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 /* 2608864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2618864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 26224acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 263b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2648864b4e5STejun Heo */ 2658864b4e5STejun Heo struct work_struct unbound_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 */ 324*ee1ceef7STejun Heo struct pool_workqueue __percpu *cpu_pwq; /* I: per-cpu pwqs */ 3255b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 3261da177e4SLinus Torvalds }; 3271da177e4SLinus Torvalds 328e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 329e904e6c2STejun Heo 330bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 331bce90380STejun Heo /* possible CPUs of each node */ 332bce90380STejun Heo 333616db877STejun Heo /* 334616db877STejun Heo * Per-cpu work items which run for longer than the following threshold are 335616db877STejun Heo * automatically considered CPU intensive and excluded from concurrency 336616db877STejun Heo * management to prevent them from noticeably delaying other per-cpu work items. 337aa6fde93STejun Heo * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 338aa6fde93STejun Heo * The actual value is initialized in wq_cpu_intensive_thresh_init(). 339616db877STejun Heo */ 340aa6fde93STejun Heo static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 341616db877STejun Heo module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 342616db877STejun Heo 343d55262c4STejun Heo static bool wq_disable_numa; 344d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 345d55262c4STejun Heo 346cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 347552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 348cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 349cee22a15SViresh Kumar 350863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3513347fa09STejun Heo 352bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 353bce90380STejun Heo 3544c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3554c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3564c16bd32STejun Heo 35768e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3581258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 359a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 360d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 361d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3625bcab335STejun Heo 363e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 36468e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3657d19c5ceSTejun Heo 36699c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 367ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 368ef557180SMike Galbraith 369ace3c549Stiozhang /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 370ace3c549Stiozhang static struct cpumask wq_cmdline_cpumask __initdata; 371ace3c549Stiozhang 372ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 373ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 374b05a7928SFrederic Weisbecker 375f303fccbSTejun Heo /* 376f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 377f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 378f303fccbSTejun Heo * to uncover usages which depend on it. 379f303fccbSTejun Heo */ 380f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 381f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 382f303fccbSTejun Heo #else 383f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 384f303fccbSTejun Heo #endif 385f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 386f303fccbSTejun Heo 3877d19c5ceSTejun Heo /* the per-cpu worker pools */ 38825528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3897d19c5ceSTejun Heo 39068e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3917d19c5ceSTejun Heo 39268e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 39329c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 39429c91e99STejun Heo 395c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 39629c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 39729c91e99STejun Heo 3988a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3998a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 4008a2b7538STejun Heo 401d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 402ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 403044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 4041aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 405044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 406d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 407044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 408f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 409044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 41024d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 4110668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 4120668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 4130668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 4140668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 415d320c038STejun Heo 4167d19c5ceSTejun Heo static int worker_thread(void *__worker); 4176ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 418c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 41955df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4207d19c5ceSTejun Heo 42197bd2347STejun Heo #define CREATE_TRACE_POINTS 42297bd2347STejun Heo #include <trace/events/workqueue.h> 42397bd2347STejun Heo 42468e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 42524acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 426f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 42724acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4285bcab335STejun Heo 4295b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 43024acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 431f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 432f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 43324acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4345b95e1afSLai Jiangshan 435f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 436f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 437f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4387a62c2c8STejun Heo (pool)++) 4394ce62e9eSTejun Heo 44049e3cf44STejun Heo /** 44117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 44217116969STejun Heo * @pool: iteration cursor 443611c92a0STejun Heo * @pi: integer used for iteration 444fa1b54e6STejun Heo * 44524acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 44668e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 44768e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 448fa1b54e6STejun Heo * 449fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 450fa1b54e6STejun Heo * ignored. 45117116969STejun Heo */ 452611c92a0STejun Heo #define for_each_pool(pool, pi) \ 453611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 45468e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 455fa1b54e6STejun Heo else 45617116969STejun Heo 45717116969STejun Heo /** 458822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 459822d8405STejun Heo * @worker: iteration cursor 460822d8405STejun Heo * @pool: worker_pool to iterate workers of 461822d8405STejun Heo * 4621258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 463822d8405STejun Heo * 464822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 465822d8405STejun Heo * ignored. 466822d8405STejun Heo */ 467da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 468da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4691258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 470822d8405STejun Heo else 471822d8405STejun Heo 472822d8405STejun Heo /** 47349e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 47449e3cf44STejun Heo * @pwq: iteration cursor 47549e3cf44STejun Heo * @wq: the target workqueue 47676af4d93STejun Heo * 47724acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 478794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 479794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 48076af4d93STejun Heo * 48176af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 48276af4d93STejun Heo * ignored. 48349e3cf44STejun Heo */ 48449e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 48549e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4865a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 487f3421797STejun Heo 488dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 489dc186ad7SThomas Gleixner 490f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 491dc186ad7SThomas Gleixner 49299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 49399777288SStanislaw Gruszka { 49499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 49599777288SStanislaw Gruszka } 49699777288SStanislaw Gruszka 497b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 498b9fdac7fSDu, Changbin { 499b9fdac7fSDu, Changbin struct work_struct *work = addr; 500b9fdac7fSDu, Changbin 501b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 502b9fdac7fSDu, Changbin } 503b9fdac7fSDu, Changbin 504dc186ad7SThomas Gleixner /* 505dc186ad7SThomas Gleixner * fixup_init is called when: 506dc186ad7SThomas Gleixner * - an active object is initialized 507dc186ad7SThomas Gleixner */ 50802a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 509dc186ad7SThomas Gleixner { 510dc186ad7SThomas Gleixner struct work_struct *work = addr; 511dc186ad7SThomas Gleixner 512dc186ad7SThomas Gleixner switch (state) { 513dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 514dc186ad7SThomas Gleixner cancel_work_sync(work); 515dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 51602a982a6SDu, Changbin return true; 517dc186ad7SThomas Gleixner default: 51802a982a6SDu, Changbin return false; 519dc186ad7SThomas Gleixner } 520dc186ad7SThomas Gleixner } 521dc186ad7SThomas Gleixner 522dc186ad7SThomas Gleixner /* 523dc186ad7SThomas Gleixner * fixup_free is called when: 524dc186ad7SThomas Gleixner * - an active object is freed 525dc186ad7SThomas Gleixner */ 52602a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 527dc186ad7SThomas Gleixner { 528dc186ad7SThomas Gleixner struct work_struct *work = addr; 529dc186ad7SThomas Gleixner 530dc186ad7SThomas Gleixner switch (state) { 531dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 532dc186ad7SThomas Gleixner cancel_work_sync(work); 533dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 53402a982a6SDu, Changbin return true; 535dc186ad7SThomas Gleixner default: 53602a982a6SDu, Changbin return false; 537dc186ad7SThomas Gleixner } 538dc186ad7SThomas Gleixner } 539dc186ad7SThomas Gleixner 540f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 541dc186ad7SThomas Gleixner .name = "work_struct", 54299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 543b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 544dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 545dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 546dc186ad7SThomas Gleixner }; 547dc186ad7SThomas Gleixner 548dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 549dc186ad7SThomas Gleixner { 550dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 551dc186ad7SThomas Gleixner } 552dc186ad7SThomas Gleixner 553dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 554dc186ad7SThomas Gleixner { 555dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 556dc186ad7SThomas Gleixner } 557dc186ad7SThomas Gleixner 558dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 559dc186ad7SThomas Gleixner { 560dc186ad7SThomas Gleixner if (onstack) 561dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 562dc186ad7SThomas Gleixner else 563dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 564dc186ad7SThomas Gleixner } 565dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 566dc186ad7SThomas Gleixner 567dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 568dc186ad7SThomas Gleixner { 569dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 570dc186ad7SThomas Gleixner } 571dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 572dc186ad7SThomas Gleixner 573ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 574ea2e64f2SThomas Gleixner { 575ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 576ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 577ea2e64f2SThomas Gleixner } 578ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 579ea2e64f2SThomas Gleixner 580dc186ad7SThomas Gleixner #else 581dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 582dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 583dc186ad7SThomas Gleixner #endif 584dc186ad7SThomas Gleixner 5854e8b22bdSLi Bin /** 58667dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5874e8b22bdSLi Bin * @pool: the pool pointer of interest 5884e8b22bdSLi Bin * 5894e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5904e8b22bdSLi Bin * successfully, -errno on failure. 5914e8b22bdSLi Bin */ 5929daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5939daf9e67STejun Heo { 5949daf9e67STejun Heo int ret; 5959daf9e67STejun Heo 59668e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5975bcab335STejun Heo 5984e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5994e8b22bdSLi Bin GFP_KERNEL); 600229641a6STejun Heo if (ret >= 0) { 601e68035fbSTejun Heo pool->id = ret; 602229641a6STejun Heo return 0; 603229641a6STejun Heo } 6049daf9e67STejun Heo return ret; 6059daf9e67STejun Heo } 6069daf9e67STejun Heo 60776af4d93STejun Heo /** 608df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 609df2d5ae4STejun Heo * @wq: the target workqueue 610df2d5ae4STejun Heo * @node: the node ID 611df2d5ae4STejun Heo * 61224acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 6135b95e1afSLai Jiangshan * read locked. 614df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 615df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 616d185af30SYacine Belkadi * 617d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 618df2d5ae4STejun Heo */ 619df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 620df2d5ae4STejun Heo int node) 621df2d5ae4STejun Heo { 6225b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 623d6e022f1STejun Heo 624d6e022f1STejun Heo /* 625d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 626d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 627d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 628d6e022f1STejun Heo * happens, this workaround can be removed. 629d6e022f1STejun Heo */ 630d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 631d6e022f1STejun Heo return wq->dfl_pwq; 632d6e022f1STejun Heo 633df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 634df2d5ae4STejun Heo } 635df2d5ae4STejun Heo 63673f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 63773f53c4aSTejun Heo { 63873f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 63973f53c4aSTejun Heo } 64073f53c4aSTejun Heo 641c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 64273f53c4aSTejun Heo { 643c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 64473f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 64573f53c4aSTejun Heo } 64673f53c4aSTejun Heo 64773f53c4aSTejun Heo static int work_next_color(int color) 64873f53c4aSTejun Heo { 64973f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 650a848e3b6SOleg Nesterov } 651a848e3b6SOleg Nesterov 6524594bf15SDavid Howells /* 653112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 654112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6557c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6567a22ad75STejun Heo * 657112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 658112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 659bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 660bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6617a22ad75STejun Heo * 662112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6637c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 664112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6657c3eed5cSTejun Heo * available only while the work item is queued. 666bbb68dfaSTejun Heo * 667bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 668bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 669bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 670bbb68dfaSTejun Heo * try to steal the PENDING bit. 6714594bf15SDavid Howells */ 6727a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6737a22ad75STejun Heo unsigned long flags) 6747a22ad75STejun Heo { 6756183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6767a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6777a22ad75STejun Heo } 6787a22ad75STejun Heo 679112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6804690c4abSTejun Heo unsigned long extra_flags) 681365970a1SDavid Howells { 682112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 683112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 684365970a1SDavid Howells } 685365970a1SDavid Howells 6864468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6874468a00fSLai Jiangshan int pool_id) 6884468a00fSLai Jiangshan { 6894468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6904468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6914468a00fSLai Jiangshan } 6924468a00fSLai Jiangshan 6937c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6947c3eed5cSTejun Heo int pool_id) 6954d707b9fSOleg Nesterov { 69623657bb1STejun Heo /* 69723657bb1STejun Heo * The following wmb is paired with the implied mb in 69823657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 69923657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 70023657bb1STejun Heo * owner. 70123657bb1STejun Heo */ 70223657bb1STejun Heo smp_wmb(); 7037c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 704346c09f8SRoman Pen /* 705346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 706346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 707346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 7088bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 709346c09f8SRoman Pen * the same @work. E.g. consider this case: 710346c09f8SRoman Pen * 711346c09f8SRoman Pen * CPU#0 CPU#1 712346c09f8SRoman Pen * ---------------------------- -------------------------------- 713346c09f8SRoman Pen * 714346c09f8SRoman Pen * 1 STORE event_indicated 715346c09f8SRoman Pen * 2 queue_work_on() { 716346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 717346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 718346c09f8SRoman Pen * 5 set_work_data() # clear bit 719346c09f8SRoman Pen * 6 smp_mb() 720346c09f8SRoman Pen * 7 work->current_func() { 721346c09f8SRoman Pen * 8 LOAD event_indicated 722346c09f8SRoman Pen * } 723346c09f8SRoman Pen * 724346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 725346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 726346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 727346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 728346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 729346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 730346c09f8SRoman Pen * before actual STORE. 731346c09f8SRoman Pen */ 732346c09f8SRoman Pen smp_mb(); 7334d707b9fSOleg Nesterov } 7344d707b9fSOleg Nesterov 7357a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 736365970a1SDavid Howells { 7377c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7387c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7397a22ad75STejun Heo } 7407a22ad75STejun Heo 741afa4bb77SLinus Torvalds static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 742afa4bb77SLinus Torvalds { 743afa4bb77SLinus Torvalds return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 744afa4bb77SLinus Torvalds } 745afa4bb77SLinus Torvalds 746112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7477a22ad75STejun Heo { 748e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7497a22ad75STejun Heo 750112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 751afa4bb77SLinus Torvalds return work_struct_pwq(data); 752e120153dSTejun Heo else 753e120153dSTejun Heo return NULL; 7547a22ad75STejun Heo } 7557a22ad75STejun Heo 7567c3eed5cSTejun Heo /** 7577c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7587c3eed5cSTejun Heo * @work: the work item of interest 7597c3eed5cSTejun Heo * 76068e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 76124acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 76224acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 763fa1b54e6STejun Heo * 764fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 765fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 766fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 767fa1b54e6STejun Heo * returned pool is and stays online. 768d185af30SYacine Belkadi * 769d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7707c3eed5cSTejun Heo */ 7717c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7727a22ad75STejun Heo { 773e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7747c3eed5cSTejun Heo int pool_id; 7757a22ad75STejun Heo 77668e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 777fa1b54e6STejun Heo 778112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 779afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool; 7807a22ad75STejun Heo 7817c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7827c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7837a22ad75STejun Heo return NULL; 7847a22ad75STejun Heo 785fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7867c3eed5cSTejun Heo } 7877c3eed5cSTejun Heo 7887c3eed5cSTejun Heo /** 7897c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7907c3eed5cSTejun Heo * @work: the work item of interest 7917c3eed5cSTejun Heo * 792d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7937c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7947c3eed5cSTejun Heo */ 7957c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7967c3eed5cSTejun Heo { 79754d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7987c3eed5cSTejun Heo 799112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 800afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool->id; 80154d5b7d0SLai Jiangshan 80254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 8037c3eed5cSTejun Heo } 8047c3eed5cSTejun Heo 805bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 806bbb68dfaSTejun Heo { 8077c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 808bbb68dfaSTejun Heo 8097c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 8107c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 811bbb68dfaSTejun Heo } 812bbb68dfaSTejun Heo 813bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 814bbb68dfaSTejun Heo { 815bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 816bbb68dfaSTejun Heo 817112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 818bbb68dfaSTejun Heo } 819bbb68dfaSTejun Heo 820e22bee78STejun Heo /* 8213270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8223270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 823d565ed63STejun Heo * they're being called with pool->lock held. 824e22bee78STejun Heo */ 825e22bee78STejun Heo 82663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 827649027d7STejun Heo { 828bc35f7efSLai Jiangshan return !pool->nr_running; 829649027d7STejun Heo } 830649027d7STejun Heo 831e22bee78STejun Heo /* 832e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 833e22bee78STejun Heo * running workers. 834974271c4STejun Heo * 835974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 836706026c2STejun Heo * function will always return %true for unbound pools as long as the 837974271c4STejun Heo * worklist isn't empty. 838e22bee78STejun Heo */ 83963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 840e22bee78STejun Heo { 84163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 842e22bee78STejun Heo } 843e22bee78STejun Heo 844e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 84563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 846e22bee78STejun Heo { 84763d95a91STejun Heo return pool->nr_idle; 848e22bee78STejun Heo } 849e22bee78STejun Heo 850e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 85163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 852e22bee78STejun Heo { 853bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 854e22bee78STejun Heo } 855e22bee78STejun Heo 856e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 85763d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 858e22bee78STejun Heo { 85963d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 860e22bee78STejun Heo } 861e22bee78STejun Heo 862e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 86363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 864e22bee78STejun Heo { 865692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 86663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 86763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 868e22bee78STejun Heo 869e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 870e22bee78STejun Heo } 871e22bee78STejun Heo 872e22bee78STejun Heo /* 873e22bee78STejun Heo * Wake up functions. 874e22bee78STejun Heo */ 875e22bee78STejun Heo 8762c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8771037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8787e11629dSTejun Heo { 87963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8807e11629dSTejun Heo return NULL; 8817e11629dSTejun Heo 88263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8837e11629dSTejun Heo } 8847e11629dSTejun Heo 8857e11629dSTejun Heo /** 8867e11629dSTejun Heo * wake_up_worker - wake up an idle worker 88763d95a91STejun Heo * @pool: worker pool to wake worker from 8887e11629dSTejun Heo * 88963d95a91STejun Heo * Wake up the first idle worker of @pool. 8907e11629dSTejun Heo * 8917e11629dSTejun Heo * CONTEXT: 892a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8937e11629dSTejun Heo */ 89463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8957e11629dSTejun Heo { 8961037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8977e11629dSTejun Heo 8987e11629dSTejun Heo if (likely(worker)) 8997e11629dSTejun Heo wake_up_process(worker->task); 9007e11629dSTejun Heo } 9017e11629dSTejun Heo 9024690c4abSTejun Heo /** 903e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 904cb444766STejun Heo * @worker: self 905d302f017STejun Heo * @flags: flags to set 906d302f017STejun Heo * 907228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 908d302f017STejun Heo */ 909228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 910d302f017STejun Heo { 911bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 912e22bee78STejun Heo 913bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 914cb444766STejun Heo 915228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 916e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 917e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 918bc35f7efSLai Jiangshan pool->nr_running--; 919e22bee78STejun Heo } 920e22bee78STejun Heo 921d302f017STejun Heo worker->flags |= flags; 922d302f017STejun Heo } 923d302f017STejun Heo 924d302f017STejun Heo /** 925e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 926cb444766STejun Heo * @worker: self 927d302f017STejun Heo * @flags: flags to clear 928d302f017STejun Heo * 929e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 930d302f017STejun Heo */ 931d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 932d302f017STejun Heo { 93363d95a91STejun Heo struct worker_pool *pool = worker->pool; 934e22bee78STejun Heo unsigned int oflags = worker->flags; 935e22bee78STejun Heo 936bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 937cb444766STejun Heo 938d302f017STejun Heo worker->flags &= ~flags; 939e22bee78STejun Heo 94042c025f3STejun Heo /* 94142c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 94242c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 94342c025f3STejun Heo * of multiple flags, not a single flag. 94442c025f3STejun Heo */ 945e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 946e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 947bc35f7efSLai Jiangshan pool->nr_running++; 948d302f017STejun Heo } 949d302f017STejun Heo 95063638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 95163638450STejun Heo 95263638450STejun Heo /* 95363638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 95463638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 95563638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 95663638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 95763638450STejun Heo * should be using an unbound workqueue instead. 95863638450STejun Heo * 95963638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 96063638450STejun Heo * and report them so that they can be examined and converted to use unbound 96163638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 96263638450STejun Heo * function is tracked and reported with exponential backoff. 96363638450STejun Heo */ 96463638450STejun Heo #define WCI_MAX_ENTS 128 96563638450STejun Heo 96663638450STejun Heo struct wci_ent { 96763638450STejun Heo work_func_t func; 96863638450STejun Heo atomic64_t cnt; 96963638450STejun Heo struct hlist_node hash_node; 97063638450STejun Heo }; 97163638450STejun Heo 97263638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 97363638450STejun Heo static int wci_nr_ents; 97463638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 97563638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 97663638450STejun Heo 97763638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 97863638450STejun Heo { 97963638450STejun Heo struct wci_ent *ent; 98063638450STejun Heo 98163638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 98263638450STejun Heo (unsigned long)func) { 98363638450STejun Heo if (ent->func == func) 98463638450STejun Heo return ent; 98563638450STejun Heo } 98663638450STejun Heo return NULL; 98763638450STejun Heo } 98863638450STejun Heo 98963638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 99063638450STejun Heo { 99163638450STejun Heo struct wci_ent *ent; 99263638450STejun Heo 99363638450STejun Heo restart: 99463638450STejun Heo ent = wci_find_ent(func); 99563638450STejun Heo if (ent) { 99663638450STejun Heo u64 cnt; 99763638450STejun Heo 99863638450STejun Heo /* 99963638450STejun Heo * Start reporting from the fourth time and back off 100063638450STejun Heo * exponentially. 100163638450STejun Heo */ 100263638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 100363638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 100463638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 100563638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 100663638450STejun Heo atomic64_read(&ent->cnt)); 100763638450STejun Heo return; 100863638450STejun Heo } 100963638450STejun Heo 101063638450STejun Heo /* 101163638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 101263638450STejun Heo * is exhausted, something went really wrong and we probably made enough 101363638450STejun Heo * noise already. 101463638450STejun Heo */ 101563638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 101663638450STejun Heo return; 101763638450STejun Heo 101863638450STejun Heo raw_spin_lock(&wci_lock); 101963638450STejun Heo 102063638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 102163638450STejun Heo raw_spin_unlock(&wci_lock); 102263638450STejun Heo return; 102363638450STejun Heo } 102463638450STejun Heo 102563638450STejun Heo if (wci_find_ent(func)) { 102663638450STejun Heo raw_spin_unlock(&wci_lock); 102763638450STejun Heo goto restart; 102863638450STejun Heo } 102963638450STejun Heo 103063638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 103163638450STejun Heo ent->func = func; 103263638450STejun Heo atomic64_set(&ent->cnt, 1); 103363638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 103463638450STejun Heo 103563638450STejun Heo raw_spin_unlock(&wci_lock); 103663638450STejun Heo } 103763638450STejun Heo 103863638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 103963638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 104063638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 104163638450STejun Heo 1042c54d5046STejun Heo /** 10431da177e4SLinus Torvalds * wq_worker_running - a worker is running again 10441da177e4SLinus Torvalds * @task: task waking up 10451da177e4SLinus Torvalds * 10461da177e4SLinus Torvalds * This function is called when a worker returns from schedule() 10471da177e4SLinus Torvalds */ 10481da177e4SLinus Torvalds void wq_worker_running(struct task_struct *task) 10491da177e4SLinus Torvalds { 10501da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 10511da177e4SLinus Torvalds 1052c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 10531da177e4SLinus Torvalds return; 10541da177e4SLinus Torvalds 10551da177e4SLinus Torvalds /* 10561da177e4SLinus Torvalds * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 10571da177e4SLinus Torvalds * and the nr_running increment below, we may ruin the nr_running reset 10581da177e4SLinus Torvalds * and leave with an unexpected pool->nr_running == 1 on the newly unbound 10591da177e4SLinus Torvalds * pool. Protect against such race. 10601da177e4SLinus Torvalds */ 10611da177e4SLinus Torvalds preempt_disable(); 10621da177e4SLinus Torvalds if (!(worker->flags & WORKER_NOT_RUNNING)) 10631da177e4SLinus Torvalds worker->pool->nr_running++; 10641da177e4SLinus Torvalds preempt_enable(); 1065616db877STejun Heo 1066616db877STejun Heo /* 1067616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1068616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1069616db877STejun Heo */ 1070616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1071616db877STejun Heo 1072c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 10731da177e4SLinus Torvalds } 10741da177e4SLinus Torvalds 10751da177e4SLinus Torvalds /** 10761da177e4SLinus Torvalds * wq_worker_sleeping - a worker is going to sleep 10771da177e4SLinus Torvalds * @task: task going to sleep 10781da177e4SLinus Torvalds * 10791da177e4SLinus Torvalds * This function is called from schedule() when a busy worker is 10801da177e4SLinus Torvalds * going to sleep. 10811da177e4SLinus Torvalds */ 10821da177e4SLinus Torvalds void wq_worker_sleeping(struct task_struct *task) 10831da177e4SLinus Torvalds { 10841da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 10851da177e4SLinus Torvalds struct worker_pool *pool; 10861da177e4SLinus Torvalds 10871da177e4SLinus Torvalds /* 10881da177e4SLinus Torvalds * Rescuers, which may not have all the fields set up like normal 10891da177e4SLinus Torvalds * workers, also reach here, let's not access anything before 10901da177e4SLinus Torvalds * checking NOT_RUNNING. 10911da177e4SLinus Torvalds */ 10921da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) 10931da177e4SLinus Torvalds return; 10941da177e4SLinus Torvalds 10951da177e4SLinus Torvalds pool = worker->pool; 10961da177e4SLinus Torvalds 10971da177e4SLinus Torvalds /* Return if preempted before wq_worker_running() was reached */ 1098c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 10991da177e4SLinus Torvalds return; 11001da177e4SLinus Torvalds 1101c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 11021da177e4SLinus Torvalds raw_spin_lock_irq(&pool->lock); 11031da177e4SLinus Torvalds 11041da177e4SLinus Torvalds /* 11051da177e4SLinus Torvalds * Recheck in case unbind_workers() preempted us. We don't 11061da177e4SLinus Torvalds * want to decrement nr_running after the worker is unbound 11071da177e4SLinus Torvalds * and nr_running has been reset. 