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 157bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 158ea1abd61SLai Jiangshan 1595826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1605826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 161bd7bdd43STejun Heo 162bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 163bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 164bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 165bd7bdd43STejun Heo 166c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 167c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 168c9e7cf27STejun Heo /* L: hash of busy workers */ 169c9e7cf27STejun Heo 1702607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 17192f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 17260f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 173e19e397aSTejun Heo 1747cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 175e19e397aSTejun Heo 1767a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 17768e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 17868e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1797a4e344cSTejun Heo 180e19e397aSTejun Heo /* 181e19e397aSTejun Heo * The current concurrency level. As it's likely to be accessed 182e19e397aSTejun Heo * from other CPUs during try_to_wake_up(), put it in a separate 183e19e397aSTejun Heo * cacheline. 184e19e397aSTejun Heo */ 185e19e397aSTejun Heo atomic_t nr_running ____cacheline_aligned_in_smp; 18629c91e99STejun Heo 18729c91e99STejun Heo /* 18824acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 18929c91e99STejun Heo * from get_work_pool(). 19029c91e99STejun Heo */ 19129c91e99STejun Heo struct rcu_head rcu; 1928b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 1938b03ae3cSTejun Heo 1948b03ae3cSTejun Heo /* 195112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 196112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 197112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 198112202d9STejun Heo * number of flag bits. 1991da177e4SLinus Torvalds */ 200112202d9STejun Heo struct pool_workqueue { 201bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2024690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 20373f53c4aSTejun Heo int work_color; /* L: current color */ 20473f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2058864b4e5STejun Heo int refcnt; /* L: reference count */ 20673f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 20773f53c4aSTejun Heo /* L: nr of in_flight works */ 208018f3a13SLai Jiangshan 209018f3a13SLai Jiangshan /* 210018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 211018f3a13SLai Jiangshan * 212018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 213018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 214018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 215018f3a13SLai Jiangshan * 216018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 217018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 218018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 219018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 220018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 221018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 222018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 223018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 224018f3a13SLai Jiangshan */ 2251e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 226a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 227f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2283c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2292e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2308864b4e5STejun Heo 2318864b4e5STejun Heo /* 2328864b4e5STejun Heo * Release of unbound pwq is punted to system_wq. See put_pwq() 2338864b4e5STejun Heo * and pwq_unbound_release_workfn() for details. pool_workqueue 23424acfb71SThomas Gleixner * itself is also RCU protected so that the first pwq can be 235b09f4fd3SLai Jiangshan * determined without grabbing wq->mutex. 2368864b4e5STejun Heo */ 2378864b4e5STejun Heo struct work_struct unbound_release_work; 2388864b4e5STejun Heo struct rcu_head rcu; 239e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2401da177e4SLinus Torvalds 2411da177e4SLinus Torvalds /* 24273f53c4aSTejun Heo * Structure used to wait for workqueue flush. 24373f53c4aSTejun Heo */ 24473f53c4aSTejun Heo struct wq_flusher { 2453c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2463c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 24773f53c4aSTejun Heo struct completion done; /* flush completion */ 24873f53c4aSTejun Heo }; 2491da177e4SLinus Torvalds 250226223abSTejun Heo struct wq_device; 251226223abSTejun Heo 25273f53c4aSTejun Heo /* 253c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 254c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2551da177e4SLinus Torvalds */ 2561da177e4SLinus Torvalds struct workqueue_struct { 2573c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 258e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 25973f53c4aSTejun Heo 2603c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2613c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2623c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 263112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2643c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2653c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2663c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 26773f53c4aSTejun Heo 2682e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 26930ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 270e22bee78STejun Heo 27187fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 272a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 273226223abSTejun Heo 2745b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 2755b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 2766029a918STejun Heo 277226223abSTejun Heo #ifdef CONFIG_SYSFS 278226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 279226223abSTejun Heo #endif 2804e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 281669de8bdSBart Van Assche char *lock_name; 282669de8bdSBart Van Assche struct lock_class_key key; 2834e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2844e6045f1SJohannes Berg #endif 285ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 2862728fd2fSTejun Heo 287e2dca7adSTejun Heo /* 28824acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 28924acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 290e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 291e2dca7adSTejun Heo */ 292e2dca7adSTejun Heo struct rcu_head rcu; 293e2dca7adSTejun Heo 2942728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 2952728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 2962728fd2fSTejun Heo struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */ 2975b95e1afSLai Jiangshan struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */ 2981da177e4SLinus Torvalds }; 2991da177e4SLinus Torvalds 300e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 301e904e6c2STejun Heo 302bce90380STejun Heo static cpumask_var_t *wq_numa_possible_cpumask; 303bce90380STejun Heo /* possible CPUs of each node */ 304bce90380STejun Heo 305d55262c4STejun Heo static bool wq_disable_numa; 306d55262c4STejun Heo module_param_named(disable_numa, wq_disable_numa, bool, 0444); 307d55262c4STejun Heo 308cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 309552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 310cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 311cee22a15SViresh Kumar 312863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3133347fa09STejun Heo 314bce90380STejun Heo static bool wq_numa_enabled; /* unbound NUMA affinity enabled */ 315bce90380STejun Heo 3164c16bd32STejun Heo /* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */ 3174c16bd32STejun Heo static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf; 3184c16bd32STejun Heo 31968e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3201258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 321a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 322d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 323d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3245bcab335STejun Heo 325e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 32668e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3277d19c5ceSTejun Heo 328ef557180SMike Galbraith /* PL: allowable cpus for unbound wqs and work items */ 329ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 330ef557180SMike Galbraith 331ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 332ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 333b05a7928SFrederic Weisbecker 334f303fccbSTejun Heo /* 335f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 336f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 337f303fccbSTejun Heo * to uncover usages which depend on it. 338f303fccbSTejun Heo */ 339f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 340f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 341f303fccbSTejun Heo #else 342f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 343f303fccbSTejun Heo #endif 344f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 345f303fccbSTejun Heo 3467d19c5ceSTejun Heo /* the per-cpu worker pools */ 34725528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 3487d19c5ceSTejun Heo 34968e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 3507d19c5ceSTejun Heo 35168e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 35229c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 35329c91e99STejun Heo 354c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 35529c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 35629c91e99STejun Heo 3578a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 3588a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 3598a2b7538STejun Heo 360d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 361ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 362044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 3631aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 364044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 365d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 366044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 367f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 368044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 36924d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 3700668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 3710668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 3720668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 3730668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 374d320c038STejun Heo 3757d19c5ceSTejun Heo static int worker_thread(void *__worker); 3766ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 377c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 37855df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 3797d19c5ceSTejun Heo 38097bd2347STejun Heo #define CREATE_TRACE_POINTS 38197bd2347STejun Heo #include <trace/events/workqueue.h> 38297bd2347STejun Heo 38368e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 38424acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 385f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 38624acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 3875bcab335STejun Heo 3885b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 38924acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 390f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 391f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 39224acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 3935b95e1afSLai Jiangshan 394f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 395f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 396f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 3977a62c2c8STejun Heo (pool)++) 3984ce62e9eSTejun Heo 39949e3cf44STejun Heo /** 40017116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 40117116969STejun Heo * @pool: iteration cursor 402611c92a0STejun Heo * @pi: integer used for iteration 403fa1b54e6STejun Heo * 40424acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 40568e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 40668e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 407fa1b54e6STejun Heo * 408fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 409fa1b54e6STejun Heo * ignored. 41017116969STejun Heo */ 411611c92a0STejun Heo #define for_each_pool(pool, pi) \ 412611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 41368e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 414fa1b54e6STejun Heo else 41517116969STejun Heo 41617116969STejun Heo /** 417822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 418822d8405STejun Heo * @worker: iteration cursor 419822d8405STejun Heo * @pool: worker_pool to iterate workers of 420822d8405STejun Heo * 4211258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 422822d8405STejun Heo * 423822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 424822d8405STejun Heo * ignored. 425822d8405STejun Heo */ 426da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 427da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4281258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 429822d8405STejun Heo else 430822d8405STejun Heo 431822d8405STejun Heo /** 43249e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 43349e3cf44STejun Heo * @pwq: iteration cursor 43449e3cf44STejun Heo * @wq: the target workqueue 43576af4d93STejun Heo * 43624acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 437794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 438794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 43976af4d93STejun Heo * 44076af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 44176af4d93STejun Heo * ignored. 44249e3cf44STejun Heo */ 44349e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 44449e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 4455a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 446f3421797STejun Heo 447dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 448dc186ad7SThomas Gleixner 449f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 450dc186ad7SThomas Gleixner 45199777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 45299777288SStanislaw Gruszka { 45399777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 45499777288SStanislaw Gruszka } 45599777288SStanislaw Gruszka 456b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 457b9fdac7fSDu, Changbin { 458b9fdac7fSDu, Changbin struct work_struct *work = addr; 459b9fdac7fSDu, Changbin 460b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 461b9fdac7fSDu, Changbin } 462b9fdac7fSDu, Changbin 463dc186ad7SThomas Gleixner /* 464dc186ad7SThomas Gleixner * fixup_init is called when: 465dc186ad7SThomas Gleixner * - an active object is initialized 466dc186ad7SThomas Gleixner */ 46702a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 468dc186ad7SThomas Gleixner { 469dc186ad7SThomas Gleixner struct work_struct *work = addr; 470dc186ad7SThomas Gleixner 471dc186ad7SThomas Gleixner switch (state) { 472dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 473dc186ad7SThomas Gleixner cancel_work_sync(work); 474dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 47502a982a6SDu, Changbin return true; 476dc186ad7SThomas Gleixner default: 47702a982a6SDu, Changbin return false; 478dc186ad7SThomas Gleixner } 479dc186ad7SThomas Gleixner } 480dc186ad7SThomas Gleixner 481dc186ad7SThomas Gleixner /* 482dc186ad7SThomas Gleixner * fixup_free is called when: 483dc186ad7SThomas Gleixner * - an active object is freed 484dc186ad7SThomas Gleixner */ 48502a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 486dc186ad7SThomas Gleixner { 487dc186ad7SThomas Gleixner struct work_struct *work = addr; 488dc186ad7SThomas Gleixner 489dc186ad7SThomas Gleixner switch (state) { 490dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 491dc186ad7SThomas Gleixner cancel_work_sync(work); 492dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 49302a982a6SDu, Changbin return true; 494dc186ad7SThomas Gleixner default: 49502a982a6SDu, Changbin return false; 496dc186ad7SThomas Gleixner } 497dc186ad7SThomas Gleixner } 498dc186ad7SThomas Gleixner 499f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 500dc186ad7SThomas Gleixner .name = "work_struct", 50199777288SStanislaw Gruszka .debug_hint = work_debug_hint, 502b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 503dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 504dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 505dc186ad7SThomas Gleixner }; 506dc186ad7SThomas Gleixner 507dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 508dc186ad7SThomas Gleixner { 509dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 510dc186ad7SThomas Gleixner } 511dc186ad7SThomas Gleixner 512dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 513dc186ad7SThomas Gleixner { 514dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 515dc186ad7SThomas Gleixner } 516dc186ad7SThomas Gleixner 517dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 518dc186ad7SThomas Gleixner { 519dc186ad7SThomas Gleixner if (onstack) 520dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 521dc186ad7SThomas Gleixner else 522dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 523dc186ad7SThomas Gleixner } 524dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 525dc186ad7SThomas Gleixner 526dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 527dc186ad7SThomas Gleixner { 528dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 529dc186ad7SThomas Gleixner } 530dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 531dc186ad7SThomas Gleixner 532ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 533ea2e64f2SThomas Gleixner { 534ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 535ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 536ea2e64f2SThomas Gleixner } 537ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 538ea2e64f2SThomas Gleixner 539dc186ad7SThomas Gleixner #else 540dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 541dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 542dc186ad7SThomas Gleixner #endif 543dc186ad7SThomas Gleixner 5444e8b22bdSLi Bin /** 54567dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 5464e8b22bdSLi Bin * @pool: the pool pointer of interest 5474e8b22bdSLi Bin * 5484e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 5494e8b22bdSLi Bin * successfully, -errno on failure. 5504e8b22bdSLi Bin */ 5519daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 5529daf9e67STejun Heo { 5539daf9e67STejun Heo int ret; 5549daf9e67STejun Heo 55568e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5565bcab335STejun Heo 5574e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 5584e8b22bdSLi Bin GFP_KERNEL); 559229641a6STejun Heo if (ret >= 0) { 560e68035fbSTejun Heo pool->id = ret; 561229641a6STejun Heo return 0; 562229641a6STejun Heo } 5639daf9e67STejun Heo return ret; 5649daf9e67STejun Heo } 5659daf9e67STejun Heo 56676af4d93STejun Heo /** 567df2d5ae4STejun Heo * unbound_pwq_by_node - return the unbound pool_workqueue for the given node 568df2d5ae4STejun Heo * @wq: the target workqueue 569df2d5ae4STejun Heo * @node: the node ID 570df2d5ae4STejun Heo * 57124acfb71SThomas Gleixner * This must be called with any of wq_pool_mutex, wq->mutex or RCU 5725b95e1afSLai Jiangshan * read locked. 573df2d5ae4STejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 574df2d5ae4STejun Heo * responsible for guaranteeing that the pwq stays online. 575d185af30SYacine Belkadi * 576d185af30SYacine Belkadi * Return: The unbound pool_workqueue for @node. 577df2d5ae4STejun Heo */ 578df2d5ae4STejun Heo static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, 579df2d5ae4STejun Heo int node) 580df2d5ae4STejun Heo { 5815b95e1afSLai Jiangshan assert_rcu_or_wq_mutex_or_pool_mutex(wq); 582d6e022f1STejun Heo 583d6e022f1STejun Heo /* 584d6e022f1STejun Heo * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a 585d6e022f1STejun Heo * delayed item is pending. The plan is to keep CPU -> NODE 586d6e022f1STejun Heo * mapping valid and stable across CPU on/offlines. Once that 587d6e022f1STejun Heo * happens, this workaround can be removed. 588d6e022f1STejun Heo */ 589d6e022f1STejun Heo if (unlikely(node == NUMA_NO_NODE)) 590d6e022f1STejun Heo return wq->dfl_pwq; 591d6e022f1STejun Heo 592df2d5ae4STejun Heo return rcu_dereference_raw(wq->numa_pwq_tbl[node]); 593df2d5ae4STejun Heo } 594df2d5ae4STejun Heo 59573f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 59673f53c4aSTejun Heo { 59773f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 59873f53c4aSTejun Heo } 59973f53c4aSTejun Heo 600c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 60173f53c4aSTejun Heo { 602c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 60373f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 60473f53c4aSTejun Heo } 60573f53c4aSTejun Heo 60673f53c4aSTejun Heo static int work_next_color(int color) 60773f53c4aSTejun Heo { 60873f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 609a848e3b6SOleg Nesterov } 610a848e3b6SOleg Nesterov 6114594bf15SDavid Howells /* 612112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 613112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6147c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6157a22ad75STejun Heo * 616112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 617112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 618bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 619bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6207a22ad75STejun Heo * 621112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6227c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 623112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6247c3eed5cSTejun Heo * available only while the work item is queued. 625bbb68dfaSTejun Heo * 626bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 627bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 628bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 629bbb68dfaSTejun Heo * try to steal the PENDING bit. 6304594bf15SDavid Howells */ 6317a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6327a22ad75STejun Heo unsigned long flags) 6337a22ad75STejun Heo { 6346183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6357a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6367a22ad75STejun Heo } 6377a22ad75STejun Heo 638112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6394690c4abSTejun Heo unsigned long extra_flags) 640365970a1SDavid Howells { 641112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 642112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 643365970a1SDavid Howells } 644365970a1SDavid Howells 6454468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6464468a00fSLai Jiangshan int pool_id) 6474468a00fSLai Jiangshan { 6484468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6494468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6504468a00fSLai Jiangshan } 6514468a00fSLai Jiangshan 6527c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6537c3eed5cSTejun Heo int pool_id) 6544d707b9fSOleg Nesterov { 65523657bb1STejun Heo /* 65623657bb1STejun Heo * The following wmb is paired with the implied mb in 65723657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 65823657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 65923657bb1STejun Heo * owner. 66023657bb1STejun Heo */ 66123657bb1STejun Heo smp_wmb(); 6627c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 663346c09f8SRoman Pen /* 664346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 665346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 666346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6678bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 668346c09f8SRoman Pen * the same @work. E.g. consider this case: 669346c09f8SRoman Pen * 670346c09f8SRoman Pen * CPU#0 CPU#1 671346c09f8SRoman Pen * ---------------------------- -------------------------------- 672346c09f8SRoman Pen * 673346c09f8SRoman Pen * 1 STORE event_indicated 674346c09f8SRoman Pen * 2 queue_work_on() { 675346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 676346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 677346c09f8SRoman Pen * 5 set_work_data() # clear bit 678346c09f8SRoman Pen * 6 smp_mb() 679346c09f8SRoman Pen * 7 work->current_func() { 680346c09f8SRoman Pen * 8 LOAD event_indicated 681346c09f8SRoman Pen * } 682346c09f8SRoman Pen * 683346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 684346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 685346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 686346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 687346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 688346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 689346c09f8SRoman Pen * before actual STORE. 690346c09f8SRoman Pen */ 691346c09f8SRoman Pen smp_mb(); 6924d707b9fSOleg Nesterov } 6934d707b9fSOleg Nesterov 6947a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 695365970a1SDavid Howells { 6967c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 6977c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 6987a22ad75STejun Heo } 6997a22ad75STejun Heo 700112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7017a22ad75STejun Heo { 702e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7037a22ad75STejun Heo 704112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 705e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 706e120153dSTejun Heo else 707e120153dSTejun Heo return NULL; 7087a22ad75STejun Heo } 7097a22ad75STejun Heo 7107c3eed5cSTejun Heo /** 7117c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7127c3eed5cSTejun Heo * @work: the work item of interest 7137c3eed5cSTejun Heo * 71468e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 71524acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 71624acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 717fa1b54e6STejun Heo * 718fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 719fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 720fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 721fa1b54e6STejun Heo * returned pool is and stays online. 722d185af30SYacine Belkadi * 723d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7247c3eed5cSTejun Heo */ 7257c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7267a22ad75STejun Heo { 727e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7287c3eed5cSTejun Heo int pool_id; 7297a22ad75STejun Heo 73068e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 731fa1b54e6STejun Heo 732112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 733112202d9STejun Heo return ((struct pool_workqueue *) 7347c3eed5cSTejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool; 7357a22ad75STejun Heo 7367c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7377c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7387a22ad75STejun Heo return NULL; 7397a22ad75STejun Heo 740fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7417c3eed5cSTejun Heo } 7427c3eed5cSTejun Heo 7437c3eed5cSTejun Heo /** 7447c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7457c3eed5cSTejun Heo * @work: the work item of interest 7467c3eed5cSTejun Heo * 747d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7487c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7497c3eed5cSTejun Heo */ 7507c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7517c3eed5cSTejun Heo { 75254d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7537c3eed5cSTejun Heo 754112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 755112202d9STejun Heo return ((struct pool_workqueue *) 75654d5b7d0SLai Jiangshan (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id; 75754d5b7d0SLai Jiangshan 75854d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7597c3eed5cSTejun Heo } 7607c3eed5cSTejun Heo 761bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 762bbb68dfaSTejun Heo { 7637c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 764bbb68dfaSTejun Heo 7657c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 7667c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 767bbb68dfaSTejun Heo } 768bbb68dfaSTejun Heo 769bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 770bbb68dfaSTejun Heo { 771bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 772bbb68dfaSTejun Heo 773112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 774bbb68dfaSTejun Heo } 775bbb68dfaSTejun Heo 776e22bee78STejun Heo /* 7773270476aSTejun Heo * Policy functions. These define the policies on how the global worker 7783270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 779d565ed63STejun Heo * they're being called with pool->lock held. 780e22bee78STejun Heo */ 781e22bee78STejun Heo 78263d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 783649027d7STejun Heo { 784e19e397aSTejun Heo return !atomic_read(&pool->nr_running); 785649027d7STejun Heo } 786649027d7STejun Heo 787e22bee78STejun Heo /* 788e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 789e22bee78STejun Heo * running workers. 790974271c4STejun Heo * 791974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 792706026c2STejun Heo * function will always return %true for unbound pools as long as the 793974271c4STejun Heo * worklist isn't empty. 794e22bee78STejun Heo */ 79563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 796e22bee78STejun Heo { 79763d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 798e22bee78STejun Heo } 799e22bee78STejun Heo 800e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 80163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 802e22bee78STejun Heo { 80363d95a91STejun Heo return pool->nr_idle; 804e22bee78STejun Heo } 805e22bee78STejun Heo 806e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 80763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 808e22bee78STejun Heo { 809e19e397aSTejun Heo return !list_empty(&pool->worklist) && 810e19e397aSTejun Heo atomic_read(&pool->nr_running) <= 1; 811e22bee78STejun Heo } 812e22bee78STejun Heo 813e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 81463d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 815e22bee78STejun Heo { 81663d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 817e22bee78STejun Heo } 818e22bee78STejun Heo 819e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 82063d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 821e22bee78STejun Heo { 822692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 82363d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 82463d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 825e22bee78STejun Heo 826e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 827e22bee78STejun Heo } 828e22bee78STejun Heo 829e22bee78STejun Heo /* 830e22bee78STejun Heo * Wake up functions. 831e22bee78STejun Heo */ 832e22bee78STejun Heo 8331037de36SLai Jiangshan /* Return the first idle worker. Safe with preemption disabled */ 8341037de36SLai Jiangshan static struct worker *first_idle_worker(struct worker_pool *pool) 8357e11629dSTejun Heo { 83663d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 8377e11629dSTejun Heo return NULL; 8387e11629dSTejun Heo 83963d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 8407e11629dSTejun Heo } 8417e11629dSTejun Heo 8427e11629dSTejun Heo /** 8437e11629dSTejun Heo * wake_up_worker - wake up an idle worker 84463d95a91STejun Heo * @pool: worker pool to wake worker from 8457e11629dSTejun Heo * 84663d95a91STejun Heo * Wake up the first idle worker of @pool. 8477e11629dSTejun Heo * 8487e11629dSTejun Heo * CONTEXT: 849a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 8507e11629dSTejun Heo */ 85163d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 8527e11629dSTejun Heo { 8531037de36SLai Jiangshan struct worker *worker = first_idle_worker(pool); 8547e11629dSTejun Heo 8557e11629dSTejun Heo if (likely(worker)) 8567e11629dSTejun Heo wake_up_process(worker->task); 8577e11629dSTejun Heo } 8587e11629dSTejun Heo 8594690c4abSTejun Heo /** 8606d25be57SThomas Gleixner * wq_worker_running - a worker is running again 861e22bee78STejun Heo * @task: task waking up 862e22bee78STejun Heo * 8636d25be57SThomas Gleixner * This function is called when a worker returns from schedule() 864e22bee78STejun Heo */ 8656d25be57SThomas Gleixner void wq_worker_running(struct task_struct *task) 866e22bee78STejun Heo { 867e22bee78STejun Heo struct worker *worker = kthread_data(task); 868e22bee78STejun Heo 8696d25be57SThomas Gleixner if (!worker->sleeping) 8706d25be57SThomas Gleixner return; 8716d25be57SThomas Gleixner if (!(worker->flags & WORKER_NOT_RUNNING)) 872e19e397aSTejun Heo atomic_inc(&worker->pool->nr_running); 8736d25be57SThomas Gleixner worker->sleeping = 0; 87436576000SJoonsoo Kim } 875e22bee78STejun Heo 876e22bee78STejun Heo /** 877e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 878e22bee78STejun Heo * @task: task going to sleep 879e22bee78STejun Heo * 8806d25be57SThomas Gleixner * This function is called from schedule() when a busy worker is 88162849a96SSebastian Andrzej Siewior * going to sleep. Preemption needs to be disabled to protect ->sleeping 88262849a96SSebastian Andrzej Siewior * assignment. 883e22bee78STejun Heo */ 8846d25be57SThomas Gleixner void wq_worker_sleeping(struct task_struct *task) 885e22bee78STejun Heo { 8866d25be57SThomas Gleixner struct worker *next, *worker = kthread_data(task); 887111c225aSTejun Heo struct worker_pool *pool; 888e22bee78STejun Heo 889111c225aSTejun Heo /* 890111c225aSTejun Heo * Rescuers, which may not have all the fields set up like normal 891111c225aSTejun Heo * workers, also reach here, let's not access anything before 892111c225aSTejun Heo * checking NOT_RUNNING. 