078aSLai Jiangshan /* 4044951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 404501341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 404601341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 404701341fbdSYunfeng Ye * let's kick iff work items were activated. 4048951a078aSLai Jiangshan */ 404901341fbdSYunfeng Ye if (kick) 4050951a078aSLai Jiangshan wake_up_worker(pwq->pool); 4051699ce097STejun Heo } else { 4052699ce097STejun Heo pwq->max_active = 0; 4053699ce097STejun Heo } 4054699ce097STejun Heo 4055a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 40560fbd95aaSTejun Heo } 40570fbd95aaSTejun Heo 405867dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4059f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4060f147f29eSTejun Heo struct worker_pool *pool) 4061d2c1d404STejun Heo { 4062d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4063d2c1d404STejun Heo 4064e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4065e50aba9aSTejun Heo 4066d2c1d404STejun Heo pwq->pool = pool; 4067d2c1d404STejun Heo pwq->wq = wq; 4068d2c1d404STejun Heo pwq->flush_color = -1; 40698864b4e5STejun Heo pwq->refcnt = 1; 4070f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 40711befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4072d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 4073687a9aa5STejun Heo kthread_init_work(&pwq->release_work, pwq_release_workfn); 4074f147f29eSTejun Heo } 4075d2c1d404STejun Heo 4076f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 40771befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4078f147f29eSTejun Heo { 4079f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4080f147f29eSTejun Heo 4081f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 408275ccf595STejun Heo 40831befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 40841befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 40851befcf30STejun Heo return; 40861befcf30STejun Heo 408729b1cb41SLai Jiangshan /* set the matching work_color */ 408875ccf595STejun Heo pwq->work_color = wq->work_color; 4089983ca25eSTejun Heo 4090983ca25eSTejun Heo /* sync max_active to the current setting */ 4091983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4092983ca25eSTejun Heo 4093983ca25eSTejun Heo /* link in @pwq */ 40949e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4095df2d5ae4STejun Heo } 40966029a918STejun Heo 4097f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4098f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4099f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4100f147f29eSTejun Heo { 4101f147f29eSTejun Heo struct worker_pool *pool; 4102f147f29eSTejun Heo struct pool_workqueue *pwq; 4103f147f29eSTejun Heo 4104f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4105f147f29eSTejun Heo 4106f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4107f147f29eSTejun Heo if (!pool) 4108f147f29eSTejun Heo return NULL; 4109f147f29eSTejun Heo 4110e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4111f147f29eSTejun Heo if (!pwq) { 4112f147f29eSTejun Heo put_unbound_pool(pool); 4113f147f29eSTejun Heo return NULL; 4114f147f29eSTejun Heo } 4115f147f29eSTejun Heo 4116f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4117f147f29eSTejun Heo return pwq; 4118d2c1d404STejun Heo } 4119d2c1d404STejun Heo 41204c16bd32STejun Heo /** 4121fef59c9cSTejun Heo * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 4122042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 4123fef59c9cSTejun Heo * @pod: the target CPU pod 41244c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 41254c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 41264c16bd32STejun Heo * 4127fef59c9cSTejun Heo * Calculate the cpumask a workqueue with @attrs should use on @pod. If 4128fef59c9cSTejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during calculation. 4129fef59c9cSTejun Heo * The result is stored in @cpumask. 41304c16bd32STejun Heo * 4131fef59c9cSTejun Heo * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 4132fef59c9cSTejun Heo * and @pod has online CPUs requested by @attrs, the returned cpumask is the 4133fef59c9cSTejun Heo * intersection of the possible CPUs of @pod and @attrs->cpumask. 41344c16bd32STejun Heo * 4135fef59c9cSTejun Heo * The caller is responsible for ensuring that the cpumask of @pod stays stable. 41364c16bd32STejun Heo */ 4137fef59c9cSTejun Heo static void wq_calc_pod_cpumask(const struct workqueue_attrs *attrs, int pod, 41384c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 41394c16bd32STejun Heo { 4140fef59c9cSTejun Heo if (!wq_pod_enabled || attrs->ordered) 41414c16bd32STejun Heo goto use_dfl; 41424c16bd32STejun Heo 4143fef59c9cSTejun Heo /* does @pod have any online CPUs @attrs wants? */ 4144fef59c9cSTejun Heo cpumask_and(cpumask, cpumask_of_node(pod), attrs->cpumask); 41454c16bd32STejun Heo if (cpu_going_down >= 0) 41464c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 41474c16bd32STejun Heo 41484c16bd32STejun Heo if (cpumask_empty(cpumask)) 41494c16bd32STejun Heo goto use_dfl; 41504c16bd32STejun Heo 4151fef59c9cSTejun Heo /* yeap, return possible CPUs in @pod that @attrs wants */ 4152fef59c9cSTejun Heo cpumask_and(cpumask, attrs->cpumask, wq_pod_cpus[pod]); 41531ad0f0a7SMichael Bringmann 4154636b927eSTejun Heo if (cpumask_empty(cpumask)) 41551ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 41561ad0f0a7SMichael Bringmann "possible intersect\n"); 4157636b927eSTejun Heo return; 41584c16bd32STejun Heo 41594c16bd32STejun Heo use_dfl: 41604c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 41614c16bd32STejun Heo } 41624c16bd32STejun Heo 4163636b927eSTejun Heo /* install @pwq into @wq's cpu_pwq and return the old pwq */ 4164636b927eSTejun Heo static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 4165636b927eSTejun Heo int cpu, struct pool_workqueue *pwq) 41661befcf30STejun Heo { 41671befcf30STejun Heo struct pool_workqueue *old_pwq; 41681befcf30STejun Heo 41695b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 41701befcf30STejun Heo lockdep_assert_held(&wq->mutex); 41711befcf30STejun Heo 41721befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 41731befcf30STejun Heo link_pwq(pwq); 41741befcf30STejun Heo 4175636b927eSTejun Heo old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4176636b927eSTejun Heo rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq); 41771befcf30STejun Heo return old_pwq; 41781befcf30STejun Heo } 41791befcf30STejun Heo 41802d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 41812d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 41822d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 41832d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4184042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 41852d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 41862d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 41872d5f0764SLai Jiangshan }; 41882d5f0764SLai Jiangshan 41892d5f0764SLai Jiangshan /* free the resources after success or abort */ 41902d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 41912d5f0764SLai Jiangshan { 41922d5f0764SLai Jiangshan if (ctx) { 4193636b927eSTejun Heo int cpu; 41942d5f0764SLai Jiangshan 4195636b927eSTejun Heo for_each_possible_cpu(cpu) 4196636b927eSTejun Heo put_pwq_unlocked(ctx->pwq_tbl[cpu]); 41972d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 41982d5f0764SLai Jiangshan 41992d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 42002d5f0764SLai Jiangshan 42012d5f0764SLai Jiangshan kfree(ctx); 42022d5f0764SLai Jiangshan } 42032d5f0764SLai Jiangshan } 42042d5f0764SLai Jiangshan 42052d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 42062d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 42072d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 420899c621efSLai Jiangshan const struct workqueue_attrs *attrs, 420999c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 42102d5f0764SLai Jiangshan { 42112d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 42122d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 4213636b927eSTejun Heo int cpu; 42142d5f0764SLai Jiangshan 42152d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42162d5f0764SLai Jiangshan 4217636b927eSTejun Heo ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 42182d5f0764SLai Jiangshan 4219be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4220be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 42212d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 42222d5f0764SLai Jiangshan goto out_free; 42232d5f0764SLai Jiangshan 4224042f7df1SLai Jiangshan /* 422599c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 422699c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4227042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4228042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4229042f7df1SLai Jiangshan */ 42302d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 423199c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4232042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 423399c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 42342d5f0764SLai Jiangshan 42352d5f0764SLai Jiangshan /* 42362d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 42372d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 42382d5f0764SLai Jiangshan * pools. 42392d5f0764SLai Jiangshan */ 42402d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 42412d5f0764SLai Jiangshan 42422d5f0764SLai Jiangshan /* 42432d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 42442d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 42452d5f0764SLai Jiangshan * it even if we don't use it immediately. 42462d5f0764SLai Jiangshan */ 42472d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 42482d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 42492d5f0764SLai Jiangshan goto out_free; 42502d5f0764SLai Jiangshan 4251636b927eSTejun Heo for_each_possible_cpu(cpu) { 4252af73f5c9STejun Heo if (new_attrs->ordered) { 42532d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 4254636b927eSTejun Heo ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 4255636b927eSTejun Heo } else { 4256fef59c9cSTejun Heo wq_calc_pod_cpumask(new_attrs, cpu_to_node(cpu), -1, 4257636b927eSTejun Heo tmp_attrs->cpumask); 4258636b927eSTejun Heo ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, tmp_attrs); 4259636b927eSTejun Heo if (!ctx->pwq_tbl[cpu]) 4260636b927eSTejun Heo goto out_free; 42612d5f0764SLai Jiangshan } 42622d5f0764SLai Jiangshan } 42632d5f0764SLai Jiangshan 4264042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4265042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4266042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 42672d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4268042f7df1SLai Jiangshan 42692d5f0764SLai Jiangshan ctx->wq = wq; 42702d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 42712d5f0764SLai Jiangshan return ctx; 42722d5f0764SLai Jiangshan 42732d5f0764SLai Jiangshan out_free: 42742d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 42752d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 42762d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 42772d5f0764SLai Jiangshan return NULL; 42782d5f0764SLai Jiangshan } 42792d5f0764SLai Jiangshan 42802d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 42812d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 42822d5f0764SLai Jiangshan { 4283636b927eSTejun Heo int cpu; 42842d5f0764SLai Jiangshan 42852d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 42862d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 42872d5f0764SLai Jiangshan 42882d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 42892d5f0764SLai Jiangshan 42902d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 4291636b927eSTejun Heo for_each_possible_cpu(cpu) 4292636b927eSTejun Heo ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 4293636b927eSTejun Heo ctx->pwq_tbl[cpu]); 42942d5f0764SLai Jiangshan 42952d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 42962d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 42972d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 42982d5f0764SLai Jiangshan 42992d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 43002d5f0764SLai Jiangshan } 43012d5f0764SLai Jiangshan 4302a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4303a0111cf6SLai Jiangshan { 4304a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4305ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4306a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4307a0111cf6SLai Jiangshan } 4308a0111cf6SLai Jiangshan 4309a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4310a0111cf6SLai Jiangshan { 4311a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4312ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4313a0111cf6SLai Jiangshan } 4314a0111cf6SLai Jiangshan 4315a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4316a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4317a0111cf6SLai Jiangshan { 4318a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4319a0111cf6SLai Jiangshan 4320a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4321a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4322a0111cf6SLai Jiangshan return -EINVAL; 4323a0111cf6SLai Jiangshan 4324a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 43250a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 43260a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4327a0111cf6SLai Jiangshan return -EINVAL; 4328a0111cf6SLai Jiangshan 43290a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 43300a94efb5STejun Heo } 43310a94efb5STejun Heo 433299c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 43336201171eSwanghaibin if (!ctx) 43346201171eSwanghaibin return -ENOMEM; 4335a0111cf6SLai Jiangshan 4336a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4337a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4338a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4339a0111cf6SLai Jiangshan 