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> 5262635ea8SSergey Senozhatsky #include <linux/nmi.h> 53940d71c6SSergey Senozhatsky #include <linux/kvm_para.h> 54e22bee78STejun Heo 55ea138446STejun Heo #include "workqueue_internal.h" 561da177e4SLinus Torvalds 57c8e55f36STejun Heo enum { 58bc2ae0f5STejun Heo /* 5924647570STejun Heo * worker_pool flags 60bc2ae0f5STejun Heo * 6124647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 62bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 63bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 64bc2ae0f5STejun Heo * is in effect. 65bc2ae0f5STejun Heo * 66bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 67bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 6824647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 69bc2ae0f5STejun Heo * 70bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 711258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 724736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 73bc2ae0f5STejun Heo */ 74692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7524647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 76db7bccf4STejun Heo 77c8e55f36STejun Heo /* worker flags */ 78c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 79c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 80e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 81fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 82f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 83a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 84e22bee78STejun Heo 85a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 86a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 87db7bccf4STejun Heo 88e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 894ce62e9eSTejun Heo 9029c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 91c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 92db7bccf4STejun Heo 93e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 94e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 95e22bee78STejun Heo 963233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 973233cdbdSTejun Heo /* call for help after 10ms 983233cdbdSTejun Heo (min two ticks) */ 99e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 100e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* 103e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1048698a745SDongsheng Yang * all cpus. Give MIN_NICE. 105e22bee78STejun Heo */ 1068698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1078698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 108ecf6881fSTejun Heo 109ecf6881fSTejun Heo WQ_NAME_LEN = 24, 110c8e55f36STejun Heo }; 111c8e55f36STejun Heo 1121da177e4SLinus Torvalds /* 1134690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1144690c4abSTejun Heo * 115e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 116e41e704bSTejun Heo * everyone else. 1174690c4abSTejun Heo * 118e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 119e22bee78STejun Heo * only be modified and accessed from the local cpu. 120e22bee78STejun Heo * 121d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1224690c4abSTejun Heo * 123d565ed63STejun Heo * X: During normal operation, modification requires pool->lock and should 124d565ed63STejun Heo * be done only from local cpu. Either disabling preemption on local 125d565ed63STejun Heo * cpu or grabbing pool->lock is enough for read access. If 126d565ed63STejun Heo * POOL_DISASSOCIATED is set, it's identical to L. 127e22bee78STejun Heo * 1281258fae7STejun Heo * A: wq_pool_attach_mutex protected. 129822d8405STejun Heo * 13068e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13176af4d93STejun Heo * 13224acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1335bcab335STejun Heo * 1345b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1355b95e1afSLai Jiangshan * 1365b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 13724acfb71SThomas Gleixner * RCU for reads. 1385b95e1afSLai Jiangshan * 1393c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1403c25a55dSLai Jiangshan * 14124acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1422e109a28STejun Heo * 1432e109a28STejun Heo * MD: wq_mayday_lock protected. 1444690c4abSTejun Heo */ 1454690c4abSTejun Heo 1462eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 147c34056a3STejun Heo 148bd7bdd43STejun Heo struct worker_pool { 149a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 150d84ff051STejun Heo int cpu; /* I: the associated cpu */ 151f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1529daf9e67STejun Heo int id; /* I: pool ID */ 15311ebea50STejun Heo unsigned int flags; /* X: flags */ 154bd7bdd43STejun Heo 15582607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 15682607adcSTejun Heo 157bc35f7efSLai Jiangshan /* 158bc35f7efSLai Jiangshan * The counter is incremented in a process context on the associated CPU 159bc35f7efSLai Jiangshan * w/ preemption disabled, and decremented or reset in the same context 160bc35f7efSLai Jiangshan * but w/ pool->lock held. The readers grab pool->lock and are 161bc35f7efSLai Jiangshan * guaranteed to see if the counter reached zero. 162bc35f7efSLai Jiangshan */ 163bc35f7efSLai Jiangshan int nr_running; 16484f91c62SLai Jiangshan 165bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 166ea1abd61SLai Jiangshan 1675826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1685826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 169bd7bdd43STejun Heo 1702c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 171bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 1723f959aa3SValentin Schneider struct work_struct idle_cull_work; /* L: worker idle cleanup */ 1733f959aa3SValentin Schneider 174bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 175bd7bdd43STejun Heo 176c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 177c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 178c9e7cf27STejun Heo /* L: hash of busy workers */ 179c9e7cf27STejun Heo 1802607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 18192f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 182*e02b9312SValentin Schneider struct list_head dying_workers; /* A: workers about to die */ 18360f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 184e19e397aSTejun Heo 1857cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 186e19e397aSTejun Heo 1877a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 18868e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 18968e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1907a4e344cSTejun Heo 191e19e397aSTejun Heo /* 19224acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 19329c91e99STejun Heo * from get_work_pool(). 19429c91e99STejun Heo */ 19529c91e99STejun Heo struct rcu_head rcu; 19684f91c62SLai Jiangshan }; 1978b03ae3cSTejun Heo 1988b03ae3cSTejun Heo /* 199112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 200112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 201112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 202112202d9STejun Heo * number of flag bits. 2031da177e4SLinus Torvalds */ 204112202d9STejun Heo struct pool_workqueue { 205bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2064690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20773f53c4aSTejun Heo int work_color; /* L: current color */ 20873f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2098864b4e5STejun Heo int refcnt; /* L: reference count */ 21073f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 21173f53c4aSTejun Heo /* L: nr of in_flight works */ 212018f3a13SLai Jiangshan 213018f3a13SLai Jiangshan /* 214018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 215018f3a13SLai Jiangshan * 216018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 217018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 218018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 219018f3a13SLai Jiangshan * 220018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 221018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 222018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 223018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 224018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 225018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 226018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 227018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 228018f3a13SLai Jiangshan */ 2291e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 230a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 231f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2323c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2332e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2348864b4e5STejun Heo 2358864b4e5STejun Heo /* 2368864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2378864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 23824acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 239b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2408864b4e5STejun Heo */ 2418864b4e5STejun Heo struct work_struct unbound_release_work; 2428864b4e5STejun Heo struct rcu_head rcu; 243e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2441da177e4SLinus Torvalds 2451da177e4SLinus Torvalds /* 24673f53c4aSTejun Heo * Structure used to wait for workqueue flush. 24773f53c4aSTejun Heo */ 24873f53c4aSTejun Heo struct wq_flusher { 2493c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2503c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 25173f53c4aSTejun Heo struct completion done; /* flush completion */ 25273f53c4aSTejun Heo }; 2531da177e4SLinus Torvalds 254226223abSTejun Heo struct wq_device; 255226223abSTejun Heo 25673f53c4aSTejun Heo /* 257c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 258c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2591da177e4SLinus Torvalds */ 2601da177e4SLinus Torvalds struct workqueue_struct { 2613c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 262e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 26373f53c4aSTejun Heo 2643c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2653c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2663c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 267112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2683c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2693c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2703c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 27173f53c4aSTejun Heo 2722e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 27330ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 274e22bee78STejun Heo 27587fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 276a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 277226223abSTejun Heo 2785b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2795b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2806029a918STejun Heo 281226223abSTejun Heo #ifdef CONFIG_SYSFS 282226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 283226223abSTejun Heo #endif 2844e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 285669de8bdSBart Van Assche char *lock_name; 286669de8bdSBart Van Assche struct lock_class_key key; 2874e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2884e6045f1SJohannes Berg #endif 289ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2902728fd2fSTejun Heo 291e2dca7adSTejun Heo /* 29224acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 29324acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 294e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 295e2dca7adSTejun Heo */ 296e2dca7adSTejun Heo struct rcu_head rcu; 297e2dca7adSTejun Heo 2982728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2992728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 3002728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 3015b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 3021da177e4SLinus Torvalds }; 3031da177e4SLinus Torvalds 304e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 305e904e6c2STejun Heo 306bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 307bce90380STejun Heo /* possible CPUs of each node */ 308bce90380STejun Heo 309d55262c4STejun Heo static bool wq_disable_numa; 310d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 311d55262c4STejun Heo 312cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 313552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 314cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 315cee22a15SViresh Kumar 316863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3173347fa09STejun Heo 318bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 319bce90380STejun Heo 3204c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3214c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3224c16bd32STejun Heo 32368e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3241258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 325a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 326d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 327d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3285bcab335STejun Heo 329e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 33068e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3317d19c5ceSTejun Heo 33299c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 333ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 334ef557180SMike Galbraith 335ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 336ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 337b05a7928SFrederic Weisbecker 338f303fccbSTejun Heo /* 339f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 340f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 341f303fccbSTejun Heo * to uncover usages which depend on it. 342f303fccbSTejun Heo */ 343f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 344f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 345f303fccbSTejun Heo #else 346f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 347f303fccbSTejun Heo #endif 348f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 349f303fccbSTejun Heo 3507d19c5ceSTejun Heo /* the per-cpu worker pools */ 35125528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3527d19c5ceSTejun Heo 35368e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3547d19c5ceSTejun Heo 35568e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 35629c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 35729c91e99STejun Heo 358c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 35929c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 36029c91e99STejun Heo 3618a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3628a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3638a2b7538STejun Heo 364d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 365ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 366044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3671aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 368044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 369d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 370044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 371f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 372044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 37324d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3740668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3750668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3760668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3770668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 378d320c038STejun Heo 3797d19c5ceSTejun Heo static int worker_thread(void *__worker); 3806ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 381c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 38255df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 3837d19c5ceSTejun Heo 38497bd2347STejun Heo #define CREATE_TRACE_POINTS 38597bd2347STejun Heo #include <trace/events/workqueue.h> 38697bd2347STejun Heo 38768e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 38824acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 389f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 39024acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 3915bcab335STejun Heo 3925b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 39324acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 394f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 395f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 39624acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 3975b95e1afSLai Jiangshan 398f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 399f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 400f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4017a62c2c8STejun Heo (pool)++) 4024ce62e9eSTejun Heo 40349e3cf44STejun Heo /** 40417116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 40517116969STejun Heo * @pool: iteration cursor 406611c92a0STejun Heo * @pi: integer used for iteration 407fa1b54e6STejun Heo * 40824acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 40968e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 41068e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 411fa1b54e6STejun Heo * 412fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 413fa1b54e6STejun Heo * ignored. 41417116969STejun Heo */ 415611c92a0STejun Heo #define for_each_pool(pool, pi) \ 416611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 41768e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 418fa1b54e6STejun Heo else 41917116969STejun Heo 42017116969STejun Heo /** 421822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 422822d8405STejun Heo * @worker: iteration cursor 423822d8405STejun Heo * @pool: worker_pool to iterate workers of 424822d8405STejun Heo * 4251258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 426822d8405STejun Heo * 427822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 428822d8405STejun Heo * ignored. 429822d8405STejun Heo */ 430da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 431da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4321258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 433822d8405STejun Heo else 434822d8405STejun Heo 435822d8405STejun Heo /** 43649e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 43749e3cf44STejun Heo * @pwq: iteration cursor 43849e3cf44STejun Heo * @wq: the target workqueue 43976af4d93STejun Heo * 44024acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 441794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 442794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 44376af4d93STejun Heo * 44476af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 44576af4d93STejun Heo * ignored. 44649e3cf44STejun Heo */ 44749e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 44849e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4495a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 450f3421797STejun Heo 451dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 452dc186ad7SThomas Gleixner 453f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 454dc186ad7SThomas Gleixner 45599777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 45699777288SStanislaw Gruszka { 45799777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 45899777288SStanislaw Gruszka } 45999777288SStanislaw Gruszka 460b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 461b9fdac7fSDu, Changbin { 462b9fdac7fSDu, Changbin struct work_struct *work = addr; 463b9fdac7fSDu, Changbin 464b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 465b9fdac7fSDu, Changbin } 466b9fdac7fSDu, Changbin 467dc186ad7SThomas Gleixner /* 468dc186ad7SThomas Gleixner * fixup_init is called when: 469dc186ad7SThomas Gleixner * - an active object is initialized 470dc186ad7SThomas Gleixner */ 47102a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 472dc186ad7SThomas Gleixner { 473dc186ad7SThomas Gleixner struct work_struct *work = addr; 474dc186ad7SThomas Gleixner 475dc186ad7SThomas Gleixner switch (state) { 476dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 477dc186ad7SThomas Gleixner cancel_work_sync(work); 478dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 47902a982a6SDu, Changbin return true; 480dc186ad7SThomas Gleixner default: 48102a982a6SDu, Changbin return false; 482dc186ad7SThomas Gleixner } 483dc186ad7SThomas Gleixner } 484dc186ad7SThomas Gleixner 485dc186ad7SThomas Gleixner /* 486dc186ad7SThomas Gleixner * fixup_free is called when: 487dc186ad7SThomas Gleixner * - an active object is freed 488dc186ad7SThomas Gleixner */ 48902a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 490dc186ad7SThomas Gleixner { 491dc186ad7SThomas Gleixner struct work_struct *work = addr; 492dc186ad7SThomas Gleixner 493dc186ad7SThomas Gleixner switch (state) { 494dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 495dc186ad7SThomas Gleixner cancel_work_sync(work); 496dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 49702a982a6SDu, Changbin return true; 498dc186ad7SThomas Gleixner default: 49902a982a6SDu, Changbin return false; 500dc186ad7SThomas Gleixner } 501dc186ad7SThomas Gleixner } 502dc186ad7SThomas Gleixner 503f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 504dc186ad7SThomas Gleixner .name = "work_struct", 50599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 506b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 507dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 508dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 509dc186ad7SThomas Gleixner }; 510dc186ad7SThomas Gleixner 511dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 512dc186ad7SThomas Gleixner { 513dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 514dc186ad7SThomas Gleixner } 515dc186ad7SThomas Gleixner 516dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 517dc186ad7SThomas Gleixner { 518dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 519dc186ad7SThomas Gleixner } 520dc186ad7SThomas Gleixner 521dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 522dc186ad7SThomas Gleixner { 523dc186ad7SThomas Gleixner if (onstack) 524dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 525dc186ad7SThomas Gleixner else 526dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 527dc186ad7SThomas Gleixner } 528dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 529dc186ad7SThomas Gleixner 530dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 531dc186ad7SThomas Gleixner { 532dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 533dc186ad7SThomas Gleixner } 534dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 535dc186ad7SThomas Gleixner 536ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 537ea2e64f2SThomas Gleixner { 538ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 539ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 540ea2e64f2SThomas Gleixner } 541ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 542ea2e64f2SThomas Gleixner 543dc186ad7SThomas Gleixner #else 544dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 545dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 546dc186ad7SThomas Gleixner #endif 547dc186ad7SThomas Gleixner 5484e8b22bdSLi Bin /** 54967dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5504e8b22bdSLi Bin * @pool: the pool pointer of interest 5514e8b22bdSLi Bin * 5524e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5534e8b22bdSLi Bin * successfully, -errno on failure. 5544e8b22bdSLi Bin */ 5559daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5569daf9e67STejun Heo { 5579daf9e67STejun Heo int ret; 5589daf9e67STejun Heo 55968e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5605bcab335STejun Heo 5614e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5624e8b22bdSLi Bin GFP_KERNEL); 563229641a6STejun Heo if (ret >= 0) { 564e68035fbSTejun Heo pool->id = ret; 565229641a6STejun Heo return 0; 566229641a6STejun Heo } 5679daf9e67STejun Heo return ret; 5689daf9e67STejun Heo } 5699daf9e67STejun Heo 57076af4d93STejun Heo /** 571df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 572df2d5ae4STejun Heo * @wq: the target workqueue 573df2d5ae4STejun Heo * @node: the node ID 574df2d5ae4STejun Heo * 57524acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5765b95e1afSLai Jiangshan * read locked. 577df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 578df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 579d185af30SYacine Belkadi * 580d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 581df2d5ae4STejun Heo */ 582df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 583df2d5ae4STejun Heo int node) 584df2d5ae4STejun Heo { 5855b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 586d6e022f1STejun Heo 587d6e022f1STejun Heo /* 588d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 589d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 590d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 591d6e022f1STejun Heo * happens, this workaround can be removed. 592d6e022f1STejun Heo */ 593d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 594d6e022f1STejun Heo return wq->dfl_pwq; 595d6e022f1STejun Heo 596df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 597df2d5ae4STejun Heo } 598df2d5ae4STejun Heo 59973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 60073f53c4aSTejun Heo { 60173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 60273f53c4aSTejun Heo } 60373f53c4aSTejun Heo 604c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 60573f53c4aSTejun Heo { 606c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 60773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 60873f53c4aSTejun Heo } 60973f53c4aSTejun Heo 61073f53c4aSTejun Heo static int work_next_color(int color) 61173f53c4aSTejun Heo { 61273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 613a848e3b6SOleg Nesterov } 614a848e3b6SOleg Nesterov 6154594bf15SDavid Howells /* 616112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 617112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6187c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6197a22ad75STejun Heo * 620112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 621112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 622bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 623bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6247a22ad75STejun Heo * 625112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6267c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 627112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6287c3eed5cSTejun Heo * available only while the work item is queued. 629bbb68dfaSTejun Heo * 630bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 631bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 632bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 633bbb68dfaSTejun Heo * try to steal the PENDING bit. 6344594bf15SDavid Howells */ 6357a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6367a22ad75STejun Heo unsigned long flags) 6377a22ad75STejun Heo { 6386183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6397a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6407a22ad75STejun Heo } 6417a22ad75STejun Heo 642112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6434690c4abSTejun Heo unsigned long extra_flags) 644365970a1SDavid Howells { 645112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 646112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 647365970a1SDavid Howells } 648365970a1SDavid Howells 6494468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6504468a00fSLai Jiangshan int pool_id) 6514468a00fSLai Jiangshan { 6524468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6534468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6544468a00fSLai Jiangshan } 6554468a00fSLai Jiangshan 6567c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6577c3eed5cSTejun Heo int pool_id) 6584d707b9fSOleg Nesterov { 65923657bb1STejun Heo /* 66023657bb1STejun Heo * The following wmb is paired with the implied mb in 66123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 66223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 66323657bb1STejun Heo * owner. 66423657bb1STejun Heo */ 66523657bb1STejun Heo smp_wmb(); 6667c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 667346c09f8SRoman Pen /* 668346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 669346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 670346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6718bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 672346c09f8SRoman Pen * the same @work. E.g. consider this case: 673346c09f8SRoman Pen * 674346c09f8SRoman Pen * CPU#0 CPU#1 675346c09f8SRoman Pen * ---------------------------- -------------------------------- 676346c09f8SRoman Pen * 677346c09f8SRoman Pen * 1 STORE event_indicated 678346c09f8SRoman Pen * 2 queue_work_on() { 679346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 680346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 681346c09f8SRoman Pen * 5 set_work_data() # clear bit 682346c09f8SRoman Pen * 6 smp_mb() 683346c09f8SRoman Pen * 7 work->current_func() { 684346c09f8SRoman Pen * 8 LOAD event_indicated 685346c09f8SRoman Pen * } 686346c09f8SRoman Pen * 687346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 688346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 689346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 690346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 691346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 692346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 693346c09f8SRoman Pen * before actual STORE. 694346c09f8SRoman Pen */ 695346c09f8SRoman Pen smp_mb(); 6964d707b9fSOleg Nesterov } 6974d707b9fSOleg Nesterov 6987a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 699365970a1SDavid Howells { 7007c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7017c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7027a22ad75STejun Heo } 7037a22ad75STejun Heo 704112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7057a22ad75STejun Heo { 706e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7077a22ad75STejun Heo 708112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 709e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 710e120153dSTejun Heo else 711e120153dSTejun Heo return NULL; 7127a22ad75STejun Heo } 7137a22ad75STejun Heo 7147c3eed5cSTejun Heo /** 7157c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7167c3eed5cSTejun Heo * @work: the work item of interest 7177c3eed5cSTejun Heo * 71868e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 71924acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 72024acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 721fa1b54e6STejun Heo * 722fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 723fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 724fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 725fa1b54e6STejun Heo * returned pool is and stays online. 726d185af30SYacine Belkadi * 727d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7287c3eed5cSTejun Heo */ 7297c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7307a22ad75STejun Heo { 731e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7327c3eed5cSTejun Heo int pool_id; 7337a22ad75STejun Heo 73468e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 735fa1b54e6STejun Heo 736112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 737112202d9STejun Heo return ((struct pool_workqueue *) 7387c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7397a22ad75STejun Heo 7407c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7417c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7427a22ad75STejun Heo return NULL; 7437a22ad75STejun Heo 744fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7457c3eed5cSTejun Heo } 7467c3eed5cSTejun Heo 7477c3eed5cSTejun Heo /** 7487c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7497c3eed5cSTejun Heo * @work: the work item of interest 7507c3eed5cSTejun Heo * 751d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7527c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7537c3eed5cSTejun Heo */ 7547c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7557c3eed5cSTejun Heo { 75654d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7577c3eed5cSTejun Heo 758112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 759112202d9STejun Heo return ((struct pool_workqueue *) 76054d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 76154d5b7d0SLai Jiangshan 76254d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7637c3eed5cSTejun Heo } 7647c3eed5cSTejun Heo 765bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 766bbb68dfaSTejun Heo { 7677c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 768bbb68dfaSTejun Heo 7697c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7707c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 771bbb68dfaSTejun Heo } 772bbb68dfaSTejun Heo 773bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 774bbb68dfaSTejun Heo { 775bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 776bbb68dfaSTejun Heo 777112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 778bbb68dfaSTejun Heo } 779bbb68dfaSTejun Heo 780e22bee78STejun Heo /* 7813270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7823270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 783d565ed63STejun Heo * they're being called with pool->lock held. 784e22bee78STejun Heo */ 785e22bee78STejun Heo 78663d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 787649027d7STejun Heo { 788bc35f7efSLai Jiangshan return !pool->nr_running; 789649027d7STejun Heo } 790649027d7STejun Heo 791e22bee78STejun Heo /* 792e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 793e22bee78STejun Heo * running workers. 794974271c4STejun Heo * 795974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 796706026c2STejun Heo * function will always return %true for unbound pools as long as the 797974271c4STejun Heo * worklist isn't empty. 798e22bee78STejun Heo */ 79963d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 800e22bee78STejun Heo { 80163d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 802e22bee78STejun Heo } 803e22bee78STejun Heo 804e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 80563d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 806e22bee78STejun Heo { 80763d95a91STejun Heo return pool->nr_idle; 808e22bee78STejun Heo } 809e22bee78STejun Heo 810e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 81163d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 812e22bee78STejun Heo { 813bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 814e22bee78STejun Heo } 815e22bee78STejun Heo 816e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 81763d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 818e22bee78STejun Heo { 81963d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 820e22bee78STejun Heo } 821e22bee78STejun Heo 822e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 82363d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 824e22bee78STejun Heo { 825692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 82663d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 82763d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 828e22bee78STejun Heo 829e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 830e22bee78STejun Heo } 831e22bee78STejun Heo 832e22bee78STejun Heo /* 833e22bee78STejun Heo * Wake up functions. 