893111c225aSTejun Heo */ 8942d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 8956d25be57SThomas Gleixner return; 896e22bee78STejun Heo 897111c225aSTejun Heo pool = worker->pool; 898111c225aSTejun Heo 89962849a96SSebastian Andrzej Siewior /* Return if preempted before wq_worker_running() was reached */ 90062849a96SSebastian Andrzej Siewior if (worker->sleeping) 9016d25be57SThomas Gleixner return; 9026d25be57SThomas Gleixner 9036d25be57SThomas Gleixner worker->sleeping = 1; 904a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 905e22bee78STejun Heo 906e22bee78STejun Heo /* 907e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 908e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 909e22bee78STejun Heo * Please read comment there. 910e22bee78STejun Heo * 911628c78e7STejun Heo * NOT_RUNNING is clear. This means that we're bound to and 912628c78e7STejun Heo * running on the local cpu w/ rq lock held and preemption 913628c78e7STejun Heo * disabled, which in turn means that none else could be 914d565ed63STejun Heo * manipulating idle_list, so dereferencing idle_list without pool 915628c78e7STejun Heo * lock is safe. 916e22bee78STejun Heo */ 917e19e397aSTejun Heo if (atomic_dec_and_test(&pool->nr_running) && 9186d25be57SThomas Gleixner !list_empty(&pool->worklist)) { 9196d25be57SThomas Gleixner next = first_idle_worker(pool); 9206d25be57SThomas Gleixner if (next) 9216d25be57SThomas Gleixner wake_up_process(next->task); 9226d25be57SThomas Gleixner } 923a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 924e22bee78STejun Heo } 925e22bee78STejun Heo 926e22bee78STejun Heo /** 9271b69ac6bSJohannes Weiner * wq_worker_last_func - retrieve worker's last work function 9288194fe94SBart Van Assche * @task: Task to retrieve last work function of. 9291b69ac6bSJohannes Weiner * 9301b69ac6bSJohannes Weiner * Determine the last function a worker executed. This is called from 9311b69ac6bSJohannes Weiner * the scheduler to get a worker's last known identity. 9321b69ac6bSJohannes Weiner * 9331b69ac6bSJohannes Weiner * CONTEXT: 934a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(rq->lock) 9351b69ac6bSJohannes Weiner * 9364b047002SJohannes Weiner * This function is called during schedule() when a kworker is going 9374b047002SJohannes Weiner * to sleep. It's used by psi to identify aggregation workers during 9384b047002SJohannes Weiner * dequeuing, to allow periodic aggregation to shut-off when that 9394b047002SJohannes Weiner * worker is the last task in the system or cgroup to go to sleep. 9404b047002SJohannes Weiner * 9414b047002SJohannes Weiner * As this function doesn't involve any workqueue-related locking, it 9424b047002SJohannes Weiner * only returns stable values when called from inside the scheduler's 9434b047002SJohannes Weiner * queuing and dequeuing paths, when @task, which must be a kworker, 9444b047002SJohannes Weiner * is guaranteed to not be processing any works. 9454b047002SJohannes Weiner * 9461b69ac6bSJohannes Weiner * Return: 9471b69ac6bSJohannes Weiner * The last work function %current executed as a worker, NULL if it 9481b69ac6bSJohannes Weiner * hasn't executed any work yet. 9491b69ac6bSJohannes Weiner */ 9501b69ac6bSJohannes Weiner work_func_t wq_worker_last_func(struct task_struct *task) 9511b69ac6bSJohannes Weiner { 9521b69ac6bSJohannes Weiner struct worker *worker = kthread_data(task); 9531b69ac6bSJohannes Weiner 9541b69ac6bSJohannes Weiner return worker->last_func; 9551b69ac6bSJohannes Weiner } 9561b69ac6bSJohannes Weiner 9571b69ac6bSJohannes Weiner /** 958e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 959cb444766STejun Heo * @worker: self 960d302f017STejun Heo * @flags: flags to set 961d302f017STejun Heo * 962228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 963d302f017STejun Heo * 964cb444766STejun Heo * CONTEXT: 965a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 966d302f017STejun Heo */ 967228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 968d302f017STejun Heo { 969bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 970e22bee78STejun Heo 971cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 972cb444766STejun Heo 973228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 974e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 975e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 976e19e397aSTejun Heo atomic_dec(&pool->nr_running); 977e22bee78STejun Heo } 978e22bee78STejun Heo 979d302f017STejun Heo worker->flags |= flags; 980d302f017STejun Heo } 981d302f017STejun Heo 982d302f017STejun Heo /** 983e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 984cb444766STejun Heo * @worker: self 985d302f017STejun Heo * @flags: flags to clear 986d302f017STejun Heo * 987e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 988d302f017STejun Heo * 989cb444766STejun Heo * CONTEXT: 990a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) 991d302f017STejun Heo */ 992d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 993d302f017STejun Heo { 99463d95a91STejun Heo struct worker_pool *pool = worker->pool; 995e22bee78STejun Heo unsigned int oflags = worker->flags; 996e22bee78STejun Heo 997cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 998cb444766STejun Heo 999d302f017STejun Heo worker->flags &= ~flags; 1000e22bee78STejun Heo 100142c025f3STejun Heo /* 100242c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 100342c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 100442c025f3STejun Heo * of multiple flags, not a single flag. 100542c025f3STejun Heo */ 1006e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 1007e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 1008e19e397aSTejun Heo atomic_inc(&pool->nr_running); 1009d302f017STejun Heo } 1010d302f017STejun Heo 1011d302f017STejun Heo /** 10128cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 1013c9e7cf27STejun Heo * @pool: pool of interest 10148cca0eeaSTejun Heo * @work: work to find worker for 10158cca0eeaSTejun Heo * 1016c9e7cf27STejun Heo * Find a worker which is executing @work on @pool by searching 1017c9e7cf27STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 1018a2c1c57bSTejun Heo * to match, its current execution should match the address of @work and 1019a2c1c57bSTejun Heo * its work function. This is to avoid unwanted dependency between 1020a2c1c57bSTejun Heo * unrelated work executions through a work item being recycled while still 1021a2c1c57bSTejun Heo * being executed. 1022a2c1c57bSTejun Heo * 1023a2c1c57bSTejun Heo * This is a bit tricky. A work item may be freed once its execution 1024a2c1c57bSTejun Heo * starts and nothing prevents the freed area from being recycled for 1025a2c1c57bSTejun Heo * another work item. If the same work item address ends up being reused 1026a2c1c57bSTejun Heo * before the original execution finishes, workqueue will identify the 1027a2c1c57bSTejun Heo * recycled work item as currently executing and make it wait until the 1028a2c1c57bSTejun Heo * current execution finishes, introducing an unwanted dependency. 1029a2c1c57bSTejun Heo * 1030c5aa87bbSTejun Heo * This function checks the work item address and work function to avoid 1031c5aa87bbSTejun Heo * false positives. Note that this isn't complete as one may construct a 1032c5aa87bbSTejun Heo * work function which can introduce dependency onto itself through a 1033c5aa87bbSTejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 1034c5aa87bbSTejun Heo * foot that badly, there's only so much we can do, and if such deadlock 1035c5aa87bbSTejun Heo * actually occurs, it should be easy to locate the culprit work function. 10368cca0eeaSTejun Heo * 10378cca0eeaSTejun Heo * CONTEXT: 1038a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 10398cca0eeaSTejun Heo * 1040d185af30SYacine Belkadi * Return: 1041d185af30SYacine Belkadi * Pointer to worker which is executing @work if found, %NULL 10428cca0eeaSTejun Heo * otherwise. 10438cca0eeaSTejun Heo */ 1044c9e7cf27STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 10458cca0eeaSTejun Heo struct work_struct *work) 10468cca0eeaSTejun Heo { 104742f8570fSSasha Levin struct worker *worker; 104842f8570fSSasha Levin 1049b67bfe0dSSasha Levin hash_for_each_possible(pool->busy_hash, worker, hentry, 1050a2c1c57bSTejun Heo (unsigned long)work) 1051a2c1c57bSTejun Heo if (worker->current_work == work && 1052a2c1c57bSTejun Heo worker->current_func == work->func) 105342f8570fSSasha Levin return worker; 105442f8570fSSasha Levin 105542f8570fSSasha Levin return NULL; 10568cca0eeaSTejun Heo } 10578cca0eeaSTejun Heo 10588cca0eeaSTejun Heo /** 1059bf4ede01STejun Heo * move_linked_works - move linked works to a list 1060bf4ede01STejun Heo * @work: start of series of works to be scheduled 1061bf4ede01STejun Heo * @head: target list to append @work to 1062402dd89dSShailendra Verma * @nextp: out parameter for nested worklist walking 1063bf4ede01STejun Heo * 1064bf4ede01STejun Heo * Schedule linked works starting from @work to @head. Work series to 1065bf4ede01STejun Heo * be scheduled starts at @work and includes any consecutive work with 1066bf4ede01STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1067bf4ede01STejun Heo * 1068bf4ede01STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1069bf4ede01STejun Heo * the last scheduled work. This allows move_linked_works() to be 1070bf4ede01STejun Heo * nested inside outer list_for_each_entry_safe(). 1071bf4ede01STejun Heo * 1072bf4ede01STejun Heo * CONTEXT: 1073a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1074bf4ede01STejun Heo */ 1075bf4ede01STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1076bf4ede01STejun Heo struct work_struct **nextp) 1077bf4ede01STejun Heo { 1078bf4ede01STejun Heo struct work_struct *n; 1079bf4ede01STejun Heo 1080bf4ede01STejun Heo /* 1081bf4ede01STejun Heo * Linked worklist will always end before the end of the list, 1082bf4ede01STejun Heo * use NULL for list head. 1083bf4ede01STejun Heo */ 1084bf4ede01STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1085bf4ede01STejun Heo list_move_tail(&work->entry, head); 1086bf4ede01STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1087bf4ede01STejun Heo break; 1088bf4ede01STejun Heo } 1089bf4ede01STejun Heo 1090bf4ede01STejun Heo /* 1091bf4ede01STejun Heo * If we're already inside safe list traversal and have moved 1092bf4ede01STejun Heo * multiple works to the scheduled queue, the next position 1093bf4ede01STejun Heo * needs to be updated. 1094bf4ede01STejun Heo */ 1095bf4ede01STejun Heo if (nextp) 1096bf4ede01STejun Heo *nextp = n; 1097bf4ede01STejun Heo } 1098bf4ede01STejun Heo 10998864b4e5STejun Heo /** 11008864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 11018864b4e5STejun Heo * @pwq: pool_workqueue to get 11028864b4e5STejun Heo * 11038864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 11048864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 11058864b4e5STejun Heo */ 11068864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 11078864b4e5STejun Heo { 11088864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11098864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 11108864b4e5STejun Heo pwq->refcnt++; 11118864b4e5STejun Heo } 11128864b4e5STejun Heo 11138864b4e5STejun Heo /** 11148864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 11158864b4e5STejun Heo * @pwq: pool_workqueue to put 11168864b4e5STejun Heo * 11178864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 11188864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 11198864b4e5STejun Heo */ 11208864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 11218864b4e5STejun Heo { 11228864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 11238864b4e5STejun Heo if (likely(--pwq->refcnt)) 11248864b4e5STejun Heo return; 11258864b4e5STejun Heo if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND))) 11268864b4e5STejun Heo return; 11278864b4e5STejun Heo /* 11288864b4e5STejun Heo * @pwq can't be released under pool->lock, bounce to 11298864b4e5STejun Heo * pwq_unbound_release_workfn(). This never recurses on the same 11308864b4e5STejun Heo * pool->lock as this path is taken only for unbound workqueues and 11318864b4e5STejun Heo * the release work item is scheduled on a per-cpu workqueue. To 11328864b4e5STejun Heo * avoid lockdep warning, unbound pool->locks are given lockdep 11338864b4e5STejun Heo * subclass of 1 in get_unbound_pool(). 11348864b4e5STejun Heo */ 11358864b4e5STejun Heo schedule_work(&pwq->unbound_release_work); 11368864b4e5STejun Heo } 11378864b4e5STejun Heo 1138dce90d47STejun Heo /** 1139dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1140dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1141dce90d47STejun Heo * 1142dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1143dce90d47STejun Heo */ 1144dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1145dce90d47STejun Heo { 1146dce90d47STejun Heo if (pwq) { 1147dce90d47STejun Heo /* 114824acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1149dce90d47STejun Heo * following lock operations are safe. 1150dce90d47STejun Heo */ 1151a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1152dce90d47STejun Heo put_pwq(pwq); 1153a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1154dce90d47STejun Heo } 1155dce90d47STejun Heo } 1156dce90d47STejun Heo 1157f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1158bf4ede01STejun Heo { 1159112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1160bf4ede01STejun Heo 1161bf4ede01STejun Heo trace_workqueue_activate_work(work); 116282607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 116382607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1164112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1165f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1166112202d9STejun Heo pwq->nr_active++; 1167bf4ede01STejun Heo } 1168bf4ede01STejun Heo 1169f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 11703aa62497SLai Jiangshan { 1171f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 11723aa62497SLai Jiangshan struct work_struct, entry); 11733aa62497SLai Jiangshan 1174f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 11753aa62497SLai Jiangshan } 11763aa62497SLai Jiangshan 1177bf4ede01STejun Heo /** 1178112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1179112202d9STejun Heo * @pwq: pwq of interest 1180c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1181bf4ede01STejun Heo * 1182bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1183112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1184bf4ede01STejun Heo * 1185bf4ede01STejun Heo * CONTEXT: 1186a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1187bf4ede01STejun Heo */ 1188c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1189bf4ede01STejun Heo { 1190c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1191c4560c2cSLai Jiangshan 1192018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1193112202d9STejun Heo pwq->nr_active--; 1194f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1195f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1196112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1197f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1198bf4ede01STejun Heo } 1199018f3a13SLai Jiangshan } 1200018f3a13SLai Jiangshan 1201018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1202bf4ede01STejun Heo 1203bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1204112202d9STejun Heo if (likely(pwq->flush_color != color)) 12058864b4e5STejun Heo goto out_put; 1206bf4ede01STejun Heo 1207bf4ede01STejun Heo /* are there still in-flight works? */ 1208112202d9STejun Heo if (pwq->nr_in_flight[color]) 12098864b4e5STejun Heo goto out_put; 1210bf4ede01STejun Heo 1211112202d9STejun Heo /* this pwq is done, clear flush_color */ 1212112202d9STejun Heo pwq->flush_color = -1; 1213bf4ede01STejun Heo 1214bf4ede01STejun Heo /* 1215112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1216bf4ede01STejun Heo * will handle the rest. 1217bf4ede01STejun Heo */ 1218112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1219112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 12208864b4e5STejun Heo out_put: 12218864b4e5STejun Heo put_pwq(pwq); 1222bf4ede01STejun Heo } 1223bf4ede01STejun Heo 122436e227d2STejun Heo /** 1225bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 122636e227d2STejun Heo * @work: work item to steal 122736e227d2STejun Heo * @is_dwork: @work is a delayed_work 1228bbb68dfaSTejun Heo * @flags: place to store irq state 122936e227d2STejun Heo * 123036e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1231d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 123236e227d2STejun Heo * 1233d185af30SYacine Belkadi * Return: 12343eb6b31bSMauro Carvalho Chehab * 12353eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 123636e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 123736e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 123836e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1239bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1240bbb68dfaSTejun Heo * for arbitrarily long 12413eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 124236e227d2STejun Heo * 1243d185af30SYacine Belkadi * Note: 1244bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1245e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1246e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1247e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1248bbb68dfaSTejun Heo * 1249bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1250bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1251bbb68dfaSTejun Heo * 1252e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1253bf4ede01STejun Heo */ 1254bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1255bbb68dfaSTejun Heo unsigned long *flags) 1256bf4ede01STejun Heo { 1257d565ed63STejun Heo struct worker_pool *pool; 1258112202d9STejun Heo struct pool_workqueue *pwq; 1259bf4ede01STejun Heo 1260bbb68dfaSTejun Heo local_irq_save(*flags); 1261bbb68dfaSTejun Heo 126236e227d2STejun Heo /* try to steal the timer if it exists */ 126336e227d2STejun Heo if (is_dwork) { 126436e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 126536e227d2STejun Heo 1266e0aecdd8STejun Heo /* 1267e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1268e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1269e0aecdd8STejun Heo * running on the local CPU. 1270e0aecdd8STejun Heo */ 127136e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 127236e227d2STejun Heo return 1; 127336e227d2STejun Heo } 127436e227d2STejun Heo 127536e227d2STejun Heo /* try to claim PENDING the normal way */ 1276bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1277bf4ede01STejun Heo return 0; 1278bf4ede01STejun Heo 127924acfb71SThomas Gleixner rcu_read_lock(); 1280bf4ede01STejun Heo /* 1281bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1282bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1283bf4ede01STejun Heo */ 1284d565ed63STejun Heo pool = get_work_pool(work); 1285d565ed63STejun Heo if (!pool) 1286bbb68dfaSTejun Heo goto fail; 1287bf4ede01STejun Heo 1288a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1289bf4ede01STejun Heo /* 1290112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1291112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1292112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1293112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1294112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 12950b3dae68SLai Jiangshan * item is currently queued on that pool. 1296bf4ede01STejun Heo */ 1297112202d9STejun Heo pwq = get_work_pwq(work); 1298112202d9STejun Heo if (pwq && pwq->pool == pool) { 1299bf4ede01STejun Heo debug_work_deactivate(work); 13003aa62497SLai Jiangshan 13013aa62497SLai Jiangshan /* 1302018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1303018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1304018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1305018f3a13SLai Jiangshan * 1306f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1307d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1308f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 130916062836STejun Heo * management later on and cause stall. Make sure the work 131016062836STejun Heo * item is activated before grabbing. 13113aa62497SLai Jiangshan */ 1312f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1313f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 13143aa62497SLai Jiangshan 1315bf4ede01STejun Heo list_del_init(&work->entry); 1316c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 131736e227d2STejun Heo 1318112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 13194468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 13204468a00fSLai Jiangshan 1321a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 132224acfb71SThomas Gleixner rcu_read_unlock(); 132336e227d2STejun Heo return 1; 1324bf4ede01STejun Heo } 1325a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1326bbb68dfaSTejun Heo fail: 132724acfb71SThomas Gleixner rcu_read_unlock(); 1328bbb68dfaSTejun Heo local_irq_restore(*flags); 1329bbb68dfaSTejun Heo if (work_is_canceling(work)) 1330bbb68dfaSTejun Heo return -ENOENT; 1331bbb68dfaSTejun Heo cpu_relax(); 133236e227d2STejun Heo return -EAGAIN; 1333bf4ede01STejun Heo } 1334bf4ede01STejun Heo 1335bf4ede01STejun Heo /** 1336706026c2STejun Heo * insert_work - insert a work into a pool 1337112202d9STejun Heo * @pwq: pwq @work belongs to 13384690c4abSTejun Heo * @work: work to insert 13394690c4abSTejun Heo * @head: insertion point 13404690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 13414690c4abSTejun Heo * 1342112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1343706026c2STejun Heo * work_struct flags. 13444690c4abSTejun Heo * 13454690c4abSTejun Heo * CONTEXT: 1346a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1347365970a1SDavid Howells */ 1348112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1349112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1350b89deed3SOleg Nesterov { 1351112202d9STejun Heo struct worker_pool *pool = pwq->pool; 1352e1d8aa9fSFrederic Weisbecker 1353e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1354f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1355e89a85d6SWalter Wu 13564690c4abSTejun Heo /* we own @work, set data and link */ 1357112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 13581a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 13598864b4e5STejun Heo get_pwq(pwq); 1360e22bee78STejun Heo 1361e22bee78STejun Heo /* 1362c5aa87bbSTejun Heo * Ensure either wq_worker_sleeping() sees the above 1363c5aa87bbSTejun Heo * list_add_tail() or we see zero nr_running to avoid workers lying 1364c5aa87bbSTejun Heo * around lazily while there are works to be processed. 1365e22bee78STejun Heo */ 1366e22bee78STejun Heo smp_mb(); 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 14399ef28a73SLi Bin /* if draining, only works from the same workqueue are allowed */ 1440618b01ebSTejun Heo if (unlikely(wq->flags & __WQ_DRAINING) && 1441c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1442e41e704bSTejun Heo return; 144324acfb71SThomas Gleixner rcu_read_lock(); 14449e8cd2f5STejun Heo retry: 1445aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1446aa202f1fSHillf Danton if (wq->flags & WQ_UNBOUND) { 1447df2d5ae4STejun Heo if (req_cpu == WORK_CPU_UNBOUND) 1448ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1449df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 1450aa202f1fSHillf Danton } else { 1451aa202f1fSHillf Danton if (req_cpu == WORK_CPU_UNBOUND) 1452aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1453aa202f1fSHillf Danton pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 1454aa202f1fSHillf Danton } 1455f3421797STejun Heo 145618aa9effSTejun Heo /* 1457c9178087STejun Heo * If @work was previously on a different pool, it might still be 1458c9178087STejun Heo * running there, in which case the work needs to be queued on that 1459c9178087STejun Heo * pool to guarantee non-reentrancy. 146018aa9effSTejun Heo */ 1461c9e7cf27STejun Heo last_pool = get_work_pool(work); 1462112202d9STejun Heo if (last_pool && last_pool != pwq->pool) { 146318aa9effSTejun Heo struct worker *worker; 146418aa9effSTejun Heo 1465a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 146618aa9effSTejun Heo 1467c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 146818aa9effSTejun Heo 1469112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1470c9178087STejun Heo pwq = worker->current_pwq; 14718594fadeSLai Jiangshan } else { 147218aa9effSTejun Heo /* meh... not running there, queue here */ 1473a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1474a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 147518aa9effSTejun Heo } 14768930cabaSTejun Heo } else { 1477a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pwq->pool->lock); 14788930cabaSTejun Heo } 1479502ca9d8STejun Heo 14809e8cd2f5STejun Heo /* 14819e8cd2f5STejun Heo * pwq is determined and locked. For unbound pools, we could have 14829e8cd2f5STejun Heo * raced with pwq release and it could already be dead. If its 14839e8cd2f5STejun Heo * refcnt is zero, repeat pwq selection. Note that pwqs never die 1484df2d5ae4STejun Heo * without another pwq replacing it in the numa_pwq_tbl or while 1485df2d5ae4STejun Heo * work items are executing on it, so the retrying is guaranteed to 14869e8cd2f5STejun Heo * make forward-progress. 14879e8cd2f5STejun Heo */ 14889e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 14899e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1490a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 14919e8cd2f5STejun Heo cpu_relax(); 14929e8cd2f5STejun Heo goto retry; 14939e8cd2f5STejun Heo } 14949e8cd2f5STejun Heo /* oops */ 14959e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 14969e8cd2f5STejun Heo wq->name, cpu); 14979e8cd2f5STejun Heo } 14989e8cd2f5STejun Heo 1499112202d9STejun Heo /* pwq determined, queue */ 1500112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1501502ca9d8STejun Heo 150224acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 150324acfb71SThomas Gleixner goto out; 15041e19ffc6STejun Heo 1505112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1506112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 15071e19ffc6STejun Heo 1508112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1509cdadf009STejun Heo trace_workqueue_activate_work(work); 1510112202d9STejun Heo pwq->nr_active++; 1511112202d9STejun Heo worklist = &pwq->pool->worklist; 151282607adcSTejun Heo if (list_empty(worklist)) 151382607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 15148a2e8e5dSTejun Heo } else { 1515f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1516f97a4a1aSLai Jiangshan worklist = &pwq->inactive_works; 15178a2e8e5dSTejun Heo } 15181e19ffc6STejun Heo 15190687c66bSZqiang debug_work_activate(work); 1520112202d9STejun Heo insert_work(pwq, work, worklist, work_flags); 15211e19ffc6STejun Heo 152224acfb71SThomas Gleixner out: 1523a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pwq->pool->lock); 152424acfb71SThomas Gleixner rcu_read_unlock(); 15251da177e4SLinus Torvalds } 15261da177e4SLinus Torvalds 15270fcb78c2SRolf Eike Beer /** 1528c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1529c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1530c1a220e7SZhang Rui * @wq: workqueue to use 1531c1a220e7SZhang Rui * @work: work to queue 1532c1a220e7SZhang Rui * 1533c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1534*443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1535*443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1536d185af30SYacine Belkadi * 1537d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1538c1a220e7SZhang Rui */ 1539d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1540d4283e93STejun Heo struct work_struct *work) 1541c1a220e7SZhang Rui { 1542d4283e93STejun Heo bool ret = false; 15438930cabaSTejun Heo unsigned long flags; 15448930cabaSTejun Heo 15458930cabaSTejun Heo local_irq_save(flags); 1546c1a220e7SZhang Rui 154722df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 15484690c4abSTejun Heo __queue_work(cpu, wq, work); 1549d4283e93STejun Heo ret = true; 1550c1a220e7SZhang Rui } 15518930cabaSTejun Heo 15528930cabaSTejun Heo local_irq_restore(flags); 1553c1a220e7SZhang Rui return ret; 1554c1a220e7SZhang Rui } 1555ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1556c1a220e7SZhang Rui 15578204e0c1SAlexander Duyck /** 15588204e0c1SAlexander Duyck * workqueue_select_cpu_near - Select a CPU based on NUMA node 15598204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 15608204e0c1SAlexander Duyck * 15618204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 15628204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 15638204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 15648204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 15658204e0c1SAlexander Duyck */ 15668204e0c1SAlexander Duyck static int workqueue_select_cpu_near(int node) 15678204e0c1SAlexander Duyck { 15688204e0c1SAlexander Duyck int cpu; 15698204e0c1SAlexander Duyck 15708204e0c1SAlexander Duyck /* No point in doing this if NUMA isn't enabled for workqueues */ 15718204e0c1SAlexander Duyck if (!wq_numa_enabled) 15728204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15738204e0c1SAlexander Duyck 15748204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 15758204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 15768204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 15778204e0c1SAlexander Duyck 15788204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 15798204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 15808204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 15818204e0c1SAlexander Duyck return cpu; 15828204e0c1SAlexander Duyck 15838204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 15848204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 15858204e0c1SAlexander Duyck 15868204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 15878204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 15888204e0c1SAlexander Duyck } 15898204e0c1SAlexander Duyck 15908204e0c1SAlexander Duyck /** 15918204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 15928204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 15938204e0c1SAlexander Duyck * @wq: workqueue to use 15948204e0c1SAlexander Duyck * @work: work to queue 15958204e0c1SAlexander Duyck * 15968204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 15978204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 15988204e0c1SAlexander Duyck * NUMA node. 15998204e0c1SAlexander Duyck * 16008204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 16018204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 16028204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 16038204e0c1SAlexander Duyck * 16048204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 16058204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 16068204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 16078204e0c1SAlexander Duyck * 16088204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 16098204e0c1SAlexander Duyck */ 16108204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 16118204e0c1SAlexander Duyck struct work_struct *work) 16128204e0c1SAlexander Duyck { 16138204e0c1SAlexander Duyck unsigned long flags; 16148204e0c1SAlexander Duyck bool ret = false; 16158204e0c1SAlexander Duyck 16168204e0c1SAlexander Duyck /* 16178204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 16188204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 16198204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 16208204e0c1SAlexander Duyck * 16218204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 16228204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 16238204e0c1SAlexander Duyck * some round robin type logic. 16248204e0c1SAlexander Duyck */ 16258204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 16268204e0c1SAlexander Duyck 16278204e0c1SAlexander Duyck local_irq_save(flags); 16288204e0c1SAlexander Duyck 16298204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 16308204e0c1SAlexander Duyck int cpu = workqueue_select_cpu_near(node); 16318204e0c1SAlexander Duyck 16328204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 16338204e0c1SAlexander Duyck ret = true; 16348204e0c1SAlexander Duyck } 16358204e0c1SAlexander Duyck 16368204e0c1SAlexander Duyck local_irq_restore(flags); 16378204e0c1SAlexander Duyck return ret; 16388204e0c1SAlexander Duyck } 16398204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 16408204e0c1SAlexander Duyck 16418c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 16421da177e4SLinus Torvalds { 16438c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 16441da177e4SLinus Torvalds 1645e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 164660c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 16471da177e4SLinus Torvalds } 16481438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 16491da177e4SLinus Torvalds 16507beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 165152bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16521da177e4SLinus Torvalds { 16537beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 16547beb2edfSTejun Heo struct work_struct *work = &dwork->work; 16551da177e4SLinus Torvalds 1656637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 165798173112SSami Tolvanen WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn); 1658fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1659fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 16607beb2edfSTejun Heo 16618852aac2STejun Heo /* 16628852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 16638852aac2STejun Heo * both optimization and correctness. The earliest @timer can 16648852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 16658852aac2STejun Heo * on that there's no such delay when @delay is 0. 