43406201171eSwanghaibin return 0; 4341a0111cf6SLai Jiangshan } 4342a0111cf6SLai Jiangshan 43439e8cd2f5STejun Heo /** 43449e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 43459e8cd2f5STejun Heo * @wq: the target workqueue 43469e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 43479e8cd2f5STejun Heo * 4348fef59c9cSTejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 4349fef59c9cSTejun Heo * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 4350fef59c9cSTejun Heo * work items are affine to the pod it was issued on. Older pwqs are released as 4351fef59c9cSTejun Heo * in-flight work items finish. Note that a work item which repeatedly requeues 4352fef59c9cSTejun Heo * itself back-to-back will stay on its current pwq. 43539e8cd2f5STejun Heo * 4354d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4355d185af30SYacine Belkadi * 4356ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4357509b3204SDaniel Jordan * 4358d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 43599e8cd2f5STejun Heo */ 4360513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 43619e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 43629e8cd2f5STejun Heo { 4363a0111cf6SLai Jiangshan int ret; 43649e8cd2f5STejun Heo 4365509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4366509b3204SDaniel Jordan 4367509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4368a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4369509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 43702d5f0764SLai Jiangshan 43712d5f0764SLai Jiangshan return ret; 43729e8cd2f5STejun Heo } 43739e8cd2f5STejun Heo 43744c16bd32STejun Heo /** 4375fef59c9cSTejun Heo * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug 43764c16bd32STejun Heo * @wq: the target workqueue 43774cbfd3deSTejun Heo * @cpu: the CPU to update pool association for 43784cbfd3deSTejun Heo * @hotplug_cpu: the CPU coming up or going down 43794c16bd32STejun Heo * @online: whether @cpu is coming up or going down 43804c16bd32STejun Heo * 43814c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 4382fef59c9cSTejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of 43834c16bd32STejun Heo * @wq accordingly. 43844c16bd32STejun Heo * 43854c16bd32STejun Heo * 4386fef59c9cSTejun Heo * If pod affinity can't be adjusted due to memory allocation failure, it falls 4387fef59c9cSTejun Heo * back to @wq->dfl_pwq which may not be optimal but is always correct. 4388fef59c9cSTejun Heo * 4389fef59c9cSTejun Heo * Note that when the last allowed CPU of a pod goes offline for a workqueue 4390fef59c9cSTejun Heo * with a cpumask spanning multiple pods, the workers which were already 4391fef59c9cSTejun Heo * executing the work items for the workqueue will lose their CPU affinity and 4392fef59c9cSTejun Heo * may execute on any CPU. This is similar to how per-cpu workqueues behave on 4393fef59c9cSTejun Heo * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 4394fef59c9cSTejun Heo * responsibility to flush the work item from CPU_DOWN_PREPARE. 43954c16bd32STejun Heo */ 4396fef59c9cSTejun Heo static void wq_update_pod(struct workqueue_struct *wq, int cpu, 43974cbfd3deSTejun Heo int hotplug_cpu, bool online) 43984c16bd32STejun Heo { 4399fef59c9cSTejun Heo int pod = cpu_to_node(cpu); 44004cbfd3deSTejun Heo int off_cpu = online ? -1 : hotplug_cpu; 44014c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 44024c16bd32STejun Heo struct workqueue_attrs *target_attrs; 44034c16bd32STejun Heo cpumask_t *cpumask; 44044c16bd32STejun Heo 44054c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 44064c16bd32STejun Heo 4407fef59c9cSTejun Heo if (!wq_pod_enabled || !(wq->flags & WQ_UNBOUND) || 4408af73f5c9STejun Heo wq->unbound_attrs->ordered) 44094c16bd32STejun Heo return; 44104c16bd32STejun Heo 44114c16bd32STejun Heo /* 44124c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 44134c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 44144c16bd32STejun Heo * CPU hotplug exclusion. 44154c16bd32STejun Heo */ 4416fef59c9cSTejun Heo target_attrs = wq_update_pod_attrs_buf; 44174c16bd32STejun Heo cpumask = target_attrs->cpumask; 44184c16bd32STejun Heo 44194c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 44204c16bd32STejun Heo 4421636b927eSTejun Heo /* nothing to do if the target cpumask matches the current pwq */ 4422fef59c9cSTejun Heo wq_calc_pod_cpumask(wq->dfl_pwq->pool->attrs, pod, off_cpu, cpumask); 4423636b927eSTejun Heo pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu), 4424636b927eSTejun Heo lockdep_is_held(&wq_pool_mutex)); 44254c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4426f7142ed4SLai Jiangshan return; 44274c16bd32STejun Heo 44284c16bd32STejun Heo /* create a new pwq */ 44294c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 44304c16bd32STejun Heo if (!pwq) { 4431fef59c9cSTejun Heo pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 44324c16bd32STejun Heo wq->name); 443377f300b1SDaeseok Youn goto use_dfl_pwq; 44344c16bd32STejun Heo } 44354c16bd32STejun Heo 4436f7142ed4SLai Jiangshan /* Install the new pwq. */ 44374c16bd32STejun Heo mutex_lock(&wq->mutex); 4438636b927eSTejun Heo old_pwq = install_unbound_pwq(wq, cpu, pwq); 44394c16bd32STejun Heo goto out_unlock; 44404c16bd32STejun Heo 44414c16bd32STejun Heo use_dfl_pwq: 4442f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4443a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 44444c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4445a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 4446636b927eSTejun Heo old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq); 44474c16bd32STejun Heo out_unlock: 44484c16bd32STejun Heo mutex_unlock(&wq->mutex); 44494c16bd32STejun Heo put_pwq_unlocked(old_pwq); 44504c16bd32STejun Heo } 44514c16bd32STejun Heo 445230cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 44531da177e4SLinus Torvalds { 445449e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 44558a2b7538STejun Heo int cpu, ret; 4456e1d8aa9fSFrederic Weisbecker 4457687a9aa5STejun Heo wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 4458ee1ceef7STejun Heo if (!wq->cpu_pwq) 4459687a9aa5STejun Heo goto enomem; 446030cdf249STejun Heo 4461636b927eSTejun Heo if (!(wq->flags & WQ_UNBOUND)) { 446230cdf249STejun Heo for_each_possible_cpu(cpu) { 4463687a9aa5STejun Heo struct pool_workqueue **pwq_p = 4464ee1ceef7STejun Heo per_cpu_ptr(wq->cpu_pwq, cpu); 4465687a9aa5STejun Heo struct worker_pool *pool = 4466687a9aa5STejun Heo &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); 446730cdf249STejun Heo 4468687a9aa5STejun Heo *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 4469687a9aa5STejun Heo pool->node); 4470687a9aa5STejun Heo if (!*pwq_p) 4471687a9aa5STejun Heo goto enomem; 4472687a9aa5STejun Heo 4473687a9aa5STejun Heo init_pwq(*pwq_p, wq, pool); 4474f147f29eSTejun Heo 4475f147f29eSTejun Heo mutex_lock(&wq->mutex); 4476687a9aa5STejun Heo link_pwq(*pwq_p); 4477f147f29eSTejun Heo mutex_unlock(&wq->mutex); 447830cdf249STejun Heo } 447930cdf249STejun Heo return 0; 4480509b3204SDaniel Jordan } 4481509b3204SDaniel Jordan 4482ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4483509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 44848a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 44858a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 44868a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 44878a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 44888a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 44899e8cd2f5STejun Heo } else { 4490509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 44919e8cd2f5STejun Heo } 4492ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4493509b3204SDaniel Jordan 4494509b3204SDaniel Jordan return ret; 4495687a9aa5STejun Heo 4496687a9aa5STejun Heo enomem: 4497687a9aa5STejun Heo if (wq->cpu_pwq) { 4498687a9aa5STejun Heo for_each_possible_cpu(cpu) 4499687a9aa5STejun Heo kfree(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4500687a9aa5STejun Heo free_percpu(wq->cpu_pwq); 4501687a9aa5STejun Heo wq->cpu_pwq = NULL; 4502687a9aa5STejun Heo } 4503687a9aa5STejun Heo return -ENOMEM; 45040f900049STejun Heo } 45050f900049STejun Heo 4506f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4507f3421797STejun Heo const char *name) 4508b71ab8c2STejun Heo { 4509636b927eSTejun Heo if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 4510044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4511636b927eSTejun Heo max_active, name, 1, WQ_MAX_ACTIVE); 4512b71ab8c2STejun Heo 4513636b927eSTejun Heo return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 4514b71ab8c2STejun Heo } 4515b71ab8c2STejun Heo 4516983c7515STejun Heo /* 4517983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4518983c7515STejun Heo * to guarantee forward progress. 4519983c7515STejun Heo */ 4520983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4521983c7515STejun Heo { 4522983c7515STejun Heo struct worker *rescuer; 4523b92b36eaSDan Carpenter int ret; 4524983c7515STejun Heo 4525983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4526983c7515STejun Heo return 0; 4527983c7515STejun Heo 4528983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 45294c0736a7SPetr Mladek if (!rescuer) { 45304c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 45314c0736a7SPetr Mladek wq->name); 4532983c7515STejun Heo return -ENOMEM; 45334c0736a7SPetr Mladek } 4534983c7515STejun Heo 4535983c7515STejun Heo rescuer->rescue_wq = wq; 4536983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4537f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4538b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 45394c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 45404c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4541983c7515STejun Heo kfree(rescuer); 4542b92b36eaSDan Carpenter return ret; 4543983c7515STejun Heo } 4544983c7515STejun Heo 4545983c7515STejun Heo wq->rescuer = rescuer; 4546983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4547983c7515STejun Heo wake_up_process(rescuer->task); 4548983c7515STejun Heo 4549983c7515STejun Heo return 0; 4550983c7515STejun Heo } 4551983c7515STejun Heo 4552a2775bbcSMathieu Malaterre __printf(1, 4) 4553669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 455497e37d7bSTejun Heo unsigned int flags, 4555669de8bdSBart Van Assche int max_active, ...) 45563af24433SOleg Nesterov { 4557ecf6881fSTejun Heo va_list args; 45583af24433SOleg Nesterov struct workqueue_struct *wq; 455949e3cf44STejun Heo struct pool_workqueue *pwq; 4560b196be89STejun Heo 45615c0338c6STejun Heo /* 4562fef59c9cSTejun Heo * Unbound && max_active == 1 used to imply ordered, which is no longer 4563fef59c9cSTejun Heo * the case on many machines due to per-pod pools. While 45645c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 4565fef59c9cSTejun Heo * workqueue, keep the previous behavior to avoid subtle breakages. 45665c0338c6STejun Heo */ 45675c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 45685c0338c6STejun Heo flags |= __WQ_ORDERED; 45695c0338c6STejun Heo 4570cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4571cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4572cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4573cee22a15SViresh Kumar 4574ecf6881fSTejun Heo /* allocate wq and format name */ 4575636b927eSTejun Heo wq = kzalloc(sizeof(*wq), GFP_KERNEL); 4576b196be89STejun Heo if (!wq) 4577d2c1d404STejun Heo return NULL; 4578b196be89STejun Heo 45796029a918STejun Heo if (flags & WQ_UNBOUND) { 4580be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 45816029a918STejun Heo if (!wq->unbound_attrs) 45826029a918STejun Heo goto err_free_wq; 45836029a918STejun Heo } 45846029a918STejun Heo 4585669de8bdSBart Van Assche va_start(args, max_active); 4586ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4587b196be89STejun Heo va_end(args); 45883af24433SOleg Nesterov 4589d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4590b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 45913af24433SOleg Nesterov 4592b196be89STejun Heo /* init wq */ 459397e37d7bSTejun Heo wq->flags = flags; 4594a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 45953c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4596112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 459730cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 459873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 459973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4600493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 46013af24433SOleg Nesterov 4602669de8bdSBart Van Assche wq_init_lockdep(wq); 4603cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 46043af24433SOleg Nesterov 460530cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 460682efcab3SBart Van Assche goto err_unreg_lockdep; 46071537663fSTejun Heo 460840c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4609d2c1d404STejun Heo goto err_destroy; 4610e22bee78STejun Heo 4611226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4612226223abSTejun Heo goto err_destroy; 4613226223abSTejun Heo 46146af8bf3dSOleg Nesterov /* 461568e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 461668e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 461768e13a67SLai Jiangshan * list. 46186af8bf3dSOleg Nesterov */ 461968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4620a0a1a5fdSTejun Heo 4621a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 462249e3cf44STejun Heo for_each_pwq(pwq, wq) 4623699ce097STejun Heo pwq_adjust_max_active(pwq); 4624a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4625a0a1a5fdSTejun Heo 4626e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4627a0a1a5fdSTejun Heo 462868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46293af24433SOleg Nesterov 46303af24433SOleg Nesterov return wq; 4631d2c1d404STejun Heo 463282efcab3SBart Van Assche err_unreg_lockdep: 4633009bb421SBart Van Assche wq_unregister_lockdep(wq); 4634009bb421SBart Van Assche wq_free_lockdep(wq); 463582efcab3SBart Van Assche err_free_wq: 46366029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 46374690c4abSTejun Heo kfree(wq); 4638d2c1d404STejun Heo return NULL; 4639d2c1d404STejun Heo err_destroy: 4640d2c1d404STejun Heo destroy_workqueue(wq); 46414690c4abSTejun Heo return NULL; 46421da177e4SLinus Torvalds } 4643669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 46441da177e4SLinus Torvalds 4645c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4646c29eb853STejun Heo { 4647c29eb853STejun Heo int i; 4648c29eb853STejun Heo 4649c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4650c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4651c29eb853STejun Heo return true; 4652c29eb853STejun Heo 4653c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4654c29eb853STejun Heo return true; 4655f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4656c29eb853STejun Heo return true; 4657c29eb853STejun Heo 4658c29eb853STejun Heo return false; 4659c29eb853STejun Heo } 4660c29eb853STejun Heo 46613af24433SOleg Nesterov /** 46623af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 46633af24433SOleg Nesterov * @wq: target workqueue 46643af24433SOleg Nesterov * 46653af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 46663af24433SOleg Nesterov */ 46673af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 46683af24433SOleg Nesterov { 466949e3cf44STejun Heo struct pool_workqueue *pwq; 4670636b927eSTejun Heo int cpu; 46713af24433SOleg Nesterov 4672def98c84STejun Heo /* 4673def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4674def98c84STejun Heo * lead to sysfs name conflicts. 