834e22bee78STejun Heo */ 835e22bee78STejun Heo 8362c1f1a91SLai Jiangshan /* Return the first idle worker. Called with pool->lock held. */ 8371037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8387e11629dSTejun Heo { 83963d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8407e11629dSTejun Heo return NULL; 8417e11629dSTejun Heo 84263d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8437e11629dSTejun Heo } 8447e11629dSTejun Heo 8457e11629dSTejun Heo /** 8467e11629dSTejun Heo * wake_up_worker - wake up an idle worker 84763d95a91STejun Heo * @pool: worker pool to wake worker from 8487e11629dSTejun Heo * 84963d95a91STejun Heo * Wake up the first idle worker of @pool. 8507e11629dSTejun Heo * 8517e11629dSTejun Heo * CONTEXT: 852a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8537e11629dSTejun Heo */ 85463d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8557e11629dSTejun Heo { 8561037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8577e11629dSTejun Heo 8587e11629dSTejun Heo if (likely(worker)) 8597e11629dSTejun Heo wake_up_process(worker->task); 8607e11629dSTejun Heo } 8617e11629dSTejun Heo 8624690c4abSTejun Heo /** 8636d25be57SThomas Gleixner * wq_worker_running - a worker is running again 864e22bee78STejun Heo * @task: task waking up 865e22bee78STejun Heo * 8666d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 867e22bee78STejun Heo */ 8686d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 869e22bee78STejun Heo { 870e22bee78STejun Heo struct worker *worker = kthread_data(task); 871e22bee78STejun Heo 8726d25be57SThomas Gleixner if (!worker->sleeping) 8736d25be57SThomas Gleixner return; 87407edfeceSFrederic Weisbecker 87507edfeceSFrederic Weisbecker /* 87607edfeceSFrederic Weisbecker * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 87707edfeceSFrederic Weisbecker * and the nr_running increment below, we may ruin the nr_running reset 87807edfeceSFrederic Weisbecker * and leave with an unexpected pool->nr_running == 1 on the newly unbound 87907edfeceSFrederic Weisbecker * pool. Protect against such race. 88007edfeceSFrederic Weisbecker */ 88107edfeceSFrederic Weisbecker preempt_disable(); 8826d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 883bc35f7efSLai Jiangshan worker->pool->nr_running++; 88407edfeceSFrederic Weisbecker preempt_enable(); 8856d25be57SThomas Gleixner worker->sleeping = 0; 88636576000SJoonsoo Kim } 887e22bee78STejun Heo 888e22bee78STejun Heo /** 889e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 890e22bee78STejun Heo * @task: task going to sleep 891e22bee78STejun Heo * 8926d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 893ccf45156SLai Jiangshan * going to sleep. 894e22bee78STejun Heo */ 8956d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 896e22bee78STejun Heo { 897cc5bff38SLai Jiangshan struct worker *worker = kthread_data(task); 898111c225aSTejun Heo struct worker_pool *pool; 899e22bee78STejun Heo 900111c225aSTejun Heo /* 901111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 902111c225aSTejun Heo * workers, also reach here, let's not access anything before 903111c225aSTejun Heo * checking NOT_RUNNING. 904111c225aSTejun Heo */ 9052d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 9066d25be57SThomas Gleixner return; 907e22bee78STejun Heo 908111c225aSTejun Heo pool = worker->pool; 909111c225aSTejun Heo 91062849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 91162849a96SSebastian Andrzej Siewior if (worker->sleeping) 9126d25be57SThomas Gleixner return; 9136d25be57SThomas Gleixner 9146d25be57SThomas Gleixner worker->sleeping = 1; 915a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 916e22bee78STejun Heo 917e22bee78STejun Heo /* 91845c753f5SFrederic Weisbecker * Recheck in case unbind_workers() preempted us. We don't 91945c753f5SFrederic Weisbecker * want to decrement nr_running after the worker is unbound 92045c753f5SFrederic Weisbecker * and nr_running has been reset. 92145c753f5SFrederic Weisbecker */ 92245c753f5SFrederic Weisbecker if (worker->flags & WORKER_NOT_RUNNING) { 92345c753f5SFrederic Weisbecker raw_spin_unlock_irq(&pool->lock); 92445c753f5SFrederic Weisbecker return; 92545c753f5SFrederic Weisbecker } 92645c753f5SFrederic Weisbecker 927bc35f7efSLai Jiangshan pool->nr_running--; 928bc35f7efSLai Jiangshan if (need_more_worker(pool)) 929cc5bff38SLai Jiangshan wake_up_worker(pool); 930a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 931e22bee78STejun Heo } 932e22bee78STejun Heo 933e22bee78STejun Heo /** 9341b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9358194fe94SBart Van Assche * @task: Task to retrieve last work function of. 9361b69ac6bSJohannes Weiner * 9371b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9381b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9391b69ac6bSJohannes Weiner * 9401b69ac6bSJohannes Weiner * CONTEXT: 941a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 9421b69ac6bSJohannes Weiner * 9434b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9444b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9454b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9464b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9474b047002SJohannes Weiner * 9484b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9494b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9504b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9514b047002SJohannes Weiner * is guaranteed to not be processing any works. 9524b047002SJohannes Weiner * 9531b69ac6bSJohannes Weiner * Return: 9541b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9551b69ac6bSJohannes Weiner * hasn't executed any work yet. 9561b69ac6bSJohannes Weiner */ 9571b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9581b69ac6bSJohannes Weiner { 9591b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9601b69ac6bSJohannes Weiner 9611b69ac6bSJohannes Weiner return worker->last_func; 9621b69ac6bSJohannes Weiner } 9631b69ac6bSJohannes Weiner 9641b69ac6bSJohannes Weiner /** 965e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 966cb444766STejun Heo * @worker: self 967d302f017STejun Heo * @flags: flags to set 968d302f017STejun Heo * 969228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 970d302f017STejun Heo * 971cb444766STejun Heo * CONTEXT: 972a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 973d302f017STejun Heo */ 974228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 975d302f017STejun Heo { 976bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 977e22bee78STejun Heo 978cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 979cb444766STejun Heo 980228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 981e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 982e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 983bc35f7efSLai Jiangshan pool->nr_running--; 984e22bee78STejun Heo } 985e22bee78STejun Heo 986d302f017STejun Heo worker->flags |= flags; 987d302f017STejun Heo } 988d302f017STejun Heo 989d302f017STejun Heo /** 990e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 991cb444766STejun Heo * @worker: self 992d302f017STejun Heo * @flags: flags to clear 993d302f017STejun Heo * 994e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 995d302f017STejun Heo * 996cb444766STejun Heo * CONTEXT: 997a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 998d302f017STejun Heo */ 999d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 1000d302f017STejun Heo { 100163d95a91STejun Heo struct worker_pool *pool = worker->pool; 1002e22bee78STejun Heo unsigned int oflags = worker->flags; 1003e22bee78STejun Heo 1004cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 1005cb444766STejun Heo 1006d302f017STejun Heo worker->flags &= ~flags; 1007e22bee78STejun Heo 100842c025f3STejun Heo /* 100942c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 101042c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 101142c025f3STejun Heo * of multiple flags, not a single flag. 101242c025f3STejun Heo */ 1013e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1014e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 1015bc35f7efSLai Jiangshan pool->nr_running++; 1016d302f017STejun Heo } 1017d302f017STejun Heo 1018d302f017STejun Heo /** 10198cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1020c9e7cf27STejun Heo * @pool: pool of interest 10218cca0eeaSTejun Heo * @work: work to find worker for 10228cca0eeaSTejun Heo * 1023c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1024c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1025a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1026a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1027a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1028a2c1c57bSTejun Heo * being executed. 1029a2c1c57bSTejun Heo * 1030a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1031a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1032a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1033a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1034a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1035a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1036a2c1c57bSTejun Heo * 1037c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1038c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1039c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1040c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1041c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1042c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10438cca0eeaSTejun Heo * 10448cca0eeaSTejun Heo * CONTEXT: 1045a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10468cca0eeaSTejun Heo * 1047d185af30SYacine Belkadi * Return: 1048d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10498cca0eeaSTejun Heo * otherwise. 10508cca0eeaSTejun Heo */ 1051c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10528cca0eeaSTejun Heo struct work_struct *work) 10538cca0eeaSTejun Heo { 105442f8570fSSasha Levin struct worker *worker; 105542f8570fSSasha Levin 1056b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1057a2c1c57bSTejun Heo (unsigned long)work) 1058a2c1c57bSTejun Heo if (worker->current_work == work && 1059a2c1c57bSTejun Heo worker->current_func == work->func) 106042f8570fSSasha Levin return worker; 106142f8570fSSasha Levin 106242f8570fSSasha Levin return NULL; 10638cca0eeaSTejun Heo } 10648cca0eeaSTejun Heo 10658cca0eeaSTejun Heo /** 1066bf4ede01STejun Heo * move_linked_works - move linked works to a list 1067bf4ede01STejun Heo * @work: start of series of works to be scheduled 1068bf4ede01STejun Heo * @head: target list to append @work to 1069402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1070bf4ede01STejun Heo * 1071bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1072bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1073bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1074bf4ede01STejun Heo * 1075bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1076bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1077bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1078bf4ede01STejun Heo * 1079bf4ede01STejun Heo * CONTEXT: 1080a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1081bf4ede01STejun Heo */ 1082bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1083bf4ede01STejun Heo struct work_struct **nextp) 1084bf4ede01STejun Heo { 1085bf4ede01STejun Heo struct work_struct *n; 1086bf4ede01STejun Heo 1087bf4ede01STejun Heo /* 1088bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1089bf4ede01STejun Heo * use NULL for list head. 1090bf4ede01STejun Heo */ 1091bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1092bf4ede01STejun Heo list_move_tail(&work->entry, head); 1093bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1094bf4ede01STejun Heo break; 1095bf4ede01STejun Heo } 1096bf4ede01STejun Heo 1097bf4ede01STejun Heo /* 1098bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1099bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1100bf4ede01STejun Heo * needs to be updated. 1101bf4ede01STejun Heo */ 1102bf4ede01STejun Heo if (nextp) 1103bf4ede01STejun Heo *nextp = n; 1104bf4ede01STejun Heo } 1105bf4ede01STejun Heo 11068864b4e5STejun Heo /** 11078864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 11088864b4e5STejun Heo * @pwq: pool_workqueue to get 11098864b4e5STejun Heo * 11108864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 11118864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 11128864b4e5STejun Heo */ 11138864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 11148864b4e5STejun Heo { 11158864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11168864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 11178864b4e5STejun Heo pwq->refcnt++; 11188864b4e5STejun Heo } 11198864b4e5STejun Heo 11208864b4e5STejun Heo /** 11218864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 11228864b4e5STejun Heo * @pwq: pool_workqueue to put 11238864b4e5STejun Heo * 11248864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11258864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11268864b4e5STejun Heo */ 11278864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11288864b4e5STejun Heo { 11298864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11308864b4e5STejun Heo if (likely(--pwq->refcnt)) 11318864b4e5STejun Heo return; 11328864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11338864b4e5STejun Heo return; 11348864b4e5STejun Heo /* 11358864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11368864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11378864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11388864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11398864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11408864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11418864b4e5STejun Heo */ 11428864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11438864b4e5STejun Heo } 11448864b4e5STejun Heo 1145dce90d47STejun Heo /** 1146dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1147dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1148dce90d47STejun Heo * 1149dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1150dce90d47STejun Heo */ 1151dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1152dce90d47STejun Heo { 1153dce90d47STejun Heo if (pwq) { 1154dce90d47STejun Heo /* 115524acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1156dce90d47STejun Heo * following lock operations are safe. 1157dce90d47STejun Heo */ 1158a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1159dce90d47STejun Heo put_pwq(pwq); 1160a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1161dce90d47STejun Heo } 1162dce90d47STejun Heo } 1163dce90d47STejun Heo 1164f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1165bf4ede01STejun Heo { 1166112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1167bf4ede01STejun Heo 1168bf4ede01STejun Heo trace_workqueue_activate_work(work); 116982607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 117082607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1171112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1172f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1173112202d9STejun Heo pwq->nr_active++; 1174bf4ede01STejun Heo } 1175bf4ede01STejun Heo 1176f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 11773aa62497SLai Jiangshan { 1178f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 11793aa62497SLai Jiangshan struct work_struct, entry); 11803aa62497SLai Jiangshan 1181f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 11823aa62497SLai Jiangshan } 11833aa62497SLai Jiangshan 1184bf4ede01STejun Heo /** 1185112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1186112202d9STejun Heo * @pwq: pwq of interest 1187c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1188bf4ede01STejun Heo * 1189bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1190112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1191bf4ede01STejun Heo * 1192bf4ede01STejun Heo * CONTEXT: 1193a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1194bf4ede01STejun Heo */ 1195c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1196bf4ede01STejun Heo { 1197c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1198c4560c2cSLai Jiangshan 1199018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1200112202d9STejun Heo pwq->nr_active--; 1201f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1202f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1203112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1204f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1205bf4ede01STejun Heo } 1206018f3a13SLai Jiangshan } 1207018f3a13SLai Jiangshan 1208018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1209bf4ede01STejun Heo 1210bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1211112202d9STejun Heo if (likely(pwq->flush_color != color)) 12128864b4e5STejun Heo goto out_put; 1213bf4ede01STejun Heo 1214bf4ede01STejun Heo /* are there still in-flight works? */ 1215112202d9STejun Heo if (pwq->nr_in_flight[color]) 12168864b4e5STejun Heo goto out_put; 1217bf4ede01STejun Heo 1218112202d9STejun Heo /* this pwq is done, clear flush_color */ 1219112202d9STejun Heo pwq->flush_color = -1; 1220bf4ede01STejun Heo 1221bf4ede01STejun Heo /* 1222112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1223bf4ede01STejun Heo * will handle the rest. 1224bf4ede01STejun Heo */ 1225112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1226112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12278864b4e5STejun Heo out_put: 12288864b4e5STejun Heo put_pwq(pwq); 1229bf4ede01STejun Heo } 1230bf4ede01STejun Heo 123136e227d2STejun Heo /** 1232bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 123336e227d2STejun Heo * @work: work item to steal 123436e227d2STejun Heo * @is_dwork: @work is a delayed_work 1235bbb68dfaSTejun Heo * @flags: place to store irq state 123636e227d2STejun Heo * 123736e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1238d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 123936e227d2STejun Heo * 1240d185af30SYacine Belkadi * Return: 12413eb6b31bSMauro Carvalho Chehab * 12423eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 124336e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 124436e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 124536e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1246bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1247bbb68dfaSTejun Heo * for arbitrarily long 12483eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 124936e227d2STejun Heo * 1250d185af30SYacine Belkadi * Note: 1251bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1252e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1253e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1254e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1255bbb68dfaSTejun Heo * 1256bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1257bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1258bbb68dfaSTejun Heo * 1259e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1260bf4ede01STejun Heo */ 1261bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1262bbb68dfaSTejun Heo unsigned long *flags) 1263bf4ede01STejun Heo { 1264d565ed63STejun Heo struct worker_pool *pool; 1265112202d9STejun Heo struct pool_workqueue *pwq; 1266bf4ede01STejun Heo 1267bbb68dfaSTejun Heo local_irq_save(*flags); 1268bbb68dfaSTejun Heo 126936e227d2STejun Heo /* try to steal the timer if it exists */ 127036e227d2STejun Heo if (is_dwork) { 127136e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 127236e227d2STejun Heo 1273e0aecdd8STejun Heo /* 1274e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1275e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1276e0aecdd8STejun Heo * running on the local CPU. 1277e0aecdd8STejun Heo */ 127836e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 127936e227d2STejun Heo return 1; 128036e227d2STejun Heo } 128136e227d2STejun Heo 128236e227d2STejun Heo /* try to claim PENDING the normal way */ 1283bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1284bf4ede01STejun Heo return 0; 1285bf4ede01STejun Heo 128624acfb71SThomas Gleixner rcu_read_lock(); 1287bf4ede01STejun Heo /* 1288bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1289bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1290bf4ede01STejun Heo */ 1291d565ed63STejun Heo pool = get_work_pool(work); 1292d565ed63STejun Heo if (!pool) 1293bbb68dfaSTejun Heo goto fail; 1294bf4ede01STejun Heo 1295a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1296bf4ede01STejun Heo /* 1297112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1298112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1299112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1300112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1301112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 13020b3dae68SLai Jiangshan * item is currently queued on that pool. 1303bf4ede01STejun Heo */ 1304112202d9STejun Heo pwq = get_work_pwq(work); 1305112202d9STejun Heo if (pwq && pwq->pool == pool) { 1306bf4ede01STejun Heo debug_work_deactivate(work); 13073aa62497SLai Jiangshan 13083aa62497SLai Jiangshan /* 1309018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1310018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1311018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1312018f3a13SLai Jiangshan * 1313f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1314d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1315f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 131616062836STejun Heo * management later on and cause stall. Make sure the work 131716062836STejun Heo * item is activated before grabbing. 13183aa62497SLai Jiangshan */ 1319f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1320f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13213aa62497SLai Jiangshan 1322bf4ede01STejun Heo list_del_init(&work->entry); 1323c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 132436e227d2STejun Heo 1325112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 13264468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 13274468a00fSLai Jiangshan 1328a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 132924acfb71SThomas Gleixner rcu_read_unlock(); 133036e227d2STejun Heo return 1; 1331bf4ede01STejun Heo } 1332a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1333bbb68dfaSTejun Heo fail: 133424acfb71SThomas Gleixner rcu_read_unlock(); 1335bbb68dfaSTejun Heo local_irq_restore(*flags); 1336bbb68dfaSTejun Heo if (work_is_canceling(work)) 1337bbb68dfaSTejun Heo return -ENOENT; 1338bbb68dfaSTejun Heo cpu_relax(); 133936e227d2STejun Heo return -EAGAIN; 1340bf4ede01STejun Heo } 1341bf4ede01STejun Heo 1342bf4ede01STejun Heo /** 1343706026c2STejun Heo * insert_work - insert a work into a pool 1344112202d9STejun Heo * @pwq: pwq @work belongs to 13454690c4abSTejun Heo * @work: work to insert 13464690c4abSTejun Heo * @head: insertion point 13474690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13484690c4abSTejun Heo * 1349112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1350706026c2STejun Heo * work_struct flags. 13514690c4abSTejun Heo * 13524690c4abSTejun Heo * CONTEXT: 1353a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1354365970a1SDavid Howells */ 1355112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1356112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1357b89deed3SOleg Nesterov { 1358112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1359e1d8aa9fSFrederic Weisbecker 1360e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1361f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1362e89a85d6SWalter Wu 13634690c4abSTejun Heo /* we own @work, set data and link */ 1364112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13651a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13668864b4e5STejun Heo get_pwq(pwq); 1367e22bee78STejun Heo 136863d95a91STejun Heo if (__need_more_worker(pool)) 136963d95a91STejun Heo wake_up_worker(pool); 1370b89deed3SOleg Nesterov } 1371b89deed3SOleg Nesterov 1372c8efcc25STejun Heo /* 1373c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 13748d03ecfeSTejun Heo * same workqueue. 1375c8efcc25STejun Heo */ 1376c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1377c8efcc25STejun Heo { 1378c8efcc25STejun Heo struct worker *worker; 1379c8efcc25STejun Heo 13808d03ecfeSTejun Heo worker = current_wq_worker(); 1381c8efcc25STejun Heo /* 1382bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 13838d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1384c8efcc25STejun Heo */ 1385112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1386c8efcc25STejun Heo } 1387c8efcc25STejun Heo 1388ef557180SMike Galbraith /* 1389ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1390ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1391ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1392ef557180SMike Galbraith */ 1393ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1394ef557180SMike Galbraith { 1395f303fccbSTejun Heo static bool printed_dbg_warning; 1396ef557180SMike Galbraith int new_cpu; 1397ef557180SMike Galbraith 1398f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1399ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1400ef557180SMike Galbraith return cpu; 1401f303fccbSTejun Heo } else if (!printed_dbg_warning) { 1402f303fccbSTejun Heo pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1403f303fccbSTejun Heo printed_dbg_warning = true; 1404f303fccbSTejun Heo } 1405f303fccbSTejun Heo 1406ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1407ef557180SMike Galbraith return cpu; 1408ef557180SMike Galbraith 1409ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1410ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1411ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1412ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1413ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1414ef557180SMike Galbraith return cpu; 1415ef557180SMike Galbraith } 1416ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1417ef557180SMike Galbraith 1418ef557180SMike Galbraith return new_cpu; 1419ef557180SMike Galbraith } 1420ef557180SMike Galbraith 1421d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 14221da177e4SLinus Torvalds struct work_struct *work) 14231da177e4SLinus Torvalds { 1424112202d9STejun Heo struct pool_workqueue *pwq; 1425c9178087STejun Heo struct worker_pool *last_pool; 14261e19ffc6STejun Heo struct list_head *worklist; 14278a2e8e5dSTejun Heo unsigned int work_flags; 1428b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 14298930cabaSTejun Heo 14308930cabaSTejun Heo /* 14318930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 14328930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 14338930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 14348930cabaSTejun Heo * happen with IRQ disabled. 14358930cabaSTejun Heo */ 14368e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 14371da177e4SLinus Torvalds 14381e19ffc6STejun Heo 143933e3f0a3SRichard Clark /* 144033e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 144133e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 144233e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 144333e3f0a3SRichard Clark */ 144433e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 144533e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1446e41e704bSTejun Heo return; 144724acfb71SThomas Gleixner rcu_read_lock(); 14489e8cd2f5STejun Heo retry: 1449aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1450aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1451df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1452ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1453df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1454aa202f1fSHillf Danton } else { 1455aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1456aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1457aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1458aa202f1fSHillf Danton } 1459f3421797STejun Heo 146018aa9effSTejun Heo /* 1461c9178087STejun Heo * If @work was previously on a different pool, it might still be 1462c9178087STejun Heo * running there, in which case the work needs to be queued on that 1463c9178087STejun Heo * pool to guarantee non-reentrancy. 146418aa9effSTejun Heo */ 1465c9e7cf27STejun Heo last_pool = get_work_pool(work); 1466112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 146718aa9effSTejun Heo struct worker *worker; 146818aa9effSTejun Heo 1469a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 147018aa9effSTejun Heo 1471c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 147218aa9effSTejun Heo 1473112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1474c9178087STejun Heo pwq = worker->current_pwq; 14758594fadeSLai Jiangshan } else { 147618aa9effSTejun Heo /* meh... not running there, queue here */ 1477a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1478a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 147918aa9effSTejun Heo } 14808930cabaSTejun Heo } else { 1481a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14828930cabaSTejun Heo } 1483502ca9d8STejun Heo 14849e8cd2f5STejun Heo /* 14859e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14869e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14879e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1488df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1489df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14909e8cd2f5STejun Heo * make forward-progress. 14919e8cd2f5STejun Heo */ 14929e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14939e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1494a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14959e8cd2f5STejun Heo cpu_relax(); 14969e8cd2f5STejun Heo goto retry; 14979e8cd2f5STejun Heo } 14989e8cd2f5STejun Heo /* oops */ 14999e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 15009e8cd2f5STejun Heo wq->name, cpu); 15019e8cd2f5STejun Heo } 15029e8cd2f5STejun Heo 1503112202d9STejun Heo /* pwq determined, queue */ 1504112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1505502ca9d8STejun Heo 150624acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 150724acfb71SThomas Gleixner goto out; 15081e19ffc6STejun Heo 1509112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1510112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 15111e19ffc6STejun Heo 1512112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1513cdadf009STejun Heo trace_workqueue_activate_work(work); 1514112202d9STejun Heo pwq->nr_active++; 1515112202d9STejun Heo worklist = &pwq->pool->worklist; 151682607adcSTejun Heo if (list_empty(worklist)) 151782607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 15188a2e8e5dSTejun Heo } else { 1519f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1520f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 15218a2e8e5dSTejun Heo } 15221e19ffc6STejun Heo 15230687c66bSZqiang debug_work_activate(work); 1524112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15251e19ffc6STejun Heo 152624acfb71SThomas Gleixner out: 1527a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 152824acfb71SThomas Gleixner rcu_read_unlock(); 15291da177e4SLinus Torvalds } 15301da177e4SLinus Torvalds 15310fcb78c2SRolf Eike Beer /** 1532c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1533c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1534c1a220e7SZhang Rui * @wq: workqueue to use 1535c1a220e7SZhang Rui * @work: work to queue 1536c1a220e7SZhang Rui * 1537c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1538443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1539443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1540d185af30SYacine Belkadi * 1541d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1542c1a220e7SZhang Rui */ 1543d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1544d4283e93STejun Heo struct work_struct *work) 1545c1a220e7SZhang Rui { 1546d4283e93STejun Heo bool ret = false; 15478930cabaSTejun Heo unsigned long flags; 15488930cabaSTejun Heo 15498930cabaSTejun Heo local_irq_save(flags); 1550c1a220e7SZhang Rui 155122df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15524690c4abSTejun Heo __queue_work(cpu, wq, work); 1553d4283e93STejun Heo ret = true; 1554c1a220e7SZhang Rui } 15558930cabaSTejun Heo 15568930cabaSTejun Heo local_irq_restore(flags); 1557c1a220e7SZhang Rui return ret; 1558c1a220e7SZhang Rui } 1559ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1560c1a220e7SZhang Rui 15618204e0c1SAlexander Duyck /** 15628204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15638204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15648204e0c1SAlexander Duyck * 15658204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15668204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15678204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15688204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15698204e0c1SAlexander Duyck */ 15708204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15718204e0c1SAlexander Duyck { 15728204e0c1SAlexander Duyck int cpu; 15738204e0c1SAlexander Duyck 15748204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15758204e0c1SAlexander Duyck if (!wq_numa_enabled) 15768204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15778204e0c1SAlexander Duyck 15788204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15798204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15808204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15818204e0c1SAlexander Duyck 15828204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15838204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15848204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15858204e0c1SAlexander Duyck return cpu; 15868204e0c1SAlexander Duyck 15878204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15888204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15898204e0c1SAlexander Duyck 15908204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15918204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15928204e0c1SAlexander Duyck } 15938204e0c1SAlexander Duyck 15948204e0c1SAlexander Duyck /** 15958204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15968204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15978204e0c1SAlexander Duyck * @wq: workqueue to use 15988204e0c1SAlexander Duyck * @work: work to queue 15998204e0c1SAlexander Duyck * 16008204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 16018204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 16028204e0c1SAlexander Duyck * NUMA node. 16038204e0c1SAlexander Duyck * 16048204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 16058204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 16068204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 16078204e0c1SAlexander Duyck * 16088204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 16098204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 16108204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 16118204e0c1SAlexander Duyck * 16128204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 16138204e0c1SAlexander Duyck */ 16148204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 16158204e0c1SAlexander Duyck struct work_struct *work) 16168204e0c1SAlexander Duyck { 16178204e0c1SAlexander Duyck unsigned long flags; 16188204e0c1SAlexander Duyck bool ret = false; 16198204e0c1SAlexander Duyck 16208204e0c1SAlexander Duyck /* 16218204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 16228204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 16238204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 16248204e0c1SAlexander Duyck * 16258204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16268204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16278204e0c1SAlexander Duyck * some round robin type logic. 