16668852aac2STejun Heo */ 16678852aac2STejun Heo if (!delay) { 16688852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 16698852aac2STejun Heo return; 16708852aac2STejun Heo } 16718852aac2STejun Heo 167260c057bcSLai Jiangshan dwork->wq = wq; 16731265057fSTejun Heo dwork->cpu = cpu; 16747beb2edfSTejun Heo timer->expires = jiffies + delay; 16757beb2edfSTejun Heo 1676041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 16777beb2edfSTejun Heo add_timer_on(timer, cpu); 1678041bd12eSTejun Heo else 1679041bd12eSTejun Heo add_timer(timer); 16807beb2edfSTejun Heo } 16811da177e4SLinus Torvalds 16820fcb78c2SRolf Eike Beer /** 16830fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 16840fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 16850fcb78c2SRolf Eike Beer * @wq: workqueue to use 1686af9997e4SRandy Dunlap * @dwork: work to queue 16870fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 16880fcb78c2SRolf Eike Beer * 1689d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1690715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1691715f1300STejun Heo * execution. 16920fcb78c2SRolf Eike Beer */ 1693d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 169452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 16957a6bc1cdSVenkatesh Pallipadi { 169652bad64dSDavid Howells struct work_struct *work = &dwork->work; 1697d4283e93STejun Heo bool ret = false; 16988930cabaSTejun Heo unsigned long flags; 16998930cabaSTejun Heo 17008930cabaSTejun Heo /* read the comment in __queue_work() */ 17018930cabaSTejun Heo local_irq_save(flags); 17027a6bc1cdSVenkatesh Pallipadi 170322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 17047beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1705d4283e93STejun Heo ret = true; 17067a6bc1cdSVenkatesh Pallipadi } 17078930cabaSTejun Heo 17088930cabaSTejun Heo local_irq_restore(flags); 17097a6bc1cdSVenkatesh Pallipadi return ret; 17107a6bc1cdSVenkatesh Pallipadi } 1711ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 17121da177e4SLinus Torvalds 1713c8e55f36STejun Heo /** 17148376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 17158376fe22STejun Heo * @cpu: CPU number to execute work on 17168376fe22STejun Heo * @wq: workqueue to use 17178376fe22STejun Heo * @dwork: work to queue 17188376fe22STejun Heo * @delay: number of jiffies to wait before queueing 17198376fe22STejun Heo * 17208376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 17218376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 17228376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 17238376fe22STejun Heo * current state. 17248376fe22STejun Heo * 1725d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 17268376fe22STejun Heo * pending and its timer was modified. 17278376fe22STejun Heo * 1728e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 17298376fe22STejun Heo * See try_to_grab_pending() for details. 17308376fe22STejun Heo */ 17318376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 17328376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 17338376fe22STejun Heo { 17348376fe22STejun Heo unsigned long flags; 17358376fe22STejun Heo int ret; 17368376fe22STejun Heo 17378376fe22STejun Heo do { 17388376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 17398376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 17408376fe22STejun Heo 17418376fe22STejun Heo if (likely(ret >= 0)) { 17428376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 17438376fe22STejun Heo local_irq_restore(flags); 17448376fe22STejun Heo } 17458376fe22STejun Heo 17468376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 17478376fe22STejun Heo return ret; 17488376fe22STejun Heo } 17498376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 17508376fe22STejun Heo 175105f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 175205f0fe6bSTejun Heo { 175305f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 175405f0fe6bSTejun Heo 175505f0fe6bSTejun Heo /* read the comment in __queue_work() */ 175605f0fe6bSTejun Heo local_irq_disable(); 175705f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 175805f0fe6bSTejun Heo local_irq_enable(); 175905f0fe6bSTejun Heo } 176005f0fe6bSTejun Heo 176105f0fe6bSTejun Heo /** 176205f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 176305f0fe6bSTejun Heo * @wq: workqueue to use 176405f0fe6bSTejun Heo * @rwork: work to queue 176505f0fe6bSTejun Heo * 176605f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 176705f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 1768bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 176905f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 177005f0fe6bSTejun Heo */ 177105f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 177205f0fe6bSTejun Heo { 177305f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 177405f0fe6bSTejun Heo 177505f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 177605f0fe6bSTejun Heo rwork->wq = wq; 177705f0fe6bSTejun Heo call_rcu(&rwork->rcu, rcu_work_rcufn); 177805f0fe6bSTejun Heo return true; 177905f0fe6bSTejun Heo } 178005f0fe6bSTejun Heo 178105f0fe6bSTejun Heo return false; 178205f0fe6bSTejun Heo } 178305f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 178405f0fe6bSTejun Heo 17858376fe22STejun Heo /** 1786c8e55f36STejun Heo * worker_enter_idle - enter idle state 1787c8e55f36STejun Heo * @worker: worker which is entering idle state 1788c8e55f36STejun Heo * 1789c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1790c8e55f36STejun Heo * necessary. 1791c8e55f36STejun Heo * 1792c8e55f36STejun Heo * LOCKING: 1793a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1794c8e55f36STejun Heo */ 1795c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 17961da177e4SLinus Torvalds { 1797bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1798c8e55f36STejun Heo 17996183c009STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 18006183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 18016183c009STejun Heo (worker->hentry.next || worker->hentry.pprev))) 18026183c009STejun Heo return; 1803c8e55f36STejun Heo 1804051e1850SLai Jiangshan /* can't use worker_set_flags(), also called from create_worker() */ 1805cb444766STejun Heo worker->flags |= WORKER_IDLE; 1806bd7bdd43STejun Heo pool->nr_idle++; 1807e22bee78STejun Heo worker->last_active = jiffies; 1808c8e55f36STejun Heo 1809c8e55f36STejun Heo /* idle_list is LIFO */ 1810bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1811db7bccf4STejun Heo 181263d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1813628c78e7STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 1814cb444766STejun Heo 1815544ecf31STejun Heo /* 1816e8b3f8dbSLai Jiangshan * Sanity check nr_running. Because unbind_workers() releases 1817d565ed63STejun Heo * pool->lock between setting %WORKER_UNBOUND and zapping 1818628c78e7STejun Heo * nr_running, the warning may trigger spuriously. Check iff 1819628c78e7STejun Heo * unbind is not in progress. 1820544ecf31STejun Heo */ 182124647570STejun Heo WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 1822bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 1823e19e397aSTejun Heo atomic_read(&pool->nr_running)); 1824c8e55f36STejun Heo } 1825c8e55f36STejun Heo 1826c8e55f36STejun Heo /** 1827c8e55f36STejun Heo * worker_leave_idle - leave idle state 1828c8e55f36STejun Heo * @worker: worker which is leaving idle state 1829c8e55f36STejun Heo * 1830c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1831c8e55f36STejun Heo * 1832c8e55f36STejun Heo * LOCKING: 1833a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1834c8e55f36STejun Heo */ 1835c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1836c8e55f36STejun Heo { 1837bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1838c8e55f36STejun Heo 18396183c009STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 18406183c009STejun Heo return; 1841d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1842bd7bdd43STejun Heo pool->nr_idle--; 1843c8e55f36STejun Heo list_del_init(&worker->entry); 1844c8e55f36STejun Heo } 1845c8e55f36STejun Heo 1846f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 1847c34056a3STejun Heo { 1848c34056a3STejun Heo struct worker *worker; 1849c34056a3STejun Heo 1850f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 1851c8e55f36STejun Heo if (worker) { 1852c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1853affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 1854da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 1855e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1856e22bee78STejun Heo worker->flags = WORKER_PREP; 1857c8e55f36STejun Heo } 1858c34056a3STejun Heo return worker; 1859c34056a3STejun Heo } 1860c34056a3STejun Heo 1861c34056a3STejun Heo /** 18624736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 18634736cbf7SLai Jiangshan * @worker: worker to be attached 18644736cbf7SLai Jiangshan * @pool: the target pool 18654736cbf7SLai Jiangshan * 18664736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 18674736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 18684736cbf7SLai Jiangshan * cpu-[un]hotplugs. 18694736cbf7SLai Jiangshan */ 18704736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 18714736cbf7SLai Jiangshan struct worker_pool *pool) 18724736cbf7SLai Jiangshan { 18731258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 18744736cbf7SLai Jiangshan 18754736cbf7SLai Jiangshan /* 18761258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 18771258fae7STejun Heo * stable across this function. See the comments above the flag 18781258fae7STejun Heo * definition for details. 18794736cbf7SLai Jiangshan */ 18804736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 18814736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 18825c25b5ffSPeter Zijlstra else 18835c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 18844736cbf7SLai Jiangshan 1885640f17c8SPeter Zijlstra if (worker->rescue_wq) 1886640f17c8SPeter Zijlstra set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask); 1887640f17c8SPeter Zijlstra 18884736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 1889a2d812a2STejun Heo worker->pool = pool; 18904736cbf7SLai Jiangshan 18911258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 18924736cbf7SLai Jiangshan } 18934736cbf7SLai Jiangshan 18944736cbf7SLai Jiangshan /** 189560f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 189660f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 189760f5a4bcSLai Jiangshan * 18984736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 18994736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 19004736cbf7SLai Jiangshan * other reference to the pool. 190160f5a4bcSLai Jiangshan */ 1902a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 190360f5a4bcSLai Jiangshan { 1904a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 190560f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 190660f5a4bcSLai Jiangshan 19071258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 1908a2d812a2STejun Heo 19095c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 1910da028469SLai Jiangshan list_del(&worker->node); 1911a2d812a2STejun Heo worker->pool = NULL; 1912a2d812a2STejun Heo 1913da028469SLai Jiangshan if (list_empty(&pool->workers)) 191460f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 19151258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 191660f5a4bcSLai Jiangshan 1917b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 1918b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 1919b62c0751SLai Jiangshan 192060f5a4bcSLai Jiangshan if (detach_completion) 192160f5a4bcSLai Jiangshan complete(detach_completion); 192260f5a4bcSLai Jiangshan } 192360f5a4bcSLai Jiangshan 192460f5a4bcSLai Jiangshan /** 1925c34056a3STejun Heo * create_worker - create a new workqueue worker 192663d95a91STejun Heo * @pool: pool the new worker will belong to 1927c34056a3STejun Heo * 1928051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 1929c34056a3STejun Heo * 1930c34056a3STejun Heo * CONTEXT: 1931c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1932c34056a3STejun Heo * 1933d185af30SYacine Belkadi * Return: 1934c34056a3STejun Heo * Pointer to the newly created worker. 1935c34056a3STejun Heo */ 1936bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1937c34056a3STejun Heo { 1938e441b56fSZhen Lei struct worker *worker; 1939e441b56fSZhen Lei int id; 1940e3c916a4STejun Heo char id_buf[16]; 1941c34056a3STejun Heo 19427cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 1943e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 1944822d8405STejun Heo if (id < 0) 1945e441b56fSZhen Lei return NULL; 1946c34056a3STejun Heo 1947f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 1948c34056a3STejun Heo if (!worker) 1949c34056a3STejun Heo goto fail; 1950c34056a3STejun Heo 1951c34056a3STejun Heo worker->id = id; 1952c34056a3STejun Heo 195329c91e99STejun Heo if (pool->cpu >= 0) 1954e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 1955e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 1956f3421797STejun Heo else 1957e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 1958e3c916a4STejun Heo 1959f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 1960e3c916a4STejun Heo "kworker/%s", id_buf); 1961c34056a3STejun Heo if (IS_ERR(worker->task)) 1962c34056a3STejun Heo goto fail; 1963c34056a3STejun Heo 196491151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 196525834c73SPeter Zijlstra kthread_bind_mask(worker->task, pool->attrs->cpumask); 196691151228SOleg Nesterov 1967da028469SLai Jiangshan /* successful, attach the worker to the pool */ 19684736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 1969822d8405STejun Heo 1970051e1850SLai Jiangshan /* start the newly created worker */ 1971a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 1972051e1850SLai Jiangshan worker->pool->nr_workers++; 1973051e1850SLai Jiangshan worker_enter_idle(worker); 1974051e1850SLai Jiangshan wake_up_process(worker->task); 1975a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 1976051e1850SLai Jiangshan 1977c34056a3STejun Heo return worker; 1978822d8405STejun Heo 1979c34056a3STejun Heo fail: 1980e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 1981c34056a3STejun Heo kfree(worker); 1982c34056a3STejun Heo return NULL; 1983c34056a3STejun Heo } 1984c34056a3STejun Heo 1985c34056a3STejun Heo /** 1986c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1987c34056a3STejun Heo * @worker: worker to be destroyed 1988c34056a3STejun Heo * 198973eb7fe7SLai Jiangshan * Destroy @worker and adjust @pool stats accordingly. The worker should 199073eb7fe7SLai Jiangshan * be idle. 1991c8e55f36STejun Heo * 1992c8e55f36STejun Heo * CONTEXT: 1993a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1994c34056a3STejun Heo */ 1995c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1996c34056a3STejun Heo { 1997bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1998c34056a3STejun Heo 1999cd549687STejun Heo lockdep_assert_held(&pool->lock); 2000cd549687STejun Heo 2001c34056a3STejun Heo /* sanity check frenzy */ 20026183c009STejun Heo if (WARN_ON(worker->current_work) || 200373eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 200473eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 20056183c009STejun Heo return; 2006c34056a3STejun Heo 2007bd7bdd43STejun Heo pool->nr_workers--; 2008bd7bdd43STejun Heo pool->nr_idle--; 2009c8e55f36STejun Heo 2010c8e55f36STejun Heo list_del_init(&worker->entry); 2011cb444766STejun Heo worker->flags |= WORKER_DIE; 201260f5a4bcSLai Jiangshan wake_up_process(worker->task); 2013c34056a3STejun Heo } 2014c34056a3STejun Heo 201532a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2016e22bee78STejun Heo { 201732a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 2018e22bee78STejun Heo 2019a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2020e22bee78STejun Heo 20213347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2022e22bee78STejun Heo struct worker *worker; 2023e22bee78STejun Heo unsigned long expires; 2024e22bee78STejun Heo 2025e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 202663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2027e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2028e22bee78STejun Heo 20293347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 203063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 20313347fc9fSLai Jiangshan break; 2032e22bee78STejun Heo } 20333347fc9fSLai Jiangshan 20343347fc9fSLai Jiangshan destroy_worker(worker); 2035e22bee78STejun Heo } 2036e22bee78STejun Heo 2037a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2038e22bee78STejun Heo } 2039e22bee78STejun Heo 2040493a1724STejun Heo static void send_mayday(struct work_struct *work) 2041e22bee78STejun Heo { 2042112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2043112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2044493a1724STejun Heo 20452e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2046e22bee78STejun Heo 2047493008a8STejun Heo if (!wq->rescuer) 2048493a1724STejun Heo return; 2049e22bee78STejun Heo 2050e22bee78STejun Heo /* mayday mayday mayday */ 2051493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 205277668c8bSLai Jiangshan /* 205377668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 205477668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 205577668c8bSLai Jiangshan * rescuer is done with it. 205677668c8bSLai Jiangshan */ 205777668c8bSLai Jiangshan get_pwq(pwq); 2058493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2059e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2060493a1724STejun Heo } 2061e22bee78STejun Heo } 2062e22bee78STejun Heo 206332a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2064e22bee78STejun Heo { 206532a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2066e22bee78STejun Heo struct work_struct *work; 2067e22bee78STejun Heo 2068a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2069a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2070e22bee78STejun Heo 207163d95a91STejun Heo if (need_to_create_worker(pool)) { 2072e22bee78STejun Heo /* 2073e22bee78STejun Heo * We've been trying to create a new worker but 2074e22bee78STejun Heo * haven't been successful. We might be hitting an 2075e22bee78STejun Heo * allocation deadlock. Send distress signals to 2076e22bee78STejun Heo * rescuers. 2077e22bee78STejun Heo */ 207863d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2079e22bee78STejun Heo send_mayday(work); 2080e22bee78STejun Heo } 2081e22bee78STejun Heo 2082a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2083a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2084e22bee78STejun Heo 208563d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2086e22bee78STejun Heo } 2087e22bee78STejun Heo 2088e22bee78STejun Heo /** 2089e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 209063d95a91STejun Heo * @pool: pool to create a new worker for 2091e22bee78STejun Heo * 209263d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2093e22bee78STejun Heo * have at least one idle worker on return from this function. If 2094e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 209563d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2096e22bee78STejun Heo * possible allocation deadlock. 2097e22bee78STejun Heo * 2098c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2099c5aa87bbSTejun Heo * may_start_working() %true. 2100e22bee78STejun Heo * 2101e22bee78STejun Heo * LOCKING: 2102a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2103e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2104e22bee78STejun Heo * manager. 2105e22bee78STejun Heo */ 210629187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2107d565ed63STejun Heo __releases(&pool->lock) 2108d565ed63STejun Heo __acquires(&pool->lock) 2109e22bee78STejun Heo { 2110e22bee78STejun Heo restart: 2111a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 21129f9c2364STejun Heo 2113e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 211463d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2115e22bee78STejun Heo 2116e22bee78STejun Heo while (true) { 2117051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2118e22bee78STejun Heo break; 2119e22bee78STejun Heo 2120e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 21219f9c2364STejun Heo 212263d95a91STejun Heo if (!need_to_create_worker(pool)) 2123e22bee78STejun Heo break; 2124e22bee78STejun Heo } 2125e22bee78STejun Heo 212663d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2127a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2128051e1850SLai Jiangshan /* 2129051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2130051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2131051e1850SLai Jiangshan * already become busy. 2132051e1850SLai Jiangshan */ 213363d95a91STejun Heo if (need_to_create_worker(pool)) 2134e22bee78STejun Heo goto restart; 2135e22bee78STejun Heo } 2136e22bee78STejun Heo 2137e22bee78STejun Heo /** 2138e22bee78STejun Heo * manage_workers - manage worker pool 2139e22bee78STejun Heo * @worker: self 2140e22bee78STejun Heo * 2141706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2142e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2143706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2144e22bee78STejun Heo * 2145e22bee78STejun Heo * The caller can safely start processing works on false return. On 2146e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2147e22bee78STejun Heo * and may_start_working() is true. 2148e22bee78STejun Heo * 2149e22bee78STejun Heo * CONTEXT: 2150a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2151e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2152e22bee78STejun Heo * 2153d185af30SYacine Belkadi * Return: 215429187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 215529187a9eSTejun Heo * start processing works, %true if management function was performed and 215629187a9eSTejun Heo * the conditions that the caller verified before calling the function may 215729187a9eSTejun Heo * no longer be true. 2158e22bee78STejun Heo */ 2159e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2160e22bee78STejun Heo { 216163d95a91STejun Heo struct worker_pool *pool = worker->pool; 2162e22bee78STejun Heo 2163692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 216429187a9eSTejun Heo return false; 2165692b4825STejun Heo 2166692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 21672607d7a6STejun Heo pool->manager = worker; 2168e22bee78STejun Heo 216929187a9eSTejun Heo maybe_create_worker(pool); 2170e22bee78STejun Heo 21712607d7a6STejun Heo pool->manager = NULL; 2172692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2173d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 217429187a9eSTejun Heo return true; 2175e22bee78STejun Heo } 2176e22bee78STejun Heo 2177a62428c0STejun Heo /** 2178a62428c0STejun Heo * process_one_work - process single work 2179c34056a3STejun Heo * @worker: self 2180a62428c0STejun Heo * @work: work to process 2181a62428c0STejun Heo * 2182a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2183a62428c0STejun Heo * process a single work including synchronization against and 2184a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2185a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2186a62428c0STejun Heo * call this function to process a work. 2187a62428c0STejun Heo * 2188a62428c0STejun Heo * CONTEXT: 2189a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2190a62428c0STejun Heo */ 2191c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2192d565ed63STejun Heo __releases(&pool->lock) 2193d565ed63STejun Heo __acquires(&pool->lock) 21941da177e4SLinus Torvalds { 2195112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2196bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2197112202d9STejun Heo bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE; 2198c4560c2cSLai Jiangshan unsigned long work_data; 21997e11629dSTejun Heo struct worker *collision; 22004e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 22014e6045f1SJohannes Berg /* 2202a62428c0STejun Heo * It is permissible to free the struct work_struct from 2203a62428c0STejun Heo * inside the function that is called from it, this we need to 2204a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2205a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2206a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 22074e6045f1SJohannes Berg */ 22084d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 22094d82a1deSPeter Zijlstra 22104d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 22114e6045f1SJohannes Berg #endif 2212807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 221385327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2214ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 221525511a47STejun Heo 22167e11629dSTejun Heo /* 22177e11629dSTejun Heo * A single work shouldn't be executed concurrently by 22187e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 22197e11629dSTejun Heo * already processing the work. If so, defer the work to the 22207e11629dSTejun Heo * currently executing one. 22217e11629dSTejun Heo */ 2222c9e7cf27STejun Heo collision = find_worker_executing_work(pool, work); 22237e11629dSTejun Heo if (unlikely(collision)) { 22247e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 22257e11629dSTejun Heo return; 22267e11629dSTejun Heo } 22271da177e4SLinus Torvalds 22288930cabaSTejun Heo /* claim and dequeue */ 22291da177e4SLinus Torvalds debug_work_deactivate(work); 2230c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2231c34056a3STejun Heo worker->current_work = work; 2232a2c1c57bSTejun Heo worker->current_func = work->func; 2233112202d9STejun Heo worker->current_pwq = pwq; 2234c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2235d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 22367a22ad75STejun Heo 22378bf89593STejun Heo /* 22388bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 22398bf89593STejun Heo * overridden through set_worker_desc(). 22408bf89593STejun Heo */ 22418bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 22428bf89593STejun Heo 2243a62428c0STejun Heo list_del_init(&work->entry); 2244a62428c0STejun Heo 2245649027d7STejun Heo /* 2246228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2247228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2248228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2249228f1d00SLai Jiangshan * execution of the pending work items. 2250fb0e7bebSTejun Heo */ 2251fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2252228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2253fb0e7bebSTejun Heo 2254974271c4STejun Heo /* 2255a489a03eSLai Jiangshan * Wake up another worker if necessary. The condition is always 2256a489a03eSLai Jiangshan * false for normal per-cpu workers since nr_running would always 2257a489a03eSLai Jiangshan * be >= 1 at this point. This is used to chain execution of the 2258a489a03eSLai Jiangshan * pending work items for WORKER_NOT_RUNNING workers such as the 2259228f1d00SLai Jiangshan * UNBOUND and CPU_INTENSIVE ones. 2260974271c4STejun Heo */ 2261a489a03eSLai Jiangshan if (need_more_worker(pool)) 226263d95a91STejun Heo wake_up_worker(pool); 2263974271c4STejun Heo 22648930cabaSTejun Heo /* 22657c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2266d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 226723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 226823657bb1STejun Heo * disabled. 22698930cabaSTejun Heo */ 22707c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 22711da177e4SLinus Torvalds 2272a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2273365970a1SDavid Howells 2274a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 22753295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2276e6f3faa7SPeter Zijlstra /* 2277f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2278f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2279e6f3faa7SPeter Zijlstra * 2280e6f3faa7SPeter Zijlstra * However, that would result in: 2281e6f3faa7SPeter Zijlstra * 2282e6f3faa7SPeter Zijlstra * A(W1) 2283e6f3faa7SPeter Zijlstra * WFC(C) 2284e6f3faa7SPeter Zijlstra * A(W1) 2285e6f3faa7SPeter Zijlstra * C(C) 2286e6f3faa7SPeter Zijlstra * 2287e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2288e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2289e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2290f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2291e6f3faa7SPeter Zijlstra * these locks. 2292e6f3faa7SPeter Zijlstra * 2293e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2294e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2295e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2296e6f3faa7SPeter Zijlstra */ 2297f52be570SPeter Zijlstra lockdep_invariant_state(true); 2298e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2299a2c1c57bSTejun Heo worker->current_func(work); 2300e36c886aSArjan van de Ven /* 2301e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2302e36c886aSArjan van de Ven * point will only record its address. 2303e36c886aSArjan van de Ven */ 23041c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 23053295f0efSIngo Molnar lock_map_release(&lockdep_map); 2306112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 23071da177e4SLinus Torvalds 2308d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2309044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2310d75f773cSSakari Ailus " last function: %ps\n", 2311a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2312a2c1c57bSTejun Heo worker->current_func); 2313d5abe669SPeter Zijlstra debug_show_held_locks(current); 2314d5abe669SPeter Zijlstra dump_stack(); 2315d5abe669SPeter Zijlstra } 2316d5abe669SPeter Zijlstra 2317b22ce278STejun Heo /* 2318025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2319b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2320b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2321b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2322789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2323789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2324b22ce278STejun Heo */ 2325a7e6425eSPaul E. McKenney cond_resched(); 2326b22ce278STejun Heo 2327a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2328a62428c0STejun Heo 2329fb0e7bebSTejun Heo /* clear cpu intensive status */ 2330fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2331fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2332fb0e7bebSTejun Heo 23331b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 23341b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 23351b69ac6bSJohannes Weiner 2336a62428c0STejun Heo /* we're done with it, release */ 233742f8570fSSasha Levin hash_del(&worker->hentry); 2338c34056a3STejun Heo worker->current_work = NULL; 2339a2c1c57bSTejun Heo worker->current_func = NULL; 2340112202d9STejun Heo worker->current_pwq = NULL; 2341d812796eSLai Jiangshan worker->current_color = INT_MAX; 2342c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 23431da177e4SLinus Torvalds } 23441da177e4SLinus Torvalds 2345affee4b2STejun Heo /** 2346affee4b2STejun Heo * process_scheduled_works - process scheduled works 2347affee4b2STejun Heo * @worker: self 2348affee4b2STejun Heo * 2349affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2350affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2351affee4b2STejun Heo * fetches a work from the top and executes it. 2352affee4b2STejun Heo * 2353affee4b2STejun Heo * CONTEXT: 2354a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2355affee4b2STejun Heo * multiple times. 2356affee4b2STejun Heo */ 2357affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 23581da177e4SLinus Torvalds { 2359affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2360affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2361a62428c0STejun Heo struct work_struct, entry); 2362c34056a3STejun Heo process_one_work(worker, work); 2363a62428c0STejun Heo } 23641da177e4SLinus Torvalds } 23651da177e4SLinus Torvalds 2366197f6accSTejun Heo static void set_pf_worker(bool val) 2367197f6accSTejun Heo { 2368197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2369197f6accSTejun Heo if (val) 2370197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2371197f6accSTejun Heo else 2372197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2373197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2374197f6accSTejun Heo } 2375197f6accSTejun Heo 23764690c4abSTejun Heo /** 23774690c4abSTejun Heo * worker_thread - the worker thread function 2378c34056a3STejun Heo * @__worker: self 23794690c4abSTejun Heo * 2380c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2381c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2382c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2383c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2384c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2385d185af30SYacine Belkadi * 2386d185af30SYacine Belkadi * Return: 0 23874690c4abSTejun Heo */ 2388c34056a3STejun Heo static int worker_thread(void *__worker) 23891da177e4SLinus Torvalds { 2390c34056a3STejun Heo struct worker *worker = __worker; 2391bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 23921da177e4SLinus Torvalds 2393e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2394197f6accSTejun Heo set_pf_worker(true); 2395c8e55f36STejun Heo woke_up: 2396a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2397affee4b2STejun Heo 2398a9ab775bSTejun Heo /* am I supposed to die? */ 2399a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2400a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2401a9ab775bSTejun Heo WARN_ON_ONCE(!list_empty(&worker->entry)); 2402197f6accSTejun Heo set_pf_worker(false); 240360f5a4bcSLai Jiangshan 240460f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2405e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2406a2d812a2STejun Heo worker_detach_from_pool(worker); 240760f5a4bcSLai Jiangshan kfree(worker); 2408c8e55f36STejun Heo return 0; 2409c8e55f36STejun Heo } 2410c8e55f36STejun Heo 2411c8e55f36STejun Heo worker_leave_idle(worker); 2412db7bccf4STejun Heo recheck: 2413e22bee78STejun Heo /* no more worker necessary? */ 241463d95a91STejun Heo if (!need_more_worker(pool)) 2415e22bee78STejun Heo goto sleep; 2416e22bee78STejun Heo 2417e22bee78STejun Heo /* do we need to manage? */ 241863d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2419e22bee78STejun Heo goto recheck; 2420e22bee78STejun Heo 2421c8e55f36STejun Heo /* 2422c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2423c8e55f36STejun Heo * preparing to process a work or actually processing it. 2424c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2425c8e55f36STejun Heo */ 24266183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2427c8e55f36STejun Heo 2428e22bee78STejun Heo /* 2429a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2430a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2431a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2432a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2433a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2434e22bee78STejun Heo */ 2435a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2436e22bee78STejun Heo 2437e22bee78STejun Heo do { 2438affee4b2STejun Heo struct work_struct *work = 2439bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2440affee4b2STejun Heo struct work_struct, entry); 2441affee4b2STejun Heo 244282607adcSTejun Heo pool->watchdog_ts = jiffies; 244382607adcSTejun Heo 2444c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2445affee4b2STejun Heo /* optimization path, not strictly necessary */ 2446affee4b2STejun Heo process_one_work(worker, work); 2447affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2448affee4b2STejun Heo process_scheduled_works(worker); 2449affee4b2STejun Heo } else { 2450c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2451affee4b2STejun Heo process_scheduled_works(worker); 2452affee4b2STejun Heo } 245363d95a91STejun Heo } while (keep_working(pool)); 2454affee4b2STejun Heo 2455228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2456d313dd85STejun Heo sleep: 2457c8e55f36STejun Heo /* 2458d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2459d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2460d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2461d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2462d565ed63STejun Heo * event. 