4675def98c84STejun Heo */ 4676def98c84STejun Heo workqueue_sysfs_unregister(wq); 4677def98c84STejun Heo 467833e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 467933e3f0a3SRichard Clark mutex_lock(&wq->mutex); 468033e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 468133e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 468233e3f0a3SRichard Clark 46839c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 46849c5a2ba7STejun Heo drain_workqueue(wq); 4685c8efcc25STejun Heo 4686def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4687def98c84STejun Heo if (wq->rescuer) { 4688def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4689def98c84STejun Heo 4690def98c84STejun Heo /* this prevents new queueing */ 4691a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4692def98c84STejun Heo wq->rescuer = NULL; 4693a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4694def98c84STejun Heo 4695def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4696def98c84STejun Heo kthread_stop(rescuer->task); 46978efe1223STejun Heo kfree(rescuer); 4698def98c84STejun Heo } 4699def98c84STejun Heo 4700c29eb853STejun Heo /* 4701c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4702c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4703c29eb853STejun Heo */ 4704c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4705b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 470649e3cf44STejun Heo for_each_pwq(pwq, wq) { 4707a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4708c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 47091d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4710e66b39afSTejun Heo __func__, wq->name); 4711c29eb853STejun Heo show_pwq(pwq); 4712a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4713b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4714c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 471555df0933SImran Khan show_one_workqueue(wq); 47166183c009STejun Heo return; 471776af4d93STejun Heo } 4718a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 471976af4d93STejun Heo } 4720b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 47216183c009STejun Heo 4722a0a1a5fdSTejun Heo /* 4723a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4724a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4725a0a1a5fdSTejun Heo */ 4726e2dca7adSTejun Heo list_del_rcu(&wq->list); 472768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47283af24433SOleg Nesterov 47298864b4e5STejun Heo /* 4730636b927eSTejun Heo * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 4731636b927eSTejun Heo * to put the base refs. @wq will be auto-destroyed from the last 4732636b927eSTejun Heo * pwq_put. RCU read lock prevents @wq from going away from under us. 47338864b4e5STejun Heo */ 4734636b927eSTejun Heo rcu_read_lock(); 4735636b927eSTejun Heo 4736636b927eSTejun Heo for_each_possible_cpu(cpu) { 4737636b927eSTejun Heo pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4738636b927eSTejun Heo RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL); 47394c16bd32STejun Heo put_pwq_unlocked(pwq); 47404c16bd32STejun Heo } 47414c16bd32STejun Heo 4742636b927eSTejun Heo put_pwq_unlocked(wq->dfl_pwq); 47434c16bd32STejun Heo wq->dfl_pwq = NULL; 4744636b927eSTejun Heo 4745636b927eSTejun Heo rcu_read_unlock(); 47463af24433SOleg Nesterov } 47473af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 47483af24433SOleg Nesterov 4749dcd989cbSTejun Heo /** 4750dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4751dcd989cbSTejun Heo * @wq: target workqueue 4752dcd989cbSTejun Heo * @max_active: new max_active value. 4753dcd989cbSTejun Heo * 4754dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4755dcd989cbSTejun Heo * 4756dcd989cbSTejun Heo * CONTEXT: 4757dcd989cbSTejun Heo * Don't call from IRQ context. 4758dcd989cbSTejun Heo */ 4759dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4760dcd989cbSTejun Heo { 476149e3cf44STejun Heo struct pool_workqueue *pwq; 4762dcd989cbSTejun Heo 47638719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 47640a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 47658719dceaSTejun Heo return; 47668719dceaSTejun Heo 4767f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4768dcd989cbSTejun Heo 4769a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4770dcd989cbSTejun Heo 47710a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4772dcd989cbSTejun Heo wq->saved_max_active = max_active; 4773dcd989cbSTejun Heo 4774699ce097STejun Heo for_each_pwq(pwq, wq) 4775699ce097STejun Heo pwq_adjust_max_active(pwq); 4776dcd989cbSTejun Heo 4777a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4778dcd989cbSTejun Heo } 4779dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4780dcd989cbSTejun Heo 4781dcd989cbSTejun Heo /** 478227d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 478327d4ee03SLukas Wunner * 478427d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 478527d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 478627d4ee03SLukas Wunner * 478727d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 478827d4ee03SLukas Wunner */ 478927d4ee03SLukas Wunner struct work_struct *current_work(void) 479027d4ee03SLukas Wunner { 479127d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 479227d4ee03SLukas Wunner 479327d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 479427d4ee03SLukas Wunner } 479527d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 479627d4ee03SLukas Wunner 479727d4ee03SLukas Wunner /** 4798e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4799e6267616STejun Heo * 4800e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4801e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4802d185af30SYacine Belkadi * 4803d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4804e6267616STejun Heo */ 4805e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4806e6267616STejun Heo { 4807e6267616STejun Heo struct worker *worker = current_wq_worker(); 4808e6267616STejun Heo 48096a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4810e6267616STejun Heo } 4811e6267616STejun Heo 4812e6267616STejun Heo /** 4813dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4814dcd989cbSTejun Heo * @cpu: CPU in question 4815dcd989cbSTejun Heo * @wq: target workqueue 4816dcd989cbSTejun Heo * 4817dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4818dcd989cbSTejun Heo * no synchronization around this function and the test result is 4819dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4820dcd989cbSTejun Heo * 4821d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4822636b927eSTejun Heo * 4823636b927eSTejun Heo * With the exception of ordered workqueues, all workqueues have per-cpu 4824636b927eSTejun Heo * pool_workqueues, each with its own congested state. A workqueue being 4825636b927eSTejun Heo * congested on one CPU doesn't mean that the workqueue is contested on any 4826636b927eSTejun Heo * other CPUs. 4827d3251859STejun Heo * 4828d185af30SYacine Belkadi * Return: 4829dcd989cbSTejun Heo * %true if congested, %false otherwise. 4830dcd989cbSTejun Heo */ 4831d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4832dcd989cbSTejun Heo { 48337fb98ea7STejun Heo struct pool_workqueue *pwq; 483476af4d93STejun Heo bool ret; 483576af4d93STejun Heo 483624acfb71SThomas Gleixner rcu_read_lock(); 483724acfb71SThomas Gleixner preempt_disable(); 48387fb98ea7STejun Heo 4839d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4840d3251859STejun Heo cpu = smp_processor_id(); 4841d3251859STejun Heo 4842687a9aa5STejun Heo pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4843f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 4844636b927eSTejun Heo 484524acfb71SThomas Gleixner preempt_enable(); 484624acfb71SThomas Gleixner rcu_read_unlock(); 484776af4d93STejun Heo 484876af4d93STejun Heo return ret; 4849dcd989cbSTejun Heo } 4850dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4851dcd989cbSTejun Heo 4852dcd989cbSTejun Heo /** 4853dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4854dcd989cbSTejun Heo * @work: the work to be tested 4855dcd989cbSTejun Heo * 4856dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4857dcd989cbSTejun Heo * synchronization around this function and the test result is 4858dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4859dcd989cbSTejun Heo * 4860d185af30SYacine Belkadi * Return: 4861dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4862dcd989cbSTejun Heo */ 4863dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4864dcd989cbSTejun Heo { 4865fa1b54e6STejun Heo struct worker_pool *pool; 4866dcd989cbSTejun Heo unsigned long flags; 4867dcd989cbSTejun Heo unsigned int ret = 0; 4868dcd989cbSTejun Heo 4869dcd989cbSTejun Heo if (work_pending(work)) 4870dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4871038366c5SLai Jiangshan 487224acfb71SThomas Gleixner rcu_read_lock(); 4873fa1b54e6STejun Heo pool = get_work_pool(work); 4874038366c5SLai Jiangshan if (pool) { 4875a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4876c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4877dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4878a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4879038366c5SLai Jiangshan } 488024acfb71SThomas Gleixner rcu_read_unlock(); 4881dcd989cbSTejun Heo 4882dcd989cbSTejun Heo return ret; 4883dcd989cbSTejun Heo } 4884dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4885dcd989cbSTejun Heo 48863d1cb205STejun Heo /** 48873d1cb205STejun Heo * set_worker_desc - set description for the current work item 48883d1cb205STejun Heo * @fmt: printf-style format string 48893d1cb205STejun Heo * @...: arguments for the format string 48903d1cb205STejun Heo * 48913d1cb205STejun Heo * This function can be called by a running work function to describe what 48923d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 48933d1cb205STejun Heo * information will be printed out together to help debugging. The 48943d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 48953d1cb205STejun Heo */ 48963d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 48973d1cb205STejun Heo { 48983d1cb205STejun Heo struct worker *worker = current_wq_worker(); 48993d1cb205STejun Heo va_list args; 49003d1cb205STejun Heo 49013d1cb205STejun Heo if (worker) { 49023d1cb205STejun Heo va_start(args, fmt); 49033d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 49043d1cb205STejun Heo va_end(args); 49053d1cb205STejun Heo } 49063d1cb205STejun Heo } 49075c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 49083d1cb205STejun Heo 49093d1cb205STejun Heo /** 49103d1cb205STejun Heo * print_worker_info - print out worker information and description 49113d1cb205STejun Heo * @log_lvl: the log level to use when printing 49123d1cb205STejun Heo * @task: target task 49133d1cb205STejun Heo * 49143d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 49153d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 49163d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 49173d1cb205STejun Heo * 49183d1cb205STejun Heo * This function can be safely called on any task as long as the 49193d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 49203d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 49213d1cb205STejun Heo */ 49223d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 49233d1cb205STejun Heo { 49243d1cb205STejun Heo work_func_t *fn = NULL; 49253d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 49263d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 49273d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 49283d1cb205STejun Heo struct workqueue_struct *wq = NULL; 49293d1cb205STejun Heo struct worker *worker; 49303d1cb205STejun Heo 49313d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 49323d1cb205STejun Heo return; 49333d1cb205STejun Heo 49343d1cb205STejun Heo /* 49353d1cb205STejun Heo * This function is called without any synchronization and @task 49363d1cb205STejun Heo * could be in any state. Be careful with dereferences. 