11081da177e4SLinus Torvalds */ 11091da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) { 11101da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 11111da177e4SLinus Torvalds return; 11121da177e4SLinus Torvalds } 11131da177e4SLinus Torvalds 11141da177e4SLinus Torvalds pool->nr_running--; 1115725e8ec5STejun Heo if (need_more_worker(pool)) { 1116725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 11171da177e4SLinus Torvalds wake_up_worker(pool); 1118725e8ec5STejun Heo } 11191da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 11201da177e4SLinus Torvalds } 11211da177e4SLinus Torvalds 11221da177e4SLinus Torvalds /** 1123616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1124616db877STejun Heo * @task: task currently running 1125616db877STejun Heo * 1126616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1127616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1128616db877STejun Heo */ 1129616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1130616db877STejun Heo { 1131616db877STejun Heo struct worker *worker = kthread_data(task); 1132616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1133616db877STejun Heo struct worker_pool *pool = worker->pool; 1134616db877STejun Heo 1135616db877STejun Heo if (!pwq) 1136616db877STejun Heo return; 1137616db877STejun Heo 11388a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 11398a1dd1e5STejun Heo 114018c8ae81SZqiang if (!wq_cpu_intensive_thresh_us) 114118c8ae81SZqiang return; 114218c8ae81SZqiang 1143616db877STejun Heo /* 1144616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1145616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1146616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1147c8f6219bSZqiang * 1148c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1149c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1150c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1151c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1152c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1153c8f6219bSZqiang * We probably want to make this prettier in the future. 1154616db877STejun Heo */ 1155c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1156616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1157616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1158616db877STejun Heo return; 1159616db877STejun Heo 1160616db877STejun Heo raw_spin_lock(&pool->lock); 1161616db877STejun Heo 1162616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 116363638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1164616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1165616db877STejun Heo 1166616db877STejun Heo if (need_more_worker(pool)) { 1167616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1168616db877STejun Heo wake_up_worker(pool); 1169616db877STejun Heo } 1170616db877STejun Heo 1171616db877STejun Heo raw_spin_unlock(&pool->lock); 1172616db877STejun Heo } 1173616db877STejun Heo 1174616db877STejun Heo /** 11751da177e4SLinus Torvalds * wq_worker_last_func - retrieve worker's last work function 11761da177e4SLinus Torvalds * @task: Task to retrieve last work function of. 11771da177e4SLinus Torvalds * 11781da177e4SLinus Torvalds * Determine the last function a worker executed. This is called from 11791da177e4SLinus Torvalds * the scheduler to get a worker's last known identity. 11801da177e4SLinus Torvalds * 11811da177e4SLinus Torvalds * CONTEXT: 11829b41ea72SAndrew Morton * raw_spin_lock_irq(rq->lock) 11831da177e4SLinus Torvalds * 11841da177e4SLinus Torvalds * This function is called during schedule() when a kworker is going 1185f756d5e2SNathan Lynch * to sleep. It's used by psi to identify aggregation workers during 1186f756d5e2SNathan Lynch * dequeuing, to allow periodic aggregation to shut-off when that 11871da177e4SLinus Torvalds * worker is the last task in the system or cgroup to go to sleep. 11881da177e4SLinus Torvalds * 11891da177e4SLinus Torvalds * As this function doesn't involve any workqueue-related locking, it 11901da177e4SLinus Torvalds * only returns stable values when called from inside the scheduler's 11911da177e4SLinus Torvalds * queuing and dequeuing paths, when @task, which must be a kworker, 11921da177e4SLinus Torvalds * is guaranteed to not be processing any works. 1193365970a1SDavid Howells * 1194365970a1SDavid Howells * Return: 1195365970a1SDavid Howells * The last work function %current executed as a worker, NULL if it 1196365970a1SDavid Howells * hasn't executed any work yet. 1197365970a1SDavid Howells */ 1198365970a1SDavid Howells work_func_t wq_worker_last_func(struct task_struct *task) 1199365970a1SDavid Howells { 1200365970a1SDavid Howells struct worker *worker = kthread_data(task); 1201365970a1SDavid Howells 1202365970a1SDavid Howells return worker->last_func; 1203365970a1SDavid Howells } 1204365970a1SDavid Howells 1205d302f017STejun Heo /** 12068cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1207c9e7cf27STejun Heo * @pool: pool of interest 12088cca0eeaSTejun Heo * @work: work to find worker for 12098cca0eeaSTejun Heo * 1210c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1211c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1212a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1213a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1214a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1215a2c1c57bSTejun Heo * being executed. 1216a2c1c57bSTejun Heo * 1217a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1218a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1219a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1220a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1221a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1222a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1223a2c1c57bSTejun Heo * 1224c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1225c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1226c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1227c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1228c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1229c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 12308cca0eeaSTejun Heo * 12318cca0eeaSTejun Heo * CONTEXT: 1232a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 12338cca0eeaSTejun Heo * 1234d185af30SYacine Belkadi * Return: 1235d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 12368cca0eeaSTejun Heo * otherwise. 12378cca0eeaSTejun Heo */ 1238c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 12398cca0eeaSTejun Heo struct work_struct *work) 12408cca0eeaSTejun Heo { 124142f8570fSSasha Levin struct worker *worker; 124242f8570fSSasha Levin 1243b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1244a2c1c57bSTejun Heo (unsigned long)work) 1245a2c1c57bSTejun Heo if (worker->current_work == work && 1246a2c1c57bSTejun Heo worker->current_func == work->func) 124742f8570fSSasha Levin return worker; 124842f8570fSSasha Levin 124942f8570fSSasha Levin return NULL; 12508cca0eeaSTejun Heo } 12518cca0eeaSTejun Heo 12528cca0eeaSTejun Heo /** 1253bf4ede01STejun Heo * move_linked_works - move linked works to a list 1254bf4ede01STejun Heo * @work: start of series of works to be scheduled 1255bf4ede01STejun Heo * @head: target list to append @work to 1256402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1257bf4ede01STejun Heo * 1258bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1259bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1260bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1261bf4ede01STejun Heo * 1262bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1263bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1264bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1265bf4ede01STejun Heo * 1266bf4ede01STejun Heo * CONTEXT: 1267a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1268bf4ede01STejun Heo */ 1269bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1270bf4ede01STejun Heo struct work_struct **nextp) 1271bf4ede01STejun Heo { 1272bf4ede01STejun Heo struct work_struct *n; 1273bf4ede01STejun Heo 1274bf4ede01STejun Heo /* 1275bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1276bf4ede01STejun Heo * use NULL for list head. 1277bf4ede01STejun Heo */ 1278bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1279bf4ede01STejun Heo list_move_tail(&work->entry, head); 1280bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1281bf4ede01STejun Heo break; 1282bf4ede01STejun Heo } 1283bf4ede01STejun Heo 1284bf4ede01STejun Heo /* 1285bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1286bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1287bf4ede01STejun Heo * needs to be updated. 1288bf4ede01STejun Heo */ 1289bf4ede01STejun Heo if (nextp) 1290bf4ede01STejun Heo *nextp = n; 1291bf4ede01STejun Heo } 1292bf4ede01STejun Heo 12938864b4e5STejun Heo /** 12948864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 12958864b4e5STejun Heo * @pwq: pool_workqueue to get 12968864b4e5STejun Heo * 12978864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 12988864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 12998864b4e5STejun Heo */ 13008864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 13018864b4e5STejun Heo { 13028864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13038864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 13048864b4e5STejun Heo pwq->refcnt++; 13058864b4e5STejun Heo } 13068864b4e5STejun Heo 13078864b4e5STejun Heo /** 13088864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 13098864b4e5STejun Heo * @pwq: pool_workqueue to put 13108864b4e5STejun Heo * 13118864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 13128864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 13138864b4e5STejun Heo */ 13148864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 13158864b4e5STejun Heo { 13168864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13178864b4e5STejun Heo if (likely(--pwq->refcnt)) 13188864b4e5STejun Heo return; 13198864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 13208864b4e5STejun Heo return; 13218864b4e5STejun Heo /* 13228864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 13238864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 13248864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 13258864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 13268864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 13278864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 13288864b4e5STejun Heo */ 13298864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 13308864b4e5STejun Heo } 13318864b4e5STejun Heo 1332dce90d47STejun Heo /** 1333dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1334dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1335dce90d47STejun Heo * 1336dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1337dce90d47STejun Heo */ 1338dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1339dce90d47STejun Heo { 1340dce90d47STejun Heo if (pwq) { 1341dce90d47STejun Heo /* 134224acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1343dce90d47STejun Heo * following lock operations are safe. 1344dce90d47STejun Heo */ 1345a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1346dce90d47STejun Heo put_pwq(pwq); 1347a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1348dce90d47STejun Heo } 1349dce90d47STejun Heo } 1350dce90d47STejun Heo 1351f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1352bf4ede01STejun Heo { 1353112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1354bf4ede01STejun Heo 1355bf4ede01STejun Heo trace_workqueue_activate_work(work); 135682607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 135782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1358112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1359f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1360112202d9STejun Heo pwq->nr_active++; 1361bf4ede01STejun Heo } 1362bf4ede01STejun Heo 1363f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 13643aa62497SLai Jiangshan { 1365f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 13663aa62497SLai Jiangshan struct work_struct, entry); 13673aa62497SLai Jiangshan 1368f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13693aa62497SLai Jiangshan } 13703aa62497SLai Jiangshan 1371bf4ede01STejun Heo /** 1372112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1373112202d9STejun Heo * @pwq: pwq of interest 1374c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1375bf4ede01STejun Heo * 1376bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1377112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1378bf4ede01STejun Heo * 1379bf4ede01STejun Heo * CONTEXT: 1380a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1381bf4ede01STejun Heo */ 1382c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1383bf4ede01STejun Heo { 1384c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1385c4560c2cSLai Jiangshan 1386018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1387112202d9STejun Heo pwq->nr_active--; 1388f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1389f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1390112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1391f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1392bf4ede01STejun Heo } 1393018f3a13SLai Jiangshan } 1394018f3a13SLai Jiangshan 1395018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1396bf4ede01STejun Heo 1397bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1398112202d9STejun Heo if (likely(pwq->flush_color != color)) 13998864b4e5STejun Heo goto out_put; 1400bf4ede01STejun Heo 1401bf4ede01STejun Heo /* are there still in-flight works? */ 1402112202d9STejun Heo if (pwq->nr_in_flight[color]) 14038864b4e5STejun Heo goto out_put; 1404bf4ede01STejun Heo 1405112202d9STejun Heo /* this pwq is done, clear flush_color */ 1406112202d9STejun Heo pwq->flush_color = -1; 1407bf4ede01STejun Heo 1408bf4ede01STejun Heo /* 1409112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1410bf4ede01STejun Heo * will handle the rest. 1411bf4ede01STejun Heo */ 1412112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1413112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 14148864b4e5STejun Heo out_put: 14158864b4e5STejun Heo put_pwq(pwq); 1416bf4ede01STejun Heo } 1417bf4ede01STejun Heo 141836e227d2STejun Heo /** 1419bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 142036e227d2STejun Heo * @work: work item to steal 142136e227d2STejun Heo * @is_dwork: @work is a delayed_work 1422bbb68dfaSTejun Heo * @flags: place to store irq state 142336e227d2STejun Heo * 142436e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1425d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 142636e227d2STejun Heo * 1427d185af30SYacine Belkadi * Return: 14283eb6b31bSMauro Carvalho Chehab * 14293eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 143036e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 143136e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 143236e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1433bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1434bbb68dfaSTejun Heo * for arbitrarily long 14353eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 143636e227d2STejun Heo * 1437d185af30SYacine Belkadi * Note: 1438bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1439e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1440e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1441e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1442bbb68dfaSTejun Heo * 1443bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1444bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1445bbb68dfaSTejun Heo * 1446e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1447bf4ede01STejun Heo */ 1448bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1449bbb68dfaSTejun Heo unsigned long *flags) 1450bf4ede01STejun Heo { 1451d565ed63STejun Heo struct worker_pool *pool; 1452112202d9STejun Heo struct pool_workqueue *pwq; 1453bf4ede01STejun Heo 1454bbb68dfaSTejun Heo local_irq_save(*flags); 1455bbb68dfaSTejun Heo 145636e227d2STejun Heo /* try to steal the timer if it exists */ 145736e227d2STejun Heo if (is_dwork) { 145836e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 145936e227d2STejun Heo 1460e0aecdd8STejun Heo /* 1461e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1462e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1463e0aecdd8STejun Heo * running on the local CPU. 1464e0aecdd8STejun Heo */ 146536e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 146636e227d2STejun Heo return 1; 146736e227d2STejun Heo } 146836e227d2STejun Heo 146936e227d2STejun Heo /* try to claim PENDING the normal way */ 1470bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1471bf4ede01STejun Heo return 0; 1472bf4ede01STejun Heo 147324acfb71SThomas Gleixner rcu_read_lock(); 1474bf4ede01STejun Heo /* 1475bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1476bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1477bf4ede01STejun Heo */ 1478d565ed63STejun Heo pool = get_work_pool(work); 1479d565ed63STejun Heo if (!pool) 1480bbb68dfaSTejun Heo goto fail; 1481bf4ede01STejun Heo 1482a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1483bf4ede01STejun Heo /* 1484112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1485112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1486112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1487112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1488112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 14890b3dae68SLai Jiangshan * item is currently queued on that pool. 1490bf4ede01STejun Heo */ 1491112202d9STejun Heo pwq = get_work_pwq(work); 1492112202d9STejun Heo if (pwq && pwq->pool == pool) { 1493bf4ede01STejun Heo debug_work_deactivate(work); 14943aa62497SLai Jiangshan 14953aa62497SLai Jiangshan /* 1496018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1497018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1498018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1499018f3a13SLai Jiangshan * 1500f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1501d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1502f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 150316062836STejun Heo * management later on and cause stall. Make sure the work 150416062836STejun Heo * item is activated before grabbing. 15053aa62497SLai Jiangshan */ 1506f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1507f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 15083aa62497SLai Jiangshan 1509bf4ede01STejun Heo list_del_init(&work->entry); 1510c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 151136e227d2STejun Heo 1512112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 15134468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 15144468a00fSLai Jiangshan 1515a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 151624acfb71SThomas Gleixner rcu_read_unlock(); 151736e227d2STejun Heo return 1; 1518bf4ede01STejun Heo } 1519a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1520bbb68dfaSTejun Heo fail: 152124acfb71SThomas Gleixner rcu_read_unlock(); 1522bbb68dfaSTejun Heo local_irq_restore(*flags); 1523bbb68dfaSTejun Heo if (work_is_canceling(work)) 1524bbb68dfaSTejun Heo return -ENOENT; 1525bbb68dfaSTejun Heo cpu_relax(); 152636e227d2STejun Heo return -EAGAIN; 1527bf4ede01STejun Heo } 1528bf4ede01STejun Heo 1529bf4ede01STejun Heo /** 1530706026c2STejun Heo * insert_work - insert a work into a pool 1531112202d9STejun Heo * @pwq: pwq @work belongs to 15324690c4abSTejun Heo * @work: work to insert 15334690c4abSTejun Heo * @head: insertion point 15344690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 15354690c4abSTejun Heo * 1536112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1537706026c2STejun Heo * work_struct flags. 15384690c4abSTejun Heo * 15394690c4abSTejun Heo * CONTEXT: 1540a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1541365970a1SDavid Howells */ 1542112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1543112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1544b89deed3SOleg Nesterov { 1545fe089f87STejun Heo debug_work_activate(work); 1546e1d8aa9fSFrederic Weisbecker 1547e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1548f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1549e89a85d6SWalter Wu 15504690c4abSTejun Heo /* we own @work, set data and link */ 1551112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 15521a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 15538864b4e5STejun Heo get_pwq(pwq); 1554b89deed3SOleg Nesterov } 1555b89deed3SOleg Nesterov 1556c8efcc25STejun Heo /* 1557c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 15588d03ecfeSTejun Heo * same workqueue. 1559c8efcc25STejun Heo */ 1560c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1561c8efcc25STejun Heo { 1562c8efcc25STejun Heo struct worker *worker; 1563c8efcc25STejun Heo 15648d03ecfeSTejun Heo worker = current_wq_worker(); 1565c8efcc25STejun Heo /* 1566bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 15678d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1568c8efcc25STejun Heo */ 1569112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1570c8efcc25STejun Heo } 1571c8efcc25STejun Heo 1572ef557180SMike Galbraith /* 1573ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1574ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1575ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1576ef557180SMike Galbraith */ 1577ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1578ef557180SMike Galbraith { 1579ef557180SMike Galbraith int new_cpu; 1580ef557180SMike Galbraith 1581f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1582ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1583ef557180SMike Galbraith return cpu; 1584a8ec5880SAmmar Faizi } else { 1585a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1586f303fccbSTejun Heo } 1587f303fccbSTejun Heo 1588ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1589ef557180SMike Galbraith return cpu; 1590ef557180SMike Galbraith 1591ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1592ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1593ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1594ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1595ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1596ef557180SMike Galbraith return cpu; 1597ef557180SMike Galbraith } 1598ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1599ef557180SMike Galbraith 1600ef557180SMike Galbraith return new_cpu; 1601ef557180SMike Galbraith } 1602ef557180SMike Galbraith 1603d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 16041da177e4SLinus Torvalds struct work_struct *work) 16051da177e4SLinus Torvalds { 1606112202d9STejun Heo struct pool_workqueue *pwq; 1607fe089f87STejun Heo struct worker_pool *last_pool, *pool; 16088a2e8e5dSTejun Heo unsigned int work_flags; 1609b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 16108930cabaSTejun Heo 16118930cabaSTejun Heo /* 16128930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 16138930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 16148930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 16158930cabaSTejun Heo * happen with IRQ disabled. 16168930cabaSTejun Heo */ 16178e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 16181da177e4SLinus Torvalds 16191e19ffc6STejun Heo 162033e3f0a3SRichard Clark /* 162133e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 162233e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 162333e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 162433e3f0a3SRichard Clark */ 162533e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 162633e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1627e41e704bSTejun Heo return; 162824acfb71SThomas Gleixner rcu_read_lock(); 16299e8cd2f5STejun Heo retry: 1630aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1631aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1632df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1633ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1634df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1635aa202f1fSHillf Danton } else { 1636aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1637aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1638*ee1ceef7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwq, cpu); 1639aa202f1fSHillf Danton } 1640f3421797STejun Heo 1641fe089f87STejun Heo pool = pwq->pool; 1642fe089f87STejun Heo 164318aa9effSTejun Heo /* 1644c9178087STejun Heo * If @work was previously on a different pool, it might still be 1645c9178087STejun Heo * running there, in which case the work needs to be queued on that 1646c9178087STejun Heo * pool to guarantee non-reentrancy. 164718aa9effSTejun Heo */ 1648c9e7cf27STejun Heo last_pool = get_work_pool(work); 1649fe089f87STejun Heo if (last_pool && last_pool != pool) { 165018aa9effSTejun Heo struct worker *worker; 165118aa9effSTejun Heo 1652a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 165318aa9effSTejun Heo 1654c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 165518aa9effSTejun Heo 1656112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1657c9178087STejun Heo pwq = worker->current_pwq; 1658fe089f87STejun Heo pool = pwq->pool; 1659fe089f87STejun Heo WARN_ON_ONCE(pool != last_pool); 16608594fadeSLai Jiangshan } else { 166118aa9effSTejun Heo /* meh... not running there, queue here */ 1662a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1663fe089f87STejun Heo raw_spin_lock(&pool->lock); 166418aa9effSTejun Heo } 16658930cabaSTejun Heo } else { 1666fe089f87STejun Heo raw_spin_lock(&pool->lock); 16678930cabaSTejun Heo } 1668502ca9d8STejun Heo 16699e8cd2f5STejun Heo /* 16709e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 16719e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 16729e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1673df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1674df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 16759e8cd2f5STejun Heo * make forward-progress. 16769e8cd2f5STejun Heo */ 16779e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 16789e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1679fe089f87STejun Heo raw_spin_unlock(&pool->lock); 16809e8cd2f5STejun Heo cpu_relax(); 16819e8cd2f5STejun Heo goto retry; 16829e8cd2f5STejun Heo } 16839e8cd2f5STejun Heo /* oops */ 16849e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 16859e8cd2f5STejun Heo wq->name, cpu); 16869e8cd2f5STejun Heo } 16879e8cd2f5STejun Heo 1688112202d9STejun Heo /* pwq determined, queue */ 1689112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1690502ca9d8STejun Heo 169124acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 169224acfb71SThomas Gleixner goto out; 16931e19ffc6STejun Heo 1694112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1695112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 16961e19ffc6STejun Heo 1697112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1698fe089f87STejun Heo if (list_empty(&pool->worklist)) 1699fe089f87STejun Heo pool->watchdog_ts = jiffies; 1700fe089f87STejun Heo 1701cdadf009STejun Heo trace_workqueue_activate_work(work); 1702112202d9STejun Heo pwq->nr_active++; 1703fe089f87STejun Heo insert_work(pwq, work, &pool->worklist, work_flags); 1704fe089f87STejun Heo 1705fe089f87STejun Heo if (__need_more_worker(pool)) 1706fe089f87STejun Heo wake_up_worker(pool); 17078a2e8e5dSTejun Heo } else { 1708f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1709fe089f87STejun Heo insert_work(pwq, work, &pwq->inactive_works, work_flags); 17108a2e8e5dSTejun Heo } 17111e19ffc6STejun Heo 171224acfb71SThomas Gleixner out: 1713fe089f87STejun Heo raw_spin_unlock(&pool->lock); 171424acfb71SThomas Gleixner rcu_read_unlock(); 17151da177e4SLinus Torvalds } 17161da177e4SLinus Torvalds 17170fcb78c2SRolf Eike Beer /** 1718c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1719c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1720c1a220e7SZhang Rui * @wq: workqueue to use 1721c1a220e7SZhang Rui * @work: work to queue 1722c1a220e7SZhang Rui * 1723c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1724443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1725443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1726854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1727854f5cc5SPaul E. McKenney * online will get a splat. 1728d185af30SYacine Belkadi * 1729d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1730c1a220e7SZhang Rui */ 1731d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1732d4283e93STejun Heo struct work_struct *work) 1733c1a220e7SZhang Rui { 1734d4283e93STejun Heo bool ret = false; 17358930cabaSTejun Heo unsigned long flags; 17368930cabaSTejun Heo 17378930cabaSTejun Heo local_irq_save(flags); 1738c1a220e7SZhang Rui 173922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17404690c4abSTejun Heo __queue_work(cpu, wq, work); 1741d4283e93STejun Heo ret = true; 1742c1a220e7SZhang Rui } 17438930cabaSTejun Heo 17448930cabaSTejun Heo local_irq_restore(flags); 1745c1a220e7SZhang Rui return ret; 1746c1a220e7SZhang Rui } 1747ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1748c1a220e7SZhang Rui 17498204e0c1SAlexander Duyck /** 17508204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 17518204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 17528204e0c1SAlexander Duyck * 17538204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 17548204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 17558204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 17568204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 17578204e0c1SAlexander Duyck */ 17588204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 17598204e0c1SAlexander Duyck { 17608204e0c1SAlexander Duyck int cpu; 17618204e0c1SAlexander Duyck 17628204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 17638204e0c1SAlexander Duyck if (!wq_numa_enabled) 17648204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17658204e0c1SAlexander Duyck 17668204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 17678204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 17688204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 17698204e0c1SAlexander Duyck 17708204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 17718204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 17728204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 17738204e0c1SAlexander Duyck return cpu; 17748204e0c1SAlexander Duyck 17758204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 17768204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 17778204e0c1SAlexander Duyck 17788204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 17798204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 17808204e0c1SAlexander Duyck } 17818204e0c1SAlexander Duyck 17828204e0c1SAlexander Duyck /** 17838204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 17848204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 17858204e0c1SAlexander Duyck * @wq: workqueue to use 17868204e0c1SAlexander Duyck * @work: work to queue 17878204e0c1SAlexander Duyck * 17888204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 17898204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 17908204e0c1SAlexander Duyck * NUMA node. 17918204e0c1SAlexander Duyck * 17928204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 17938204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 17948204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 17958204e0c1SAlexander Duyck * 17968204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 17978204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 17988204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 17998204e0c1SAlexander Duyck * 18008204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 18018204e0c1SAlexander Duyck */ 18028204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18038204e0c1SAlexander Duyck struct work_struct *work) 18048204e0c1SAlexander Duyck { 18058204e0c1SAlexander Duyck unsigned long flags; 18068204e0c1SAlexander Duyck bool ret = false; 18078204e0c1SAlexander Duyck 18088204e0c1SAlexander Duyck /* 18098204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 18108204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 18118204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 18128204e0c1SAlexander Duyck * 18138204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 18148204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 18158204e0c1SAlexander Duyck * some round robin type logic. 18168204e0c1SAlexander Duyck */ 18178204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 18188204e0c1SAlexander Duyck 18198204e0c1SAlexander Duyck local_irq_save(flags); 18208204e0c1SAlexander Duyck 18218204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18228204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 18238204e0c1SAlexander Duyck 18248204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 18258204e0c1SAlexander Duyck ret = true; 18268204e0c1SAlexander Duyck } 18278204e0c1SAlexander Duyck 18288204e0c1SAlexander Duyck local_irq_restore(flags); 18298204e0c1SAlexander Duyck return ret; 18308204e0c1SAlexander Duyck } 18318204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 18328204e0c1SAlexander Duyck 18338c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 18341da177e4SLinus Torvalds { 18358c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 18361da177e4SLinus Torvalds 1837e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 183860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 18391da177e4SLinus Torvalds } 18401438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 18411da177e4SLinus Torvalds 18427beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 184352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18441da177e4SLinus Torvalds { 18457beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 18467beb2edfSTejun Heo struct work_struct *work = &dwork->work; 18471da177e4SLinus Torvalds 1848637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 18494b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1850fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1851fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 18527beb2edfSTejun Heo 18538852aac2STejun Heo /* 18548852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 18558852aac2STejun Heo * both optimization and correctness. The earliest @timer can 18568852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 18578852aac2STejun Heo * on that there's no such delay when @delay is 0. 18588852aac2STejun Heo */ 18598852aac2STejun Heo if (!delay) { 18608852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 18618852aac2STejun Heo return; 18628852aac2STejun Heo } 18638852aac2STejun Heo 186460c057bcSLai Jiangshan dwork->wq = wq; 18651265057fSTejun Heo dwork->cpu = cpu; 18667beb2edfSTejun Heo timer->expires = jiffies + delay; 18677beb2edfSTejun Heo 1868041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 18697beb2edfSTejun Heo add_timer_on(timer, cpu); 1870041bd12eSTejun Heo else 1871041bd12eSTejun Heo add_timer(timer); 18727beb2edfSTejun Heo } 18731da177e4SLinus Torvalds 18740fcb78c2SRolf Eike Beer /** 18750fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 18760fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 18770fcb78c2SRolf Eike Beer * @wq: workqueue to use 1878af9997e4SRandy Dunlap * @dwork: work to queue 18790fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 18800fcb78c2SRolf Eike Beer * 1881d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1882715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1883715f1300STejun Heo * execution. 