2463c8e55f36STejun Heo */ 2464c8e55f36STejun Heo worker_enter_idle(worker); 2465c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2466a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24671da177e4SLinus Torvalds schedule(); 2468c8e55f36STejun Heo goto woke_up; 24691da177e4SLinus Torvalds } 24701da177e4SLinus Torvalds 2471e22bee78STejun Heo /** 2472e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2473111c225aSTejun Heo * @__rescuer: self 2474e22bee78STejun Heo * 2475e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2476493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2477e22bee78STejun Heo * 2478706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2479e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2480e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2481e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2482e22bee78STejun Heo * the problem rescuer solves. 2483e22bee78STejun Heo * 2484706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2485706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2486e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2487e22bee78STejun Heo * 2488e22bee78STejun Heo * This should happen rarely. 2489d185af30SYacine Belkadi * 2490d185af30SYacine Belkadi * Return: 0 2491e22bee78STejun Heo */ 2492111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2493e22bee78STejun Heo { 2494111c225aSTejun Heo struct worker *rescuer = __rescuer; 2495111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 2496e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 24974d595b86SLai Jiangshan bool should_stop; 2498e22bee78STejun Heo 2499e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2500111c225aSTejun Heo 2501111c225aSTejun Heo /* 2502111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2503111c225aSTejun Heo * doesn't participate in concurrency management. 2504111c225aSTejun Heo */ 2505197f6accSTejun Heo set_pf_worker(true); 2506e22bee78STejun Heo repeat: 2507c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 25081da177e4SLinus Torvalds 25094d595b86SLai Jiangshan /* 25104d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 25114d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 25124d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 25134d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 25144d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 25154d595b86SLai Jiangshan * list is always empty on exit. 25164d595b86SLai Jiangshan */ 25174d595b86SLai Jiangshan should_stop = kthread_should_stop(); 25181da177e4SLinus Torvalds 2519493a1724STejun Heo /* see whether any pwq is asking for help */ 2520a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2521493a1724STejun Heo 2522493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2523493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2524493a1724STejun Heo struct pool_workqueue, mayday_node); 2525112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2526e22bee78STejun Heo struct work_struct *work, *n; 252782607adcSTejun Heo bool first = true; 2528e22bee78STejun Heo 2529e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2530493a1724STejun Heo list_del_init(&pwq->mayday_node); 2531493a1724STejun Heo 2532a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2533e22bee78STejun Heo 253451697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 253551697d39SLai Jiangshan 2536a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2537e22bee78STejun Heo 2538e22bee78STejun Heo /* 2539e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2540e22bee78STejun Heo * process'em. 2541e22bee78STejun Heo */ 25420479c8c5STejun Heo WARN_ON_ONCE(!list_empty(scheduled)); 254382607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 254482607adcSTejun Heo if (get_work_pwq(work) == pwq) { 254582607adcSTejun Heo if (first) 254682607adcSTejun Heo pool->watchdog_ts = jiffies; 2547e22bee78STejun Heo move_linked_works(work, scheduled, &n); 254882607adcSTejun Heo } 254982607adcSTejun Heo first = false; 255082607adcSTejun Heo } 2551e22bee78STejun Heo 2552008847f6SNeilBrown if (!list_empty(scheduled)) { 2553e22bee78STejun Heo process_scheduled_works(rescuer); 25547576958aSTejun Heo 25557576958aSTejun Heo /* 2556008847f6SNeilBrown * The above execution of rescued work items could 2557008847f6SNeilBrown * have created more to rescue through 2558f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2559008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2560008847f6SNeilBrown * that such back-to-back work items, which may be 2561008847f6SNeilBrown * being used to relieve memory pressure, don't 2562008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2563008847f6SNeilBrown */ 25644f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2565a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2566e66b39afSTejun Heo /* 2567e66b39afSTejun Heo * Queue iff we aren't racing destruction 2568e66b39afSTejun Heo * and somebody else hasn't queued it already. 2569e66b39afSTejun Heo */ 2570e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2571008847f6SNeilBrown get_pwq(pwq); 2572e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2573e66b39afSTejun Heo } 2574a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2575008847f6SNeilBrown } 2576008847f6SNeilBrown } 2577008847f6SNeilBrown 2578008847f6SNeilBrown /* 257977668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 258013b1d625SLai Jiangshan * go away while we're still attached to it. 258177668c8bSLai Jiangshan */ 258277668c8bSLai Jiangshan put_pwq(pwq); 258377668c8bSLai Jiangshan 258477668c8bSLai Jiangshan /* 2585d8ca83e6SLai Jiangshan * Leave this pool. If need_more_worker() is %true, notify a 25867576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 25877576958aSTejun Heo * and stalling the execution. 25887576958aSTejun Heo */ 2589d8ca83e6SLai Jiangshan if (need_more_worker(pool)) 259063d95a91STejun Heo wake_up_worker(pool); 25917576958aSTejun Heo 2592a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 259313b1d625SLai Jiangshan 2594a2d812a2STejun Heo worker_detach_from_pool(rescuer); 259513b1d625SLai Jiangshan 2596a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 25971da177e4SLinus Torvalds } 25981da177e4SLinus Torvalds 2599a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2600493a1724STejun Heo 26014d595b86SLai Jiangshan if (should_stop) { 26024d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2603197f6accSTejun Heo set_pf_worker(false); 26044d595b86SLai Jiangshan return 0; 26054d595b86SLai Jiangshan } 26064d595b86SLai Jiangshan 2607111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2608111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2609e22bee78STejun Heo schedule(); 2610e22bee78STejun Heo goto repeat; 26111da177e4SLinus Torvalds } 26121da177e4SLinus Torvalds 2613fca839c0STejun Heo /** 2614fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2615fca839c0STejun Heo * @target_wq: workqueue being flushed 2616fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2617fca839c0STejun Heo * 2618fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2619fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2620fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2621fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2622fca839c0STejun Heo * a deadlock. 2623fca839c0STejun Heo */ 2624fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2625fca839c0STejun Heo struct work_struct *target_work) 2626fca839c0STejun Heo { 2627fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2628fca839c0STejun Heo struct worker *worker; 2629fca839c0STejun Heo 2630fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2631fca839c0STejun Heo return; 2632fca839c0STejun Heo 2633fca839c0STejun Heo worker = current_wq_worker(); 2634fca839c0STejun Heo 2635fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2636d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2637fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 263823d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 263923d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2640d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2641fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2642fca839c0STejun Heo target_wq->name, target_func); 2643fca839c0STejun Heo } 2644fca839c0STejun Heo 2645fc2e4d70SOleg Nesterov struct wq_barrier { 2646fc2e4d70SOleg Nesterov struct work_struct work; 2647fc2e4d70SOleg Nesterov struct completion done; 26482607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2649fc2e4d70SOleg Nesterov }; 2650fc2e4d70SOleg Nesterov 2651fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2652fc2e4d70SOleg Nesterov { 2653fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2654fc2e4d70SOleg Nesterov complete(&barr->done); 2655fc2e4d70SOleg Nesterov } 2656fc2e4d70SOleg Nesterov 26574690c4abSTejun Heo /** 26584690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2659112202d9STejun Heo * @pwq: pwq to insert barrier into 26604690c4abSTejun Heo * @barr: wq_barrier to insert 2661affee4b2STejun Heo * @target: target work to attach @barr to 2662affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 26634690c4abSTejun Heo * 2664affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2665affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2666affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2667affee4b2STejun Heo * cpu. 2668affee4b2STejun Heo * 2669affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2670affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2671affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2672affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2673affee4b2STejun Heo * after a work with LINKED flag set. 2674affee4b2STejun Heo * 2675affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2676112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 26774690c4abSTejun Heo * 26784690c4abSTejun Heo * CONTEXT: 2679a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 26804690c4abSTejun Heo */ 2681112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2682affee4b2STejun Heo struct wq_barrier *barr, 2683affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2684fc2e4d70SOleg Nesterov { 2685d812796eSLai Jiangshan unsigned int work_flags = 0; 2686d812796eSLai Jiangshan unsigned int work_color; 2687affee4b2STejun Heo struct list_head *head; 2688affee4b2STejun Heo 2689dc186ad7SThomas Gleixner /* 2690d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2691dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2692dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2693dc186ad7SThomas Gleixner * might deadlock. 2694dc186ad7SThomas Gleixner */ 2695ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 269622df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 269752fa5bc5SBoqun Feng 2698fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2699fd1a5b04SByungchul Park 27002607d7a6STejun Heo barr->task = current; 270183c22520SOleg Nesterov 2702018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2703018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2704018f3a13SLai Jiangshan 2705affee4b2STejun Heo /* 2706affee4b2STejun Heo * If @target is currently being executed, schedule the 2707affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2708affee4b2STejun Heo */ 2709d812796eSLai Jiangshan if (worker) { 2710affee4b2STejun Heo head = worker->scheduled.next; 2711d812796eSLai Jiangshan work_color = worker->current_color; 2712d812796eSLai Jiangshan } else { 2713affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2714affee4b2STejun Heo 2715affee4b2STejun Heo head = target->entry.next; 2716affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2717d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 2718d812796eSLai Jiangshan work_color = get_work_color(*bits); 2719affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2720affee4b2STejun Heo } 2721affee4b2STejun Heo 2722d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 2723d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 2724d812796eSLai Jiangshan 2725dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2726d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 2727fc2e4d70SOleg Nesterov } 2728fc2e4d70SOleg Nesterov 272973f53c4aSTejun Heo /** 2730112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 273173f53c4aSTejun Heo * @wq: workqueue being flushed 273273f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 273373f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 273473f53c4aSTejun Heo * 2735112202d9STejun Heo * Prepare pwqs for workqueue flushing. 273673f53c4aSTejun Heo * 2737112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 2738112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 2739112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 2740112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 2741112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 274273f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 274373f53c4aSTejun Heo * 274473f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 274573f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 274673f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 274773f53c4aSTejun Heo * is returned. 274873f53c4aSTejun Heo * 2749112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 275073f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 275173f53c4aSTejun Heo * advanced to @work_color. 275273f53c4aSTejun Heo * 275373f53c4aSTejun Heo * CONTEXT: 27543c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 275573f53c4aSTejun Heo * 2756d185af30SYacine Belkadi * Return: 275773f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 275873f53c4aSTejun Heo * otherwise. 275973f53c4aSTejun Heo */ 2760112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 276173f53c4aSTejun Heo int flush_color, int work_color) 27621da177e4SLinus Torvalds { 276373f53c4aSTejun Heo bool wait = false; 276449e3cf44STejun Heo struct pool_workqueue *pwq; 27651da177e4SLinus Torvalds 276673f53c4aSTejun Heo if (flush_color >= 0) { 27676183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 2768112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 2769dc186ad7SThomas Gleixner } 277014441960SOleg Nesterov 277149e3cf44STejun Heo for_each_pwq(pwq, wq) { 2772112202d9STejun Heo struct worker_pool *pool = pwq->pool; 27731da177e4SLinus Torvalds 2774a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 277573f53c4aSTejun Heo 277673f53c4aSTejun Heo if (flush_color >= 0) { 27776183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 277873f53c4aSTejun Heo 2779112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 2780112202d9STejun Heo pwq->flush_color = flush_color; 2781112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 278273f53c4aSTejun Heo wait = true; 27831da177e4SLinus Torvalds } 278473f53c4aSTejun Heo } 278573f53c4aSTejun Heo 278673f53c4aSTejun Heo if (work_color >= 0) { 27876183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 2788112202d9STejun Heo pwq->work_color = work_color; 278973f53c4aSTejun Heo } 279073f53c4aSTejun Heo 2791a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27921da177e4SLinus Torvalds } 27931da177e4SLinus Torvalds 2794112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 279573f53c4aSTejun Heo complete(&wq->first_flusher->done); 279673f53c4aSTejun Heo 279773f53c4aSTejun Heo return wait; 279883c22520SOleg Nesterov } 27991da177e4SLinus Torvalds 28000fcb78c2SRolf Eike Beer /** 28011da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 28020fcb78c2SRolf Eike Beer * @wq: workqueue to flush 28031da177e4SLinus Torvalds * 2804c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 2805c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 28061da177e4SLinus Torvalds */ 28077ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 28081da177e4SLinus Torvalds { 280973f53c4aSTejun Heo struct wq_flusher this_flusher = { 281073f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 281173f53c4aSTejun Heo .flush_color = -1, 2812fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 281373f53c4aSTejun Heo }; 281473f53c4aSTejun Heo int next_color; 2815b1f4ec17SOleg Nesterov 28163347fa09STejun Heo if (WARN_ON(!wq_online)) 28173347fa09STejun Heo return; 28183347fa09STejun Heo 281987915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 282087915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 282187915adcSJohannes Berg 28223c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 282373f53c4aSTejun Heo 282473f53c4aSTejun Heo /* 282573f53c4aSTejun Heo * Start-to-wait phase 282673f53c4aSTejun Heo */ 282773f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 282873f53c4aSTejun Heo 282973f53c4aSTejun Heo if (next_color != wq->flush_color) { 283073f53c4aSTejun Heo /* 283173f53c4aSTejun Heo * Color space is not full. The current work_color 283273f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 283373f53c4aSTejun Heo * by one. 283473f53c4aSTejun Heo */ 28356183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 283673f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 283773f53c4aSTejun Heo wq->work_color = next_color; 283873f53c4aSTejun Heo 283973f53c4aSTejun Heo if (!wq->first_flusher) { 284073f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 28416183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 284273f53c4aSTejun Heo 284373f53c4aSTejun Heo wq->first_flusher = &this_flusher; 284473f53c4aSTejun Heo 2845112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 284673f53c4aSTejun Heo wq->work_color)) { 284773f53c4aSTejun Heo /* nothing to flush, done */ 284873f53c4aSTejun Heo wq->flush_color = next_color; 284973f53c4aSTejun Heo wq->first_flusher = NULL; 285073f53c4aSTejun Heo goto out_unlock; 285173f53c4aSTejun Heo } 285273f53c4aSTejun Heo } else { 285373f53c4aSTejun Heo /* wait in queue */ 28546183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 285573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 2856112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 285773f53c4aSTejun Heo } 285873f53c4aSTejun Heo } else { 285973f53c4aSTejun Heo /* 286073f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 286173f53c4aSTejun Heo * The next flush completion will assign us 286273f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 286373f53c4aSTejun Heo */ 286473f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 286573f53c4aSTejun Heo } 286673f53c4aSTejun Heo 2867fca839c0STejun Heo check_flush_dependency(wq, NULL); 2868fca839c0STejun Heo 28693c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 287073f53c4aSTejun Heo 287173f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 287273f53c4aSTejun Heo 287373f53c4aSTejun Heo /* 287473f53c4aSTejun Heo * Wake-up-and-cascade phase 287573f53c4aSTejun Heo * 287673f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 287773f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 287873f53c4aSTejun Heo */ 287900d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 288073f53c4aSTejun Heo return; 288173f53c4aSTejun Heo 28823c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 288373f53c4aSTejun Heo 28844ce48b37STejun Heo /* we might have raced, check again with mutex held */ 28854ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 28864ce48b37STejun Heo goto out_unlock; 28874ce48b37STejun Heo 288800d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 288973f53c4aSTejun Heo 28906183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 28916183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 289273f53c4aSTejun Heo 289373f53c4aSTejun Heo while (true) { 289473f53c4aSTejun Heo struct wq_flusher *next, *tmp; 289573f53c4aSTejun Heo 289673f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 289773f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 289873f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 289973f53c4aSTejun Heo break; 290073f53c4aSTejun Heo list_del_init(&next->list); 290173f53c4aSTejun Heo complete(&next->done); 290273f53c4aSTejun Heo } 290373f53c4aSTejun Heo 29046183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 290573f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 290673f53c4aSTejun Heo 290773f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 290873f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 290973f53c4aSTejun Heo 291073f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 291173f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 291273f53c4aSTejun Heo /* 291373f53c4aSTejun Heo * Assign the same color to all overflowed 291473f53c4aSTejun Heo * flushers, advance work_color and append to 291573f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 291673f53c4aSTejun Heo * phase for these overflowed flushers. 291773f53c4aSTejun Heo */ 291873f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 291973f53c4aSTejun Heo tmp->flush_color = wq->work_color; 292073f53c4aSTejun Heo 292173f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 292273f53c4aSTejun Heo 292373f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 292473f53c4aSTejun Heo &wq->flusher_queue); 2925112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 292673f53c4aSTejun Heo } 292773f53c4aSTejun Heo 292873f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 29296183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 293073f53c4aSTejun Heo break; 293173f53c4aSTejun Heo } 293273f53c4aSTejun Heo 293373f53c4aSTejun Heo /* 293473f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 2935112202d9STejun Heo * the new first flusher and arm pwqs. 293673f53c4aSTejun Heo */ 29376183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 29386183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 293973f53c4aSTejun Heo 294073f53c4aSTejun Heo list_del_init(&next->list); 294173f53c4aSTejun Heo wq->first_flusher = next; 294273f53c4aSTejun Heo 2943112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 294473f53c4aSTejun Heo break; 294573f53c4aSTejun Heo 294673f53c4aSTejun Heo /* 294773f53c4aSTejun Heo * Meh... this color is already done, clear first 294873f53c4aSTejun Heo * flusher and repeat cascading. 294973f53c4aSTejun Heo */ 295073f53c4aSTejun Heo wq->first_flusher = NULL; 295173f53c4aSTejun Heo } 295273f53c4aSTejun Heo 295373f53c4aSTejun Heo out_unlock: 29543c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 29551da177e4SLinus Torvalds } 29561dadafa8STim Gardner EXPORT_SYMBOL(flush_workqueue); 29571da177e4SLinus Torvalds 29589c5a2ba7STejun Heo /** 29599c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 29609c5a2ba7STejun Heo * @wq: workqueue to drain 29619c5a2ba7STejun Heo * 29629c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 29639c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 29649c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 2965b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 29669c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 29679c5a2ba7STejun Heo * takes too long. 29689c5a2ba7STejun Heo */ 29699c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 29709c5a2ba7STejun Heo { 29719c5a2ba7STejun Heo unsigned int flush_cnt = 0; 297249e3cf44STejun Heo struct pool_workqueue *pwq; 29739c5a2ba7STejun Heo 29749c5a2ba7STejun Heo /* 29759c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 29769c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 2977618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 29789c5a2ba7STejun Heo */ 297987fc741eSLai Jiangshan mutex_lock(&wq->mutex); 29809c5a2ba7STejun Heo if (!wq->nr_drainers++) 2981618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 298287fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 29839c5a2ba7STejun Heo reflush: 29849c5a2ba7STejun Heo flush_workqueue(wq); 29859c5a2ba7STejun Heo 2986b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 298776af4d93STejun Heo 298849e3cf44STejun Heo for_each_pwq(pwq, wq) { 2989fa2563e4SThomas Tuttle bool drained; 29909c5a2ba7STejun Heo 2991a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 2992f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 2993a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 2994fa2563e4SThomas Tuttle 2995fa2563e4SThomas Tuttle if (drained) 29969c5a2ba7STejun Heo continue; 29979c5a2ba7STejun Heo 29989c5a2ba7STejun Heo if (++flush_cnt == 10 || 29999c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3000e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3001e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 300276af4d93STejun Heo 3003b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 30049c5a2ba7STejun Heo goto reflush; 30059c5a2ba7STejun Heo } 30069c5a2ba7STejun Heo 30079c5a2ba7STejun Heo if (!--wq->nr_drainers) 3008618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 300987fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 30109c5a2ba7STejun Heo } 30119c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 30129c5a2ba7STejun Heo 3013d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3014d6e89786SJohannes Berg bool from_cancel) 3015baf59022STejun Heo { 3016baf59022STejun Heo struct worker *worker = NULL; 3017c9e7cf27STejun Heo struct worker_pool *pool; 3018112202d9STejun Heo struct pool_workqueue *pwq; 3019baf59022STejun Heo 3020baf59022STejun Heo might_sleep(); 3021baf59022STejun Heo 302224acfb71SThomas Gleixner rcu_read_lock(); 3023fa1b54e6STejun Heo pool = get_work_pool(work); 3024fa1b54e6STejun Heo if (!pool) { 302524acfb71SThomas Gleixner rcu_read_unlock(); 3026fa1b54e6STejun Heo return false; 3027fa1b54e6STejun Heo } 3028fa1b54e6STejun Heo 3029a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 30300b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3031112202d9STejun Heo pwq = get_work_pwq(work); 3032112202d9STejun Heo if (pwq) { 3033112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3034baf59022STejun Heo goto already_gone; 3035606a5020STejun Heo } else { 3036c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3037baf59022STejun Heo if (!worker) 3038baf59022STejun Heo goto already_gone; 3039112202d9STejun Heo pwq = worker->current_pwq; 3040606a5020STejun Heo } 3041baf59022STejun Heo 3042fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3043fca839c0STejun Heo 3044112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3045a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3046baf59022STejun Heo 3047e159489bSTejun Heo /* 3048a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3049a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3050a1d14934SPeter Zijlstra * 3051a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3052a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3053a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3054a1d14934SPeter Zijlstra * forward progress. 3055e159489bSTejun Heo */ 3056d6e89786SJohannes Berg if (!from_cancel && 3057d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3058112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3059112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3060a1d14934SPeter Zijlstra } 306124acfb71SThomas Gleixner rcu_read_unlock(); 3062baf59022STejun Heo return true; 3063baf59022STejun Heo already_gone: 3064a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 306524acfb71SThomas Gleixner rcu_read_unlock(); 3066baf59022STejun Heo return false; 3067baf59022STejun Heo } 3068baf59022STejun Heo 3069d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3070d6e89786SJohannes Berg { 3071d6e89786SJohannes Berg struct wq_barrier barr; 3072d6e89786SJohannes Berg 3073d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3074d6e89786SJohannes Berg return false; 3075d6e89786SJohannes Berg 30764d43d395STetsuo Handa if (WARN_ON(!work->func)) 30774d43d395STetsuo Handa return false; 30784d43d395STetsuo Handa 307987915adcSJohannes Berg if (!from_cancel) { 308087915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 308187915adcSJohannes Berg lock_map_release(&work->lockdep_map); 308287915adcSJohannes Berg } 308387915adcSJohannes Berg 3084d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3085d6e89786SJohannes Berg wait_for_completion(&barr.done); 3086d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3087d6e89786SJohannes Berg return true; 3088d6e89786SJohannes Berg } else { 3089d6e89786SJohannes Berg return false; 3090d6e89786SJohannes Berg } 3091d6e89786SJohannes Berg } 3092d6e89786SJohannes Berg 3093db700897SOleg Nesterov /** 3094401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3095401a8d04STejun Heo * @work: the work to flush 3096db700897SOleg Nesterov * 3097606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3098606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3099401a8d04STejun Heo * 3100d185af30SYacine Belkadi * Return: 3101401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3102401a8d04STejun Heo * %false if it was already idle. 3103db700897SOleg Nesterov */ 3104401a8d04STejun Heo bool flush_work(struct work_struct *work) 3105db700897SOleg Nesterov { 3106d6e89786SJohannes Berg return __flush_work(work, false); 3107606a5020STejun Heo } 3108db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3109db700897SOleg Nesterov 31108603e1b3STejun Heo struct cwt_wait { 3111ac6424b9SIngo Molnar wait_queue_entry_t wait; 31128603e1b3STejun Heo struct work_struct *work; 31138603e1b3STejun Heo }; 31148603e1b3STejun Heo 3115ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 31168603e1b3STejun Heo { 31178603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 31188603e1b3STejun Heo 31198603e1b3STejun Heo if (cwait->work != key) 31208603e1b3STejun Heo return 0; 31218603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 31228603e1b3STejun Heo } 31238603e1b3STejun Heo 312436e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3125401a8d04STejun Heo { 31268603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3127bbb68dfaSTejun Heo unsigned long flags; 31281f1f642eSOleg Nesterov int ret; 31291f1f642eSOleg Nesterov 31301f1f642eSOleg Nesterov do { 3131bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3132bbb68dfaSTejun Heo /* 31338603e1b3STejun Heo * If someone else is already canceling, wait for it to 31348603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 31358603e1b3STejun Heo * because we may get scheduled between @work's completion 31368603e1b3STejun Heo * and the other canceling task resuming and clearing 31378603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 31388603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 31398603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 31408603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 31418603e1b3STejun Heo * we're hogging the CPU. 31428603e1b3STejun Heo * 31438603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 31448603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 31458603e1b3STejun Heo * wake function which matches @work along with exclusive 31468603e1b3STejun Heo * wait and wakeup. 3147bbb68dfaSTejun Heo */ 31488603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 31498603e1b3STejun Heo struct cwt_wait cwait; 31508603e1b3STejun Heo 31518603e1b3STejun Heo init_wait(&cwait.wait); 31528603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 31538603e1b3STejun Heo cwait.work = work; 31548603e1b3STejun Heo 31558603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 31568603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 31578603e1b3STejun Heo if (work_is_canceling(work)) 31588603e1b3STejun Heo schedule(); 31598603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 31608603e1b3STejun Heo } 31611f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 31621f1f642eSOleg Nesterov 3163bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3164bbb68dfaSTejun Heo mark_work_canceling(work); 3165bbb68dfaSTejun Heo local_irq_restore(flags); 3166bbb68dfaSTejun Heo 31673347fa09STejun Heo /* 31683347fa09STejun Heo * This allows canceling during early boot. We know that @work 31693347fa09STejun Heo * isn't executing. 31703347fa09STejun Heo */ 31713347fa09STejun Heo if (wq_online) 3172d6e89786SJohannes Berg __flush_work(work, true); 31733347fa09STejun Heo 31747a22ad75STejun Heo clear_work_data(work); 31758603e1b3STejun Heo 31768603e1b3STejun Heo /* 31778603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 31788603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 31798603e1b3STejun Heo * visible there. 