49373d1cb205STejun Heo */ 4938e700591aSPetr Mladek worker = kthread_probe_data(task); 49393d1cb205STejun Heo 49403d1cb205STejun Heo /* 49418bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 49428bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 49433d1cb205STejun Heo */ 4944fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4945fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4946fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4947fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4948fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 49493d1cb205STejun Heo 49503d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4951d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 49528bf89593STejun Heo if (strcmp(name, desc)) 49533d1cb205STejun Heo pr_cont(" (%s)", desc); 49543d1cb205STejun Heo pr_cont("\n"); 49553d1cb205STejun Heo } 49563d1cb205STejun Heo } 49573d1cb205STejun Heo 49583494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 49593494fc30STejun Heo { 49603494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 49613494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 49623494fc30STejun Heo pr_cont(" node=%d", pool->node); 49633494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 49643494fc30STejun Heo } 49653494fc30STejun Heo 4966c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 4967c76feb0dSPaul E. McKenney bool comma; 4968c76feb0dSPaul E. McKenney work_func_t func; 4969c76feb0dSPaul E. McKenney long ctr; 4970c76feb0dSPaul E. McKenney }; 4971c76feb0dSPaul E. McKenney 4972c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 4973c76feb0dSPaul E. McKenney { 4974c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 4975c76feb0dSPaul E. McKenney goto out_record; 4976c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 4977c76feb0dSPaul E. McKenney pcwsp->ctr++; 4978c76feb0dSPaul E. McKenney return; 4979c76feb0dSPaul E. McKenney } 4980c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 4981c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 4982c76feb0dSPaul E. McKenney else 4983c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 4984c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 4985c76feb0dSPaul E. McKenney out_record: 4986c76feb0dSPaul E. McKenney if ((long)func == -1L) 4987c76feb0dSPaul E. McKenney return; 4988c76feb0dSPaul E. McKenney pcwsp->comma = comma; 4989c76feb0dSPaul E. McKenney pcwsp->func = func; 4990c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 4991c76feb0dSPaul E. McKenney } 4992c76feb0dSPaul E. McKenney 4993c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 49943494fc30STejun Heo { 49953494fc30STejun Heo if (work->func == wq_barrier_func) { 49963494fc30STejun Heo struct wq_barrier *barr; 49973494fc30STejun Heo 49983494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 49993494fc30STejun Heo 5000c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 50013494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 50023494fc30STejun Heo task_pid_nr(barr->task)); 50033494fc30STejun Heo } else { 5004c76feb0dSPaul E. McKenney if (!comma) 5005c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5006c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 50073494fc30STejun Heo } 50083494fc30STejun Heo } 50093494fc30STejun Heo 50103494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 50113494fc30STejun Heo { 5012c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 50133494fc30STejun Heo struct worker_pool *pool = pwq->pool; 50143494fc30STejun Heo struct work_struct *work; 50153494fc30STejun Heo struct worker *worker; 50163494fc30STejun Heo bool has_in_flight = false, has_pending = false; 50173494fc30STejun Heo int bkt; 50183494fc30STejun Heo 50193494fc30STejun Heo pr_info(" pwq %d:", pool->id); 50203494fc30STejun Heo pr_cont_pool_info(pool); 50213494fc30STejun Heo 5022e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5023e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 50243494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 50253494fc30STejun Heo 50263494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 50273494fc30STejun Heo if (worker->current_pwq == pwq) { 50283494fc30STejun Heo has_in_flight = true; 50293494fc30STejun Heo break; 50303494fc30STejun Heo } 50313494fc30STejun Heo } 50323494fc30STejun Heo if (has_in_flight) { 50333494fc30STejun Heo bool comma = false; 50343494fc30STejun Heo 50353494fc30STejun Heo pr_info(" in-flight:"); 50363494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 50373494fc30STejun Heo if (worker->current_pwq != pwq) 50383494fc30STejun Heo continue; 50393494fc30STejun Heo 5040d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 50413494fc30STejun Heo task_pid_nr(worker->task), 504230ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 50433494fc30STejun Heo worker->current_func); 50443494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5045c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5046c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 50473494fc30STejun Heo comma = true; 50483494fc30STejun Heo } 50493494fc30STejun Heo pr_cont("\n"); 50503494fc30STejun Heo } 50513494fc30STejun Heo 50523494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 50533494fc30STejun Heo if (get_work_pwq(work) == pwq) { 50543494fc30STejun Heo has_pending = true; 50553494fc30STejun Heo break; 50563494fc30STejun Heo } 50573494fc30STejun Heo } 50583494fc30STejun Heo if (has_pending) { 50593494fc30STejun Heo bool comma = false; 50603494fc30STejun Heo 50613494fc30STejun Heo pr_info(" pending:"); 50623494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 50633494fc30STejun Heo if (get_work_pwq(work) != pwq) 50643494fc30STejun Heo continue; 50653494fc30STejun Heo 5066c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 50673494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 50683494fc30STejun Heo } 5069c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 50703494fc30STejun Heo pr_cont("\n"); 50713494fc30STejun Heo } 50723494fc30STejun Heo 5073f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 50743494fc30STejun Heo bool comma = false; 50753494fc30STejun Heo 5076f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5077f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5078c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 50793494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 50803494fc30STejun Heo } 5081c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 50823494fc30STejun Heo pr_cont("\n"); 50833494fc30STejun Heo } 50843494fc30STejun Heo } 50853494fc30STejun Heo 50863494fc30STejun Heo /** 508755df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 508855df0933SImran Khan * @wq: workqueue whose state will be printed 50893494fc30STejun Heo */ 509055df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 50913494fc30STejun Heo { 50923494fc30STejun Heo struct pool_workqueue *pwq; 50933494fc30STejun Heo bool idle = true; 509455df0933SImran Khan unsigned long flags; 50953494fc30STejun Heo 50963494fc30STejun Heo for_each_pwq(pwq, wq) { 5097f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 50983494fc30STejun Heo idle = false; 50993494fc30STejun Heo break; 51003494fc30STejun Heo } 51013494fc30STejun Heo } 510255df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 510355df0933SImran Khan return; 51043494fc30STejun Heo 51053494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 51063494fc30STejun Heo 51073494fc30STejun Heo for_each_pwq(pwq, wq) { 5108a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 510957116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 511057116ce1SJohan Hovold /* 511157116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 511257116ce1SJohan Hovold * drivers that queue work while holding locks 511357116ce1SJohan Hovold * also taken in their write paths. 511457116ce1SJohan Hovold */ 511557116ce1SJohan Hovold printk_deferred_enter(); 51163494fc30STejun Heo show_pwq(pwq); 511757116ce1SJohan Hovold printk_deferred_exit(); 511857116ce1SJohan Hovold } 5119a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 512062635ea8SSergey Senozhatsky /* 512162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 512255df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 512362635ea8SSergey Senozhatsky * hard lockup. 512462635ea8SSergey Senozhatsky */ 512562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 51263494fc30STejun Heo } 512755df0933SImran Khan 51283494fc30STejun Heo } 51293494fc30STejun Heo 513055df0933SImran Khan /** 513155df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 513255df0933SImran Khan * @pool: worker pool whose state will be printed 513355df0933SImran Khan */ 513455df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 513555df0933SImran Khan { 51363494fc30STejun Heo struct worker *worker; 51373494fc30STejun Heo bool first = true; 513855df0933SImran Khan unsigned long flags; 5139335a42ebSPetr Mladek unsigned long hung = 0; 51403494fc30STejun Heo 5141a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 51423494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 51433494fc30STejun Heo goto next_pool; 5144335a42ebSPetr Mladek 5145335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5146335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5147335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5148335a42ebSPetr Mladek 514957116ce1SJohan Hovold /* 515057116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 515157116ce1SJohan Hovold * queue work while holding locks also taken in their write 515257116ce1SJohan Hovold * paths. 515357116ce1SJohan Hovold */ 515457116ce1SJohan Hovold printk_deferred_enter(); 51553494fc30STejun Heo pr_info("pool %d:", pool->id); 51563494fc30STejun Heo pr_cont_pool_info(pool); 5157335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 51583494fc30STejun Heo if (pool->manager) 51593494fc30STejun Heo pr_cont(" manager: %d", 51603494fc30STejun Heo task_pid_nr(pool->manager->task)); 51613494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 51623494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 51633494fc30STejun Heo task_pid_nr(worker->task)); 51643494fc30STejun Heo first = false; 51653494fc30STejun Heo } 51663494fc30STejun Heo pr_cont("\n"); 516757116ce1SJohan Hovold printk_deferred_exit(); 51683494fc30STejun Heo next_pool: 5169a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 517062635ea8SSergey Senozhatsky /* 517162635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 517255df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 517362635ea8SSergey Senozhatsky * hard lockup. 517462635ea8SSergey Senozhatsky */ 517562635ea8SSergey Senozhatsky touch_nmi_watchdog(); 517655df0933SImran Khan 51773494fc30STejun Heo } 51783494fc30STejun Heo 517955df0933SImran Khan /** 518055df0933SImran Khan * show_all_workqueues - dump workqueue state 518155df0933SImran Khan * 5182704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 518355df0933SImran Khan */ 518455df0933SImran Khan void show_all_workqueues(void) 518555df0933SImran Khan { 518655df0933SImran Khan struct workqueue_struct *wq; 518755df0933SImran Khan struct worker_pool *pool; 518855df0933SImran Khan int pi; 518955df0933SImran Khan 519055df0933SImran Khan rcu_read_lock(); 519155df0933SImran Khan 519255df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 519355df0933SImran Khan 519455df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 519555df0933SImran Khan show_one_workqueue(wq); 519655df0933SImran Khan 519755df0933SImran Khan for_each_pool(pool, pi) 519855df0933SImran Khan show_one_worker_pool(pool); 519955df0933SImran Khan 520024acfb71SThomas Gleixner rcu_read_unlock(); 52013494fc30STejun Heo } 52023494fc30STejun Heo 5203704bc669SJungseung Lee /** 5204704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5205704bc669SJungseung Lee * 5206704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5207704bc669SJungseung Lee * still busy. 5208704bc669SJungseung Lee */ 5209704bc669SJungseung Lee void show_freezable_workqueues(void) 5210704bc669SJungseung Lee { 5211704bc669SJungseung Lee struct workqueue_struct *wq; 5212704bc669SJungseung Lee 5213704bc669SJungseung Lee rcu_read_lock(); 5214704bc669SJungseung Lee 5215704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5216704bc669SJungseung Lee 5217704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5218704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5219704bc669SJungseung Lee continue; 5220704bc669SJungseung Lee show_one_workqueue(wq); 5221704bc669SJungseung Lee } 5222704bc669SJungseung Lee 5223704bc669SJungseung Lee rcu_read_unlock(); 5224704bc669SJungseung Lee } 5225704bc669SJungseung Lee 52266b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 52276b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 52286b59808bSTejun Heo { 52296b59808bSTejun Heo int off; 52306b59808bSTejun Heo 52316b59808bSTejun Heo /* always show the actual comm */ 52326b59808bSTejun Heo off = strscpy(buf, task->comm, size); 52336b59808bSTejun Heo if (off < 0) 52346b59808bSTejun Heo return; 52356b59808bSTejun Heo 5236197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 52376b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 52386b59808bSTejun Heo 5239197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5240197f6accSTejun Heo struct worker *worker = kthread_data(task); 5241197f6accSTejun Heo struct worker_pool *pool = worker->pool; 52426b59808bSTejun Heo 52436b59808bSTejun Heo if (pool) { 5244a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 52456b59808bSTejun Heo /* 5246197f6accSTejun Heo * ->desc tracks information (wq name or 5247197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5248197f6accSTejun Heo * current, prepend '+', otherwise '-'. 52496b59808bSTejun Heo */ 52506b59808bSTejun Heo if (worker->desc[0] != '\0') { 52516b59808bSTejun Heo if (worker->current_work) 52526b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 52536b59808bSTejun Heo worker->desc); 52546b59808bSTejun Heo else 52556b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 52566b59808bSTejun Heo worker->desc); 52576b59808bSTejun Heo } 5258a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 52596b59808bSTejun Heo } 5260197f6accSTejun Heo } 52616b59808bSTejun Heo 52626b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 52636b59808bSTejun Heo } 52646b59808bSTejun Heo 526566448bc2SMathieu Malaterre #ifdef CONFIG_SMP 526666448bc2SMathieu Malaterre 5267db7bccf4STejun Heo /* 5268db7bccf4STejun Heo * CPU hotplug. 