18840fcb78c2SRolf Eike Beer */ 1885d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 188652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 18877a6bc1cdSVenkatesh Pallipadi { 188852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1889d4283e93STejun Heo bool ret = false; 18908930cabaSTejun Heo unsigned long flags; 18918930cabaSTejun Heo 18928930cabaSTejun Heo /* read the comment in __queue_work() */ 18938930cabaSTejun Heo local_irq_save(flags); 18947a6bc1cdSVenkatesh Pallipadi 189522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18967beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1897d4283e93STejun Heo ret = true; 18987a6bc1cdSVenkatesh Pallipadi } 18998930cabaSTejun Heo 19008930cabaSTejun Heo local_irq_restore(flags); 19017a6bc1cdSVenkatesh Pallipadi return ret; 19027a6bc1cdSVenkatesh Pallipadi } 1903ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19041da177e4SLinus Torvalds 1905c8e55f36STejun Heo /** 19068376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19078376fe22STejun Heo * @cpu: CPU number to execute work on 19088376fe22STejun Heo * @wq: workqueue to use 19098376fe22STejun Heo * @dwork: work to queue 19108376fe22STejun Heo * @delay: number of jiffies to wait before queueing 19118376fe22STejun Heo * 19128376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 19138376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 19148376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 19158376fe22STejun Heo * current state. 19168376fe22STejun Heo * 1917d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 19188376fe22STejun Heo * pending and its timer was modified. 19198376fe22STejun Heo * 1920e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 19218376fe22STejun Heo * See try_to_grab_pending() for details. 19228376fe22STejun Heo */ 19238376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 19248376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 19258376fe22STejun Heo { 19268376fe22STejun Heo unsigned long flags; 19278376fe22STejun Heo int ret; 19288376fe22STejun Heo 19298376fe22STejun Heo do { 19308376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 19318376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 19328376fe22STejun Heo 19338376fe22STejun Heo if (likely(ret >= 0)) { 19348376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 19358376fe22STejun Heo local_irq_restore(flags); 19368376fe22STejun Heo } 19378376fe22STejun Heo 19388376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 19398376fe22STejun Heo return ret; 19408376fe22STejun Heo } 19418376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 19428376fe22STejun Heo 194305f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 194405f0fe6bSTejun Heo { 194505f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 194605f0fe6bSTejun Heo 194705f0fe6bSTejun Heo /* read the comment in __queue_work() */ 194805f0fe6bSTejun Heo local_irq_disable(); 194905f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 195005f0fe6bSTejun Heo local_irq_enable(); 195105f0fe6bSTejun Heo } 195205f0fe6bSTejun Heo 195305f0fe6bSTejun Heo /** 195405f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 195505f0fe6bSTejun Heo * @wq: workqueue to use 195605f0fe6bSTejun Heo * @rwork: work to queue 195705f0fe6bSTejun Heo * 195805f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 195905f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1960bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 196105f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 196205f0fe6bSTejun Heo */ 196305f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 196405f0fe6bSTejun Heo { 196505f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 196605f0fe6bSTejun Heo 196705f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 196805f0fe6bSTejun Heo rwork->wq = wq; 1969a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 197005f0fe6bSTejun Heo return true; 197105f0fe6bSTejun Heo } 197205f0fe6bSTejun Heo 197305f0fe6bSTejun Heo return false; 197405f0fe6bSTejun Heo } 197505f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 197605f0fe6bSTejun Heo 19778376fe22STejun Heo /** 1978c8e55f36STejun Heo * worker_enter_idle - enter idle state 1979c8e55f36STejun Heo * @worker: worker which is entering idle state 1980c8e55f36STejun Heo * 1981c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1982c8e55f36STejun Heo * necessary. 1983c8e55f36STejun Heo * 1984c8e55f36STejun Heo * LOCKING: 1985a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1986c8e55f36STejun Heo */ 1987c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 19881da177e4SLinus Torvalds { 1989bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1990c8e55f36STejun Heo 19916183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 19926183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 19936183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 19946183c009STejun Heo return; 1995c8e55f36STejun Heo 1996051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1997cb444766STejun Heo worker->flags |= WORKER_IDLE; 1998bd7bdd43STejun Heo pool->nr_idle++; 1999e22bee78STejun Heo worker->last_active = jiffies; 2000c8e55f36STejun Heo 2001c8e55f36STejun Heo /* idle_list is LIFO */ 2002bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 2003db7bccf4STejun Heo 200463d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 2005628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 2006cb444766STejun Heo 2007989442d7SLai Jiangshan /* Sanity check nr_running. */ 2008bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 2009c8e55f36STejun Heo } 2010c8e55f36STejun Heo 2011c8e55f36STejun Heo /** 2012c8e55f36STejun Heo * worker_leave_idle - leave idle state 2013c8e55f36STejun Heo * @worker: worker which is leaving idle state 2014c8e55f36STejun Heo * 2015c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 2016c8e55f36STejun Heo * 2017c8e55f36STejun Heo * LOCKING: 2018a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2019c8e55f36STejun Heo */ 2020c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 2021c8e55f36STejun Heo { 2022bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2023c8e55f36STejun Heo 20246183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 20256183c009STejun Heo return; 2026d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 2027bd7bdd43STejun Heo pool->nr_idle--; 2028c8e55f36STejun Heo list_del_init(&worker->entry); 2029c8e55f36STejun Heo } 2030c8e55f36STejun Heo 2031f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 2032c34056a3STejun Heo { 2033c34056a3STejun Heo struct worker *worker; 2034c34056a3STejun Heo 2035f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2036c8e55f36STejun Heo if (worker) { 2037c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 2038affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 2039da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2040e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2041e22bee78STejun Heo worker->flags = WORKER_PREP; 2042c8e55f36STejun Heo } 2043c34056a3STejun Heo return worker; 2044c34056a3STejun Heo } 2045c34056a3STejun Heo 2046c34056a3STejun Heo /** 20474736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20484736cbf7SLai Jiangshan * @worker: worker to be attached 20494736cbf7SLai Jiangshan * @pool: the target pool 20504736cbf7SLai Jiangshan * 20514736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20524736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20534736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20544736cbf7SLai Jiangshan */ 20554736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 20564736cbf7SLai Jiangshan struct worker_pool *pool) 20574736cbf7SLai Jiangshan { 20581258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 20594736cbf7SLai Jiangshan 20604736cbf7SLai Jiangshan /* 20611258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 20621258fae7STejun Heo * stable across this function. See the comments above the flag 20631258fae7STejun Heo * definition for details. 20644736cbf7SLai Jiangshan */ 20654736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 20664736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 20675c25b5ffSPeter Zijlstra else 20685c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 20694736cbf7SLai Jiangshan 2070640f17c8SPeter Zijlstra if (worker->rescue_wq) 2071640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 2072640f17c8SPeter Zijlstra 20734736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2074a2d812a2STejun Heo worker->pool = pool; 20754736cbf7SLai Jiangshan 20761258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 20774736cbf7SLai Jiangshan } 20784736cbf7SLai Jiangshan 20794736cbf7SLai Jiangshan /** 208060f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 208160f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 208260f5a4bcSLai Jiangshan * 20834736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 20844736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 20854736cbf7SLai Jiangshan * other reference to the pool. 208660f5a4bcSLai Jiangshan */ 2087a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 208860f5a4bcSLai Jiangshan { 2089a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 209060f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 209160f5a4bcSLai Jiangshan 20921258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2093a2d812a2STejun Heo 20945c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2095da028469SLai Jiangshan list_del(&worker->node); 2096a2d812a2STejun Heo worker->pool = NULL; 2097a2d812a2STejun Heo 2098e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 209960f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 21001258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 210160f5a4bcSLai Jiangshan 2102b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2103b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2104b62c0751SLai Jiangshan 210560f5a4bcSLai Jiangshan if (detach_completion) 210660f5a4bcSLai Jiangshan complete(detach_completion); 210760f5a4bcSLai Jiangshan } 210860f5a4bcSLai Jiangshan 210960f5a4bcSLai Jiangshan /** 2110c34056a3STejun Heo * create_worker - create a new workqueue worker 211163d95a91STejun Heo * @pool: pool the new worker will belong to 2112c34056a3STejun Heo * 2113051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2114c34056a3STejun Heo * 2115c34056a3STejun Heo * CONTEXT: 2116c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2117c34056a3STejun Heo * 2118d185af30SYacine Belkadi * Return: 2119c34056a3STejun Heo * Pointer to the newly created worker. 2120c34056a3STejun Heo */ 2121bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2122c34056a3STejun Heo { 2123e441b56fSZhen Lei struct worker *worker; 2124e441b56fSZhen Lei int id; 2125e3c916a4STejun Heo char id_buf[16]; 2126c34056a3STejun Heo 21277cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2128e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 21293f0ea0b8SPetr Mladek if (id < 0) { 21303f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 21313f0ea0b8SPetr Mladek ERR_PTR(id)); 2132e441b56fSZhen Lei return NULL; 21333f0ea0b8SPetr Mladek } 2134c34056a3STejun Heo 2135f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 21363f0ea0b8SPetr Mladek if (!worker) { 21373f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2138c34056a3STejun Heo goto fail; 21393f0ea0b8SPetr Mladek } 2140c34056a3STejun Heo 2141c34056a3STejun Heo worker->id = id; 2142c34056a3STejun Heo 214329c91e99STejun Heo if (pool->cpu >= 0) 2144e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2145e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2146f3421797STejun Heo else 2147e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2148e3c916a4STejun Heo 2149f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2150e3c916a4STejun Heo "kworker/%s", id_buf); 21513f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 215260f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 215360f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 215460f54038SPetr Mladek id_buf); 215560f54038SPetr Mladek } else { 21563f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 21573f0ea0b8SPetr Mladek worker->task); 215860f54038SPetr Mladek } 2159c34056a3STejun Heo goto fail; 21603f0ea0b8SPetr Mladek } 2161c34056a3STejun Heo 216291151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 216325834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 216491151228SOleg Nesterov 2165da028469SLai Jiangshan /* successful, attach the worker to the pool */ 21664736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2167822d8405STejun Heo 2168051e1850SLai Jiangshan /* start the newly created worker */ 2169a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2170051e1850SLai Jiangshan worker->pool->nr_workers++; 2171051e1850SLai Jiangshan worker_enter_idle(worker); 2172051e1850SLai Jiangshan wake_up_process(worker->task); 2173a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2174051e1850SLai Jiangshan 2175c34056a3STejun Heo return worker; 2176822d8405STejun Heo 2177c34056a3STejun Heo fail: 2178e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2179c34056a3STejun Heo kfree(worker); 2180c34056a3STejun Heo return NULL; 2181c34056a3STejun Heo } 2182c34056a3STejun Heo 2183793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2184793777bcSValentin Schneider { 2185793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2186793777bcSValentin Schneider 2187793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2188793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2189793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2190793777bcSValentin Schneider else 2191793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2192793777bcSValentin Schneider } 2193793777bcSValentin Schneider 2194e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2195e02b9312SValentin Schneider { 2196e02b9312SValentin Schneider struct worker *worker, *tmp; 2197e02b9312SValentin Schneider 2198e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2199e02b9312SValentin Schneider list_del_init(&worker->entry); 2200e02b9312SValentin Schneider unbind_worker(worker); 2201e02b9312SValentin Schneider /* 2202e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2203e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2204e02b9312SValentin Schneider * wouldn't have gotten here. 2205c34056a3STejun Heo * 2206e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2207e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2208e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2209e02b9312SValentin Schneider * outside of pool->lock. 2210e02b9312SValentin Schneider */ 2211e02b9312SValentin Schneider wake_up_process(worker->task); 2212e02b9312SValentin Schneider } 2213e02b9312SValentin Schneider } 2214e02b9312SValentin Schneider 2215e02b9312SValentin Schneider /** 2216e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2217e02b9312SValentin Schneider * @worker: worker to be destroyed 2218e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2219e02b9312SValentin Schneider * 2220e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2221e02b9312SValentin Schneider * should be idle. 2222c8e55f36STejun Heo * 2223c8e55f36STejun Heo * CONTEXT: 2224a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2225c34056a3STejun Heo */ 2226e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2227c34056a3STejun Heo { 2228bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2229c34056a3STejun Heo 2230cd549687STejun Heo lockdep_assert_held(&pool->lock); 2231e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2232cd549687STejun Heo 2233c34056a3STejun Heo /* sanity check frenzy */ 22346183c009STejun Heo if (WARN_ON(worker->current_work) || 223573eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 223673eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 22376183c009STejun Heo return; 2238c34056a3STejun Heo 2239bd7bdd43STejun Heo pool->nr_workers--; 2240bd7bdd43STejun Heo pool->nr_idle--; 2241c8e55f36STejun Heo 2242cb444766STejun Heo worker->flags |= WORKER_DIE; 2243e02b9312SValentin Schneider 2244e02b9312SValentin Schneider list_move(&worker->entry, list); 2245e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2246c34056a3STejun Heo } 2247c34056a3STejun Heo 22483f959aa3SValentin Schneider /** 22493f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 22503f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 22513f959aa3SValentin Schneider * 22523f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 22533f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 22543f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 22553f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 22563f959aa3SValentin Schneider * it expire and re-evaluate things from there. 22573f959aa3SValentin Schneider */ 225832a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2259e22bee78STejun Heo { 226032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 22613f959aa3SValentin Schneider bool do_cull = false; 22623f959aa3SValentin Schneider 22633f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 22643f959aa3SValentin Schneider return; 22653f959aa3SValentin Schneider 22663f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 22673f959aa3SValentin Schneider 22683f959aa3SValentin Schneider if (too_many_workers(pool)) { 22693f959aa3SValentin Schneider struct worker *worker; 22703f959aa3SValentin Schneider unsigned long expires; 22713f959aa3SValentin Schneider 22723f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 22733f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 22743f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 22753f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 22763f959aa3SValentin Schneider 22773f959aa3SValentin Schneider if (!do_cull) 22783f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 22793f959aa3SValentin Schneider } 22803f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 22813f959aa3SValentin Schneider 22823f959aa3SValentin Schneider if (do_cull) 22833f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 22843f959aa3SValentin Schneider } 22853f959aa3SValentin Schneider 22863f959aa3SValentin Schneider /** 22873f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 22883f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 22893f959aa3SValentin Schneider * 22903f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 22913f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2292e02b9312SValentin Schneider * 2293e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2294e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2295e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 22963f959aa3SValentin Schneider */ 22973f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 22983f959aa3SValentin Schneider { 22993f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 23009680540cSYang Yingliang LIST_HEAD(cull_list); 2301e22bee78STejun Heo 2302e02b9312SValentin Schneider /* 2303e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2304e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2305e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2306e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2307e02b9312SValentin Schneider */ 2308e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2309a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2310e22bee78STejun Heo 23113347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2312e22bee78STejun Heo struct worker *worker; 2313e22bee78STejun Heo unsigned long expires; 2314e22bee78STejun Heo 231563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2316e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2317e22bee78STejun Heo 23183347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 231963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 23203347fc9fSLai Jiangshan break; 2321e22bee78STejun Heo } 23223347fc9fSLai Jiangshan 2323e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2324e22bee78STejun Heo } 2325e22bee78STejun Heo 2326a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2327e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2328e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2329e22bee78STejun Heo } 2330e22bee78STejun Heo 2331493a1724STejun Heo static void send_mayday(struct work_struct *work) 2332e22bee78STejun Heo { 2333112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2334112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2335493a1724STejun Heo 23362e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2337e22bee78STejun Heo 2338493008a8STejun Heo if (!wq->rescuer) 2339493a1724STejun Heo return; 2340e22bee78STejun Heo 2341e22bee78STejun Heo /* mayday mayday mayday */ 2342493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 234377668c8bSLai Jiangshan /* 234477668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 234577668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 234677668c8bSLai Jiangshan * rescuer is done with it. 234777668c8bSLai Jiangshan */ 234877668c8bSLai Jiangshan get_pwq(pwq); 2349493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2350e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2351725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2352493a1724STejun Heo } 2353e22bee78STejun Heo } 2354e22bee78STejun Heo 235532a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2356e22bee78STejun Heo { 235732a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2358e22bee78STejun Heo struct work_struct *work; 2359e22bee78STejun Heo 2360a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2361a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2362e22bee78STejun Heo 236363d95a91STejun Heo if (need_to_create_worker(pool)) { 2364e22bee78STejun Heo /* 2365e22bee78STejun Heo * We've been trying to create a new worker but 2366e22bee78STejun Heo * haven't been successful. We might be hitting an 2367e22bee78STejun Heo * allocation deadlock. Send distress signals to 2368e22bee78STejun Heo * rescuers. 2369e22bee78STejun Heo */ 237063d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2371e22bee78STejun Heo send_mayday(work); 2372e22bee78STejun Heo } 2373e22bee78STejun Heo 2374a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2375a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2376e22bee78STejun Heo 237763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2378e22bee78STejun Heo } 2379e22bee78STejun Heo 2380e22bee78STejun Heo /** 2381e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 238263d95a91STejun Heo * @pool: pool to create a new worker for 2383e22bee78STejun Heo * 238463d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2385e22bee78STejun Heo * have at least one idle worker on return from this function. If 2386e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 238763d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2388e22bee78STejun Heo * possible allocation deadlock. 2389e22bee78STejun Heo * 2390c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2391c5aa87bbSTejun Heo * may_start_working() %true. 2392e22bee78STejun Heo * 2393e22bee78STejun Heo * LOCKING: 2394a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2395e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2396e22bee78STejun Heo * manager. 2397e22bee78STejun Heo */ 239829187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2399d565ed63STejun Heo __releases(&pool->lock) 2400d565ed63STejun Heo __acquires(&pool->lock) 2401e22bee78STejun Heo { 2402e22bee78STejun Heo restart: 2403a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24049f9c2364STejun Heo 2405e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 240663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2407e22bee78STejun Heo 2408e22bee78STejun Heo while (true) { 2409051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2410e22bee78STejun Heo break; 2411e22bee78STejun Heo 2412e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 24139f9c2364STejun Heo 241463d95a91STejun Heo if (!need_to_create_worker(pool)) 2415e22bee78STejun Heo break; 2416e22bee78STejun Heo } 2417e22bee78STejun Heo 241863d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2419a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2420051e1850SLai Jiangshan /* 2421051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2422051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2423051e1850SLai Jiangshan * already become busy. 2424051e1850SLai Jiangshan */ 242563d95a91STejun Heo if (need_to_create_worker(pool)) 2426e22bee78STejun Heo goto restart; 2427e22bee78STejun Heo } 2428e22bee78STejun Heo 2429e22bee78STejun Heo /** 2430e22bee78STejun Heo * manage_workers - manage worker pool 2431e22bee78STejun Heo * @worker: self 2432e22bee78STejun Heo * 2433706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2434e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2435706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2436e22bee78STejun Heo * 2437e22bee78STejun Heo * The caller can safely start processing works on false return. On 2438e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2439e22bee78STejun Heo * and may_start_working() is true. 2440e22bee78STejun Heo * 2441e22bee78STejun Heo * CONTEXT: 2442a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2443e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2444e22bee78STejun Heo * 2445d185af30SYacine Belkadi * Return: 244629187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 244729187a9eSTejun Heo * start processing works, %true if management function was performed and 244829187a9eSTejun Heo * the conditions that the caller verified before calling the function may 244929187a9eSTejun Heo * no longer be true. 2450e22bee78STejun Heo */ 2451e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2452e22bee78STejun Heo { 245363d95a91STejun Heo struct worker_pool *pool = worker->pool; 2454e22bee78STejun Heo 2455692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 245629187a9eSTejun Heo return false; 2457692b4825STejun Heo 2458692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 24592607d7a6STejun Heo pool->manager = worker; 2460e22bee78STejun Heo 246129187a9eSTejun Heo maybe_create_worker(pool); 2462e22bee78STejun Heo 24632607d7a6STejun Heo pool->manager = NULL; 2464692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2465d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 246629187a9eSTejun Heo return true; 2467e22bee78STejun Heo } 2468e22bee78STejun Heo 2469a62428c0STejun Heo /** 2470a62428c0STejun Heo * process_one_work - process single work 2471c34056a3STejun Heo * @worker: self 2472a62428c0STejun Heo * @work: work to process 2473a62428c0STejun Heo * 2474a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2475a62428c0STejun Heo * process a single work including synchronization against and 2476a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2477a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2478a62428c0STejun Heo * call this function to process a work. 2479a62428c0STejun Heo * 2480a62428c0STejun Heo * CONTEXT: 2481a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2482a62428c0STejun Heo */ 2483c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2484d565ed63STejun Heo __releases(&pool->lock) 2485d565ed63STejun Heo __acquires(&pool->lock) 24861da177e4SLinus Torvalds { 2487112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2488bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2489c4560c2cSLai Jiangshan unsigned long work_data; 24907e11629dSTejun Heo struct worker *collision; 24914e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 24924e6045f1SJohannes Berg /* 2493a62428c0STejun Heo * It is permissible to free the struct work_struct from 2494a62428c0STejun Heo * inside the function that is called from it, this we need to 2495a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2496a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2497a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 24984e6045f1SJohannes Berg */ 24994d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 25004d82a1deSPeter Zijlstra 25014d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 25024e6045f1SJohannes Berg #endif 2503807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 250485327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2505ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 250625511a47STejun Heo 25077e11629dSTejun Heo /* 25087e11629dSTejun Heo * A single work shouldn't be executed concurrently by 25097e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 25107e11629dSTejun Heo * already processing the work. If so, defer the work to the 25117e11629dSTejun Heo * currently executing one. 25127e11629dSTejun Heo */ 2513c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 25147e11629dSTejun Heo if (unlikely(collision)) { 25157e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 25167e11629dSTejun Heo return; 25177e11629dSTejun Heo } 25181da177e4SLinus Torvalds 25198930cabaSTejun Heo /* claim and dequeue */ 25201da177e4SLinus Torvalds debug_work_deactivate(work); 2521c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2522c34056a3STejun Heo worker->current_work = work; 2523a2c1c57bSTejun Heo worker->current_func = work->func; 2524112202d9STejun Heo worker->current_pwq = pwq; 2525616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2526c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2527d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 25287a22ad75STejun Heo 25298bf89593STejun Heo /* 25308bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 25318bf89593STejun Heo * overridden through set_worker_desc(). 25328bf89593STejun Heo */ 25338bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 25348bf89593STejun Heo 2535a62428c0STejun Heo list_del_init(&work->entry); 2536a62428c0STejun Heo 2537649027d7STejun Heo /* 2538228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2539228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2540228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2541228f1d00SLai Jiangshan * execution of the pending work items. 2542fb0e7bebSTejun Heo */ 2543616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2544228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2545fb0e7bebSTejun Heo 2546974271c4STejun Heo /* 2547a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2548a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2549a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2550a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2551228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2552974271c4STejun Heo */ 2553a489a03eSLai Jiangshan if (need_more_worker(pool)) 255463d95a91STejun Heo wake_up_worker(pool); 2555974271c4STejun Heo 25568930cabaSTejun Heo /* 25577c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2558d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 255923657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 256023657bb1STejun Heo * disabled. 25618930cabaSTejun Heo */ 25627c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 25631da177e4SLinus Torvalds 2564a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2565365970a1SDavid Howells 2566a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 25673295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2568e6f3faa7SPeter Zijlstra /* 2569f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2570f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2571e6f3faa7SPeter Zijlstra * 2572e6f3faa7SPeter Zijlstra * However, that would result in: 2573e6f3faa7SPeter Zijlstra * 2574e6f3faa7SPeter Zijlstra * A(W1) 2575e6f3faa7SPeter Zijlstra * WFC(C) 2576e6f3faa7SPeter Zijlstra * A(W1) 2577e6f3faa7SPeter Zijlstra * C(C) 2578e6f3faa7SPeter Zijlstra * 2579e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2580e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2581e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2582f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2583e6f3faa7SPeter Zijlstra * these locks. 2584e6f3faa7SPeter Zijlstra * 2585e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2586e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2587e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2588e6f3faa7SPeter Zijlstra */ 2589f52be570SPeter Zijlstra lockdep_invariant_state(true); 2590725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2591e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2592a2c1c57bSTejun Heo worker->current_func(work); 2593e36c886aSArjan van de Ven /* 2594e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2595e36c886aSArjan van de Ven * point will only record its address. 2596e36c886aSArjan van de Ven */ 25971c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2598725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 25993295f0efSIngo Molnar lock_map_release(&lockdep_map); 2600112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 26011da177e4SLinus Torvalds 2602d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2603044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2604d75f773cSSakari Ailus " last function: %ps\n", 2605a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2606a2c1c57bSTejun Heo worker->current_func); 2607d5abe669SPeter Zijlstra debug_show_held_locks(current); 2608d5abe669SPeter Zijlstra dump_stack(); 2609d5abe669SPeter Zijlstra } 2610d5abe669SPeter Zijlstra 2611b22ce278STejun Heo /* 2612025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2613b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2614b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2615b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2616789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2617789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2618b22ce278STejun Heo */ 2619a7e6425eSPaul E. McKenney cond_resched(); 2620b22ce278STejun Heo 2621a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2622a62428c0STejun Heo 2623616db877STejun Heo /* 2624616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2625616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2626616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2627616db877STejun Heo */ 2628fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2629fb0e7bebSTejun Heo 26301b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 26311b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 26321b69ac6bSJohannes Weiner 2633a62428c0STejun Heo /* we're done with it, release */ 263442f8570fSSasha Levin hash_del(&worker->hentry); 2635c34056a3STejun Heo worker->current_work = NULL; 2636a2c1c57bSTejun Heo worker->current_func = NULL; 2637112202d9STejun Heo worker->current_pwq = NULL; 2638d812796eSLai Jiangshan worker->current_color = INT_MAX; 2639c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26401da177e4SLinus Torvalds } 26411da177e4SLinus Torvalds 2642affee4b2STejun Heo /** 2643affee4b2STejun Heo * process_scheduled_works - process scheduled works 2644affee4b2STejun Heo * @worker: self 2645affee4b2STejun Heo * 2646affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2647affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2648affee4b2STejun Heo * fetches a work from the top and executes it. 2649affee4b2STejun Heo * 2650affee4b2STejun Heo * CONTEXT: 2651a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2652affee4b2STejun Heo * multiple times. 