16288204e0c1SAlexander Duyck */ 16298204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16308204e0c1SAlexander Duyck 16318204e0c1SAlexander Duyck local_irq_save(flags); 16328204e0c1SAlexander Duyck 16338204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16348204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16358204e0c1SAlexander Duyck 16368204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16378204e0c1SAlexander Duyck ret = true; 16388204e0c1SAlexander Duyck } 16398204e0c1SAlexander Duyck 16408204e0c1SAlexander Duyck local_irq_restore(flags); 16418204e0c1SAlexander Duyck return ret; 16428204e0c1SAlexander Duyck } 16438204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16448204e0c1SAlexander Duyck 16458c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16461da177e4SLinus Torvalds { 16478c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16481da177e4SLinus Torvalds 1649e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 165060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16511da177e4SLinus Torvalds } 16521438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16531da177e4SLinus Torvalds 16547beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 165552bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16561da177e4SLinus Torvalds { 16577beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16587beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16591da177e4SLinus Torvalds 1660637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 16614b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1662fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1663fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16647beb2edfSTejun Heo 16658852aac2STejun Heo /* 16668852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16678852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16688852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16698852aac2STejun Heo * on that there's no such delay when @delay is 0. 16708852aac2STejun Heo */ 16718852aac2STejun Heo if (!delay) { 16728852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16738852aac2STejun Heo return; 16748852aac2STejun Heo } 16758852aac2STejun Heo 167660c057bcSLai Jiangshan dwork->wq = wq; 16771265057fSTejun Heo dwork->cpu = cpu; 16787beb2edfSTejun Heo timer->expires = jiffies + delay; 16797beb2edfSTejun Heo 1680041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16817beb2edfSTejun Heo add_timer_on(timer, cpu); 1682041bd12eSTejun Heo else 1683041bd12eSTejun Heo add_timer(timer); 16847beb2edfSTejun Heo } 16851da177e4SLinus Torvalds 16860fcb78c2SRolf Eike Beer /** 16870fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16880fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16890fcb78c2SRolf Eike Beer * @wq: workqueue to use 1690af9997e4SRandy Dunlap * @dwork: work to queue 16910fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16920fcb78c2SRolf Eike Beer * 1693d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1694715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1695715f1300STejun Heo * execution. 16960fcb78c2SRolf Eike Beer */ 1697d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 169852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16997a6bc1cdSVenkatesh Pallipadi { 170052bad64dSDavid Howells struct work_struct *work = &dwork->work; 1701d4283e93STejun Heo bool ret = false; 17028930cabaSTejun Heo unsigned long flags; 17038930cabaSTejun Heo 17048930cabaSTejun Heo /* read the comment in __queue_work() */ 17058930cabaSTejun Heo local_irq_save(flags); 17067a6bc1cdSVenkatesh Pallipadi 170722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17087beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1709d4283e93STejun Heo ret = true; 17107a6bc1cdSVenkatesh Pallipadi } 17118930cabaSTejun Heo 17128930cabaSTejun Heo local_irq_restore(flags); 17137a6bc1cdSVenkatesh Pallipadi return ret; 17147a6bc1cdSVenkatesh Pallipadi } 1715ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 17161da177e4SLinus Torvalds 1717c8e55f36STejun Heo /** 17188376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 17198376fe22STejun Heo * @cpu: CPU number to execute work on 17208376fe22STejun Heo * @wq: workqueue to use 17218376fe22STejun Heo * @dwork: work to queue 17228376fe22STejun Heo * @delay: number of jiffies to wait before queueing 17238376fe22STejun Heo * 17248376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17258376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17268376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17278376fe22STejun Heo * current state. 17288376fe22STejun Heo * 1729d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17308376fe22STejun Heo * pending and its timer was modified. 17318376fe22STejun Heo * 1732e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17338376fe22STejun Heo * See try_to_grab_pending() for details. 17348376fe22STejun Heo */ 17358376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17368376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17378376fe22STejun Heo { 17388376fe22STejun Heo unsigned long flags; 17398376fe22STejun Heo int ret; 17408376fe22STejun Heo 17418376fe22STejun Heo do { 17428376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17438376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17448376fe22STejun Heo 17458376fe22STejun Heo if (likely(ret >= 0)) { 17468376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17478376fe22STejun Heo local_irq_restore(flags); 17488376fe22STejun Heo } 17498376fe22STejun Heo 17508376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17518376fe22STejun Heo return ret; 17528376fe22STejun Heo } 17538376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17548376fe22STejun Heo 175505f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 175605f0fe6bSTejun Heo { 175705f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 175805f0fe6bSTejun Heo 175905f0fe6bSTejun Heo /* read the comment in __queue_work() */ 176005f0fe6bSTejun Heo local_irq_disable(); 176105f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 176205f0fe6bSTejun Heo local_irq_enable(); 176305f0fe6bSTejun Heo } 176405f0fe6bSTejun Heo 176505f0fe6bSTejun Heo /** 176605f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 176705f0fe6bSTejun Heo * @wq: workqueue to use 176805f0fe6bSTejun Heo * @rwork: work to queue 176905f0fe6bSTejun Heo * 177005f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 177105f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1772bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 177305f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 177405f0fe6bSTejun Heo */ 177505f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 177605f0fe6bSTejun Heo { 177705f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 177805f0fe6bSTejun Heo 177905f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 178005f0fe6bSTejun Heo rwork->wq = wq; 1781a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 178205f0fe6bSTejun Heo return true; 178305f0fe6bSTejun Heo } 178405f0fe6bSTejun Heo 178505f0fe6bSTejun Heo return false; 178605f0fe6bSTejun Heo } 178705f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 178805f0fe6bSTejun Heo 17898376fe22STejun Heo /** 1790c8e55f36STejun Heo * worker_enter_idle - enter idle state 1791c8e55f36STejun Heo * @worker: worker which is entering idle state 1792c8e55f36STejun Heo * 1793c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1794c8e55f36STejun Heo * necessary. 1795c8e55f36STejun Heo * 1796c8e55f36STejun Heo * LOCKING: 1797a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1798c8e55f36STejun Heo */ 1799c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 18001da177e4SLinus Torvalds { 1801bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1802c8e55f36STejun Heo 18036183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 18046183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 18056183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 18066183c009STejun Heo return; 1807c8e55f36STejun Heo 1808051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1809cb444766STejun Heo worker->flags |= WORKER_IDLE; 1810bd7bdd43STejun Heo pool->nr_idle++; 1811e22bee78STejun Heo worker->last_active = jiffies; 1812c8e55f36STejun Heo 1813c8e55f36STejun Heo /* idle_list is LIFO */ 1814bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1815db7bccf4STejun Heo 181663d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1817628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1818cb444766STejun Heo 1819989442d7SLai Jiangshan /* Sanity check nr_running. */ 1820bc35f7efSLai Jiangshan WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 1821c8e55f36STejun Heo } 1822c8e55f36STejun Heo 1823c8e55f36STejun Heo /** 1824c8e55f36STejun Heo * worker_leave_idle - leave idle state 1825c8e55f36STejun Heo * @worker: worker which is leaving idle state 1826c8e55f36STejun Heo * 1827c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1828c8e55f36STejun Heo * 1829c8e55f36STejun Heo * LOCKING: 1830a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1831c8e55f36STejun Heo */ 1832c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1833c8e55f36STejun Heo { 1834bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1835c8e55f36STejun Heo 18366183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18376183c009STejun Heo return; 1838d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1839bd7bdd43STejun Heo pool->nr_idle--; 1840c8e55f36STejun Heo list_del_init(&worker->entry); 1841c8e55f36STejun Heo } 1842c8e55f36STejun Heo 1843f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1844c34056a3STejun Heo { 1845c34056a3STejun Heo struct worker *worker; 1846c34056a3STejun Heo 1847f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1848c8e55f36STejun Heo if (worker) { 1849c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1850affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1851da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1852e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1853e22bee78STejun Heo worker->flags = WORKER_PREP; 1854c8e55f36STejun Heo } 1855c34056a3STejun Heo return worker; 1856c34056a3STejun Heo } 1857c34056a3STejun Heo 1858c34056a3STejun Heo /** 18594736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18604736cbf7SLai Jiangshan * @worker: worker to be attached 18614736cbf7SLai Jiangshan * @pool: the target pool 18624736cbf7SLai Jiangshan * 18634736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18644736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18654736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18664736cbf7SLai Jiangshan */ 18674736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18684736cbf7SLai Jiangshan struct worker_pool *pool) 18694736cbf7SLai Jiangshan { 18701258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18714736cbf7SLai Jiangshan 18724736cbf7SLai Jiangshan /* 18731258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18741258fae7STejun Heo * stable across this function. See the comments above the flag 18751258fae7STejun Heo * definition for details. 18764736cbf7SLai Jiangshan */ 18774736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18784736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18795c25b5ffSPeter Zijlstra else 18805c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 18814736cbf7SLai Jiangshan 1882640f17c8SPeter Zijlstra if (worker->rescue_wq) 1883640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1884640f17c8SPeter Zijlstra 18854736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1886a2d812a2STejun Heo worker->pool = pool; 18874736cbf7SLai Jiangshan 18881258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18894736cbf7SLai Jiangshan } 18904736cbf7SLai Jiangshan 18914736cbf7SLai Jiangshan /** 189260f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 189360f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 189460f5a4bcSLai Jiangshan * 18954736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18964736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 18974736cbf7SLai Jiangshan * other reference to the pool. 189860f5a4bcSLai Jiangshan */ 1899a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 190060f5a4bcSLai Jiangshan { 1901a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 190260f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 190360f5a4bcSLai Jiangshan 19041258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1905a2d812a2STejun Heo 19065c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 1907da028469SLai Jiangshan list_del(&worker->node); 1908a2d812a2STejun Heo worker->pool = NULL; 1909a2d812a2STejun Heo 1910*e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 191160f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 19121258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 191360f5a4bcSLai Jiangshan 1914b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1915b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1916b62c0751SLai Jiangshan 191760f5a4bcSLai Jiangshan if (detach_completion) 191860f5a4bcSLai Jiangshan complete(detach_completion); 191960f5a4bcSLai Jiangshan } 192060f5a4bcSLai Jiangshan 192160f5a4bcSLai Jiangshan /** 1922c34056a3STejun Heo * create_worker - create a new workqueue worker 192363d95a91STejun Heo * @pool: pool the new worker will belong to 1924c34056a3STejun Heo * 1925051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1926c34056a3STejun Heo * 1927c34056a3STejun Heo * CONTEXT: 1928c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1929c34056a3STejun Heo * 1930d185af30SYacine Belkadi * Return: 1931c34056a3STejun Heo * Pointer to the newly created worker. 1932c34056a3STejun Heo */ 1933bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1934c34056a3STejun Heo { 1935e441b56fSZhen Lei struct worker *worker; 1936e441b56fSZhen Lei int id; 1937e3c916a4STejun Heo char id_buf[16]; 1938c34056a3STejun Heo 19397cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 1940e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 1941822d8405STejun Heo if (id < 0) 1942e441b56fSZhen Lei return NULL; 1943c34056a3STejun Heo 1944f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1945c34056a3STejun Heo if (!worker) 1946c34056a3STejun Heo goto fail; 1947c34056a3STejun Heo 1948c34056a3STejun Heo worker->id = id; 1949c34056a3STejun Heo 195029c91e99STejun Heo if (pool->cpu >= 0) 1951e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1952e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1953f3421797STejun Heo else 1954e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1955e3c916a4STejun Heo 1956f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1957e3c916a4STejun Heo "kworker/%s", id_buf); 1958c34056a3STejun Heo if (IS_ERR(worker->task)) 1959c34056a3STejun Heo goto fail; 1960c34056a3STejun Heo 196191151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 196225834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 196391151228SOleg Nesterov 1964da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19654736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1966822d8405STejun Heo 1967051e1850SLai Jiangshan /* start the newly created worker */ 1968a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1969051e1850SLai Jiangshan worker->pool->nr_workers++; 1970051e1850SLai Jiangshan worker_enter_idle(worker); 1971051e1850SLai Jiangshan wake_up_process(worker->task); 1972a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1973051e1850SLai Jiangshan 1974c34056a3STejun Heo return worker; 1975822d8405STejun Heo 1976c34056a3STejun Heo fail: 1977e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 1978c34056a3STejun Heo kfree(worker); 1979c34056a3STejun Heo return NULL; 1980c34056a3STejun Heo } 1981c34056a3STejun Heo 1982793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 1983793777bcSValentin Schneider { 1984793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 1985793777bcSValentin Schneider 1986793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 1987793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 1988793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 1989793777bcSValentin Schneider else 1990793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 1991793777bcSValentin Schneider } 1992793777bcSValentin Schneider 1993793777bcSValentin Schneider static void rebind_worker(struct worker *worker, struct worker_pool *pool) 1994793777bcSValentin Schneider { 1995793777bcSValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 1996793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask) < 0); 1997793777bcSValentin Schneider } 1998793777bcSValentin Schneider 1999*e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2000*e02b9312SValentin Schneider { 2001*e02b9312SValentin Schneider struct worker *worker, *tmp; 2002*e02b9312SValentin Schneider 2003*e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2004*e02b9312SValentin Schneider list_del_init(&worker->entry); 2005*e02b9312SValentin Schneider unbind_worker(worker); 2006*e02b9312SValentin Schneider /* 2007*e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2008*e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2009*e02b9312SValentin Schneider * wouldn't have gotten here. 2010c34056a3STejun Heo * 2011*e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2012*e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2013*e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2014*e02b9312SValentin Schneider * outside of pool->lock. 2015*e02b9312SValentin Schneider */ 2016*e02b9312SValentin Schneider wake_up_process(worker->task); 2017*e02b9312SValentin Schneider } 2018*e02b9312SValentin Schneider } 2019*e02b9312SValentin Schneider 2020*e02b9312SValentin Schneider /** 2021*e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2022*e02b9312SValentin Schneider * @worker: worker to be destroyed 2023*e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2024*e02b9312SValentin Schneider * 2025*e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2026*e02b9312SValentin Schneider * should be idle. 2027c8e55f36STejun Heo * 2028c8e55f36STejun Heo * CONTEXT: 2029a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2030c34056a3STejun Heo */ 2031*e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2032c34056a3STejun Heo { 2033bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2034c34056a3STejun Heo 2035cd549687STejun Heo lockdep_assert_held(&pool->lock); 2036*e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2037cd549687STejun Heo 2038c34056a3STejun Heo /* sanity check frenzy */ 20396183c009STejun Heo if (WARN_ON(worker->current_work) || 204073eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 204173eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 20426183c009STejun Heo return; 2043c34056a3STejun Heo 2044bd7bdd43STejun Heo pool->nr_workers--; 2045bd7bdd43STejun Heo pool->nr_idle--; 2046c8e55f36STejun Heo 2047cb444766STejun Heo worker->flags |= WORKER_DIE; 2048*e02b9312SValentin Schneider 2049*e02b9312SValentin Schneider list_move(&worker->entry, list); 2050*e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2051c34056a3STejun Heo } 2052c34056a3STejun Heo 20533f959aa3SValentin Schneider /** 20543f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 20553f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 20563f959aa3SValentin Schneider * 20573f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 20583f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 20593f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 20603f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 20613f959aa3SValentin Schneider * it expire and re-evaluate things from there. 20623f959aa3SValentin Schneider */ 206332a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2064e22bee78STejun Heo { 206532a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 20663f959aa3SValentin Schneider bool do_cull = false; 20673f959aa3SValentin Schneider 20683f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 20693f959aa3SValentin Schneider return; 20703f959aa3SValentin Schneider 20713f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 20723f959aa3SValentin Schneider 20733f959aa3SValentin Schneider if (too_many_workers(pool)) { 20743f959aa3SValentin Schneider struct worker *worker; 20753f959aa3SValentin Schneider unsigned long expires; 20763f959aa3SValentin Schneider 20773f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 20783f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 20793f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 20803f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 20813f959aa3SValentin Schneider 20823f959aa3SValentin Schneider if (!do_cull) 20833f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 20843f959aa3SValentin Schneider } 20853f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 20863f959aa3SValentin Schneider 20873f959aa3SValentin Schneider if (do_cull) 20883f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 20893f959aa3SValentin Schneider } 20903f959aa3SValentin Schneider 20913f959aa3SValentin Schneider /** 20923f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 20933f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 20943f959aa3SValentin Schneider * 20953f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 20963f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2097*e02b9312SValentin Schneider * 2098*e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2099*e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2100*e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 21013f959aa3SValentin Schneider */ 21023f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 21033f959aa3SValentin Schneider { 21043f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 2105*e02b9312SValentin Schneider struct list_head cull_list; 2106e22bee78STejun Heo 2107*e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 2108*e02b9312SValentin Schneider /* 2109*e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2110*e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2111*e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2112*e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2113*e02b9312SValentin Schneider */ 2114*e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2115a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2116e22bee78STejun Heo 21173347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2118e22bee78STejun Heo struct worker *worker; 2119e22bee78STejun Heo unsigned long expires; 2120e22bee78STejun Heo 212163d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2122e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2123e22bee78STejun Heo 21243347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 212563d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 21263347fc9fSLai Jiangshan break; 2127e22bee78STejun Heo } 21283347fc9fSLai Jiangshan 2129*e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2130e22bee78STejun Heo } 2131e22bee78STejun Heo 2132a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2133*e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2134*e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2135e22bee78STejun Heo } 2136e22bee78STejun Heo 2137493a1724STejun Heo static void send_mayday(struct work_struct *work) 2138e22bee78STejun Heo { 2139112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2140112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2141493a1724STejun Heo 21422e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2143e22bee78STejun Heo 2144493008a8STejun Heo if (!wq->rescuer) 2145493a1724STejun Heo return; 2146e22bee78STejun Heo 2147e22bee78STejun Heo /* mayday mayday mayday */ 2148493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 214977668c8bSLai Jiangshan /* 215077668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 215177668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 215277668c8bSLai Jiangshan * rescuer is done with it. 215377668c8bSLai Jiangshan */ 215477668c8bSLai Jiangshan get_pwq(pwq); 2155493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2156e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2157493a1724STejun Heo } 2158e22bee78STejun Heo } 2159e22bee78STejun Heo 216032a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2161e22bee78STejun Heo { 216232a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2163e22bee78STejun Heo struct work_struct *work; 2164e22bee78STejun Heo 2165a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2166a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2167e22bee78STejun Heo 216863d95a91STejun Heo if (need_to_create_worker(pool)) { 2169e22bee78STejun Heo /* 2170e22bee78STejun Heo * We've been trying to create a new worker but 2171e22bee78STejun Heo * haven't been successful. We might be hitting an 2172e22bee78STejun Heo * allocation deadlock. Send distress signals to 2173e22bee78STejun Heo * rescuers. 2174e22bee78STejun Heo */ 217563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2176e22bee78STejun Heo send_mayday(work); 2177e22bee78STejun Heo } 2178e22bee78STejun Heo 2179a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2180a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2181e22bee78STejun Heo 218263d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2183e22bee78STejun Heo } 2184e22bee78STejun Heo 2185e22bee78STejun Heo /** 2186e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 218763d95a91STejun Heo * @pool: pool to create a new worker for 2188e22bee78STejun Heo * 218963d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2190e22bee78STejun Heo * have at least one idle worker on return from this function. If 2191e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 219263d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2193e22bee78STejun Heo * possible allocation deadlock. 2194e22bee78STejun Heo * 2195c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2196c5aa87bbSTejun Heo * may_start_working() %true. 2197e22bee78STejun Heo * 2198e22bee78STejun Heo * LOCKING: 2199a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2200e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2201e22bee78STejun Heo * manager. 2202e22bee78STejun Heo */ 220329187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2204d565ed63STejun Heo __releases(&pool->lock) 2205d565ed63STejun Heo __acquires(&pool->lock) 2206e22bee78STejun Heo { 2207e22bee78STejun Heo restart: 2208a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 22099f9c2364STejun Heo 2210e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 221163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2212e22bee78STejun Heo 2213e22bee78STejun Heo while (true) { 2214051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2215e22bee78STejun Heo break; 2216e22bee78STejun Heo 2217e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 22189f9c2364STejun Heo 221963d95a91STejun Heo if (!need_to_create_worker(pool)) 2220e22bee78STejun Heo break; 2221e22bee78STejun Heo } 2222e22bee78STejun Heo 222363d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2224a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2225051e1850SLai Jiangshan /* 2226051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2227051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2228051e1850SLai Jiangshan * already become busy. 2229051e1850SLai Jiangshan */ 223063d95a91STejun Heo if (need_to_create_worker(pool)) 2231e22bee78STejun Heo goto restart; 2232e22bee78STejun Heo } 2233e22bee78STejun Heo 2234e22bee78STejun Heo /** 2235e22bee78STejun Heo * manage_workers - manage worker pool 2236e22bee78STejun Heo * @worker: self 2237e22bee78STejun Heo * 2238706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2239e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2240706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2241e22bee78STejun Heo * 2242e22bee78STejun Heo * The caller can safely start processing works on false return. On 2243e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2244e22bee78STejun Heo * and may_start_working() is true. 2245e22bee78STejun Heo * 2246e22bee78STejun Heo * CONTEXT: 2247a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2248e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2249e22bee78STejun Heo * 2250d185af30SYacine Belkadi * Return: 225129187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 225229187a9eSTejun Heo * start processing works, %true if management function was performed and 225329187a9eSTejun Heo * the conditions that the caller verified before calling the function may 225429187a9eSTejun Heo * no longer be true. 2255e22bee78STejun Heo */ 2256e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2257e22bee78STejun Heo { 225863d95a91STejun Heo struct worker_pool *pool = worker->pool; 2259e22bee78STejun Heo 2260692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 226129187a9eSTejun Heo return false; 2262692b4825STejun Heo 2263692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 22642607d7a6STejun Heo pool->manager = worker; 2265e22bee78STejun Heo 226629187a9eSTejun Heo maybe_create_worker(pool); 2267e22bee78STejun Heo 22682607d7a6STejun Heo pool->manager = NULL; 2269692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2270d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 227129187a9eSTejun Heo return true; 2272e22bee78STejun Heo } 2273e22bee78STejun Heo 2274a62428c0STejun Heo /** 2275a62428c0STejun Heo * process_one_work - process single work 2276c34056a3STejun Heo * @worker: self 2277a62428c0STejun Heo * @work: work to process 2278a62428c0STejun Heo * 2279a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2280a62428c0STejun Heo * process a single work including synchronization against and 2281a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2282a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2283a62428c0STejun Heo * call this function to process a work. 2284a62428c0STejun Heo * 2285a62428c0STejun Heo * CONTEXT: 2286a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2287a62428c0STejun Heo */ 2288c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2289d565ed63STejun Heo __releases(&pool->lock) 2290d565ed63STejun Heo __acquires(&pool->lock) 22911da177e4SLinus Torvalds { 2292112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2293bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2294112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2295c4560c2cSLai Jiangshan unsigned long work_data; 22967e11629dSTejun Heo struct worker *collision; 22974e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 22984e6045f1SJohannes Berg /* 2299a62428c0STejun Heo * It is permissible to free the struct work_struct from 2300a62428c0STejun Heo * inside the function that is called from it, this we need to 2301a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2302a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2303a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 23044e6045f1SJohannes Berg */ 23054d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 23064d82a1deSPeter Zijlstra 23074d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 23084e6045f1SJohannes Berg #endif 2309807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 231085327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2311ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 231225511a47STejun Heo 23137e11629dSTejun Heo /* 23147e11629dSTejun Heo * A single work shouldn't be executed concurrently by 23157e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 23167e11629dSTejun Heo * already processing the work. If so, defer the work to the 23177e11629dSTejun Heo * currently executing one. 23187e11629dSTejun Heo */ 2319c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 23207e11629dSTejun Heo if (unlikely(collision)) { 23217e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 23227e11629dSTejun Heo return; 23237e11629dSTejun Heo } 23241da177e4SLinus Torvalds 23258930cabaSTejun Heo /* claim and dequeue */ 23261da177e4SLinus Torvalds debug_work_deactivate(work); 2327c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2328c34056a3STejun Heo worker->current_work = work; 2329a2c1c57bSTejun Heo worker->current_func = work->func; 2330112202d9STejun Heo worker->current_pwq = pwq; 2331c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2332d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 23337a22ad75STejun Heo 23348bf89593STejun Heo /* 23358bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 23368bf89593STejun Heo * overridden through set_worker_desc(). 23378bf89593STejun Heo */ 23388bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 23398bf89593STejun Heo 2340a62428c0STejun Heo list_del_init(&work->entry); 2341a62428c0STejun Heo 2342649027d7STejun Heo /* 2343228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2344228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2345228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2346228f1d00SLai Jiangshan * execution of the pending work items. 2347fb0e7bebSTejun Heo */ 2348fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2349228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2350fb0e7bebSTejun Heo 2351974271c4STejun Heo /* 2352a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2353a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2354a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2355a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2356228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2357974271c4STejun Heo */ 2358a489a03eSLai Jiangshan if (need_more_worker(pool)) 235963d95a91STejun Heo wake_up_worker(pool); 2360974271c4STejun Heo 23618930cabaSTejun Heo /* 23627c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2363d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 236423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 236523657bb1STejun Heo * disabled. 