5269db7bccf4STejun Heo * 5270e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5271112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5272706026c2STejun Heo * pool which make migrating pending and scheduled works very 5273e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 527494cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5275e22bee78STejun Heo * blocked draining impractical. 5276e22bee78STejun Heo * 527724647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5278628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5279628c78e7STejun Heo * cpu comes back online. 5280db7bccf4STejun Heo */ 5281db7bccf4STejun Heo 5282e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5283db7bccf4STejun Heo { 52844ce62e9eSTejun Heo struct worker_pool *pool; 5285db7bccf4STejun Heo struct worker *worker; 5286db7bccf4STejun Heo 5287f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 52881258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5289a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5290e22bee78STejun Heo 5291f2d5a0eeSTejun Heo /* 529292f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 529394cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 529411b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5295b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5296b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5297b4ac9384SLai Jiangshan * is on the same cpu. 5298f2d5a0eeSTejun Heo */ 5299da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5300403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5301db7bccf4STejun Heo 530224647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5303f2d5a0eeSTejun Heo 5304e22bee78STejun Heo /* 5305989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5306989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5307989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5308989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5309989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5310eb283428SLai Jiangshan * are served by workers tied to the pool. 5311e22bee78STejun Heo */ 5312bc35f7efSLai Jiangshan pool->nr_running = 0; 5313eb283428SLai Jiangshan 5314eb283428SLai Jiangshan /* 5315eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5316eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5317eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5318eb283428SLai Jiangshan */ 5319eb283428SLai Jiangshan wake_up_worker(pool); 5320989442d7SLai Jiangshan 5321a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5322989442d7SLai Jiangshan 5323793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5324793777bcSValentin Schneider unbind_worker(worker); 5325989442d7SLai Jiangshan 5326989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5327eb283428SLai Jiangshan } 5328db7bccf4STejun Heo } 5329db7bccf4STejun Heo 5330bd7c089eSTejun Heo /** 5331bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5332bd7c089eSTejun Heo * @pool: pool of interest 5333bd7c089eSTejun Heo * 5334a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5335bd7c089eSTejun Heo */ 5336bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5337bd7c089eSTejun Heo { 5338a9ab775bSTejun Heo struct worker *worker; 5339bd7c089eSTejun Heo 53401258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5341bd7c089eSTejun Heo 5342bd7c089eSTejun Heo /* 5343a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5344a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5345402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5346a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5347a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5348bd7c089eSTejun Heo */ 5349c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5350c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5351c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5352c63a2e52SValentin Schneider pool->attrs->cpumask) < 0); 5353c63a2e52SValentin Schneider } 5354a9ab775bSTejun Heo 5355a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5356f7c17d26SWanpeng Li 53573de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5358a9ab775bSTejun Heo 5359da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5360a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5361a9ab775bSTejun Heo 5362a9ab775bSTejun Heo /* 5363a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5364a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5365a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5366a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5367a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5368a9ab775bSTejun Heo * concurrency management. Note that when or whether 5369a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5370a9ab775bSTejun Heo * 5371c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5372a9ab775bSTejun Heo * tested without holding any lock in 53736d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5374a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5375a9ab775bSTejun Heo * management operations. 5376bd7c089eSTejun Heo */ 5377a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5378a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5379a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5380c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5381bd7c089eSTejun Heo } 5382a9ab775bSTejun Heo 5383a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5384bd7c089eSTejun Heo } 5385bd7c089eSTejun Heo 53867dbc725eSTejun Heo /** 53877dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 53887dbc725eSTejun Heo * @pool: unbound pool of interest 53897dbc725eSTejun Heo * @cpu: the CPU which is coming up 53907dbc725eSTejun Heo * 53917dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 53927dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 53937dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 53947dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 53957dbc725eSTejun Heo */ 53967dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 53977dbc725eSTejun Heo { 53987dbc725eSTejun Heo static cpumask_t cpumask; 53997dbc725eSTejun Heo struct worker *worker; 54007dbc725eSTejun Heo 54011258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 54027dbc725eSTejun Heo 54037dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 54047dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 54057dbc725eSTejun Heo return; 54067dbc725eSTejun Heo 54077dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 54087dbc725eSTejun Heo 54097dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5410da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5411d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 54127dbc725eSTejun Heo } 54137dbc725eSTejun Heo 54147ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 54151da177e4SLinus Torvalds { 54164ce62e9eSTejun Heo struct worker_pool *pool; 54171da177e4SLinus Torvalds 5418f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54193ce63377STejun Heo if (pool->nr_workers) 54203ce63377STejun Heo continue; 5421051e1850SLai Jiangshan if (!create_worker(pool)) 54227ee681b2SThomas Gleixner return -ENOMEM; 54233af24433SOleg Nesterov } 54247ee681b2SThomas Gleixner return 0; 54257ee681b2SThomas Gleixner } 54261da177e4SLinus Torvalds 54277ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 54287ee681b2SThomas Gleixner { 54297ee681b2SThomas Gleixner struct worker_pool *pool; 54307ee681b2SThomas Gleixner struct workqueue_struct *wq; 54317ee681b2SThomas Gleixner int pi; 54327ee681b2SThomas Gleixner 543368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 54347dbc725eSTejun Heo 54357dbc725eSTejun Heo for_each_pool(pool, pi) { 54361258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 543794cf58bbSTejun Heo 5438f05b558dSLai Jiangshan if (pool->cpu == cpu) 543994cf58bbSTejun Heo rebind_workers(pool); 5440f05b558dSLai Jiangshan else if (pool->cpu < 0) 54417dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 544294cf58bbSTejun Heo 54431258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 544494cf58bbSTejun Heo } 54457dbc725eSTejun Heo 5446fef59c9cSTejun Heo /* update pod affinity of unbound workqueues */ 54474cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 54484cbfd3deSTejun Heo int tcpu; 54494cbfd3deSTejun Heo 54504cbfd3deSTejun Heo for_each_possible_cpu(tcpu) { 54514cbfd3deSTejun Heo if (cpu_to_node(tcpu) == cpu_to_node(cpu)) { 5452fef59c9cSTejun Heo wq_update_pod(wq, tcpu, cpu, true); 54534cbfd3deSTejun Heo } 54544cbfd3deSTejun Heo } 54554cbfd3deSTejun Heo } 54564c16bd32STejun Heo 545768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 54587ee681b2SThomas Gleixner return 0; 545965758202STejun Heo } 546065758202STejun Heo 54617ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 546265758202STejun Heo { 54634c16bd32STejun Heo struct workqueue_struct *wq; 54648db25e78STejun Heo 54654c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5466e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5467e8b3f8dbSLai Jiangshan return -1; 5468e8b3f8dbSLai Jiangshan 5469e8b3f8dbSLai Jiangshan unbind_workers(cpu); 54704c16bd32STejun Heo 5471fef59c9cSTejun Heo /* update pod affinity of unbound workqueues */ 54724c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 54734cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 54744cbfd3deSTejun Heo int tcpu; 54754cbfd3deSTejun Heo 54764cbfd3deSTejun Heo for_each_possible_cpu(tcpu) { 54774cbfd3deSTejun Heo if (cpu_to_node(tcpu) == cpu_to_node(cpu)) { 5478fef59c9cSTejun Heo wq_update_pod(wq, tcpu, cpu, false); 54794cbfd3deSTejun Heo } 54804cbfd3deSTejun Heo } 54814cbfd3deSTejun Heo } 54824c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 54834c16bd32STejun Heo 54847ee681b2SThomas Gleixner return 0; 548565758202STejun Heo } 548665758202STejun Heo 54872d3854a3SRusty Russell struct work_for_cpu { 5488ed48ece2STejun Heo struct work_struct work; 54892d3854a3SRusty Russell long (*fn)(void *); 54902d3854a3SRusty Russell void *arg; 54912d3854a3SRusty Russell long ret; 54922d3854a3SRusty Russell }; 54932d3854a3SRusty Russell 5494ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 54952d3854a3SRusty Russell { 5496ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5497ed48ece2STejun Heo 54982d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 54992d3854a3SRusty Russell } 55002d3854a3SRusty Russell 55012d3854a3SRusty Russell /** 550222aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 55032d3854a3SRusty Russell * @cpu: the cpu to run on 55042d3854a3SRusty Russell * @fn: the function to run 55052d3854a3SRusty Russell * @arg: the function arg 55062d3854a3SRusty Russell * 550731ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 55086b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5509d185af30SYacine Belkadi * 5510d185af30SYacine Belkadi * Return: The value @fn returns. 55112d3854a3SRusty Russell */ 5512d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 55132d3854a3SRusty Russell { 5514ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 55152d3854a3SRusty Russell 5516ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5517ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 551812997d1aSBjorn Helgaas flush_work(&wfc.work); 5519440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 55202d3854a3SRusty Russell return wfc.ret; 55212d3854a3SRusty Russell } 55222d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 55230e8d6a93SThomas Gleixner 55240e8d6a93SThomas Gleixner /** 55250e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 55260e8d6a93SThomas Gleixner * @cpu: the cpu to run on 55270e8d6a93SThomas Gleixner * @fn: the function to run 55280e8d6a93SThomas Gleixner * @arg: the function argument 55290e8d6a93SThomas Gleixner * 55300e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 55310e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 55320e8d6a93SThomas Gleixner * 55330e8d6a93SThomas Gleixner * Return: The value @fn returns. 55340e8d6a93SThomas Gleixner */ 55350e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 55360e8d6a93SThomas Gleixner { 55370e8d6a93SThomas Gleixner long ret = -ENODEV; 55380e8d6a93SThomas Gleixner 5539ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 55400e8d6a93SThomas Gleixner if (cpu_online(cpu)) 55410e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5542ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 55430e8d6a93SThomas Gleixner return ret; 55440e8d6a93SThomas Gleixner } 55450e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 55462d3854a3SRusty Russell #endif /* CONFIG_SMP */ 55472d3854a3SRusty Russell 5548a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5549e7577c50SRusty Russell 5550a0a1a5fdSTejun Heo /** 5551a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5552a0a1a5fdSTejun Heo * 555358a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5554f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5555706026c2STejun Heo * pool->worklist. 5556a0a1a5fdSTejun Heo * 5557a0a1a5fdSTejun Heo * CONTEXT: 5558a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5559a0a1a5fdSTejun Heo */ 5560a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5561a0a1a5fdSTejun Heo { 556224b8a847STejun Heo struct workqueue_struct *wq; 556324b8a847STejun Heo struct pool_workqueue *pwq; 5564a0a1a5fdSTejun Heo 556568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5566a0a1a5fdSTejun Heo 55676183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5568a0a1a5fdSTejun Heo workqueue_freezing = true; 5569a0a1a5fdSTejun Heo 557024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5571a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5572699ce097STejun Heo for_each_pwq(pwq, wq) 5573699ce097STejun Heo pwq_adjust_max_active(pwq); 5574a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5575a1056305STejun Heo } 55765bcab335STejun Heo 557768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5578a0a1a5fdSTejun Heo } 5579a0a1a5fdSTejun Heo 5580a0a1a5fdSTejun Heo /** 558158a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5582a0a1a5fdSTejun Heo * 5583a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5584a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5585a0a1a5fdSTejun Heo * 5586a0a1a5fdSTejun Heo * CONTEXT: 558768e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5588a0a1a5fdSTejun Heo * 5589d185af30SYacine Belkadi * Return: 559058a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 559158a69cb4STejun Heo * is complete. 