2653affee4b2STejun Heo */ 2654affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26551da177e4SLinus Torvalds { 2656c0ab017dSTejun Heo struct work_struct *work; 2657c0ab017dSTejun Heo bool first = true; 2658c0ab017dSTejun Heo 2659c0ab017dSTejun Heo while ((work = list_first_entry_or_null(&worker->scheduled, 2660c0ab017dSTejun Heo struct work_struct, entry))) { 2661c0ab017dSTejun Heo if (first) { 2662c0ab017dSTejun Heo worker->pool->watchdog_ts = jiffies; 2663c0ab017dSTejun Heo first = false; 2664c0ab017dSTejun Heo } 2665c34056a3STejun Heo process_one_work(worker, work); 2666a62428c0STejun Heo } 26671da177e4SLinus Torvalds } 26681da177e4SLinus Torvalds 2669197f6accSTejun Heo static void set_pf_worker(bool val) 2670197f6accSTejun Heo { 2671197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2672197f6accSTejun Heo if (val) 2673197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2674197f6accSTejun Heo else 2675197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2676197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2677197f6accSTejun Heo } 2678197f6accSTejun Heo 26794690c4abSTejun Heo /** 26804690c4abSTejun Heo * worker_thread - the worker thread function 2681c34056a3STejun Heo * @__worker: self 26824690c4abSTejun Heo * 2683c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2684c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2685c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2686c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2687c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2688d185af30SYacine Belkadi * 2689d185af30SYacine Belkadi * Return: 0 26904690c4abSTejun Heo */ 2691c34056a3STejun Heo static int worker_thread(void *__worker) 26921da177e4SLinus Torvalds { 2693c34056a3STejun Heo struct worker *worker = __worker; 2694bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 26951da177e4SLinus Torvalds 2696e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2697197f6accSTejun Heo set_pf_worker(true); 2698c8e55f36STejun Heo woke_up: 2699a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2700affee4b2STejun Heo 2701a9ab775bSTejun Heo /* am I supposed to die? */ 2702a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2703a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2704197f6accSTejun Heo set_pf_worker(false); 270560f5a4bcSLai Jiangshan 270660f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2707e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2708a2d812a2STejun Heo worker_detach_from_pool(worker); 2709e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 271060f5a4bcSLai Jiangshan kfree(worker); 2711c8e55f36STejun Heo return 0; 2712c8e55f36STejun Heo } 2713c8e55f36STejun Heo 2714c8e55f36STejun Heo worker_leave_idle(worker); 2715db7bccf4STejun Heo recheck: 2716e22bee78STejun Heo /* no more worker necessary? */ 271763d95a91STejun Heo if (!need_more_worker(pool)) 2718e22bee78STejun Heo goto sleep; 2719e22bee78STejun Heo 2720e22bee78STejun Heo /* do we need to manage? */ 272163d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2722e22bee78STejun Heo goto recheck; 2723e22bee78STejun Heo 2724c8e55f36STejun Heo /* 2725c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2726c8e55f36STejun Heo * preparing to process a work or actually processing it. 2727c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2728c8e55f36STejun Heo */ 27296183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2730c8e55f36STejun Heo 2731e22bee78STejun Heo /* 2732a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2733a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2734a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2735a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2736a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2737e22bee78STejun Heo */ 2738a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2739e22bee78STejun Heo 2740e22bee78STejun Heo do { 2741affee4b2STejun Heo struct work_struct *work = 2742bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2743affee4b2STejun Heo struct work_struct, entry); 2744affee4b2STejun Heo 2745c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2746affee4b2STejun Heo process_scheduled_works(worker); 274763d95a91STejun Heo } while (keep_working(pool)); 2748affee4b2STejun Heo 2749228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2750d313dd85STejun Heo sleep: 2751c8e55f36STejun Heo /* 2752d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2753d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2754d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2755d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2756d565ed63STejun Heo * event. 2757c8e55f36STejun Heo */ 2758c8e55f36STejun Heo worker_enter_idle(worker); 2759c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2760a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27611da177e4SLinus Torvalds schedule(); 2762c8e55f36STejun Heo goto woke_up; 27631da177e4SLinus Torvalds } 27641da177e4SLinus Torvalds 2765e22bee78STejun Heo /** 2766e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2767111c225aSTejun Heo * @__rescuer: self 2768e22bee78STejun Heo * 2769e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2770493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2771e22bee78STejun Heo * 2772706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2773e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2774e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2775e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2776e22bee78STejun Heo * the problem rescuer solves. 2777e22bee78STejun Heo * 2778706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2779706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2780e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2781e22bee78STejun Heo * 2782e22bee78STejun Heo * This should happen rarely. 2783d185af30SYacine Belkadi * 2784d185af30SYacine Belkadi * Return: 0 2785e22bee78STejun Heo */ 2786111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2787e22bee78STejun Heo { 2788111c225aSTejun Heo struct worker *rescuer = __rescuer; 2789111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2790e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 27914d595b86SLai Jiangshan bool should_stop; 2792e22bee78STejun Heo 2793e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2794111c225aSTejun Heo 2795111c225aSTejun Heo /* 2796111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2797111c225aSTejun Heo * doesn't participate in concurrency management. 2798111c225aSTejun Heo */ 2799197f6accSTejun Heo set_pf_worker(true); 2800e22bee78STejun Heo repeat: 2801c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 28021da177e4SLinus Torvalds 28034d595b86SLai Jiangshan /* 28044d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 28054d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 28064d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 28074d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 28084d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 28094d595b86SLai Jiangshan * list is always empty on exit. 28104d595b86SLai Jiangshan */ 28114d595b86SLai Jiangshan should_stop = kthread_should_stop(); 28121da177e4SLinus Torvalds 2813493a1724STejun Heo /* see whether any pwq is asking for help */ 2814a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2815493a1724STejun Heo 2816493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2817493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2818493a1724STejun Heo struct pool_workqueue, mayday_node); 2819112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2820e22bee78STejun Heo struct work_struct *work, *n; 2821e22bee78STejun Heo 2822e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2823493a1724STejun Heo list_del_init(&pwq->mayday_node); 2824493a1724STejun Heo 2825a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2826e22bee78STejun Heo 282751697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 282851697d39SLai Jiangshan 2829a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2830e22bee78STejun Heo 2831e22bee78STejun Heo /* 2832e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2833e22bee78STejun Heo * process'em. 2834e22bee78STejun Heo */ 28350479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 283682607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 283782607adcSTejun Heo if (get_work_pwq(work) == pwq) { 2838e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2839725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 284082607adcSTejun Heo } 284182607adcSTejun Heo } 2842e22bee78STejun Heo 2843008847f6SNeilBrown if (!list_empty(scheduled)) { 2844e22bee78STejun Heo process_scheduled_works(rescuer); 28457576958aSTejun Heo 28467576958aSTejun Heo /* 2847008847f6SNeilBrown * The above execution of rescued work items could 2848008847f6SNeilBrown * have created more to rescue through 2849f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2850008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2851008847f6SNeilBrown * that such back-to-back work items, which may be 2852008847f6SNeilBrown * being used to relieve memory pressure, don't 2853008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2854008847f6SNeilBrown */ 28554f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2856a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2857e66b39afSTejun Heo /* 2858e66b39afSTejun Heo * Queue iff we aren't racing destruction 2859e66b39afSTejun Heo * and somebody else hasn't queued it already. 2860e66b39afSTejun Heo */ 2861e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2862008847f6SNeilBrown get_pwq(pwq); 2863e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2864e66b39afSTejun Heo } 2865a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2866008847f6SNeilBrown } 2867008847f6SNeilBrown } 2868008847f6SNeilBrown 2869008847f6SNeilBrown /* 287077668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 287113b1d625SLai Jiangshan * go away while we're still attached to it. 287277668c8bSLai Jiangshan */ 287377668c8bSLai Jiangshan put_pwq(pwq); 287477668c8bSLai Jiangshan 287577668c8bSLai Jiangshan /* 2876d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 28777576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 28787576958aSTejun Heo * and stalling the execution. 28797576958aSTejun Heo */ 2880d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 288163d95a91STejun Heo wake_up_worker(pool); 28827576958aSTejun Heo 2883a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 288413b1d625SLai Jiangshan 2885a2d812a2STejun Heo worker_detach_from_pool(rescuer); 288613b1d625SLai Jiangshan 2887a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 28881da177e4SLinus Torvalds } 28891da177e4SLinus Torvalds 2890a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2891493a1724STejun Heo 28924d595b86SLai Jiangshan if (should_stop) { 28934d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2894197f6accSTejun Heo set_pf_worker(false); 28954d595b86SLai Jiangshan return 0; 28964d595b86SLai Jiangshan } 28974d595b86SLai Jiangshan 2898111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2899111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2900e22bee78STejun Heo schedule(); 2901e22bee78STejun Heo goto repeat; 29021da177e4SLinus Torvalds } 29031da177e4SLinus Torvalds 2904fca839c0STejun Heo /** 2905fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2906fca839c0STejun Heo * @target_wq: workqueue being flushed 2907fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2908fca839c0STejun Heo * 2909fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2910fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2911fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2912fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2913fca839c0STejun Heo * a deadlock. 2914fca839c0STejun Heo */ 2915fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2916fca839c0STejun Heo struct work_struct *target_work) 2917fca839c0STejun Heo { 2918fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2919fca839c0STejun Heo struct worker *worker; 2920fca839c0STejun Heo 2921fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2922fca839c0STejun Heo return; 2923fca839c0STejun Heo 2924fca839c0STejun Heo worker = current_wq_worker(); 2925fca839c0STejun Heo 2926fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2927d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2928fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 292923d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 293023d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2931d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2932fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2933fca839c0STejun Heo target_wq->name, target_func); 2934fca839c0STejun Heo } 2935fca839c0STejun Heo 2936fc2e4d70SOleg Nesterov struct wq_barrier { 2937fc2e4d70SOleg Nesterov struct work_struct work; 2938fc2e4d70SOleg Nesterov struct completion done; 29392607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2940fc2e4d70SOleg Nesterov }; 2941fc2e4d70SOleg Nesterov 2942fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2943fc2e4d70SOleg Nesterov { 2944fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2945fc2e4d70SOleg Nesterov complete(&barr->done); 2946fc2e4d70SOleg Nesterov } 2947fc2e4d70SOleg Nesterov 29484690c4abSTejun Heo /** 29494690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2950112202d9STejun Heo * @pwq: pwq to insert barrier into 29514690c4abSTejun Heo * @barr: wq_barrier to insert 2952affee4b2STejun Heo * @target: target work to attach @barr to 2953affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29544690c4abSTejun Heo * 2955affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2956affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2957affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2958affee4b2STejun Heo * cpu. 2959affee4b2STejun Heo * 2960affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2961affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2962affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2963affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2964affee4b2STejun Heo * after a work with LINKED flag set. 2965affee4b2STejun Heo * 2966affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2967112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 29684690c4abSTejun Heo * 29694690c4abSTejun Heo * CONTEXT: 2970a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 29714690c4abSTejun Heo */ 2972112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2973affee4b2STejun Heo struct wq_barrier *barr, 2974affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2975fc2e4d70SOleg Nesterov { 2976d812796eSLai Jiangshan unsigned int work_flags = 0; 2977d812796eSLai Jiangshan unsigned int work_color; 2978affee4b2STejun Heo struct list_head *head; 2979affee4b2STejun Heo 2980dc186ad7SThomas Gleixner /* 2981d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2982dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2983dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2984dc186ad7SThomas Gleixner * might deadlock. 2985dc186ad7SThomas Gleixner */ 2986ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 298722df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 298852fa5bc5SBoqun Feng 2989fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2990fd1a5b04SByungchul Park 29912607d7a6STejun Heo barr->task = current; 299283c22520SOleg Nesterov 2993018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2994018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2995018f3a13SLai Jiangshan 2996affee4b2STejun Heo /* 2997affee4b2STejun Heo * If @target is currently being executed, schedule the 2998affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2999affee4b2STejun Heo */ 3000d812796eSLai Jiangshan if (worker) { 3001affee4b2STejun Heo head = worker->scheduled.next; 3002d812796eSLai Jiangshan work_color = worker->current_color; 3003d812796eSLai Jiangshan } else { 3004affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 3005affee4b2STejun Heo 3006affee4b2STejun Heo head = target->entry.next; 3007affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 3008d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 3009d812796eSLai Jiangshan work_color = get_work_color(*bits); 3010affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3011affee4b2STejun Heo } 3012affee4b2STejun Heo 3013d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 3014d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 3015d812796eSLai Jiangshan 3016d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 3017fc2e4d70SOleg Nesterov } 3018fc2e4d70SOleg Nesterov 301973f53c4aSTejun Heo /** 3020112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 302173f53c4aSTejun Heo * @wq: workqueue being flushed 302273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 302373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 302473f53c4aSTejun Heo * 3025112202d9STejun Heo * Prepare pwqs for workqueue flushing. 302673f53c4aSTejun Heo * 3027112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 3028112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 3029112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3030112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 3031112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 303273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 303373f53c4aSTejun Heo * 303473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 303573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 303673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 303773f53c4aSTejun Heo * is returned. 303873f53c4aSTejun Heo * 3039112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 304073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 304173f53c4aSTejun Heo * advanced to @work_color. 304273f53c4aSTejun Heo * 304373f53c4aSTejun Heo * CONTEXT: 30443c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 304573f53c4aSTejun Heo * 3046d185af30SYacine Belkadi * Return: 304773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 304873f53c4aSTejun Heo * otherwise. 304973f53c4aSTejun Heo */ 3050112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 305173f53c4aSTejun Heo int flush_color, int work_color) 30521da177e4SLinus Torvalds { 305373f53c4aSTejun Heo bool wait = false; 305449e3cf44STejun Heo struct pool_workqueue *pwq; 30551da177e4SLinus Torvalds 305673f53c4aSTejun Heo if (flush_color >= 0) { 30576183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3058112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3059dc186ad7SThomas Gleixner } 306014441960SOleg Nesterov 306149e3cf44STejun Heo for_each_pwq(pwq, wq) { 3062112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30631da177e4SLinus Torvalds 3064a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 306573f53c4aSTejun Heo 306673f53c4aSTejun Heo if (flush_color >= 0) { 30676183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 306873f53c4aSTejun Heo 3069112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3070112202d9STejun Heo pwq->flush_color = flush_color; 3071112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 307273f53c4aSTejun Heo wait = true; 30731da177e4SLinus Torvalds } 307473f53c4aSTejun Heo } 307573f53c4aSTejun Heo 307673f53c4aSTejun Heo if (work_color >= 0) { 30776183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3078112202d9STejun Heo pwq->work_color = work_color; 307973f53c4aSTejun Heo } 308073f53c4aSTejun Heo 3081a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 30821da177e4SLinus Torvalds } 30831da177e4SLinus Torvalds 3084112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 308573f53c4aSTejun Heo complete(&wq->first_flusher->done); 308673f53c4aSTejun Heo 308773f53c4aSTejun Heo return wait; 308883c22520SOleg Nesterov } 30891da177e4SLinus Torvalds 30900fcb78c2SRolf Eike Beer /** 3091c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 30920fcb78c2SRolf Eike Beer * @wq: workqueue to flush 30931da177e4SLinus Torvalds * 3094c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3095c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 30961da177e4SLinus Torvalds */ 3097c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 30981da177e4SLinus Torvalds { 309973f53c4aSTejun Heo struct wq_flusher this_flusher = { 310073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 310173f53c4aSTejun Heo .flush_color = -1, 3102fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 310373f53c4aSTejun Heo }; 310473f53c4aSTejun Heo int next_color; 3105b1f4ec17SOleg Nesterov 31063347fa09STejun Heo if (WARN_ON(!wq_online)) 31073347fa09STejun Heo return; 31083347fa09STejun Heo 310987915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 311087915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 311187915adcSJohannes Berg 31123c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 311373f53c4aSTejun Heo 311473f53c4aSTejun Heo /* 311573f53c4aSTejun Heo * Start-to-wait phase 311673f53c4aSTejun Heo */ 311773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 311873f53c4aSTejun Heo 311973f53c4aSTejun Heo if (next_color != wq->flush_color) { 312073f53c4aSTejun Heo /* 312173f53c4aSTejun Heo * Color space is not full. The current work_color 312273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 312373f53c4aSTejun Heo * by one. 312473f53c4aSTejun Heo */ 31256183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 312673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 312773f53c4aSTejun Heo wq->work_color = next_color; 312873f53c4aSTejun Heo 312973f53c4aSTejun Heo if (!wq->first_flusher) { 313073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 31316183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 313273f53c4aSTejun Heo 313373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 313473f53c4aSTejun Heo 3135112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 313673f53c4aSTejun Heo wq->work_color)) { 313773f53c4aSTejun Heo /* nothing to flush, done */ 313873f53c4aSTejun Heo wq->flush_color = next_color; 313973f53c4aSTejun Heo wq->first_flusher = NULL; 314073f53c4aSTejun Heo goto out_unlock; 314173f53c4aSTejun Heo } 314273f53c4aSTejun Heo } else { 314373f53c4aSTejun Heo /* wait in queue */ 31446183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 314573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3146112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 314773f53c4aSTejun Heo } 314873f53c4aSTejun Heo } else { 314973f53c4aSTejun Heo /* 315073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 315173f53c4aSTejun Heo * The next flush completion will assign us 315273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 315373f53c4aSTejun Heo */ 315473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 315573f53c4aSTejun Heo } 315673f53c4aSTejun Heo 3157fca839c0STejun Heo check_flush_dependency(wq, NULL); 3158fca839c0STejun Heo 31593c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 316073f53c4aSTejun Heo 316173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 316273f53c4aSTejun Heo 316373f53c4aSTejun Heo /* 316473f53c4aSTejun Heo * Wake-up-and-cascade phase 316573f53c4aSTejun Heo * 316673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 316773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 316873f53c4aSTejun Heo */ 316900d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 317073f53c4aSTejun Heo return; 317173f53c4aSTejun Heo 31723c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 317373f53c4aSTejun Heo 31744ce48b37STejun Heo /* we might have raced, check again with mutex held */ 31754ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 31764ce48b37STejun Heo goto out_unlock; 31774ce48b37STejun Heo 317800d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 317973f53c4aSTejun Heo 31806183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 31816183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 318273f53c4aSTejun Heo 318373f53c4aSTejun Heo while (true) { 318473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 318573f53c4aSTejun Heo 318673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 318773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 318873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 318973f53c4aSTejun Heo break; 319073f53c4aSTejun Heo list_del_init(&next->list); 319173f53c4aSTejun Heo complete(&next->done); 319273f53c4aSTejun Heo } 319373f53c4aSTejun Heo 31946183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 319573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 319673f53c4aSTejun Heo 319773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 319873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 319973f53c4aSTejun Heo 320073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 320173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 320273f53c4aSTejun Heo /* 320373f53c4aSTejun Heo * Assign the same color to all overflowed 320473f53c4aSTejun Heo * flushers, advance work_color and append to 320573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 320673f53c4aSTejun Heo * phase for these overflowed flushers. 320773f53c4aSTejun Heo */ 320873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 320973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 321073f53c4aSTejun Heo 321173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 321273f53c4aSTejun Heo 321373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 321473f53c4aSTejun Heo &wq->flusher_queue); 3215112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 321673f53c4aSTejun Heo } 321773f53c4aSTejun Heo 321873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 32196183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 322073f53c4aSTejun Heo break; 322173f53c4aSTejun Heo } 322273f53c4aSTejun Heo 322373f53c4aSTejun Heo /* 322473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3225112202d9STejun Heo * the new first flusher and arm pwqs. 322673f53c4aSTejun Heo */ 32276183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 32286183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 322973f53c4aSTejun Heo 323073f53c4aSTejun Heo list_del_init(&next->list); 323173f53c4aSTejun Heo wq->first_flusher = next; 323273f53c4aSTejun Heo 3233112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 323473f53c4aSTejun Heo break; 323573f53c4aSTejun Heo 323673f53c4aSTejun Heo /* 323773f53c4aSTejun Heo * Meh... this color is already done, clear first 323873f53c4aSTejun Heo * flusher and repeat cascading. 323973f53c4aSTejun Heo */ 324073f53c4aSTejun Heo wq->first_flusher = NULL; 324173f53c4aSTejun Heo } 324273f53c4aSTejun Heo 324373f53c4aSTejun Heo out_unlock: 32443c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32451da177e4SLinus Torvalds } 3246c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32471da177e4SLinus Torvalds 32489c5a2ba7STejun Heo /** 32499c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32509c5a2ba7STejun Heo * @wq: workqueue to drain 32519c5a2ba7STejun Heo * 32529c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32539c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32549c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3255b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32569c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32579c5a2ba7STejun Heo * takes too long. 32589c5a2ba7STejun Heo */ 32599c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32609c5a2ba7STejun Heo { 32619c5a2ba7STejun Heo unsigned int flush_cnt = 0; 326249e3cf44STejun Heo struct pool_workqueue *pwq; 32639c5a2ba7STejun Heo 32649c5a2ba7STejun Heo /* 32659c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 32669c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3267618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 32689c5a2ba7STejun Heo */ 326987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 32709c5a2ba7STejun Heo if (!wq->nr_drainers++) 3271618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 327287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32739c5a2ba7STejun Heo reflush: 3274c4f135d6STetsuo Handa __flush_workqueue(wq); 32759c5a2ba7STejun Heo 3276b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 327776af4d93STejun Heo 327849e3cf44STejun Heo for_each_pwq(pwq, wq) { 3279fa2563e4SThomas Tuttle bool drained; 32809c5a2ba7STejun Heo 3281a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3282f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3283a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3284fa2563e4SThomas Tuttle 3285fa2563e4SThomas Tuttle if (drained) 32869c5a2ba7STejun Heo continue; 32879c5a2ba7STejun Heo 32889c5a2ba7STejun Heo if (++flush_cnt == 10 || 32899c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3290e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3291e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 329276af4d93STejun Heo 3293b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 32949c5a2ba7STejun Heo goto reflush; 32959c5a2ba7STejun Heo } 32969c5a2ba7STejun Heo 32979c5a2ba7STejun Heo if (!--wq->nr_drainers) 3298618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 329987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 33009c5a2ba7STejun Heo } 33019c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 33029c5a2ba7STejun Heo 3303d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3304d6e89786SJohannes Berg bool from_cancel) 3305baf59022STejun Heo { 3306baf59022STejun Heo struct worker *worker = NULL; 3307c9e7cf27STejun Heo struct worker_pool *pool; 3308112202d9STejun Heo struct pool_workqueue *pwq; 3309baf59022STejun Heo 3310baf59022STejun Heo might_sleep(); 3311baf59022STejun Heo 331224acfb71SThomas Gleixner rcu_read_lock(); 3313fa1b54e6STejun Heo pool = get_work_pool(work); 3314fa1b54e6STejun Heo if (!pool) { 331524acfb71SThomas Gleixner rcu_read_unlock(); 3316fa1b54e6STejun Heo return false; 3317fa1b54e6STejun Heo } 3318fa1b54e6STejun Heo 3319a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 33200b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3321112202d9STejun Heo pwq = get_work_pwq(work); 3322112202d9STejun Heo if (pwq) { 3323112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3324baf59022STejun Heo goto already_gone; 3325606a5020STejun Heo } else { 3326c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3327baf59022STejun Heo if (!worker) 3328baf59022STejun Heo goto already_gone; 3329112202d9STejun Heo pwq = worker->current_pwq; 3330606a5020STejun Heo } 3331baf59022STejun Heo 3332fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3333fca839c0STejun Heo 3334112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3335a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3336baf59022STejun Heo 3337e159489bSTejun Heo /* 3338a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3339a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3340a1d14934SPeter Zijlstra * 3341a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3342a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3343a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3344a1d14934SPeter Zijlstra * forward progress. 3345e159489bSTejun Heo */ 3346d6e89786SJohannes Berg if (!from_cancel && 3347d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3348112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3349112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3350a1d14934SPeter Zijlstra } 335124acfb71SThomas Gleixner rcu_read_unlock(); 3352baf59022STejun Heo return true; 3353baf59022STejun Heo already_gone: 3354a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 335524acfb71SThomas Gleixner rcu_read_unlock(); 3356baf59022STejun Heo return false; 3357baf59022STejun Heo } 3358baf59022STejun Heo 3359d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3360d6e89786SJohannes Berg { 3361d6e89786SJohannes Berg struct wq_barrier barr; 3362d6e89786SJohannes Berg 3363d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3364d6e89786SJohannes Berg return false; 3365d6e89786SJohannes Berg 33664d43d395STetsuo Handa if (WARN_ON(!work->func)) 33674d43d395STetsuo Handa return false; 33684d43d395STetsuo Handa 336987915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 337087915adcSJohannes Berg lock_map_release(&work->lockdep_map); 337187915adcSJohannes Berg 3372d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3373d6e89786SJohannes Berg wait_for_completion(&barr.done); 3374d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3375d6e89786SJohannes Berg return true; 3376d6e89786SJohannes Berg } else { 3377d6e89786SJohannes Berg return false; 3378d6e89786SJohannes Berg } 3379d6e89786SJohannes Berg } 3380d6e89786SJohannes Berg 3381db700897SOleg Nesterov /** 3382401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3383401a8d04STejun Heo * @work: the work to flush 3384db700897SOleg Nesterov * 3385606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3386606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3387401a8d04STejun Heo * 3388d185af30SYacine Belkadi * Return: 3389401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3390401a8d04STejun Heo * %false if it was already idle. 3391db700897SOleg Nesterov */ 3392401a8d04STejun Heo bool flush_work(struct work_struct *work) 3393db700897SOleg Nesterov { 3394d6e89786SJohannes Berg return __flush_work(work, false); 3395606a5020STejun Heo } 3396db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3397db700897SOleg Nesterov 33988603e1b3STejun Heo struct cwt_wait { 3399ac6424b9SIngo Molnar wait_queue_entry_t wait; 34008603e1b3STejun Heo struct work_struct *work; 34018603e1b3STejun Heo }; 34028603e1b3STejun Heo 3403ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 34048603e1b3STejun Heo { 34058603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 34068603e1b3STejun Heo 34078603e1b3STejun Heo if (cwait->work != key) 34088603e1b3STejun Heo return 0; 34098603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 34108603e1b3STejun Heo } 34118603e1b3STejun Heo 341236e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3413401a8d04STejun Heo { 34148603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3415bbb68dfaSTejun Heo unsigned long flags; 34161f1f642eSOleg Nesterov int ret; 34171f1f642eSOleg Nesterov 34181f1f642eSOleg Nesterov do { 3419bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3420bbb68dfaSTejun Heo /* 34218603e1b3STejun Heo * If someone else is already canceling, wait for it to 34228603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 34238603e1b3STejun Heo * because we may get scheduled between @work's completion 34248603e1b3STejun Heo * and the other canceling task resuming and clearing 34258603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 34268603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 34278603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 34288603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 34298603e1b3STejun Heo * we're hogging the CPU. 34308603e1b3STejun Heo * 34318603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 34328603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 34338603e1b3STejun Heo * wake function which matches @work along with exclusive 34348603e1b3STejun Heo * wait and wakeup. 