23668930cabaSTejun Heo */ 23677c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 23681da177e4SLinus Torvalds 2369a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2370365970a1SDavid Howells 2371a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 23723295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2373e6f3faa7SPeter Zijlstra /* 2374f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2375f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2376e6f3faa7SPeter Zijlstra * 2377e6f3faa7SPeter Zijlstra * However, that would result in: 2378e6f3faa7SPeter Zijlstra * 2379e6f3faa7SPeter Zijlstra * A(W1) 2380e6f3faa7SPeter Zijlstra * WFC(C) 2381e6f3faa7SPeter Zijlstra * A(W1) 2382e6f3faa7SPeter Zijlstra * C(C) 2383e6f3faa7SPeter Zijlstra * 2384e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2385e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2386e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2387f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2388e6f3faa7SPeter Zijlstra * these locks. 2389e6f3faa7SPeter Zijlstra * 2390e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2391e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2392e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2393e6f3faa7SPeter Zijlstra */ 2394f52be570SPeter Zijlstra lockdep_invariant_state(true); 2395e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2396a2c1c57bSTejun Heo worker->current_func(work); 2397e36c886aSArjan van de Ven /* 2398e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2399e36c886aSArjan van de Ven * point will only record its address. 2400e36c886aSArjan van de Ven */ 24011c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 24023295f0efSIngo Molnar lock_map_release(&lockdep_map); 2403112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 24041da177e4SLinus Torvalds 2405d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2406044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2407d75f773cSSakari Ailus " last function: %ps\n", 2408a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2409a2c1c57bSTejun Heo worker->current_func); 2410d5abe669SPeter Zijlstra debug_show_held_locks(current); 2411d5abe669SPeter Zijlstra dump_stack(); 2412d5abe669SPeter Zijlstra } 2413d5abe669SPeter Zijlstra 2414b22ce278STejun Heo /* 2415025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2416b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2417b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2418b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2419789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2420789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2421b22ce278STejun Heo */ 2422a7e6425eSPaul E. McKenney cond_resched(); 2423b22ce278STejun Heo 2424a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2425a62428c0STejun Heo 2426fb0e7bebSTejun Heo /* clear cpu intensive status */ 2427fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2428fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2429fb0e7bebSTejun Heo 24301b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 24311b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 24321b69ac6bSJohannes Weiner 2433a62428c0STejun Heo /* we're done with it, release */ 243442f8570fSSasha Levin hash_del(&worker->hentry); 2435c34056a3STejun Heo worker->current_work = NULL; 2436a2c1c57bSTejun Heo worker->current_func = NULL; 2437112202d9STejun Heo worker->current_pwq = NULL; 2438d812796eSLai Jiangshan worker->current_color = INT_MAX; 2439c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 24401da177e4SLinus Torvalds } 24411da177e4SLinus Torvalds 2442affee4b2STejun Heo /** 2443affee4b2STejun Heo * process_scheduled_works - process scheduled works 2444affee4b2STejun Heo * @worker: self 2445affee4b2STejun Heo * 2446affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2447affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2448affee4b2STejun Heo * fetches a work from the top and executes it. 2449affee4b2STejun Heo * 2450affee4b2STejun Heo * CONTEXT: 2451a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2452affee4b2STejun Heo * multiple times. 2453affee4b2STejun Heo */ 2454affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 24551da177e4SLinus Torvalds { 2456affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2457affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2458a62428c0STejun Heo struct work_struct, entry); 2459c34056a3STejun Heo process_one_work(worker, work); 2460a62428c0STejun Heo } 24611da177e4SLinus Torvalds } 24621da177e4SLinus Torvalds 2463197f6accSTejun Heo static void set_pf_worker(bool val) 2464197f6accSTejun Heo { 2465197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2466197f6accSTejun Heo if (val) 2467197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2468197f6accSTejun Heo else 2469197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2470197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2471197f6accSTejun Heo } 2472197f6accSTejun Heo 24734690c4abSTejun Heo /** 24744690c4abSTejun Heo * worker_thread - the worker thread function 2475c34056a3STejun Heo * @__worker: self 24764690c4abSTejun Heo * 2477c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2478c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2479c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2480c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2481c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2482d185af30SYacine Belkadi * 2483d185af30SYacine Belkadi * Return: 0 24844690c4abSTejun Heo */ 2485c34056a3STejun Heo static int worker_thread(void *__worker) 24861da177e4SLinus Torvalds { 2487c34056a3STejun Heo struct worker *worker = __worker; 2488bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 24891da177e4SLinus Torvalds 2490e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2491197f6accSTejun Heo set_pf_worker(true); 2492c8e55f36STejun Heo woke_up: 2493a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2494affee4b2STejun Heo 2495a9ab775bSTejun Heo /* am I supposed to die? */ 2496a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2497a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2498197f6accSTejun Heo set_pf_worker(false); 249960f5a4bcSLai Jiangshan 250060f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2501e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2502a2d812a2STejun Heo worker_detach_from_pool(worker); 2503*e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 250460f5a4bcSLai Jiangshan kfree(worker); 2505c8e55f36STejun Heo return 0; 2506c8e55f36STejun Heo } 2507c8e55f36STejun Heo 2508c8e55f36STejun Heo worker_leave_idle(worker); 2509db7bccf4STejun Heo recheck: 2510e22bee78STejun Heo /* no more worker necessary? */ 251163d95a91STejun Heo if (!need_more_worker(pool)) 2512e22bee78STejun Heo goto sleep; 2513e22bee78STejun Heo 2514e22bee78STejun Heo /* do we need to manage? */ 251563d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2516e22bee78STejun Heo goto recheck; 2517e22bee78STejun Heo 2518c8e55f36STejun Heo /* 2519c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2520c8e55f36STejun Heo * preparing to process a work or actually processing it. 2521c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2522c8e55f36STejun Heo */ 25236183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2524c8e55f36STejun Heo 2525e22bee78STejun Heo /* 2526a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2527a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2528a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2529a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2530a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2531e22bee78STejun Heo */ 2532a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2533e22bee78STejun Heo 2534e22bee78STejun Heo do { 2535affee4b2STejun Heo struct work_struct *work = 2536bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2537affee4b2STejun Heo struct work_struct, entry); 2538affee4b2STejun Heo 253982607adcSTejun Heo pool->watchdog_ts = jiffies; 254082607adcSTejun Heo 2541c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2542affee4b2STejun Heo /* optimization path, not strictly necessary */ 2543affee4b2STejun Heo process_one_work(worker, work); 2544affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2545affee4b2STejun Heo process_scheduled_works(worker); 2546affee4b2STejun Heo } else { 2547c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2548affee4b2STejun Heo process_scheduled_works(worker); 2549affee4b2STejun Heo } 255063d95a91STejun Heo } while (keep_working(pool)); 2551affee4b2STejun Heo 2552228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2553d313dd85STejun Heo sleep: 2554c8e55f36STejun Heo /* 2555d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2556d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2557d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2558d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2559d565ed63STejun Heo * event. 2560c8e55f36STejun Heo */ 2561c8e55f36STejun Heo worker_enter_idle(worker); 2562c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2563a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 25641da177e4SLinus Torvalds schedule(); 2565c8e55f36STejun Heo goto woke_up; 25661da177e4SLinus Torvalds } 25671da177e4SLinus Torvalds 2568e22bee78STejun Heo /** 2569e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2570111c225aSTejun Heo * @__rescuer: self 2571e22bee78STejun Heo * 2572e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2573493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2574e22bee78STejun Heo * 2575706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2576e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2577e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2578e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2579e22bee78STejun Heo * the problem rescuer solves. 2580e22bee78STejun Heo * 2581706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2582706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2583e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2584e22bee78STejun Heo * 2585e22bee78STejun Heo * This should happen rarely. 2586d185af30SYacine Belkadi * 2587d185af30SYacine Belkadi * Return: 0 2588e22bee78STejun Heo */ 2589111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2590e22bee78STejun Heo { 2591111c225aSTejun Heo struct worker *rescuer = __rescuer; 2592111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2593e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 25944d595b86SLai Jiangshan bool should_stop; 2595e22bee78STejun Heo 2596e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2597111c225aSTejun Heo 2598111c225aSTejun Heo /* 2599111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2600111c225aSTejun Heo * doesn't participate in concurrency management. 2601111c225aSTejun Heo */ 2602197f6accSTejun Heo set_pf_worker(true); 2603e22bee78STejun Heo repeat: 2604c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 26051da177e4SLinus Torvalds 26064d595b86SLai Jiangshan /* 26074d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 26084d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 26094d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 26104d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 26114d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 26124d595b86SLai Jiangshan * list is always empty on exit. 26134d595b86SLai Jiangshan */ 26144d595b86SLai Jiangshan should_stop = kthread_should_stop(); 26151da177e4SLinus Torvalds 2616493a1724STejun Heo /* see whether any pwq is asking for help */ 2617a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2618493a1724STejun Heo 2619493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2620493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2621493a1724STejun Heo struct pool_workqueue, mayday_node); 2622112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2623e22bee78STejun Heo struct work_struct *work, *n; 262482607adcSTejun Heo bool first = true; 2625e22bee78STejun Heo 2626e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2627493a1724STejun Heo list_del_init(&pwq->mayday_node); 2628493a1724STejun Heo 2629a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2630e22bee78STejun Heo 263151697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 263251697d39SLai Jiangshan 2633a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2634e22bee78STejun Heo 2635e22bee78STejun Heo /* 2636e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2637e22bee78STejun Heo * process'em. 2638e22bee78STejun Heo */ 26390479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 264082607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 264182607adcSTejun Heo if (get_work_pwq(work) == pwq) { 264282607adcSTejun Heo if (first) 264382607adcSTejun Heo pool->watchdog_ts = jiffies; 2644e22bee78STejun Heo move_linked_works(work, scheduled, &n); 264582607adcSTejun Heo } 264682607adcSTejun Heo first = false; 264782607adcSTejun Heo } 2648e22bee78STejun Heo 2649008847f6SNeilBrown if (!list_empty(scheduled)) { 2650e22bee78STejun Heo process_scheduled_works(rescuer); 26517576958aSTejun Heo 26527576958aSTejun Heo /* 2653008847f6SNeilBrown * The above execution of rescued work items could 2654008847f6SNeilBrown * have created more to rescue through 2655f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2656008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2657008847f6SNeilBrown * that such back-to-back work items, which may be 2658008847f6SNeilBrown * being used to relieve memory pressure, don't 2659008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2660008847f6SNeilBrown */ 26614f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2662a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2663e66b39afSTejun Heo /* 2664e66b39afSTejun Heo * Queue iff we aren't racing destruction 2665e66b39afSTejun Heo * and somebody else hasn't queued it already. 2666e66b39afSTejun Heo */ 2667e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2668008847f6SNeilBrown get_pwq(pwq); 2669e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2670e66b39afSTejun Heo } 2671a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2672008847f6SNeilBrown } 2673008847f6SNeilBrown } 2674008847f6SNeilBrown 2675008847f6SNeilBrown /* 267677668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 267713b1d625SLai Jiangshan * go away while we're still attached to it. 267877668c8bSLai Jiangshan */ 267977668c8bSLai Jiangshan put_pwq(pwq); 268077668c8bSLai Jiangshan 268177668c8bSLai Jiangshan /* 2682d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 26837576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 26847576958aSTejun Heo * and stalling the execution. 26857576958aSTejun Heo */ 2686d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 268763d95a91STejun Heo wake_up_worker(pool); 26887576958aSTejun Heo 2689a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 269013b1d625SLai Jiangshan 2691a2d812a2STejun Heo worker_detach_from_pool(rescuer); 269213b1d625SLai Jiangshan 2693a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 26941da177e4SLinus Torvalds } 26951da177e4SLinus Torvalds 2696a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2697493a1724STejun Heo 26984d595b86SLai Jiangshan if (should_stop) { 26994d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2700197f6accSTejun Heo set_pf_worker(false); 27014d595b86SLai Jiangshan return 0; 27024d595b86SLai Jiangshan } 27034d595b86SLai Jiangshan 2704111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2705111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2706e22bee78STejun Heo schedule(); 2707e22bee78STejun Heo goto repeat; 27081da177e4SLinus Torvalds } 27091da177e4SLinus Torvalds 2710fca839c0STejun Heo /** 2711fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2712fca839c0STejun Heo * @target_wq: workqueue being flushed 2713fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2714fca839c0STejun Heo * 2715fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2716fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2717fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2718fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2719fca839c0STejun Heo * a deadlock. 2720fca839c0STejun Heo */ 2721fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2722fca839c0STejun Heo struct work_struct *target_work) 2723fca839c0STejun Heo { 2724fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2725fca839c0STejun Heo struct worker *worker; 2726fca839c0STejun Heo 2727fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2728fca839c0STejun Heo return; 2729fca839c0STejun Heo 2730fca839c0STejun Heo worker = current_wq_worker(); 2731fca839c0STejun Heo 2732fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2733d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2734fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 273523d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 273623d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2737d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2738fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2739fca839c0STejun Heo target_wq->name, target_func); 2740fca839c0STejun Heo } 2741fca839c0STejun Heo 2742fc2e4d70SOleg Nesterov struct wq_barrier { 2743fc2e4d70SOleg Nesterov struct work_struct work; 2744fc2e4d70SOleg Nesterov struct completion done; 27452607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2746fc2e4d70SOleg Nesterov }; 2747fc2e4d70SOleg Nesterov 2748fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2749fc2e4d70SOleg Nesterov { 2750fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2751fc2e4d70SOleg Nesterov complete(&barr->done); 2752fc2e4d70SOleg Nesterov } 2753fc2e4d70SOleg Nesterov 27544690c4abSTejun Heo /** 27554690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2756112202d9STejun Heo * @pwq: pwq to insert barrier into 27574690c4abSTejun Heo * @barr: wq_barrier to insert 2758affee4b2STejun Heo * @target: target work to attach @barr to 2759affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 27604690c4abSTejun Heo * 2761affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2762affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2763affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2764affee4b2STejun Heo * cpu. 2765affee4b2STejun Heo * 2766affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2767affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2768affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2769affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2770affee4b2STejun Heo * after a work with LINKED flag set. 2771affee4b2STejun Heo * 2772affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2773112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 27744690c4abSTejun Heo * 27754690c4abSTejun Heo * CONTEXT: 2776a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 27774690c4abSTejun Heo */ 2778112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2779affee4b2STejun Heo struct wq_barrier *barr, 2780affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2781fc2e4d70SOleg Nesterov { 2782d812796eSLai Jiangshan unsigned int work_flags = 0; 2783d812796eSLai Jiangshan unsigned int work_color; 2784affee4b2STejun Heo struct list_head *head; 2785affee4b2STejun Heo 2786dc186ad7SThomas Gleixner /* 2787d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2788dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2789dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2790dc186ad7SThomas Gleixner * might deadlock. 2791dc186ad7SThomas Gleixner */ 2792ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 279322df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 279452fa5bc5SBoqun Feng 2795fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2796fd1a5b04SByungchul Park 27972607d7a6STejun Heo barr->task = current; 279883c22520SOleg Nesterov 2799018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2800018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2801018f3a13SLai Jiangshan 2802affee4b2STejun Heo /* 2803affee4b2STejun Heo * If @target is currently being executed, schedule the 2804affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2805affee4b2STejun Heo */ 2806d812796eSLai Jiangshan if (worker) { 2807affee4b2STejun Heo head = worker->scheduled.next; 2808d812796eSLai Jiangshan work_color = worker->current_color; 2809d812796eSLai Jiangshan } else { 2810affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2811affee4b2STejun Heo 2812affee4b2STejun Heo head = target->entry.next; 2813affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2814d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 2815d812796eSLai Jiangshan work_color = get_work_color(*bits); 2816affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2817affee4b2STejun Heo } 2818affee4b2STejun Heo 2819d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 2820d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 2821d812796eSLai Jiangshan 2822dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2823d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 2824fc2e4d70SOleg Nesterov } 2825fc2e4d70SOleg Nesterov 282673f53c4aSTejun Heo /** 2827112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 282873f53c4aSTejun Heo * @wq: workqueue being flushed 282973f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 283073f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 283173f53c4aSTejun Heo * 2832112202d9STejun Heo * Prepare pwqs for workqueue flushing. 283373f53c4aSTejun Heo * 2834112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2835112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2836112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2837112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2838112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 283973f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 284073f53c4aSTejun Heo * 284173f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 284273f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 284373f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 284473f53c4aSTejun Heo * is returned. 284573f53c4aSTejun Heo * 2846112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 284773f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 284873f53c4aSTejun Heo * advanced to @work_color. 284973f53c4aSTejun Heo * 285073f53c4aSTejun Heo * CONTEXT: 28513c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 285273f53c4aSTejun Heo * 2853d185af30SYacine Belkadi * Return: 285473f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 285573f53c4aSTejun Heo * otherwise. 285673f53c4aSTejun Heo */ 2857112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 285873f53c4aSTejun Heo int flush_color, int work_color) 28591da177e4SLinus Torvalds { 286073f53c4aSTejun Heo bool wait = false; 286149e3cf44STejun Heo struct pool_workqueue *pwq; 28621da177e4SLinus Torvalds 286373f53c4aSTejun Heo if (flush_color >= 0) { 28646183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2865112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2866dc186ad7SThomas Gleixner } 286714441960SOleg Nesterov 286849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2869112202d9STejun Heo struct worker_pool *pool = pwq->pool; 28701da177e4SLinus Torvalds 2871a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 287273f53c4aSTejun Heo 287373f53c4aSTejun Heo if (flush_color >= 0) { 28746183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 287573f53c4aSTejun Heo 2876112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2877112202d9STejun Heo pwq->flush_color = flush_color; 2878112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 287973f53c4aSTejun Heo wait = true; 28801da177e4SLinus Torvalds } 288173f53c4aSTejun Heo } 288273f53c4aSTejun Heo 288373f53c4aSTejun Heo if (work_color >= 0) { 28846183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2885112202d9STejun Heo pwq->work_color = work_color; 288673f53c4aSTejun Heo } 288773f53c4aSTejun Heo 2888a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 28891da177e4SLinus Torvalds } 28901da177e4SLinus Torvalds 2891112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 289273f53c4aSTejun Heo complete(&wq->first_flusher->done); 289373f53c4aSTejun Heo 289473f53c4aSTejun Heo return wait; 289583c22520SOleg Nesterov } 28961da177e4SLinus Torvalds 28970fcb78c2SRolf Eike Beer /** 2898c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 28990fcb78c2SRolf Eike Beer * @wq: workqueue to flush 29001da177e4SLinus Torvalds * 2901c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2902c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 29031da177e4SLinus Torvalds */ 2904c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 29051da177e4SLinus Torvalds { 290673f53c4aSTejun Heo struct wq_flusher this_flusher = { 290773f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 290873f53c4aSTejun Heo .flush_color = -1, 2909fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 291073f53c4aSTejun Heo }; 291173f53c4aSTejun Heo int next_color; 2912b1f4ec17SOleg Nesterov 29133347fa09STejun Heo if (WARN_ON(!wq_online)) 29143347fa09STejun Heo return; 29153347fa09STejun Heo 291687915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 291787915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 291887915adcSJohannes Berg 29193c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 292073f53c4aSTejun Heo 292173f53c4aSTejun Heo /* 292273f53c4aSTejun Heo * Start-to-wait phase 292373f53c4aSTejun Heo */ 292473f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 292573f53c4aSTejun Heo 292673f53c4aSTejun Heo if (next_color != wq->flush_color) { 292773f53c4aSTejun Heo /* 292873f53c4aSTejun Heo * Color space is not full. The current work_color 292973f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 293073f53c4aSTejun Heo * by one. 293173f53c4aSTejun Heo */ 29326183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 293373f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 293473f53c4aSTejun Heo wq->work_color = next_color; 293573f53c4aSTejun Heo 293673f53c4aSTejun Heo if (!wq->first_flusher) { 293773f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 29386183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 293973f53c4aSTejun Heo 294073f53c4aSTejun Heo wq->first_flusher = &this_flusher; 294173f53c4aSTejun Heo 2942112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 294373f53c4aSTejun Heo wq->work_color)) { 294473f53c4aSTejun Heo /* nothing to flush, done */ 294573f53c4aSTejun Heo wq->flush_color = next_color; 294673f53c4aSTejun Heo wq->first_flusher = NULL; 294773f53c4aSTejun Heo goto out_unlock; 294873f53c4aSTejun Heo } 294973f53c4aSTejun Heo } else { 295073f53c4aSTejun Heo /* wait in queue */ 29516183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 295273f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2953112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 295473f53c4aSTejun Heo } 295573f53c4aSTejun Heo } else { 295673f53c4aSTejun Heo /* 295773f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 295873f53c4aSTejun Heo * The next flush completion will assign us 295973f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 296073f53c4aSTejun Heo */ 296173f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 296273f53c4aSTejun Heo } 296373f53c4aSTejun Heo 2964fca839c0STejun Heo check_flush_dependency(wq, NULL); 2965fca839c0STejun Heo 29663c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 296773f53c4aSTejun Heo 296873f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 296973f53c4aSTejun Heo 297073f53c4aSTejun Heo /* 297173f53c4aSTejun Heo * Wake-up-and-cascade phase 297273f53c4aSTejun Heo * 297373f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 297473f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 297573f53c4aSTejun Heo */ 297600d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 297773f53c4aSTejun Heo return; 297873f53c4aSTejun Heo 29793c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 298073f53c4aSTejun Heo 29814ce48b37STejun Heo /* we might have raced, check again with mutex held */ 29824ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 29834ce48b37STejun Heo goto out_unlock; 29844ce48b37STejun Heo 298500d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 298673f53c4aSTejun Heo 29876183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 29886183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 298973f53c4aSTejun Heo 299073f53c4aSTejun Heo while (true) { 299173f53c4aSTejun Heo struct wq_flusher *next, *tmp; 299273f53c4aSTejun Heo 299373f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 299473f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 299573f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 299673f53c4aSTejun Heo break; 299773f53c4aSTejun Heo list_del_init(&next->list); 299873f53c4aSTejun Heo complete(&next->done); 299973f53c4aSTejun Heo } 300073f53c4aSTejun Heo 30016183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 300273f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 300373f53c4aSTejun Heo 300473f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 300573f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 300673f53c4aSTejun Heo 300773f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 300873f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 300973f53c4aSTejun Heo /* 301073f53c4aSTejun Heo * Assign the same color to all overflowed 301173f53c4aSTejun Heo * flushers, advance work_color and append to 301273f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 301373f53c4aSTejun Heo * phase for these overflowed flushers. 301473f53c4aSTejun Heo */ 301573f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 301673f53c4aSTejun Heo tmp->flush_color = wq->work_color; 301773f53c4aSTejun Heo 301873f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 301973f53c4aSTejun Heo 302073f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 302173f53c4aSTejun Heo &wq->flusher_queue); 3022112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 302373f53c4aSTejun Heo } 302473f53c4aSTejun Heo 302573f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 30266183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 302773f53c4aSTejun Heo break; 302873f53c4aSTejun Heo } 302973f53c4aSTejun Heo 303073f53c4aSTejun Heo /* 303173f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3032112202d9STejun Heo * the new first flusher and arm pwqs. 303373f53c4aSTejun Heo */ 30346183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 30356183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 303673f53c4aSTejun Heo 303773f53c4aSTejun Heo list_del_init(&next->list); 303873f53c4aSTejun Heo wq->first_flusher = next; 303973f53c4aSTejun Heo 3040112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 304173f53c4aSTejun Heo break; 304273f53c4aSTejun Heo 304373f53c4aSTejun Heo /* 304473f53c4aSTejun Heo * Meh... this color is already done, clear first 304573f53c4aSTejun Heo * flusher and repeat cascading. 304673f53c4aSTejun Heo */ 304773f53c4aSTejun Heo wq->first_flusher = NULL; 304873f53c4aSTejun Heo } 304973f53c4aSTejun Heo 305073f53c4aSTejun Heo out_unlock: 30513c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 30521da177e4SLinus Torvalds } 3053c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 30541da177e4SLinus Torvalds 30559c5a2ba7STejun Heo /** 30569c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 30579c5a2ba7STejun Heo * @wq: workqueue to drain 30589c5a2ba7STejun Heo * 30599c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 30609c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 30619c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3062b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 30639c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 30649c5a2ba7STejun Heo * takes too long. 30659c5a2ba7STejun Heo */ 30669c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 30679c5a2ba7STejun Heo { 30689c5a2ba7STejun Heo unsigned int flush_cnt = 0; 306949e3cf44STejun Heo struct pool_workqueue *pwq; 30709c5a2ba7STejun Heo 30719c5a2ba7STejun Heo /* 30729c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 30739c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3074618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 30759c5a2ba7STejun Heo */ 307687fc741eSLai Jiangshan mutex_lock(&wq->mutex); 30779c5a2ba7STejun Heo if (!wq->nr_drainers++) 3078618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 307987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 30809c5a2ba7STejun Heo reflush: 3081c4f135d6STetsuo Handa __flush_workqueue(wq); 30829c5a2ba7STejun Heo 3083b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 308476af4d93STejun Heo 308549e3cf44STejun Heo for_each_pwq(pwq, wq) { 3086fa2563e4SThomas Tuttle bool drained; 30879c5a2ba7STejun Heo 3088a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3089f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3090a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3091fa2563e4SThomas Tuttle 3092fa2563e4SThomas Tuttle if (drained) 30939c5a2ba7STejun Heo continue; 30949c5a2ba7STejun Heo 30959c5a2ba7STejun Heo if (++flush_cnt == 10 || 30969c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3097e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3098e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 309976af4d93STejun Heo 3100b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 31019c5a2ba7STejun Heo goto reflush; 31029c5a2ba7STejun Heo } 31039c5a2ba7STejun Heo 31049c5a2ba7STejun Heo if (!--wq->nr_drainers) 3105618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 310687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 31079c5a2ba7STejun Heo } 31089c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 31099c5a2ba7STejun Heo 3110d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3111d6e89786SJohannes Berg bool from_cancel) 3112baf59022STejun Heo { 3113baf59022STejun Heo struct worker *worker = NULL; 3114c9e7cf27STejun Heo struct worker_pool *pool; 3115112202d9STejun Heo struct pool_workqueue *pwq; 3116baf59022STejun Heo 3117baf59022STejun Heo might_sleep(); 3118baf59022STejun Heo 311924acfb71SThomas Gleixner rcu_read_lock(); 3120fa1b54e6STejun Heo pool = get_work_pool(work); 3121fa1b54e6STejun Heo if (!pool) { 312224acfb71SThomas Gleixner rcu_read_unlock(); 3123fa1b54e6STejun Heo return false; 3124fa1b54e6STejun Heo } 3125fa1b54e6STejun Heo 3126a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 31270b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3128112202d9STejun Heo pwq = get_work_pwq(work); 3129112202d9STejun Heo if (pwq) { 3130112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3131baf59022STejun Heo goto already_gone; 3132606a5020STejun Heo } else { 3133c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3134baf59022STejun Heo if (!worker) 3135baf59022STejun Heo goto already_gone; 3136112202d9STejun Heo pwq = worker->current_pwq; 3137606a5020STejun Heo } 3138baf59022STejun Heo 3139fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3140fca839c0STejun Heo 3141112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3142a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3143baf59022STejun Heo 3144e159489bSTejun Heo /* 3145a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3146a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3147a1d14934SPeter Zijlstra * 3148a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3149a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3150a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3151a1d14934SPeter Zijlstra * forward progress. 