31808603e1b3STejun Heo */ 31818603e1b3STejun Heo smp_mb(); 31828603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 31838603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 31848603e1b3STejun Heo 31851f1f642eSOleg Nesterov return ret; 31861f1f642eSOleg Nesterov } 31871f1f642eSOleg Nesterov 31886e84d644SOleg Nesterov /** 3189401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3190401a8d04STejun Heo * @work: the work to cancel 31916e84d644SOleg Nesterov * 3192401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3193401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3194401a8d04STejun Heo * another workqueue. On return from this function, @work is 3195401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 31961f1f642eSOleg Nesterov * 3197401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3198401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 31996e84d644SOleg Nesterov * 3200401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 32016e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3202401a8d04STejun Heo * 3203d185af30SYacine Belkadi * Return: 3204401a8d04STejun Heo * %true if @work was pending, %false otherwise. 32056e84d644SOleg Nesterov */ 3206401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 32076e84d644SOleg Nesterov { 320836e227d2STejun Heo return __cancel_work_timer(work, false); 3209b89deed3SOleg Nesterov } 321028e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3211b89deed3SOleg Nesterov 32126e84d644SOleg Nesterov /** 3213401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3214401a8d04STejun Heo * @dwork: the delayed work to flush 32156e84d644SOleg Nesterov * 3216401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3217401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3218401a8d04STejun Heo * considers the last queueing instance of @dwork. 32191f1f642eSOleg Nesterov * 3220d185af30SYacine Belkadi * Return: 3221401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3222401a8d04STejun Heo * %false if it was already idle. 32236e84d644SOleg Nesterov */ 3224401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3225401a8d04STejun Heo { 32268930cabaSTejun Heo local_irq_disable(); 3227401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 322860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 32298930cabaSTejun Heo local_irq_enable(); 3230401a8d04STejun Heo return flush_work(&dwork->work); 3231401a8d04STejun Heo } 3232401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3233401a8d04STejun Heo 323405f0fe6bSTejun Heo /** 323505f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 323605f0fe6bSTejun Heo * @rwork: the rcu work to flush 323705f0fe6bSTejun Heo * 323805f0fe6bSTejun Heo * Return: 323905f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 324005f0fe6bSTejun Heo * %false if it was already idle. 324105f0fe6bSTejun Heo */ 324205f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 324305f0fe6bSTejun Heo { 324405f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 324505f0fe6bSTejun Heo rcu_barrier(); 324605f0fe6bSTejun Heo flush_work(&rwork->work); 324705f0fe6bSTejun Heo return true; 324805f0fe6bSTejun Heo } else { 324905f0fe6bSTejun Heo return flush_work(&rwork->work); 325005f0fe6bSTejun Heo } 325105f0fe6bSTejun Heo } 325205f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 325305f0fe6bSTejun Heo 3254f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3255f72b8792SJens Axboe { 3256f72b8792SJens Axboe unsigned long flags; 3257f72b8792SJens Axboe int ret; 3258f72b8792SJens Axboe 3259f72b8792SJens Axboe do { 3260f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3261f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3262f72b8792SJens Axboe 3263f72b8792SJens Axboe if (unlikely(ret < 0)) 3264f72b8792SJens Axboe return false; 3265f72b8792SJens Axboe 3266f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3267f72b8792SJens Axboe local_irq_restore(flags); 3268f72b8792SJens Axboe return ret; 3269f72b8792SJens Axboe } 3270f72b8792SJens Axboe 3271401a8d04STejun Heo /** 327257b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 327357b30ae7STejun Heo * @dwork: delayed_work to cancel 327409383498STejun Heo * 3275d185af30SYacine Belkadi * Kill off a pending delayed_work. 3276d185af30SYacine Belkadi * 3277d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3278d185af30SYacine Belkadi * pending. 3279d185af30SYacine Belkadi * 3280d185af30SYacine Belkadi * Note: 3281d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3282d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3283d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 328409383498STejun Heo * 328557b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 328609383498STejun Heo */ 328757b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 328809383498STejun Heo { 3289f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 329009383498STejun Heo } 329157b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 329209383498STejun Heo 329309383498STejun Heo /** 3294401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3295401a8d04STejun Heo * @dwork: the delayed work cancel 3296401a8d04STejun Heo * 3297401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3298401a8d04STejun Heo * 3299d185af30SYacine Belkadi * Return: 3300401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3301401a8d04STejun Heo */ 3302401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 33036e84d644SOleg Nesterov { 330436e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 33056e84d644SOleg Nesterov } 3306f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 33071da177e4SLinus Torvalds 33080fcb78c2SRolf Eike Beer /** 330931ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3310b6136773SAndrew Morton * @func: the function to call 3311b6136773SAndrew Morton * 331231ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 331331ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3314b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 331531ddd871STejun Heo * 3316d185af30SYacine Belkadi * Return: 331731ddd871STejun Heo * 0 on success, -errno on failure. 3318b6136773SAndrew Morton */ 331965f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 332015316ba8SChristoph Lameter { 332115316ba8SChristoph Lameter int cpu; 332238f51568SNamhyung Kim struct work_struct __percpu *works; 332315316ba8SChristoph Lameter 3324b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3325b6136773SAndrew Morton if (!works) 332615316ba8SChristoph Lameter return -ENOMEM; 3327b6136773SAndrew Morton 3328ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 332993981800STejun Heo 333015316ba8SChristoph Lameter for_each_online_cpu(cpu) { 33319bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 33329bfb1839SIngo Molnar 33339bfb1839SIngo Molnar INIT_WORK(work, func); 33348de6d308SOleg Nesterov schedule_work_on(cpu, work); 333515316ba8SChristoph Lameter } 333693981800STejun Heo 333793981800STejun Heo for_each_online_cpu(cpu) 33388616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 333993981800STejun Heo 3340ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3341b6136773SAndrew Morton free_percpu(works); 334215316ba8SChristoph Lameter return 0; 334315316ba8SChristoph Lameter } 334415316ba8SChristoph Lameter 3345eef6a7d5SAlan Stern /** 33461fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 33471fa44ecaSJames Bottomley * @fn: the function to execute 33481fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 33491fa44ecaSJames Bottomley * be available when the work executes) 33501fa44ecaSJames Bottomley * 33511fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 33521fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 33531fa44ecaSJames Bottomley * 3354d185af30SYacine Belkadi * Return: 0 - function was executed 33551fa44ecaSJames Bottomley * 1 - function was scheduled for execution 33561fa44ecaSJames Bottomley */ 335765f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 33581fa44ecaSJames Bottomley { 33591fa44ecaSJames Bottomley if (!in_interrupt()) { 336065f27f38SDavid Howells fn(&ew->work); 33611fa44ecaSJames Bottomley return 0; 33621fa44ecaSJames Bottomley } 33631fa44ecaSJames Bottomley 336465f27f38SDavid Howells INIT_WORK(&ew->work, fn); 33651fa44ecaSJames Bottomley schedule_work(&ew->work); 33661fa44ecaSJames Bottomley 33671fa44ecaSJames Bottomley return 1; 33681fa44ecaSJames Bottomley } 33691fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 33701fa44ecaSJames Bottomley 33717a4e344cSTejun Heo /** 33727a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 33737a4e344cSTejun Heo * @attrs: workqueue_attrs to free 33747a4e344cSTejun Heo * 33757a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 33767a4e344cSTejun Heo */ 3377513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 33787a4e344cSTejun Heo { 33797a4e344cSTejun Heo if (attrs) { 33807a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 33817a4e344cSTejun Heo kfree(attrs); 33827a4e344cSTejun Heo } 33837a4e344cSTejun Heo } 33847a4e344cSTejun Heo 33857a4e344cSTejun Heo /** 33867a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 33877a4e344cSTejun Heo * 33887a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3389d185af30SYacine Belkadi * return it. 3390d185af30SYacine Belkadi * 3391d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 33927a4e344cSTejun Heo */ 3393513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 33947a4e344cSTejun Heo { 33957a4e344cSTejun Heo struct workqueue_attrs *attrs; 33967a4e344cSTejun Heo 3397be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 33987a4e344cSTejun Heo if (!attrs) 33997a4e344cSTejun Heo goto fail; 3400be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 34017a4e344cSTejun Heo goto fail; 34027a4e344cSTejun Heo 340313e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 34047a4e344cSTejun Heo return attrs; 34057a4e344cSTejun Heo fail: 34067a4e344cSTejun Heo free_workqueue_attrs(attrs); 34077a4e344cSTejun Heo return NULL; 34087a4e344cSTejun Heo } 34097a4e344cSTejun Heo 341029c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 341129c91e99STejun Heo const struct workqueue_attrs *from) 341229c91e99STejun Heo { 341329c91e99STejun Heo to->nice = from->nice; 341429c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 34152865a8fbSShaohua Li /* 34162865a8fbSShaohua Li * Unlike hash and equality test, this function doesn't ignore 34172865a8fbSShaohua Li * ->no_numa as it is used for both pool and wq attrs. Instead, 34182865a8fbSShaohua Li * get_unbound_pool() explicitly clears ->no_numa after copying. 34192865a8fbSShaohua Li */ 34202865a8fbSShaohua Li to->no_numa = from->no_numa; 342129c91e99STejun Heo } 342229c91e99STejun Heo 342329c91e99STejun Heo /* hash value of the content of @attr */ 342429c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 342529c91e99STejun Heo { 342629c91e99STejun Heo u32 hash = 0; 342729c91e99STejun Heo 342829c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 342913e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 343013e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 343129c91e99STejun Heo return hash; 343229c91e99STejun Heo } 343329c91e99STejun Heo 343429c91e99STejun Heo /* content equality test */ 343529c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 343629c91e99STejun Heo const struct workqueue_attrs *b) 343729c91e99STejun Heo { 343829c91e99STejun Heo if (a->nice != b->nice) 343929c91e99STejun Heo return false; 344029c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 344129c91e99STejun Heo return false; 344229c91e99STejun Heo return true; 344329c91e99STejun Heo } 344429c91e99STejun Heo 34457a4e344cSTejun Heo /** 34467a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 34477a4e344cSTejun Heo * @pool: worker_pool to initialize 34487a4e344cSTejun Heo * 3449402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3450d185af30SYacine Belkadi * 3451d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 345229c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 345329c91e99STejun Heo * on @pool safely to release it. 34547a4e344cSTejun Heo */ 34557a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 34564e1a1f9aSTejun Heo { 3457a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 345829c91e99STejun Heo pool->id = -1; 345929c91e99STejun Heo pool->cpu = -1; 3460f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 34614e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 346282607adcSTejun Heo pool->watchdog_ts = jiffies; 34634e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 34644e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 34654e1a1f9aSTejun Heo hash_init(pool->busy_hash); 34664e1a1f9aSTejun Heo 346732a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 34684e1a1f9aSTejun Heo 346932a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 34704e1a1f9aSTejun Heo 3471da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 34727a4e344cSTejun Heo 34737cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 347429c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 347529c91e99STejun Heo pool->refcnt = 1; 347629c91e99STejun Heo 347729c91e99STejun Heo /* shouldn't fail above this point */ 3478be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 34797a4e344cSTejun Heo if (!pool->attrs) 34807a4e344cSTejun Heo return -ENOMEM; 34817a4e344cSTejun Heo return 0; 34824e1a1f9aSTejun Heo } 34834e1a1f9aSTejun Heo 3484669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3485669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3486669de8bdSBart Van Assche { 3487669de8bdSBart Van Assche char *lock_name; 3488669de8bdSBart Van Assche 3489669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3490669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3491669de8bdSBart Van Assche if (!lock_name) 3492669de8bdSBart Van Assche lock_name = wq->name; 349369a106c0SQian Cai 349469a106c0SQian Cai wq->lock_name = lock_name; 3495669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3496669de8bdSBart Van Assche } 3497669de8bdSBart Van Assche 3498669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3499669de8bdSBart Van Assche { 3500669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3501669de8bdSBart Van Assche } 3502669de8bdSBart Van Assche 3503669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3504669de8bdSBart Van Assche { 3505669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3506669de8bdSBart Van Assche kfree(wq->lock_name); 3507669de8bdSBart Van Assche } 3508669de8bdSBart Van Assche #else 3509669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3510669de8bdSBart Van Assche { 3511669de8bdSBart Van Assche } 3512669de8bdSBart Van Assche 3513669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3514669de8bdSBart Van Assche { 3515669de8bdSBart Van Assche } 3516669de8bdSBart Van Assche 3517669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3518669de8bdSBart Van Assche { 3519669de8bdSBart Van Assche } 3520669de8bdSBart Van Assche #endif 3521669de8bdSBart Van Assche 3522e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3523e2dca7adSTejun Heo { 3524e2dca7adSTejun Heo struct workqueue_struct *wq = 3525e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3526e2dca7adSTejun Heo 3527669de8bdSBart Van Assche wq_free_lockdep(wq); 3528669de8bdSBart Van Assche 3529e2dca7adSTejun Heo if (!(wq->flags & WQ_UNBOUND)) 3530e2dca7adSTejun Heo free_percpu(wq->cpu_pwqs); 3531e2dca7adSTejun Heo else 3532e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3533e2dca7adSTejun Heo 3534e2dca7adSTejun Heo kfree(wq); 3535e2dca7adSTejun Heo } 3536e2dca7adSTejun Heo 353729c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 353829c91e99STejun Heo { 353929c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 354029c91e99STejun Heo 35417cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 354229c91e99STejun Heo free_workqueue_attrs(pool->attrs); 354329c91e99STejun Heo kfree(pool); 354429c91e99STejun Heo } 354529c91e99STejun Heo 3546d8bb65abSSebastian Andrzej Siewior /* This returns with the lock held on success (pool manager is inactive). */ 3547d8bb65abSSebastian Andrzej Siewior static bool wq_manager_inactive(struct worker_pool *pool) 3548d8bb65abSSebastian Andrzej Siewior { 3549a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 3550d8bb65abSSebastian Andrzej Siewior 3551d8bb65abSSebastian Andrzej Siewior if (pool->flags & POOL_MANAGER_ACTIVE) { 3552a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3553d8bb65abSSebastian Andrzej Siewior return false; 3554d8bb65abSSebastian Andrzej Siewior } 3555d8bb65abSSebastian Andrzej Siewior return true; 3556d8bb65abSSebastian Andrzej Siewior } 3557d8bb65abSSebastian Andrzej Siewior 355829c91e99STejun Heo /** 355929c91e99STejun Heo * put_unbound_pool - put a worker_pool 356029c91e99STejun Heo * @pool: worker_pool to put 356129c91e99STejun Heo * 356224acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3563c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3564c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3565c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3566a892caccSTejun Heo * 3567a892caccSTejun Heo * Should be called with wq_pool_mutex held. 356829c91e99STejun Heo */ 356929c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 357029c91e99STejun Heo { 357160f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 357229c91e99STejun Heo struct worker *worker; 357329c91e99STejun Heo 3574a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3575a892caccSTejun Heo 3576a892caccSTejun Heo if (--pool->refcnt) 357729c91e99STejun Heo return; 357829c91e99STejun Heo 357929c91e99STejun Heo /* sanity checks */ 358061d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3581a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 358229c91e99STejun Heo return; 358329c91e99STejun Heo 358429c91e99STejun Heo /* release id and unhash */ 358529c91e99STejun Heo if (pool->id >= 0) 358629c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 358729c91e99STejun Heo hash_del(&pool->hash_node); 358829c91e99STejun Heo 3589c5aa87bbSTejun Heo /* 3590692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3591692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3592692b4825STejun Heo * manager and @pool gets freed with the flag set. 3593d8bb65abSSebastian Andrzej Siewior * Because of how wq_manager_inactive() works, we will hold the 3594d8bb65abSSebastian Andrzej Siewior * spinlock after a successful wait. 3595c5aa87bbSTejun Heo */ 3596d8bb65abSSebastian Andrzej Siewior rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool), 3597d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3598692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 3599692b4825STejun Heo 36001037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 360129c91e99STejun Heo destroy_worker(worker); 360229c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3603a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 360460f5a4bcSLai Jiangshan 36051258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 3606da028469SLai Jiangshan if (!list_empty(&pool->workers)) 360760f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 36081258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 360960f5a4bcSLai Jiangshan 361060f5a4bcSLai Jiangshan if (pool->detach_completion) 361160f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 361260f5a4bcSLai Jiangshan 361329c91e99STejun Heo /* shut down the timers */ 361429c91e99STejun Heo del_timer_sync(&pool->idle_timer); 361529c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 361629c91e99STejun Heo 361724acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 361825b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 361929c91e99STejun Heo } 362029c91e99STejun Heo 362129c91e99STejun Heo /** 362229c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 362329c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 362429c91e99STejun Heo * 362529c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 362629c91e99STejun Heo * reference count and return it. If there already is a matching 362729c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3628d185af30SYacine Belkadi * create a new one. 3629a892caccSTejun Heo * 3630a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3631d185af30SYacine Belkadi * 3632d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3633d185af30SYacine Belkadi * On failure, %NULL. 363429c91e99STejun Heo */ 363529c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 363629c91e99STejun Heo { 363729c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 363829c91e99STejun Heo struct worker_pool *pool; 3639f3f90ad4STejun Heo int node; 3640e2273584SXunlei Pang int target_node = NUMA_NO_NODE; 364129c91e99STejun Heo 3642a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 364329c91e99STejun Heo 364429c91e99STejun Heo /* do we already have a matching pool? */ 364529c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 364629c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 364729c91e99STejun Heo pool->refcnt++; 36483fb1823cSLai Jiangshan return pool; 364929c91e99STejun Heo } 365029c91e99STejun Heo } 365129c91e99STejun Heo 3652e2273584SXunlei Pang /* if cpumask is contained inside a NUMA node, we belong to that node */ 3653e2273584SXunlei Pang if (wq_numa_enabled) { 3654e2273584SXunlei Pang for_each_node(node) { 3655e2273584SXunlei Pang if (cpumask_subset(attrs->cpumask, 3656e2273584SXunlei Pang wq_numa_possible_cpumask[node])) { 3657e2273584SXunlei Pang target_node = node; 3658e2273584SXunlei Pang break; 3659e2273584SXunlei Pang } 3660e2273584SXunlei Pang } 3661e2273584SXunlei Pang } 3662e2273584SXunlei Pang 366329c91e99STejun Heo /* nope, create a new one */ 3664e2273584SXunlei Pang pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node); 366529c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 366629c91e99STejun Heo goto fail; 366729c91e99STejun Heo 36688864b4e5STejun Heo lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */ 366929c91e99STejun Heo copy_workqueue_attrs(pool->attrs, attrs); 3670e2273584SXunlei Pang pool->node = target_node; 367129c91e99STejun Heo 36722865a8fbSShaohua Li /* 36732865a8fbSShaohua Li * no_numa isn't a worker_pool attribute, always clear it. See 36742865a8fbSShaohua Li * 'struct workqueue_attrs' comments for detail. 36752865a8fbSShaohua Li */ 36762865a8fbSShaohua Li pool->attrs->no_numa = false; 36772865a8fbSShaohua Li 367829c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 367929c91e99STejun Heo goto fail; 368029c91e99STejun Heo 368129c91e99STejun Heo /* create and start the initial worker */ 36823347fa09STejun Heo if (wq_online && !create_worker(pool)) 368329c91e99STejun Heo goto fail; 368429c91e99STejun Heo 368529c91e99STejun Heo /* install */ 368629c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 36873fb1823cSLai Jiangshan 368829c91e99STejun Heo return pool; 368929c91e99STejun Heo fail: 369029c91e99STejun Heo if (pool) 369129c91e99STejun Heo put_unbound_pool(pool); 369229c91e99STejun Heo return NULL; 369329c91e99STejun Heo } 369429c91e99STejun Heo 36958864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 36968864b4e5STejun Heo { 36978864b4e5STejun Heo kmem_cache_free(pwq_cache, 36988864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 36998864b4e5STejun Heo } 37008864b4e5STejun Heo 37018864b4e5STejun Heo /* 37028864b4e5STejun Heo * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt 37038864b4e5STejun Heo * and needs to be destroyed. 37048864b4e5STejun Heo */ 37058864b4e5STejun Heo static void pwq_unbound_release_workfn(struct work_struct *work) 37068864b4e5STejun Heo { 37078864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 37088864b4e5STejun Heo unbound_release_work); 37098864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 37108864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 3711b42b0bddSYang Yingliang bool is_last = false; 37128864b4e5STejun Heo 3713b42b0bddSYang Yingliang /* 3714b42b0bddSYang Yingliang * when @pwq is not linked, it doesn't hold any reference to the 3715b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 3716b42b0bddSYang Yingliang */ 3717b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 37188864b4e5STejun Heo if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND))) 37198864b4e5STejun Heo return; 37208864b4e5STejun Heo 37213c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 37228864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 3723bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 37243c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 3725b42b0bddSYang Yingliang } 37268864b4e5STejun Heo 3727a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 37288864b4e5STejun Heo put_unbound_pool(pool); 3729a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 3730a892caccSTejun Heo 373125b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 37328864b4e5STejun Heo 37338864b4e5STejun Heo /* 37348864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 3735e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 37368864b4e5STejun Heo */ 3737669de8bdSBart Van Assche if (is_last) { 3738669de8bdSBart Van Assche wq_unregister_lockdep(wq); 373925b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 37406029a918STejun Heo } 3741669de8bdSBart Van Assche } 37428864b4e5STejun Heo 37430fbd95aaSTejun Heo /** 3744699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 37450fbd95aaSTejun Heo * @pwq: target pool_workqueue 37460fbd95aaSTejun Heo * 3747699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 3748f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 3749699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 37500fbd95aaSTejun Heo */ 3751699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 37520fbd95aaSTejun Heo { 3753699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 3754699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 37553347fa09STejun Heo unsigned long flags; 3756699ce097STejun Heo 3757699ce097STejun Heo /* for @wq->saved_max_active */ 3758a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 3759699ce097STejun Heo 3760699ce097STejun Heo /* fast exit for non-freezable wqs */ 3761699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 3762699ce097STejun Heo return; 3763699ce097STejun Heo 37643347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 3765a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 3766699ce097STejun Heo 376774b414eaSLai Jiangshan /* 376874b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 376974b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 377074b414eaSLai Jiangshan * is updated and visible. 377174b414eaSLai Jiangshan */ 377274b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 377301341fbdSYunfeng Ye bool kick = false; 377401341fbdSYunfeng Ye 3775699ce097STejun Heo pwq->max_active = wq->saved_max_active; 37760fbd95aaSTejun Heo 3777f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 377801341fbdSYunfeng Ye pwq->nr_active < pwq->max_active) { 3779f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 378001341fbdSYunfeng Ye kick = true; 378101341fbdSYunfeng Ye } 3782951a078aSLai Jiangshan 3783951a078aSLai Jiangshan /* 3784951a078aSLai Jiangshan * Need to kick a worker after thawed or an unbound wq's 378501341fbdSYunfeng Ye * max_active is bumped. In realtime scenarios, always kicking a 378601341fbdSYunfeng Ye * worker will cause interference on the isolated cpu cores, so 378701341fbdSYunfeng Ye * let's kick iff work items were activated. 3788951a078aSLai Jiangshan */ 378901341fbdSYunfeng Ye if (kick) 3790951a078aSLai Jiangshan wake_up_worker(pwq->pool); 3791699ce097STejun Heo } else { 3792699ce097STejun Heo pwq->max_active = 0; 3793699ce097STejun Heo } 3794699ce097STejun Heo 3795a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 37960fbd95aaSTejun Heo } 37970fbd95aaSTejun Heo 379867dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 3799f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 3800f147f29eSTejun Heo struct worker_pool *pool) 3801d2c1d404STejun Heo { 3802d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 3803d2c1d404STejun Heo 3804e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 3805e50aba9aSTejun Heo 3806d2c1d404STejun Heo pwq->pool = pool; 3807d2c1d404STejun Heo pwq->wq = wq; 3808d2c1d404STejun Heo pwq->flush_color = -1; 38098864b4e5STejun Heo pwq->refcnt = 1; 3810f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 38111befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 3812d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 38138864b4e5STejun Heo INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn); 3814f147f29eSTejun Heo } 3815d2c1d404STejun Heo 3816f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 38171befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 3818f147f29eSTejun Heo { 3819f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 3820f147f29eSTejun Heo 3821f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 382275ccf595STejun Heo 38231befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 38241befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 38251befcf30STejun Heo return; 38261befcf30STejun Heo 382729b1cb41SLai Jiangshan /* set the matching work_color */ 382875ccf595STejun Heo pwq->work_color = wq->work_color; 3829983ca25eSTejun Heo 3830983ca25eSTejun Heo /* sync max_active to the current setting */ 3831983ca25eSTejun Heo pwq_adjust_max_active(pwq); 3832983ca25eSTejun Heo 3833983ca25eSTejun Heo /* link in @pwq */ 38349e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 3835df2d5ae4STejun Heo } 38366029a918STejun Heo 3837f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 3838f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 3839f147f29eSTejun Heo const struct workqueue_attrs *attrs) 3840f147f29eSTejun Heo { 3841f147f29eSTejun Heo struct worker_pool *pool; 3842f147f29eSTejun Heo struct pool_workqueue *pwq; 3843f147f29eSTejun Heo 3844f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3845f147f29eSTejun Heo 3846f147f29eSTejun Heo pool = get_unbound_pool(attrs); 3847f147f29eSTejun Heo if (!pool) 3848f147f29eSTejun Heo return NULL; 3849f147f29eSTejun Heo 3850e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 3851f147f29eSTejun Heo if (!pwq) { 3852f147f29eSTejun Heo put_unbound_pool(pool); 3853f147f29eSTejun Heo return NULL; 3854f147f29eSTejun Heo } 3855f147f29eSTejun Heo 3856f147f29eSTejun Heo init_pwq(pwq, wq, pool); 3857f147f29eSTejun Heo return pwq; 3858d2c1d404STejun Heo } 3859d2c1d404STejun Heo 38604c16bd32STejun Heo /** 386130186c6fSGong Zhaogang * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node 3862042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 38634c16bd32STejun Heo * @node: the target NUMA node 38644c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 38654c16bd32STejun Heo * @cpumask: outarg, the resulting cpumask 38664c16bd32STejun Heo * 38674c16bd32STejun Heo * Calculate the cpumask a workqueue with @attrs should use on @node. If 38684c16bd32STejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during 3869d185af30SYacine Belkadi * calculation. The result is stored in @cpumask. 38704c16bd32STejun Heo * 38714c16bd32STejun Heo * If NUMA affinity is not enabled, @attrs->cpumask is always used. If 38724c16bd32STejun Heo * enabled and @node has online CPUs requested by @attrs, the returned 38734c16bd32STejun Heo * cpumask is the intersection of the possible CPUs of @node and 38744c16bd32STejun Heo * @attrs->cpumask. 38754c16bd32STejun Heo * 38764c16bd32STejun Heo * The caller is responsible for ensuring that the cpumask of @node stays 38774c16bd32STejun Heo * stable. 3878d185af30SYacine Belkadi * 3879d185af30SYacine Belkadi * Return: %true if the resulting @cpumask is different from @attrs->cpumask, 3880d185af30SYacine Belkadi * %false if equal. 