3435bbb68dfaSTejun Heo */ 34368603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 34378603e1b3STejun Heo struct cwt_wait cwait; 34388603e1b3STejun Heo 34398603e1b3STejun Heo init_wait(&cwait.wait); 34408603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34418603e1b3STejun Heo cwait.work = work; 34428603e1b3STejun Heo 34438603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34448603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34458603e1b3STejun Heo if (work_is_canceling(work)) 34468603e1b3STejun Heo schedule(); 34478603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34488603e1b3STejun Heo } 34491f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34501f1f642eSOleg Nesterov 3451bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3452bbb68dfaSTejun Heo mark_work_canceling(work); 3453bbb68dfaSTejun Heo local_irq_restore(flags); 3454bbb68dfaSTejun Heo 34553347fa09STejun Heo /* 34563347fa09STejun Heo * This allows canceling during early boot. We know that @work 34573347fa09STejun Heo * isn't executing. 34583347fa09STejun Heo */ 34593347fa09STejun Heo if (wq_online) 3460d6e89786SJohannes Berg __flush_work(work, true); 34613347fa09STejun Heo 34627a22ad75STejun Heo clear_work_data(work); 34638603e1b3STejun Heo 34648603e1b3STejun Heo /* 34658603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 34668603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 34678603e1b3STejun Heo * visible there. 34688603e1b3STejun Heo */ 34698603e1b3STejun Heo smp_mb(); 34708603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 34718603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 34728603e1b3STejun Heo 34731f1f642eSOleg Nesterov return ret; 34741f1f642eSOleg Nesterov } 34751f1f642eSOleg Nesterov 34766e84d644SOleg Nesterov /** 3477401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3478401a8d04STejun Heo * @work: the work to cancel 34796e84d644SOleg Nesterov * 3480401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3481401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3482401a8d04STejun Heo * another workqueue. On return from this function, @work is 3483401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 34841f1f642eSOleg Nesterov * 3485401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3486401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 34876e84d644SOleg Nesterov * 3488401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 34896e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3490401a8d04STejun Heo * 3491d185af30SYacine Belkadi * Return: 3492401a8d04STejun Heo * %true if @work was pending, %false otherwise. 34936e84d644SOleg Nesterov */ 3494401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 34956e84d644SOleg Nesterov { 349636e227d2STejun Heo return __cancel_work_timer(work, false); 3497b89deed3SOleg Nesterov } 349828e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3499b89deed3SOleg Nesterov 35006e84d644SOleg Nesterov /** 3501401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3502401a8d04STejun Heo * @dwork: the delayed work to flush 35036e84d644SOleg Nesterov * 3504401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3505401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3506401a8d04STejun Heo * considers the last queueing instance of @dwork. 35071f1f642eSOleg Nesterov * 3508d185af30SYacine Belkadi * Return: 3509401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3510401a8d04STejun Heo * %false if it was already idle. 35116e84d644SOleg Nesterov */ 3512401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3513401a8d04STejun Heo { 35148930cabaSTejun Heo local_irq_disable(); 3515401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 351660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 35178930cabaSTejun Heo local_irq_enable(); 3518401a8d04STejun Heo return flush_work(&dwork->work); 3519401a8d04STejun Heo } 3520401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3521401a8d04STejun Heo 352205f0fe6bSTejun Heo /** 352305f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 352405f0fe6bSTejun Heo * @rwork: the rcu work to flush 352505f0fe6bSTejun Heo * 352605f0fe6bSTejun Heo * Return: 352705f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 352805f0fe6bSTejun Heo * %false if it was already idle. 352905f0fe6bSTejun Heo */ 353005f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 353105f0fe6bSTejun Heo { 353205f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 353305f0fe6bSTejun Heo rcu_barrier(); 353405f0fe6bSTejun Heo flush_work(&rwork->work); 353505f0fe6bSTejun Heo return true; 353605f0fe6bSTejun Heo } else { 353705f0fe6bSTejun Heo return flush_work(&rwork->work); 353805f0fe6bSTejun Heo } 353905f0fe6bSTejun Heo } 354005f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 354105f0fe6bSTejun Heo 3542f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3543f72b8792SJens Axboe { 3544f72b8792SJens Axboe unsigned long flags; 3545f72b8792SJens Axboe int ret; 3546f72b8792SJens Axboe 3547f72b8792SJens Axboe do { 3548f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3549f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3550f72b8792SJens Axboe 3551f72b8792SJens Axboe if (unlikely(ret < 0)) 3552f72b8792SJens Axboe return false; 3553f72b8792SJens Axboe 3554f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3555f72b8792SJens Axboe local_irq_restore(flags); 3556f72b8792SJens Axboe return ret; 3557f72b8792SJens Axboe } 3558f72b8792SJens Axboe 355973b4b532SAndrey Grodzovsky /* 356073b4b532SAndrey Grodzovsky * See cancel_delayed_work() 356173b4b532SAndrey Grodzovsky */ 356273b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 356373b4b532SAndrey Grodzovsky { 356473b4b532SAndrey Grodzovsky return __cancel_work(work, false); 356573b4b532SAndrey Grodzovsky } 356673b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 356773b4b532SAndrey Grodzovsky 3568401a8d04STejun Heo /** 356957b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 357057b30ae7STejun Heo * @dwork: delayed_work to cancel 357109383498STejun Heo * 3572d185af30SYacine Belkadi * Kill off a pending delayed_work. 3573d185af30SYacine Belkadi * 3574d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3575d185af30SYacine Belkadi * pending. 3576d185af30SYacine Belkadi * 3577d185af30SYacine Belkadi * Note: 3578d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3579d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3580d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 358109383498STejun Heo * 358257b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 358309383498STejun Heo */ 358457b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 358509383498STejun Heo { 3586f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 358709383498STejun Heo } 358857b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 358909383498STejun Heo 359009383498STejun Heo /** 3591401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3592401a8d04STejun Heo * @dwork: the delayed work cancel 3593401a8d04STejun Heo * 3594401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3595401a8d04STejun Heo * 3596d185af30SYacine Belkadi * Return: 3597401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3598401a8d04STejun Heo */ 3599401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 36006e84d644SOleg Nesterov { 360136e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 36026e84d644SOleg Nesterov } 3603f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 36041da177e4SLinus Torvalds 36050fcb78c2SRolf Eike Beer /** 360631ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3607b6136773SAndrew Morton * @func: the function to call 3608b6136773SAndrew Morton * 360931ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 361031ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3611b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 361231ddd871STejun Heo * 3613d185af30SYacine Belkadi * Return: 361431ddd871STejun Heo * 0 on success, -errno on failure. 3615b6136773SAndrew Morton */ 361665f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 361715316ba8SChristoph Lameter { 361815316ba8SChristoph Lameter int cpu; 361938f51568SNamhyung Kim struct work_struct __percpu *works; 362015316ba8SChristoph Lameter 3621b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3622b6136773SAndrew Morton if (!works) 362315316ba8SChristoph Lameter return -ENOMEM; 3624b6136773SAndrew Morton 3625ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 362693981800STejun Heo 362715316ba8SChristoph Lameter for_each_online_cpu(cpu) { 36289bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 36299bfb1839SIngo Molnar 36309bfb1839SIngo Molnar INIT_WORK(work, func); 36318de6d308SOleg Nesterov schedule_work_on(cpu, work); 363215316ba8SChristoph Lameter } 363393981800STejun Heo 363493981800STejun Heo for_each_online_cpu(cpu) 36358616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 363693981800STejun Heo 3637ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3638b6136773SAndrew Morton free_percpu(works); 363915316ba8SChristoph Lameter return 0; 364015316ba8SChristoph Lameter } 364115316ba8SChristoph Lameter 3642eef6a7d5SAlan Stern /** 36431fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36441fa44ecaSJames Bottomley * @fn: the function to execute 36451fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36461fa44ecaSJames Bottomley * be available when the work executes) 36471fa44ecaSJames Bottomley * 36481fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36491fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36501fa44ecaSJames Bottomley * 3651d185af30SYacine Belkadi * Return: 0 - function was executed 36521fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36531fa44ecaSJames Bottomley */ 365465f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36551fa44ecaSJames Bottomley { 36561fa44ecaSJames Bottomley if (!in_interrupt()) { 365765f27f38SDavid Howells fn(&ew->work); 36581fa44ecaSJames Bottomley return 0; 36591fa44ecaSJames Bottomley } 36601fa44ecaSJames Bottomley 366165f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36621fa44ecaSJames Bottomley schedule_work(&ew->work); 36631fa44ecaSJames Bottomley 36641fa44ecaSJames Bottomley return 1; 36651fa44ecaSJames Bottomley } 36661fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 36671fa44ecaSJames Bottomley 36687a4e344cSTejun Heo /** 36697a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 36707a4e344cSTejun Heo * @attrs: workqueue_attrs to free 36717a4e344cSTejun Heo * 36727a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 36737a4e344cSTejun Heo */ 3674513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 36757a4e344cSTejun Heo { 36767a4e344cSTejun Heo if (attrs) { 36777a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 36787a4e344cSTejun Heo kfree(attrs); 36797a4e344cSTejun Heo } 36807a4e344cSTejun Heo } 36817a4e344cSTejun Heo 36827a4e344cSTejun Heo /** 36837a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 36847a4e344cSTejun Heo * 36857a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3686d185af30SYacine Belkadi * return it. 3687d185af30SYacine Belkadi * 3688d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 36897a4e344cSTejun Heo */ 3690513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 36917a4e344cSTejun Heo { 36927a4e344cSTejun Heo struct workqueue_attrs *attrs; 36937a4e344cSTejun Heo 3694be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 36957a4e344cSTejun Heo if (!attrs) 36967a4e344cSTejun Heo goto fail; 3697be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 36987a4e344cSTejun Heo goto fail; 36997a4e344cSTejun Heo 370013e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 37017a4e344cSTejun Heo return attrs; 37027a4e344cSTejun Heo fail: 37037a4e344cSTejun Heo free_workqueue_attrs(attrs); 37047a4e344cSTejun Heo return NULL; 37057a4e344cSTejun Heo } 37067a4e344cSTejun Heo 370729c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 370829c91e99STejun Heo const struct workqueue_attrs *from) 370929c91e99STejun Heo { 371029c91e99STejun Heo to->nice = from->nice; 371129c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 37122865a8fbSShaohua Li /* 37132865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 37142865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 37152865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 37162865a8fbSShaohua Li */ 37172865a8fbSShaohua Li to->no_numa = from->no_numa; 371829c91e99STejun Heo } 371929c91e99STejun Heo 372029c91e99STejun Heo /* hash value of the content of @attr */ 372129c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 372229c91e99STejun Heo { 372329c91e99STejun Heo u32 hash = 0; 372429c91e99STejun Heo 372529c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 372613e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 372713e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 372829c91e99STejun Heo return hash; 372929c91e99STejun Heo } 373029c91e99STejun Heo 373129c91e99STejun Heo /* content equality test */ 373229c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 373329c91e99STejun Heo const struct workqueue_attrs *b) 373429c91e99STejun Heo { 373529c91e99STejun Heo if (a->nice != b->nice) 373629c91e99STejun Heo return false; 373729c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 373829c91e99STejun Heo return false; 373929c91e99STejun Heo return true; 374029c91e99STejun Heo } 374129c91e99STejun Heo 37427a4e344cSTejun Heo /** 37437a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 37447a4e344cSTejun Heo * @pool: worker_pool to initialize 37457a4e344cSTejun Heo * 3746402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3747d185af30SYacine Belkadi * 3748d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 374929c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 375029c91e99STejun Heo * on @pool safely to release it. 37517a4e344cSTejun Heo */ 37527a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 37534e1a1f9aSTejun Heo { 3754a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 375529c91e99STejun Heo pool->id = -1; 375629c91e99STejun Heo pool->cpu = -1; 3757f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 37584e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 375982607adcSTejun Heo pool->watchdog_ts = jiffies; 37604e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 37614e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 37624e1a1f9aSTejun Heo hash_init(pool->busy_hash); 37634e1a1f9aSTejun Heo 376432a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 37653f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 37664e1a1f9aSTejun Heo 376732a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 37684e1a1f9aSTejun Heo 3769da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3770e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 37717a4e344cSTejun Heo 37727cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 377329c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 377429c91e99STejun Heo pool->refcnt = 1; 377529c91e99STejun Heo 377629c91e99STejun Heo /* shouldn't fail above this point */ 3777be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 37787a4e344cSTejun Heo if (!pool->attrs) 37797a4e344cSTejun Heo return -ENOMEM; 37807a4e344cSTejun Heo return 0; 37814e1a1f9aSTejun Heo } 37824e1a1f9aSTejun Heo 3783669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3784669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3785669de8bdSBart Van Assche { 3786669de8bdSBart Van Assche char *lock_name; 3787669de8bdSBart Van Assche 3788669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3789669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3790669de8bdSBart Van Assche if (!lock_name) 3791669de8bdSBart Van Assche lock_name = wq->name; 379269a106c0SQian Cai 379369a106c0SQian Cai wq->lock_name = lock_name; 3794669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3795669de8bdSBart Van Assche } 3796669de8bdSBart Van Assche 3797669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3798669de8bdSBart Van Assche { 3799669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3800669de8bdSBart Van Assche } 3801669de8bdSBart Van Assche 3802669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3803669de8bdSBart Van Assche { 3804669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3805669de8bdSBart Van Assche kfree(wq->lock_name); 3806669de8bdSBart Van Assche } 3807669de8bdSBart Van Assche #else 3808669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3809669de8bdSBart Van Assche { 3810669de8bdSBart Van Assche } 3811669de8bdSBart Van Assche 3812669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3813669de8bdSBart Van Assche { 3814669de8bdSBart Van Assche } 3815669de8bdSBart Van Assche 3816669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3817669de8bdSBart Van Assche { 3818669de8bdSBart Van Assche } 3819669de8bdSBart Van Assche #endif 3820669de8bdSBart Van Assche 3821e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3822e2dca7adSTejun Heo { 3823e2dca7adSTejun Heo struct workqueue_struct *wq = 3824e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3825e2dca7adSTejun Heo 3826669de8bdSBart Van Assche wq_free_lockdep(wq); 3827669de8bdSBart Van Assche 3828e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3829*ee1ceef7STejun Heo free_percpu(wq->cpu_pwq); 3830e2dca7adSTejun Heo else 3831e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3832e2dca7adSTejun Heo 3833e2dca7adSTejun Heo kfree(wq); 3834e2dca7adSTejun Heo } 3835e2dca7adSTejun Heo 383629c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 383729c91e99STejun Heo { 383829c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 383929c91e99STejun Heo 38407cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 384129c91e99STejun Heo free_workqueue_attrs(pool->attrs); 384229c91e99STejun Heo kfree(pool); 384329c91e99STejun Heo } 384429c91e99STejun Heo 384529c91e99STejun Heo /** 384629c91e99STejun Heo * put_unbound_pool - put a worker_pool 384729c91e99STejun Heo * @pool: worker_pool to put 384829c91e99STejun Heo * 384924acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3850c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3851c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3852c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3853a892caccSTejun Heo * 3854a892caccSTejun Heo * Should be called with wq_pool_mutex held. 385529c91e99STejun Heo */ 385629c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 385729c91e99STejun Heo { 385860f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 385929c91e99STejun Heo struct worker *worker; 38609680540cSYang Yingliang LIST_HEAD(cull_list); 3861e02b9312SValentin Schneider 3862a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3863a892caccSTejun Heo 3864a892caccSTejun Heo if (--pool->refcnt) 386529c91e99STejun Heo return; 386629c91e99STejun Heo 386729c91e99STejun Heo /* sanity checks */ 386861d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3869a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 387029c91e99STejun Heo return; 387129c91e99STejun Heo 387229c91e99STejun Heo /* release id and unhash */ 387329c91e99STejun Heo if (pool->id >= 0) 387429c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 387529c91e99STejun Heo hash_del(&pool->hash_node); 387629c91e99STejun Heo 3877c5aa87bbSTejun Heo /* 3878692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3879692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3880692b4825STejun Heo * manager and @pool gets freed with the flag set. 38819ab03be4SValentin Schneider * 38829ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 38839ab03be4SValentin Schneider * only get here with 38849ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 38859ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 38869ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 38879ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 38889ab03be4SValentin Schneider * drops pool->lock 3889c5aa87bbSTejun Heo */ 38909ab03be4SValentin Schneider while (true) { 38919ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 38929ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3893d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3894e02b9312SValentin Schneider 3895e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 38969ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 38979ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3898692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 38999ab03be4SValentin Schneider break; 39009ab03be4SValentin Schneider } 39019ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3902e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 39039ab03be4SValentin Schneider } 3904692b4825STejun Heo 39051037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3906e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 390729c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3908a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 390960f5a4bcSLai Jiangshan 3910e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3911e02b9312SValentin Schneider 3912e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 391360f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 39141258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 391560f5a4bcSLai Jiangshan 391660f5a4bcSLai Jiangshan if (pool->detach_completion) 391760f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 391860f5a4bcSLai Jiangshan 391929c91e99STejun Heo /* shut down the timers */ 392029c91e99STejun Heo del_timer_sync(&pool->idle_timer); 39213f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 392229c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 392329c91e99STejun Heo 392424acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 392525b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 392629c91e99STejun Heo } 392729c91e99STejun Heo 392829c91e99STejun Heo /** 392929c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 393029c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 393129c91e99STejun Heo * 393229c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 393329c91e99STejun Heo * reference count and return it. If there already is a matching 393429c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3935d185af30SYacine Belkadi * create a new one. 3936a892caccSTejun Heo * 3937a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3938d185af30SYacine Belkadi * 3939d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3940d185af30SYacine Belkadi * On failure, %NULL. 394129c91e99STejun Heo */ 394229c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 394329c91e99STejun Heo { 394429c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 394529c91e99STejun Heo struct worker_pool *pool; 3946f3f90ad4STejun Heo int node; 3947e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 394829c91e99STejun Heo 3949a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 395029c91e99STejun Heo 395129c91e99STejun Heo /* do we already have a matching pool? */ 395229c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 395329c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 395429c91e99STejun Heo pool->refcnt++; 39553fb1823cSLai Jiangshan return pool; 395629c91e99STejun Heo } 395729c91e99STejun Heo } 395829c91e99STejun Heo 3959e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3960e2273584SXunlei Pang if (wq_numa_enabled) { 3961e2273584SXunlei Pang for_each_node(node) { 3962e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3963e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3964e2273584SXunlei Pang target_node = node; 3965e2273584SXunlei Pang break; 3966e2273584SXunlei Pang } 3967e2273584SXunlei Pang } 3968e2273584SXunlei Pang } 3969e2273584SXunlei Pang 397029c91e99STejun Heo /* nope, create a new one */ 3971e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 397229c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 397329c91e99STejun Heo goto fail; 397429c91e99STejun Heo 39758864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 397629c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3977e2273584SXunlei Pang pool->node = target_node; 397829c91e99STejun Heo 39792865a8fbSShaohua Li /* 39802865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 39812865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 39822865a8fbSShaohua Li */ 39832865a8fbSShaohua Li pool->attrs->no_numa = false; 39842865a8fbSShaohua Li 398529c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 398629c91e99STejun Heo goto fail; 398729c91e99STejun Heo 398829c91e99STejun Heo /* create and start the initial worker */ 39893347fa09STejun Heo if (wq_online && !create_worker(pool)) 399029c91e99STejun Heo goto fail; 399129c91e99STejun Heo 399229c91e99STejun Heo /* install */ 399329c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 39943fb1823cSLai Jiangshan 399529c91e99STejun Heo return pool; 399629c91e99STejun Heo fail: 399729c91e99STejun Heo if (pool) 399829c91e99STejun Heo put_unbound_pool(pool); 399929c91e99STejun Heo return NULL; 400029c91e99STejun Heo } 400129c91e99STejun Heo 40028864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 40038864b4e5STejun Heo { 40048864b4e5STejun Heo kmem_cache_free(pwq_cache, 40058864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 40068864b4e5STejun Heo } 40078864b4e5STejun Heo 40088864b4e5STejun Heo /* 40098864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 40108864b4e5STejun Heo * and needs to be destroyed. 40118864b4e5STejun Heo */ 40128864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 40138864b4e5STejun Heo { 40148864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 40158864b4e5STejun Heo unbound_release_work); 40168864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 40178864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 4018b42b0bddSYang Yingliang bool is_last = false; 40198864b4e5STejun Heo 4020b42b0bddSYang Yingliang /* 4021b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 4022b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 4023b42b0bddSYang Yingliang */ 4024b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 40258864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 40268864b4e5STejun Heo return; 40278864b4e5STejun Heo 40283c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 40298864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 4030bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 40313c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 4032b42b0bddSYang Yingliang } 40338864b4e5STejun Heo 4034a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 40358864b4e5STejun Heo put_unbound_pool(pool); 4036a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 4037a892caccSTejun Heo 403825b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 40398864b4e5STejun Heo 40408864b4e5STejun Heo /* 40418864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 4042e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 40438864b4e5STejun Heo */ 4044669de8bdSBart Van Assche if (is_last) { 4045669de8bdSBart Van Assche wq_unregister_lockdep(wq); 404625b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 40476029a918STejun Heo } 4048669de8bdSBart Van Assche } 40498864b4e5STejun Heo 40500fbd95aaSTejun Heo /** 4051699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 40520fbd95aaSTejun Heo * @pwq: target pool_workqueue 40530fbd95aaSTejun Heo * 4054699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4055f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4056699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 40570fbd95aaSTejun Heo */ 4058699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 40590fbd95aaSTejun Heo { 4060699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4061699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 40623347fa09STejun Heo unsigned long flags; 4063699ce097STejun Heo 4064699ce097STejun Heo /* for @wq->saved_max_active */ 4065a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4066699ce097STejun Heo 4067699ce097STejun Heo /* fast exit for non-freezable wqs */ 4068699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4069699ce097STejun Heo return; 4070699ce097STejun Heo 40713347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4072a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4073699ce097STejun Heo 407474b414eaSLai Jiangshan /* 407574b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 407674b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 407774b414eaSLai Jiangshan * is updated and visible. 407874b414eaSLai Jiangshan */ 407974b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 408001341fbdSYunfeng Ye bool kick = false; 408101341fbdSYunfeng Ye 4082699ce097STejun Heo pwq->max_active = wq->saved_max_active; 40830fbd95aaSTejun Heo 4084f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 408501341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 4086f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 408701341fbdSYunfeng Ye kick = true; 408801341fbdSYunfeng Ye } 4089951a078aSLai Jiangshan 4090951a078aSLai Jiangshan /* 4091951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 409201341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 409301341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 409401341fbdSYunfeng Ye * let's kick iff work items were activated. 4095951a078aSLai Jiangshan */ 409601341fbdSYunfeng Ye if (kick) 4097951a078aSLai Jiangshan wake_up_worker(pwq->pool); 4098699ce097STejun Heo } else { 4099699ce097STejun Heo pwq->max_active = 0; 4100699ce097STejun Heo } 4101699ce097STejun Heo 4102a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 41030fbd95aaSTejun Heo } 41040fbd95aaSTejun Heo 410567dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4106f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4107f147f29eSTejun Heo struct worker_pool *pool) 4108d2c1d404STejun Heo { 4109d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4110d2c1d404STejun Heo 4111e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4112e50aba9aSTejun Heo 4113d2c1d404STejun Heo pwq->pool = pool; 4114d2c1d404STejun Heo pwq->wq = wq; 4115d2c1d404STejun Heo pwq->flush_color = -1; 41168864b4e5STejun Heo pwq->refcnt = 1; 4117f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 41181befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4119d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 41208864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 4121f147f29eSTejun Heo } 4122d2c1d404STejun Heo 4123f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 41241befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4125f147f29eSTejun Heo { 4126f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4127f147f29eSTejun Heo 4128f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 412975ccf595STejun Heo 41301befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 41311befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 41321befcf30STejun Heo return; 41331befcf30STejun Heo 413429b1cb41SLai Jiangshan /* set the matching work_color */ 413575ccf595STejun Heo pwq->work_color = wq->work_color; 4136983ca25eSTejun Heo 4137983ca25eSTejun Heo /* sync max_active to the current setting */ 4138983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4139983ca25eSTejun Heo 4140983ca25eSTejun Heo /* link in @pwq */ 41419e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4142df2d5ae4STejun Heo } 41436029a918STejun Heo 4144f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4145f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4146f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4147f147f29eSTejun Heo { 4148f147f29eSTejun Heo struct worker_pool *pool; 4149f147f29eSTejun Heo struct pool_workqueue *pwq; 4150f147f29eSTejun Heo 4151f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4152f147f29eSTejun Heo 4153f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4154f147f29eSTejun Heo if (!pool) 4155f147f29eSTejun Heo return NULL; 4156f147f29eSTejun Heo 4157e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4158f147f29eSTejun Heo if (!pwq) { 4159f147f29eSTejun Heo put_unbound_pool(pool); 4160f147f29eSTejun Heo return NULL; 4161f147f29eSTejun Heo } 4162f147f29eSTejun Heo 4163f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4164f147f29eSTejun Heo return pwq; 4165d2c1d404STejun Heo } 4166d2c1d404STejun Heo 41674c16bd32STejun Heo /** 416830186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 4169042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 41704c16bd32STejun Heo * @node: the target NUMA node 41714c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 41724c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 41734c16bd32STejun Heo * 41744c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 41754c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 4176d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 41774c16bd32STejun Heo * 41784c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 41794c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 41804c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 41814c16bd32STejun Heo * @attrs->cpumask. 41824c16bd32STejun Heo * 41834c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 41844c16bd32STejun Heo * stable. 4185d185af30SYacine Belkadi * 4186d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 4187d185af30SYacine Belkadi * %false if equal. 41884c16bd32STejun Heo */ 41894c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 41904c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 41914c16bd32STejun Heo { 4192d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 41934c16bd32STejun Heo goto use_dfl; 41944c16bd32STejun Heo 41954c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 41964c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 41974c16bd32STejun Heo if (cpu_going_down >= 0) 41984c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 41994c16bd32STejun Heo 42004c16bd32STejun Heo if (cpumask_empty(cpumask)) 42014c16bd32STejun Heo goto use_dfl; 42024c16bd32STejun Heo 42034c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 42044c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 42051ad0f0a7SMichael Bringmann 42061ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 42071ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 42081ad0f0a7SMichael Bringmann "possible intersect\n"); 42091ad0f0a7SMichael Bringmann return false; 42101ad0f0a7SMichael Bringmann } 42111ad0f0a7SMichael Bringmann 42124c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 42134c16bd32STejun Heo 42144c16bd32STejun Heo use_dfl: 42154c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 42164c16bd32STejun Heo return false; 42174c16bd32STejun Heo } 42184c16bd32STejun Heo 42191befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 42201befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 42211befcf30STejun Heo int node, 42221befcf30STejun Heo struct pool_workqueue *pwq) 42231befcf30STejun Heo { 42241befcf30STejun Heo struct pool_workqueue *old_pwq; 42251befcf30STejun Heo 42265b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42271befcf30STejun Heo lockdep_assert_held(&wq->mutex); 42281befcf30STejun Heo 42291befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 42301befcf30STejun Heo link_pwq(pwq); 42311befcf30STejun Heo 42321befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 42331befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 42341befcf30STejun Heo return old_pwq; 42351befcf30STejun Heo } 42361befcf30STejun Heo 42372d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 42382d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 42392d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 42402d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4241042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 42422d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 42432d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 42442d5f0764SLai Jiangshan }; 42452d5f0764SLai Jiangshan 42462d5f0764SLai Jiangshan /* free the resources after success or abort */ 42472d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 42482d5f0764SLai Jiangshan { 42492d5f0764SLai Jiangshan if (ctx) { 42502d5f0764SLai Jiangshan int node; 42512d5f0764SLai Jiangshan 42522d5f0764SLai Jiangshan for_each_node(node) 42532d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 42542d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 42552d5f0764SLai Jiangshan 42562d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 42572d5f0764SLai Jiangshan 42582d5f0764SLai Jiangshan kfree(ctx); 42592d5f0764SLai Jiangshan } 42602d5f0764SLai Jiangshan } 42612d5f0764SLai Jiangshan 42622d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 42632d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 42642d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 426599c621efSLai Jiangshan const struct workqueue_attrs *attrs, 426699c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 42672d5f0764SLai Jiangshan { 42682d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 42692d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 42702d5f0764SLai Jiangshan int node; 42712d5f0764SLai Jiangshan 42722d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42732d5f0764SLai Jiangshan 4274acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 42752d5f0764SLai Jiangshan 4276be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4277be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 42782d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 42792d5f0764SLai Jiangshan goto out_free; 42802d5f0764SLai Jiangshan 4281042f7df1SLai Jiangshan /* 428299c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 428399c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4284042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4285042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4286042f7df1SLai Jiangshan */ 42872d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 428899c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4289042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 429099c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 42912d5f0764SLai Jiangshan 42922d5f0764SLai Jiangshan /* 42932d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 42942d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 42952d5f0764SLai Jiangshan * pools. 