3152e159489bSTejun Heo */ 3153d6e89786SJohannes Berg if (!from_cancel && 3154d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3155112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3156112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3157a1d14934SPeter Zijlstra } 315824acfb71SThomas Gleixner rcu_read_unlock(); 3159baf59022STejun Heo return true; 3160baf59022STejun Heo already_gone: 3161a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 316224acfb71SThomas Gleixner rcu_read_unlock(); 3163baf59022STejun Heo return false; 3164baf59022STejun Heo } 3165baf59022STejun Heo 3166d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3167d6e89786SJohannes Berg { 3168d6e89786SJohannes Berg struct wq_barrier barr; 3169d6e89786SJohannes Berg 3170d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3171d6e89786SJohannes Berg return false; 3172d6e89786SJohannes Berg 31734d43d395STetsuo Handa if (WARN_ON(!work->func)) 31744d43d395STetsuo Handa return false; 31754d43d395STetsuo Handa 317687915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 317787915adcSJohannes Berg lock_map_release(&work->lockdep_map); 317887915adcSJohannes Berg 3179d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3180d6e89786SJohannes Berg wait_for_completion(&barr.done); 3181d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3182d6e89786SJohannes Berg return true; 3183d6e89786SJohannes Berg } else { 3184d6e89786SJohannes Berg return false; 3185d6e89786SJohannes Berg } 3186d6e89786SJohannes Berg } 3187d6e89786SJohannes Berg 3188db700897SOleg Nesterov /** 3189401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3190401a8d04STejun Heo * @work: the work to flush 3191db700897SOleg Nesterov * 3192606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3193606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3194401a8d04STejun Heo * 3195d185af30SYacine Belkadi * Return: 3196401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3197401a8d04STejun Heo * %false if it was already idle. 3198db700897SOleg Nesterov */ 3199401a8d04STejun Heo bool flush_work(struct work_struct *work) 3200db700897SOleg Nesterov { 3201d6e89786SJohannes Berg return __flush_work(work, false); 3202606a5020STejun Heo } 3203db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3204db700897SOleg Nesterov 32058603e1b3STejun Heo struct cwt_wait { 3206ac6424b9SIngo Molnar wait_queue_entry_t wait; 32078603e1b3STejun Heo struct work_struct *work; 32088603e1b3STejun Heo }; 32098603e1b3STejun Heo 3210ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 32118603e1b3STejun Heo { 32128603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 32138603e1b3STejun Heo 32148603e1b3STejun Heo if (cwait->work != key) 32158603e1b3STejun Heo return 0; 32168603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 32178603e1b3STejun Heo } 32188603e1b3STejun Heo 321936e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3220401a8d04STejun Heo { 32218603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3222bbb68dfaSTejun Heo unsigned long flags; 32231f1f642eSOleg Nesterov int ret; 32241f1f642eSOleg Nesterov 32251f1f642eSOleg Nesterov do { 3226bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3227bbb68dfaSTejun Heo /* 32288603e1b3STejun Heo * If someone else is already canceling, wait for it to 32298603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 32308603e1b3STejun Heo * because we may get scheduled between @work's completion 32318603e1b3STejun Heo * and the other canceling task resuming and clearing 32328603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 32338603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 32348603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 32358603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 32368603e1b3STejun Heo * we're hogging the CPU. 32378603e1b3STejun Heo * 32388603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 32398603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 32408603e1b3STejun Heo * wake function which matches @work along with exclusive 32418603e1b3STejun Heo * wait and wakeup. 3242bbb68dfaSTejun Heo */ 32438603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 32448603e1b3STejun Heo struct cwt_wait cwait; 32458603e1b3STejun Heo 32468603e1b3STejun Heo init_wait(&cwait.wait); 32478603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 32488603e1b3STejun Heo cwait.work = work; 32498603e1b3STejun Heo 32508603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 32518603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 32528603e1b3STejun Heo if (work_is_canceling(work)) 32538603e1b3STejun Heo schedule(); 32548603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 32558603e1b3STejun Heo } 32561f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 32571f1f642eSOleg Nesterov 3258bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3259bbb68dfaSTejun Heo mark_work_canceling(work); 3260bbb68dfaSTejun Heo local_irq_restore(flags); 3261bbb68dfaSTejun Heo 32623347fa09STejun Heo /* 32633347fa09STejun Heo * This allows canceling during early boot. We know that @work 32643347fa09STejun Heo * isn't executing. 32653347fa09STejun Heo */ 32663347fa09STejun Heo if (wq_online) 3267d6e89786SJohannes Berg __flush_work(work, true); 32683347fa09STejun Heo 32697a22ad75STejun Heo clear_work_data(work); 32708603e1b3STejun Heo 32718603e1b3STejun Heo /* 32728603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 32738603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 32748603e1b3STejun Heo * visible there. 32758603e1b3STejun Heo */ 32768603e1b3STejun Heo smp_mb(); 32778603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 32788603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 32798603e1b3STejun Heo 32801f1f642eSOleg Nesterov return ret; 32811f1f642eSOleg Nesterov } 32821f1f642eSOleg Nesterov 32836e84d644SOleg Nesterov /** 3284401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3285401a8d04STejun Heo * @work: the work to cancel 32866e84d644SOleg Nesterov * 3287401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3288401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3289401a8d04STejun Heo * another workqueue. On return from this function, @work is 3290401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 32911f1f642eSOleg Nesterov * 3292401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3293401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 32946e84d644SOleg Nesterov * 3295401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 32966e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3297401a8d04STejun Heo * 3298d185af30SYacine Belkadi * Return: 3299401a8d04STejun Heo * %true if @work was pending, %false otherwise. 33006e84d644SOleg Nesterov */ 3301401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 33026e84d644SOleg Nesterov { 330336e227d2STejun Heo return __cancel_work_timer(work, false); 3304b89deed3SOleg Nesterov } 330528e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3306b89deed3SOleg Nesterov 33076e84d644SOleg Nesterov /** 3308401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3309401a8d04STejun Heo * @dwork: the delayed work to flush 33106e84d644SOleg Nesterov * 3311401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3312401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3313401a8d04STejun Heo * considers the last queueing instance of @dwork. 33141f1f642eSOleg Nesterov * 3315d185af30SYacine Belkadi * Return: 3316401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3317401a8d04STejun Heo * %false if it was already idle. 33186e84d644SOleg Nesterov */ 3319401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3320401a8d04STejun Heo { 33218930cabaSTejun Heo local_irq_disable(); 3322401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 332360c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 33248930cabaSTejun Heo local_irq_enable(); 3325401a8d04STejun Heo return flush_work(&dwork->work); 3326401a8d04STejun Heo } 3327401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3328401a8d04STejun Heo 332905f0fe6bSTejun Heo /** 333005f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 333105f0fe6bSTejun Heo * @rwork: the rcu work to flush 333205f0fe6bSTejun Heo * 333305f0fe6bSTejun Heo * Return: 333405f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 333505f0fe6bSTejun Heo * %false if it was already idle. 333605f0fe6bSTejun Heo */ 333705f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 333805f0fe6bSTejun Heo { 333905f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 334005f0fe6bSTejun Heo rcu_barrier(); 334105f0fe6bSTejun Heo flush_work(&rwork->work); 334205f0fe6bSTejun Heo return true; 334305f0fe6bSTejun Heo } else { 334405f0fe6bSTejun Heo return flush_work(&rwork->work); 334505f0fe6bSTejun Heo } 334605f0fe6bSTejun Heo } 334705f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 334805f0fe6bSTejun Heo 3349f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3350f72b8792SJens Axboe { 3351f72b8792SJens Axboe unsigned long flags; 3352f72b8792SJens Axboe int ret; 3353f72b8792SJens Axboe 3354f72b8792SJens Axboe do { 3355f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3356f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3357f72b8792SJens Axboe 3358f72b8792SJens Axboe if (unlikely(ret < 0)) 3359f72b8792SJens Axboe return false; 3360f72b8792SJens Axboe 3361f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3362f72b8792SJens Axboe local_irq_restore(flags); 3363f72b8792SJens Axboe return ret; 3364f72b8792SJens Axboe } 3365f72b8792SJens Axboe 336673b4b532SAndrey Grodzovsky /* 336773b4b532SAndrey Grodzovsky * See cancel_delayed_work() 336873b4b532SAndrey Grodzovsky */ 336973b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 337073b4b532SAndrey Grodzovsky { 337173b4b532SAndrey Grodzovsky return __cancel_work(work, false); 337273b4b532SAndrey Grodzovsky } 337373b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 337473b4b532SAndrey Grodzovsky 3375401a8d04STejun Heo /** 337657b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 337757b30ae7STejun Heo * @dwork: delayed_work to cancel 337809383498STejun Heo * 3379d185af30SYacine Belkadi * Kill off a pending delayed_work. 3380d185af30SYacine Belkadi * 3381d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3382d185af30SYacine Belkadi * pending. 3383d185af30SYacine Belkadi * 3384d185af30SYacine Belkadi * Note: 3385d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3386d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3387d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 338809383498STejun Heo * 338957b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 339009383498STejun Heo */ 339157b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 339209383498STejun Heo { 3393f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 339409383498STejun Heo } 339557b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 339609383498STejun Heo 339709383498STejun Heo /** 3398401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3399401a8d04STejun Heo * @dwork: the delayed work cancel 3400401a8d04STejun Heo * 3401401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3402401a8d04STejun Heo * 3403d185af30SYacine Belkadi * Return: 3404401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3405401a8d04STejun Heo */ 3406401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 34076e84d644SOleg Nesterov { 340836e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 34096e84d644SOleg Nesterov } 3410f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 34111da177e4SLinus Torvalds 34120fcb78c2SRolf Eike Beer /** 341331ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3414b6136773SAndrew Morton * @func: the function to call 3415b6136773SAndrew Morton * 341631ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 341731ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3418b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 341931ddd871STejun Heo * 3420d185af30SYacine Belkadi * Return: 342131ddd871STejun Heo * 0 on success, -errno on failure. 3422b6136773SAndrew Morton */ 342365f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 342415316ba8SChristoph Lameter { 342515316ba8SChristoph Lameter int cpu; 342638f51568SNamhyung Kim struct work_struct __percpu *works; 342715316ba8SChristoph Lameter 3428b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3429b6136773SAndrew Morton if (!works) 343015316ba8SChristoph Lameter return -ENOMEM; 3431b6136773SAndrew Morton 3432ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 343393981800STejun Heo 343415316ba8SChristoph Lameter for_each_online_cpu(cpu) { 34359bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 34369bfb1839SIngo Molnar 34379bfb1839SIngo Molnar INIT_WORK(work, func); 34388de6d308SOleg Nesterov schedule_work_on(cpu, work); 343915316ba8SChristoph Lameter } 344093981800STejun Heo 344193981800STejun Heo for_each_online_cpu(cpu) 34428616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 344393981800STejun Heo 3444ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3445b6136773SAndrew Morton free_percpu(works); 344615316ba8SChristoph Lameter return 0; 344715316ba8SChristoph Lameter } 344815316ba8SChristoph Lameter 3449eef6a7d5SAlan Stern /** 34501fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 34511fa44ecaSJames Bottomley * @fn: the function to execute 34521fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 34531fa44ecaSJames Bottomley * be available when the work executes) 34541fa44ecaSJames Bottomley * 34551fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 34561fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 34571fa44ecaSJames Bottomley * 3458d185af30SYacine Belkadi * Return: 0 - function was executed 34591fa44ecaSJames Bottomley * 1 - function was scheduled for execution 34601fa44ecaSJames Bottomley */ 346165f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 34621fa44ecaSJames Bottomley { 34631fa44ecaSJames Bottomley if (!in_interrupt()) { 346465f27f38SDavid Howells fn(&ew->work); 34651fa44ecaSJames Bottomley return 0; 34661fa44ecaSJames Bottomley } 34671fa44ecaSJames Bottomley 346865f27f38SDavid Howells INIT_WORK(&ew->work, fn); 34691fa44ecaSJames Bottomley schedule_work(&ew->work); 34701fa44ecaSJames Bottomley 34711fa44ecaSJames Bottomley return 1; 34721fa44ecaSJames Bottomley } 34731fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 34741fa44ecaSJames Bottomley 34757a4e344cSTejun Heo /** 34767a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 34777a4e344cSTejun Heo * @attrs: workqueue_attrs to free 34787a4e344cSTejun Heo * 34797a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 34807a4e344cSTejun Heo */ 3481513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 34827a4e344cSTejun Heo { 34837a4e344cSTejun Heo if (attrs) { 34847a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 34857a4e344cSTejun Heo kfree(attrs); 34867a4e344cSTejun Heo } 34877a4e344cSTejun Heo } 34887a4e344cSTejun Heo 34897a4e344cSTejun Heo /** 34907a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 34917a4e344cSTejun Heo * 34927a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3493d185af30SYacine Belkadi * return it. 3494d185af30SYacine Belkadi * 3495d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 34967a4e344cSTejun Heo */ 3497513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 34987a4e344cSTejun Heo { 34997a4e344cSTejun Heo struct workqueue_attrs *attrs; 35007a4e344cSTejun Heo 3501be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 35027a4e344cSTejun Heo if (!attrs) 35037a4e344cSTejun Heo goto fail; 3504be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 35057a4e344cSTejun Heo goto fail; 35067a4e344cSTejun Heo 350713e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 35087a4e344cSTejun Heo return attrs; 35097a4e344cSTejun Heo fail: 35107a4e344cSTejun Heo free_workqueue_attrs(attrs); 35117a4e344cSTejun Heo return NULL; 35127a4e344cSTejun Heo } 35137a4e344cSTejun Heo 351429c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 351529c91e99STejun Heo const struct workqueue_attrs *from) 351629c91e99STejun Heo { 351729c91e99STejun Heo to->nice = from->nice; 351829c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 35192865a8fbSShaohua Li /* 35202865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 35212865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 35222865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 35232865a8fbSShaohua Li */ 35242865a8fbSShaohua Li to->no_numa = from->no_numa; 352529c91e99STejun Heo } 352629c91e99STejun Heo 352729c91e99STejun Heo /* hash value of the content of @attr */ 352829c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 352929c91e99STejun Heo { 353029c91e99STejun Heo u32 hash = 0; 353129c91e99STejun Heo 353229c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 353313e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 353413e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 353529c91e99STejun Heo return hash; 353629c91e99STejun Heo } 353729c91e99STejun Heo 353829c91e99STejun Heo /* content equality test */ 353929c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 354029c91e99STejun Heo const struct workqueue_attrs *b) 354129c91e99STejun Heo { 354229c91e99STejun Heo if (a->nice != b->nice) 354329c91e99STejun Heo return false; 354429c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 354529c91e99STejun Heo return false; 354629c91e99STejun Heo return true; 354729c91e99STejun Heo } 354829c91e99STejun Heo 35497a4e344cSTejun Heo /** 35507a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 35517a4e344cSTejun Heo * @pool: worker_pool to initialize 35527a4e344cSTejun Heo * 3553402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3554d185af30SYacine Belkadi * 3555d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 355629c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 355729c91e99STejun Heo * on @pool safely to release it. 35587a4e344cSTejun Heo */ 35597a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 35604e1a1f9aSTejun Heo { 3561a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 356229c91e99STejun Heo pool->id = -1; 356329c91e99STejun Heo pool->cpu = -1; 3564f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 35654e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 356682607adcSTejun Heo pool->watchdog_ts = jiffies; 35674e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 35684e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 35694e1a1f9aSTejun Heo hash_init(pool->busy_hash); 35704e1a1f9aSTejun Heo 357132a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 35723f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 35734e1a1f9aSTejun Heo 357432a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 35754e1a1f9aSTejun Heo 3576da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3577*e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 35787a4e344cSTejun Heo 35797cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 358029c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 358129c91e99STejun Heo pool->refcnt = 1; 358229c91e99STejun Heo 358329c91e99STejun Heo /* shouldn't fail above this point */ 3584be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 35857a4e344cSTejun Heo if (!pool->attrs) 35867a4e344cSTejun Heo return -ENOMEM; 35877a4e344cSTejun Heo return 0; 35884e1a1f9aSTejun Heo } 35894e1a1f9aSTejun Heo 3590669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3591669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3592669de8bdSBart Van Assche { 3593669de8bdSBart Van Assche char *lock_name; 3594669de8bdSBart Van Assche 3595669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3596669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3597669de8bdSBart Van Assche if (!lock_name) 3598669de8bdSBart Van Assche lock_name = wq->name; 359969a106c0SQian Cai 360069a106c0SQian Cai wq->lock_name = lock_name; 3601669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3602669de8bdSBart Van Assche } 3603669de8bdSBart Van Assche 3604669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3605669de8bdSBart Van Assche { 3606669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3607669de8bdSBart Van Assche } 3608669de8bdSBart Van Assche 3609669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3610669de8bdSBart Van Assche { 3611669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3612669de8bdSBart Van Assche kfree(wq->lock_name); 3613669de8bdSBart Van Assche } 3614669de8bdSBart Van Assche #else 3615669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3616669de8bdSBart Van Assche { 3617669de8bdSBart Van Assche } 3618669de8bdSBart Van Assche 3619669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3620669de8bdSBart Van Assche { 3621669de8bdSBart Van Assche } 3622669de8bdSBart Van Assche 3623669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3624669de8bdSBart Van Assche { 3625669de8bdSBart Van Assche } 3626669de8bdSBart Van Assche #endif 3627669de8bdSBart Van Assche 3628e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3629e2dca7adSTejun Heo { 3630e2dca7adSTejun Heo struct workqueue_struct *wq = 3631e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3632e2dca7adSTejun Heo 3633669de8bdSBart Van Assche wq_free_lockdep(wq); 3634669de8bdSBart Van Assche 3635e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3636e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3637e2dca7adSTejun Heo else 3638e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3639e2dca7adSTejun Heo 3640e2dca7adSTejun Heo kfree(wq); 3641e2dca7adSTejun Heo } 3642e2dca7adSTejun Heo 364329c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 364429c91e99STejun Heo { 364529c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 364629c91e99STejun Heo 36477cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 364829c91e99STejun Heo free_workqueue_attrs(pool->attrs); 364929c91e99STejun Heo kfree(pool); 365029c91e99STejun Heo } 365129c91e99STejun Heo 365229c91e99STejun Heo /** 365329c91e99STejun Heo * put_unbound_pool - put a worker_pool 365429c91e99STejun Heo * @pool: worker_pool to put 365529c91e99STejun Heo * 365624acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3657c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3658c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3659c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3660a892caccSTejun Heo * 3661a892caccSTejun Heo * Should be called with wq_pool_mutex held. 366229c91e99STejun Heo */ 366329c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 366429c91e99STejun Heo { 366560f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 3666*e02b9312SValentin Schneider struct list_head cull_list; 366729c91e99STejun Heo struct worker *worker; 366829c91e99STejun Heo 3669*e02b9312SValentin Schneider INIT_LIST_HEAD(&cull_list); 3670*e02b9312SValentin Schneider 3671a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3672a892caccSTejun Heo 3673a892caccSTejun Heo if (--pool->refcnt) 367429c91e99STejun Heo return; 367529c91e99STejun Heo 367629c91e99STejun Heo /* sanity checks */ 367761d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3678a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 367929c91e99STejun Heo return; 368029c91e99STejun Heo 368129c91e99STejun Heo /* release id and unhash */ 368229c91e99STejun Heo if (pool->id >= 0) 368329c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 368429c91e99STejun Heo hash_del(&pool->hash_node); 368529c91e99STejun Heo 3686c5aa87bbSTejun Heo /* 3687692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3688692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3689692b4825STejun Heo * manager and @pool gets freed with the flag set. 36909ab03be4SValentin Schneider * 36919ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 36929ab03be4SValentin Schneider * only get here with 36939ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 36949ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 36959ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 36969ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 36979ab03be4SValentin Schneider * drops pool->lock 3698c5aa87bbSTejun Heo */ 36999ab03be4SValentin Schneider while (true) { 37009ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 37019ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3702d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3703*e02b9312SValentin Schneider 3704*e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 37059ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 37069ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3707692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 37089ab03be4SValentin Schneider break; 37099ab03be4SValentin Schneider } 37109ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3711*e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 37129ab03be4SValentin Schneider } 3713692b4825STejun Heo 37141037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3715*e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 371629c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3717a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 371860f5a4bcSLai Jiangshan 3719*e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3720*e02b9312SValentin Schneider 3721*e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 372260f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 37231258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 372460f5a4bcSLai Jiangshan 372560f5a4bcSLai Jiangshan if (pool->detach_completion) 372660f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 372760f5a4bcSLai Jiangshan 372829c91e99STejun Heo /* shut down the timers */ 372929c91e99STejun Heo del_timer_sync(&pool->idle_timer); 37303f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 373129c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 373229c91e99STejun Heo 373324acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 373425b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 373529c91e99STejun Heo } 373629c91e99STejun Heo 373729c91e99STejun Heo /** 373829c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 373929c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 374029c91e99STejun Heo * 374129c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 374229c91e99STejun Heo * reference count and return it. If there already is a matching 374329c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3744d185af30SYacine Belkadi * create a new one. 3745a892caccSTejun Heo * 3746a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3747d185af30SYacine Belkadi * 3748d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3749d185af30SYacine Belkadi * On failure, %NULL. 375029c91e99STejun Heo */ 375129c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 375229c91e99STejun Heo { 375329c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 375429c91e99STejun Heo struct worker_pool *pool; 3755f3f90ad4STejun Heo int node; 3756e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 375729c91e99STejun Heo 3758a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 375929c91e99STejun Heo 376029c91e99STejun Heo /* do we already have a matching pool? */ 376129c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 376229c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 376329c91e99STejun Heo pool->refcnt++; 37643fb1823cSLai Jiangshan return pool; 376529c91e99STejun Heo } 376629c91e99STejun Heo } 376729c91e99STejun Heo 3768e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3769e2273584SXunlei Pang if (wq_numa_enabled) { 3770e2273584SXunlei Pang for_each_node(node) { 3771e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3772e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3773e2273584SXunlei Pang target_node = node; 3774e2273584SXunlei Pang break; 3775e2273584SXunlei Pang } 3776e2273584SXunlei Pang } 3777e2273584SXunlei Pang } 3778e2273584SXunlei Pang 377929c91e99STejun Heo /* nope, create a new one */ 3780e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 378129c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 378229c91e99STejun Heo goto fail; 378329c91e99STejun Heo 37848864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 378529c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3786e2273584SXunlei Pang pool->node = target_node; 378729c91e99STejun Heo 37882865a8fbSShaohua Li /* 37892865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 37902865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 37912865a8fbSShaohua Li */ 37922865a8fbSShaohua Li pool->attrs->no_numa = false; 37932865a8fbSShaohua Li 379429c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 379529c91e99STejun Heo goto fail; 379629c91e99STejun Heo 379729c91e99STejun Heo /* create and start the initial worker */ 37983347fa09STejun Heo if (wq_online && !create_worker(pool)) 379929c91e99STejun Heo goto fail; 380029c91e99STejun Heo 380129c91e99STejun Heo /* install */ 380229c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 38033fb1823cSLai Jiangshan 380429c91e99STejun Heo return pool; 380529c91e99STejun Heo fail: 380629c91e99STejun Heo if (pool) 380729c91e99STejun Heo put_unbound_pool(pool); 380829c91e99STejun Heo return NULL; 380929c91e99STejun Heo } 381029c91e99STejun Heo 38118864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 38128864b4e5STejun Heo { 38138864b4e5STejun Heo kmem_cache_free(pwq_cache, 38148864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 38158864b4e5STejun Heo } 38168864b4e5STejun Heo 38178864b4e5STejun Heo /* 38188864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 38198864b4e5STejun Heo * and needs to be destroyed. 38208864b4e5STejun Heo */ 38218864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 38228864b4e5STejun Heo { 38238864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 38248864b4e5STejun Heo unbound_release_work); 38258864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 38268864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3827b42b0bddSYang Yingliang bool is_last = false; 38288864b4e5STejun Heo 3829b42b0bddSYang Yingliang /* 3830b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3831b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3832b42b0bddSYang Yingliang */ 3833b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 38348864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 38358864b4e5STejun Heo return; 38368864b4e5STejun Heo 38373c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 38388864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3839bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 38403c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3841b42b0bddSYang Yingliang } 38428864b4e5STejun Heo 3843a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 38448864b4e5STejun Heo put_unbound_pool(pool); 3845a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3846a892caccSTejun Heo 384725b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 38488864b4e5STejun Heo 38498864b4e5STejun Heo /* 38508864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3851e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 38528864b4e5STejun Heo */ 3853669de8bdSBart Van Assche if (is_last) { 3854669de8bdSBart Van Assche wq_unregister_lockdep(wq); 385525b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 38566029a918STejun Heo } 3857669de8bdSBart Van Assche } 38588864b4e5STejun Heo 38590fbd95aaSTejun Heo /** 3860699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 38610fbd95aaSTejun Heo * @pwq: target pool_workqueue 38620fbd95aaSTejun Heo * 3863699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3864f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3865699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 38660fbd95aaSTejun Heo */ 3867699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 38680fbd95aaSTejun Heo { 3869699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3870699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 38713347fa09STejun Heo unsigned long flags; 3872699ce097STejun Heo 3873699ce097STejun Heo /* for @wq->saved_max_active */ 3874a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3875699ce097STejun Heo 3876699ce097STejun Heo /* fast exit for non-freezable wqs */ 3877699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3878699ce097STejun Heo return; 3879699ce097STejun Heo 38803347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3881a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3882699ce097STejun Heo 388374b414eaSLai Jiangshan /* 388474b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 388574b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 388674b414eaSLai Jiangshan * is updated and visible. 388774b414eaSLai Jiangshan */ 388874b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 388901341fbdSYunfeng Ye bool kick = false; 389001341fbdSYunfeng Ye 3891699ce097STejun Heo pwq->max_active = wq->saved_max_active; 38920fbd95aaSTejun Heo 3893f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 389401341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3895f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 389601341fbdSYunfeng Ye kick = true; 389701341fbdSYunfeng Ye } 3898951a078aSLai Jiangshan 3899951a078aSLai Jiangshan /* 3900951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 390101341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 390201341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 390301341fbdSYunfeng Ye * let's kick iff work items were activated. 