38814c16bd32STejun Heo */ 38824c16bd32STejun Heo static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node, 38834c16bd32STejun Heo int cpu_going_down, cpumask_t *cpumask) 38844c16bd32STejun Heo { 3885d55262c4STejun Heo if (!wq_numa_enabled || attrs->no_numa) 38864c16bd32STejun Heo goto use_dfl; 38874c16bd32STejun Heo 38884c16bd32STejun Heo /* does @node have any online CPUs @attrs wants? */ 38894c16bd32STejun Heo cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask); 38904c16bd32STejun Heo if (cpu_going_down >= 0) 38914c16bd32STejun Heo cpumask_clear_cpu(cpu_going_down, cpumask); 38924c16bd32STejun Heo 38934c16bd32STejun Heo if (cpumask_empty(cpumask)) 38944c16bd32STejun Heo goto use_dfl; 38954c16bd32STejun Heo 38964c16bd32STejun Heo /* yeap, return possible CPUs in @node that @attrs wants */ 38974c16bd32STejun Heo cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]); 38981ad0f0a7SMichael Bringmann 38991ad0f0a7SMichael Bringmann if (cpumask_empty(cpumask)) { 39001ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 39011ad0f0a7SMichael Bringmann "possible intersect\n"); 39021ad0f0a7SMichael Bringmann return false; 39031ad0f0a7SMichael Bringmann } 39041ad0f0a7SMichael Bringmann 39054c16bd32STejun Heo return !cpumask_equal(cpumask, attrs->cpumask); 39064c16bd32STejun Heo 39074c16bd32STejun Heo use_dfl: 39084c16bd32STejun Heo cpumask_copy(cpumask, attrs->cpumask); 39094c16bd32STejun Heo return false; 39104c16bd32STejun Heo } 39114c16bd32STejun Heo 39121befcf30STejun Heo /* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */ 39131befcf30STejun Heo static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq, 39141befcf30STejun Heo int node, 39151befcf30STejun Heo struct pool_workqueue *pwq) 39161befcf30STejun Heo { 39171befcf30STejun Heo struct pool_workqueue *old_pwq; 39181befcf30STejun Heo 39195b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39201befcf30STejun Heo lockdep_assert_held(&wq->mutex); 39211befcf30STejun Heo 39221befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 39231befcf30STejun Heo link_pwq(pwq); 39241befcf30STejun Heo 39251befcf30STejun Heo old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 39261befcf30STejun Heo rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq); 39271befcf30STejun Heo return old_pwq; 39281befcf30STejun Heo } 39291befcf30STejun Heo 39302d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 39312d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 39322d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 39332d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 3934042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 39352d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 39362d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 39372d5f0764SLai Jiangshan }; 39382d5f0764SLai Jiangshan 39392d5f0764SLai Jiangshan /* free the resources after success or abort */ 39402d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 39412d5f0764SLai Jiangshan { 39422d5f0764SLai Jiangshan if (ctx) { 39432d5f0764SLai Jiangshan int node; 39442d5f0764SLai Jiangshan 39452d5f0764SLai Jiangshan for_each_node(node) 39462d5f0764SLai Jiangshan put_pwq_unlocked(ctx->pwq_tbl[node]); 39472d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 39482d5f0764SLai Jiangshan 39492d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 39502d5f0764SLai Jiangshan 39512d5f0764SLai Jiangshan kfree(ctx); 39522d5f0764SLai Jiangshan } 39532d5f0764SLai Jiangshan } 39542d5f0764SLai Jiangshan 39552d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 39562d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 39572d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 39582d5f0764SLai Jiangshan const struct workqueue_attrs *attrs) 39592d5f0764SLai Jiangshan { 39602d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 39612d5f0764SLai Jiangshan struct workqueue_attrs *new_attrs, *tmp_attrs; 39622d5f0764SLai Jiangshan int node; 39632d5f0764SLai Jiangshan 39642d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 39652d5f0764SLai Jiangshan 3966acafe7e3SKees Cook ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL); 39672d5f0764SLai Jiangshan 3968be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 3969be69d00dSThomas Gleixner tmp_attrs = alloc_workqueue_attrs(); 39702d5f0764SLai Jiangshan if (!ctx || !new_attrs || !tmp_attrs) 39712d5f0764SLai Jiangshan goto out_free; 39722d5f0764SLai Jiangshan 3973042f7df1SLai Jiangshan /* 3974042f7df1SLai Jiangshan * Calculate the attrs of the default pwq. 3975042f7df1SLai Jiangshan * If the user configured cpumask doesn't overlap with the 3976042f7df1SLai Jiangshan * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask. 3977042f7df1SLai Jiangshan */ 39782d5f0764SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 3979b05a7928SFrederic Weisbecker cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask); 3980042f7df1SLai Jiangshan if (unlikely(cpumask_empty(new_attrs->cpumask))) 3981042f7df1SLai Jiangshan cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask); 39822d5f0764SLai Jiangshan 39832d5f0764SLai Jiangshan /* 39842d5f0764SLai Jiangshan * We may create multiple pwqs with differing cpumasks. Make a 39852d5f0764SLai Jiangshan * copy of @new_attrs which will be modified and used to obtain 39862d5f0764SLai Jiangshan * pools. 39872d5f0764SLai Jiangshan */ 39882d5f0764SLai Jiangshan copy_workqueue_attrs(tmp_attrs, new_attrs); 39892d5f0764SLai Jiangshan 39902d5f0764SLai Jiangshan /* 39912d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 39922d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 39932d5f0764SLai Jiangshan * it even if we don't use it immediately. 39942d5f0764SLai Jiangshan */ 39952d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 39962d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 39972d5f0764SLai Jiangshan goto out_free; 39982d5f0764SLai Jiangshan 39992d5f0764SLai Jiangshan for_each_node(node) { 4000042f7df1SLai Jiangshan if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) { 40012d5f0764SLai Jiangshan ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs); 40022d5f0764SLai Jiangshan if (!ctx->pwq_tbl[node]) 40032d5f0764SLai Jiangshan goto out_free; 40042d5f0764SLai Jiangshan } else { 40052d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 40062d5f0764SLai Jiangshan ctx->pwq_tbl[node] = ctx->dfl_pwq; 40072d5f0764SLai Jiangshan } 40082d5f0764SLai Jiangshan } 40092d5f0764SLai Jiangshan 4010042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4011042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4012042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 40132d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4014042f7df1SLai Jiangshan 40152d5f0764SLai Jiangshan ctx->wq = wq; 40162d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40172d5f0764SLai Jiangshan return ctx; 40182d5f0764SLai Jiangshan 40192d5f0764SLai Jiangshan out_free: 40202d5f0764SLai Jiangshan free_workqueue_attrs(tmp_attrs); 40212d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 40222d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 40232d5f0764SLai Jiangshan return NULL; 40242d5f0764SLai Jiangshan } 40252d5f0764SLai Jiangshan 40262d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 40272d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 40282d5f0764SLai Jiangshan { 40292d5f0764SLai Jiangshan int node; 40302d5f0764SLai Jiangshan 40312d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 40322d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 40332d5f0764SLai Jiangshan 40342d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 40352d5f0764SLai Jiangshan 40362d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 40372d5f0764SLai Jiangshan for_each_node(node) 40382d5f0764SLai Jiangshan ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node, 40392d5f0764SLai Jiangshan ctx->pwq_tbl[node]); 40402d5f0764SLai Jiangshan 40412d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 40422d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 40432d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 40442d5f0764SLai Jiangshan 40452d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 40462d5f0764SLai Jiangshan } 40472d5f0764SLai Jiangshan 4048a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4049a0111cf6SLai Jiangshan { 4050a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4051ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4052a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4053a0111cf6SLai Jiangshan } 4054a0111cf6SLai Jiangshan 4055a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4056a0111cf6SLai Jiangshan { 4057a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4058ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4059a0111cf6SLai Jiangshan } 4060a0111cf6SLai Jiangshan 4061a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4062a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4063a0111cf6SLai Jiangshan { 4064a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4065a0111cf6SLai Jiangshan 4066a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4067a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4068a0111cf6SLai Jiangshan return -EINVAL; 4069a0111cf6SLai Jiangshan 4070a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 40710a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 40720a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4073a0111cf6SLai Jiangshan return -EINVAL; 4074a0111cf6SLai Jiangshan 40750a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 40760a94efb5STejun Heo } 40770a94efb5STejun Heo 4078a0111cf6SLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs); 40796201171eSwanghaibin if (!ctx) 40806201171eSwanghaibin return -ENOMEM; 4081a0111cf6SLai Jiangshan 4082a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4083a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4084a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4085a0111cf6SLai Jiangshan 40866201171eSwanghaibin return 0; 4087a0111cf6SLai Jiangshan } 4088a0111cf6SLai Jiangshan 40899e8cd2f5STejun Heo /** 40909e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 40919e8cd2f5STejun Heo * @wq: the target workqueue 40929e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 40939e8cd2f5STejun Heo * 40944c16bd32STejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA 40954c16bd32STejun Heo * machines, this function maps a separate pwq to each NUMA node with 40964c16bd32STejun Heo * possibles CPUs in @attrs->cpumask so that work items are affine to the 40974c16bd32STejun Heo * NUMA node it was issued on. Older pwqs are released as in-flight work 40984c16bd32STejun Heo * items finish. Note that a work item which repeatedly requeues itself 40994c16bd32STejun Heo * back-to-back will stay on its current pwq. 41009e8cd2f5STejun Heo * 4101d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4102d185af30SYacine Belkadi * 4103ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4104509b3204SDaniel Jordan * 4105d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 41069e8cd2f5STejun Heo */ 4107513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 41089e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 41099e8cd2f5STejun Heo { 4110a0111cf6SLai Jiangshan int ret; 41119e8cd2f5STejun Heo 4112509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4113509b3204SDaniel Jordan 4114509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4115a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4116509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 41172d5f0764SLai Jiangshan 41182d5f0764SLai Jiangshan return ret; 41199e8cd2f5STejun Heo } 41209e8cd2f5STejun Heo 41214c16bd32STejun Heo /** 41224c16bd32STejun Heo * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug 41234c16bd32STejun Heo * @wq: the target workqueue 41244c16bd32STejun Heo * @cpu: the CPU coming up or going down 41254c16bd32STejun Heo * @online: whether @cpu is coming up or going down 41264c16bd32STejun Heo * 41274c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 41284c16bd32STejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of 41294c16bd32STejun Heo * @wq accordingly. 41304c16bd32STejun Heo * 41314c16bd32STejun Heo * If NUMA affinity can't be adjusted due to memory allocation failure, it 41324c16bd32STejun Heo * falls back to @wq->dfl_pwq which may not be optimal but is always 41334c16bd32STejun Heo * correct. 41344c16bd32STejun Heo * 41354c16bd32STejun Heo * Note that when the last allowed CPU of a NUMA node goes offline for a 41364c16bd32STejun Heo * workqueue with a cpumask spanning multiple nodes, the workers which were 41374c16bd32STejun Heo * already executing the work items for the workqueue will lose their CPU 41384c16bd32STejun Heo * affinity and may execute on any CPU. This is similar to how per-cpu 41394c16bd32STejun Heo * workqueues behave on CPU_DOWN. If a workqueue user wants strict 41404c16bd32STejun Heo * affinity, it's the user's responsibility to flush the work item from 41414c16bd32STejun Heo * CPU_DOWN_PREPARE. 41424c16bd32STejun Heo */ 41434c16bd32STejun Heo static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu, 41444c16bd32STejun Heo bool online) 41454c16bd32STejun Heo { 41464c16bd32STejun Heo int node = cpu_to_node(cpu); 41474c16bd32STejun Heo int cpu_off = online ? -1 : cpu; 41484c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 41494c16bd32STejun Heo struct workqueue_attrs *target_attrs; 41504c16bd32STejun Heo cpumask_t *cpumask; 41514c16bd32STejun Heo 41524c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 41534c16bd32STejun Heo 4154f7142ed4SLai Jiangshan if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) || 4155f7142ed4SLai Jiangshan wq->unbound_attrs->no_numa) 41564c16bd32STejun Heo return; 41574c16bd32STejun Heo 41584c16bd32STejun Heo /* 41594c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 41604c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 41614c16bd32STejun Heo * CPU hotplug exclusion. 41624c16bd32STejun Heo */ 41634c16bd32STejun Heo target_attrs = wq_update_unbound_numa_attrs_buf; 41644c16bd32STejun Heo cpumask = target_attrs->cpumask; 41654c16bd32STejun Heo 41664c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 41674c16bd32STejun Heo pwq = unbound_pwq_by_node(wq, node); 41684c16bd32STejun Heo 41694c16bd32STejun Heo /* 41704c16bd32STejun Heo * Let's determine what needs to be done. If the target cpumask is 4171042f7df1SLai Jiangshan * different from the default pwq's, we need to compare it to @pwq's 4172042f7df1SLai Jiangshan * and create a new one if they don't match. If the target cpumask 4173042f7df1SLai Jiangshan * equals the default pwq's, the default pwq should be used. 41744c16bd32STejun Heo */ 4175042f7df1SLai Jiangshan if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) { 41764c16bd32STejun Heo if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask)) 4177f7142ed4SLai Jiangshan return; 41784c16bd32STejun Heo } else { 41794c16bd32STejun Heo goto use_dfl_pwq; 41804c16bd32STejun Heo } 41814c16bd32STejun Heo 41824c16bd32STejun Heo /* create a new pwq */ 41834c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 41844c16bd32STejun Heo if (!pwq) { 41852d916033SFabian Frederick pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n", 41864c16bd32STejun Heo wq->name); 418777f300b1SDaeseok Youn goto use_dfl_pwq; 41884c16bd32STejun Heo } 41894c16bd32STejun Heo 4190f7142ed4SLai Jiangshan /* Install the new pwq. */ 41914c16bd32STejun Heo mutex_lock(&wq->mutex); 41924c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, pwq); 41934c16bd32STejun Heo goto out_unlock; 41944c16bd32STejun Heo 41954c16bd32STejun Heo use_dfl_pwq: 4196f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4197a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 41984c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4199a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 42004c16bd32STejun Heo old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq); 42014c16bd32STejun Heo out_unlock: 42024c16bd32STejun Heo mutex_unlock(&wq->mutex); 42034c16bd32STejun Heo put_pwq_unlocked(old_pwq); 42044c16bd32STejun Heo } 42054c16bd32STejun Heo 420630cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 42071da177e4SLinus Torvalds { 420849e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 42098a2b7538STejun Heo int cpu, ret; 4210e1d8aa9fSFrederic Weisbecker 421130cdf249STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4212420c0ddbSTejun Heo wq->cpu_pwqs = alloc_percpu(struct pool_workqueue); 4213420c0ddbSTejun Heo if (!wq->cpu_pwqs) 421430cdf249STejun Heo return -ENOMEM; 421530cdf249STejun Heo 421630cdf249STejun Heo for_each_possible_cpu(cpu) { 42177fb98ea7STejun Heo struct pool_workqueue *pwq = 42187fb98ea7STejun Heo per_cpu_ptr(wq->cpu_pwqs, cpu); 42197a62c2c8STejun Heo struct worker_pool *cpu_pools = 4220f02ae73aSTejun Heo per_cpu(cpu_worker_pools, cpu); 422130cdf249STejun Heo 4222f147f29eSTejun Heo init_pwq(pwq, wq, &cpu_pools[highpri]); 4223f147f29eSTejun Heo 4224f147f29eSTejun Heo mutex_lock(&wq->mutex); 42251befcf30STejun Heo link_pwq(pwq); 4226f147f29eSTejun Heo mutex_unlock(&wq->mutex); 422730cdf249STejun Heo } 422830cdf249STejun Heo return 0; 4229509b3204SDaniel Jordan } 4230509b3204SDaniel Jordan 4231ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4232509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 42338a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 42348a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 42358a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 42368a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 42378a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 42389e8cd2f5STejun Heo } else { 4239509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 42409e8cd2f5STejun Heo } 4241ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4242509b3204SDaniel Jordan 4243509b3204SDaniel Jordan return ret; 42440f900049STejun Heo } 42450f900049STejun Heo 4246f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4247f3421797STejun Heo const char *name) 4248b71ab8c2STejun Heo { 4249f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 4250f3421797STejun Heo 4251f3421797STejun Heo if (max_active < 1 || max_active > lim) 4252044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4253f3421797STejun Heo max_active, name, 1, lim); 4254b71ab8c2STejun Heo 4255f3421797STejun Heo return clamp_val(max_active, 1, lim); 4256b71ab8c2STejun Heo } 4257b71ab8c2STejun Heo 4258983c7515STejun Heo /* 4259983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4260983c7515STejun Heo * to guarantee forward progress. 4261983c7515STejun Heo */ 4262983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4263983c7515STejun Heo { 4264983c7515STejun Heo struct worker *rescuer; 4265b92b36eaSDan Carpenter int ret; 4266983c7515STejun Heo 4267983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4268983c7515STejun Heo return 0; 4269983c7515STejun Heo 4270983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 4271983c7515STejun Heo if (!rescuer) 4272983c7515STejun Heo return -ENOMEM; 4273983c7515STejun Heo 4274983c7515STejun Heo rescuer->rescue_wq = wq; 4275983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4276f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4277b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 4278983c7515STejun Heo kfree(rescuer); 4279b92b36eaSDan Carpenter return ret; 4280983c7515STejun Heo } 4281983c7515STejun Heo 4282983c7515STejun Heo wq->rescuer = rescuer; 4283983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4284983c7515STejun Heo wake_up_process(rescuer->task); 4285983c7515STejun Heo 4286983c7515STejun Heo return 0; 4287983c7515STejun Heo } 4288983c7515STejun Heo 4289a2775bbcSMathieu Malaterre __printf(1, 4) 4290669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 429197e37d7bSTejun Heo unsigned int flags, 4292669de8bdSBart Van Assche int max_active, ...) 42933af24433SOleg Nesterov { 4294df2d5ae4STejun Heo size_t tbl_size = 0; 4295ecf6881fSTejun Heo va_list args; 42963af24433SOleg Nesterov struct workqueue_struct *wq; 429749e3cf44STejun Heo struct pool_workqueue *pwq; 4298b196be89STejun Heo 42995c0338c6STejun Heo /* 43005c0338c6STejun Heo * Unbound && max_active == 1 used to imply ordered, which is no 43015c0338c6STejun Heo * longer the case on NUMA machines due to per-node pools. While 43025c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 43035c0338c6STejun Heo * workqueue, keep the previous behavior to avoid subtle breakages 43045c0338c6STejun Heo * on NUMA. 43055c0338c6STejun Heo */ 43065c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 43075c0338c6STejun Heo flags |= __WQ_ORDERED; 43085c0338c6STejun Heo 4309cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4310cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4311cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4312cee22a15SViresh Kumar 4313ecf6881fSTejun Heo /* allocate wq and format name */ 4314df2d5ae4STejun Heo if (flags & WQ_UNBOUND) 4315ddcb57e2SLai Jiangshan tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]); 4316df2d5ae4STejun Heo 4317df2d5ae4STejun Heo wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL); 4318b196be89STejun Heo if (!wq) 4319d2c1d404STejun Heo return NULL; 4320b196be89STejun Heo 43216029a918STejun Heo if (flags & WQ_UNBOUND) { 4322be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 43236029a918STejun Heo if (!wq->unbound_attrs) 43246029a918STejun Heo goto err_free_wq; 43256029a918STejun Heo } 43266029a918STejun Heo 4327669de8bdSBart Van Assche va_start(args, max_active); 4328ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4329b196be89STejun Heo va_end(args); 43303af24433SOleg Nesterov 4331d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4332b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 43333af24433SOleg Nesterov 4334b196be89STejun Heo /* init wq */ 433597e37d7bSTejun Heo wq->flags = flags; 4336a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 43373c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4338112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 433930cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 434073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 434173f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4342493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 43433af24433SOleg Nesterov 4344669de8bdSBart Van Assche wq_init_lockdep(wq); 4345cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 43463af24433SOleg Nesterov 434730cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 434882efcab3SBart Van Assche goto err_unreg_lockdep; 43491537663fSTejun Heo 435040c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4351d2c1d404STejun Heo goto err_destroy; 4352e22bee78STejun Heo 4353226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4354226223abSTejun Heo goto err_destroy; 4355226223abSTejun Heo 43566af8bf3dSOleg Nesterov /* 435768e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 435868e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 435968e13a67SLai Jiangshan * list. 43606af8bf3dSOleg Nesterov */ 436168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4362a0a1a5fdSTejun Heo 4363a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 436449e3cf44STejun Heo for_each_pwq(pwq, wq) 4365699ce097STejun Heo pwq_adjust_max_active(pwq); 4366a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4367a0a1a5fdSTejun Heo 4368e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4369a0a1a5fdSTejun Heo 437068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 43713af24433SOleg Nesterov 43723af24433SOleg Nesterov return wq; 4373d2c1d404STejun Heo 437482efcab3SBart Van Assche err_unreg_lockdep: 4375009bb421SBart Van Assche wq_unregister_lockdep(wq); 4376009bb421SBart Van Assche wq_free_lockdep(wq); 437782efcab3SBart Van Assche err_free_wq: 43786029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 43794690c4abSTejun Heo kfree(wq); 4380d2c1d404STejun Heo return NULL; 4381d2c1d404STejun Heo err_destroy: 4382d2c1d404STejun Heo destroy_workqueue(wq); 43834690c4abSTejun Heo return NULL; 43841da177e4SLinus Torvalds } 4385669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 43861da177e4SLinus Torvalds 4387c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4388c29eb853STejun Heo { 4389c29eb853STejun Heo int i; 4390c29eb853STejun Heo 4391c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4392c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4393c29eb853STejun Heo return true; 4394c29eb853STejun Heo 4395c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4396c29eb853STejun Heo return true; 4397f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4398c29eb853STejun Heo return true; 4399c29eb853STejun Heo 4400c29eb853STejun Heo return false; 4401c29eb853STejun Heo } 4402c29eb853STejun Heo 44033af24433SOleg Nesterov /** 44043af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 44053af24433SOleg Nesterov * @wq: target workqueue 44063af24433SOleg Nesterov * 44073af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 44083af24433SOleg Nesterov */ 44093af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 44103af24433SOleg Nesterov { 441149e3cf44STejun Heo struct pool_workqueue *pwq; 44124c16bd32STejun Heo int node; 44133af24433SOleg Nesterov 4414def98c84STejun Heo /* 4415def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4416def98c84STejun Heo * lead to sysfs name conflicts. 4417def98c84STejun Heo */ 4418def98c84STejun Heo workqueue_sysfs_unregister(wq); 4419def98c84STejun Heo 44209c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 44219c5a2ba7STejun Heo drain_workqueue(wq); 4422c8efcc25STejun Heo 4423def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4424def98c84STejun Heo if (wq->rescuer) { 4425def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4426def98c84STejun Heo 4427def98c84STejun Heo /* this prevents new queueing */ 4428a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4429def98c84STejun Heo wq->rescuer = NULL; 4430a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4431def98c84STejun Heo 4432def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4433def98c84STejun Heo kthread_stop(rescuer->task); 44348efe1223STejun Heo kfree(rescuer); 4435def98c84STejun Heo } 4436def98c84STejun Heo 4437c29eb853STejun Heo /* 4438c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4439c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4440c29eb853STejun Heo */ 4441c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4442b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 444349e3cf44STejun Heo for_each_pwq(pwq, wq) { 4444a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4445c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 44461d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4447e66b39afSTejun Heo __func__, wq->name); 4448c29eb853STejun Heo show_pwq(pwq); 4449a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4450b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4451c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 445255df0933SImran Khan show_one_workqueue(wq); 44536183c009STejun Heo return; 445476af4d93STejun Heo } 4455a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 445676af4d93STejun Heo } 4457b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 44586183c009STejun Heo 4459a0a1a5fdSTejun Heo /* 4460a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4461a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4462a0a1a5fdSTejun Heo */ 4463e2dca7adSTejun Heo list_del_rcu(&wq->list); 446468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 44653af24433SOleg Nesterov 44668864b4e5STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 4467669de8bdSBart Van Assche wq_unregister_lockdep(wq); 446829c91e99STejun Heo /* 44698864b4e5STejun Heo * The base ref is never dropped on per-cpu pwqs. Directly 4470e2dca7adSTejun Heo * schedule RCU free. 447129c91e99STejun Heo */ 447225b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 44738864b4e5STejun Heo } else { 44748864b4e5STejun Heo /* 44758864b4e5STejun Heo * We're the sole accessor of @wq at this point. Directly 44764c16bd32STejun Heo * access numa_pwq_tbl[] and dfl_pwq to put the base refs. 44774c16bd32STejun Heo * @wq will be freed when the last pwq is released. 44788864b4e5STejun Heo */ 44794c16bd32STejun Heo for_each_node(node) { 44804c16bd32STejun Heo pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]); 44814c16bd32STejun Heo RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL); 44824c16bd32STejun Heo put_pwq_unlocked(pwq); 44834c16bd32STejun Heo } 44844c16bd32STejun Heo 44854c16bd32STejun Heo /* 44864c16bd32STejun Heo * Put dfl_pwq. @wq may be freed any time after dfl_pwq is 44874c16bd32STejun Heo * put. Don't access it afterwards. 44884c16bd32STejun Heo */ 44894c16bd32STejun Heo pwq = wq->dfl_pwq; 44904c16bd32STejun Heo wq->dfl_pwq = NULL; 4491dce90d47STejun Heo put_pwq_unlocked(pwq); 449229c91e99STejun Heo } 44933af24433SOleg Nesterov } 44943af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 44953af24433SOleg Nesterov 4496dcd989cbSTejun Heo /** 4497dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4498dcd989cbSTejun Heo * @wq: target workqueue 4499dcd989cbSTejun Heo * @max_active: new max_active value. 4500dcd989cbSTejun Heo * 4501dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4502dcd989cbSTejun Heo * 4503dcd989cbSTejun Heo * CONTEXT: 4504dcd989cbSTejun Heo * Don't call from IRQ context. 4505dcd989cbSTejun Heo */ 4506dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4507dcd989cbSTejun Heo { 450849e3cf44STejun Heo struct pool_workqueue *pwq; 4509dcd989cbSTejun Heo 45108719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 45110a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 45128719dceaSTejun Heo return; 45138719dceaSTejun Heo 4514f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4515dcd989cbSTejun Heo 4516a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4517dcd989cbSTejun Heo 45180a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4519dcd989cbSTejun Heo wq->saved_max_active = max_active; 4520dcd989cbSTejun Heo 4521699ce097STejun Heo for_each_pwq(pwq, wq) 4522699ce097STejun Heo pwq_adjust_max_active(pwq); 4523dcd989cbSTejun Heo 4524a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4525dcd989cbSTejun Heo } 4526dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4527dcd989cbSTejun Heo 4528dcd989cbSTejun Heo /** 452927d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 453027d4ee03SLukas Wunner * 453127d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 453227d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 453327d4ee03SLukas Wunner * 453427d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 453527d4ee03SLukas Wunner */ 453627d4ee03SLukas Wunner struct work_struct *current_work(void) 453727d4ee03SLukas Wunner { 453827d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 453927d4ee03SLukas Wunner 454027d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 454127d4ee03SLukas Wunner } 454227d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 454327d4ee03SLukas Wunner 454427d4ee03SLukas Wunner /** 4545e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4546e6267616STejun Heo * 4547e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4548e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4549d185af30SYacine Belkadi * 4550d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4551e6267616STejun Heo */ 4552e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4553e6267616STejun Heo { 4554e6267616STejun Heo struct worker *worker = current_wq_worker(); 4555e6267616STejun Heo 45566a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4557e6267616STejun Heo } 4558e6267616STejun Heo 4559e6267616STejun Heo /** 4560dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4561dcd989cbSTejun Heo * @cpu: CPU in question 4562dcd989cbSTejun Heo * @wq: target workqueue 4563dcd989cbSTejun Heo * 4564dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4565dcd989cbSTejun Heo * no synchronization around this function and the test result is 4566dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4567dcd989cbSTejun Heo * 4568d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4569d3251859STejun Heo * Note that both per-cpu and unbound workqueues may be associated with 4570d3251859STejun Heo * multiple pool_workqueues which have separate congested states. A 4571d3251859STejun Heo * workqueue being congested on one CPU doesn't mean the workqueue is also 4572d3251859STejun Heo * contested on other CPUs / NUMA nodes. 4573d3251859STejun Heo * 4574d185af30SYacine Belkadi * Return: 4575dcd989cbSTejun Heo * %true if congested, %false otherwise. 4576dcd989cbSTejun Heo */ 4577d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4578dcd989cbSTejun Heo { 45797fb98ea7STejun Heo struct pool_workqueue *pwq; 458076af4d93STejun Heo bool ret; 458176af4d93STejun Heo 458224acfb71SThomas Gleixner rcu_read_lock(); 458324acfb71SThomas Gleixner preempt_disable(); 45847fb98ea7STejun Heo 4585d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4586d3251859STejun Heo cpu = smp_processor_id(); 4587d3251859STejun Heo 45887fb98ea7STejun Heo if (!(wq->flags & WQ_UNBOUND)) 45897fb98ea7STejun Heo pwq = per_cpu_ptr(wq->cpu_pwqs, cpu); 45907fb98ea7STejun Heo else 4591df2d5ae4STejun Heo pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu)); 4592dcd989cbSTejun Heo 4593f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 459424acfb71SThomas Gleixner preempt_enable(); 459524acfb71SThomas Gleixner rcu_read_unlock(); 459676af4d93STejun Heo 459776af4d93STejun Heo return ret; 4598dcd989cbSTejun Heo } 4599dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4600dcd989cbSTejun Heo 4601dcd989cbSTejun Heo /** 4602dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4603dcd989cbSTejun Heo * @work: the work to be tested 4604dcd989cbSTejun Heo * 4605dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4606dcd989cbSTejun Heo * synchronization around this function and the test result is 4607dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4608dcd989cbSTejun Heo * 4609d185af30SYacine Belkadi * Return: 4610dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4611dcd989cbSTejun Heo */ 4612dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4613dcd989cbSTejun Heo { 4614fa1b54e6STejun Heo struct worker_pool *pool; 4615dcd989cbSTejun Heo unsigned long flags; 4616dcd989cbSTejun Heo unsigned int ret = 0; 4617dcd989cbSTejun Heo 4618dcd989cbSTejun Heo if (work_pending(work)) 4619dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4620038366c5SLai Jiangshan 462124acfb71SThomas Gleixner rcu_read_lock(); 4622fa1b54e6STejun Heo pool = get_work_pool(work); 4623038366c5SLai Jiangshan if (pool) { 4624a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4625c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4626dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4627a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4628038366c5SLai Jiangshan } 462924acfb71SThomas Gleixner rcu_read_unlock(); 4630dcd989cbSTejun Heo 4631dcd989cbSTejun Heo return ret; 4632dcd989cbSTejun Heo } 4633dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4634dcd989cbSTejun Heo 46353d1cb205STejun Heo /** 46363d1cb205STejun Heo * set_worker_desc - set description for the current work item 46373d1cb205STejun Heo * @fmt: printf-style format string 46383d1cb205STejun Heo * @...: arguments for the format string 46393d1cb205STejun Heo * 46403d1cb205STejun Heo * This function can be called by a running work function to describe what 46413d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 46423d1cb205STejun Heo * information will be printed out together to help debugging. The 46433d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 46443d1cb205STejun Heo */ 46453d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 46463d1cb205STejun Heo { 46473d1cb205STejun Heo struct worker *worker = current_wq_worker(); 46483d1cb205STejun Heo va_list args; 46493d1cb205STejun Heo 46503d1cb205STejun Heo if (worker) { 46513d1cb205STejun Heo va_start(args, fmt); 46523d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 46533d1cb205STejun Heo va_end(args); 46543d1cb205STejun Heo } 46553d1cb205STejun Heo } 46565c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 46573d1cb205STejun Heo 46583d1cb205STejun Heo /** 46593d1cb205STejun Heo * print_worker_info - print out worker information and description 46603d1cb205STejun Heo * @log_lvl: the log level to use when printing 46613d1cb205STejun Heo * @task: target task 46623d1cb205STejun Heo * 46633d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 46643d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 46653d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 46663d1cb205STejun Heo * 46673d1cb205STejun Heo * This function can be safely called on any task as long as the 46683d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 46693d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 46703d1cb205STejun Heo */ 46713d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 46723d1cb205STejun Heo { 46733d1cb205STejun Heo work_func_t *fn = NULL; 46743d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 46753d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 46763d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 46773d1cb205STejun Heo struct workqueue_struct *wq = NULL; 46783d1cb205STejun Heo struct worker *worker; 46793d1cb205STejun Heo 46803d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 46813d1cb205STejun Heo return; 46823d1cb205STejun Heo 46833d1cb205STejun Heo /* 46843d1cb205STejun Heo * This function is called without any synchronization and @task 46853d1cb205STejun Heo * could be in any state. Be careful with dereferences. 