42962d5f0764SLai Jiangshan */ 42972d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 42982d5f0764SLai Jiangshan 42992d5f0764SLai Jiangshan /* 43002d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 43012d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 43022d5f0764SLai Jiangshan * it even if we don't use it immediately. 43032d5f0764SLai Jiangshan */ 43042d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 43052d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 43062d5f0764SLai Jiangshan goto out_free; 43072d5f0764SLai Jiangshan 43082d5f0764SLai Jiangshan for_each_node(node) { 4309042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 43102d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 43112d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 43122d5f0764SLai Jiangshan goto out_free; 43132d5f0764SLai Jiangshan } else { 43142d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 43152d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 43162d5f0764SLai Jiangshan } 43172d5f0764SLai Jiangshan } 43182d5f0764SLai Jiangshan 4319042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4320042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4321042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 43222d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4323042f7df1SLai Jiangshan 43242d5f0764SLai Jiangshan ctx->wq = wq; 43252d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43262d5f0764SLai Jiangshan return ctx; 43272d5f0764SLai Jiangshan 43282d5f0764SLai Jiangshan out_free: 43292d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 43302d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 43312d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 43322d5f0764SLai Jiangshan return NULL; 43332d5f0764SLai Jiangshan } 43342d5f0764SLai Jiangshan 43352d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 43362d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 43372d5f0764SLai Jiangshan { 43382d5f0764SLai Jiangshan int node; 43392d5f0764SLai Jiangshan 43402d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 43412d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 43422d5f0764SLai Jiangshan 43432d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 43442d5f0764SLai Jiangshan 43452d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 43462d5f0764SLai Jiangshan for_each_node(node) 43472d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 43482d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 43492d5f0764SLai Jiangshan 43502d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 43512d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 43522d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 43532d5f0764SLai Jiangshan 43542d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 43552d5f0764SLai Jiangshan } 43562d5f0764SLai Jiangshan 4357a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4358a0111cf6SLai Jiangshan { 4359a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4360ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4361a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4362a0111cf6SLai Jiangshan } 4363a0111cf6SLai Jiangshan 4364a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4365a0111cf6SLai Jiangshan { 4366a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4367ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4368a0111cf6SLai Jiangshan } 4369a0111cf6SLai Jiangshan 4370a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4371a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4372a0111cf6SLai Jiangshan { 4373a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4374a0111cf6SLai Jiangshan 4375a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4376a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4377a0111cf6SLai Jiangshan return -EINVAL; 4378a0111cf6SLai Jiangshan 4379a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 43800a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 43810a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4382a0111cf6SLai Jiangshan return -EINVAL; 4383a0111cf6SLai Jiangshan 43840a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 43850a94efb5STejun Heo } 43860a94efb5STejun Heo 438799c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 43886201171eSwanghaibin if (!ctx) 43896201171eSwanghaibin return -ENOMEM; 4390a0111cf6SLai Jiangshan 4391a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4392a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4393a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4394a0111cf6SLai Jiangshan 43956201171eSwanghaibin return 0; 4396a0111cf6SLai Jiangshan } 4397a0111cf6SLai Jiangshan 43989e8cd2f5STejun Heo /** 43999e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 44009e8cd2f5STejun Heo * @wq: the target workqueue 44019e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 44029e8cd2f5STejun Heo * 44034c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 44044c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 44054c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 44064c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 44074c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 44084c16bd32STejun Heo * back-to-back will stay on its current pwq. 44099e8cd2f5STejun Heo * 4410d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4411d185af30SYacine Belkadi * 4412ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4413509b3204SDaniel Jordan * 4414d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 44159e8cd2f5STejun Heo */ 4416513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 44179e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 44189e8cd2f5STejun Heo { 4419a0111cf6SLai Jiangshan int ret; 44209e8cd2f5STejun Heo 4421509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4422509b3204SDaniel Jordan 4423509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4424a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4425509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 44262d5f0764SLai Jiangshan 44272d5f0764SLai Jiangshan return ret; 44289e8cd2f5STejun Heo } 44299e8cd2f5STejun Heo 44304c16bd32STejun Heo /** 44314c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 44324c16bd32STejun Heo * @wq: the target workqueue 44334c16bd32STejun Heo * @cpu: the CPU coming up or going down 44344c16bd32STejun Heo * @online: whether @cpu is coming up or going down 44354c16bd32STejun Heo * 44364c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 44374c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 44384c16bd32STejun Heo * @wq accordingly. 44394c16bd32STejun Heo * 44404c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 44414c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 44424c16bd32STejun Heo * correct. 44434c16bd32STejun Heo * 44444c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 44454c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 44464c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 44474c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 44484c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 44494c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 44504c16bd32STejun Heo * CPU_DOWN_PREPARE. 44514c16bd32STejun Heo */ 44524c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 44534c16bd32STejun Heo bool online) 44544c16bd32STejun Heo { 44554c16bd32STejun Heo int node = cpu_to_node(cpu); 44564c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 44574c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 44584c16bd32STejun Heo struct workqueue_attrs *target_attrs; 44594c16bd32STejun Heo cpumask_t *cpumask; 44604c16bd32STejun Heo 44614c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 44624c16bd32STejun Heo 4463f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4464f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 44654c16bd32STejun Heo return; 44664c16bd32STejun Heo 44674c16bd32STejun Heo /* 44684c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 44694c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 44704c16bd32STejun Heo * CPU hotplug exclusion. 44714c16bd32STejun Heo */ 44724c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 44734c16bd32STejun Heo cpumask = target_attrs->cpumask; 44744c16bd32STejun Heo 44754c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 44764c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 44774c16bd32STejun Heo 44784c16bd32STejun Heo /* 44794c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4480042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4481042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4482042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 44834c16bd32STejun Heo */ 4484042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 44854c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4486f7142ed4SLai Jiangshan return; 44874c16bd32STejun Heo } else { 44884c16bd32STejun Heo goto use_dfl_pwq; 44894c16bd32STejun Heo } 44904c16bd32STejun Heo 44914c16bd32STejun Heo /* create a new pwq */ 44924c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 44934c16bd32STejun Heo if (!pwq) { 44942d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 44954c16bd32STejun Heo wq->name); 449677f300b1SDaeseok Youn goto use_dfl_pwq; 44974c16bd32STejun Heo } 44984c16bd32STejun Heo 4499f7142ed4SLai Jiangshan /* Install the new pwq. */ 45004c16bd32STejun Heo mutex_lock(&wq->mutex); 45014c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 45024c16bd32STejun Heo goto out_unlock; 45034c16bd32STejun Heo 45044c16bd32STejun Heo use_dfl_pwq: 4505f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4506a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 45074c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4508a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 45094c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 45104c16bd32STejun Heo out_unlock: 45114c16bd32STejun Heo mutex_unlock(&wq->mutex); 45124c16bd32STejun Heo put_pwq_unlocked(old_pwq); 45134c16bd32STejun Heo } 45144c16bd32STejun Heo 451530cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 45161da177e4SLinus Torvalds { 451749e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 45188a2b7538STejun Heo int cpu, ret; 4519e1d8aa9fSFrederic Weisbecker 452030cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4521*ee1ceef7STejun Heo wq->cpu_pwq = alloc_percpu(struct pool_workqueue); 4522*ee1ceef7STejun Heo if (!wq->cpu_pwq) 452330cdf249STejun Heo return -ENOMEM; 452430cdf249STejun Heo 452530cdf249STejun Heo for_each_possible_cpu(cpu) { 45267fb98ea7STejun Heo struct pool_workqueue *pwq = 4527*ee1ceef7STejun Heo per_cpu_ptr(wq->cpu_pwq, cpu); 45287a62c2c8STejun Heo struct worker_pool *cpu_pools = 4529f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 453030cdf249STejun Heo 4531f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4532f147f29eSTejun Heo 4533f147f29eSTejun Heo mutex_lock(&wq->mutex); 45341befcf30STejun Heo link_pwq(pwq); 4535f147f29eSTejun Heo mutex_unlock(&wq->mutex); 453630cdf249STejun Heo } 453730cdf249STejun Heo return 0; 4538509b3204SDaniel Jordan } 4539509b3204SDaniel Jordan 4540ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4541509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 45428a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 45438a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 45448a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 45458a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 45468a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 45479e8cd2f5STejun Heo } else { 4548509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 45499e8cd2f5STejun Heo } 4550ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4551509b3204SDaniel Jordan 4552509b3204SDaniel Jordan return ret; 45530f900049STejun Heo } 45540f900049STejun Heo 4555f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4556f3421797STejun Heo const char *name) 4557b71ab8c2STejun Heo { 4558f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4559f3421797STejun Heo 4560f3421797STejun Heo if (max_active < 1 || max_active > lim) 4561044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4562f3421797STejun Heo max_active, name, 1, lim); 4563b71ab8c2STejun Heo 4564f3421797STejun Heo return clamp_val(max_active, 1, lim); 4565b71ab8c2STejun Heo } 4566b71ab8c2STejun Heo 4567983c7515STejun Heo /* 4568983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4569983c7515STejun Heo * to guarantee forward progress. 4570983c7515STejun Heo */ 4571983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4572983c7515STejun Heo { 4573983c7515STejun Heo struct worker *rescuer; 4574b92b36eaSDan Carpenter int ret; 4575983c7515STejun Heo 4576983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4577983c7515STejun Heo return 0; 4578983c7515STejun Heo 4579983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 45804c0736a7SPetr Mladek if (!rescuer) { 45814c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 45824c0736a7SPetr Mladek wq->name); 4583983c7515STejun Heo return -ENOMEM; 45844c0736a7SPetr Mladek } 4585983c7515STejun Heo 4586983c7515STejun Heo rescuer->rescue_wq = wq; 4587983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4588f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4589b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 45904c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 45914c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4592983c7515STejun Heo kfree(rescuer); 4593b92b36eaSDan Carpenter return ret; 4594983c7515STejun Heo } 4595983c7515STejun Heo 4596983c7515STejun Heo wq->rescuer = rescuer; 4597983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4598983c7515STejun Heo wake_up_process(rescuer->task); 4599983c7515STejun Heo 4600983c7515STejun Heo return 0; 4601983c7515STejun Heo } 4602983c7515STejun Heo 4603a2775bbcSMathieu Malaterre __printf(1, 4) 4604669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 460597e37d7bSTejun Heo unsigned int flags, 4606669de8bdSBart Van Assche int max_active, ...) 46073af24433SOleg Nesterov { 4608df2d5ae4STejun Heo size_t tbl_size = 0; 4609ecf6881fSTejun Heo va_list args; 46103af24433SOleg Nesterov struct workqueue_struct *wq; 461149e3cf44STejun Heo struct pool_workqueue *pwq; 4612b196be89STejun Heo 46135c0338c6STejun Heo /* 46145c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 46155c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 46165c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 46175c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 46185c0338c6STejun Heo * on NUMA. 46195c0338c6STejun Heo */ 46205c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 46215c0338c6STejun Heo flags |= __WQ_ORDERED; 46225c0338c6STejun Heo 4623cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4624cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4625cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4626cee22a15SViresh Kumar 4627ecf6881fSTejun Heo /* allocate wq and format name */ 4628df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4629ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4630df2d5ae4STejun Heo 4631df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4632b196be89STejun Heo if (!wq) 4633d2c1d404STejun Heo return NULL; 4634b196be89STejun Heo 46356029a918STejun Heo if (flags & WQ_UNBOUND) { 4636be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 46376029a918STejun Heo if (!wq->unbound_attrs) 46386029a918STejun Heo goto err_free_wq; 46396029a918STejun Heo } 46406029a918STejun Heo 4641669de8bdSBart Van Assche va_start(args, max_active); 4642ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4643b196be89STejun Heo va_end(args); 46443af24433SOleg Nesterov 4645d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4646b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 46473af24433SOleg Nesterov 4648b196be89STejun Heo /* init wq */ 464997e37d7bSTejun Heo wq->flags = flags; 4650a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 46513c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4652112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 465330cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 465473f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 465573f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4656493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 46573af24433SOleg Nesterov 4658669de8bdSBart Van Assche wq_init_lockdep(wq); 4659cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 46603af24433SOleg Nesterov 466130cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 466282efcab3SBart Van Assche goto err_unreg_lockdep; 46631537663fSTejun Heo 466440c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4665d2c1d404STejun Heo goto err_destroy; 4666e22bee78STejun Heo 4667226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4668226223abSTejun Heo goto err_destroy; 4669226223abSTejun Heo 46706af8bf3dSOleg Nesterov /* 467168e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 467268e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 467368e13a67SLai Jiangshan * list. 46746af8bf3dSOleg Nesterov */ 467568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4676a0a1a5fdSTejun Heo 4677a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 467849e3cf44STejun Heo for_each_pwq(pwq, wq) 4679699ce097STejun Heo pwq_adjust_max_active(pwq); 4680a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4681a0a1a5fdSTejun Heo 4682e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4683a0a1a5fdSTejun Heo 468468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46853af24433SOleg Nesterov 46863af24433SOleg Nesterov return wq; 4687d2c1d404STejun Heo 468882efcab3SBart Van Assche err_unreg_lockdep: 4689009bb421SBart Van Assche wq_unregister_lockdep(wq); 4690009bb421SBart Van Assche wq_free_lockdep(wq); 469182efcab3SBart Van Assche err_free_wq: 46926029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 46934690c4abSTejun Heo kfree(wq); 4694d2c1d404STejun Heo return NULL; 4695d2c1d404STejun Heo err_destroy: 4696d2c1d404STejun Heo destroy_workqueue(wq); 46974690c4abSTejun Heo return NULL; 46981da177e4SLinus Torvalds } 4699669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 47001da177e4SLinus Torvalds 4701c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4702c29eb853STejun Heo { 4703c29eb853STejun Heo int i; 4704c29eb853STejun Heo 4705c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4706c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4707c29eb853STejun Heo return true; 4708c29eb853STejun Heo 4709c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4710c29eb853STejun Heo return true; 4711f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4712c29eb853STejun Heo return true; 4713c29eb853STejun Heo 4714c29eb853STejun Heo return false; 4715c29eb853STejun Heo } 4716c29eb853STejun Heo 47173af24433SOleg Nesterov /** 47183af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 47193af24433SOleg Nesterov * @wq: target workqueue 47203af24433SOleg Nesterov * 47213af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 47223af24433SOleg Nesterov */ 47233af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 47243af24433SOleg Nesterov { 472549e3cf44STejun Heo struct pool_workqueue *pwq; 47264c16bd32STejun Heo int node; 47273af24433SOleg Nesterov 4728def98c84STejun Heo /* 4729def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4730def98c84STejun Heo * lead to sysfs name conflicts. 4731def98c84STejun Heo */ 4732def98c84STejun Heo workqueue_sysfs_unregister(wq); 4733def98c84STejun Heo 473433e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 473533e3f0a3SRichard Clark mutex_lock(&wq->mutex); 473633e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 473733e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 473833e3f0a3SRichard Clark 47399c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 47409c5a2ba7STejun Heo drain_workqueue(wq); 4741c8efcc25STejun Heo 4742def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4743def98c84STejun Heo if (wq->rescuer) { 4744def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4745def98c84STejun Heo 4746def98c84STejun Heo /* this prevents new queueing */ 4747a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4748def98c84STejun Heo wq->rescuer = NULL; 4749a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4750def98c84STejun Heo 4751def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4752def98c84STejun Heo kthread_stop(rescuer->task); 47538efe1223STejun Heo kfree(rescuer); 4754def98c84STejun Heo } 4755def98c84STejun Heo 4756c29eb853STejun Heo /* 4757c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4758c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4759c29eb853STejun Heo */ 4760c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4761b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 476249e3cf44STejun Heo for_each_pwq(pwq, wq) { 4763a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4764c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 47651d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4766e66b39afSTejun Heo __func__, wq->name); 4767c29eb853STejun Heo show_pwq(pwq); 4768a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4769b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4770c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 477155df0933SImran Khan show_one_workqueue(wq); 47726183c009STejun Heo return; 477376af4d93STejun Heo } 4774a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 477576af4d93STejun Heo } 4776b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 47776183c009STejun Heo 4778a0a1a5fdSTejun Heo /* 4779a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4780a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4781a0a1a5fdSTejun Heo */ 4782e2dca7adSTejun Heo list_del_rcu(&wq->list); 478368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47843af24433SOleg Nesterov 47858864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4786669de8bdSBart Van Assche wq_unregister_lockdep(wq); 478729c91e99STejun Heo /* 47888864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4789e2dca7adSTejun Heo * schedule RCU free. 479029c91e99STejun Heo */ 479125b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 47928864b4e5STejun Heo } else { 47938864b4e5STejun Heo /* 47948864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 47954c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 47964c16bd32STejun Heo * @wq will be freed when the last pwq is released. 47978864b4e5STejun Heo */ 47984c16bd32STejun Heo for_each_node(node) { 47994c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 48004c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 48014c16bd32STejun Heo put_pwq_unlocked(pwq); 48024c16bd32STejun Heo } 48034c16bd32STejun Heo 48044c16bd32STejun Heo /* 48054c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 48064c16bd32STejun Heo * put. Don't access it afterwards. 48074c16bd32STejun Heo */ 48084c16bd32STejun Heo pwq = wq->dfl_pwq; 48094c16bd32STejun Heo wq->dfl_pwq = NULL; 4810dce90d47STejun Heo put_pwq_unlocked(pwq); 481129c91e99STejun Heo } 48123af24433SOleg Nesterov } 48133af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 48143af24433SOleg Nesterov 4815dcd989cbSTejun Heo /** 4816dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4817dcd989cbSTejun Heo * @wq: target workqueue 4818dcd989cbSTejun Heo * @max_active: new max_active value. 4819dcd989cbSTejun Heo * 4820dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4821dcd989cbSTejun Heo * 4822dcd989cbSTejun Heo * CONTEXT: 4823dcd989cbSTejun Heo * Don't call from IRQ context. 4824dcd989cbSTejun Heo */ 4825dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4826dcd989cbSTejun Heo { 482749e3cf44STejun Heo struct pool_workqueue *pwq; 4828dcd989cbSTejun Heo 48298719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 48300a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 48318719dceaSTejun Heo return; 48328719dceaSTejun Heo 4833f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4834dcd989cbSTejun Heo 4835a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4836dcd989cbSTejun Heo 48370a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4838dcd989cbSTejun Heo wq->saved_max_active = max_active; 4839dcd989cbSTejun Heo 4840699ce097STejun Heo for_each_pwq(pwq, wq) 4841699ce097STejun Heo pwq_adjust_max_active(pwq); 4842dcd989cbSTejun Heo 4843a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4844dcd989cbSTejun Heo } 4845dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4846dcd989cbSTejun Heo 4847dcd989cbSTejun Heo /** 484827d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 484927d4ee03SLukas Wunner * 485027d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 485127d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 485227d4ee03SLukas Wunner * 485327d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 485427d4ee03SLukas Wunner */ 485527d4ee03SLukas Wunner struct work_struct *current_work(void) 485627d4ee03SLukas Wunner { 485727d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 485827d4ee03SLukas Wunner 485927d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 486027d4ee03SLukas Wunner } 486127d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 486227d4ee03SLukas Wunner 486327d4ee03SLukas Wunner /** 4864e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4865e6267616STejun Heo * 4866e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4867e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4868d185af30SYacine Belkadi * 4869d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4870e6267616STejun Heo */ 4871e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4872e6267616STejun Heo { 4873e6267616STejun Heo struct worker *worker = current_wq_worker(); 4874e6267616STejun Heo 48756a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4876e6267616STejun Heo } 4877e6267616STejun Heo 4878e6267616STejun Heo /** 4879dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4880dcd989cbSTejun Heo * @cpu: CPU in question 4881dcd989cbSTejun Heo * @wq: target workqueue 4882dcd989cbSTejun Heo * 4883dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4884dcd989cbSTejun Heo * no synchronization around this function and the test result is 4885dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4886dcd989cbSTejun Heo * 4887d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4888d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4889d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4890d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4891d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4892d3251859STejun Heo * 4893d185af30SYacine Belkadi * Return: 4894dcd989cbSTejun Heo * %true if congested, %false otherwise. 4895dcd989cbSTejun Heo */ 4896d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4897dcd989cbSTejun Heo { 48987fb98ea7STejun Heo struct pool_workqueue *pwq; 489976af4d93STejun Heo bool ret; 490076af4d93STejun Heo 490124acfb71SThomas Gleixner rcu_read_lock(); 490224acfb71SThomas Gleixner preempt_disable(); 49037fb98ea7STejun Heo 4904d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4905d3251859STejun Heo cpu = smp_processor_id(); 4906d3251859STejun Heo 49077fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 4908*ee1ceef7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwq, cpu); 49097fb98ea7STejun Heo else 4910df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4911dcd989cbSTejun Heo 4912f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 491324acfb71SThomas Gleixner preempt_enable(); 491424acfb71SThomas Gleixner rcu_read_unlock(); 491576af4d93STejun Heo 491676af4d93STejun Heo return ret; 4917dcd989cbSTejun Heo } 4918dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4919dcd989cbSTejun Heo 4920dcd989cbSTejun Heo /** 4921dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4922dcd989cbSTejun Heo * @work: the work to be tested 4923dcd989cbSTejun Heo * 4924dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4925dcd989cbSTejun Heo * synchronization around this function and the test result is 4926dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4927dcd989cbSTejun Heo * 4928d185af30SYacine Belkadi * Return: 4929dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4930dcd989cbSTejun Heo */ 4931dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4932dcd989cbSTejun Heo { 4933fa1b54e6STejun Heo struct worker_pool *pool; 4934dcd989cbSTejun Heo unsigned long flags; 4935dcd989cbSTejun Heo unsigned int ret = 0; 4936dcd989cbSTejun Heo 4937dcd989cbSTejun Heo if (work_pending(work)) 4938dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4939038366c5SLai Jiangshan 494024acfb71SThomas Gleixner rcu_read_lock(); 4941fa1b54e6STejun Heo pool = get_work_pool(work); 4942038366c5SLai Jiangshan if (pool) { 4943a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4944c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4945dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4946a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4947038366c5SLai Jiangshan } 494824acfb71SThomas Gleixner rcu_read_unlock(); 4949dcd989cbSTejun Heo 4950dcd989cbSTejun Heo return ret; 4951dcd989cbSTejun Heo } 4952dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4953dcd989cbSTejun Heo 49543d1cb205STejun Heo /** 49553d1cb205STejun Heo * set_worker_desc - set description for the current work item 49563d1cb205STejun Heo * @fmt: printf-style format string 49573d1cb205STejun Heo * @...: arguments for the format string 49583d1cb205STejun Heo * 49593d1cb205STejun Heo * This function can be called by a running work function to describe what 49603d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 49613d1cb205STejun Heo * information will be printed out together to help debugging. The 49623d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 49633d1cb205STejun Heo */ 49643d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 49653d1cb205STejun Heo { 49663d1cb205STejun Heo struct worker *worker = current_wq_worker(); 49673d1cb205STejun Heo va_list args; 49683d1cb205STejun Heo 49693d1cb205STejun Heo if (worker) { 49703d1cb205STejun Heo va_start(args, fmt); 49713d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 49723d1cb205STejun Heo va_end(args); 49733d1cb205STejun Heo } 49743d1cb205STejun Heo } 49755c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 49763d1cb205STejun Heo 49773d1cb205STejun Heo /** 49783d1cb205STejun Heo * print_worker_info - print out worker information and description 49793d1cb205STejun Heo * @log_lvl: the log level to use when printing 49803d1cb205STejun Heo * @task: target task 49813d1cb205STejun Heo * 49823d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 49833d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 49843d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 49853d1cb205STejun Heo * 49863d1cb205STejun Heo * This function can be safely called on any task as long as the 49873d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 49883d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 49893d1cb205STejun Heo */ 49903d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 49913d1cb205STejun Heo { 49923d1cb205STejun Heo work_func_t *fn = NULL; 49933d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 49943d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 49953d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 49963d1cb205STejun Heo struct workqueue_struct *wq = NULL; 49973d1cb205STejun Heo struct worker *worker; 49983d1cb205STejun Heo 49993d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 50003d1cb205STejun Heo return; 50013d1cb205STejun Heo 50023d1cb205STejun Heo /* 50033d1cb205STejun Heo * This function is called without any synchronization and @task 50043d1cb205STejun Heo * could be in any state. Be careful with dereferences. 