3904951a078aSLai Jiangshan */ 390501341fbdSYunfeng Ye if (kick) 3906951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3907699ce097STejun Heo } else { 3908699ce097STejun Heo pwq->max_active = 0; 3909699ce097STejun Heo } 3910699ce097STejun Heo 3911a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 39120fbd95aaSTejun Heo } 39130fbd95aaSTejun Heo 391467dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3915f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3916f147f29eSTejun Heo struct worker_pool *pool) 3917d2c1d404STejun Heo { 3918d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3919d2c1d404STejun Heo 3920e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3921e50aba9aSTejun Heo 3922d2c1d404STejun Heo pwq->pool = pool; 3923d2c1d404STejun Heo pwq->wq = wq; 3924d2c1d404STejun Heo pwq->flush_color = -1; 39258864b4e5STejun Heo pwq->refcnt = 1; 3926f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 39271befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3928d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 39298864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3930f147f29eSTejun Heo } 3931d2c1d404STejun Heo 3932f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 39331befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3934f147f29eSTejun Heo { 3935f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3936f147f29eSTejun Heo 3937f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 393875ccf595STejun Heo 39391befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 39401befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 39411befcf30STejun Heo return; 39421befcf30STejun Heo 394329b1cb41SLai Jiangshan /* set the matching work_color */ 394475ccf595STejun Heo pwq->work_color = wq->work_color; 3945983ca25eSTejun Heo 3946983ca25eSTejun Heo /* sync max_active to the current setting */ 3947983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3948983ca25eSTejun Heo 3949983ca25eSTejun Heo /* link in @pwq */ 39509e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3951df2d5ae4STejun Heo } 39526029a918STejun Heo 3953f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3954f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3955f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3956f147f29eSTejun Heo { 3957f147f29eSTejun Heo struct worker_pool *pool; 3958f147f29eSTejun Heo struct pool_workqueue *pwq; 3959f147f29eSTejun Heo 3960f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3961f147f29eSTejun Heo 3962f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3963f147f29eSTejun Heo if (!pool) 3964f147f29eSTejun Heo return NULL; 3965f147f29eSTejun Heo 3966e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3967f147f29eSTejun Heo if (!pwq) { 3968f147f29eSTejun Heo put_unbound_pool(pool); 3969f147f29eSTejun Heo return NULL; 3970f147f29eSTejun Heo } 3971f147f29eSTejun Heo 3972f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3973f147f29eSTejun Heo return pwq; 3974d2c1d404STejun Heo } 3975d2c1d404STejun Heo 39764c16bd32STejun Heo /** 397730186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3978042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 39794c16bd32STejun Heo * @node: the target NUMA node 39804c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 39814c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 39824c16bd32STejun Heo * 39834c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 39844c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3985d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 39864c16bd32STejun Heo * 39874c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 39884c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 39894c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 39904c16bd32STejun Heo * @attrs->cpumask. 39914c16bd32STejun Heo * 39924c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 39934c16bd32STejun Heo * stable. 3994d185af30SYacine Belkadi * 3995d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3996d185af30SYacine Belkadi * %false if equal. 39974c16bd32STejun Heo */ 39984c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 39994c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 40004c16bd32STejun Heo { 4001d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 40024c16bd32STejun Heo goto use_dfl; 40034c16bd32STejun Heo 40044c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 40054c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 40064c16bd32STejun Heo if (cpu_going_down >= 0) 40074c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 40084c16bd32STejun Heo 40094c16bd32STejun Heo if (cpumask_empty(cpumask)) 40104c16bd32STejun Heo goto use_dfl; 40114c16bd32STejun Heo 40124c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 40134c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 40141ad0f0a7SMichael Bringmann 40151ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 40161ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 40171ad0f0a7SMichael Bringmann "possible intersect\n"); 40181ad0f0a7SMichael Bringmann return false; 40191ad0f0a7SMichael Bringmann } 40201ad0f0a7SMichael Bringmann 40214c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 40224c16bd32STejun Heo 40234c16bd32STejun Heo use_dfl: 40244c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 40254c16bd32STejun Heo return false; 40264c16bd32STejun Heo } 40274c16bd32STejun Heo 40281befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 40291befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 40301befcf30STejun Heo int node, 40311befcf30STejun Heo struct pool_workqueue *pwq) 40321befcf30STejun Heo { 40331befcf30STejun Heo struct pool_workqueue *old_pwq; 40341befcf30STejun Heo 40355b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 40361befcf30STejun Heo lockdep_assert_held(&wq->mutex); 40371befcf30STejun Heo 40381befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 40391befcf30STejun Heo link_pwq(pwq); 40401befcf30STejun Heo 40411befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 40421befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 40431befcf30STejun Heo return old_pwq; 40441befcf30STejun Heo } 40451befcf30STejun Heo 40462d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 40472d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 40482d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 40492d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4050042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 40512d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 40522d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 40532d5f0764SLai Jiangshan }; 40542d5f0764SLai Jiangshan 40552d5f0764SLai Jiangshan /* free the resources after success or abort */ 40562d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 40572d5f0764SLai Jiangshan { 40582d5f0764SLai Jiangshan if (ctx) { 40592d5f0764SLai Jiangshan int node; 40602d5f0764SLai Jiangshan 40612d5f0764SLai Jiangshan for_each_node(node) 40622d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 40632d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 40642d5f0764SLai Jiangshan 40652d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 40662d5f0764SLai Jiangshan 40672d5f0764SLai Jiangshan kfree(ctx); 40682d5f0764SLai Jiangshan } 40692d5f0764SLai Jiangshan } 40702d5f0764SLai Jiangshan 40712d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 40722d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 40732d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 407499c621efSLai Jiangshan const struct workqueue_attrs *attrs, 407599c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 40762d5f0764SLai Jiangshan { 40772d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 40782d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 40792d5f0764SLai Jiangshan int node; 40802d5f0764SLai Jiangshan 40812d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 40822d5f0764SLai Jiangshan 4083acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 40842d5f0764SLai Jiangshan 4085be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4086be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 40872d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 40882d5f0764SLai Jiangshan goto out_free; 40892d5f0764SLai Jiangshan 4090042f7df1SLai Jiangshan /* 409199c621efSLai Jiangshan * Calculate the attrs of the default pwq with unbound_cpumask 409299c621efSLai Jiangshan * which is wq_unbound_cpumask or to set to wq_unbound_cpumask. 4093042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 4094042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 4095042f7df1SLai Jiangshan */ 40962d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 409799c621efSLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, unbound_cpumask); 4098042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 409999c621efSLai Jiangshan cpumask_copy(new_attrs->cpumask, unbound_cpumask); 41002d5f0764SLai Jiangshan 41012d5f0764SLai Jiangshan /* 41022d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 41032d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 41042d5f0764SLai Jiangshan * pools. 41052d5f0764SLai Jiangshan */ 41062d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 41072d5f0764SLai Jiangshan 41082d5f0764SLai Jiangshan /* 41092d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 41102d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 41112d5f0764SLai Jiangshan * it even if we don't use it immediately. 41122d5f0764SLai Jiangshan */ 41132d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 41142d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 41152d5f0764SLai Jiangshan goto out_free; 41162d5f0764SLai Jiangshan 41172d5f0764SLai Jiangshan for_each_node(node) { 4118042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 41192d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 41202d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 41212d5f0764SLai Jiangshan goto out_free; 41222d5f0764SLai Jiangshan } else { 41232d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 41242d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 41252d5f0764SLai Jiangshan } 41262d5f0764SLai Jiangshan } 41272d5f0764SLai Jiangshan 4128042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4129042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4130042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 41312d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4132042f7df1SLai Jiangshan 41332d5f0764SLai Jiangshan ctx->wq = wq; 41342d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 41352d5f0764SLai Jiangshan return ctx; 41362d5f0764SLai Jiangshan 41372d5f0764SLai Jiangshan out_free: 41382d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 41392d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 41402d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 41412d5f0764SLai Jiangshan return NULL; 41422d5f0764SLai Jiangshan } 41432d5f0764SLai Jiangshan 41442d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 41452d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 41462d5f0764SLai Jiangshan { 41472d5f0764SLai Jiangshan int node; 41482d5f0764SLai Jiangshan 41492d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 41502d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 41512d5f0764SLai Jiangshan 41522d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 41532d5f0764SLai Jiangshan 41542d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 41552d5f0764SLai Jiangshan for_each_node(node) 41562d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 41572d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 41582d5f0764SLai Jiangshan 41592d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 41602d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 41612d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 41622d5f0764SLai Jiangshan 41632d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 41642d5f0764SLai Jiangshan } 41652d5f0764SLai Jiangshan 4166a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4167a0111cf6SLai Jiangshan { 4168a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4169ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4170a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4171a0111cf6SLai Jiangshan } 4172a0111cf6SLai Jiangshan 4173a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4174a0111cf6SLai Jiangshan { 4175a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4176ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4177a0111cf6SLai Jiangshan } 4178a0111cf6SLai Jiangshan 4179a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4180a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4181a0111cf6SLai Jiangshan { 4182a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4183a0111cf6SLai Jiangshan 4184a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4185a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4186a0111cf6SLai Jiangshan return -EINVAL; 4187a0111cf6SLai Jiangshan 4188a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 41890a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 41900a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4191a0111cf6SLai Jiangshan return -EINVAL; 4192a0111cf6SLai Jiangshan 41930a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 41940a94efb5STejun Heo } 41950a94efb5STejun Heo 419699c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 41976201171eSwanghaibin if (!ctx) 41986201171eSwanghaibin return -ENOMEM; 4199a0111cf6SLai Jiangshan 4200a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4201a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4202a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4203a0111cf6SLai Jiangshan 42046201171eSwanghaibin return 0; 4205a0111cf6SLai Jiangshan } 4206a0111cf6SLai Jiangshan 42079e8cd2f5STejun Heo /** 42089e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 42099e8cd2f5STejun Heo * @wq: the target workqueue 42109e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 42119e8cd2f5STejun Heo * 42124c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 42134c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 42144c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 42154c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 42164c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 42174c16bd32STejun Heo * back-to-back will stay on its current pwq. 42189e8cd2f5STejun Heo * 4219d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4220d185af30SYacine Belkadi * 4221ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4222509b3204SDaniel Jordan * 4223d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 42249e8cd2f5STejun Heo */ 4225513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 42269e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 42279e8cd2f5STejun Heo { 4228a0111cf6SLai Jiangshan int ret; 42299e8cd2f5STejun Heo 4230509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4231509b3204SDaniel Jordan 4232509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4233a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4234509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 42352d5f0764SLai Jiangshan 42362d5f0764SLai Jiangshan return ret; 42379e8cd2f5STejun Heo } 42389e8cd2f5STejun Heo 42394c16bd32STejun Heo /** 42404c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 42414c16bd32STejun Heo * @wq: the target workqueue 42424c16bd32STejun Heo * @cpu: the CPU coming up or going down 42434c16bd32STejun Heo * @online: whether @cpu is coming up or going down 42444c16bd32STejun Heo * 42454c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 42464c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 42474c16bd32STejun Heo * @wq accordingly. 42484c16bd32STejun Heo * 42494c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 42504c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 42514c16bd32STejun Heo * correct. 42524c16bd32STejun Heo * 42534c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 42544c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 42554c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 42564c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 42574c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 42584c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 42594c16bd32STejun Heo * CPU_DOWN_PREPARE. 42604c16bd32STejun Heo */ 42614c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 42624c16bd32STejun Heo bool online) 42634c16bd32STejun Heo { 42644c16bd32STejun Heo int node = cpu_to_node(cpu); 42654c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 42664c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 42674c16bd32STejun Heo struct workqueue_attrs *target_attrs; 42684c16bd32STejun Heo cpumask_t *cpumask; 42694c16bd32STejun Heo 42704c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 42714c16bd32STejun Heo 4272f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4273f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 42744c16bd32STejun Heo return; 42754c16bd32STejun Heo 42764c16bd32STejun Heo /* 42774c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 42784c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 42794c16bd32STejun Heo * CPU hotplug exclusion. 42804c16bd32STejun Heo */ 42814c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 42824c16bd32STejun Heo cpumask = target_attrs->cpumask; 42834c16bd32STejun Heo 42844c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 42854c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 42864c16bd32STejun Heo 42874c16bd32STejun Heo /* 42884c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4289042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4290042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4291042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 42924c16bd32STejun Heo */ 4293042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 42944c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4295f7142ed4SLai Jiangshan return; 42964c16bd32STejun Heo } else { 42974c16bd32STejun Heo goto use_dfl_pwq; 42984c16bd32STejun Heo } 42994c16bd32STejun Heo 43004c16bd32STejun Heo /* create a new pwq */ 43014c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 43024c16bd32STejun Heo if (!pwq) { 43032d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 43044c16bd32STejun Heo wq->name); 430577f300b1SDaeseok Youn goto use_dfl_pwq; 43064c16bd32STejun Heo } 43074c16bd32STejun Heo 4308f7142ed4SLai Jiangshan /* Install the new pwq. */ 43094c16bd32STejun Heo mutex_lock(&wq->mutex); 43104c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 43114c16bd32STejun Heo goto out_unlock; 43124c16bd32STejun Heo 43134c16bd32STejun Heo use_dfl_pwq: 4314f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4315a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 43164c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4317a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 43184c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 43194c16bd32STejun Heo out_unlock: 43204c16bd32STejun Heo mutex_unlock(&wq->mutex); 43214c16bd32STejun Heo put_pwq_unlocked(old_pwq); 43224c16bd32STejun Heo } 43234c16bd32STejun Heo 432430cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 43251da177e4SLinus Torvalds { 432649e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 43278a2b7538STejun Heo int cpu, ret; 4328e1d8aa9fSFrederic Weisbecker 432930cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4330420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4331420c0ddbSTejun Heo if (!wq->cpu_pwqs) 433230cdf249STejun Heo return -ENOMEM; 433330cdf249STejun Heo 433430cdf249STejun Heo for_each_possible_cpu(cpu) { 43357fb98ea7STejun Heo struct pool_workqueue *pwq = 43367fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 43377a62c2c8STejun Heo struct worker_pool *cpu_pools = 4338f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 433930cdf249STejun Heo 4340f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4341f147f29eSTejun Heo 4342f147f29eSTejun Heo mutex_lock(&wq->mutex); 43431befcf30STejun Heo link_pwq(pwq); 4344f147f29eSTejun Heo mutex_unlock(&wq->mutex); 434530cdf249STejun Heo } 434630cdf249STejun Heo return 0; 4347509b3204SDaniel Jordan } 4348509b3204SDaniel Jordan 4349ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4350509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 43518a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 43528a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 43538a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 43548a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 43558a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 43569e8cd2f5STejun Heo } else { 4357509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 43589e8cd2f5STejun Heo } 4359ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4360509b3204SDaniel Jordan 4361509b3204SDaniel Jordan return ret; 43620f900049STejun Heo } 43630f900049STejun Heo 4364f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4365f3421797STejun Heo const char *name) 4366b71ab8c2STejun Heo { 4367f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4368f3421797STejun Heo 4369f3421797STejun Heo if (max_active < 1 || max_active > lim) 4370044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4371f3421797STejun Heo max_active, name, 1, lim); 4372b71ab8c2STejun Heo 4373f3421797STejun Heo return clamp_val(max_active, 1, lim); 4374b71ab8c2STejun Heo } 4375b71ab8c2STejun Heo 4376983c7515STejun Heo /* 4377983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4378983c7515STejun Heo * to guarantee forward progress. 4379983c7515STejun Heo */ 4380983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4381983c7515STejun Heo { 4382983c7515STejun Heo struct worker *rescuer; 4383b92b36eaSDan Carpenter int ret; 4384983c7515STejun Heo 4385983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4386983c7515STejun Heo return 0; 4387983c7515STejun Heo 4388983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4389983c7515STejun Heo if (!rescuer) 4390983c7515STejun Heo return -ENOMEM; 4391983c7515STejun Heo 4392983c7515STejun Heo rescuer->rescue_wq = wq; 4393983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4394f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4395b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4396983c7515STejun Heo kfree(rescuer); 4397b92b36eaSDan Carpenter return ret; 4398983c7515STejun Heo } 4399983c7515STejun Heo 4400983c7515STejun Heo wq->rescuer = rescuer; 4401983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4402983c7515STejun Heo wake_up_process(rescuer->task); 4403983c7515STejun Heo 4404983c7515STejun Heo return 0; 4405983c7515STejun Heo } 4406983c7515STejun Heo 4407a2775bbcSMathieu Malaterre __printf(1, 4) 4408669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 440997e37d7bSTejun Heo unsigned int flags, 4410669de8bdSBart Van Assche int max_active, ...) 44113af24433SOleg Nesterov { 4412df2d5ae4STejun Heo size_t tbl_size = 0; 4413ecf6881fSTejun Heo va_list args; 44143af24433SOleg Nesterov struct workqueue_struct *wq; 441549e3cf44STejun Heo struct pool_workqueue *pwq; 4416b196be89STejun Heo 44175c0338c6STejun Heo /* 44185c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 44195c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 44205c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 44215c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 44225c0338c6STejun Heo * on NUMA. 44235c0338c6STejun Heo */ 44245c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 44255c0338c6STejun Heo flags |= __WQ_ORDERED; 44265c0338c6STejun Heo 4427cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4428cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4429cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4430cee22a15SViresh Kumar 4431ecf6881fSTejun Heo /* allocate wq and format name */ 4432df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4433ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4434df2d5ae4STejun Heo 4435df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4436b196be89STejun Heo if (!wq) 4437d2c1d404STejun Heo return NULL; 4438b196be89STejun Heo 44396029a918STejun Heo if (flags & WQ_UNBOUND) { 4440be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 44416029a918STejun Heo if (!wq->unbound_attrs) 44426029a918STejun Heo goto err_free_wq; 44436029a918STejun Heo } 44446029a918STejun Heo 4445669de8bdSBart Van Assche va_start(args, max_active); 4446ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4447b196be89STejun Heo va_end(args); 44483af24433SOleg Nesterov 4449d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4450b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 44513af24433SOleg Nesterov 4452b196be89STejun Heo /* init wq */ 445397e37d7bSTejun Heo wq->flags = flags; 4454a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 44553c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4456112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 445730cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 445873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 445973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4460493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 44613af24433SOleg Nesterov 4462669de8bdSBart Van Assche wq_init_lockdep(wq); 4463cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 44643af24433SOleg Nesterov 446530cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 446682efcab3SBart Van Assche goto err_unreg_lockdep; 44671537663fSTejun Heo 446840c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4469d2c1d404STejun Heo goto err_destroy; 4470e22bee78STejun Heo 4471226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4472226223abSTejun Heo goto err_destroy; 4473226223abSTejun Heo 44746af8bf3dSOleg Nesterov /* 447568e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 447668e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 447768e13a67SLai Jiangshan * list. 44786af8bf3dSOleg Nesterov */ 447968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4480a0a1a5fdSTejun Heo 4481a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 448249e3cf44STejun Heo for_each_pwq(pwq, wq) 4483699ce097STejun Heo pwq_adjust_max_active(pwq); 4484a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4485a0a1a5fdSTejun Heo 4486e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4487a0a1a5fdSTejun Heo 448868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44893af24433SOleg Nesterov 44903af24433SOleg Nesterov return wq; 4491d2c1d404STejun Heo 449282efcab3SBart Van Assche err_unreg_lockdep: 4493009bb421SBart Van Assche wq_unregister_lockdep(wq); 4494009bb421SBart Van Assche wq_free_lockdep(wq); 449582efcab3SBart Van Assche err_free_wq: 44966029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 44974690c4abSTejun Heo kfree(wq); 4498d2c1d404STejun Heo return NULL; 4499d2c1d404STejun Heo err_destroy: 4500d2c1d404STejun Heo destroy_workqueue(wq); 45014690c4abSTejun Heo return NULL; 45021da177e4SLinus Torvalds } 4503669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 45041da177e4SLinus Torvalds 4505c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4506c29eb853STejun Heo { 4507c29eb853STejun Heo int i; 4508c29eb853STejun Heo 4509c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4510c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4511c29eb853STejun Heo return true; 4512c29eb853STejun Heo 4513c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4514c29eb853STejun Heo return true; 4515f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4516c29eb853STejun Heo return true; 4517c29eb853STejun Heo 4518c29eb853STejun Heo return false; 4519c29eb853STejun Heo } 4520c29eb853STejun Heo 45213af24433SOleg Nesterov /** 45223af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 45233af24433SOleg Nesterov * @wq: target workqueue 45243af24433SOleg Nesterov * 45253af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 45263af24433SOleg Nesterov */ 45273af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 45283af24433SOleg Nesterov { 452949e3cf44STejun Heo struct pool_workqueue *pwq; 45304c16bd32STejun Heo int node; 45313af24433SOleg Nesterov 4532def98c84STejun Heo /* 4533def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4534def98c84STejun Heo * lead to sysfs name conflicts. 4535def98c84STejun Heo */ 4536def98c84STejun Heo workqueue_sysfs_unregister(wq); 4537def98c84STejun Heo 453833e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 453933e3f0a3SRichard Clark mutex_lock(&wq->mutex); 454033e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 454133e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 454233e3f0a3SRichard Clark 45439c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 45449c5a2ba7STejun Heo drain_workqueue(wq); 4545c8efcc25STejun Heo 4546def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4547def98c84STejun Heo if (wq->rescuer) { 4548def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4549def98c84STejun Heo 4550def98c84STejun Heo /* this prevents new queueing */ 4551a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4552def98c84STejun Heo wq->rescuer = NULL; 4553a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4554def98c84STejun Heo 4555def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4556def98c84STejun Heo kthread_stop(rescuer->task); 45578efe1223STejun Heo kfree(rescuer); 4558def98c84STejun Heo } 4559def98c84STejun Heo 4560c29eb853STejun Heo /* 4561c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4562c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4563c29eb853STejun Heo */ 4564c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4565b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 456649e3cf44STejun Heo for_each_pwq(pwq, wq) { 4567a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4568c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 45691d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4570e66b39afSTejun Heo __func__, wq->name); 4571c29eb853STejun Heo show_pwq(pwq); 4572a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4573b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4574c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 457555df0933SImran Khan show_one_workqueue(wq); 45766183c009STejun Heo return; 457776af4d93STejun Heo } 4578a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 457976af4d93STejun Heo } 4580b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 45816183c009STejun Heo 4582a0a1a5fdSTejun Heo /* 4583a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4584a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4585a0a1a5fdSTejun Heo */ 4586e2dca7adSTejun Heo list_del_rcu(&wq->list); 458768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 45883af24433SOleg Nesterov 45898864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4590669de8bdSBart Van Assche wq_unregister_lockdep(wq); 459129c91e99STejun Heo /* 45928864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4593e2dca7adSTejun Heo * schedule RCU free. 459429c91e99STejun Heo */ 459525b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 45968864b4e5STejun Heo } else { 45978864b4e5STejun Heo /* 45988864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 45994c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 46004c16bd32STejun Heo * @wq will be freed when the last pwq is released. 46018864b4e5STejun Heo */ 46024c16bd32STejun Heo for_each_node(node) { 46034c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 46044c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 46054c16bd32STejun Heo put_pwq_unlocked(pwq); 46064c16bd32STejun Heo } 46074c16bd32STejun Heo 46084c16bd32STejun Heo /* 46094c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 46104c16bd32STejun Heo * put. Don't access it afterwards. 46114c16bd32STejun Heo */ 46124c16bd32STejun Heo pwq = wq->dfl_pwq; 46134c16bd32STejun Heo wq->dfl_pwq = NULL; 4614dce90d47STejun Heo put_pwq_unlocked(pwq); 461529c91e99STejun Heo } 46163af24433SOleg Nesterov } 46173af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 46183af24433SOleg Nesterov 4619dcd989cbSTejun Heo /** 4620dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4621dcd989cbSTejun Heo * @wq: target workqueue 4622dcd989cbSTejun Heo * @max_active: new max_active value. 4623dcd989cbSTejun Heo * 4624dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4625dcd989cbSTejun Heo * 4626dcd989cbSTejun Heo * CONTEXT: 4627dcd989cbSTejun Heo * Don't call from IRQ context. 4628dcd989cbSTejun Heo */ 4629dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4630dcd989cbSTejun Heo { 463149e3cf44STejun Heo struct pool_workqueue *pwq; 4632dcd989cbSTejun Heo 46338719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 46340a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 46358719dceaSTejun Heo return; 46368719dceaSTejun Heo 4637f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4638dcd989cbSTejun Heo 4639a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4640dcd989cbSTejun Heo 46410a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4642dcd989cbSTejun Heo wq->saved_max_active = max_active; 4643dcd989cbSTejun Heo 4644699ce097STejun Heo for_each_pwq(pwq, wq) 4645699ce097STejun Heo pwq_adjust_max_active(pwq); 4646dcd989cbSTejun Heo 4647a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4648dcd989cbSTejun Heo } 4649dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4650dcd989cbSTejun Heo 4651dcd989cbSTejun Heo /** 465227d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 465327d4ee03SLukas Wunner * 465427d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 465527d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 465627d4ee03SLukas Wunner * 465727d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 465827d4ee03SLukas Wunner */ 465927d4ee03SLukas Wunner struct work_struct *current_work(void) 466027d4ee03SLukas Wunner { 466127d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 466227d4ee03SLukas Wunner 466327d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 466427d4ee03SLukas Wunner } 466527d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 466627d4ee03SLukas Wunner 466727d4ee03SLukas Wunner /** 4668e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4669e6267616STejun Heo * 4670e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4671e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4672d185af30SYacine Belkadi * 4673d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4674e6267616STejun Heo */ 4675e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4676e6267616STejun Heo { 4677e6267616STejun Heo struct worker *worker = current_wq_worker(); 4678e6267616STejun Heo 46796a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4680e6267616STejun Heo } 4681e6267616STejun Heo 4682e6267616STejun Heo /** 4683dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4684dcd989cbSTejun Heo * @cpu: CPU in question 4685dcd989cbSTejun Heo * @wq: target workqueue 4686dcd989cbSTejun Heo * 4687dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4688dcd989cbSTejun Heo * no synchronization around this function and the test result is 4689dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4690dcd989cbSTejun Heo * 4691d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4692d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4693d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4694d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4695d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4696d3251859STejun Heo * 4697d185af30SYacine Belkadi * Return: 4698dcd989cbSTejun Heo * %true if congested, %false otherwise. 