46863d1cb205STejun Heo */ 4687e700591aSPetr Mladek worker = kthread_probe_data(task); 46883d1cb205STejun Heo 46893d1cb205STejun Heo /* 46908bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 46918bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 46923d1cb205STejun Heo */ 4693fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 4694fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 4695fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 4696fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 4697fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 46983d1cb205STejun Heo 46993d1cb205STejun Heo if (fn || name[0] || desc[0]) { 4700d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 47018bf89593STejun Heo if (strcmp(name, desc)) 47023d1cb205STejun Heo pr_cont(" (%s)", desc); 47033d1cb205STejun Heo pr_cont("\n"); 47043d1cb205STejun Heo } 47053d1cb205STejun Heo } 47063d1cb205STejun Heo 47073494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 47083494fc30STejun Heo { 47093494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 47103494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 47113494fc30STejun Heo pr_cont(" node=%d", pool->node); 47123494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 47133494fc30STejun Heo } 47143494fc30STejun Heo 47153494fc30STejun Heo static void pr_cont_work(bool comma, struct work_struct *work) 47163494fc30STejun Heo { 47173494fc30STejun Heo if (work->func == wq_barrier_func) { 47183494fc30STejun Heo struct wq_barrier *barr; 47193494fc30STejun Heo 47203494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 47213494fc30STejun Heo 47223494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 47233494fc30STejun Heo task_pid_nr(barr->task)); 47243494fc30STejun Heo } else { 4725d75f773cSSakari Ailus pr_cont("%s %ps", comma ? "," : "", work->func); 47263494fc30STejun Heo } 47273494fc30STejun Heo } 47283494fc30STejun Heo 47293494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 47303494fc30STejun Heo { 47313494fc30STejun Heo struct worker_pool *pool = pwq->pool; 47323494fc30STejun Heo struct work_struct *work; 47333494fc30STejun Heo struct worker *worker; 47343494fc30STejun Heo bool has_in_flight = false, has_pending = false; 47353494fc30STejun Heo int bkt; 47363494fc30STejun Heo 47373494fc30STejun Heo pr_info(" pwq %d:", pool->id); 47383494fc30STejun Heo pr_cont_pool_info(pool); 47393494fc30STejun Heo 4740e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 4741e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 47423494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 47433494fc30STejun Heo 47443494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47453494fc30STejun Heo if (worker->current_pwq == pwq) { 47463494fc30STejun Heo has_in_flight = true; 47473494fc30STejun Heo break; 47483494fc30STejun Heo } 47493494fc30STejun Heo } 47503494fc30STejun Heo if (has_in_flight) { 47513494fc30STejun Heo bool comma = false; 47523494fc30STejun Heo 47533494fc30STejun Heo pr_info(" in-flight:"); 47543494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 47553494fc30STejun Heo if (worker->current_pwq != pwq) 47563494fc30STejun Heo continue; 47573494fc30STejun Heo 4758d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 47593494fc30STejun Heo task_pid_nr(worker->task), 476030ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 47613494fc30STejun Heo worker->current_func); 47623494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 47633494fc30STejun Heo pr_cont_work(false, work); 47643494fc30STejun Heo comma = true; 47653494fc30STejun Heo } 47663494fc30STejun Heo pr_cont("\n"); 47673494fc30STejun Heo } 47683494fc30STejun Heo 47693494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47703494fc30STejun Heo if (get_work_pwq(work) == pwq) { 47713494fc30STejun Heo has_pending = true; 47723494fc30STejun Heo break; 47733494fc30STejun Heo } 47743494fc30STejun Heo } 47753494fc30STejun Heo if (has_pending) { 47763494fc30STejun Heo bool comma = false; 47773494fc30STejun Heo 47783494fc30STejun Heo pr_info(" pending:"); 47793494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 47803494fc30STejun Heo if (get_work_pwq(work) != pwq) 47813494fc30STejun Heo continue; 47823494fc30STejun Heo 47833494fc30STejun Heo pr_cont_work(comma, work); 47843494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47853494fc30STejun Heo } 47863494fc30STejun Heo pr_cont("\n"); 47873494fc30STejun Heo } 47883494fc30STejun Heo 4789f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 47903494fc30STejun Heo bool comma = false; 47913494fc30STejun Heo 4792f97a4a1aSLai Jiangshan pr_info(" inactive:"); 4793f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 47943494fc30STejun Heo pr_cont_work(comma, work); 47953494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 47963494fc30STejun Heo } 47973494fc30STejun Heo pr_cont("\n"); 47983494fc30STejun Heo } 47993494fc30STejun Heo } 48003494fc30STejun Heo 48013494fc30STejun Heo /** 480255df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 480355df0933SImran Khan * @wq: workqueue whose state will be printed 48043494fc30STejun Heo */ 480555df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 48063494fc30STejun Heo { 48073494fc30STejun Heo struct pool_workqueue *pwq; 48083494fc30STejun Heo bool idle = true; 480955df0933SImran Khan unsigned long flags; 48103494fc30STejun Heo 48113494fc30STejun Heo for_each_pwq(pwq, wq) { 4812f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 48133494fc30STejun Heo idle = false; 48143494fc30STejun Heo break; 48153494fc30STejun Heo } 48163494fc30STejun Heo } 481755df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 481855df0933SImran Khan return; 48193494fc30STejun Heo 48203494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 48213494fc30STejun Heo 48223494fc30STejun Heo for_each_pwq(pwq, wq) { 4823a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 482457116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 482557116ce1SJohan Hovold /* 482657116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 482757116ce1SJohan Hovold * drivers that queue work while holding locks 482857116ce1SJohan Hovold * also taken in their write paths. 482957116ce1SJohan Hovold */ 483057116ce1SJohan Hovold printk_deferred_enter(); 48313494fc30STejun Heo show_pwq(pwq); 483257116ce1SJohan Hovold printk_deferred_exit(); 483357116ce1SJohan Hovold } 4834a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 483562635ea8SSergey Senozhatsky /* 483662635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 483755df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 483862635ea8SSergey Senozhatsky * hard lockup. 483962635ea8SSergey Senozhatsky */ 484062635ea8SSergey Senozhatsky touch_nmi_watchdog(); 48413494fc30STejun Heo } 484255df0933SImran Khan 48433494fc30STejun Heo } 48443494fc30STejun Heo 484555df0933SImran Khan /** 484655df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 484755df0933SImran Khan * @pool: worker pool whose state will be printed 484855df0933SImran Khan */ 484955df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 485055df0933SImran Khan { 48513494fc30STejun Heo struct worker *worker; 48523494fc30STejun Heo bool first = true; 485355df0933SImran Khan unsigned long flags; 48543494fc30STejun Heo 4855a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 48563494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 48573494fc30STejun Heo goto next_pool; 485857116ce1SJohan Hovold /* 485957116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 486057116ce1SJohan Hovold * queue work while holding locks also taken in their write 486157116ce1SJohan Hovold * paths. 486257116ce1SJohan Hovold */ 486357116ce1SJohan Hovold printk_deferred_enter(); 48643494fc30STejun Heo pr_info("pool %d:", pool->id); 48653494fc30STejun Heo pr_cont_pool_info(pool); 486682607adcSTejun Heo pr_cont(" hung=%us workers=%d", 486782607adcSTejun Heo jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000, 486882607adcSTejun Heo pool->nr_workers); 48693494fc30STejun Heo if (pool->manager) 48703494fc30STejun Heo pr_cont(" manager: %d", 48713494fc30STejun Heo task_pid_nr(pool->manager->task)); 48723494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 48733494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 48743494fc30STejun Heo task_pid_nr(worker->task)); 48753494fc30STejun Heo first = false; 48763494fc30STejun Heo } 48773494fc30STejun Heo pr_cont("\n"); 487857116ce1SJohan Hovold printk_deferred_exit(); 48793494fc30STejun Heo next_pool: 4880a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 488162635ea8SSergey Senozhatsky /* 488262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 488355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 488462635ea8SSergey Senozhatsky * hard lockup. 488562635ea8SSergey Senozhatsky */ 488662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 488755df0933SImran Khan 48883494fc30STejun Heo } 48893494fc30STejun Heo 489055df0933SImran Khan /** 489155df0933SImran Khan * show_all_workqueues - dump workqueue state 489255df0933SImran Khan * 489355df0933SImran Khan * Called from a sysrq handler or try_to_freeze_tasks() and prints out 489455df0933SImran Khan * all busy workqueues and pools. 489555df0933SImran Khan */ 489655df0933SImran Khan void show_all_workqueues(void) 489755df0933SImran Khan { 489855df0933SImran Khan struct workqueue_struct *wq; 489955df0933SImran Khan struct worker_pool *pool; 490055df0933SImran Khan int pi; 490155df0933SImran Khan 490255df0933SImran Khan rcu_read_lock(); 490355df0933SImran Khan 490455df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 490555df0933SImran Khan 490655df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 490755df0933SImran Khan show_one_workqueue(wq); 490855df0933SImran Khan 490955df0933SImran Khan for_each_pool(pool, pi) 491055df0933SImran Khan show_one_worker_pool(pool); 491155df0933SImran Khan 491224acfb71SThomas Gleixner rcu_read_unlock(); 49133494fc30STejun Heo } 49143494fc30STejun Heo 49156b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 49166b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 49176b59808bSTejun Heo { 49186b59808bSTejun Heo int off; 49196b59808bSTejun Heo 49206b59808bSTejun Heo /* always show the actual comm */ 49216b59808bSTejun Heo off = strscpy(buf, task->comm, size); 49226b59808bSTejun Heo if (off < 0) 49236b59808bSTejun Heo return; 49246b59808bSTejun Heo 4925197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 49266b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 49276b59808bSTejun Heo 4928197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 4929197f6accSTejun Heo struct worker *worker = kthread_data(task); 4930197f6accSTejun Heo struct worker_pool *pool = worker->pool; 49316b59808bSTejun Heo 49326b59808bSTejun Heo if (pool) { 4933a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 49346b59808bSTejun Heo /* 4935197f6accSTejun Heo * ->desc tracks information (wq name or 4936197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 4937197f6accSTejun Heo * current, prepend '+', otherwise '-'. 49386b59808bSTejun Heo */ 49396b59808bSTejun Heo if (worker->desc[0] != '\0') { 49406b59808bSTejun Heo if (worker->current_work) 49416b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 49426b59808bSTejun Heo worker->desc); 49436b59808bSTejun Heo else 49446b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 49456b59808bSTejun Heo worker->desc); 49466b59808bSTejun Heo } 4947a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 49486b59808bSTejun Heo } 4949197f6accSTejun Heo } 49506b59808bSTejun Heo 49516b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 49526b59808bSTejun Heo } 49536b59808bSTejun Heo 495466448bc2SMathieu Malaterre #ifdef CONFIG_SMP 495566448bc2SMathieu Malaterre 4956db7bccf4STejun Heo /* 4957db7bccf4STejun Heo * CPU hotplug. 4958db7bccf4STejun Heo * 4959e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 4960112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 4961706026c2STejun Heo * pool which make migrating pending and scheduled works very 4962e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 496394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 4964e22bee78STejun Heo * blocked draining impractical. 4965e22bee78STejun Heo * 496624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 4967628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 4968628c78e7STejun Heo * cpu comes back online. 4969db7bccf4STejun Heo */ 4970db7bccf4STejun Heo 4971e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 4972db7bccf4STejun Heo { 49734ce62e9eSTejun Heo struct worker_pool *pool; 4974db7bccf4STejun Heo struct worker *worker; 4975db7bccf4STejun Heo 4976f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 49771258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 4978a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 4979e22bee78STejun Heo 4980f2d5a0eeSTejun Heo /* 498192f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 498294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 498394cf58bbSTejun Heo * except for the ones which are still executing works from 498494cf58bbSTejun Heo * before the last CPU down must be on the cpu. After 498594cf58bbSTejun Heo * this, they may become diasporas. 4986f2d5a0eeSTejun Heo */ 4987da028469SLai Jiangshan for_each_pool_worker(worker, pool) 4988403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 4989db7bccf4STejun Heo 499024647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 4991f2d5a0eeSTejun Heo 4992a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 499306249738SPeter Zijlstra 49945c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 49955c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 4996547a77d0SLai Jiangshan WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 49975c25b5ffSPeter Zijlstra } 499806249738SPeter Zijlstra 49991258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 5000e22bee78STejun Heo 5001e22bee78STejun Heo /* 5002eb283428SLai Jiangshan * Call schedule() so that we cross rq->lock and thus can 5003eb283428SLai Jiangshan * guarantee sched callbacks see the %WORKER_UNBOUND flag. 5004eb283428SLai Jiangshan * This is necessary as scheduler callbacks may be invoked 5005eb283428SLai Jiangshan * from other cpus. 5006628c78e7STejun Heo */ 5007628c78e7STejun Heo schedule(); 5008628c78e7STejun Heo 5009628c78e7STejun Heo /* 5010eb283428SLai Jiangshan * Sched callbacks are disabled now. Zap nr_running. 5011eb283428SLai Jiangshan * After this, nr_running stays zero and need_more_worker() 5012eb283428SLai Jiangshan * and keep_working() are always true as long as the 5013eb283428SLai Jiangshan * worklist is not empty. This pool now behaves as an 5014eb283428SLai Jiangshan * unbound (in terms of concurrency management) pool which 5015eb283428SLai Jiangshan * are served by workers tied to the pool. 5016e22bee78STejun Heo */ 5017e19e397aSTejun Heo atomic_set(&pool->nr_running, 0); 5018eb283428SLai Jiangshan 5019eb283428SLai Jiangshan /* 5020eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5021eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5022eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5023eb283428SLai Jiangshan */ 5024a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5025eb283428SLai Jiangshan wake_up_worker(pool); 5026a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5027eb283428SLai Jiangshan } 5028db7bccf4STejun Heo } 5029db7bccf4STejun Heo 5030bd7c089eSTejun Heo /** 5031bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5032bd7c089eSTejun Heo * @pool: pool of interest 5033bd7c089eSTejun Heo * 5034a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5035bd7c089eSTejun Heo */ 5036bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5037bd7c089eSTejun Heo { 5038a9ab775bSTejun Heo struct worker *worker; 5039bd7c089eSTejun Heo 50401258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5041bd7c089eSTejun Heo 5042bd7c089eSTejun Heo /* 5043a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5044a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5045402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5046a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5047a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5048bd7c089eSTejun Heo */ 50495c25b5ffSPeter Zijlstra for_each_pool_worker(worker, pool) { 50505c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 5051a9ab775bSTejun Heo WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5052a9ab775bSTejun Heo pool->attrs->cpumask) < 0); 50535c25b5ffSPeter Zijlstra } 5054a9ab775bSTejun Heo 5055a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5056f7c17d26SWanpeng Li 50573de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5058a9ab775bSTejun Heo 5059da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5060a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5061a9ab775bSTejun Heo 5062a9ab775bSTejun Heo /* 5063a9ab775bSTejun Heo * A bound idle worker should actually be on the runqueue 5064a9ab775bSTejun Heo * of the associated CPU for local wake-ups targeting it to 5065a9ab775bSTejun Heo * work. Kick all idle workers so that they migrate to the 5066a9ab775bSTejun Heo * associated CPU. Doing this in the same loop as 5067a9ab775bSTejun Heo * replacing UNBOUND with REBOUND is safe as no worker will 5068a9ab775bSTejun Heo * be bound before @pool->lock is released. 5069a9ab775bSTejun Heo */ 5070a9ab775bSTejun Heo if (worker_flags & WORKER_IDLE) 5071bd7c089eSTejun Heo wake_up_process(worker->task); 5072bd7c089eSTejun Heo 5073bd7c089eSTejun Heo /* 5074a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5075a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5076a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5077a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5078a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5079a9ab775bSTejun Heo * concurrency management. Note that when or whether 5080a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5081a9ab775bSTejun Heo * 5082c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5083a9ab775bSTejun Heo * tested without holding any lock in 50846d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5085a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5086a9ab775bSTejun Heo * management operations. 5087bd7c089eSTejun Heo */ 5088a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5089a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5090a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5091c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5092bd7c089eSTejun Heo } 5093a9ab775bSTejun Heo 5094a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5095bd7c089eSTejun Heo } 5096bd7c089eSTejun Heo 50977dbc725eSTejun Heo /** 50987dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 50997dbc725eSTejun Heo * @pool: unbound pool of interest 51007dbc725eSTejun Heo * @cpu: the CPU which is coming up 51017dbc725eSTejun Heo * 51027dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 51037dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 51047dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 51057dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 51067dbc725eSTejun Heo */ 51077dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 51087dbc725eSTejun Heo { 51097dbc725eSTejun Heo static cpumask_t cpumask; 51107dbc725eSTejun Heo struct worker *worker; 51117dbc725eSTejun Heo 51121258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 51137dbc725eSTejun Heo 51147dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 51157dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 51167dbc725eSTejun Heo return; 51177dbc725eSTejun Heo 51187dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 51197dbc725eSTejun Heo 51207dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5121da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5122d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 51237dbc725eSTejun Heo } 51247dbc725eSTejun Heo 51257ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 51261da177e4SLinus Torvalds { 51274ce62e9eSTejun Heo struct worker_pool *pool; 51281da177e4SLinus Torvalds 5129f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 51303ce63377STejun Heo if (pool->nr_workers) 51313ce63377STejun Heo continue; 5132051e1850SLai Jiangshan if (!create_worker(pool)) 51337ee681b2SThomas Gleixner return -ENOMEM; 51343af24433SOleg Nesterov } 51357ee681b2SThomas Gleixner return 0; 51367ee681b2SThomas Gleixner } 51371da177e4SLinus Torvalds 51387ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 51397ee681b2SThomas Gleixner { 51407ee681b2SThomas Gleixner struct worker_pool *pool; 51417ee681b2SThomas Gleixner struct workqueue_struct *wq; 51427ee681b2SThomas Gleixner int pi; 51437ee681b2SThomas Gleixner 514468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 51457dbc725eSTejun Heo 51467dbc725eSTejun Heo for_each_pool(pool, pi) { 51471258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 514894cf58bbSTejun Heo 5149f05b558dSLai Jiangshan if (pool->cpu == cpu) 515094cf58bbSTejun Heo rebind_workers(pool); 5151f05b558dSLai Jiangshan else if (pool->cpu < 0) 51527dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 515394cf58bbSTejun Heo 51541258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 515594cf58bbSTejun Heo } 51567dbc725eSTejun Heo 51574c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51584c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51594c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, true); 51604c16bd32STejun Heo 516168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 51627ee681b2SThomas Gleixner return 0; 516365758202STejun Heo } 516465758202STejun Heo 51657ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 516665758202STejun Heo { 51674c16bd32STejun Heo struct workqueue_struct *wq; 51688db25e78STejun Heo 51694c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5170e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5171e8b3f8dbSLai Jiangshan return -1; 5172e8b3f8dbSLai Jiangshan 5173e8b3f8dbSLai Jiangshan unbind_workers(cpu); 51744c16bd32STejun Heo 51754c16bd32STejun Heo /* update NUMA affinity of unbound workqueues */ 51764c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 51774c16bd32STejun Heo list_for_each_entry(wq, &workqueues, list) 51784c16bd32STejun Heo wq_update_unbound_numa(wq, cpu, false); 51794c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 51804c16bd32STejun Heo 51817ee681b2SThomas Gleixner return 0; 518265758202STejun Heo } 518365758202STejun Heo 51842d3854a3SRusty Russell struct work_for_cpu { 5185ed48ece2STejun Heo struct work_struct work; 51862d3854a3SRusty Russell long (*fn)(void *); 51872d3854a3SRusty Russell void *arg; 51882d3854a3SRusty Russell long ret; 51892d3854a3SRusty Russell }; 51902d3854a3SRusty Russell 5191ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 51922d3854a3SRusty Russell { 5193ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5194ed48ece2STejun Heo 51952d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 51962d3854a3SRusty Russell } 51972d3854a3SRusty Russell 51982d3854a3SRusty Russell /** 519922aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 52002d3854a3SRusty Russell * @cpu: the cpu to run on 52012d3854a3SRusty Russell * @fn: the function to run 52022d3854a3SRusty Russell * @arg: the function arg 52032d3854a3SRusty Russell * 520431ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 52056b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5206d185af30SYacine Belkadi * 5207d185af30SYacine Belkadi * Return: The value @fn returns. 52082d3854a3SRusty Russell */ 5209d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 52102d3854a3SRusty Russell { 5211ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 52122d3854a3SRusty Russell 5213ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5214ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 521512997d1aSBjorn Helgaas flush_work(&wfc.work); 5216440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 52172d3854a3SRusty Russell return wfc.ret; 52182d3854a3SRusty Russell } 52192d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 52200e8d6a93SThomas Gleixner 52210e8d6a93SThomas Gleixner /** 52220e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 52230e8d6a93SThomas Gleixner * @cpu: the cpu to run on 52240e8d6a93SThomas Gleixner * @fn: the function to run 52250e8d6a93SThomas Gleixner * @arg: the function argument 52260e8d6a93SThomas Gleixner * 52270e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 52280e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 52290e8d6a93SThomas Gleixner * 52300e8d6a93SThomas Gleixner * Return: The value @fn returns. 52310e8d6a93SThomas Gleixner */ 52320e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 52330e8d6a93SThomas Gleixner { 52340e8d6a93SThomas Gleixner long ret = -ENODEV; 52350e8d6a93SThomas Gleixner 5236ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 52370e8d6a93SThomas Gleixner if (cpu_online(cpu)) 52380e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5239ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 52400e8d6a93SThomas Gleixner return ret; 52410e8d6a93SThomas Gleixner } 52420e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 52432d3854a3SRusty Russell #endif /* CONFIG_SMP */ 52442d3854a3SRusty Russell 5245a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5246e7577c50SRusty Russell 5247a0a1a5fdSTejun Heo /** 5248a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5249a0a1a5fdSTejun Heo * 525058a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5251f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5252706026c2STejun Heo * pool->worklist. 5253a0a1a5fdSTejun Heo * 5254a0a1a5fdSTejun Heo * CONTEXT: 5255a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5256a0a1a5fdSTejun Heo */ 5257a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5258a0a1a5fdSTejun Heo { 525924b8a847STejun Heo struct workqueue_struct *wq; 526024b8a847STejun Heo struct pool_workqueue *pwq; 5261a0a1a5fdSTejun Heo 526268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5263a0a1a5fdSTejun Heo 52646183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5265a0a1a5fdSTejun Heo workqueue_freezing = true; 5266a0a1a5fdSTejun Heo 526724b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5268a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5269699ce097STejun Heo for_each_pwq(pwq, wq) 5270699ce097STejun Heo pwq_adjust_max_active(pwq); 5271a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5272a1056305STejun Heo } 52735bcab335STejun Heo 527468e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5275a0a1a5fdSTejun Heo } 5276a0a1a5fdSTejun Heo 5277a0a1a5fdSTejun Heo /** 527858a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5279a0a1a5fdSTejun Heo * 5280a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5281a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5282a0a1a5fdSTejun Heo * 5283a0a1a5fdSTejun Heo * CONTEXT: 528468e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5285a0a1a5fdSTejun Heo * 5286d185af30SYacine Belkadi * Return: 528758a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 528858a69cb4STejun Heo * is complete. 5289a0a1a5fdSTejun Heo */ 5290a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5291a0a1a5fdSTejun Heo { 5292a0a1a5fdSTejun Heo bool busy = false; 529324b8a847STejun Heo struct workqueue_struct *wq; 529424b8a847STejun Heo struct pool_workqueue *pwq; 5295a0a1a5fdSTejun Heo 529668e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5297a0a1a5fdSTejun Heo 52986183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5299a0a1a5fdSTejun Heo 530024b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 530124b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 530224b8a847STejun Heo continue; 5303a0a1a5fdSTejun Heo /* 5304a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5305a0a1a5fdSTejun Heo * to peek without lock. 5306a0a1a5fdSTejun Heo */ 530724acfb71SThomas Gleixner rcu_read_lock(); 530824b8a847STejun Heo for_each_pwq(pwq, wq) { 53096183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5310112202d9STejun Heo if (pwq->nr_active) { 5311a0a1a5fdSTejun Heo busy = true; 531224acfb71SThomas Gleixner rcu_read_unlock(); 5313a0a1a5fdSTejun Heo goto out_unlock; 5314a0a1a5fdSTejun Heo } 5315a0a1a5fdSTejun Heo } 531624acfb71SThomas Gleixner rcu_read_unlock(); 5317a0a1a5fdSTejun Heo } 5318a0a1a5fdSTejun Heo out_unlock: 531968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5320a0a1a5fdSTejun Heo return busy; 5321a0a1a5fdSTejun Heo } 5322a0a1a5fdSTejun Heo 5323a0a1a5fdSTejun Heo /** 5324a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5325a0a1a5fdSTejun Heo * 5326a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5327706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5328a0a1a5fdSTejun Heo * 5329a0a1a5fdSTejun Heo * CONTEXT: 5330a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5331a0a1a5fdSTejun Heo */ 5332a0a1a5fdSTejun Heo void thaw_workqueues(void) 5333a0a1a5fdSTejun Heo { 533424b8a847STejun Heo struct workqueue_struct *wq; 533524b8a847STejun Heo struct pool_workqueue *pwq; 5336a0a1a5fdSTejun Heo 533768e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5338a0a1a5fdSTejun Heo 5339a0a1a5fdSTejun Heo if (!workqueue_freezing) 5340a0a1a5fdSTejun Heo goto out_unlock; 5341a0a1a5fdSTejun Heo 534274b414eaSLai Jiangshan workqueue_freezing = false; 534324b8a847STejun Heo 534424b8a847STejun Heo /* restore max_active and repopulate worklist */ 534524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5346a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5347699ce097STejun Heo for_each_pwq(pwq, wq) 5348699ce097STejun Heo pwq_adjust_max_active(pwq); 5349a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 535024b8a847STejun Heo } 535124b8a847STejun Heo 5352a0a1a5fdSTejun Heo out_unlock: 535368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5354a0a1a5fdSTejun Heo } 5355a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5356a0a1a5fdSTejun Heo 5357042f7df1SLai Jiangshan static int workqueue_apply_unbound_cpumask(void) 5358042f7df1SLai Jiangshan { 5359042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5360042f7df1SLai Jiangshan int ret = 0; 5361042f7df1SLai Jiangshan struct workqueue_struct *wq; 5362042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5363042f7df1SLai Jiangshan 5364042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5365042f7df1SLai Jiangshan 5366042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5367042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5368042f7df1SLai Jiangshan continue; 5369042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5370042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5371042f7df1SLai Jiangshan continue; 5372042f7df1SLai Jiangshan 5373042f7df1SLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs); 5374042f7df1SLai Jiangshan if (!ctx) { 5375042f7df1SLai Jiangshan ret = -ENOMEM; 5376042f7df1SLai Jiangshan break; 5377042f7df1SLai Jiangshan } 5378042f7df1SLai Jiangshan 5379042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5380042f7df1SLai Jiangshan } 5381042f7df1SLai Jiangshan 5382042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5383042f7df1SLai Jiangshan if (!ret) 5384042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5385042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5386042f7df1SLai Jiangshan } 5387042f7df1SLai Jiangshan 5388042f7df1SLai Jiangshan return ret; 5389042f7df1SLai Jiangshan } 5390042f7df1SLai Jiangshan 5391042f7df1SLai Jiangshan /** 5392042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5393042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5394042f7df1SLai Jiangshan * 5395042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5396042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5397042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5398042f7df1SLai Jiangshan * 539967dc8325SCai Huoqing * Return: 0 - Success 5400042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5401042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5402042f7df1SLai Jiangshan */ 5403042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5404042f7df1SLai Jiangshan { 5405042f7df1SLai Jiangshan int ret = -EINVAL; 5406042f7df1SLai Jiangshan cpumask_var_t saved_cpumask; 5407042f7df1SLai Jiangshan 5408c98a9805STal Shorer /* 5409c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5410c98a9805STal Shorer * If the user wishes to include them, we allow that. 5411c98a9805STal Shorer */ 5412042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5413042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5414a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5415d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5416d25302e4SMenglong Dong ret = 0; 5417d25302e4SMenglong Dong goto out_unlock; 5418d25302e4SMenglong Dong } 5419d25302e4SMenglong Dong 5420d25302e4SMenglong Dong if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) { 5421d25302e4SMenglong Dong ret = -ENOMEM; 5422d25302e4SMenglong Dong goto out_unlock; 5423d25302e4SMenglong Dong } 5424042f7df1SLai Jiangshan 5425042f7df1SLai Jiangshan /* save the old wq_unbound_cpumask. */ 5426042f7df1SLai Jiangshan cpumask_copy(saved_cpumask, wq_unbound_cpumask); 5427042f7df1SLai Jiangshan 5428042f7df1SLai Jiangshan /* update wq_unbound_cpumask at first and apply it to wqs. */ 5429042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, cpumask); 5430042f7df1SLai Jiangshan ret = workqueue_apply_unbound_cpumask(); 5431042f7df1SLai Jiangshan 5432042f7df1SLai Jiangshan /* restore the wq_unbound_cpumask when failed. */ 5433042f7df1SLai Jiangshan if (ret < 0) 5434042f7df1SLai Jiangshan cpumask_copy(wq_unbound_cpumask, saved_cpumask); 5435042f7df1SLai Jiangshan 5436d25302e4SMenglong Dong free_cpumask_var(saved_cpumask); 5437d25302e4SMenglong Dong out_unlock: 5438a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5439042f7df1SLai Jiangshan } 5440042f7df1SLai Jiangshan 5441042f7df1SLai Jiangshan return ret; 5442042f7df1SLai Jiangshan } 5443042f7df1SLai Jiangshan 54446ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 54456ba94429SFrederic Weisbecker /* 54466ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 54476ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 54486ba94429SFrederic Weisbecker * following attributes. 