50053d1cb205STejun Heo */ 5006e700591aSPetr Mladek worker = kthread_probe_data(task); 50073d1cb205STejun Heo 50083d1cb205STejun Heo /* 50098bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 50108bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 50113d1cb205STejun Heo */ 5012fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5013fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5014fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5015fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5016fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 50173d1cb205STejun Heo 50183d1cb205STejun Heo if (fn || name[0] || desc[0]) { 5019d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 50208bf89593STejun Heo if (strcmp(name, desc)) 50213d1cb205STejun Heo pr_cont(" (%s)", desc); 50223d1cb205STejun Heo pr_cont("\n"); 50233d1cb205STejun Heo } 50243d1cb205STejun Heo } 50253d1cb205STejun Heo 50263494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 50273494fc30STejun Heo { 50283494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 50293494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 50303494fc30STejun Heo pr_cont(" node=%d", pool->node); 50313494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 50323494fc30STejun Heo } 50333494fc30STejun Heo 5034c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 5035c76feb0dSPaul E. McKenney bool comma; 5036c76feb0dSPaul E. McKenney work_func_t func; 5037c76feb0dSPaul E. McKenney long ctr; 5038c76feb0dSPaul E. McKenney }; 5039c76feb0dSPaul E. McKenney 5040c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5041c76feb0dSPaul E. McKenney { 5042c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 5043c76feb0dSPaul E. McKenney goto out_record; 5044c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 5045c76feb0dSPaul E. McKenney pcwsp->ctr++; 5046c76feb0dSPaul E. McKenney return; 5047c76feb0dSPaul E. McKenney } 5048c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 5049c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5050c76feb0dSPaul E. McKenney else 5051c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5052c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 5053c76feb0dSPaul E. McKenney out_record: 5054c76feb0dSPaul E. McKenney if ((long)func == -1L) 5055c76feb0dSPaul E. McKenney return; 5056c76feb0dSPaul E. McKenney pcwsp->comma = comma; 5057c76feb0dSPaul E. McKenney pcwsp->func = func; 5058c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 5059c76feb0dSPaul E. McKenney } 5060c76feb0dSPaul E. McKenney 5061c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 50623494fc30STejun Heo { 50633494fc30STejun Heo if (work->func == wq_barrier_func) { 50643494fc30STejun Heo struct wq_barrier *barr; 50653494fc30STejun Heo 50663494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 50673494fc30STejun Heo 5068c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 50693494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 50703494fc30STejun Heo task_pid_nr(barr->task)); 50713494fc30STejun Heo } else { 5072c76feb0dSPaul E. McKenney if (!comma) 5073c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5074c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 50753494fc30STejun Heo } 50763494fc30STejun Heo } 50773494fc30STejun Heo 50783494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 50793494fc30STejun Heo { 5080c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 50813494fc30STejun Heo struct worker_pool *pool = pwq->pool; 50823494fc30STejun Heo struct work_struct *work; 50833494fc30STejun Heo struct worker *worker; 50843494fc30STejun Heo bool has_in_flight = false, has_pending = false; 50853494fc30STejun Heo int bkt; 50863494fc30STejun Heo 50873494fc30STejun Heo pr_info(" pwq %d:", pool->id); 50883494fc30STejun Heo pr_cont_pool_info(pool); 50893494fc30STejun Heo 5090e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5091e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 50923494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 50933494fc30STejun Heo 50943494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 50953494fc30STejun Heo if (worker->current_pwq == pwq) { 50963494fc30STejun Heo has_in_flight = true; 50973494fc30STejun Heo break; 50983494fc30STejun Heo } 50993494fc30STejun Heo } 51003494fc30STejun Heo if (has_in_flight) { 51013494fc30STejun Heo bool comma = false; 51023494fc30STejun Heo 51033494fc30STejun Heo pr_info(" in-flight:"); 51043494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51053494fc30STejun Heo if (worker->current_pwq != pwq) 51063494fc30STejun Heo continue; 51073494fc30STejun Heo 5108d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 51093494fc30STejun Heo task_pid_nr(worker->task), 511030ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 51113494fc30STejun Heo worker->current_func); 51123494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5113c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5114c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51153494fc30STejun Heo comma = true; 51163494fc30STejun Heo } 51173494fc30STejun Heo pr_cont("\n"); 51183494fc30STejun Heo } 51193494fc30STejun Heo 51203494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51213494fc30STejun Heo if (get_work_pwq(work) == pwq) { 51223494fc30STejun Heo has_pending = true; 51233494fc30STejun Heo break; 51243494fc30STejun Heo } 51253494fc30STejun Heo } 51263494fc30STejun Heo if (has_pending) { 51273494fc30STejun Heo bool comma = false; 51283494fc30STejun Heo 51293494fc30STejun Heo pr_info(" pending:"); 51303494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51313494fc30STejun Heo if (get_work_pwq(work) != pwq) 51323494fc30STejun Heo continue; 51333494fc30STejun Heo 5134c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51353494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51363494fc30STejun Heo } 5137c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51383494fc30STejun Heo pr_cont("\n"); 51393494fc30STejun Heo } 51403494fc30STejun Heo 5141f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 51423494fc30STejun Heo bool comma = false; 51433494fc30STejun Heo 5144f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5145f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5146c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51473494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51483494fc30STejun Heo } 5149c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51503494fc30STejun Heo pr_cont("\n"); 51513494fc30STejun Heo } 51523494fc30STejun Heo } 51533494fc30STejun Heo 51543494fc30STejun Heo /** 515555df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 515655df0933SImran Khan * @wq: workqueue whose state will be printed 51573494fc30STejun Heo */ 515855df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 51593494fc30STejun Heo { 51603494fc30STejun Heo struct pool_workqueue *pwq; 51613494fc30STejun Heo bool idle = true; 516255df0933SImran Khan unsigned long flags; 51633494fc30STejun Heo 51643494fc30STejun Heo for_each_pwq(pwq, wq) { 5165f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 51663494fc30STejun Heo idle = false; 51673494fc30STejun Heo break; 51683494fc30STejun Heo } 51693494fc30STejun Heo } 517055df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 517155df0933SImran Khan return; 51723494fc30STejun Heo 51733494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 51743494fc30STejun Heo 51753494fc30STejun Heo for_each_pwq(pwq, wq) { 5176a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 517757116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 517857116ce1SJohan Hovold /* 517957116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 518057116ce1SJohan Hovold * drivers that queue work while holding locks 518157116ce1SJohan Hovold * also taken in their write paths. 518257116ce1SJohan Hovold */ 518357116ce1SJohan Hovold printk_deferred_enter(); 51843494fc30STejun Heo show_pwq(pwq); 518557116ce1SJohan Hovold printk_deferred_exit(); 518657116ce1SJohan Hovold } 5187a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 518862635ea8SSergey Senozhatsky /* 518962635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 519055df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 519162635ea8SSergey Senozhatsky * hard lockup. 519262635ea8SSergey Senozhatsky */ 519362635ea8SSergey Senozhatsky touch_nmi_watchdog(); 51943494fc30STejun Heo } 519555df0933SImran Khan 51963494fc30STejun Heo } 51973494fc30STejun Heo 519855df0933SImran Khan /** 519955df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 520055df0933SImran Khan * @pool: worker pool whose state will be printed 520155df0933SImran Khan */ 520255df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 520355df0933SImran Khan { 52043494fc30STejun Heo struct worker *worker; 52053494fc30STejun Heo bool first = true; 520655df0933SImran Khan unsigned long flags; 5207335a42ebSPetr Mladek unsigned long hung = 0; 52083494fc30STejun Heo 5209a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 52103494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 52113494fc30STejun Heo goto next_pool; 5212335a42ebSPetr Mladek 5213335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5214335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5215335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5216335a42ebSPetr Mladek 521757116ce1SJohan Hovold /* 521857116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 521957116ce1SJohan Hovold * queue work while holding locks also taken in their write 522057116ce1SJohan Hovold * paths. 522157116ce1SJohan Hovold */ 522257116ce1SJohan Hovold printk_deferred_enter(); 52233494fc30STejun Heo pr_info("pool %d:", pool->id); 52243494fc30STejun Heo pr_cont_pool_info(pool); 5225335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 52263494fc30STejun Heo if (pool->manager) 52273494fc30STejun Heo pr_cont(" manager: %d", 52283494fc30STejun Heo task_pid_nr(pool->manager->task)); 52293494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 52303494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 52313494fc30STejun Heo task_pid_nr(worker->task)); 52323494fc30STejun Heo first = false; 52333494fc30STejun Heo } 52343494fc30STejun Heo pr_cont("\n"); 523557116ce1SJohan Hovold printk_deferred_exit(); 52363494fc30STejun Heo next_pool: 5237a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 523862635ea8SSergey Senozhatsky /* 523962635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 524055df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 524162635ea8SSergey Senozhatsky * hard lockup. 524262635ea8SSergey Senozhatsky */ 524362635ea8SSergey Senozhatsky touch_nmi_watchdog(); 524455df0933SImran Khan 52453494fc30STejun Heo } 52463494fc30STejun Heo 524755df0933SImran Khan /** 524855df0933SImran Khan * show_all_workqueues - dump workqueue state 524955df0933SImran Khan * 5250704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 525155df0933SImran Khan */ 525255df0933SImran Khan void show_all_workqueues(void) 525355df0933SImran Khan { 525455df0933SImran Khan struct workqueue_struct *wq; 525555df0933SImran Khan struct worker_pool *pool; 525655df0933SImran Khan int pi; 525755df0933SImran Khan 525855df0933SImran Khan rcu_read_lock(); 525955df0933SImran Khan 526055df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 526155df0933SImran Khan 526255df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 526355df0933SImran Khan show_one_workqueue(wq); 526455df0933SImran Khan 526555df0933SImran Khan for_each_pool(pool, pi) 526655df0933SImran Khan show_one_worker_pool(pool); 526755df0933SImran Khan 526824acfb71SThomas Gleixner rcu_read_unlock(); 52693494fc30STejun Heo } 52703494fc30STejun Heo 5271704bc669SJungseung Lee /** 5272704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5273704bc669SJungseung Lee * 5274704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5275704bc669SJungseung Lee * still busy. 5276704bc669SJungseung Lee */ 5277704bc669SJungseung Lee void show_freezable_workqueues(void) 5278704bc669SJungseung Lee { 5279704bc669SJungseung Lee struct workqueue_struct *wq; 5280704bc669SJungseung Lee 5281704bc669SJungseung Lee rcu_read_lock(); 5282704bc669SJungseung Lee 5283704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5284704bc669SJungseung Lee 5285704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5286704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5287704bc669SJungseung Lee continue; 5288704bc669SJungseung Lee show_one_workqueue(wq); 5289704bc669SJungseung Lee } 5290704bc669SJungseung Lee 5291704bc669SJungseung Lee rcu_read_unlock(); 5292704bc669SJungseung Lee } 5293704bc669SJungseung Lee 52946b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 52956b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 52966b59808bSTejun Heo { 52976b59808bSTejun Heo int off; 52986b59808bSTejun Heo 52996b59808bSTejun Heo /* always show the actual comm */ 53006b59808bSTejun Heo off = strscpy(buf, task->comm, size); 53016b59808bSTejun Heo if (off < 0) 53026b59808bSTejun Heo return; 53036b59808bSTejun Heo 5304197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 53056b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 53066b59808bSTejun Heo 5307197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5308197f6accSTejun Heo struct worker *worker = kthread_data(task); 5309197f6accSTejun Heo struct worker_pool *pool = worker->pool; 53106b59808bSTejun Heo 53116b59808bSTejun Heo if (pool) { 5312a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 53136b59808bSTejun Heo /* 5314197f6accSTejun Heo * ->desc tracks information (wq name or 5315197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5316197f6accSTejun Heo * current, prepend '+', otherwise '-'. 53176b59808bSTejun Heo */ 53186b59808bSTejun Heo if (worker->desc[0] != '\0') { 53196b59808bSTejun Heo if (worker->current_work) 53206b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 53216b59808bSTejun Heo worker->desc); 53226b59808bSTejun Heo else 53236b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 53246b59808bSTejun Heo worker->desc); 53256b59808bSTejun Heo } 5326a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 53276b59808bSTejun Heo } 5328197f6accSTejun Heo } 53296b59808bSTejun Heo 53306b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 53316b59808bSTejun Heo } 53326b59808bSTejun Heo 533366448bc2SMathieu Malaterre #ifdef CONFIG_SMP 533466448bc2SMathieu Malaterre 5335db7bccf4STejun Heo /* 5336db7bccf4STejun Heo * CPU hotplug. 5337db7bccf4STejun Heo * 5338e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5339112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5340706026c2STejun Heo * pool which make migrating pending and scheduled works very 5341e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 534294cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5343e22bee78STejun Heo * blocked draining impractical. 5344e22bee78STejun Heo * 534524647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5346628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5347628c78e7STejun Heo * cpu comes back online. 5348db7bccf4STejun Heo */ 5349db7bccf4STejun Heo 5350e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5351db7bccf4STejun Heo { 53524ce62e9eSTejun Heo struct worker_pool *pool; 5353db7bccf4STejun Heo struct worker *worker; 5354db7bccf4STejun Heo 5355f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53561258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5357a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5358e22bee78STejun Heo 5359f2d5a0eeSTejun Heo /* 536092f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 536194cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 536211b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5363b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5364b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5365b4ac9384SLai Jiangshan * is on the same cpu. 5366f2d5a0eeSTejun Heo */ 5367da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5368403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5369db7bccf4STejun Heo 537024647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5371f2d5a0eeSTejun Heo 5372e22bee78STejun Heo /* 5373989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5374989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5375989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5376989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5377989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5378eb283428SLai Jiangshan * are served by workers tied to the pool. 5379e22bee78STejun Heo */ 5380bc35f7efSLai Jiangshan pool->nr_running = 0; 5381eb283428SLai Jiangshan 5382eb283428SLai Jiangshan /* 5383eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5384eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5385eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5386eb283428SLai Jiangshan */ 5387eb283428SLai Jiangshan wake_up_worker(pool); 5388989442d7SLai Jiangshan 5389a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5390989442d7SLai Jiangshan 5391793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5392793777bcSValentin Schneider unbind_worker(worker); 5393989442d7SLai Jiangshan 5394989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5395eb283428SLai Jiangshan } 5396db7bccf4STejun Heo } 5397db7bccf4STejun Heo 5398bd7c089eSTejun Heo /** 5399bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5400bd7c089eSTejun Heo * @pool: pool of interest 5401bd7c089eSTejun Heo * 5402a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5403bd7c089eSTejun Heo */ 5404bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5405bd7c089eSTejun Heo { 5406a9ab775bSTejun Heo struct worker *worker; 5407bd7c089eSTejun Heo 54081258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5409bd7c089eSTejun Heo 5410bd7c089eSTejun Heo /* 5411a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5412a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5413402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5414a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5415a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5416bd7c089eSTejun Heo */ 5417c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5418c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5419c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5420c63a2e52SValentin Schneider pool->attrs->cpumask) < 0); 5421c63a2e52SValentin Schneider } 5422a9ab775bSTejun Heo 5423a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5424f7c17d26SWanpeng Li 54253de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5426a9ab775bSTejun Heo 5427da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5428a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5429a9ab775bSTejun Heo 5430a9ab775bSTejun Heo /* 5431a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5432a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5433a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5434a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5435a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5436a9ab775bSTejun Heo * concurrency management. Note that when or whether 5437a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5438a9ab775bSTejun Heo * 5439c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5440a9ab775bSTejun Heo * tested without holding any lock in 54416d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5442a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5443a9ab775bSTejun Heo * management operations. 5444bd7c089eSTejun Heo */ 5445a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5446a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5447a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5448c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5449bd7c089eSTejun Heo } 5450a9ab775bSTejun Heo 5451a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5452bd7c089eSTejun Heo } 5453bd7c089eSTejun Heo 54547dbc725eSTejun Heo /** 54557dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 54567dbc725eSTejun Heo * @pool: unbound pool of interest 54577dbc725eSTejun Heo * @cpu: the CPU which is coming up 54587dbc725eSTejun Heo * 54597dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 54607dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 54617dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 54627dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 54637dbc725eSTejun Heo */ 54647dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 54657dbc725eSTejun Heo { 54667dbc725eSTejun Heo static cpumask_t cpumask; 54677dbc725eSTejun Heo struct worker *worker; 54687dbc725eSTejun Heo 54691258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 54707dbc725eSTejun Heo 54717dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 54727dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 54737dbc725eSTejun Heo return; 54747dbc725eSTejun Heo 54757dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 54767dbc725eSTejun Heo 54777dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5478da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5479d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 54807dbc725eSTejun Heo } 54817dbc725eSTejun Heo 54827ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 54831da177e4SLinus Torvalds { 54844ce62e9eSTejun Heo struct worker_pool *pool; 54851da177e4SLinus Torvalds 5486f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54873ce63377STejun Heo if (pool->nr_workers) 54883ce63377STejun Heo continue; 5489051e1850SLai Jiangshan if (!create_worker(pool)) 54907ee681b2SThomas Gleixner return -ENOMEM; 54913af24433SOleg Nesterov } 54927ee681b2SThomas Gleixner return 0; 54937ee681b2SThomas Gleixner } 54941da177e4SLinus Torvalds 54957ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 54967ee681b2SThomas Gleixner { 54977ee681b2SThomas Gleixner struct worker_pool *pool; 54987ee681b2SThomas Gleixner struct workqueue_struct *wq; 54997ee681b2SThomas Gleixner int pi; 55007ee681b2SThomas Gleixner 550168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55027dbc725eSTejun Heo 55037dbc725eSTejun Heo for_each_pool(pool, pi) { 55041258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 550594cf58bbSTejun Heo 5506f05b558dSLai Jiangshan if (pool->cpu == cpu) 550794cf58bbSTejun Heo rebind_workers(pool); 5508f05b558dSLai Jiangshan else if (pool->cpu < 0) 55097dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 551094cf58bbSTejun Heo 55111258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 551294cf58bbSTejun Heo } 55137dbc725eSTejun Heo 55144c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55154c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 55164c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 55174c16bd32STejun Heo 551868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55197ee681b2SThomas Gleixner return 0; 552065758202STejun Heo } 552165758202STejun Heo 55227ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 552365758202STejun Heo { 55244c16bd32STejun Heo struct workqueue_struct *wq; 55258db25e78STejun Heo 55264c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5527e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5528e8b3f8dbSLai Jiangshan return -1; 5529e8b3f8dbSLai Jiangshan 5530e8b3f8dbSLai Jiangshan unbind_workers(cpu); 55314c16bd32STejun Heo 55324c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 55334c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 55344c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 55354c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 55364c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 55374c16bd32STejun Heo 55387ee681b2SThomas Gleixner return 0; 553965758202STejun Heo } 554065758202STejun Heo 55412d3854a3SRusty Russell struct work_for_cpu { 5542ed48ece2STejun Heo struct work_struct work; 55432d3854a3SRusty Russell long (*fn)(void *); 55442d3854a3SRusty Russell void *arg; 55452d3854a3SRusty Russell long ret; 55462d3854a3SRusty Russell }; 55472d3854a3SRusty Russell 5548ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 55492d3854a3SRusty Russell { 5550ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5551ed48ece2STejun Heo 55522d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 55532d3854a3SRusty Russell } 55542d3854a3SRusty Russell 55552d3854a3SRusty Russell /** 555622aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 55572d3854a3SRusty Russell * @cpu: the cpu to run on 55582d3854a3SRusty Russell * @fn: the function to run 55592d3854a3SRusty Russell * @arg: the function arg 55602d3854a3SRusty Russell * 556131ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 55626b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5563d185af30SYacine Belkadi * 5564d185af30SYacine Belkadi * Return: The value @fn returns. 55652d3854a3SRusty Russell */ 5566d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 55672d3854a3SRusty Russell { 5568ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 55692d3854a3SRusty Russell 5570ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5571ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 557212997d1aSBjorn Helgaas flush_work(&wfc.work); 5573440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 55742d3854a3SRusty Russell return wfc.ret; 55752d3854a3SRusty Russell } 55762d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 55770e8d6a93SThomas Gleixner 55780e8d6a93SThomas Gleixner /** 55790e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 55800e8d6a93SThomas Gleixner * @cpu: the cpu to run on 55810e8d6a93SThomas Gleixner * @fn: the function to run 55820e8d6a93SThomas Gleixner * @arg: the function argument 55830e8d6a93SThomas Gleixner * 55840e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 55850e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 55860e8d6a93SThomas Gleixner * 55870e8d6a93SThomas Gleixner * Return: The value @fn returns. 55880e8d6a93SThomas Gleixner */ 55890e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 55900e8d6a93SThomas Gleixner { 55910e8d6a93SThomas Gleixner long ret = -ENODEV; 55920e8d6a93SThomas Gleixner 5593ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 55940e8d6a93SThomas Gleixner if (cpu_online(cpu)) 55950e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5596ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 55970e8d6a93SThomas Gleixner return ret; 55980e8d6a93SThomas Gleixner } 55990e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 56002d3854a3SRusty Russell #endif /* CONFIG_SMP */ 56012d3854a3SRusty Russell 5602a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5603e7577c50SRusty Russell 5604a0a1a5fdSTejun Heo /** 5605a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5606a0a1a5fdSTejun Heo * 560758a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5608f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5609706026c2STejun Heo * pool->worklist. 5610a0a1a5fdSTejun Heo * 5611a0a1a5fdSTejun Heo * CONTEXT: 5612a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5613a0a1a5fdSTejun Heo */ 5614a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5615a0a1a5fdSTejun Heo { 561624b8a847STejun Heo struct workqueue_struct *wq; 561724b8a847STejun Heo struct pool_workqueue *pwq; 5618a0a1a5fdSTejun Heo 561968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5620a0a1a5fdSTejun Heo 56216183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5622a0a1a5fdSTejun Heo workqueue_freezing = true; 5623a0a1a5fdSTejun Heo 562424b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5625a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5626699ce097STejun Heo for_each_pwq(pwq, wq) 5627699ce097STejun Heo pwq_adjust_max_active(pwq); 5628a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5629a1056305STejun Heo } 56305bcab335STejun Heo 563168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5632a0a1a5fdSTejun Heo } 5633a0a1a5fdSTejun Heo 5634a0a1a5fdSTejun Heo /** 563558a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5636a0a1a5fdSTejun Heo * 5637a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5638a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5639a0a1a5fdSTejun Heo * 5640a0a1a5fdSTejun Heo * CONTEXT: 564168e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5642a0a1a5fdSTejun Heo * 5643d185af30SYacine Belkadi * Return: 564458a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 564558a69cb4STejun Heo * is complete. 5646a0a1a5fdSTejun Heo */ 5647a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5648a0a1a5fdSTejun Heo { 5649a0a1a5fdSTejun Heo bool busy = false; 565024b8a847STejun Heo struct workqueue_struct *wq; 565124b8a847STejun Heo struct pool_workqueue *pwq; 5652a0a1a5fdSTejun Heo 565368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5654a0a1a5fdSTejun Heo 56556183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5656a0a1a5fdSTejun Heo 565724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 565824b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 565924b8a847STejun Heo continue; 5660a0a1a5fdSTejun Heo /* 5661a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5662a0a1a5fdSTejun Heo * to peek without lock. 5663a0a1a5fdSTejun Heo */ 566424acfb71SThomas Gleixner rcu_read_lock(); 566524b8a847STejun Heo for_each_pwq(pwq, wq) { 56666183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5667112202d9STejun Heo if (pwq->nr_active) { 5668a0a1a5fdSTejun Heo busy = true; 566924acfb71SThomas Gleixner rcu_read_unlock(); 5670a0a1a5fdSTejun Heo goto out_unlock; 5671a0a1a5fdSTejun Heo } 5672a0a1a5fdSTejun Heo } 567324acfb71SThomas Gleixner rcu_read_unlock(); 5674a0a1a5fdSTejun Heo } 5675a0a1a5fdSTejun Heo out_unlock: 567668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5677a0a1a5fdSTejun Heo return busy; 5678a0a1a5fdSTejun Heo } 5679a0a1a5fdSTejun Heo 5680a0a1a5fdSTejun Heo /** 5681a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5682a0a1a5fdSTejun Heo * 5683a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5684706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5685a0a1a5fdSTejun Heo * 5686a0a1a5fdSTejun Heo * CONTEXT: 5687a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5688a0a1a5fdSTejun Heo */ 5689a0a1a5fdSTejun Heo void thaw_workqueues(void) 5690a0a1a5fdSTejun Heo { 569124b8a847STejun Heo struct workqueue_struct *wq; 569224b8a847STejun Heo struct pool_workqueue *pwq; 5693a0a1a5fdSTejun Heo 569468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5695a0a1a5fdSTejun Heo 5696a0a1a5fdSTejun Heo if (!workqueue_freezing) 5697a0a1a5fdSTejun Heo goto out_unlock; 5698a0a1a5fdSTejun Heo 569974b414eaSLai Jiangshan workqueue_freezing = false; 570024b8a847STejun Heo 570124b8a847STejun Heo /* restore max_active and repopulate worklist */ 570224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5703a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5704699ce097STejun Heo for_each_pwq(pwq, wq) 5705699ce097STejun Heo pwq_adjust_max_active(pwq); 5706a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 570724b8a847STejun Heo } 570824b8a847STejun Heo 5709a0a1a5fdSTejun Heo out_unlock: 571068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5711a0a1a5fdSTejun Heo } 5712a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5713a0a1a5fdSTejun Heo 571499c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5715042f7df1SLai Jiangshan { 5716042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5717042f7df1SLai Jiangshan int ret = 0; 5718042f7df1SLai Jiangshan struct workqueue_struct *wq; 5719042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5720042f7df1SLai Jiangshan 5721042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5722042f7df1SLai Jiangshan 5723042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5724042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5725042f7df1SLai Jiangshan continue; 5726042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5727042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5728042f7df1SLai Jiangshan continue; 5729042f7df1SLai Jiangshan 573099c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5731042f7df1SLai Jiangshan if (!ctx) { 5732042f7df1SLai Jiangshan ret = -ENOMEM; 5733042f7df1SLai Jiangshan break; 5734042f7df1SLai Jiangshan } 5735042f7df1SLai Jiangshan 5736042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5737042f7df1SLai Jiangshan } 5738042f7df1SLai Jiangshan 5739042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5740042f7df1SLai Jiangshan if (!ret) 5741042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5742042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5743042f7df1SLai Jiangshan } 5744042f7df1SLai Jiangshan 574599c621efSLai Jiangshan if (!ret) { 574699c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 574799c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 574899c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 574999c621efSLai Jiangshan } 5750042f7df1SLai Jiangshan return ret; 5751042f7df1SLai Jiangshan } 5752042f7df1SLai Jiangshan 5753042f7df1SLai Jiangshan /** 5754042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5755042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5756042f7df1SLai Jiangshan * 5757042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5758042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5759042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5760042f7df1SLai Jiangshan * 576167dc8325SCai Huoqing * Return: 0 - Success 5762042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5763042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5764042f7df1SLai Jiangshan */ 5765042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5766042f7df1SLai Jiangshan { 5767042f7df1SLai Jiangshan int ret = -EINVAL; 5768042f7df1SLai Jiangshan 5769c98a9805STal Shorer /* 5770c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5771c98a9805STal Shorer * If the user wishes to include them, we allow that. 5772c98a9805STal Shorer */ 5773042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5774042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5775a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5776d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5777d25302e4SMenglong Dong ret = 0; 5778d25302e4SMenglong Dong goto out_unlock; 5779d25302e4SMenglong Dong } 5780d25302e4SMenglong Dong 578199c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5782042f7df1SLai Jiangshan 5783d25302e4SMenglong Dong out_unlock: 5784a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5785042f7df1SLai Jiangshan } 5786042f7df1SLai Jiangshan 5787042f7df1SLai Jiangshan return ret; 5788042f7df1SLai Jiangshan } 5789042f7df1SLai Jiangshan 57906ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 57916ba94429SFrederic Weisbecker /* 57926ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 57936ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 57946ba94429SFrederic Weisbecker * following attributes. 