4699dcd989cbSTejun Heo */ 4700d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4701dcd989cbSTejun Heo { 47027fb98ea7STejun Heo struct pool_workqueue *pwq; 470376af4d93STejun Heo bool ret; 470476af4d93STejun Heo 470524acfb71SThomas Gleixner rcu_read_lock(); 470624acfb71SThomas Gleixner preempt_disable(); 47077fb98ea7STejun Heo 4708d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4709d3251859STejun Heo cpu = smp_processor_id(); 4710d3251859STejun Heo 47117fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 47127fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 47137fb98ea7STejun Heo else 4714df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4715dcd989cbSTejun Heo 4716f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 471724acfb71SThomas Gleixner preempt_enable(); 471824acfb71SThomas Gleixner rcu_read_unlock(); 471976af4d93STejun Heo 472076af4d93STejun Heo return ret; 4721dcd989cbSTejun Heo } 4722dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4723dcd989cbSTejun Heo 4724dcd989cbSTejun Heo /** 4725dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4726dcd989cbSTejun Heo * @work: the work to be tested 4727dcd989cbSTejun Heo * 4728dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4729dcd989cbSTejun Heo * synchronization around this function and the test result is 4730dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4731dcd989cbSTejun Heo * 4732d185af30SYacine Belkadi * Return: 4733dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4734dcd989cbSTejun Heo */ 4735dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4736dcd989cbSTejun Heo { 4737fa1b54e6STejun Heo struct worker_pool *pool; 4738dcd989cbSTejun Heo unsigned long flags; 4739dcd989cbSTejun Heo unsigned int ret = 0; 4740dcd989cbSTejun Heo 4741dcd989cbSTejun Heo if (work_pending(work)) 4742dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4743038366c5SLai Jiangshan 474424acfb71SThomas Gleixner rcu_read_lock(); 4745fa1b54e6STejun Heo pool = get_work_pool(work); 4746038366c5SLai Jiangshan if (pool) { 4747a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4748c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4749dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4750a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4751038366c5SLai Jiangshan } 475224acfb71SThomas Gleixner rcu_read_unlock(); 4753dcd989cbSTejun Heo 4754dcd989cbSTejun Heo return ret; 4755dcd989cbSTejun Heo } 4756dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4757dcd989cbSTejun Heo 47583d1cb205STejun Heo /** 47593d1cb205STejun Heo * set_worker_desc - set description for the current work item 47603d1cb205STejun Heo * @fmt: printf-style format string 47613d1cb205STejun Heo * @...: arguments for the format string 47623d1cb205STejun Heo * 47633d1cb205STejun Heo * This function can be called by a running work function to describe what 47643d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 47653d1cb205STejun Heo * information will be printed out together to help debugging. The 47663d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 47673d1cb205STejun Heo */ 47683d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 47693d1cb205STejun Heo { 47703d1cb205STejun Heo struct worker *worker = current_wq_worker(); 47713d1cb205STejun Heo va_list args; 47723d1cb205STejun Heo 47733d1cb205STejun Heo if (worker) { 47743d1cb205STejun Heo va_start(args, fmt); 47753d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 47763d1cb205STejun Heo va_end(args); 47773d1cb205STejun Heo } 47783d1cb205STejun Heo } 47795c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 47803d1cb205STejun Heo 47813d1cb205STejun Heo /** 47823d1cb205STejun Heo * print_worker_info - print out worker information and description 47833d1cb205STejun Heo * @log_lvl: the log level to use when printing 47843d1cb205STejun Heo * @task: target task 47853d1cb205STejun Heo * 47863d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 47873d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 47883d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 47893d1cb205STejun Heo * 47903d1cb205STejun Heo * This function can be safely called on any task as long as the 47913d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 47923d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 47933d1cb205STejun Heo */ 47943d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 47953d1cb205STejun Heo { 47963d1cb205STejun Heo work_func_t *fn = NULL; 47973d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 47983d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 47993d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 48003d1cb205STejun Heo struct workqueue_struct *wq = NULL; 48013d1cb205STejun Heo struct worker *worker; 48023d1cb205STejun Heo 48033d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 48043d1cb205STejun Heo return; 48053d1cb205STejun Heo 48063d1cb205STejun Heo /* 48073d1cb205STejun Heo * This function is called without any synchronization and @task 48083d1cb205STejun Heo * could be in any state. Be careful with dereferences. 48093d1cb205STejun Heo */ 4810e700591aSPetr Mladek worker = kthread_probe_data(task); 48113d1cb205STejun Heo 48123d1cb205STejun Heo /* 48138bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 48148bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 48153d1cb205STejun Heo */ 4816fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4817fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4818fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4819fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4820fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 48213d1cb205STejun Heo 48223d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4823d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 48248bf89593STejun Heo if (strcmp(name, desc)) 48253d1cb205STejun Heo pr_cont(" (%s)", desc); 48263d1cb205STejun Heo pr_cont("\n"); 48273d1cb205STejun Heo } 48283d1cb205STejun Heo } 48293d1cb205STejun Heo 48303494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 48313494fc30STejun Heo { 48323494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 48333494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 48343494fc30STejun Heo pr_cont(" node=%d", pool->node); 48353494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 48363494fc30STejun Heo } 48373494fc30STejun Heo 4838c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 4839c76feb0dSPaul E. McKenney bool comma; 4840c76feb0dSPaul E. McKenney work_func_t func; 4841c76feb0dSPaul E. McKenney long ctr; 4842c76feb0dSPaul E. McKenney }; 4843c76feb0dSPaul E. McKenney 4844c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 4845c76feb0dSPaul E. McKenney { 4846c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 4847c76feb0dSPaul E. McKenney goto out_record; 4848c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 4849c76feb0dSPaul E. McKenney pcwsp->ctr++; 4850c76feb0dSPaul E. McKenney return; 4851c76feb0dSPaul E. McKenney } 4852c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 4853c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 4854c76feb0dSPaul E. McKenney else 4855c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 4856c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 4857c76feb0dSPaul E. McKenney out_record: 4858c76feb0dSPaul E. McKenney if ((long)func == -1L) 4859c76feb0dSPaul E. McKenney return; 4860c76feb0dSPaul E. McKenney pcwsp->comma = comma; 4861c76feb0dSPaul E. McKenney pcwsp->func = func; 4862c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 4863c76feb0dSPaul E. McKenney } 4864c76feb0dSPaul E. McKenney 4865c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 48663494fc30STejun Heo { 48673494fc30STejun Heo if (work->func == wq_barrier_func) { 48683494fc30STejun Heo struct wq_barrier *barr; 48693494fc30STejun Heo 48703494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 48713494fc30STejun Heo 4872c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 48733494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 48743494fc30STejun Heo task_pid_nr(barr->task)); 48753494fc30STejun Heo } else { 4876c76feb0dSPaul E. McKenney if (!comma) 4877c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 4878c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 48793494fc30STejun Heo } 48803494fc30STejun Heo } 48813494fc30STejun Heo 48823494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 48833494fc30STejun Heo { 4884c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 48853494fc30STejun Heo struct worker_pool *pool = pwq->pool; 48863494fc30STejun Heo struct work_struct *work; 48873494fc30STejun Heo struct worker *worker; 48883494fc30STejun Heo bool has_in_flight = false, has_pending = false; 48893494fc30STejun Heo int bkt; 48903494fc30STejun Heo 48913494fc30STejun Heo pr_info(" pwq %d:", pool->id); 48923494fc30STejun Heo pr_cont_pool_info(pool); 48933494fc30STejun Heo 4894e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4895e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 48963494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 48973494fc30STejun Heo 48983494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 48993494fc30STejun Heo if (worker->current_pwq == pwq) { 49003494fc30STejun Heo has_in_flight = true; 49013494fc30STejun Heo break; 49023494fc30STejun Heo } 49033494fc30STejun Heo } 49043494fc30STejun Heo if (has_in_flight) { 49053494fc30STejun Heo bool comma = false; 49063494fc30STejun Heo 49073494fc30STejun Heo pr_info(" in-flight:"); 49083494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 49093494fc30STejun Heo if (worker->current_pwq != pwq) 49103494fc30STejun Heo continue; 49113494fc30STejun Heo 4912d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 49133494fc30STejun Heo task_pid_nr(worker->task), 491430ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 49153494fc30STejun Heo worker->current_func); 49163494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 4917c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 4918c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 49193494fc30STejun Heo comma = true; 49203494fc30STejun Heo } 49213494fc30STejun Heo pr_cont("\n"); 49223494fc30STejun Heo } 49233494fc30STejun Heo 49243494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 49253494fc30STejun Heo if (get_work_pwq(work) == pwq) { 49263494fc30STejun Heo has_pending = true; 49273494fc30STejun Heo break; 49283494fc30STejun Heo } 49293494fc30STejun Heo } 49303494fc30STejun Heo if (has_pending) { 49313494fc30STejun Heo bool comma = false; 49323494fc30STejun Heo 49333494fc30STejun Heo pr_info(" pending:"); 49343494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 49353494fc30STejun Heo if (get_work_pwq(work) != pwq) 49363494fc30STejun Heo continue; 49373494fc30STejun Heo 4938c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 49393494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 49403494fc30STejun Heo } 4941c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 49423494fc30STejun Heo pr_cont("\n"); 49433494fc30STejun Heo } 49443494fc30STejun Heo 4945f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 49463494fc30STejun Heo bool comma = false; 49473494fc30STejun Heo 4948f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4949f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 4950c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 49513494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 49523494fc30STejun Heo } 4953c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 49543494fc30STejun Heo pr_cont("\n"); 49553494fc30STejun Heo } 49563494fc30STejun Heo } 49573494fc30STejun Heo 49583494fc30STejun Heo /** 495955df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 496055df0933SImran Khan * @wq: workqueue whose state will be printed 49613494fc30STejun Heo */ 496255df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 49633494fc30STejun Heo { 49643494fc30STejun Heo struct pool_workqueue *pwq; 49653494fc30STejun Heo bool idle = true; 496655df0933SImran Khan unsigned long flags; 49673494fc30STejun Heo 49683494fc30STejun Heo for_each_pwq(pwq, wq) { 4969f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 49703494fc30STejun Heo idle = false; 49713494fc30STejun Heo break; 49723494fc30STejun Heo } 49733494fc30STejun Heo } 497455df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 497555df0933SImran Khan return; 49763494fc30STejun Heo 49773494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 49783494fc30STejun Heo 49793494fc30STejun Heo for_each_pwq(pwq, wq) { 4980a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 498157116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 498257116ce1SJohan Hovold /* 498357116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 498457116ce1SJohan Hovold * drivers that queue work while holding locks 498557116ce1SJohan Hovold * also taken in their write paths. 498657116ce1SJohan Hovold */ 498757116ce1SJohan Hovold printk_deferred_enter(); 49883494fc30STejun Heo show_pwq(pwq); 498957116ce1SJohan Hovold printk_deferred_exit(); 499057116ce1SJohan Hovold } 4991a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 499262635ea8SSergey Senozhatsky /* 499362635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 499455df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 499562635ea8SSergey Senozhatsky * hard lockup. 499662635ea8SSergey Senozhatsky */ 499762635ea8SSergey Senozhatsky touch_nmi_watchdog(); 49983494fc30STejun Heo } 499955df0933SImran Khan 50003494fc30STejun Heo } 50013494fc30STejun Heo 500255df0933SImran Khan /** 500355df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 500455df0933SImran Khan * @pool: worker pool whose state will be printed 500555df0933SImran Khan */ 500655df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 500755df0933SImran Khan { 50083494fc30STejun Heo struct worker *worker; 50093494fc30STejun Heo bool first = true; 501055df0933SImran Khan unsigned long flags; 50113494fc30STejun Heo 5012a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 50133494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 50143494fc30STejun Heo goto next_pool; 501557116ce1SJohan Hovold /* 501657116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 501757116ce1SJohan Hovold * queue work while holding locks also taken in their write 501857116ce1SJohan Hovold * paths. 501957116ce1SJohan Hovold */ 502057116ce1SJohan Hovold printk_deferred_enter(); 50213494fc30STejun Heo pr_info("pool %d:", pool->id); 50223494fc30STejun Heo pr_cont_pool_info(pool); 502382607adcSTejun Heo pr_cont(" hung=%us workers=%d", 502482607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 502582607adcSTejun Heo pool->nr_workers); 50263494fc30STejun Heo if (pool->manager) 50273494fc30STejun Heo pr_cont(" manager: %d", 50283494fc30STejun Heo task_pid_nr(pool->manager->task)); 50293494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 50303494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 50313494fc30STejun Heo task_pid_nr(worker->task)); 50323494fc30STejun Heo first = false; 50333494fc30STejun Heo } 50343494fc30STejun Heo pr_cont("\n"); 503557116ce1SJohan Hovold printk_deferred_exit(); 50363494fc30STejun Heo next_pool: 5037a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 503862635ea8SSergey Senozhatsky /* 503962635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 504055df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 504162635ea8SSergey Senozhatsky * hard lockup. 504262635ea8SSergey Senozhatsky */ 504362635ea8SSergey Senozhatsky touch_nmi_watchdog(); 504455df0933SImran Khan 50453494fc30STejun Heo } 50463494fc30STejun Heo 504755df0933SImran Khan /** 504855df0933SImran Khan * show_all_workqueues - dump workqueue state 504955df0933SImran Khan * 505055df0933SImran Khan * Called from a sysrq handler or try_to_freeze_tasks() and prints out 505155df0933SImran Khan * all busy workqueues and pools. 505255df0933SImran Khan */ 505355df0933SImran Khan void show_all_workqueues(void) 505455df0933SImran Khan { 505555df0933SImran Khan struct workqueue_struct *wq; 505655df0933SImran Khan struct worker_pool *pool; 505755df0933SImran Khan int pi; 505855df0933SImran Khan 505955df0933SImran Khan rcu_read_lock(); 506055df0933SImran Khan 506155df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 506255df0933SImran Khan 506355df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 506455df0933SImran Khan show_one_workqueue(wq); 506555df0933SImran Khan 506655df0933SImran Khan for_each_pool(pool, pi) 506755df0933SImran Khan show_one_worker_pool(pool); 506855df0933SImran Khan 506924acfb71SThomas Gleixner rcu_read_unlock(); 50703494fc30STejun Heo } 50713494fc30STejun Heo 50726b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 50736b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 50746b59808bSTejun Heo { 50756b59808bSTejun Heo int off; 50766b59808bSTejun Heo 50776b59808bSTejun Heo /* always show the actual comm */ 50786b59808bSTejun Heo off = strscpy(buf, task->comm, size); 50796b59808bSTejun Heo if (off < 0) 50806b59808bSTejun Heo return; 50816b59808bSTejun Heo 5082197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 50836b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 50846b59808bSTejun Heo 5085197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5086197f6accSTejun Heo struct worker *worker = kthread_data(task); 5087197f6accSTejun Heo struct worker_pool *pool = worker->pool; 50886b59808bSTejun Heo 50896b59808bSTejun Heo if (pool) { 5090a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 50916b59808bSTejun Heo /* 5092197f6accSTejun Heo * ->desc tracks information (wq name or 5093197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5094197f6accSTejun Heo * current, prepend '+', otherwise '-'. 50956b59808bSTejun Heo */ 50966b59808bSTejun Heo if (worker->desc[0] != '\0') { 50976b59808bSTejun Heo if (worker->current_work) 50986b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 50996b59808bSTejun Heo worker->desc); 51006b59808bSTejun Heo else 51016b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 51026b59808bSTejun Heo worker->desc); 51036b59808bSTejun Heo } 5104a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 51056b59808bSTejun Heo } 5106197f6accSTejun Heo } 51076b59808bSTejun Heo 51086b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 51096b59808bSTejun Heo } 51106b59808bSTejun Heo 511166448bc2SMathieu Malaterre #ifdef CONFIG_SMP 511266448bc2SMathieu Malaterre 5113db7bccf4STejun Heo /* 5114db7bccf4STejun Heo * CPU hotplug. 5115db7bccf4STejun Heo * 5116e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5117112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5118706026c2STejun Heo * pool which make migrating pending and scheduled works very 5119e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 512094cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5121e22bee78STejun Heo * blocked draining impractical. 5122e22bee78STejun Heo * 512324647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5124628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5125628c78e7STejun Heo * cpu comes back online. 5126db7bccf4STejun Heo */ 5127db7bccf4STejun Heo 5128e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5129db7bccf4STejun Heo { 51304ce62e9eSTejun Heo struct worker_pool *pool; 5131db7bccf4STejun Heo struct worker *worker; 5132db7bccf4STejun Heo 5133f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51341258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5135a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5136e22bee78STejun Heo 5137f2d5a0eeSTejun Heo /* 513892f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 513994cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 514011b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5141b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5142b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5143b4ac9384SLai Jiangshan * is on the same cpu. 5144f2d5a0eeSTejun Heo */ 5145da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5146403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5147db7bccf4STejun Heo 514824647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5149f2d5a0eeSTejun Heo 5150e22bee78STejun Heo /* 5151989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5152989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5153989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5154989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5155989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5156eb283428SLai Jiangshan * are served by workers tied to the pool. 5157e22bee78STejun Heo */ 5158bc35f7efSLai Jiangshan pool->nr_running = 0; 5159eb283428SLai Jiangshan 5160eb283428SLai Jiangshan /* 5161eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5162eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5163eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5164eb283428SLai Jiangshan */ 5165eb283428SLai Jiangshan wake_up_worker(pool); 5166989442d7SLai Jiangshan 5167a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5168989442d7SLai Jiangshan 5169793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5170793777bcSValentin Schneider unbind_worker(worker); 5171989442d7SLai Jiangshan 5172989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5173eb283428SLai Jiangshan } 5174db7bccf4STejun Heo } 5175db7bccf4STejun Heo 5176bd7c089eSTejun Heo /** 5177bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5178bd7c089eSTejun Heo * @pool: pool of interest 5179bd7c089eSTejun Heo * 5180a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5181bd7c089eSTejun Heo */ 5182bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5183bd7c089eSTejun Heo { 5184a9ab775bSTejun Heo struct worker *worker; 5185bd7c089eSTejun Heo 51861258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5187bd7c089eSTejun Heo 5188bd7c089eSTejun Heo /* 5189a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5190a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5191402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5192a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5193a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5194bd7c089eSTejun Heo */ 5195793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5196793777bcSValentin Schneider rebind_worker(worker, pool); 5197a9ab775bSTejun Heo 5198a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5199f7c17d26SWanpeng Li 52003de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5201a9ab775bSTejun Heo 5202da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5203a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5204a9ab775bSTejun Heo 5205a9ab775bSTejun Heo /* 5206a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5207a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5208a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5209a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5210a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5211a9ab775bSTejun Heo * concurrency management. Note that when or whether 5212a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5213a9ab775bSTejun Heo * 5214c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5215a9ab775bSTejun Heo * tested without holding any lock in 52166d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5217a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5218a9ab775bSTejun Heo * management operations. 5219bd7c089eSTejun Heo */ 5220a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5221a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5222a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5223c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5224bd7c089eSTejun Heo } 5225a9ab775bSTejun Heo 5226a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5227bd7c089eSTejun Heo } 5228bd7c089eSTejun Heo 52297dbc725eSTejun Heo /** 52307dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 52317dbc725eSTejun Heo * @pool: unbound pool of interest 52327dbc725eSTejun Heo * @cpu: the CPU which is coming up 52337dbc725eSTejun Heo * 52347dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 52357dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 52367dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 52377dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 52387dbc725eSTejun Heo */ 52397dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 52407dbc725eSTejun Heo { 52417dbc725eSTejun Heo static cpumask_t cpumask; 52427dbc725eSTejun Heo struct worker *worker; 52437dbc725eSTejun Heo 52441258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 52457dbc725eSTejun Heo 52467dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 52477dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 52487dbc725eSTejun Heo return; 52497dbc725eSTejun Heo 52507dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 52517dbc725eSTejun Heo 52527dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5253da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5254d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 52557dbc725eSTejun Heo } 52567dbc725eSTejun Heo 52577ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 52581da177e4SLinus Torvalds { 52594ce62e9eSTejun Heo struct worker_pool *pool; 52601da177e4SLinus Torvalds 5261f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 52623ce63377STejun Heo if (pool->nr_workers) 52633ce63377STejun Heo continue; 5264051e1850SLai Jiangshan if (!create_worker(pool)) 52657ee681b2SThomas Gleixner return -ENOMEM; 52663af24433SOleg Nesterov } 52677ee681b2SThomas Gleixner return 0; 52687ee681b2SThomas Gleixner } 52691da177e4SLinus Torvalds 52707ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 52717ee681b2SThomas Gleixner { 52727ee681b2SThomas Gleixner struct worker_pool *pool; 52737ee681b2SThomas Gleixner struct workqueue_struct *wq; 52747ee681b2SThomas Gleixner int pi; 52757ee681b2SThomas Gleixner 527668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 52777dbc725eSTejun Heo 52787dbc725eSTejun Heo for_each_pool(pool, pi) { 52791258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 528094cf58bbSTejun Heo 5281f05b558dSLai Jiangshan if (pool->cpu == cpu) 528294cf58bbSTejun Heo rebind_workers(pool); 5283f05b558dSLai Jiangshan else if (pool->cpu < 0) 52847dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 528594cf58bbSTejun Heo 52861258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 528794cf58bbSTejun Heo } 52887dbc725eSTejun Heo 52894c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 52904c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 52914c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 52924c16bd32STejun Heo 529368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 52947ee681b2SThomas Gleixner return 0; 529565758202STejun Heo } 529665758202STejun Heo 52977ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 529865758202STejun Heo { 52994c16bd32STejun Heo struct workqueue_struct *wq; 53008db25e78STejun Heo 53014c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5302e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5303e8b3f8dbSLai Jiangshan return -1; 5304e8b3f8dbSLai Jiangshan 5305e8b3f8dbSLai Jiangshan unbind_workers(cpu); 53064c16bd32STejun Heo 53074c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 53084c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 53094c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 53104c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 53114c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 53124c16bd32STejun Heo 53137ee681b2SThomas Gleixner return 0; 531465758202STejun Heo } 531565758202STejun Heo 53162d3854a3SRusty Russell struct work_for_cpu { 5317ed48ece2STejun Heo struct work_struct work; 53182d3854a3SRusty Russell long (*fn)(void *); 53192d3854a3SRusty Russell void *arg; 53202d3854a3SRusty Russell long ret; 53212d3854a3SRusty Russell }; 53222d3854a3SRusty Russell 5323ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 53242d3854a3SRusty Russell { 5325ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5326ed48ece2STejun Heo 53272d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 53282d3854a3SRusty Russell } 53292d3854a3SRusty Russell 53302d3854a3SRusty Russell /** 533122aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 53322d3854a3SRusty Russell * @cpu: the cpu to run on 53332d3854a3SRusty Russell * @fn: the function to run 53342d3854a3SRusty Russell * @arg: the function arg 53352d3854a3SRusty Russell * 533631ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 53376b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5338d185af30SYacine Belkadi * 5339d185af30SYacine Belkadi * Return: The value @fn returns. 53402d3854a3SRusty Russell */ 5341d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 53422d3854a3SRusty Russell { 5343ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 53442d3854a3SRusty Russell 5345ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5346ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 534712997d1aSBjorn Helgaas flush_work(&wfc.work); 5348440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 53492d3854a3SRusty Russell return wfc.ret; 53502d3854a3SRusty Russell } 53512d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 53520e8d6a93SThomas Gleixner 53530e8d6a93SThomas Gleixner /** 53540e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 53550e8d6a93SThomas Gleixner * @cpu: the cpu to run on 53560e8d6a93SThomas Gleixner * @fn: the function to run 53570e8d6a93SThomas Gleixner * @arg: the function argument 53580e8d6a93SThomas Gleixner * 53590e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 53600e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 53610e8d6a93SThomas Gleixner * 53620e8d6a93SThomas Gleixner * Return: The value @fn returns. 53630e8d6a93SThomas Gleixner */ 53640e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 53650e8d6a93SThomas Gleixner { 53660e8d6a93SThomas Gleixner long ret = -ENODEV; 53670e8d6a93SThomas Gleixner 5368ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 53690e8d6a93SThomas Gleixner if (cpu_online(cpu)) 53700e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5371ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 53720e8d6a93SThomas Gleixner return ret; 53730e8d6a93SThomas Gleixner } 53740e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 53752d3854a3SRusty Russell #endif /* CONFIG_SMP */ 53762d3854a3SRusty Russell 5377a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5378e7577c50SRusty Russell 5379a0a1a5fdSTejun Heo /** 5380a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5381a0a1a5fdSTejun Heo * 538258a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5383f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5384706026c2STejun Heo * pool->worklist. 5385a0a1a5fdSTejun Heo * 5386a0a1a5fdSTejun Heo * CONTEXT: 5387a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5388a0a1a5fdSTejun Heo */ 5389a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5390a0a1a5fdSTejun Heo { 539124b8a847STejun Heo struct workqueue_struct *wq; 539224b8a847STejun Heo struct pool_workqueue *pwq; 5393a0a1a5fdSTejun Heo 539468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5395a0a1a5fdSTejun Heo 53966183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5397a0a1a5fdSTejun Heo workqueue_freezing = true; 5398a0a1a5fdSTejun Heo 539924b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5400a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5401699ce097STejun Heo for_each_pwq(pwq, wq) 5402699ce097STejun Heo pwq_adjust_max_active(pwq); 5403a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5404a1056305STejun Heo } 54055bcab335STejun Heo 540668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5407a0a1a5fdSTejun Heo } 5408a0a1a5fdSTejun Heo 5409a0a1a5fdSTejun Heo /** 541058a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5411a0a1a5fdSTejun Heo * 5412a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5413a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5414a0a1a5fdSTejun Heo * 5415a0a1a5fdSTejun Heo * CONTEXT: 541668e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5417a0a1a5fdSTejun Heo * 5418d185af30SYacine Belkadi * Return: 541958a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 542058a69cb4STejun Heo * is complete. 5421a0a1a5fdSTejun Heo */ 5422a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5423a0a1a5fdSTejun Heo { 5424a0a1a5fdSTejun Heo bool busy = false; 542524b8a847STejun Heo struct workqueue_struct *wq; 542624b8a847STejun Heo struct pool_workqueue *pwq; 5427a0a1a5fdSTejun Heo 542868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5429a0a1a5fdSTejun Heo 54306183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5431a0a1a5fdSTejun Heo 543224b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 543324b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 543424b8a847STejun Heo continue; 5435a0a1a5fdSTejun Heo /* 5436a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5437a0a1a5fdSTejun Heo * to peek without lock. 5438a0a1a5fdSTejun Heo */ 543924acfb71SThomas Gleixner rcu_read_lock(); 544024b8a847STejun Heo for_each_pwq(pwq, wq) { 54416183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5442112202d9STejun Heo if (pwq->nr_active) { 5443a0a1a5fdSTejun Heo busy = true; 544424acfb71SThomas Gleixner rcu_read_unlock(); 5445a0a1a5fdSTejun Heo goto out_unlock; 5446a0a1a5fdSTejun Heo } 5447a0a1a5fdSTejun Heo } 544824acfb71SThomas Gleixner rcu_read_unlock(); 5449a0a1a5fdSTejun Heo } 5450a0a1a5fdSTejun Heo out_unlock: 545168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5452a0a1a5fdSTejun Heo return busy; 5453a0a1a5fdSTejun Heo } 5454a0a1a5fdSTejun Heo 5455a0a1a5fdSTejun Heo /** 5456a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5457a0a1a5fdSTejun Heo * 5458a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5459706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5460a0a1a5fdSTejun Heo * 5461a0a1a5fdSTejun Heo * CONTEXT: 5462a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5463a0a1a5fdSTejun Heo */ 5464a0a1a5fdSTejun Heo void thaw_workqueues(void) 5465a0a1a5fdSTejun Heo { 546624b8a847STejun Heo struct workqueue_struct *wq; 546724b8a847STejun Heo struct pool_workqueue *pwq; 5468a0a1a5fdSTejun Heo 546968e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5470a0a1a5fdSTejun Heo 5471a0a1a5fdSTejun Heo if (!workqueue_freezing) 5472a0a1a5fdSTejun Heo goto out_unlock; 5473a0a1a5fdSTejun Heo 547474b414eaSLai Jiangshan workqueue_freezing = false; 547524b8a847STejun Heo 547624b8a847STejun Heo /* restore max_active and repopulate worklist */ 547724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5478a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5479699ce097STejun Heo for_each_pwq(pwq, wq) 5480699ce097STejun Heo pwq_adjust_max_active(pwq); 5481a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 548224b8a847STejun Heo } 548324b8a847STejun Heo 5484a0a1a5fdSTejun Heo out_unlock: 548568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5486a0a1a5fdSTejun Heo } 5487a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5488a0a1a5fdSTejun Heo 548999c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5490042f7df1SLai Jiangshan { 5491042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5492042f7df1SLai Jiangshan int ret = 0; 5493042f7df1SLai Jiangshan struct workqueue_struct *wq; 5494042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5495042f7df1SLai Jiangshan 5496042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5497042f7df1SLai Jiangshan 5498042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5499042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5500042f7df1SLai Jiangshan continue; 5501042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5502042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5503042f7df1SLai Jiangshan continue; 5504042f7df1SLai Jiangshan 550599c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 5506042f7df1SLai Jiangshan if (!ctx) { 5507042f7df1SLai Jiangshan ret = -ENOMEM; 5508042f7df1SLai Jiangshan break; 5509042f7df1SLai Jiangshan } 5510042f7df1SLai Jiangshan 5511042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5512042f7df1SLai Jiangshan } 5513042f7df1SLai Jiangshan 5514042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5515042f7df1SLai Jiangshan if (!ret) 5516042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5517042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5518042f7df1SLai Jiangshan } 5519042f7df1SLai Jiangshan 552099c621efSLai Jiangshan if (!ret) { 552199c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 552299c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 552399c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 552499c621efSLai Jiangshan } 5525042f7df1SLai Jiangshan return ret; 5526042f7df1SLai Jiangshan } 5527042f7df1SLai Jiangshan 5528042f7df1SLai Jiangshan /** 5529042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5530042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5531042f7df1SLai Jiangshan * 5532042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5533042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5534042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5535042f7df1SLai Jiangshan * 553667dc8325SCai Huoqing * Return: 0 - Success 5537042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5538042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5539042f7df1SLai Jiangshan */ 5540042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5541042f7df1SLai Jiangshan { 5542042f7df1SLai Jiangshan int ret = -EINVAL; 5543042f7df1SLai Jiangshan 5544c98a9805STal Shorer /* 5545c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5546c98a9805STal Shorer * If the user wishes to include them, we allow that. 5547c98a9805STal Shorer */ 5548042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5549042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5550a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5551d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5552d25302e4SMenglong Dong ret = 0; 5553d25302e4SMenglong Dong goto out_unlock; 5554d25302e4SMenglong Dong } 5555d25302e4SMenglong Dong 555699c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5557042f7df1SLai Jiangshan 5558d25302e4SMenglong Dong out_unlock: 5559a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5560042f7df1SLai Jiangshan } 5561042f7df1SLai Jiangshan 5562042f7df1SLai Jiangshan return ret; 5563042f7df1SLai Jiangshan } 5564042f7df1SLai Jiangshan 55656ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 55666ba94429SFrederic Weisbecker /* 55676ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 55686ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 55696ba94429SFrederic Weisbecker * following attributes. 