5592a0a1a5fdSTejun Heo */ 5593a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5594a0a1a5fdSTejun Heo { 5595a0a1a5fdSTejun Heo bool busy = false; 559624b8a847STejun Heo struct workqueue_struct *wq; 559724b8a847STejun Heo struct pool_workqueue *pwq; 5598a0a1a5fdSTejun Heo 559968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5600a0a1a5fdSTejun Heo 56016183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5602a0a1a5fdSTejun Heo 560324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 560424b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 560524b8a847STejun Heo continue; 5606a0a1a5fdSTejun Heo /* 5607a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5608a0a1a5fdSTejun Heo * to peek without lock. 5609a0a1a5fdSTejun Heo */ 561024acfb71SThomas Gleixner rcu_read_lock(); 561124b8a847STejun Heo for_each_pwq(pwq, wq) { 56126183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5613112202d9STejun Heo if (pwq->nr_active) { 5614a0a1a5fdSTejun Heo busy = true; 561524acfb71SThomas Gleixner rcu_read_unlock(); 5616a0a1a5fdSTejun Heo goto out_unlock; 5617a0a1a5fdSTejun Heo } 5618a0a1a5fdSTejun Heo } 561924acfb71SThomas Gleixner rcu_read_unlock(); 5620a0a1a5fdSTejun Heo } 5621a0a1a5fdSTejun Heo out_unlock: 562268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5623a0a1a5fdSTejun Heo return busy; 5624a0a1a5fdSTejun Heo } 5625a0a1a5fdSTejun Heo 5626a0a1a5fdSTejun Heo /** 5627a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5628a0a1a5fdSTejun Heo * 5629a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5630706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5631a0a1a5fdSTejun Heo * 5632a0a1a5fdSTejun Heo * CONTEXT: 5633a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5634a0a1a5fdSTejun Heo */ 5635a0a1a5fdSTejun Heo void thaw_workqueues(void) 5636a0a1a5fdSTejun Heo { 563724b8a847STejun Heo struct workqueue_struct *wq; 563824b8a847STejun Heo struct pool_workqueue *pwq; 5639a0a1a5fdSTejun Heo 564068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5641a0a1a5fdSTejun Heo 5642a0a1a5fdSTejun Heo if (!workqueue_freezing) 5643a0a1a5fdSTejun Heo goto out_unlock; 5644a0a1a5fdSTejun Heo 564574b414eaSLai Jiangshan workqueue_freezing = false; 564624b8a847STejun Heo 564724b8a847STejun Heo /* restore max_active and repopulate worklist */ 564824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5649a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5650699ce097STejun Heo for_each_pwq(pwq, wq) 5651699ce097STejun Heo pwq_adjust_max_active(pwq); 5652a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 565324b8a847STejun Heo } 565424b8a847STejun Heo 5655a0a1a5fdSTejun Heo out_unlock: 565668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5657a0a1a5fdSTejun Heo } 5658a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5659a0a1a5fdSTejun Heo 566099c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5661042f7df1SLai Jiangshan { 5662042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5663042f7df1SLai Jiangshan int ret = 0; 5664042f7df1SLai Jiangshan struct workqueue_struct *wq; 5665042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5666042f7df1SLai Jiangshan 5667042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5668042f7df1SLai Jiangshan 5669042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5670042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5671042f7df1SLai Jiangshan continue; 5672042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5673042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5674042f7df1SLai Jiangshan continue; 5675042f7df1SLai Jiangshan 567699c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5677042f7df1SLai Jiangshan if (!ctx) { 5678042f7df1SLai Jiangshan ret = -ENOMEM; 5679042f7df1SLai Jiangshan break; 5680042f7df1SLai Jiangshan } 5681042f7df1SLai Jiangshan 5682042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5683042f7df1SLai Jiangshan } 5684042f7df1SLai Jiangshan 5685042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5686042f7df1SLai Jiangshan if (!ret) 5687042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5688042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5689042f7df1SLai Jiangshan } 5690042f7df1SLai Jiangshan 569199c621efSLai Jiangshan if (!ret) { 569299c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 569399c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 569499c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 569599c621efSLai Jiangshan } 5696042f7df1SLai Jiangshan return ret; 5697042f7df1SLai Jiangshan } 5698042f7df1SLai Jiangshan 5699042f7df1SLai Jiangshan /** 5700042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5701042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5702042f7df1SLai Jiangshan * 5703042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5704042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5705042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5706042f7df1SLai Jiangshan * 570767dc8325SCai Huoqing * Return: 0 - Success 5708042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5709042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5710042f7df1SLai Jiangshan */ 5711042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5712042f7df1SLai Jiangshan { 5713042f7df1SLai Jiangshan int ret = -EINVAL; 5714042f7df1SLai Jiangshan 5715c98a9805STal Shorer /* 5716c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5717c98a9805STal Shorer * If the user wishes to include them, we allow that. 5718c98a9805STal Shorer */ 5719042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5720042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5721a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5722d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5723d25302e4SMenglong Dong ret = 0; 5724d25302e4SMenglong Dong goto out_unlock; 5725d25302e4SMenglong Dong } 5726d25302e4SMenglong Dong 572799c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5728042f7df1SLai Jiangshan 5729d25302e4SMenglong Dong out_unlock: 5730a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5731042f7df1SLai Jiangshan } 5732042f7df1SLai Jiangshan 5733042f7df1SLai Jiangshan return ret; 5734042f7df1SLai Jiangshan } 5735042f7df1SLai Jiangshan 57366ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 57376ba94429SFrederic Weisbecker /* 57386ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 57396ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 57406ba94429SFrederic Weisbecker * following attributes. 57416ba94429SFrederic Weisbecker * 57426ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 57436ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 57446ba94429SFrederic Weisbecker * 57456ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 57466ba94429SFrederic Weisbecker * 57476ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 57486ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 57496ba94429SFrederic Weisbecker */ 57506ba94429SFrederic Weisbecker struct wq_device { 57516ba94429SFrederic Weisbecker struct workqueue_struct *wq; 57526ba94429SFrederic Weisbecker struct device dev; 57536ba94429SFrederic Weisbecker }; 57546ba94429SFrederic Weisbecker 57556ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 57566ba94429SFrederic Weisbecker { 57576ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 57586ba94429SFrederic Weisbecker 57596ba94429SFrederic Weisbecker return wq_dev->wq; 57606ba94429SFrederic Weisbecker } 57616ba94429SFrederic Weisbecker 57626ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 57636ba94429SFrederic Weisbecker char *buf) 57646ba94429SFrederic Weisbecker { 57656ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57666ba94429SFrederic Weisbecker 57676ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 57686ba94429SFrederic Weisbecker } 57696ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 57706ba94429SFrederic Weisbecker 57716ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 57726ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 57736ba94429SFrederic Weisbecker { 57746ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57756ba94429SFrederic Weisbecker 57766ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 57776ba94429SFrederic Weisbecker } 57786ba94429SFrederic Weisbecker 57796ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 57806ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 57816ba94429SFrederic Weisbecker size_t count) 57826ba94429SFrederic Weisbecker { 57836ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57846ba94429SFrederic Weisbecker int val; 57856ba94429SFrederic Weisbecker 57866ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 57876ba94429SFrederic Weisbecker return -EINVAL; 57886ba94429SFrederic Weisbecker 57896ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 57906ba94429SFrederic Weisbecker return count; 57916ba94429SFrederic Weisbecker } 57926ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 57936ba94429SFrederic Weisbecker 57946ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 57956ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 57966ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 57976ba94429SFrederic Weisbecker NULL, 57986ba94429SFrederic Weisbecker }; 57996ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 58006ba94429SFrederic Weisbecker 58016ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 58026ba94429SFrederic Weisbecker char *buf) 58036ba94429SFrederic Weisbecker { 58046ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58056ba94429SFrederic Weisbecker int written; 58066ba94429SFrederic Weisbecker 58076ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 58086ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 58096ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 58106ba94429SFrederic Weisbecker 58116ba94429SFrederic Weisbecker return written; 58126ba94429SFrederic Weisbecker } 58136ba94429SFrederic Weisbecker 58146ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 58156ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 58166ba94429SFrederic Weisbecker { 58176ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 58186ba94429SFrederic Weisbecker 5819899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5820899a94feSLai Jiangshan 5821be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 58226ba94429SFrederic Weisbecker if (!attrs) 58236ba94429SFrederic Weisbecker return NULL; 58246ba94429SFrederic Weisbecker 58256ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 58266ba94429SFrederic Weisbecker return attrs; 58276ba94429SFrederic Weisbecker } 58286ba94429SFrederic Weisbecker 58296ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 58306ba94429SFrederic Weisbecker const char *buf, size_t count) 58316ba94429SFrederic Weisbecker { 58326ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58336ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5834d4d3e257SLai Jiangshan int ret = -ENOMEM; 5835d4d3e257SLai Jiangshan 5836d4d3e257SLai Jiangshan apply_wqattrs_lock(); 58376ba94429SFrederic Weisbecker 58386ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 58396ba94429SFrederic Weisbecker if (!attrs) 5840d4d3e257SLai Jiangshan goto out_unlock; 58416ba94429SFrederic Weisbecker 58426ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 58436ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5844d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 58456ba94429SFrederic Weisbecker else 58466ba94429SFrederic Weisbecker ret = -EINVAL; 58476ba94429SFrederic Weisbecker 5848d4d3e257SLai Jiangshan out_unlock: 5849d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 58506ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 58516ba94429SFrederic Weisbecker return ret ?: count; 58526ba94429SFrederic Weisbecker } 58536ba94429SFrederic Weisbecker 58546ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 58556ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58566ba94429SFrederic Weisbecker { 58576ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58586ba94429SFrederic Weisbecker int written; 58596ba94429SFrederic Weisbecker 58606ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 58616ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 58626ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 58636ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 58646ba94429SFrederic Weisbecker return written; 58656ba94429SFrederic Weisbecker } 58666ba94429SFrederic Weisbecker 58676ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 58686ba94429SFrederic Weisbecker struct device_attribute *attr, 58696ba94429SFrederic Weisbecker const char *buf, size_t count) 58706ba94429SFrederic Weisbecker { 58716ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58726ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5873d4d3e257SLai Jiangshan int ret = -ENOMEM; 5874d4d3e257SLai Jiangshan 5875d4d3e257SLai Jiangshan apply_wqattrs_lock(); 58766ba94429SFrederic Weisbecker 58776ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 58786ba94429SFrederic Weisbecker if (!attrs) 5879d4d3e257SLai Jiangshan goto out_unlock; 58806ba94429SFrederic Weisbecker 58816ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 58826ba94429SFrederic Weisbecker if (!ret) 5883d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 58846ba94429SFrederic Weisbecker 5885d4d3e257SLai Jiangshan out_unlock: 5886d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 58876ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 58886ba94429SFrederic Weisbecker return ret ?: count; 