57956ba94429SFrederic Weisbecker * 57966ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 57976ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 57986ba94429SFrederic Weisbecker * 57996ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 58006ba94429SFrederic Weisbecker * 58019a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 58026ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 58036ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 58049a19b463SWang Long * numa RW bool : whether enable NUMA affinity 58056ba94429SFrederic Weisbecker */ 58066ba94429SFrederic Weisbecker struct wq_device { 58076ba94429SFrederic Weisbecker struct workqueue_struct *wq; 58086ba94429SFrederic Weisbecker struct device dev; 58096ba94429SFrederic Weisbecker }; 58106ba94429SFrederic Weisbecker 58116ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 58126ba94429SFrederic Weisbecker { 58136ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 58146ba94429SFrederic Weisbecker 58156ba94429SFrederic Weisbecker return wq_dev->wq; 58166ba94429SFrederic Weisbecker } 58176ba94429SFrederic Weisbecker 58186ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 58196ba94429SFrederic Weisbecker char *buf) 58206ba94429SFrederic Weisbecker { 58216ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58226ba94429SFrederic Weisbecker 58236ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 58246ba94429SFrederic Weisbecker } 58256ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 58266ba94429SFrederic Weisbecker 58276ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 58286ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58296ba94429SFrederic Weisbecker { 58306ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58316ba94429SFrederic Weisbecker 58326ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 58336ba94429SFrederic Weisbecker } 58346ba94429SFrederic Weisbecker 58356ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 58366ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 58376ba94429SFrederic Weisbecker size_t count) 58386ba94429SFrederic Weisbecker { 58396ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58406ba94429SFrederic Weisbecker int val; 58416ba94429SFrederic Weisbecker 58426ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 58436ba94429SFrederic Weisbecker return -EINVAL; 58446ba94429SFrederic Weisbecker 58456ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 58466ba94429SFrederic Weisbecker return count; 58476ba94429SFrederic Weisbecker } 58486ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 58496ba94429SFrederic Weisbecker 58506ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 58516ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 58526ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 58536ba94429SFrederic Weisbecker NULL, 58546ba94429SFrederic Weisbecker }; 58556ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 58566ba94429SFrederic Weisbecker 58576ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 58586ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58596ba94429SFrederic Weisbecker { 58606ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58616ba94429SFrederic Weisbecker const char *delim = ""; 58626ba94429SFrederic Weisbecker int node, written = 0; 58636ba94429SFrederic Weisbecker 5864ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 586524acfb71SThomas Gleixner rcu_read_lock(); 58666ba94429SFrederic Weisbecker for_each_node(node) { 58676ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 58686ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 58696ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 58706ba94429SFrederic Weisbecker delim = " "; 58716ba94429SFrederic Weisbecker } 58726ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 587324acfb71SThomas Gleixner rcu_read_unlock(); 5874ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 58756ba94429SFrederic Weisbecker 58766ba94429SFrederic Weisbecker return written; 58776ba94429SFrederic Weisbecker } 58786ba94429SFrederic Weisbecker 58796ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 58806ba94429SFrederic Weisbecker char *buf) 58816ba94429SFrederic Weisbecker { 58826ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58836ba94429SFrederic Weisbecker int written; 58846ba94429SFrederic Weisbecker 58856ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 58866ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 58876ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 58886ba94429SFrederic Weisbecker 58896ba94429SFrederic Weisbecker return written; 58906ba94429SFrederic Weisbecker } 58916ba94429SFrederic Weisbecker 58926ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 58936ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 58946ba94429SFrederic Weisbecker { 58956ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 58966ba94429SFrederic Weisbecker 5897899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5898899a94feSLai Jiangshan 5899be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 59006ba94429SFrederic Weisbecker if (!attrs) 59016ba94429SFrederic Weisbecker return NULL; 59026ba94429SFrederic Weisbecker 59036ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 59046ba94429SFrederic Weisbecker return attrs; 59056ba94429SFrederic Weisbecker } 59066ba94429SFrederic Weisbecker 59076ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 59086ba94429SFrederic Weisbecker const char *buf, size_t count) 59096ba94429SFrederic Weisbecker { 59106ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59116ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5912d4d3e257SLai Jiangshan int ret = -ENOMEM; 5913d4d3e257SLai Jiangshan 5914d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59156ba94429SFrederic Weisbecker 59166ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59176ba94429SFrederic Weisbecker if (!attrs) 5918d4d3e257SLai Jiangshan goto out_unlock; 59196ba94429SFrederic Weisbecker 59206ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 59216ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5922d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59236ba94429SFrederic Weisbecker else 59246ba94429SFrederic Weisbecker ret = -EINVAL; 59256ba94429SFrederic Weisbecker 5926d4d3e257SLai Jiangshan out_unlock: 5927d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59286ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59296ba94429SFrederic Weisbecker return ret ?: count; 59306ba94429SFrederic Weisbecker } 59316ba94429SFrederic Weisbecker 59326ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 59336ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 59346ba94429SFrederic Weisbecker { 59356ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59366ba94429SFrederic Weisbecker int written; 59376ba94429SFrederic Weisbecker 59386ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59396ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 59406ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 59416ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59426ba94429SFrederic Weisbecker return written; 59436ba94429SFrederic Weisbecker } 59446ba94429SFrederic Weisbecker 59456ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 59466ba94429SFrederic Weisbecker struct device_attribute *attr, 59476ba94429SFrederic Weisbecker const char *buf, size_t count) 59486ba94429SFrederic Weisbecker { 59496ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59506ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5951d4d3e257SLai Jiangshan int ret = -ENOMEM; 5952d4d3e257SLai Jiangshan 5953d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59546ba94429SFrederic Weisbecker 59556ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59566ba94429SFrederic Weisbecker if (!attrs) 5957d4d3e257SLai Jiangshan goto out_unlock; 59586ba94429SFrederic Weisbecker 59596ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 59606ba94429SFrederic Weisbecker if (!ret) 5961d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59626ba94429SFrederic Weisbecker 5963d4d3e257SLai Jiangshan out_unlock: 5964d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59656ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59666ba94429SFrederic Weisbecker return ret ?: count; 59676ba94429SFrederic Weisbecker } 59686ba94429SFrederic Weisbecker 59696ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 59706ba94429SFrederic Weisbecker char *buf) 59716ba94429SFrederic Weisbecker { 59726ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59736ba94429SFrederic Weisbecker int written; 59746ba94429SFrederic Weisbecker 59756ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59766ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 59776ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 59786ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59796ba94429SFrederic Weisbecker 59806ba94429SFrederic Weisbecker return written; 59816ba94429SFrederic Weisbecker } 59826ba94429SFrederic Weisbecker 59836ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 59846ba94429SFrederic Weisbecker const char *buf, size_t count) 59856ba94429SFrederic Weisbecker { 59866ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59876ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5988d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5989d4d3e257SLai Jiangshan 5990d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59916ba94429SFrederic Weisbecker 59926ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59936ba94429SFrederic Weisbecker if (!attrs) 5994d4d3e257SLai Jiangshan goto out_unlock; 59956ba94429SFrederic Weisbecker 59966ba94429SFrederic Weisbecker ret = -EINVAL; 59976ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 59986ba94429SFrederic Weisbecker attrs->no_numa = !v; 5999d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 60006ba94429SFrederic Weisbecker } 60016ba94429SFrederic Weisbecker 6002d4d3e257SLai Jiangshan out_unlock: 6003d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 60046ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 60056ba94429SFrederic Weisbecker return ret ?: count; 60066ba94429SFrederic Weisbecker } 60076ba94429SFrederic Weisbecker 60086ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 60096ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 60106ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 60116ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 60126ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 60136ba94429SFrederic Weisbecker __ATTR_NULL, 60146ba94429SFrederic Weisbecker }; 60156ba94429SFrederic Weisbecker 60166ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 60176ba94429SFrederic Weisbecker .name = "workqueue", 60186ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 60196ba94429SFrederic Weisbecker }; 60206ba94429SFrederic Weisbecker 6021b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 6022b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 6023b05a7928SFrederic Weisbecker { 6024b05a7928SFrederic Weisbecker int written; 6025b05a7928SFrederic Weisbecker 6026042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 6027b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6028b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 6029042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 6030b05a7928SFrederic Weisbecker 6031b05a7928SFrederic Weisbecker return written; 6032b05a7928SFrederic Weisbecker } 6033b05a7928SFrederic Weisbecker 6034042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 6035042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 6036042f7df1SLai Jiangshan { 6037042f7df1SLai Jiangshan cpumask_var_t cpumask; 6038042f7df1SLai Jiangshan int ret; 6039042f7df1SLai Jiangshan 6040042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6041042f7df1SLai Jiangshan return -ENOMEM; 6042042f7df1SLai Jiangshan 6043042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 6044042f7df1SLai Jiangshan if (!ret) 6045042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 6046042f7df1SLai Jiangshan 6047042f7df1SLai Jiangshan free_cpumask_var(cpumask); 6048042f7df1SLai Jiangshan return ret ? ret : count; 6049042f7df1SLai Jiangshan } 6050042f7df1SLai Jiangshan 6051b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 6052042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6053042f7df1SLai Jiangshan wq_unbound_cpumask_store); 6054b05a7928SFrederic Weisbecker 60556ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 60566ba94429SFrederic Weisbecker { 6057686f6697SGreg Kroah-Hartman struct device *dev_root; 6058b05a7928SFrederic Weisbecker int err; 6059b05a7928SFrederic Weisbecker 6060b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 6061b05a7928SFrederic Weisbecker if (err) 6062b05a7928SFrederic Weisbecker return err; 6063b05a7928SFrederic Weisbecker 6064686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 6065686f6697SGreg Kroah-Hartman if (dev_root) { 6066686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6067686f6697SGreg Kroah-Hartman put_device(dev_root); 6068686f6697SGreg Kroah-Hartman } 6069686f6697SGreg Kroah-Hartman return err; 60706ba94429SFrederic Weisbecker } 60716ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 60726ba94429SFrederic Weisbecker 60736ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 60746ba94429SFrederic Weisbecker { 60756ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 60766ba94429SFrederic Weisbecker 60776ba94429SFrederic Weisbecker kfree(wq_dev); 60786ba94429SFrederic Weisbecker } 60796ba94429SFrederic Weisbecker 60806ba94429SFrederic Weisbecker /** 60816ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 60826ba94429SFrederic Weisbecker * @wq: the workqueue to register 60836ba94429SFrederic Weisbecker * 60846ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 60856ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 60866ba94429SFrederic Weisbecker * which is the preferred method. 60876ba94429SFrederic Weisbecker * 60886ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 60896ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 60906ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 60916ba94429SFrederic Weisbecker * attributes. 60926ba94429SFrederic Weisbecker * 60936ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 60946ba94429SFrederic Weisbecker */ 60956ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 60966ba94429SFrederic Weisbecker { 60976ba94429SFrederic Weisbecker struct wq_device *wq_dev; 60986ba94429SFrederic Weisbecker int ret; 60996ba94429SFrederic Weisbecker 61006ba94429SFrederic Weisbecker /* 6101402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 61026ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 61036ba94429SFrederic Weisbecker * workqueues. 61046ba94429SFrederic Weisbecker */ 61050a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 61066ba94429SFrederic Weisbecker return -EINVAL; 61076ba94429SFrederic Weisbecker 61086ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 61096ba94429SFrederic Weisbecker if (!wq_dev) 61106ba94429SFrederic Weisbecker return -ENOMEM; 61116ba94429SFrederic Weisbecker 61126ba94429SFrederic Weisbecker wq_dev->wq = wq; 61136ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 61146ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 611523217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 61166ba94429SFrederic Weisbecker 61176ba94429SFrederic Weisbecker /* 61186ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 61196ba94429SFrederic Weisbecker * everything is ready. 61206ba94429SFrederic Weisbecker */ 61216ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 61226ba94429SFrederic Weisbecker 61236ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 61246ba94429SFrederic Weisbecker if (ret) { 6125537f4146SArvind Yadav put_device(&wq_dev->dev); 61266ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61276ba94429SFrederic Weisbecker return ret; 61286ba94429SFrederic Weisbecker } 61296ba94429SFrederic Weisbecker 61306ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 61316ba94429SFrederic Weisbecker struct device_attribute *attr; 61326ba94429SFrederic Weisbecker 61336ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 61346ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 61356ba94429SFrederic Weisbecker if (ret) { 61366ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61376ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61386ba94429SFrederic Weisbecker return ret; 61396ba94429SFrederic Weisbecker } 61406ba94429SFrederic Weisbecker } 61416ba94429SFrederic Weisbecker } 61426ba94429SFrederic Weisbecker 61436ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 61446ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 61456ba94429SFrederic Weisbecker return 0; 61466ba94429SFrederic Weisbecker } 61476ba94429SFrederic Weisbecker 61486ba94429SFrederic Weisbecker /** 61496ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 61506ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 61516ba94429SFrederic Weisbecker * 61526ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 61536ba94429SFrederic Weisbecker */ 61546ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 61556ba94429SFrederic Weisbecker { 61566ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 61576ba94429SFrederic Weisbecker 61586ba94429SFrederic Weisbecker if (!wq->wq_dev) 61596ba94429SFrederic Weisbecker return; 61606ba94429SFrederic Weisbecker 61616ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61626ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61636ba94429SFrederic Weisbecker } 61646ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 61656ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 61666ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 61676ba94429SFrederic Weisbecker 616882607adcSTejun Heo /* 616982607adcSTejun Heo * Workqueue watchdog. 617082607adcSTejun Heo * 617182607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 617282607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 617382607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 617482607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 617582607adcSTejun Heo * largely opaque. 617682607adcSTejun Heo * 617782607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 617882607adcSTejun Heo * state if some pools failed to make forward progress for a while where 617982607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 618082607adcSTejun Heo * 618182607adcSTejun Heo * This mechanism is controlled through the kernel parameter 618282607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 618382607adcSTejun Heo * corresponding sysfs parameter file. 618482607adcSTejun Heo */ 618582607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 618682607adcSTejun Heo 618782607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 61885cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 618982607adcSTejun Heo 619082607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 619182607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 619282607adcSTejun Heo 6193cd2440d6SPetr Mladek /* 6194cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6195cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6196cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6197cd2440d6SPetr Mladek * in all other situations. 6198cd2440d6SPetr Mladek */ 6199cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6200cd2440d6SPetr Mladek { 6201cd2440d6SPetr Mladek struct worker *worker; 6202cd2440d6SPetr Mladek unsigned long flags; 6203cd2440d6SPetr Mladek int bkt; 6204cd2440d6SPetr Mladek 6205cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6206cd2440d6SPetr Mladek 6207cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6208cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6209cd2440d6SPetr Mladek /* 6210cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6211cd2440d6SPetr Mladek * drivers that queue work while holding locks 6212cd2440d6SPetr Mladek * also taken in their write paths. 6213cd2440d6SPetr Mladek */ 6214cd2440d6SPetr Mladek printk_deferred_enter(); 6215cd2440d6SPetr Mladek 6216cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6217cd2440d6SPetr Mladek sched_show_task(worker->task); 6218cd2440d6SPetr Mladek 6219cd2440d6SPetr Mladek printk_deferred_exit(); 6220cd2440d6SPetr Mladek } 6221cd2440d6SPetr Mladek } 6222cd2440d6SPetr Mladek 6223cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6224cd2440d6SPetr Mladek } 6225cd2440d6SPetr Mladek 6226cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6227cd2440d6SPetr Mladek { 6228cd2440d6SPetr Mladek struct worker_pool *pool; 6229cd2440d6SPetr Mladek int pi; 6230cd2440d6SPetr Mladek 6231cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6232cd2440d6SPetr Mladek 6233cd2440d6SPetr Mladek rcu_read_lock(); 6234cd2440d6SPetr Mladek 6235cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6236cd2440d6SPetr Mladek if (pool->cpu_stall) 6237cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6238cd2440d6SPetr Mladek 6239cd2440d6SPetr Mladek } 6240cd2440d6SPetr Mladek 6241cd2440d6SPetr Mladek rcu_read_unlock(); 6242cd2440d6SPetr Mladek } 6243cd2440d6SPetr Mladek 624482607adcSTejun Heo static void wq_watchdog_reset_touched(void) 624582607adcSTejun Heo { 624682607adcSTejun Heo int cpu; 624782607adcSTejun Heo 624882607adcSTejun Heo wq_watchdog_touched = jiffies; 624982607adcSTejun Heo for_each_possible_cpu(cpu) 625082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 625182607adcSTejun Heo } 625282607adcSTejun Heo 62535cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 625482607adcSTejun Heo { 625582607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 625682607adcSTejun Heo bool lockup_detected = false; 6257cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6258940d71c6SSergey Senozhatsky unsigned long now = jiffies; 625982607adcSTejun Heo struct worker_pool *pool; 626082607adcSTejun Heo int pi; 626182607adcSTejun Heo 626282607adcSTejun Heo if (!thresh) 626382607adcSTejun Heo return; 626482607adcSTejun Heo 626582607adcSTejun Heo rcu_read_lock(); 626682607adcSTejun Heo 626782607adcSTejun Heo for_each_pool(pool, pi) { 626882607adcSTejun Heo unsigned long pool_ts, touched, ts; 626982607adcSTejun Heo 6270cd2440d6SPetr Mladek pool->cpu_stall = false; 627182607adcSTejun Heo if (list_empty(&pool->worklist)) 627282607adcSTejun Heo continue; 627382607adcSTejun Heo 6274940d71c6SSergey Senozhatsky /* 6275940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6276940d71c6SSergey Senozhatsky * the watchdog like a stall. 6277940d71c6SSergey Senozhatsky */ 6278940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6279940d71c6SSergey Senozhatsky 628082607adcSTejun Heo /* get the latest of pool and touched timestamps */ 628189e28ce6SWang Qing if (pool->cpu >= 0) 628289e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 628389e28ce6SWang Qing else 628482607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 628589e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 628682607adcSTejun Heo 628782607adcSTejun Heo if (time_after(pool_ts, touched)) 628882607adcSTejun Heo ts = pool_ts; 628982607adcSTejun Heo else 629082607adcSTejun Heo ts = touched; 629182607adcSTejun Heo 629282607adcSTejun Heo /* did we stall? */ 6293940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 629482607adcSTejun Heo lockup_detected = true; 6295cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6296cd2440d6SPetr Mladek pool->cpu_stall = true; 6297cd2440d6SPetr Mladek cpu_pool_stall = true; 6298cd2440d6SPetr Mladek } 629982607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 630082607adcSTejun Heo pr_cont_pool_info(pool); 630182607adcSTejun Heo pr_cont(" stuck for %us!\n", 6302940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 630382607adcSTejun Heo } 6304cd2440d6SPetr Mladek 6305cd2440d6SPetr Mladek 630682607adcSTejun Heo } 630782607adcSTejun Heo 630882607adcSTejun Heo rcu_read_unlock(); 630982607adcSTejun Heo 631082607adcSTejun Heo if (lockup_detected) 631155df0933SImran Khan show_all_workqueues(); 631282607adcSTejun Heo 6313cd2440d6SPetr Mladek if (cpu_pool_stall) 6314cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6315cd2440d6SPetr Mladek 631682607adcSTejun Heo wq_watchdog_reset_touched(); 631782607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 631882607adcSTejun Heo } 631982607adcSTejun Heo 6320cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 632182607adcSTejun Heo { 632282607adcSTejun Heo if (cpu >= 0) 632382607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 632489e28ce6SWang Qing 632582607adcSTejun Heo wq_watchdog_touched = jiffies; 632682607adcSTejun Heo } 632782607adcSTejun Heo 632882607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 632982607adcSTejun Heo { 633082607adcSTejun Heo wq_watchdog_thresh = 0; 633182607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 633282607adcSTejun Heo 633382607adcSTejun Heo if (thresh) { 633482607adcSTejun Heo wq_watchdog_thresh = thresh; 633582607adcSTejun Heo wq_watchdog_reset_touched(); 633682607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 633782607adcSTejun Heo } 633882607adcSTejun Heo } 633982607adcSTejun Heo 634082607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 634182607adcSTejun Heo const struct kernel_param *kp) 634282607adcSTejun Heo { 634382607adcSTejun Heo unsigned long thresh; 634482607adcSTejun Heo int ret; 634582607adcSTejun Heo 634682607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 634782607adcSTejun Heo if (ret) 634882607adcSTejun Heo return ret; 634982607adcSTejun Heo 635082607adcSTejun Heo if (system_wq) 635182607adcSTejun Heo wq_watchdog_set_thresh(thresh); 635282607adcSTejun Heo else 635382607adcSTejun Heo wq_watchdog_thresh = thresh; 635482607adcSTejun Heo 635582607adcSTejun Heo return 0; 635682607adcSTejun Heo } 635782607adcSTejun Heo 635882607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 635982607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 636082607adcSTejun Heo .get = param_get_ulong, 636182607adcSTejun Heo }; 636282607adcSTejun Heo 636382607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 636482607adcSTejun Heo 0644); 636582607adcSTejun Heo 636682607adcSTejun Heo static void wq_watchdog_init(void) 636782607adcSTejun Heo { 63685cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 636982607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 637082607adcSTejun Heo } 637182607adcSTejun Heo 637282607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 637382607adcSTejun Heo 637482607adcSTejun Heo static inline void wq_watchdog_init(void) { } 637582607adcSTejun Heo 637682607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 637782607adcSTejun Heo 6378bce90380STejun Heo static void __init wq_numa_init(void) 6379bce90380STejun Heo { 6380bce90380STejun Heo cpumask_var_t *tbl; 6381bce90380STejun Heo int node, cpu; 6382bce90380STejun Heo 6383bce90380STejun Heo if (num_possible_nodes() <= 1) 6384bce90380STejun Heo return; 6385bce90380STejun Heo 6386d55262c4STejun Heo if (wq_disable_numa) { 6387d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 6388d55262c4STejun Heo return; 6389d55262c4STejun Heo } 6390d55262c4STejun Heo 6391f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6392f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6393f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6394f728c4a9SZhen Lei return; 6395f728c4a9SZhen Lei } 6396f728c4a9SZhen Lei } 6397f728c4a9SZhen Lei 6398be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 63994c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 64004c16bd32STejun Heo 6401bce90380STejun Heo /* 6402bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6403bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6404bce90380STejun Heo * fully initialized by now. 6405bce90380STejun Heo */ 64066396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6407bce90380STejun Heo BUG_ON(!tbl); 6408bce90380STejun Heo 6409bce90380STejun Heo for_each_node(node) 64105a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 64111be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6412bce90380STejun Heo 6413bce90380STejun Heo for_each_possible_cpu(cpu) { 6414bce90380STejun Heo node = cpu_to_node(cpu); 6415bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6416bce90380STejun Heo } 6417bce90380STejun Heo 6418bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6419bce90380STejun Heo wq_numa_enabled = true; 6420bce90380STejun Heo } 6421bce90380STejun Heo 64223347fa09STejun Heo /** 64233347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 64243347fa09STejun Heo * 64253347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 64263347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 64273347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 64283347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 64293347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 64303347fa09STejun Heo * created and scheduled right before early initcalls. 64313347fa09STejun Heo */ 64322333e829SYu Chen void __init workqueue_init_early(void) 64331da177e4SLinus Torvalds { 64347a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 64357a4e344cSTejun Heo int i, cpu; 6436c34056a3STejun Heo 643710cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6438e904e6c2STejun Heo 6439b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 644004d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 644104d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6442b05a7928SFrederic Weisbecker 6443ace3c549Stiozhang if (!cpumask_empty(&wq_cmdline_cpumask)) 6444ace3c549Stiozhang cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask); 6445ace3c549Stiozhang 6446e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6447e904e6c2STejun Heo 6448706026c2STejun Heo /* initialize CPU pools */ 644929c91e99STejun Heo for_each_possible_cpu(cpu) { 64504ce62e9eSTejun Heo struct worker_pool *pool; 64518b03ae3cSTejun Heo 64527a4e344cSTejun Heo i = 0; 6453f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 64547a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6455ec22ca5eSTejun Heo pool->cpu = cpu; 64567a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 64577a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6458f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 64597a4e344cSTejun Heo 64609daf9e67STejun Heo /* alloc pool ID */ 646168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 64629daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 646368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 64644ce62e9eSTejun Heo } 64658b03ae3cSTejun Heo } 64668b03ae3cSTejun Heo 64678a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 646829c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 646929c91e99STejun Heo struct workqueue_attrs *attrs; 647029c91e99STejun Heo 6471be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 647229c91e99STejun Heo attrs->nice = std_nice[i]; 647329c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 64748a2b7538STejun Heo 64758a2b7538STejun Heo /* 64768a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 64778a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 64788a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 64798a2b7538STejun Heo */ 6480be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 64818a2b7538STejun Heo attrs->nice = std_nice[i]; 64828a2b7538STejun Heo attrs->no_numa = true; 64838a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 648429c91e99STejun Heo } 648529c91e99STejun Heo 6486d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 64871aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6488d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6489f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6490f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 649124d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 649224d51addSTejun Heo WQ_FREEZABLE, 0); 64930668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 64940668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 64950668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 64960668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 64970668106cSViresh Kumar 0); 64981aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 64990668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 65000668106cSViresh Kumar !system_power_efficient_wq || 65010668106cSViresh Kumar !system_freezable_power_efficient_wq); 65023347fa09STejun Heo } 65033347fa09STejun Heo 6504aa6fde93STejun Heo static void __init wq_cpu_intensive_thresh_init(void) 6505aa6fde93STejun Heo { 6506aa6fde93STejun Heo unsigned long thresh; 6507aa6fde93STejun Heo unsigned long bogo; 6508aa6fde93STejun Heo 6509aa6fde93STejun Heo /* if the user set it to a specific value, keep it */ 6510aa6fde93STejun Heo if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6511aa6fde93STejun Heo return; 6512aa6fde93STejun Heo 6513aa6fde93STejun Heo /* 6514aa6fde93STejun Heo * The default of 10ms is derived from the fact that most modern (as of 6515aa6fde93STejun Heo * 2023) processors can do a lot in 10ms and that it's just below what 6516aa6fde93STejun Heo * most consider human-perceivable. However, the kernel also runs on a 6517aa6fde93STejun Heo * lot slower CPUs including microcontrollers where the threshold is way 6518aa6fde93STejun Heo * too low. 6519aa6fde93STejun Heo * 6520aa6fde93STejun Heo * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6521aa6fde93STejun Heo * This is by no means accurate but it doesn't have to be. The mechanism 6522aa6fde93STejun Heo * is still useful even when the threshold is fully scaled up. Also, as 6523aa6fde93STejun Heo * the reports would usually be applicable to everyone, some machines 6524aa6fde93STejun Heo * operating on longer thresholds won't significantly diminish their 6525aa6fde93STejun Heo * usefulness. 6526aa6fde93STejun Heo */ 6527aa6fde93STejun Heo thresh = 10 * USEC_PER_MSEC; 6528aa6fde93STejun Heo 6529aa6fde93STejun Heo /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6530aa6fde93STejun Heo bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6531aa6fde93STejun Heo if (bogo < 4000) 6532aa6fde93STejun Heo thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6533aa6fde93STejun Heo 6534aa6fde93STejun Heo pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6535aa6fde93STejun Heo loops_per_jiffy, bogo, thresh); 6536aa6fde93STejun Heo 6537aa6fde93STejun Heo wq_cpu_intensive_thresh_us = thresh; 6538aa6fde93STejun Heo } 6539aa6fde93STejun Heo 65403347fa09STejun Heo /** 65413347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 65423347fa09STejun Heo * 65433347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 65443347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 65453347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 65463347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 65473347fa09STejun Heo * with the initial workers and enable future kworker creations. 65483347fa09STejun Heo */ 65492333e829SYu Chen void __init workqueue_init(void) 65503347fa09STejun Heo { 65512186d9f9STejun Heo struct workqueue_struct *wq; 65523347fa09STejun Heo struct worker_pool *pool; 65533347fa09STejun Heo int cpu, bkt; 65543347fa09STejun Heo 6555aa6fde93STejun Heo wq_cpu_intensive_thresh_init(); 6556aa6fde93STejun Heo 65572186d9f9STejun Heo /* 65582186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 65592186d9f9STejun Heo * CPU to node mapping may not be available that early on some 65602186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 65612186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 65622186d9f9STejun Heo * affinity, fix them up. 656340c17f75STejun Heo * 656440c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 65652186d9f9STejun Heo */ 65662186d9f9STejun Heo wq_numa_init(); 65672186d9f9STejun Heo 65682186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 65692186d9f9STejun Heo 65702186d9f9STejun Heo for_each_possible_cpu(cpu) { 65712186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65722186d9f9STejun Heo pool->node = cpu_to_node(cpu); 65732186d9f9STejun Heo } 65742186d9f9STejun Heo } 65752186d9f9STejun Heo 657640c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 65772186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 657840c17f75STejun Heo WARN(init_rescuer(wq), 657940c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 658040c17f75STejun Heo wq->name); 658140c17f75STejun Heo } 65822186d9f9STejun Heo 65832186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 65842186d9f9STejun Heo 65853347fa09STejun Heo /* create the initial workers */ 65863347fa09STejun Heo for_each_online_cpu(cpu) { 65873347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65883347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 65893347fa09STejun Heo BUG_ON(!create_worker(pool)); 65903347fa09STejun Heo } 65913347fa09STejun Heo } 65923347fa09STejun Heo 65933347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 65943347fa09STejun Heo BUG_ON(!create_worker(pool)); 65953347fa09STejun Heo 65963347fa09STejun Heo wq_online = true; 659782607adcSTejun Heo wq_watchdog_init(); 65981da177e4SLinus Torvalds } 6599c4f135d6STetsuo Handa 660020bdedafSTetsuo Handa void __warn_flushing_systemwide_wq(void) 660120bdedafSTetsuo Handa { 660220bdedafSTetsuo Handa pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 660320bdedafSTetsuo Handa dump_stack(); 660420bdedafSTetsuo Handa } 6605c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6606ace3c549Stiozhang 6607ace3c549Stiozhang static int __init workqueue_unbound_cpus_setup(char *str) 6608ace3c549Stiozhang { 6609ace3c549Stiozhang if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6610ace3c549Stiozhang cpumask_clear(&wq_cmdline_cpumask); 6611ace3c549Stiozhang pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6612ace3c549Stiozhang } 6613ace3c549Stiozhang 6614ace3c549Stiozhang return 1; 6615ace3c549Stiozhang } 6616ace3c549Stiozhang __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6617