55706ba94429SFrederic Weisbecker * 55716ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 55726ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 55736ba94429SFrederic Weisbecker * 55746ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 55756ba94429SFrederic Weisbecker * 55769a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 55776ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 55786ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 55799a19b463SWang Long * numa RW bool : whether enable NUMA affinity 55806ba94429SFrederic Weisbecker */ 55816ba94429SFrederic Weisbecker struct wq_device { 55826ba94429SFrederic Weisbecker struct workqueue_struct *wq; 55836ba94429SFrederic Weisbecker struct device dev; 55846ba94429SFrederic Weisbecker }; 55856ba94429SFrederic Weisbecker 55866ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 55876ba94429SFrederic Weisbecker { 55886ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 55896ba94429SFrederic Weisbecker 55906ba94429SFrederic Weisbecker return wq_dev->wq; 55916ba94429SFrederic Weisbecker } 55926ba94429SFrederic Weisbecker 55936ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 55946ba94429SFrederic Weisbecker char *buf) 55956ba94429SFrederic Weisbecker { 55966ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55976ba94429SFrederic Weisbecker 55986ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 55996ba94429SFrederic Weisbecker } 56006ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 56016ba94429SFrederic Weisbecker 56026ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 56036ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 56046ba94429SFrederic Weisbecker { 56056ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56066ba94429SFrederic Weisbecker 56076ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 56086ba94429SFrederic Weisbecker } 56096ba94429SFrederic Weisbecker 56106ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 56116ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 56126ba94429SFrederic Weisbecker size_t count) 56136ba94429SFrederic Weisbecker { 56146ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56156ba94429SFrederic Weisbecker int val; 56166ba94429SFrederic Weisbecker 56176ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 56186ba94429SFrederic Weisbecker return -EINVAL; 56196ba94429SFrederic Weisbecker 56206ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 56216ba94429SFrederic Weisbecker return count; 56226ba94429SFrederic Weisbecker } 56236ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 56246ba94429SFrederic Weisbecker 56256ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 56266ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 56276ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 56286ba94429SFrederic Weisbecker NULL, 56296ba94429SFrederic Weisbecker }; 56306ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 56316ba94429SFrederic Weisbecker 56326ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 56336ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 56346ba94429SFrederic Weisbecker { 56356ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56366ba94429SFrederic Weisbecker const char *delim = ""; 56376ba94429SFrederic Weisbecker int node, written = 0; 56386ba94429SFrederic Weisbecker 5639ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 564024acfb71SThomas Gleixner rcu_read_lock(); 56416ba94429SFrederic Weisbecker for_each_node(node) { 56426ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 56436ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 56446ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 56456ba94429SFrederic Weisbecker delim = " "; 56466ba94429SFrederic Weisbecker } 56476ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 564824acfb71SThomas Gleixner rcu_read_unlock(); 5649ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 56506ba94429SFrederic Weisbecker 56516ba94429SFrederic Weisbecker return written; 56526ba94429SFrederic Weisbecker } 56536ba94429SFrederic Weisbecker 56546ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 56556ba94429SFrederic Weisbecker char *buf) 56566ba94429SFrederic Weisbecker { 56576ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56586ba94429SFrederic Weisbecker int written; 56596ba94429SFrederic Weisbecker 56606ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 56616ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 56626ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 56636ba94429SFrederic Weisbecker 56646ba94429SFrederic Weisbecker return written; 56656ba94429SFrederic Weisbecker } 56666ba94429SFrederic Weisbecker 56676ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 56686ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 56696ba94429SFrederic Weisbecker { 56706ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 56716ba94429SFrederic Weisbecker 5672899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5673899a94feSLai Jiangshan 5674be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 56756ba94429SFrederic Weisbecker if (!attrs) 56766ba94429SFrederic Weisbecker return NULL; 56776ba94429SFrederic Weisbecker 56786ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 56796ba94429SFrederic Weisbecker return attrs; 56806ba94429SFrederic Weisbecker } 56816ba94429SFrederic Weisbecker 56826ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 56836ba94429SFrederic Weisbecker const char *buf, size_t count) 56846ba94429SFrederic Weisbecker { 56856ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56866ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5687d4d3e257SLai Jiangshan int ret = -ENOMEM; 5688d4d3e257SLai Jiangshan 5689d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56906ba94429SFrederic Weisbecker 56916ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56926ba94429SFrederic Weisbecker if (!attrs) 5693d4d3e257SLai Jiangshan goto out_unlock; 56946ba94429SFrederic Weisbecker 56956ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 56966ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5697d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56986ba94429SFrederic Weisbecker else 56996ba94429SFrederic Weisbecker ret = -EINVAL; 57006ba94429SFrederic Weisbecker 5701d4d3e257SLai Jiangshan out_unlock: 5702d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 57036ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 57046ba94429SFrederic Weisbecker return ret ?: count; 57056ba94429SFrederic Weisbecker } 57066ba94429SFrederic Weisbecker 57076ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 57086ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 57096ba94429SFrederic Weisbecker { 57106ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57116ba94429SFrederic Weisbecker int written; 57126ba94429SFrederic Weisbecker 57136ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 57146ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 57156ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 57166ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 57176ba94429SFrederic Weisbecker return written; 57186ba94429SFrederic Weisbecker } 57196ba94429SFrederic Weisbecker 57206ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 57216ba94429SFrederic Weisbecker struct device_attribute *attr, 57226ba94429SFrederic Weisbecker const char *buf, size_t count) 57236ba94429SFrederic Weisbecker { 57246ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57256ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5726d4d3e257SLai Jiangshan int ret = -ENOMEM; 5727d4d3e257SLai Jiangshan 5728d4d3e257SLai Jiangshan apply_wqattrs_lock(); 57296ba94429SFrederic Weisbecker 57306ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 57316ba94429SFrederic Weisbecker if (!attrs) 5732d4d3e257SLai Jiangshan goto out_unlock; 57336ba94429SFrederic Weisbecker 57346ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 57356ba94429SFrederic Weisbecker if (!ret) 5736d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 57376ba94429SFrederic Weisbecker 5738d4d3e257SLai Jiangshan out_unlock: 5739d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 57406ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 57416ba94429SFrederic Weisbecker return ret ?: count; 57426ba94429SFrederic Weisbecker } 57436ba94429SFrederic Weisbecker 57446ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 57456ba94429SFrederic Weisbecker char *buf) 57466ba94429SFrederic Weisbecker { 57476ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57486ba94429SFrederic Weisbecker int written; 57496ba94429SFrederic Weisbecker 57506ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 57516ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 57526ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 57536ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 57546ba94429SFrederic Weisbecker 57556ba94429SFrederic Weisbecker return written; 57566ba94429SFrederic Weisbecker } 57576ba94429SFrederic Weisbecker 57586ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 57596ba94429SFrederic Weisbecker const char *buf, size_t count) 57606ba94429SFrederic Weisbecker { 57616ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 57626ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5763d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5764d4d3e257SLai Jiangshan 5765d4d3e257SLai Jiangshan apply_wqattrs_lock(); 57666ba94429SFrederic Weisbecker 57676ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 57686ba94429SFrederic Weisbecker if (!attrs) 5769d4d3e257SLai Jiangshan goto out_unlock; 57706ba94429SFrederic Weisbecker 57716ba94429SFrederic Weisbecker ret = -EINVAL; 57726ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 57736ba94429SFrederic Weisbecker attrs->no_numa = !v; 5774d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 57756ba94429SFrederic Weisbecker } 57766ba94429SFrederic Weisbecker 5777d4d3e257SLai Jiangshan out_unlock: 5778d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 57796ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 57806ba94429SFrederic Weisbecker return ret ?: count; 57816ba94429SFrederic Weisbecker } 57826ba94429SFrederic Weisbecker 57836ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 57846ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 57856ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 57866ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 57876ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 57886ba94429SFrederic Weisbecker __ATTR_NULL, 57896ba94429SFrederic Weisbecker }; 57906ba94429SFrederic Weisbecker 57916ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 57926ba94429SFrederic Weisbecker .name = "workqueue", 57936ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 57946ba94429SFrederic Weisbecker }; 57956ba94429SFrederic Weisbecker 5796b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5797b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5798b05a7928SFrederic Weisbecker { 5799b05a7928SFrederic Weisbecker int written; 5800b05a7928SFrederic Weisbecker 5801042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5802b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5803b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5804042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5805b05a7928SFrederic Weisbecker 5806b05a7928SFrederic Weisbecker return written; 5807b05a7928SFrederic Weisbecker } 5808b05a7928SFrederic Weisbecker 5809042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5810042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5811042f7df1SLai Jiangshan { 5812042f7df1SLai Jiangshan cpumask_var_t cpumask; 5813042f7df1SLai Jiangshan int ret; 5814042f7df1SLai Jiangshan 5815042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5816042f7df1SLai Jiangshan return -ENOMEM; 5817042f7df1SLai Jiangshan 5818042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5819042f7df1SLai Jiangshan if (!ret) 5820042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5821042f7df1SLai Jiangshan 5822042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5823042f7df1SLai Jiangshan return ret ? ret : count; 5824042f7df1SLai Jiangshan } 5825042f7df1SLai Jiangshan 5826b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5827042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5828042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5829b05a7928SFrederic Weisbecker 58306ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 58316ba94429SFrederic Weisbecker { 5832b05a7928SFrederic Weisbecker int err; 5833b05a7928SFrederic Weisbecker 5834b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5835b05a7928SFrederic Weisbecker if (err) 5836b05a7928SFrederic Weisbecker return err; 5837b05a7928SFrederic Weisbecker 5838b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 58396ba94429SFrederic Weisbecker } 58406ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 58416ba94429SFrederic Weisbecker 58426ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 58436ba94429SFrederic Weisbecker { 58446ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 58456ba94429SFrederic Weisbecker 58466ba94429SFrederic Weisbecker kfree(wq_dev); 58476ba94429SFrederic Weisbecker } 58486ba94429SFrederic Weisbecker 58496ba94429SFrederic Weisbecker /** 58506ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 58516ba94429SFrederic Weisbecker * @wq: the workqueue to register 58526ba94429SFrederic Weisbecker * 58536ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 58546ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 58556ba94429SFrederic Weisbecker * which is the preferred method. 58566ba94429SFrederic Weisbecker * 58576ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 58586ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 58596ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 58606ba94429SFrederic Weisbecker * attributes. 58616ba94429SFrederic Weisbecker * 58626ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 58636ba94429SFrederic Weisbecker */ 58646ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 58656ba94429SFrederic Weisbecker { 58666ba94429SFrederic Weisbecker struct wq_device *wq_dev; 58676ba94429SFrederic Weisbecker int ret; 58686ba94429SFrederic Weisbecker 58696ba94429SFrederic Weisbecker /* 5870402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 58716ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 58726ba94429SFrederic Weisbecker * workqueues. 58736ba94429SFrederic Weisbecker */ 58740a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 58756ba94429SFrederic Weisbecker return -EINVAL; 58766ba94429SFrederic Weisbecker 58776ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 58786ba94429SFrederic Weisbecker if (!wq_dev) 58796ba94429SFrederic Weisbecker return -ENOMEM; 58806ba94429SFrederic Weisbecker 58816ba94429SFrederic Weisbecker wq_dev->wq = wq; 58826ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 58836ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 588423217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 58856ba94429SFrederic Weisbecker 58866ba94429SFrederic Weisbecker /* 58876ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 58886ba94429SFrederic Weisbecker * everything is ready. 58896ba94429SFrederic Weisbecker */ 58906ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 58916ba94429SFrederic Weisbecker 58926ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 58936ba94429SFrederic Weisbecker if (ret) { 5894537f4146SArvind Yadav put_device(&wq_dev->dev); 58956ba94429SFrederic Weisbecker wq->wq_dev = NULL; 58966ba94429SFrederic Weisbecker return ret; 58976ba94429SFrederic Weisbecker } 58986ba94429SFrederic Weisbecker 58996ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 59006ba94429SFrederic Weisbecker struct device_attribute *attr; 59016ba94429SFrederic Weisbecker 59026ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 59036ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 59046ba94429SFrederic Weisbecker if (ret) { 59056ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 59066ba94429SFrederic Weisbecker wq->wq_dev = NULL; 59076ba94429SFrederic Weisbecker return ret; 59086ba94429SFrederic Weisbecker } 59096ba94429SFrederic Weisbecker } 59106ba94429SFrederic Weisbecker } 59116ba94429SFrederic Weisbecker 59126ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 59136ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 59146ba94429SFrederic Weisbecker return 0; 59156ba94429SFrederic Weisbecker } 59166ba94429SFrederic Weisbecker 59176ba94429SFrederic Weisbecker /** 59186ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 59196ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 59206ba94429SFrederic Weisbecker * 59216ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 59226ba94429SFrederic Weisbecker */ 59236ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 59246ba94429SFrederic Weisbecker { 59256ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 59266ba94429SFrederic Weisbecker 59276ba94429SFrederic Weisbecker if (!wq->wq_dev) 59286ba94429SFrederic Weisbecker return; 59296ba94429SFrederic Weisbecker 59306ba94429SFrederic Weisbecker wq->wq_dev = NULL; 59316ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 59326ba94429SFrederic Weisbecker } 59336ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 59346ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 59356ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 59366ba94429SFrederic Weisbecker 593782607adcSTejun Heo /* 593882607adcSTejun Heo * Workqueue watchdog. 593982607adcSTejun Heo * 594082607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 594182607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 594282607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 594382607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 594482607adcSTejun Heo * largely opaque. 594582607adcSTejun Heo * 594682607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 594782607adcSTejun Heo * state if some pools failed to make forward progress for a while where 594882607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 594982607adcSTejun Heo * 595082607adcSTejun Heo * This mechanism is controlled through the kernel parameter 595182607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 595282607adcSTejun Heo * corresponding sysfs parameter file. 595382607adcSTejun Heo */ 595482607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 595582607adcSTejun Heo 595682607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 59575cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 595882607adcSTejun Heo 595982607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 596082607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 596182607adcSTejun Heo 596282607adcSTejun Heo static void wq_watchdog_reset_touched(void) 596382607adcSTejun Heo { 596482607adcSTejun Heo int cpu; 596582607adcSTejun Heo 596682607adcSTejun Heo wq_watchdog_touched = jiffies; 596782607adcSTejun Heo for_each_possible_cpu(cpu) 596882607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 596982607adcSTejun Heo } 597082607adcSTejun Heo 59715cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 597282607adcSTejun Heo { 597382607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 597482607adcSTejun Heo bool lockup_detected = false; 5975940d71c6SSergey Senozhatsky unsigned long now = jiffies; 597682607adcSTejun Heo struct worker_pool *pool; 597782607adcSTejun Heo int pi; 597882607adcSTejun Heo 597982607adcSTejun Heo if (!thresh) 598082607adcSTejun Heo return; 598182607adcSTejun Heo 598282607adcSTejun Heo rcu_read_lock(); 598382607adcSTejun Heo 598482607adcSTejun Heo for_each_pool(pool, pi) { 598582607adcSTejun Heo unsigned long pool_ts, touched, ts; 598682607adcSTejun Heo 598782607adcSTejun Heo if (list_empty(&pool->worklist)) 598882607adcSTejun Heo continue; 598982607adcSTejun Heo 5990940d71c6SSergey Senozhatsky /* 5991940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 5992940d71c6SSergey Senozhatsky * the watchdog like a stall. 5993940d71c6SSergey Senozhatsky */ 5994940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 5995940d71c6SSergey Senozhatsky 599682607adcSTejun Heo /* get the latest of pool and touched timestamps */ 599789e28ce6SWang Qing if (pool->cpu >= 0) 599889e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 599989e28ce6SWang Qing else 600082607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 600189e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 600282607adcSTejun Heo 600382607adcSTejun Heo if (time_after(pool_ts, touched)) 600482607adcSTejun Heo ts = pool_ts; 600582607adcSTejun Heo else 600682607adcSTejun Heo ts = touched; 600782607adcSTejun Heo 600882607adcSTejun Heo /* did we stall? */ 6009940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 601082607adcSTejun Heo lockup_detected = true; 601182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 601282607adcSTejun Heo pr_cont_pool_info(pool); 601382607adcSTejun Heo pr_cont(" stuck for %us!\n", 6014940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 601582607adcSTejun Heo } 601682607adcSTejun Heo } 601782607adcSTejun Heo 601882607adcSTejun Heo rcu_read_unlock(); 601982607adcSTejun Heo 602082607adcSTejun Heo if (lockup_detected) 602155df0933SImran Khan show_all_workqueues(); 602282607adcSTejun Heo 602382607adcSTejun Heo wq_watchdog_reset_touched(); 602482607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 602582607adcSTejun Heo } 602682607adcSTejun Heo 6027cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 602882607adcSTejun Heo { 602982607adcSTejun Heo if (cpu >= 0) 603082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 603189e28ce6SWang Qing 603282607adcSTejun Heo wq_watchdog_touched = jiffies; 603382607adcSTejun Heo } 603482607adcSTejun Heo 603582607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 603682607adcSTejun Heo { 603782607adcSTejun Heo wq_watchdog_thresh = 0; 603882607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 603982607adcSTejun Heo 604082607adcSTejun Heo if (thresh) { 604182607adcSTejun Heo wq_watchdog_thresh = thresh; 604282607adcSTejun Heo wq_watchdog_reset_touched(); 604382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 604482607adcSTejun Heo } 604582607adcSTejun Heo } 604682607adcSTejun Heo 604782607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 604882607adcSTejun Heo const struct kernel_param *kp) 604982607adcSTejun Heo { 605082607adcSTejun Heo unsigned long thresh; 605182607adcSTejun Heo int ret; 605282607adcSTejun Heo 605382607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 605482607adcSTejun Heo if (ret) 605582607adcSTejun Heo return ret; 605682607adcSTejun Heo 605782607adcSTejun Heo if (system_wq) 605882607adcSTejun Heo wq_watchdog_set_thresh(thresh); 605982607adcSTejun Heo else 606082607adcSTejun Heo wq_watchdog_thresh = thresh; 606182607adcSTejun Heo 606282607adcSTejun Heo return 0; 606382607adcSTejun Heo } 606482607adcSTejun Heo 606582607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 606682607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 606782607adcSTejun Heo .get = param_get_ulong, 606882607adcSTejun Heo }; 606982607adcSTejun Heo 607082607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 607182607adcSTejun Heo 0644); 607282607adcSTejun Heo 607382607adcSTejun Heo static void wq_watchdog_init(void) 607482607adcSTejun Heo { 60755cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 607682607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 607782607adcSTejun Heo } 607882607adcSTejun Heo 607982607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 608082607adcSTejun Heo 608182607adcSTejun Heo static inline void wq_watchdog_init(void) { } 608282607adcSTejun Heo 608382607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 608482607adcSTejun Heo 6085bce90380STejun Heo static void __init wq_numa_init(void) 6086bce90380STejun Heo { 6087bce90380STejun Heo cpumask_var_t *tbl; 6088bce90380STejun Heo int node, cpu; 6089bce90380STejun Heo 6090bce90380STejun Heo if (num_possible_nodes() <= 1) 6091bce90380STejun Heo return; 6092bce90380STejun Heo 6093d55262c4STejun Heo if (wq_disable_numa) { 6094d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 6095d55262c4STejun Heo return; 6096d55262c4STejun Heo } 6097d55262c4STejun Heo 6098f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 6099f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 6100f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 6101f728c4a9SZhen Lei return; 6102f728c4a9SZhen Lei } 6103f728c4a9SZhen Lei } 6104f728c4a9SZhen Lei 6105be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 61064c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 61074c16bd32STejun Heo 6108bce90380STejun Heo /* 6109bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 6110bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 6111bce90380STejun Heo * fully initialized by now. 6112bce90380STejun Heo */ 61136396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 6114bce90380STejun Heo BUG_ON(!tbl); 6115bce90380STejun Heo 6116bce90380STejun Heo for_each_node(node) 61175a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 61181be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 6119bce90380STejun Heo 6120bce90380STejun Heo for_each_possible_cpu(cpu) { 6121bce90380STejun Heo node = cpu_to_node(cpu); 6122bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6123bce90380STejun Heo } 6124bce90380STejun Heo 6125bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6126bce90380STejun Heo wq_numa_enabled = true; 6127bce90380STejun Heo } 6128bce90380STejun Heo 61293347fa09STejun Heo /** 61303347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 61313347fa09STejun Heo * 61323347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 61333347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 61343347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 61353347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 61363347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 61373347fa09STejun Heo * created and scheduled right before early initcalls. 61383347fa09STejun Heo */ 61392333e829SYu Chen void __init workqueue_init_early(void) 61401da177e4SLinus Torvalds { 61417a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 61427a4e344cSTejun Heo int i, cpu; 6143c34056a3STejun Heo 614410cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6145e904e6c2STejun Heo 6146b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 614704d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 614804d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6149b05a7928SFrederic Weisbecker 6150e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6151e904e6c2STejun Heo 6152706026c2STejun Heo /* initialize CPU pools */ 615329c91e99STejun Heo for_each_possible_cpu(cpu) { 61544ce62e9eSTejun Heo struct worker_pool *pool; 61558b03ae3cSTejun Heo 61567a4e344cSTejun Heo i = 0; 6157f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 61587a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6159ec22ca5eSTejun Heo pool->cpu = cpu; 61607a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 61617a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6162f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 61637a4e344cSTejun Heo 61649daf9e67STejun Heo /* alloc pool ID */ 616568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 61669daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 616768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 61684ce62e9eSTejun Heo } 61698b03ae3cSTejun Heo } 61708b03ae3cSTejun Heo 61718a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 617229c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 617329c91e99STejun Heo struct workqueue_attrs *attrs; 617429c91e99STejun Heo 6175be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 617629c91e99STejun Heo attrs->nice = std_nice[i]; 617729c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 61788a2b7538STejun Heo 61798a2b7538STejun Heo /* 61808a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 61818a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 61828a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 61838a2b7538STejun Heo */ 6184be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 61858a2b7538STejun Heo attrs->nice = std_nice[i]; 61868a2b7538STejun Heo attrs->no_numa = true; 61878a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 618829c91e99STejun Heo } 618929c91e99STejun Heo 6190d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 61911aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6192d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6193f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6194f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 619524d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 619624d51addSTejun Heo WQ_FREEZABLE, 0); 61970668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 61980668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 61990668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 62000668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 62010668106cSViresh Kumar 0); 62021aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 62030668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 62040668106cSViresh Kumar !system_power_efficient_wq || 62050668106cSViresh Kumar !system_freezable_power_efficient_wq); 62063347fa09STejun Heo } 62073347fa09STejun Heo 62083347fa09STejun Heo /** 62093347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 62103347fa09STejun Heo * 62113347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 62123347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 62133347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 62143347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 62153347fa09STejun Heo * with the initial workers and enable future kworker creations. 62163347fa09STejun Heo */ 62172333e829SYu Chen void __init workqueue_init(void) 62183347fa09STejun Heo { 62192186d9f9STejun Heo struct workqueue_struct *wq; 62203347fa09STejun Heo struct worker_pool *pool; 62213347fa09STejun Heo int cpu, bkt; 62223347fa09STejun Heo 62232186d9f9STejun Heo /* 62242186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 62252186d9f9STejun Heo * CPU to node mapping may not be available that early on some 62262186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 62272186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 62282186d9f9STejun Heo * affinity, fix them up. 622940c17f75STejun Heo * 623040c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 62312186d9f9STejun Heo */ 62322186d9f9STejun Heo wq_numa_init(); 62332186d9f9STejun Heo 62342186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 62352186d9f9STejun Heo 62362186d9f9STejun Heo for_each_possible_cpu(cpu) { 62372186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 62382186d9f9STejun Heo pool->node = cpu_to_node(cpu); 62392186d9f9STejun Heo } 62402186d9f9STejun Heo } 62412186d9f9STejun Heo 624240c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 62432186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 624440c17f75STejun Heo WARN(init_rescuer(wq), 624540c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 624640c17f75STejun Heo wq->name); 624740c17f75STejun Heo } 62482186d9f9STejun Heo 62492186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 62502186d9f9STejun Heo 62513347fa09STejun Heo /* create the initial workers */ 62523347fa09STejun Heo for_each_online_cpu(cpu) { 62533347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 62543347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 62553347fa09STejun Heo BUG_ON(!create_worker(pool)); 62563347fa09STejun Heo } 62573347fa09STejun Heo } 62583347fa09STejun Heo 62593347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 62603347fa09STejun Heo BUG_ON(!create_worker(pool)); 62613347fa09STejun Heo 62623347fa09STejun Heo wq_online = true; 626382607adcSTejun Heo wq_watchdog_init(); 62641da177e4SLinus Torvalds } 6265c4f135d6STetsuo Handa 6266c4f135d6STetsuo Handa /* 6267c4f135d6STetsuo Handa * Despite the naming, this is a no-op function which is here only for avoiding 6268c4f135d6STetsuo Handa * link error. Since compile-time warning may fail to catch, we will need to 6269c4f135d6STetsuo Handa * emit run-time warning from __flush_workqueue(). 6270c4f135d6STetsuo Handa */ 6271c4f135d6STetsuo Handa void __warn_flushing_systemwide_wq(void) { } 6272c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6273