58896ba94429SFrederic Weisbecker } 58906ba94429SFrederic Weisbecker 58916ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 58926ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 58936ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 58946ba94429SFrederic Weisbecker __ATTR_NULL, 58956ba94429SFrederic Weisbecker }; 58966ba94429SFrederic Weisbecker 58976ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 58986ba94429SFrederic Weisbecker .name = "workqueue", 58996ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 59006ba94429SFrederic Weisbecker }; 59016ba94429SFrederic Weisbecker 5902b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5903b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5904b05a7928SFrederic Weisbecker { 5905b05a7928SFrederic Weisbecker int written; 5906b05a7928SFrederic Weisbecker 5907042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5908b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5909b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5910042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5911b05a7928SFrederic Weisbecker 5912b05a7928SFrederic Weisbecker return written; 5913b05a7928SFrederic Weisbecker } 5914b05a7928SFrederic Weisbecker 5915042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5916042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5917042f7df1SLai Jiangshan { 5918042f7df1SLai Jiangshan cpumask_var_t cpumask; 5919042f7df1SLai Jiangshan int ret; 5920042f7df1SLai Jiangshan 5921042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5922042f7df1SLai Jiangshan return -ENOMEM; 5923042f7df1SLai Jiangshan 5924042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5925042f7df1SLai Jiangshan if (!ret) 5926042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5927042f7df1SLai Jiangshan 5928042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5929042f7df1SLai Jiangshan return ret ? ret : count; 5930042f7df1SLai Jiangshan } 5931042f7df1SLai Jiangshan 5932b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5933042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5934042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5935b05a7928SFrederic Weisbecker 59366ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 59376ba94429SFrederic Weisbecker { 5938686f6697SGreg Kroah-Hartman struct device *dev_root; 5939b05a7928SFrederic Weisbecker int err; 5940b05a7928SFrederic Weisbecker 5941b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5942b05a7928SFrederic Weisbecker if (err) 5943b05a7928SFrederic Weisbecker return err; 5944b05a7928SFrederic Weisbecker 5945686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 5946686f6697SGreg Kroah-Hartman if (dev_root) { 5947686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 5948686f6697SGreg Kroah-Hartman put_device(dev_root); 5949686f6697SGreg Kroah-Hartman } 5950686f6697SGreg Kroah-Hartman return err; 59516ba94429SFrederic Weisbecker } 59526ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 59536ba94429SFrederic Weisbecker 59546ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 59556ba94429SFrederic Weisbecker { 59566ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 59576ba94429SFrederic Weisbecker 59586ba94429SFrederic Weisbecker kfree(wq_dev); 59596ba94429SFrederic Weisbecker } 59606ba94429SFrederic Weisbecker 59616ba94429SFrederic Weisbecker /** 59626ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 59636ba94429SFrederic Weisbecker * @wq: the workqueue to register 59646ba94429SFrederic Weisbecker * 59656ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 59666ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 59676ba94429SFrederic Weisbecker * which is the preferred method. 59686ba94429SFrederic Weisbecker * 59696ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 59706ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 59716ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 59726ba94429SFrederic Weisbecker * attributes. 59736ba94429SFrederic Weisbecker * 59746ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 59756ba94429SFrederic Weisbecker */ 59766ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 59776ba94429SFrederic Weisbecker { 59786ba94429SFrederic Weisbecker struct wq_device *wq_dev; 59796ba94429SFrederic Weisbecker int ret; 59806ba94429SFrederic Weisbecker 59816ba94429SFrederic Weisbecker /* 5982402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 59836ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 59846ba94429SFrederic Weisbecker * workqueues. 59856ba94429SFrederic Weisbecker */ 59860a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 59876ba94429SFrederic Weisbecker return -EINVAL; 59886ba94429SFrederic Weisbecker 59896ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 59906ba94429SFrederic Weisbecker if (!wq_dev) 59916ba94429SFrederic Weisbecker return -ENOMEM; 59926ba94429SFrederic Weisbecker 59936ba94429SFrederic Weisbecker wq_dev->wq = wq; 59946ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 59956ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 599623217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 59976ba94429SFrederic Weisbecker 59986ba94429SFrederic Weisbecker /* 59996ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 60006ba94429SFrederic Weisbecker * everything is ready. 60016ba94429SFrederic Weisbecker */ 60026ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 60036ba94429SFrederic Weisbecker 60046ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 60056ba94429SFrederic Weisbecker if (ret) { 6006537f4146SArvind Yadav put_device(&wq_dev->dev); 60076ba94429SFrederic Weisbecker wq->wq_dev = NULL; 60086ba94429SFrederic Weisbecker return ret; 60096ba94429SFrederic Weisbecker } 60106ba94429SFrederic Weisbecker 60116ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 60126ba94429SFrederic Weisbecker struct device_attribute *attr; 60136ba94429SFrederic Weisbecker 60146ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 60156ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 60166ba94429SFrederic Weisbecker if (ret) { 60176ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 60186ba94429SFrederic Weisbecker wq->wq_dev = NULL; 60196ba94429SFrederic Weisbecker return ret; 60206ba94429SFrederic Weisbecker } 60216ba94429SFrederic Weisbecker } 60226ba94429SFrederic Weisbecker } 60236ba94429SFrederic Weisbecker 60246ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 60256ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 60266ba94429SFrederic Weisbecker return 0; 60276ba94429SFrederic Weisbecker } 60286ba94429SFrederic Weisbecker 60296ba94429SFrederic Weisbecker /** 60306ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 60316ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 60326ba94429SFrederic Weisbecker * 60336ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 60346ba94429SFrederic Weisbecker */ 60356ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 60366ba94429SFrederic Weisbecker { 60376ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 60386ba94429SFrederic Weisbecker 60396ba94429SFrederic Weisbecker if (!wq->wq_dev) 60406ba94429SFrederic Weisbecker return; 60416ba94429SFrederic Weisbecker 60426ba94429SFrederic Weisbecker wq->wq_dev = NULL; 60436ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 60446ba94429SFrederic Weisbecker } 60456ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 60466ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 60476ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 60486ba94429SFrederic Weisbecker 604982607adcSTejun Heo /* 605082607adcSTejun Heo * Workqueue watchdog. 605182607adcSTejun Heo * 605282607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 605382607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 605482607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 605582607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 605682607adcSTejun Heo * largely opaque. 605782607adcSTejun Heo * 605882607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 605982607adcSTejun Heo * state if some pools failed to make forward progress for a while where 606082607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 606182607adcSTejun Heo * 606282607adcSTejun Heo * This mechanism is controlled through the kernel parameter 606382607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 606482607adcSTejun Heo * corresponding sysfs parameter file. 606582607adcSTejun Heo */ 606682607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 606782607adcSTejun Heo 606882607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 60695cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 607082607adcSTejun Heo 607182607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 607282607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 607382607adcSTejun Heo 6074cd2440d6SPetr Mladek /* 6075cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6076cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6077cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6078cd2440d6SPetr Mladek * in all other situations. 6079cd2440d6SPetr Mladek */ 6080cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6081cd2440d6SPetr Mladek { 6082cd2440d6SPetr Mladek struct worker *worker; 6083cd2440d6SPetr Mladek unsigned long flags; 6084cd2440d6SPetr Mladek int bkt; 6085cd2440d6SPetr Mladek 6086cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6087cd2440d6SPetr Mladek 6088cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6089cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6090cd2440d6SPetr Mladek /* 6091cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6092cd2440d6SPetr Mladek * drivers that queue work while holding locks 6093cd2440d6SPetr Mladek * also taken in their write paths. 6094cd2440d6SPetr Mladek */ 6095cd2440d6SPetr Mladek printk_deferred_enter(); 6096cd2440d6SPetr Mladek 6097cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6098cd2440d6SPetr Mladek sched_show_task(worker->task); 6099cd2440d6SPetr Mladek 6100cd2440d6SPetr Mladek printk_deferred_exit(); 6101cd2440d6SPetr Mladek } 6102cd2440d6SPetr Mladek } 6103cd2440d6SPetr Mladek 6104cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6105cd2440d6SPetr Mladek } 6106cd2440d6SPetr Mladek 6107cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6108cd2440d6SPetr Mladek { 6109cd2440d6SPetr Mladek struct worker_pool *pool; 6110cd2440d6SPetr Mladek int pi; 6111cd2440d6SPetr Mladek 6112cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6113cd2440d6SPetr Mladek 6114cd2440d6SPetr Mladek rcu_read_lock(); 6115cd2440d6SPetr Mladek 6116cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6117cd2440d6SPetr Mladek if (pool->cpu_stall) 6118cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6119cd2440d6SPetr Mladek 6120cd2440d6SPetr Mladek } 6121cd2440d6SPetr Mladek 6122cd2440d6SPetr Mladek rcu_read_unlock(); 6123cd2440d6SPetr Mladek } 6124cd2440d6SPetr Mladek 612582607adcSTejun Heo static void wq_watchdog_reset_touched(void) 612682607adcSTejun Heo { 612782607adcSTejun Heo int cpu; 612882607adcSTejun Heo 612982607adcSTejun Heo wq_watchdog_touched = jiffies; 613082607adcSTejun Heo for_each_possible_cpu(cpu) 613182607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 613282607adcSTejun Heo } 613382607adcSTejun Heo 61345cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 613582607adcSTejun Heo { 613682607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 613782607adcSTejun Heo bool lockup_detected = false; 6138cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6139940d71c6SSergey Senozhatsky unsigned long now = jiffies; 614082607adcSTejun Heo struct worker_pool *pool; 614182607adcSTejun Heo int pi; 614282607adcSTejun Heo 614382607adcSTejun Heo if (!thresh) 614482607adcSTejun Heo return; 614582607adcSTejun Heo 614682607adcSTejun Heo rcu_read_lock(); 614782607adcSTejun Heo 614882607adcSTejun Heo for_each_pool(pool, pi) { 614982607adcSTejun Heo unsigned long pool_ts, touched, ts; 615082607adcSTejun Heo 6151cd2440d6SPetr Mladek pool->cpu_stall = false; 615282607adcSTejun Heo if (list_empty(&pool->worklist)) 615382607adcSTejun Heo continue; 615482607adcSTejun Heo 6155940d71c6SSergey Senozhatsky /* 6156940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6157940d71c6SSergey Senozhatsky * the watchdog like a stall. 6158940d71c6SSergey Senozhatsky */ 6159940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6160940d71c6SSergey Senozhatsky 616182607adcSTejun Heo /* get the latest of pool and touched timestamps */ 616289e28ce6SWang Qing if (pool->cpu >= 0) 616389e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 616489e28ce6SWang Qing else 616582607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 616689e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 616782607adcSTejun Heo 616882607adcSTejun Heo if (time_after(pool_ts, touched)) 616982607adcSTejun Heo ts = pool_ts; 617082607adcSTejun Heo else 617182607adcSTejun Heo ts = touched; 617282607adcSTejun Heo 617382607adcSTejun Heo /* did we stall? */ 6174940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 617582607adcSTejun Heo lockup_detected = true; 6176cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6177cd2440d6SPetr Mladek pool->cpu_stall = true; 6178cd2440d6SPetr Mladek cpu_pool_stall = true; 6179cd2440d6SPetr Mladek } 618082607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 618182607adcSTejun Heo pr_cont_pool_info(pool); 618282607adcSTejun Heo pr_cont(" stuck for %us!