54496ba94429SFrederic Weisbecker * 54506ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 54516ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 54526ba94429SFrederic Weisbecker * 54536ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 54546ba94429SFrederic Weisbecker * 54559a19b463SWang Long * pool_ids RO int : the associated pool IDs for each node 54566ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 54576ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 54589a19b463SWang Long * numa RW bool : whether enable NUMA affinity 54596ba94429SFrederic Weisbecker */ 54606ba94429SFrederic Weisbecker struct wq_device { 54616ba94429SFrederic Weisbecker struct workqueue_struct *wq; 54626ba94429SFrederic Weisbecker struct device dev; 54636ba94429SFrederic Weisbecker }; 54646ba94429SFrederic Weisbecker 54656ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 54666ba94429SFrederic Weisbecker { 54676ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 54686ba94429SFrederic Weisbecker 54696ba94429SFrederic Weisbecker return wq_dev->wq; 54706ba94429SFrederic Weisbecker } 54716ba94429SFrederic Weisbecker 54726ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 54736ba94429SFrederic Weisbecker char *buf) 54746ba94429SFrederic Weisbecker { 54756ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54766ba94429SFrederic Weisbecker 54776ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 54786ba94429SFrederic Weisbecker } 54796ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 54806ba94429SFrederic Weisbecker 54816ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 54826ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 54836ba94429SFrederic Weisbecker { 54846ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54856ba94429SFrederic Weisbecker 54866ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 54876ba94429SFrederic Weisbecker } 54886ba94429SFrederic Weisbecker 54896ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 54906ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 54916ba94429SFrederic Weisbecker size_t count) 54926ba94429SFrederic Weisbecker { 54936ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 54946ba94429SFrederic Weisbecker int val; 54956ba94429SFrederic Weisbecker 54966ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 54976ba94429SFrederic Weisbecker return -EINVAL; 54986ba94429SFrederic Weisbecker 54996ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 55006ba94429SFrederic Weisbecker return count; 55016ba94429SFrederic Weisbecker } 55026ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 55036ba94429SFrederic Weisbecker 55046ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 55056ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 55066ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 55076ba94429SFrederic Weisbecker NULL, 55086ba94429SFrederic Weisbecker }; 55096ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 55106ba94429SFrederic Weisbecker 55116ba94429SFrederic Weisbecker static ssize_t wq_pool_ids_show(struct device *dev, 55126ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55136ba94429SFrederic Weisbecker { 55146ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55156ba94429SFrederic Weisbecker const char *delim = ""; 55166ba94429SFrederic Weisbecker int node, written = 0; 55176ba94429SFrederic Weisbecker 5518ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 551924acfb71SThomas Gleixner rcu_read_lock(); 55206ba94429SFrederic Weisbecker for_each_node(node) { 55216ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, 55226ba94429SFrederic Weisbecker "%s%d:%d", delim, node, 55236ba94429SFrederic Weisbecker unbound_pwq_by_node(wq, node)->pool->id); 55246ba94429SFrederic Weisbecker delim = " "; 55256ba94429SFrederic Weisbecker } 55266ba94429SFrederic Weisbecker written += scnprintf(buf + written, PAGE_SIZE - written, "\n"); 552724acfb71SThomas Gleixner rcu_read_unlock(); 5528ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 55296ba94429SFrederic Weisbecker 55306ba94429SFrederic Weisbecker return written; 55316ba94429SFrederic Weisbecker } 55326ba94429SFrederic Weisbecker 55336ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 55346ba94429SFrederic Weisbecker char *buf) 55356ba94429SFrederic Weisbecker { 55366ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55376ba94429SFrederic Weisbecker int written; 55386ba94429SFrederic Weisbecker 55396ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55406ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 55416ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55426ba94429SFrederic Weisbecker 55436ba94429SFrederic Weisbecker return written; 55446ba94429SFrederic Weisbecker } 55456ba94429SFrederic Weisbecker 55466ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 55476ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 55486ba94429SFrederic Weisbecker { 55496ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 55506ba94429SFrederic Weisbecker 5551899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5552899a94feSLai Jiangshan 5553be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 55546ba94429SFrederic Weisbecker if (!attrs) 55556ba94429SFrederic Weisbecker return NULL; 55566ba94429SFrederic Weisbecker 55576ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 55586ba94429SFrederic Weisbecker return attrs; 55596ba94429SFrederic Weisbecker } 55606ba94429SFrederic Weisbecker 55616ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 55626ba94429SFrederic Weisbecker const char *buf, size_t count) 55636ba94429SFrederic Weisbecker { 55646ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55656ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5566d4d3e257SLai Jiangshan int ret = -ENOMEM; 5567d4d3e257SLai Jiangshan 5568d4d3e257SLai Jiangshan apply_wqattrs_lock(); 55696ba94429SFrederic Weisbecker 55706ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 55716ba94429SFrederic Weisbecker if (!attrs) 5572d4d3e257SLai Jiangshan goto out_unlock; 55736ba94429SFrederic Weisbecker 55746ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 55756ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5576d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 55776ba94429SFrederic Weisbecker else 55786ba94429SFrederic Weisbecker ret = -EINVAL; 55796ba94429SFrederic Weisbecker 5580d4d3e257SLai Jiangshan out_unlock: 5581d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 55826ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 55836ba94429SFrederic Weisbecker return ret ?: count; 55846ba94429SFrederic Weisbecker } 55856ba94429SFrederic Weisbecker 55866ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 55876ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 55886ba94429SFrederic Weisbecker { 55896ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 55906ba94429SFrederic Weisbecker int written; 55916ba94429SFrederic Weisbecker 55926ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 55936ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 55946ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 55956ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 55966ba94429SFrederic Weisbecker return written; 55976ba94429SFrederic Weisbecker } 55986ba94429SFrederic Weisbecker 55996ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 56006ba94429SFrederic Weisbecker struct device_attribute *attr, 56016ba94429SFrederic Weisbecker const char *buf, size_t count) 56026ba94429SFrederic Weisbecker { 56036ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56046ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5605d4d3e257SLai Jiangshan int ret = -ENOMEM; 5606d4d3e257SLai Jiangshan 5607d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56086ba94429SFrederic Weisbecker 56096ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56106ba94429SFrederic Weisbecker if (!attrs) 5611d4d3e257SLai Jiangshan goto out_unlock; 56126ba94429SFrederic Weisbecker 56136ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 56146ba94429SFrederic Weisbecker if (!ret) 5615d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56166ba94429SFrederic Weisbecker 5617d4d3e257SLai Jiangshan out_unlock: 5618d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56196ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56206ba94429SFrederic Weisbecker return ret ?: count; 56216ba94429SFrederic Weisbecker } 56226ba94429SFrederic Weisbecker 56236ba94429SFrederic Weisbecker static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr, 56246ba94429SFrederic Weisbecker char *buf) 56256ba94429SFrederic Weisbecker { 56266ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56276ba94429SFrederic Weisbecker int written; 56286ba94429SFrederic Weisbecker 56296ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 56306ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", 56316ba94429SFrederic Weisbecker !wq->unbound_attrs->no_numa); 56326ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 56336ba94429SFrederic Weisbecker 56346ba94429SFrederic Weisbecker return written; 56356ba94429SFrederic Weisbecker } 56366ba94429SFrederic Weisbecker 56376ba94429SFrederic Weisbecker static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr, 56386ba94429SFrederic Weisbecker const char *buf, size_t count) 56396ba94429SFrederic Weisbecker { 56406ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 56416ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5642d4d3e257SLai Jiangshan int v, ret = -ENOMEM; 5643d4d3e257SLai Jiangshan 5644d4d3e257SLai Jiangshan apply_wqattrs_lock(); 56456ba94429SFrederic Weisbecker 56466ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 56476ba94429SFrederic Weisbecker if (!attrs) 5648d4d3e257SLai Jiangshan goto out_unlock; 56496ba94429SFrederic Weisbecker 56506ba94429SFrederic Weisbecker ret = -EINVAL; 56516ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &v) == 1) { 56526ba94429SFrederic Weisbecker attrs->no_numa = !v; 5653d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 56546ba94429SFrederic Weisbecker } 56556ba94429SFrederic Weisbecker 5656d4d3e257SLai Jiangshan out_unlock: 5657d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 56586ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 56596ba94429SFrederic Weisbecker return ret ?: count; 56606ba94429SFrederic Weisbecker } 56616ba94429SFrederic Weisbecker 56626ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 56636ba94429SFrederic Weisbecker __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL), 56646ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 56656ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 56666ba94429SFrederic Weisbecker __ATTR(numa, 0644, wq_numa_show, wq_numa_store), 56676ba94429SFrederic Weisbecker __ATTR_NULL, 56686ba94429SFrederic Weisbecker }; 56696ba94429SFrederic Weisbecker 56706ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 56716ba94429SFrederic Weisbecker .name = "workqueue", 56726ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 56736ba94429SFrederic Weisbecker }; 56746ba94429SFrederic Weisbecker 5675b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 5676b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 5677b05a7928SFrederic Weisbecker { 5678b05a7928SFrederic Weisbecker int written; 5679b05a7928SFrederic Weisbecker 5680042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 5681b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 5682b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 5683042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5684b05a7928SFrederic Weisbecker 5685b05a7928SFrederic Weisbecker return written; 5686b05a7928SFrederic Weisbecker } 5687b05a7928SFrederic Weisbecker 5688042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 5689042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 5690042f7df1SLai Jiangshan { 5691042f7df1SLai Jiangshan cpumask_var_t cpumask; 5692042f7df1SLai Jiangshan int ret; 5693042f7df1SLai Jiangshan 5694042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 5695042f7df1SLai Jiangshan return -ENOMEM; 5696042f7df1SLai Jiangshan 5697042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 5698042f7df1SLai Jiangshan if (!ret) 5699042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 5700042f7df1SLai Jiangshan 5701042f7df1SLai Jiangshan free_cpumask_var(cpumask); 5702042f7df1SLai Jiangshan return ret ? ret : count; 5703042f7df1SLai Jiangshan } 5704042f7df1SLai Jiangshan 5705b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 5706042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 5707042f7df1SLai Jiangshan wq_unbound_cpumask_store); 5708b05a7928SFrederic Weisbecker 57096ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 57106ba94429SFrederic Weisbecker { 5711b05a7928SFrederic Weisbecker int err; 5712b05a7928SFrederic Weisbecker 5713b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 5714b05a7928SFrederic Weisbecker if (err) 5715b05a7928SFrederic Weisbecker return err; 5716b05a7928SFrederic Weisbecker 5717b05a7928SFrederic Weisbecker return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr); 57186ba94429SFrederic Weisbecker } 57196ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 57206ba94429SFrederic Weisbecker 57216ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 57226ba94429SFrederic Weisbecker { 57236ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 57246ba94429SFrederic Weisbecker 57256ba94429SFrederic Weisbecker kfree(wq_dev); 57266ba94429SFrederic Weisbecker } 57276ba94429SFrederic Weisbecker 57286ba94429SFrederic Weisbecker /** 57296ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 57306ba94429SFrederic Weisbecker * @wq: the workqueue to register 57316ba94429SFrederic Weisbecker * 57326ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 57336ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 57346ba94429SFrederic Weisbecker * which is the preferred method. 57356ba94429SFrederic Weisbecker * 57366ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 57376ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 57386ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 57396ba94429SFrederic Weisbecker * attributes. 57406ba94429SFrederic Weisbecker * 57416ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 57426ba94429SFrederic Weisbecker */ 57436ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 57446ba94429SFrederic Weisbecker { 57456ba94429SFrederic Weisbecker struct wq_device *wq_dev; 57466ba94429SFrederic Weisbecker int ret; 57476ba94429SFrederic Weisbecker 57486ba94429SFrederic Weisbecker /* 5749402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 57506ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 57516ba94429SFrederic Weisbecker * workqueues. 57526ba94429SFrederic Weisbecker */ 57530a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 57546ba94429SFrederic Weisbecker return -EINVAL; 57556ba94429SFrederic Weisbecker 57566ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 57576ba94429SFrederic Weisbecker if (!wq_dev) 57586ba94429SFrederic Weisbecker return -ENOMEM; 57596ba94429SFrederic Weisbecker 57606ba94429SFrederic Weisbecker wq_dev->wq = wq; 57616ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 57626ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 576323217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 57646ba94429SFrederic Weisbecker 57656ba94429SFrederic Weisbecker /* 57666ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 57676ba94429SFrederic Weisbecker * everything is ready. 57686ba94429SFrederic Weisbecker */ 57696ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 57706ba94429SFrederic Weisbecker 57716ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 57726ba94429SFrederic Weisbecker if (ret) { 5773537f4146SArvind Yadav put_device(&wq_dev->dev); 57746ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57756ba94429SFrederic Weisbecker return ret; 57766ba94429SFrederic Weisbecker } 57776ba94429SFrederic Weisbecker 57786ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 57796ba94429SFrederic Weisbecker struct device_attribute *attr; 57806ba94429SFrederic Weisbecker 57816ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 57826ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 57836ba94429SFrederic Weisbecker if (ret) { 57846ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 57856ba94429SFrederic Weisbecker wq->wq_dev = NULL; 57866ba94429SFrederic Weisbecker return ret; 57876ba94429SFrederic Weisbecker } 57886ba94429SFrederic Weisbecker } 57896ba94429SFrederic Weisbecker } 57906ba94429SFrederic Weisbecker 57916ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 57926ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 57936ba94429SFrederic Weisbecker return 0; 57946ba94429SFrederic Weisbecker } 57956ba94429SFrederic Weisbecker 57966ba94429SFrederic Weisbecker /** 57976ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 57986ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 57996ba94429SFrederic Weisbecker * 58006ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 58016ba94429SFrederic Weisbecker */ 58026ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 58036ba94429SFrederic Weisbecker { 58046ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 58056ba94429SFrederic Weisbecker 58066ba94429SFrederic Weisbecker if (!wq->wq_dev) 58076ba94429SFrederic Weisbecker return; 58086ba94429SFrederic Weisbecker 58096ba94429SFrederic Weisbecker wq->wq_dev = NULL; 58106ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 58116ba94429SFrederic Weisbecker } 58126ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 58136ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 58146ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 58156ba94429SFrederic Weisbecker 581682607adcSTejun Heo /* 581782607adcSTejun Heo * Workqueue watchdog. 581882607adcSTejun Heo * 581982607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 582082607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 582182607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 582282607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 582382607adcSTejun Heo * largely opaque. 582482607adcSTejun Heo * 582582607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 582682607adcSTejun Heo * state if some pools failed to make forward progress for a while where 582782607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 582882607adcSTejun Heo * 582982607adcSTejun Heo * This mechanism is controlled through the kernel parameter 583082607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 583182607adcSTejun Heo * corresponding sysfs parameter file. 583282607adcSTejun Heo */ 583382607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 583482607adcSTejun Heo 583582607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 58365cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 583782607adcSTejun Heo 583882607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 583982607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 584082607adcSTejun Heo 584182607adcSTejun Heo static void wq_watchdog_reset_touched(void) 584282607adcSTejun Heo { 584382607adcSTejun Heo int cpu; 584482607adcSTejun Heo 584582607adcSTejun Heo wq_watchdog_touched = jiffies; 584682607adcSTejun Heo for_each_possible_cpu(cpu) 584782607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 584882607adcSTejun Heo } 584982607adcSTejun Heo 58505cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 585182607adcSTejun Heo { 585282607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 585382607adcSTejun Heo bool lockup_detected = false; 5854940d71c6SSergey Senozhatsky unsigned long now = jiffies; 585582607adcSTejun Heo struct worker_pool *pool; 585682607adcSTejun Heo int pi; 585782607adcSTejun Heo 585882607adcSTejun Heo if (!thresh) 585982607adcSTejun Heo return; 586082607adcSTejun Heo 586182607adcSTejun Heo rcu_read_lock(); 586282607adcSTejun Heo 586382607adcSTejun Heo for_each_pool(pool, pi) { 586482607adcSTejun Heo unsigned long pool_ts, touched, ts; 586582607adcSTejun Heo 586682607adcSTejun Heo if (list_empty(&pool->worklist)) 586782607adcSTejun Heo continue; 586882607adcSTejun Heo 5869940d71c6SSergey Senozhatsky /* 5870940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 5871940d71c6SSergey Senozhatsky * the watchdog like a stall. 5872940d71c6SSergey Senozhatsky */ 5873940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 5874940d71c6SSergey Senozhatsky 587582607adcSTejun Heo /* get the latest of pool and touched timestamps */ 587689e28ce6SWang Qing if (pool->cpu >= 0) 587789e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 587889e28ce6SWang Qing else 587982607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 588089e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 588182607adcSTejun Heo 588282607adcSTejun Heo if (time_after(pool_ts, touched)) 588382607adcSTejun Heo ts = pool_ts; 588482607adcSTejun Heo else 588582607adcSTejun Heo ts = touched; 588682607adcSTejun Heo 588782607adcSTejun Heo /* did we stall? */ 5888940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 588982607adcSTejun Heo lockup_detected = true; 589082607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 589182607adcSTejun Heo pr_cont_pool_info(pool); 589282607adcSTejun Heo pr_cont(" stuck for %us!\n", 5893940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 589482607adcSTejun Heo } 589582607adcSTejun Heo } 589682607adcSTejun Heo 589782607adcSTejun Heo rcu_read_unlock(); 589882607adcSTejun Heo 589982607adcSTejun Heo if (lockup_detected) 590055df0933SImran Khan show_all_workqueues(); 590182607adcSTejun Heo 590282607adcSTejun Heo wq_watchdog_reset_touched(); 590382607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 590482607adcSTejun Heo } 590582607adcSTejun Heo 5906cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 590782607adcSTejun Heo { 590882607adcSTejun Heo if (cpu >= 0) 590982607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 591089e28ce6SWang Qing 591182607adcSTejun Heo wq_watchdog_touched = jiffies; 591282607adcSTejun Heo } 591382607adcSTejun Heo 591482607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 591582607adcSTejun Heo { 591682607adcSTejun Heo wq_watchdog_thresh = 0; 591782607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 591882607adcSTejun Heo 591982607adcSTejun Heo if (thresh) { 592082607adcSTejun Heo wq_watchdog_thresh = thresh; 592182607adcSTejun Heo wq_watchdog_reset_touched(); 592282607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 592382607adcSTejun Heo } 592482607adcSTejun Heo } 592582607adcSTejun Heo 592682607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 592782607adcSTejun Heo const struct kernel_param *kp) 592882607adcSTejun Heo { 592982607adcSTejun Heo unsigned long thresh; 593082607adcSTejun Heo int ret; 593182607adcSTejun Heo 593282607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 593382607adcSTejun Heo if (ret) 593482607adcSTejun Heo return ret; 593582607adcSTejun Heo 593682607adcSTejun Heo if (system_wq) 593782607adcSTejun Heo wq_watchdog_set_thresh(thresh); 593882607adcSTejun Heo else 593982607adcSTejun Heo wq_watchdog_thresh = thresh; 594082607adcSTejun Heo 594182607adcSTejun Heo return 0; 594282607adcSTejun Heo } 594382607adcSTejun Heo 594482607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 594582607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 594682607adcSTejun Heo .get = param_get_ulong, 594782607adcSTejun Heo }; 594882607adcSTejun Heo 594982607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 595082607adcSTejun Heo 0644); 595182607adcSTejun Heo 595282607adcSTejun Heo static void wq_watchdog_init(void) 595382607adcSTejun Heo { 59545cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 595582607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 595682607adcSTejun Heo } 595782607adcSTejun Heo 595882607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 595982607adcSTejun Heo 596082607adcSTejun Heo static inline void wq_watchdog_init(void) { } 596182607adcSTejun Heo 596282607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 596382607adcSTejun Heo 5964bce90380STejun Heo static void __init wq_numa_init(void) 5965bce90380STejun Heo { 5966bce90380STejun Heo cpumask_var_t *tbl; 5967bce90380STejun Heo int node, cpu; 5968bce90380STejun Heo 5969bce90380STejun Heo if (num_possible_nodes() <= 1) 5970bce90380STejun Heo return; 5971bce90380STejun Heo 5972d55262c4STejun Heo if (wq_disable_numa) { 5973d55262c4STejun Heo pr_info("workqueue: NUMA affinity support disabled\n"); 5974d55262c4STejun Heo return; 5975d55262c4STejun Heo } 5976d55262c4STejun Heo 5977f728c4a9SZhen Lei for_each_possible_cpu(cpu) { 5978f728c4a9SZhen Lei if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) { 5979f728c4a9SZhen Lei pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu); 5980f728c4a9SZhen Lei return; 5981f728c4a9SZhen Lei } 5982f728c4a9SZhen Lei } 5983f728c4a9SZhen Lei 5984be69d00dSThomas Gleixner wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(); 59854c16bd32STejun Heo BUG_ON(!wq_update_unbound_numa_attrs_buf); 59864c16bd32STejun Heo 5987bce90380STejun Heo /* 5988bce90380STejun Heo * We want masks of possible CPUs of each node which isn't readily 5989bce90380STejun Heo * available. Build one from cpu_to_node() which should have been 5990bce90380STejun Heo * fully initialized by now. 5991bce90380STejun Heo */ 59926396bb22SKees Cook tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL); 5993bce90380STejun Heo BUG_ON(!tbl); 5994bce90380STejun Heo 5995bce90380STejun Heo for_each_node(node) 59965a6024f1SYasuaki Ishimatsu BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL, 59971be0c25dSTejun Heo node_online(node) ? node : NUMA_NO_NODE)); 5998bce90380STejun Heo 5999bce90380STejun Heo for_each_possible_cpu(cpu) { 6000bce90380STejun Heo node = cpu_to_node(cpu); 6001bce90380STejun Heo cpumask_set_cpu(cpu, tbl[node]); 6002bce90380STejun Heo } 6003bce90380STejun Heo 6004bce90380STejun Heo wq_numa_possible_cpumask = tbl; 6005bce90380STejun Heo wq_numa_enabled = true; 6006bce90380STejun Heo } 6007bce90380STejun Heo 60083347fa09STejun Heo /** 60093347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 60103347fa09STejun Heo * 60113347fa09STejun Heo * This is the first half of two-staged workqueue subsystem initialization 60123347fa09STejun Heo * and invoked as soon as the bare basics - memory allocation, cpumasks and 60133347fa09STejun Heo * idr are up. It sets up all the data structures and system workqueues 60143347fa09STejun Heo * and allows early boot code to create workqueues and queue/cancel work 60153347fa09STejun Heo * items. Actual work item execution starts only after kthreads can be 60163347fa09STejun Heo * created and scheduled right before early initcalls. 60173347fa09STejun Heo */ 60182333e829SYu Chen void __init workqueue_init_early(void) 60191da177e4SLinus Torvalds { 60207a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 60211bda3f80SFrederic Weisbecker int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ; 60227a4e344cSTejun Heo int i, cpu; 6023c34056a3STejun Heo 602410cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6025e904e6c2STejun Heo 6026b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 60271bda3f80SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags)); 6028b05a7928SFrederic Weisbecker 6029e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6030e904e6c2STejun Heo 6031706026c2STejun Heo /* initialize CPU pools */ 603229c91e99STejun Heo for_each_possible_cpu(cpu) { 60334ce62e9eSTejun Heo struct worker_pool *pool; 60348b03ae3cSTejun Heo 60357a4e344cSTejun Heo i = 0; 6036f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 60377a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6038ec22ca5eSTejun Heo pool->cpu = cpu; 60397a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 60407a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6041f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 60427a4e344cSTejun Heo 60439daf9e67STejun Heo /* alloc pool ID */ 604468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 60459daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 604668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 60474ce62e9eSTejun Heo } 60488b03ae3cSTejun Heo } 60498b03ae3cSTejun Heo 60508a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 605129c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 605229c91e99STejun Heo struct workqueue_attrs *attrs; 605329c91e99STejun Heo 6054be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 605529c91e99STejun Heo attrs->nice = std_nice[i]; 605629c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 60578a2b7538STejun Heo 60588a2b7538STejun Heo /* 60598a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 60608a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 60618a2b7538STejun Heo * Turn off NUMA so that dfl_pwq is used for all nodes. 60628a2b7538STejun Heo */ 6063be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 60648a2b7538STejun Heo attrs->nice = std_nice[i]; 60658a2b7538STejun Heo attrs->no_numa = true; 60668a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 606729c91e99STejun Heo } 606829c91e99STejun Heo 6069d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 60701aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6071d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6072f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6073f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 607424d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 607524d51addSTejun Heo WQ_FREEZABLE, 0); 60760668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 60770668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 60780668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 60790668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 60800668106cSViresh Kumar 0); 60811aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 60820668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 60830668106cSViresh Kumar !system_power_efficient_wq || 60840668106cSViresh Kumar !system_freezable_power_efficient_wq); 60853347fa09STejun Heo } 60863347fa09STejun Heo 60873347fa09STejun Heo /** 60883347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 60893347fa09STejun Heo * 60903347fa09STejun Heo * This is the latter half of two-staged workqueue subsystem initialization 60913347fa09STejun Heo * and invoked as soon as kthreads can be created and scheduled. 60923347fa09STejun Heo * Workqueues have been created and work items queued on them, but there 60933347fa09STejun Heo * are no kworkers executing the work items yet. Populate the worker pools 60943347fa09STejun Heo * with the initial workers and enable future kworker creations. 60953347fa09STejun Heo */ 60962333e829SYu Chen void __init workqueue_init(void) 60973347fa09STejun Heo { 60982186d9f9STejun Heo struct workqueue_struct *wq; 60993347fa09STejun Heo struct worker_pool *pool; 61003347fa09STejun Heo int cpu, bkt; 61013347fa09STejun Heo 61022186d9f9STejun Heo /* 61032186d9f9STejun Heo * It'd be simpler to initialize NUMA in workqueue_init_early() but 61042186d9f9STejun Heo * CPU to node mapping may not be available that early on some 61052186d9f9STejun Heo * archs such as power and arm64. As per-cpu pools created 61062186d9f9STejun Heo * previously could be missing node hint and unbound pools NUMA 61072186d9f9STejun Heo * affinity, fix them up. 610840c17f75STejun Heo * 610940c17f75STejun Heo * Also, while iterating workqueues, create rescuers if requested. 61102186d9f9STejun Heo */ 61112186d9f9STejun Heo wq_numa_init(); 61122186d9f9STejun Heo 61132186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 61142186d9f9STejun Heo 61152186d9f9STejun Heo for_each_possible_cpu(cpu) { 61162186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 61172186d9f9STejun Heo pool->node = cpu_to_node(cpu); 61182186d9f9STejun Heo } 61192186d9f9STejun Heo } 61202186d9f9STejun Heo 612140c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 61222186d9f9STejun Heo wq_update_unbound_numa(wq, smp_processor_id(), true); 612340c17f75STejun Heo WARN(init_rescuer(wq), 612440c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 612540c17f75STejun Heo wq->name); 612640c17f75STejun Heo } 61272186d9f9STejun Heo 61282186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 61292186d9f9STejun Heo 61303347fa09STejun Heo /* create the initial workers */ 61313347fa09STejun Heo for_each_online_cpu(cpu) { 61323347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 61333347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 61343347fa09STejun Heo BUG_ON(!create_worker(pool)); 61353347fa09STejun Heo } 61363347fa09STejun Heo } 61373347fa09STejun Heo 61383347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 61393347fa09STejun Heo BUG_ON(!create_worker(pool)); 61403347fa09STejun Heo 61413347fa09STejun Heo wq_online = true; 614282607adcSTejun Heo wq_watchdog_init(); 61431da177e4SLinus Torvalds } 6144