\n", 6183940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 618482607adcSTejun Heo } 6185cd2440d6SPetr Mladek 6186cd2440d6SPetr Mladek 618782607adcSTejun Heo } 618882607adcSTejun Heo 618982607adcSTejun Heo rcu_read_unlock(); 619082607adcSTejun Heo 619182607adcSTejun Heo if (lockup_detected) 619255df0933SImran Khan show_all_workqueues(); 619382607adcSTejun Heo 6194cd2440d6SPetr Mladek if (cpu_pool_stall) 6195cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6196cd2440d6SPetr Mladek 619782607adcSTejun Heo wq_watchdog_reset_touched(); 619882607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 619982607adcSTejun Heo } 620082607adcSTejun Heo 6201cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 620282607adcSTejun Heo { 620382607adcSTejun Heo if (cpu >= 0) 620482607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 620589e28ce6SWang Qing 620682607adcSTejun Heo wq_watchdog_touched = jiffies; 620782607adcSTejun Heo } 620882607adcSTejun Heo 620982607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 621082607adcSTejun Heo { 621182607adcSTejun Heo wq_watchdog_thresh = 0; 621282607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 621382607adcSTejun Heo 621482607adcSTejun Heo if (thresh) { 621582607adcSTejun Heo wq_watchdog_thresh = thresh; 621682607adcSTejun Heo wq_watchdog_reset_touched(); 621782607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 621882607adcSTejun Heo } 621982607adcSTejun Heo } 622082607adcSTejun Heo 622182607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 622282607adcSTejun Heo const struct kernel_param *kp) 622382607adcSTejun Heo { 622482607adcSTejun Heo unsigned long thresh; 622582607adcSTejun Heo int ret; 622682607adcSTejun Heo 622782607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 622882607adcSTejun Heo if (ret) 622982607adcSTejun Heo return ret; 623082607adcSTejun Heo 623182607adcSTejun Heo if (system_wq) 623282607adcSTejun Heo wq_watchdog_set_thresh(thresh); 623382607adcSTejun Heo else 623482607adcSTejun Heo wq_watchdog_thresh = thresh; 623582607adcSTejun Heo 623682607adcSTejun Heo return 0; 623782607adcSTejun Heo } 623882607adcSTejun Heo 623982607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 624082607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 624182607adcSTejun Heo .get = param_get_ulong, 624282607adcSTejun Heo }; 624382607adcSTejun Heo 624482607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 624582607adcSTejun Heo 0644); 624682607adcSTejun Heo 624782607adcSTejun Heo static void wq_watchdog_init(void) 624882607adcSTejun Heo { 62495cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 625082607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 625182607adcSTejun Heo } 625282607adcSTejun Heo 625382607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 625482607adcSTejun Heo 625582607adcSTejun Heo static inline void wq_watchdog_init(void) { } 625682607adcSTejun Heo 625782607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 625882607adcSTejun Heo 6259*a86feae6STejun Heo static void wq_pod_init(void); 6260bce90380STejun Heo 62613347fa09STejun Heo /** 62623347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 62633347fa09STejun Heo * 62643347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 62653347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 62663347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 62673347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 62683347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 62693347fa09STejun Heo * created and scheduled right before early initcalls. 62703347fa09STejun Heo */ 62712333e829SYu Chen void __init workqueue_init_early(void) 62721da177e4SLinus Torvalds { 62737a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 62747a4e344cSTejun Heo int i, cpu; 6275c34056a3STejun Heo 627610cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6277e904e6c2STejun Heo 6278b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 627904d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 628004d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6281b05a7928SFrederic Weisbecker 6282ace3c549Stiozhang if (!cpumask_empty(&wq_cmdline_cpumask)) 6283ace3c549Stiozhang cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask); 6284ace3c549Stiozhang 6285e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6286e904e6c2STejun Heo 6287706026c2STejun Heo /* initialize CPU pools */ 628829c91e99STejun Heo for_each_possible_cpu(cpu) { 62894ce62e9eSTejun Heo struct worker_pool *pool; 62908b03ae3cSTejun Heo 62917a4e344cSTejun Heo i = 0; 6292f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 62937a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6294ec22ca5eSTejun Heo pool->cpu = cpu; 62957a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 62967a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6297f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 62987a4e344cSTejun Heo 62999daf9e67STejun Heo /* alloc pool ID */ 630068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 63019daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 630268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 63034ce62e9eSTejun Heo } 63048b03ae3cSTejun Heo } 63058b03ae3cSTejun Heo 63068a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 630729c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 630829c91e99STejun Heo struct workqueue_attrs *attrs; 630929c91e99STejun Heo 6310be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 631129c91e99STejun Heo attrs->nice = std_nice[i]; 631229c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 63138a2b7538STejun Heo 63148a2b7538STejun Heo /* 63158a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 63168a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 63178a2b7538STejun Heo */ 6318be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 63198a2b7538STejun Heo attrs->nice = std_nice[i]; 6320af73f5c9STejun Heo attrs->ordered = true; 63218a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 632229c91e99STejun Heo } 632329c91e99STejun Heo 6324d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 63251aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6326d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6327f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6328636b927eSTejun Heo WQ_MAX_ACTIVE); 632924d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 633024d51addSTejun Heo WQ_FREEZABLE, 0); 63310668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 63320668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 63330668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 63340668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 63350668106cSViresh Kumar 0); 63361aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 63370668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 63380668106cSViresh Kumar !system_power_efficient_wq || 63390668106cSViresh Kumar !system_freezable_power_efficient_wq); 63403347fa09STejun Heo } 63413347fa09STejun Heo 6342aa6fde93STejun Heo static void __init wq_cpu_intensive_thresh_init(void) 6343aa6fde93STejun Heo { 6344aa6fde93STejun Heo unsigned long thresh; 6345aa6fde93STejun Heo unsigned long bogo; 6346aa6fde93STejun Heo 6347aa6fde93STejun Heo /* if the user set it to a specific value, keep it */ 6348aa6fde93STejun Heo if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6349aa6fde93STejun Heo return; 6350aa6fde93STejun Heo 6351967b494eSTejun Heo pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 6352967b494eSTejun Heo BUG_ON(IS_ERR(pwq_release_worker)); 6353967b494eSTejun Heo 6354aa6fde93STejun Heo /* 6355aa6fde93STejun Heo * The default of 10ms is derived from the fact that most modern (as of 6356aa6fde93STejun Heo * 2023) processors can do a lot in 10ms and that it's just below what 6357aa6fde93STejun Heo * most consider human-perceivable. However, the kernel also runs on a 6358aa6fde93STejun Heo * lot slower CPUs including microcontrollers where the threshold is way 6359aa6fde93STejun Heo * too low. 6360aa6fde93STejun Heo * 6361aa6fde93STejun Heo * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6362aa6fde93STejun Heo * This is by no means accurate but it doesn't have to be. The mechanism 6363aa6fde93STejun Heo * is still useful even when the threshold is fully scaled up. Also, as 6364aa6fde93STejun Heo * the reports would usually be applicable to everyone, some machines 6365aa6fde93STejun Heo * operating on longer thresholds won't significantly diminish their 6366aa6fde93STejun Heo * usefulness. 6367aa6fde93STejun Heo */ 6368aa6fde93STejun Heo thresh = 10 * USEC_PER_MSEC; 6369aa6fde93STejun Heo 6370aa6fde93STejun Heo /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6371aa6fde93STejun Heo bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6372aa6fde93STejun Heo if (bogo < 4000) 6373aa6fde93STejun Heo thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6374aa6fde93STejun Heo 6375aa6fde93STejun Heo pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6376aa6fde93STejun Heo loops_per_jiffy, bogo, thresh); 6377aa6fde93STejun Heo 6378aa6fde93STejun Heo wq_cpu_intensive_thresh_us = thresh; 6379aa6fde93STejun Heo } 6380aa6fde93STejun Heo 63813347fa09STejun Heo /** 63823347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 63833347fa09STejun Heo * 63843347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 63853347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 63863347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 63873347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 63883347fa09STejun Heo * with the initial workers and enable future kworker creations. 63893347fa09STejun Heo */ 63902333e829SYu Chen void __init workqueue_init(void) 63913347fa09STejun Heo { 63922186d9f9STejun Heo struct workqueue_struct *wq; 63933347fa09STejun Heo struct worker_pool *pool; 63943347fa09STejun Heo int cpu, bkt; 63953347fa09STejun Heo 6396aa6fde93STejun Heo wq_cpu_intensive_thresh_init(); 6397aa6fde93STejun Heo 63982186d9f9STejun Heo /* 6399fef59c9cSTejun Heo * It'd be simpler to initialize pods in workqueue_init_early() but CPU 6400fef59c9cSTejun Heo * to node mapping may not be available that early on some archs such as 6401fef59c9cSTejun Heo * power and arm64. As per-cpu pools created previously could be missing 6402fef59c9cSTejun Heo * node hint and unbound pool pod affinity, fix them up. 640340c17f75STejun Heo * 640440c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 64052186d9f9STejun Heo */ 6406fef59c9cSTejun Heo wq_pod_init(); 64072186d9f9STejun Heo 64082186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 64092186d9f9STejun Heo 64102186d9f9STejun Heo for_each_possible_cpu(cpu) { 64112186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 64122186d9f9STejun Heo pool->node = cpu_to_node(cpu); 64132186d9f9STejun Heo } 64142186d9f9STejun Heo } 64152186d9f9STejun Heo 641640c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 6417fef59c9cSTejun Heo wq_update_pod(wq, smp_processor_id(), smp_processor_id(), true); 641840c17f75STejun Heo WARN(init_rescuer(wq), 641940c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 642040c17f75STejun Heo wq->name); 642140c17f75STejun Heo } 64222186d9f9STejun Heo 64232186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 64242186d9f9STejun Heo 64253347fa09STejun Heo /* create the initial workers */ 64263347fa09STejun Heo for_each_online_cpu(cpu) { 64273347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 64283347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 64293347fa09STejun Heo BUG_ON(!create_worker(pool)); 64303347fa09STejun Heo } 64313347fa09STejun Heo } 64323347fa09STejun Heo 64333347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 64343347fa09STejun Heo BUG_ON(!create_worker(pool)); 64353347fa09STejun Heo 64363347fa09STejun Heo wq_online = true; 643782607adcSTejun Heo wq_watchdog_init(); 64381da177e4SLinus Torvalds } 6439c4f135d6STetsuo Handa 6440*a86feae6STejun Heo static void __init wq_pod_init(void) 6441*a86feae6STejun Heo { 6442*a86feae6STejun Heo cpumask_var_t *tbl; 6443*a86feae6STejun Heo int node, cpu; 6444*a86feae6STejun Heo 6445*a86feae6STejun Heo if (num_possible_nodes() <= 1) 6446*a86feae6STejun Heo return; 6447*a86feae6STejun Heo 6448*a86feae6STejun Heo for_each_possible_cpu(cpu) { 6449*a86feae6STejun Heo if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6450*a86feae6STejun Heo pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6451*a86feae6STejun Heo return; 6452*a86feae6STejun Heo } 6453*a86feae6STejun Heo } 6454*a86feae6STejun Heo 6455*a86feae6STejun Heo wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 6456*a86feae6STejun Heo BUG_ON(!wq_update_pod_attrs_buf); 6457*a86feae6STejun Heo 6458*a86feae6STejun Heo /* 6459*a86feae6STejun Heo * We want masks of possible CPUs of each node which isn't readily 6460*a86feae6STejun Heo * available. Build one from cpu_to_node() which should have been 6461*a86feae6STejun Heo * fully initialized by now. 6462*a86feae6STejun Heo */ 6463*a86feae6STejun Heo tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6464*a86feae6STejun Heo BUG_ON(!tbl); 6465*a86feae6STejun Heo 6466*a86feae6STejun Heo for_each_node(node) 6467*a86feae6STejun Heo BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 6468*a86feae6STejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6469*a86feae6STejun Heo 6470*a86feae6STejun Heo for_each_possible_cpu(cpu) { 6471*a86feae6STejun Heo node = cpu_to_node(cpu); 6472*a86feae6STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6473*a86feae6STejun Heo } 6474*a86feae6STejun Heo 6475*a86feae6STejun Heo wq_pod_cpus = tbl; 6476*a86feae6STejun Heo wq_pod_enabled = true; 6477*a86feae6STejun Heo } 6478*a86feae6STejun Heo 647920bdedafSTetsuo Handa void __warn_flushing_systemwide_wq(void) 648020bdedafSTetsuo Handa { 648120bdedafSTetsuo Handa pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 648220bdedafSTetsuo Handa dump_stack(); 648320bdedafSTetsuo Handa } 6484c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6485ace3c549Stiozhang 6486ace3c549Stiozhang static int __init workqueue_unbound_cpus_setup(char *str) 6487ace3c549Stiozhang { 6488ace3c549Stiozhang if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6489ace3c549Stiozhang cpumask_clear(&wq_cmdline_cpumask); 6490ace3c549Stiozhang pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6491ace3c549Stiozhang } 6492ace3c549Stiozhang 6493ace3c549Stiozhang return 1; 6494ace3c549Stiozhang } 6495ace3c549Stiozhang __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6496